Adding upstream version 115.7.0esr.upstream/115.7.0esr

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

12 files changed, 16573 insertions, 0 deletions

diff --git a/gfx/wr/swgl/src/blend.h b/gfx/wr/swgl/src/blend.h
new file mode 100644
index 0000000000..8bc1c93994
--- /dev/null
+++ b/gfx/wr/swgl/src/blend.h
@@ -0,0 +1,864 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+static ALWAYS_INLINE HalfRGBA8 packRGBA8(I32 a, I32 b) {
+#if USE_SSE2
+  return _mm_packs_epi32(a, b);
+#elif USE_NEON
+  return vcombine_u16(vqmovun_s32(a), vqmovun_s32(b));
+#else
+  return CONVERT(combine(a, b), HalfRGBA8);
+#endif
+}
+
+static ALWAYS_INLINE WideRGBA8 pack_pixels_RGBA8(const vec4& v,
+                                                 float scale = 255.0f) {
+  ivec4 i = round_pixel(v, scale);
+  HalfRGBA8 xz = packRGBA8(i.z, i.x);
+  HalfRGBA8 yw = packRGBA8(i.y, i.w);
+  HalfRGBA8 xyzwl = zipLow(xz, yw);
+  HalfRGBA8 xyzwh = zipHigh(xz, yw);
+  HalfRGBA8 lo = zip2Low(xyzwl, xyzwh);
+  HalfRGBA8 hi = zip2High(xyzwl, xyzwh);
+  return combine(lo, hi);
+}
+
+static ALWAYS_INLINE WideRGBA8 pack_pixels_RGBA8(Float alpha,
+                                                 float scale = 255.0f) {
+  I32 i = round_pixel(alpha, scale);
+  HalfRGBA8 c = packRGBA8(i, i);
+  c = zipLow(c, c);
+  return zip(c, c);
+}
+
+static ALWAYS_INLINE WideRGBA8 pack_pixels_RGBA8(float alpha,
+                                                 float scale = 255.0f) {
+  I32 i = round_pixel(alpha, scale);
+  return repeat2(packRGBA8(i, i));
+}
+
+UNUSED static ALWAYS_INLINE WideRGBA8 pack_pixels_RGBA8(const vec4_scalar& v,
+                                                        float scale = 255.0f) {
+  I32 i = round_pixel((Float){v.z, v.y, v.x, v.w}, scale);
+  return repeat2(packRGBA8(i, i));
+}
+
+static ALWAYS_INLINE WideRGBA8 pack_pixels_RGBA8() {
+  return pack_pixels_RGBA8(fragment_shader->gl_FragColor);
+}
+
+static ALWAYS_INLINE WideRGBA8 pack_pixels_RGBA8(WideRGBA32F v,
+                                                 float scale = 255.0f) {
+  ivec4 i = round_pixel(bit_cast<vec4>(v), scale);
+  return combine(packRGBA8(i.x, i.y), packRGBA8(i.z, i.w));
+}
+
+static ALWAYS_INLINE WideR8 packR8(I32 a) {
+#if USE_SSE2
+  return lowHalf(bit_cast<V8<uint16_t>>(_mm_packs_epi32(a, a)));
+#elif USE_NEON
+  return vqmovun_s32(a);
+#else
+  return CONVERT(a, WideR8);
+#endif
+}
+
+static ALWAYS_INLINE WideR8 pack_pixels_R8(Float c, float scale = 255.0f) {
+  return packR8(round_pixel(c, scale));
+}
+
+static ALWAYS_INLINE WideR8 pack_pixels_R8() {
+  return pack_pixels_R8(fragment_shader->gl_FragColor.x);
+}
+
+// Load a partial span > 0 and < 4 pixels.
+template <typename V, typename P>
+static ALWAYS_INLINE V partial_load_span(const P* src, int span) {
+  return bit_cast<V>(
+      (span >= 2
+           ? combine(unaligned_load<V2<P>>(src),
+                     V2<P>{span > 2 ? unaligned_load<P>(src + 2) : P(0), 0})
+           : V4<P>{unaligned_load<P>(src), 0, 0, 0}));
+}
+
+// Store a partial span > 0 and < 4 pixels.
+template <typename V, typename P>
+static ALWAYS_INLINE void partial_store_span(P* dst, V src, int span) {
+  auto pixels = bit_cast<V4<P>>(src);
+  if (span >= 2) {
+    unaligned_store(dst, lowHalf(pixels));
+    if (span > 2) {
+      unaligned_store(dst + 2, pixels.z);
+    }
+  } else {
+    unaligned_store(dst, pixels.x);
+  }
+}
+
+// Dispatcher that chooses when to load a full or partial span
+template <typename V, typename P>
+static ALWAYS_INLINE V load_span(const P* src, int span) {
+  if (span >= 4) {
+    return unaligned_load<V, P>(src);
+  } else {
+    return partial_load_span<V, P>(src, span);
+  }
+}
+
+// Dispatcher that chooses when to store a full or partial span
+template <typename V, typename P>
+static ALWAYS_INLINE void store_span(P* dst, V src, int span) {
+  if (span >= 4) {
+    unaligned_store<V, P>(dst, src);
+  } else {
+    partial_store_span<V, P>(dst, src, span);
+  }
+}
+
+template <typename T>
+static ALWAYS_INLINE T muldiv256(T x, T y) {
+  return (x * y) >> 8;
+}
+
+// (x*y + x) >> 8, cheap approximation of (x*y) / 255
+template <typename T>
+static ALWAYS_INLINE T muldiv255(T x, T y) {
+  return (x * y + x) >> 8;
+}
+
+template <typename V>
+static ALWAYS_INLINE WideRGBA8 pack_span(uint32_t*, const V& v,
+                                         float scale = 255.0f) {
+  return pack_pixels_RGBA8(v, scale);
+}
+
+template <typename C>
+static ALWAYS_INLINE WideR8 pack_span(uint8_t*, C c, float scale = 255.0f) {
+  return pack_pixels_R8(c, scale);
+}
+
+// Helper functions to apply a color modulus when available.
+struct NoColor {};
+
+template <typename P>
+static ALWAYS_INLINE P applyColor(P src, NoColor) {
+  return src;
+}
+
+struct InvertColor {};
+
+template <typename P>
+static ALWAYS_INLINE P applyColor(P src, InvertColor) {
+  return 255 - src;
+}
+
+template <typename P>
+static ALWAYS_INLINE P applyColor(P src, P color) {
+  return muldiv255(color, src);
+}
+
+static ALWAYS_INLINE WideRGBA8 applyColor(PackedRGBA8 src, WideRGBA8 color) {
+  return applyColor(unpack(src), color);
+}
+
+template <typename P, typename C>
+static ALWAYS_INLINE auto packColor(P* buf, C color) {
+  return pack_span(buf, color, 255.0f);
+}
+
+template <typename P>
+static ALWAYS_INLINE NoColor packColor(UNUSED P* buf, NoColor noColor) {
+  return noColor;
+}
+
+template <typename P>
+static ALWAYS_INLINE InvertColor packColor(UNUSED P* buf,
+                                           InvertColor invertColor) {
+  return invertColor;
+}
+
+// Single argument variation that takes an explicit destination buffer type.
+template <typename P, typename C>
+static ALWAYS_INLINE auto packColor(C color) {
+  // Just pass in a typed null pointer, as the pack routines never use the
+  // pointer's value, just its type.
+  return packColor((P*)0, color);
+}
+
+// Byte-wise addition for when x or y is a signed 8-bit value stored in the
+// low byte of a larger type T only with zeroed-out high bits, where T is
+// greater than 8 bits, i.e. uint16_t. This can result when muldiv255 is used
+// upon signed operands, using up all the precision in a 16 bit integer, and
+// potentially losing the sign bit in the last >> 8 shift. Due to the
+// properties of two's complement arithmetic, even though we've discarded the
+// sign bit, we can still represent a negative number under addition (without
+// requiring any extra sign bits), just that any negative number will behave
+// like a large unsigned number under addition, generating a single carry bit
+// on overflow that we need to discard. Thus, just doing a byte-wise add will
+// overflow without the troublesome carry, giving us only the remaining 8 low
+// bits we actually need while keeping the high bits at zero.
+template <typename T>
+static ALWAYS_INLINE T addlow(T x, T y) {
+  typedef VectorType<uint8_t, sizeof(T)> bytes;
+  return bit_cast<T>(bit_cast<bytes>(x) + bit_cast<bytes>(y));
+}
+
+// Replace color components of each pixel with the pixel's alpha values.
+template <typename T>
+static ALWAYS_INLINE T alphas(T c) {
+  return SHUFFLE(c, c, 3, 3, 3, 3, 7, 7, 7, 7, 11, 11, 11, 11, 15, 15, 15, 15);
+}
+
+// Replace the alpha values of the first vector with alpha values from the
+// second, while leaving the color components unmodified.
+template <typename T>
+static ALWAYS_INLINE T set_alphas(T c, T a) {
+  return SHUFFLE(c, a, 0, 1, 2, 19, 4, 5, 6, 23, 8, 9, 10, 27, 12, 13, 14, 31);
+}
+
+// Miscellaneous helper functions for working with packed RGBA8 data.
+static ALWAYS_INLINE HalfRGBA8 if_then_else(V8<int16_t> c, HalfRGBA8 t,
+                                            HalfRGBA8 e) {
+  return bit_cast<HalfRGBA8>((c & t) | (~c & e));
+}
+
+template <typename T, typename C, int N>
+static ALWAYS_INLINE VectorType<T, N> if_then_else(VectorType<C, N> c,
+                                                   VectorType<T, N> t,
+                                                   VectorType<T, N> e) {
+  return combine(if_then_else(lowHalf(c), lowHalf(t), lowHalf(e)),
+                 if_then_else(highHalf(c), highHalf(t), highHalf(e)));
+}
+
+static ALWAYS_INLINE HalfRGBA8 min(HalfRGBA8 x, HalfRGBA8 y) {
+#if USE_SSE2
+  return bit_cast<HalfRGBA8>(
+      _mm_min_epi16(bit_cast<V8<int16_t>>(x), bit_cast<V8<int16_t>>(y)));
+#elif USE_NEON
+  return vminq_u16(x, y);
+#else
+  return if_then_else(x < y, x, y);
+#endif
+}
+
+template <typename T, int N>
+static ALWAYS_INLINE VectorType<T, N> min(VectorType<T, N> x,
+                                          VectorType<T, N> y) {
+  return combine(min(lowHalf(x), lowHalf(y)), min(highHalf(x), highHalf(y)));
+}
+
+static ALWAYS_INLINE HalfRGBA8 max(HalfRGBA8 x, HalfRGBA8 y) {
+#if USE_SSE2
+  return bit_cast<HalfRGBA8>(
+      _mm_max_epi16(bit_cast<V8<int16_t>>(x), bit_cast<V8<int16_t>>(y)));
+#elif USE_NEON
+  return vmaxq_u16(x, y);
+#else
+  return if_then_else(x > y, x, y);
+#endif
+}
+
+template <typename T, int N>
+static ALWAYS_INLINE VectorType<T, N> max(VectorType<T, N> x,
+                                          VectorType<T, N> y) {
+  return combine(max(lowHalf(x), lowHalf(y)), max(highHalf(x), highHalf(y)));
+}
+
+template <typename T, int N>
+static ALWAYS_INLINE VectorType<T, N> recip(VectorType<T, N> v) {
+  return combine(recip(lowHalf(v)), recip(highHalf(v)));
+}
+
+// Helper to get the reciprocal if the value is non-zero, or otherwise default
+// to the supplied fallback value.
+template <typename V>
+static ALWAYS_INLINE V recip_or(V v, float f) {
+  return if_then_else(v != V(0.0f), recip(v), V(f));
+}
+
+template <typename T, int N>
+static ALWAYS_INLINE VectorType<T, N> inversesqrt(VectorType<T, N> v) {
+  return combine(inversesqrt(lowHalf(v)), inversesqrt(highHalf(v)));
+}
+
+// Extract the alpha components so that we can cheaply calculate the reciprocal
+// on a single SIMD register. Then multiply the duplicated alpha reciprocal with
+// the pixel data. 0 alpha is treated as transparent black.
+static ALWAYS_INLINE WideRGBA32F unpremultiply(WideRGBA32F v) {
+  Float a = recip_or((Float){v[3], v[7], v[11], v[15]}, 0.0f);
+  return v * a.xxxxyyyyzzzzwwww;
+}
+
+// Packed RGBA32F data is AoS in BGRA order. Transpose it to SoA and swizzle to
+// RGBA to unpack.
+static ALWAYS_INLINE vec4 unpack(PackedRGBA32F c) {
+  return bit_cast<vec4>(
+      SHUFFLE(c, c, 2, 6, 10, 14, 1, 5, 9, 13, 0, 4, 8, 12, 3, 7, 11, 15));
+}
+
+// The following lum/sat functions mostly follow the KHR_blend_equation_advanced
+// specification but are rearranged to work on premultiplied data.
+static ALWAYS_INLINE Float lumv3(vec3 v) {
+  return v.x * 0.30f + v.y * 0.59f + v.z * 0.11f;
+}
+
+static ALWAYS_INLINE Float minv3(vec3 v) { return min(min(v.x, v.y), v.z); }
+
+static ALWAYS_INLINE Float maxv3(vec3 v) { return max(max(v.x, v.y), v.z); }
+
+static inline vec3 clip_color(vec3 v, Float lum, Float alpha) {
+  Float mincol = max(-minv3(v), lum);
+  Float maxcol = max(maxv3(v), alpha - lum);
+  return lum + v * (lum * (alpha - lum) * recip_or(mincol * maxcol, 0.0f));
+}
+
+static inline vec3 set_lum(vec3 base, vec3 ref, Float alpha) {
+  return clip_color(base - lumv3(base), lumv3(ref), alpha);
+}
+
+static inline vec3 set_lum_sat(vec3 base, vec3 sref, vec3 lref, Float alpha) {
+  vec3 diff = base - minv3(base);
+  Float sbase = maxv3(diff);
+  Float ssat = maxv3(sref) - minv3(sref);
+  // The sbase range is rescaled to ssat. If sbase has 0 extent, then rescale
+  // to black, as per specification.
+  return set_lum(diff * ssat * recip_or(sbase, 0.0f), lref, alpha);
+}
+
+// Flags the reflect the current blend-stage clipping to be applied.
+enum SWGLClipFlag {
+  SWGL_CLIP_FLAG_MASK = 1 << 0,
+  SWGL_CLIP_FLAG_AA = 1 << 1,
+  SWGL_CLIP_FLAG_BLEND_OVERRIDE = 1 << 2,
+};
+static int swgl_ClipFlags = 0;
+static BlendKey swgl_BlendOverride = BLEND_KEY_NONE;
+static WideRGBA8 swgl_BlendColorRGBA8 = {0};
+static WideRGBA8 swgl_BlendAlphaRGBA8 = {0};
+
+// A pointer into the color buffer for the start of the span.
+static void* swgl_SpanBuf = nullptr;
+// A pointer into the clip mask for the start of the span.
+static uint8_t* swgl_ClipMaskBuf = nullptr;
+
+static ALWAYS_INLINE WideR8 expand_mask(UNUSED uint8_t* buf, WideR8 mask) {
+  return mask;
+}
+static ALWAYS_INLINE WideRGBA8 expand_mask(UNUSED uint32_t* buf, WideR8 mask) {
+  WideRG8 maskRG = zip(mask, mask);
+  return zip(maskRG, maskRG);
+}
+
+// Loads a chunk of clip masks. The current pointer into the color buffer is
+// used to reconstruct the relative position within the span. From there, the
+// pointer into the clip mask can be generated from the start of the clip mask
+// span.
+template <typename P>
+static ALWAYS_INLINE uint8_t* get_clip_mask(P* buf) {
+  return &swgl_ClipMaskBuf[buf - (P*)swgl_SpanBuf];
+}
+
+template <typename P>
+static ALWAYS_INLINE auto load_clip_mask(P* buf, int span)
+    -> decltype(expand_mask(buf, 0)) {
+  return expand_mask(buf,
+                     unpack(load_span<PackedR8>(get_clip_mask(buf), span)));
+}
+
+// Temporarily removes masking from the blend stage, assuming the caller will
+// handle it.
+static ALWAYS_INLINE void override_clip_mask() {
+  blend_key = BlendKey(blend_key - MASK_BLEND_KEY_NONE);
+}
+
+// Restores masking to the blend stage, assuming it was previously overridden.
+static ALWAYS_INLINE void restore_clip_mask() {
+  blend_key = BlendKey(MASK_BLEND_KEY_NONE + blend_key);
+}
+
+// A pointer to the start of the opaque destination region of the span for AA.
+static const uint8_t* swgl_OpaqueStart = nullptr;
+// The size, in bytes, of the opaque region.
+static uint32_t swgl_OpaqueSize = 0;
+// AA coverage distance offsets for the left and right edges.
+static Float swgl_LeftAADist = 0.0f;
+static Float swgl_RightAADist = 0.0f;
+// AA coverage slope values used for accumulating coverage for each step.
+static Float swgl_AASlope = 0.0f;
+
+// Get the amount of pixels we need to process before the start of the opaque
+// region.
+template <typename P>
+static ALWAYS_INLINE int get_aa_opaque_start(P* buf) {
+  return max(int((P*)swgl_OpaqueStart - buf), 0);
+}
+
+// Assuming we are already in the opaque part of the span, return the remaining
+// size of the opaque part.
+template <typename P>
+static ALWAYS_INLINE int get_aa_opaque_size(P* buf) {
+  return max(int((P*)&swgl_OpaqueStart[swgl_OpaqueSize] - buf), 0);
+}
+
+// Temporarily removes anti-aliasing from the blend stage, assuming the caller
+// will handle it.
+static ALWAYS_INLINE void override_aa() {
+  blend_key = BlendKey(blend_key - AA_BLEND_KEY_NONE);
+}
+
+// Restores anti-aliasing to the blend stage, assuming it was previously
+// overridden.
+static ALWAYS_INLINE void restore_aa() {
+  blend_key = BlendKey(AA_BLEND_KEY_NONE + blend_key);
+}
+
+static PREFER_INLINE WideRGBA8 blend_pixels(uint32_t* buf, PackedRGBA8 pdst,
+                                            WideRGBA8 src, int span = 4) {
+  WideRGBA8 dst = unpack(pdst);
+  const WideRGBA8 RGB_MASK = {0xFFFF, 0xFFFF, 0xFFFF, 0,      0xFFFF, 0xFFFF,
+                              0xFFFF, 0,      0xFFFF, 0xFFFF, 0xFFFF, 0,
+                              0xFFFF, 0xFFFF, 0xFFFF, 0};
+  const WideRGBA8 ALPHA_MASK = {0, 0, 0, 0xFFFF, 0, 0, 0, 0xFFFF,
+                                0, 0, 0, 0xFFFF, 0, 0, 0, 0xFFFF};
+  const WideRGBA8 ALPHA_OPAQUE = {0, 0, 0, 255, 0, 0, 0, 255,
+                                  0, 0, 0, 255, 0, 0, 0, 255};
+
+// clang-format off
+  // Computes AA for the given pixel based on the offset of the pixel within
+  // destination row. Given the initial coverage offsets for the left and right
+  // edges, the offset is scaled by the slope and accumulated to find the
+  // minimum coverage value for the pixel. A final weight is generated that
+  // can be used to scale the source pixel.
+#define DO_AA(format, body)                                   \
+  do {                                                        \
+    int offset = int((const uint8_t*)buf - swgl_OpaqueStart); \
+    if (uint32_t(offset) >= swgl_OpaqueSize) {                \
+      Float delta = swgl_AASlope * float(offset);             \
+      Float dist = clamp(min(swgl_LeftAADist + delta.x,       \
+                             swgl_RightAADist + delta.y),     \
+                         0.0f, 256.0f);                       \
+      auto aa = pack_pixels_##format(dist, 1.0f);             \
+      body;                                                   \
+    }                                                         \
+  } while (0)
+
+  // Each blend case is preceded by the MASK_ variant. The MASK_ case first
+  // loads the mask values and multiplies the source value by them. After, it
+  // falls through to the normal blending case using the masked source. The
+  // AA_ variations may further precede the blend cases, in which case the
+  // source value is further modified before use.
+#define BLEND_CASE_KEY(key)                          \
+  case AA_##key:                                     \
+    DO_AA(RGBA8, src = muldiv256(src, aa));          \
+    goto key;                                        \
+  case AA_MASK_##key:                                \
+    DO_AA(RGBA8, src = muldiv256(src, aa));          \
+    FALLTHROUGH;                                     \
+  case MASK_##key:                                   \
+    src = muldiv255(src, load_clip_mask(buf, span)); \
+    FALLTHROUGH;                                     \
+  case key: key
+
+#define BLEND_CASE(...) BLEND_CASE_KEY(BLEND_KEY(__VA_ARGS__))
+
+  switch (blend_key) {
+  BLEND_CASE(GL_ONE, GL_ZERO):
+    return src;
+  BLEND_CASE(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, GL_ONE,
+                  GL_ONE_MINUS_SRC_ALPHA):
+    // dst + src.a*(src.rgb1 - dst)
+    // use addlow for signed overflow
+    return addlow(dst, muldiv255(alphas(src), (src | ALPHA_OPAQUE) - dst));
+  BLEND_CASE(GL_ONE, GL_ONE_MINUS_SRC_ALPHA):
+    return src + dst - muldiv255(dst, alphas(src));
+  BLEND_CASE(GL_ZERO, GL_ONE_MINUS_SRC_COLOR):
+    return dst - muldiv255(dst, src);
+  BLEND_CASE(GL_ZERO, GL_ONE_MINUS_SRC_COLOR, GL_ZERO, GL_ONE):
+    return dst - (muldiv255(dst, src) & RGB_MASK);
+  BLEND_CASE(GL_ZERO, GL_ONE_MINUS_SRC_ALPHA):
+    return dst - muldiv255(dst, alphas(src));
+  BLEND_CASE(GL_ZERO, GL_SRC_COLOR):
+    return muldiv255(src, dst);
+  BLEND_CASE(GL_ONE, GL_ONE):
+    return src + dst;
+  BLEND_CASE(GL_ONE, GL_ONE, GL_ONE, GL_ONE_MINUS_SRC_ALPHA):
+    return src + dst - (muldiv255(dst, src) & ALPHA_MASK);
+  BLEND_CASE(GL_ONE_MINUS_DST_ALPHA, GL_ONE, GL_ZERO, GL_ONE):
+    // src*(1-dst.a) + dst*1 = src - src*dst.a + dst
+    return dst + ((src - muldiv255(src, alphas(dst))) & RGB_MASK);
+  BLEND_CASE(GL_CONSTANT_COLOR, GL_ONE_MINUS_SRC_COLOR):
+    // src*k + (1-src)*dst = src*k + dst -
+    // src*dst = dst + src*(k - dst) use addlow
+    // for signed overflow
+    return addlow(
+        dst, muldiv255(src, repeat2(ctx->blendcolor) - dst));
+
+  // We must explicitly handle the masked/anti-aliased secondary blend case.
+  // The secondary color as well as the source must be multiplied by the
+  // weights.
+  case BLEND_KEY(GL_ONE, GL_ONE_MINUS_SRC1_COLOR): {
+    WideRGBA8 secondary =
+        applyColor(dst,
+            packColor<uint32_t>(fragment_shader->gl_SecondaryFragColor));
+    return src + dst - secondary;
+  }
+  case MASK_BLEND_KEY(GL_ONE, GL_ONE_MINUS_SRC1_COLOR): {
+    WideRGBA8 secondary =
+        applyColor(dst,
+            packColor<uint32_t>(fragment_shader->gl_SecondaryFragColor));
+    WideRGBA8 mask = load_clip_mask(buf, span);
+    return muldiv255(src, mask) + dst - muldiv255(secondary, mask);
+  }
+  case AA_BLEND_KEY(GL_ONE, GL_ONE_MINUS_SRC1_COLOR): {
+    WideRGBA8 secondary =
+        applyColor(dst,
+            packColor<uint32_t>(fragment_shader->gl_SecondaryFragColor));
+    DO_AA(RGBA8, {
+      src = muldiv256(src, aa);
+      secondary = muldiv256(secondary, aa);
+    });
+    return src + dst - secondary;
+  }
+  case AA_MASK_BLEND_KEY(GL_ONE, GL_ONE_MINUS_SRC1_COLOR): {
+    WideRGBA8 secondary =
+        applyColor(dst,
+            packColor<uint32_t>(fragment_shader->gl_SecondaryFragColor));
+    WideRGBA8 mask = load_clip_mask(buf, span);
+    DO_AA(RGBA8, mask = muldiv256(mask, aa));
+    return muldiv255(src, mask) + dst - muldiv255(secondary, mask);
+  }
+
+  BLEND_CASE(GL_MIN):
+    return min(src, dst);
+  BLEND_CASE(GL_MAX):
+    return max(src, dst);
+
+  // The KHR_blend_equation_advanced spec describes the blend equations such
+  // that the unpremultiplied values Cs, Cd, As, Ad and function f combine to
+  // the result:
+  //     Cr = f(Cs,Cd)*As*Ad + Cs*As*(1-Ad) + Cd*AD*(1-As)
+  //     Ar = As*Ad + As*(1-Ad) + Ad*(1-As)
+  // However, working with unpremultiplied values requires expensive math to
+  // unpremultiply and premultiply again during blending. We can use the fact
+  // that premultiplied value P = C*A and simplify the equations such that no
+  // unpremultiplied colors are necessary, allowing us to stay with integer
+  // math that avoids floating-point conversions in the common case. Some of
+  // the blend modes require division or sqrt, in which case we do convert
+  // to (possibly transposed/unpacked) floating-point to implement the mode.
+  // However, most common modes can still use cheaper premultiplied integer
+  // math. As an example, the multiply mode f(Cs,Cd) = Cs*Cd is simplified
+  // to:
+  //     Cr = Cs*Cd*As*Ad + Cs*As*(1-Ad) + Cd*Ad*(1-As)
+  //     .. Pr = Ps*Pd + Ps - Ps*Ad + Pd - Pd*As
+  //     Ar = As*Ad + As - As*Ad + Ad - Ad*As
+  //     .. Ar = As + Ad - As*Ad
+  // Note that the alpha equation is the same for all blend equations, such
+  // that so long as the implementation results in As + Ad - As*Ad, we can
+  // avoid using separate instructions to compute the alpha result, which is
+  // dependent on the math used to implement each blend mode. The exact
+  // reductions used to get the final math for every blend mode are too
+  // involved to show here in comments, but mostly follows from replacing
+  // Cs*As and Cd*Ad with Ps and Ps while factoring out as many common terms
+  // as possible.
+
+  BLEND_CASE(GL_MULTIPLY_KHR): {
+    WideRGBA8 diff = muldiv255(alphas(src) - (src & RGB_MASK),
+                               alphas(dst) - (dst & RGB_MASK));
+    return src + dst + (diff & RGB_MASK) - alphas(diff);
+  }
+  BLEND_CASE(GL_SCREEN_KHR):
+    return src + dst - muldiv255(src, dst);
+  BLEND_CASE(GL_OVERLAY_KHR): {
+    WideRGBA8 srcA = alphas(src);
+    WideRGBA8 dstA = alphas(dst);
+    WideRGBA8 diff = muldiv255(src, dst) + muldiv255(srcA - src, dstA - dst);
+    return src + dst +
+           if_then_else(dst * 2 <= dstA, (diff & RGB_MASK) - alphas(diff),
+                        -diff);
+  }
+  BLEND_CASE(GL_DARKEN_KHR):
+    return src + dst -
+           max(muldiv255(src, alphas(dst)), muldiv255(dst, alphas(src)));
+  BLEND_CASE(GL_LIGHTEN_KHR):
+    return src + dst -
+           min(muldiv255(src, alphas(dst)), muldiv255(dst, alphas(src)));
+
+  BLEND_CASE(GL_COLORDODGE_KHR): {
+    // Color-dodge and color-burn require division, so we convert to FP math
+    // here, but avoid transposing to a vec4.
+    WideRGBA32F srcF = CONVERT(src, WideRGBA32F);
+    WideRGBA32F srcA = alphas(srcF);
+    WideRGBA32F dstF = CONVERT(dst, WideRGBA32F);
+    WideRGBA32F dstA = alphas(dstF);
+    return pack_pixels_RGBA8(
+        srcA * set_alphas(
+                   min(dstA, dstF * srcA * recip_or(srcA - srcF, 255.0f)),
+                   dstF) +
+            srcF * (255.0f - dstA) + dstF * (255.0f - srcA),
+        1.0f / 255.0f);
+  }
+  BLEND_CASE(GL_COLORBURN_KHR): {
+    WideRGBA32F srcF = CONVERT(src, WideRGBA32F);
+    WideRGBA32F srcA = alphas(srcF);
+    WideRGBA32F dstF = CONVERT(dst, WideRGBA32F);
+    WideRGBA32F dstA = alphas(dstF);
+    return pack_pixels_RGBA8(
+        srcA * set_alphas((dstA - min(dstA, (dstA - dstF) * srcA *
+                                                recip_or(srcF, 255.0f))),
+                          dstF) +
+            srcF * (255.0f - dstA) + dstF * (255.0f - srcA),
+        1.0f / 255.0f);
+  }
+  BLEND_CASE(GL_HARDLIGHT_KHR): {
+    WideRGBA8 srcA = alphas(src);
+    WideRGBA8 dstA = alphas(dst);
+    WideRGBA8 diff = muldiv255(src, dst) + muldiv255(srcA - src, dstA - dst);
+    return src + dst +
+           if_then_else(src * 2 <= srcA, (diff & RGB_MASK) - alphas(diff),
+                        -diff);
+  }
+
+  BLEND_CASE(GL_SOFTLIGHT_KHR): {
+    // Soft-light requires an unpremultiply that can't be factored out as
+    // well as a sqrt, so we convert to FP math here, but avoid transposing
+    // to a vec4.
+    WideRGBA32F srcF = CONVERT(src, WideRGBA32F);
+    WideRGBA32F srcA = alphas(srcF);
+    WideRGBA32F dstF = CONVERT(dst, WideRGBA32F);
+    WideRGBA32F dstA = alphas(dstF);
+    WideRGBA32F dstU = unpremultiply(dstF);
+    WideRGBA32F scale = srcF + srcF - srcA;
+    return pack_pixels_RGBA8(
+        dstF * (255.0f +
+                set_alphas(
+                    scale *
+                        if_then_else(scale < 0.0f, 1.0f - dstU,
+                                     min((16.0f * dstU - 12.0f) * dstU + 3.0f,
+                                         inversesqrt(dstU) - 1.0f)),
+                    WideRGBA32F(0.0f))) +
+            srcF * (255.0f - dstA),
+        1.0f / 255.0f);
+  }
+  BLEND_CASE(GL_DIFFERENCE_KHR): {
+    WideRGBA8 diff =
+        min(muldiv255(dst, alphas(src)), muldiv255(src, alphas(dst)));
+    return src + dst - diff - (diff & RGB_MASK);
+  }
+  BLEND_CASE(GL_EXCLUSION_KHR): {
+    WideRGBA8 diff = muldiv255(src, dst);
+    return src + dst - diff - (diff & RGB_MASK);
+  }
+
+  // The HSL blend modes are non-separable and require complicated use of
+  // division. It is advantageous to convert to FP and transpose to vec4
+  // math to more easily manipulate the individual color components.
+#define DO_HSL(rgb)                                                            \
+  do {                                                                         \
+    vec4 srcV = unpack(CONVERT(src, PackedRGBA32F));                           \
+    vec4 dstV = unpack(CONVERT(dst, PackedRGBA32F));                           \
+    Float srcA = srcV.w * (1.0f / 255.0f);                                     \
+    Float dstA = dstV.w * (1.0f / 255.0f);                                     \
+    Float srcDstA = srcV.w * dstA;                                             \
+    vec3 srcC = vec3(srcV) * dstA;                                             \
+    vec3 dstC = vec3(dstV) * srcA;                                             \
+    return pack_pixels_RGBA8(vec4(rgb + vec3(srcV) - srcC + vec3(dstV) - dstC, \
+                                  srcV.w + dstV.w - srcDstA),                  \
+                             1.0f);                                            \
+  } while (0)
+
+  BLEND_CASE(GL_HSL_HUE_KHR):
+    DO_HSL(set_lum_sat(srcC, dstC, dstC, srcDstA));
+  BLEND_CASE(GL_HSL_SATURATION_KHR):
+    DO_HSL(set_lum_sat(dstC, srcC, dstC, srcDstA));
+  BLEND_CASE(GL_HSL_COLOR_KHR):
+    DO_HSL(set_lum(srcC, dstC, srcDstA));
+  BLEND_CASE(GL_HSL_LUMINOSITY_KHR):
+    DO_HSL(set_lum(dstC, srcC, srcDstA));
+
+  // SWGL-specific extended blend modes.
+  BLEND_CASE(SWGL_BLEND_DROP_SHADOW): {
+    // Premultiplied alpha over blend, but with source color set to source alpha
+    // modulated with a constant color.
+    WideRGBA8 color = applyColor(alphas(src), swgl_BlendColorRGBA8);
+    return color + dst - muldiv255(dst, alphas(color));
+  }
+
+  BLEND_CASE(SWGL_BLEND_SUBPIXEL_TEXT):
+    // Premultiplied alpha over blend, but treats the source as a subpixel mask
+    // modulated with a constant color.
+    return applyColor(src, swgl_BlendColorRGBA8) + dst -
+           muldiv255(dst, applyColor(src, swgl_BlendAlphaRGBA8));
+
+  default:
+    UNREACHABLE;
+    // return src;
+  }
+
+#undef BLEND_CASE
+#undef BLEND_CASE_KEY
+  // clang-format on
+}
+
+static PREFER_INLINE WideR8 blend_pixels(uint8_t* buf, WideR8 dst, WideR8 src,
+                                         int span = 4) {
+// clang-format off
+#define BLEND_CASE_KEY(key)                          \
+  case AA_##key:                                     \
+    DO_AA(R8, src = muldiv256(src, aa));             \
+    goto key;                                        \
+  case AA_MASK_##key:                                \
+    DO_AA(R8, src = muldiv256(src, aa));             \
+    FALLTHROUGH;                                     \
+  case MASK_##key:                                   \
+    src = muldiv255(src, load_clip_mask(buf, span)); \
+    FALLTHROUGH;                                     \
+  case key: key
+
+#define BLEND_CASE(...) BLEND_CASE_KEY(BLEND_KEY(__VA_ARGS__))
+
+  switch (blend_key) {
+  BLEND_CASE(GL_ONE, GL_ZERO):
+    return src;
+  BLEND_CASE(GL_ZERO, GL_SRC_COLOR):
+    return muldiv255(src, dst);
+  BLEND_CASE(GL_ONE, GL_ONE):
+    return src + dst;
+  default:
+    UNREACHABLE;
+    // return src;
+  }
+
+#undef BLEND_CASE
+#undef BLEND_CASE_KEY
+  // clang-format on
+}
+
+static ALWAYS_INLINE void commit_span(uint32_t* buf, WideRGBA8 r) {
+  unaligned_store(buf, pack(r));
+}
+
+static ALWAYS_INLINE void commit_span(uint32_t* buf, WideRGBA8 r, int len) {
+  partial_store_span(buf, pack(r), len);
+}
+
+static ALWAYS_INLINE WideRGBA8 blend_span(uint32_t* buf, WideRGBA8 r) {
+  return blend_pixels(buf, unaligned_load<PackedRGBA8>(buf), r);
+}
+
+static ALWAYS_INLINE WideRGBA8 blend_span(uint32_t* buf, WideRGBA8 r, int len) {
+  return blend_pixels(buf, partial_load_span<PackedRGBA8>(buf, len), r, len);
+}
+
+static ALWAYS_INLINE void commit_span(uint32_t* buf, PackedRGBA8 r) {
+  unaligned_store(buf, r);
+}
+
+static ALWAYS_INLINE void commit_span(uint32_t* buf, PackedRGBA8 r, int len) {
+  partial_store_span(buf, r, len);
+}
+
+static ALWAYS_INLINE PackedRGBA8 blend_span(uint32_t* buf, PackedRGBA8 r) {
+  return pack(blend_span(buf, unpack(r)));
+}
+
+static ALWAYS_INLINE PackedRGBA8 blend_span(uint32_t* buf, PackedRGBA8 r,
+                                            int len) {
+  return pack(blend_span(buf, unpack(r), len));
+}
+
+static ALWAYS_INLINE void commit_span(uint8_t* buf, WideR8 r) {
+  unaligned_store(buf, pack(r));
+}
+
+static ALWAYS_INLINE void commit_span(uint8_t* buf, WideR8 r, int len) {
+  partial_store_span(buf, pack(r), len);
+}
+
+static ALWAYS_INLINE WideR8 blend_span(uint8_t* buf, WideR8 r) {
+  return blend_pixels(buf, unpack(unaligned_load<PackedR8>(buf)), r);
+}
+
+static ALWAYS_INLINE WideR8 blend_span(uint8_t* buf, WideR8 r, int len) {
+  return blend_pixels(buf, unpack(partial_load_span<PackedR8>(buf, len)), r,
+                      len);
+}
+
+static ALWAYS_INLINE void commit_span(uint8_t* buf, PackedR8 r) {
+  unaligned_store(buf, r);
+}
+
+static ALWAYS_INLINE void commit_span(uint8_t* buf, PackedR8 r, int len) {
+  partial_store_span(buf, r, len);
+}
+
+static ALWAYS_INLINE PackedR8 blend_span(uint8_t* buf, PackedR8 r) {
+  return pack(blend_span(buf, unpack(r)));
+}
+
+static ALWAYS_INLINE PackedR8 blend_span(uint8_t* buf, PackedR8 r, int len) {
+  return pack(blend_span(buf, unpack(r), len));
+}
+
+template <bool BLEND, typename P, typename R>
+static ALWAYS_INLINE void commit_blend_span(P* buf, R r) {
+  if (BLEND) {
+    commit_span(buf, blend_span(buf, r));
+  } else {
+    commit_span(buf, r);
+  }
+}
+
+template <bool BLEND, typename P, typename R>
+static ALWAYS_INLINE void commit_blend_span(P* buf, R r, int len) {
+  if (BLEND) {
+    commit_span(buf, blend_span(buf, r, len), len);
+  } else {
+    commit_span(buf, r, len);
+  }
+}
+
+template <typename P, typename R>
+static ALWAYS_INLINE void commit_blend_solid_span(P* buf, R r, int len) {
+  for (P* end = &buf[len & ~3]; buf < end; buf += 4) {
+    commit_span(buf, blend_span(buf, r));
+  }
+  len &= 3;
+  if (len > 0) {
+    partial_store_span(buf, pack(blend_span(buf, r, len)), len);
+  }
+}
+
+template <bool BLEND>
+static void commit_solid_span(uint32_t* buf, WideRGBA8 r, int len) {
+  commit_blend_solid_span(buf, r, len);
+}
+
+template <>
+ALWAYS_INLINE void commit_solid_span<false>(uint32_t* buf, WideRGBA8 r,
+                                            int len) {
+  fill_n(buf, len, bit_cast<U32>(pack(r)).x);
+}
+
+template <bool BLEND>
+static void commit_solid_span(uint8_t* buf, WideR8 r, int len) {
+  commit_blend_solid_span(buf, r, len);
+}
+
+template <>
+ALWAYS_INLINE void commit_solid_span<false>(uint8_t* buf, WideR8 r, int len) {
+  PackedR8 p = pack(r);
+  if (uintptr_t(buf) & 3) {
+    int align = 4 - (uintptr_t(buf) & 3);
+    align = min(align, len);
+    partial_store_span(buf, p, align);
+    buf += align;
+    len -= align;
+  }
+  fill_n((uint32_t*)buf, len / 4, bit_cast<uint32_t>(p));
+  buf += len & ~3;
+  len &= 3;
+  if (len > 0) {
+    partial_store_span(buf, p, len);
+  }
+}
diff --git a/gfx/wr/swgl/src/composite.h b/gfx/wr/swgl/src/composite.h
new file mode 100644
index 0000000000..70acabeca4
--- /dev/null
+++ b/gfx/wr/swgl/src/composite.h
@@ -0,0 +1,1386 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// Converts a pixel from a source format to a destination format. By default,
+// just return the value unchanged as for a simple copy.
+template <typename P, typename U>
+static ALWAYS_INLINE P convert_pixel(U src) {
+  return src;
+}
+
+// R8 format maps to BGRA value 0,0,R,1. The byte order is endian independent,
+// but the shifts unfortunately depend on endianness.
+template <>
+ALWAYS_INLINE uint32_t convert_pixel<uint32_t>(uint8_t src) {
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  return (uint32_t(src) << 16) | 0xFF000000;
+#else
+  return (uint32_t(src) << 8) | 0x000000FF;
+#endif
+}
+
+// RG8 format maps to BGRA value 0,G,R,1.
+template <>
+ALWAYS_INLINE uint32_t convert_pixel<uint32_t>(uint16_t src) {
+  uint32_t rg = src;
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  return ((rg & 0x00FF) << 16) | (rg & 0xFF00) | 0xFF000000;
+#else
+  return (rg & 0xFF00) | ((rg & 0x00FF) << 16) | 0x000000FF;
+#endif
+}
+
+// RGBA8 format maps to R.
+template <>
+ALWAYS_INLINE uint8_t convert_pixel<uint8_t>(uint32_t src) {
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  return (src >> 16) & 0xFF;
+#else
+  return (src >> 8) & 0xFF;
+#endif
+}
+
+// RGBA8 formats maps to R,G.
+template <>
+ALWAYS_INLINE uint16_t convert_pixel<uint16_t>(uint32_t src) {
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  return ((src >> 16) & 0x00FF) | (src & 0xFF00);
+#else
+  return (src & 0xFF00) | ((src >> 16) & 0x00FF);
+#endif
+}
+
+// R8 format maps to R,0.
+template <>
+ALWAYS_INLINE uint16_t convert_pixel<uint16_t>(uint8_t src) {
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  return src;
+#else
+  return uint16_t(src) << 8;
+#endif
+}
+
+// RG8 format maps to R.
+template <>
+ALWAYS_INLINE uint8_t convert_pixel<uint8_t>(uint16_t src) {
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  return src & 0xFF;
+#else
+  return src >> 8;
+#endif
+}
+
+template <bool COMPOSITE, typename P>
+static inline void copy_row(P* dst, const P* src, int span) {
+  // No scaling, so just do a fast copy.
+  memcpy(dst, src, span * sizeof(P));
+}
+
+template <>
+void copy_row<true, uint32_t>(uint32_t* dst, const uint32_t* src, int span) {
+  // No scaling, so just do a fast composite.
+  auto* end = dst + span;
+  while (dst + 4 <= end) {
+    WideRGBA8 srcpx = unpack(unaligned_load<PackedRGBA8>(src));
+    WideRGBA8 dstpx = unpack(unaligned_load<PackedRGBA8>(dst));
+    PackedRGBA8 r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
+    unaligned_store(dst, r);
+    src += 4;
+    dst += 4;
+  }
+  if (dst < end) {
+    WideRGBA8 srcpx = unpack(partial_load_span<PackedRGBA8>(src, end - dst));
+    WideRGBA8 dstpx = unpack(partial_load_span<PackedRGBA8>(dst, end - dst));
+    auto r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
+    partial_store_span(dst, r, end - dst);
+  }
+}
+
+template <bool COMPOSITE, typename P, typename U>
+static inline void scale_row(P* dst, int dstWidth, const U* src, int srcWidth,
+                             int span, int frac) {
+  // Do scaling with different source and dest widths.
+  for (P* end = dst + span; dst < end; dst++) {
+    *dst = convert_pixel<P>(*src);
+    // Step source according to width ratio.
+    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
+      src++;
+    }
+  }
+}
+
+template <>
+void scale_row<true, uint32_t, uint32_t>(uint32_t* dst, int dstWidth,
+                                         const uint32_t* src, int srcWidth,
+                                         int span, int frac) {
+  // Do scaling with different source and dest widths.
+  // Gather source pixels four at a time for better packing.
+  auto* end = dst + span;
+  for (; dst + 4 <= end; dst += 4) {
+    U32 srcn;
+    srcn.x = *src;
+    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
+      src++;
+    }
+    srcn.y = *src;
+    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
+      src++;
+    }
+    srcn.z = *src;
+    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
+      src++;
+    }
+    srcn.w = *src;
+    for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
+      src++;
+    }
+    WideRGBA8 srcpx = unpack(bit_cast<PackedRGBA8>(srcn));
+    WideRGBA8 dstpx = unpack(unaligned_load<PackedRGBA8>(dst));
+    PackedRGBA8 r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
+    unaligned_store(dst, r);
+  }
+  if (dst < end) {
+    // Process any remaining pixels. Try to gather as many pixels as possible
+    // into a single source chunk for compositing.
+    U32 srcn = {*src, 0, 0, 0};
+    if (end - dst > 1) {
+      for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
+        src++;
+      }
+      srcn.y = *src;
+      if (end - dst > 2) {
+        for (frac += srcWidth; frac >= dstWidth; frac -= dstWidth) {
+          src++;
+        }
+        srcn.z = *src;
+      }
+    }
+    WideRGBA8 srcpx = unpack(bit_cast<PackedRGBA8>(srcn));
+    WideRGBA8 dstpx = unpack(partial_load_span<PackedRGBA8>(dst, end - dst));
+    auto r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
+    partial_store_span(dst, r, end - dst);
+  }
+}
+
+template <bool COMPOSITE = false>
+static NO_INLINE void scale_blit(Texture& srctex, const IntRect& srcReq,
+                                 Texture& dsttex, const IntRect& dstReq,
+                                 bool invertY, const IntRect& clipRect) {
+  assert(!COMPOSITE || (srctex.internal_format == GL_RGBA8 &&
+                        dsttex.internal_format == GL_RGBA8));
+  // Cache scaling ratios
+  int srcWidth = srcReq.width();
+  int srcHeight = srcReq.height();
+  int dstWidth = dstReq.width();
+  int dstHeight = dstReq.height();
+  // Compute valid dest bounds
+  IntRect dstBounds = dsttex.sample_bounds(dstReq).intersect(clipRect);
+  // Compute valid source bounds
+  IntRect srcBounds = srctex.sample_bounds(srcReq, invertY);
+  // If srcReq is outside the source texture, we need to clip the sampling
+  // bounds so that we never sample outside valid source bounds. Get texture
+  // bounds relative to srcReq and scale to dest-space rounding inward, using
+  // this rect to limit the dest bounds further.
+  IntRect srcClip = srctex.bounds() - srcReq.origin();
+  if (invertY) {
+    srcClip.invert_y(srcReq.height());
+  }
+  srcClip.scale(srcWidth, srcHeight, dstWidth, dstHeight, true);
+  dstBounds.intersect(srcClip);
+  // Check if clipped sampling bounds are empty
+  if (dstBounds.is_empty()) {
+    return;
+  }
+
+  // Calculate source and dest pointers from clamped offsets
+  int srcStride = srctex.stride();
+  int destStride = dsttex.stride();
+  char* dest = dsttex.sample_ptr(dstReq, dstBounds);
+  // Clip the source bounds by the destination offset.
+  int fracX = srcWidth * dstBounds.x0;
+  int fracY = srcHeight * dstBounds.y0;
+  srcBounds.x0 = max(fracX / dstWidth, srcBounds.x0);
+  srcBounds.y0 = max(fracY / dstHeight, srcBounds.y0);
+  fracX %= dstWidth;
+  fracY %= dstHeight;
+  char* src = srctex.sample_ptr(srcReq, srcBounds, invertY);
+  // Inverted Y must step downward along source rows
+  if (invertY) {
+    srcStride = -srcStride;
+  }
+  int span = dstBounds.width();
+  for (int rows = dstBounds.height(); rows > 0; rows--) {
+    switch (srctex.bpp()) {
+      case 1:
+        switch (dsttex.bpp()) {
+          case 2:
+            scale_row<COMPOSITE>((uint16_t*)dest, dstWidth, (uint8_t*)src,
+                                 srcWidth, span, fracX);
+            break;
+          case 4:
+            scale_row<COMPOSITE>((uint32_t*)dest, dstWidth, (uint8_t*)src,
+                                 srcWidth, span, fracX);
+            break;
+          default:
+            if (srcWidth == dstWidth)
+              copy_row<COMPOSITE>((uint8_t*)dest, (uint8_t*)src, span);
+            else
+              scale_row<COMPOSITE>((uint8_t*)dest, dstWidth, (uint8_t*)src,
+                                   srcWidth, span, fracX);
+            break;
+        }
+        break;
+      case 2:
+        switch (dsttex.bpp()) {
+          case 1:
+            scale_row<COMPOSITE>((uint8_t*)dest, dstWidth, (uint16_t*)src,
+                                 srcWidth, span, fracX);
+            break;
+          case 4:
+            scale_row<COMPOSITE>((uint32_t*)dest, dstWidth, (uint16_t*)src,
+                                 srcWidth, span, fracX);
+            break;
+          default:
+            if (srcWidth == dstWidth)
+              copy_row<COMPOSITE>((uint16_t*)dest, (uint16_t*)src, span);
+            else
+              scale_row<COMPOSITE>((uint16_t*)dest, dstWidth, (uint16_t*)src,
+                                   srcWidth, span, fracX);
+            break;
+        }
+        break;
+      case 4:
+        switch (dsttex.bpp()) {
+          case 1:
+            scale_row<COMPOSITE>((uint8_t*)dest, dstWidth, (uint32_t*)src,
+                                 srcWidth, span, fracX);
+            break;
+          case 2:
+            scale_row<COMPOSITE>((uint16_t*)dest, dstWidth, (uint32_t*)src,
+                                 srcWidth, span, fracX);
+            break;
+          default:
+            if (srcWidth == dstWidth)
+              copy_row<COMPOSITE>((uint32_t*)dest, (uint32_t*)src, span);
+            else
+              scale_row<COMPOSITE>((uint32_t*)dest, dstWidth, (uint32_t*)src,
+                                   srcWidth, span, fracX);
+            break;
+        }
+        break;
+      default:
+        assert(false);
+        break;
+    }
+    dest += destStride;
+    // Step source according to height ratio.
+    for (fracY += srcHeight; fracY >= dstHeight; fracY -= dstHeight) {
+      src += srcStride;
+    }
+  }
+}
+
+template <bool COMPOSITE>
+static void linear_row_blit(uint32_t* dest, int span, const vec2_scalar& srcUV,
+                            float srcDU, sampler2D sampler) {
+  vec2 uv = init_interp(srcUV, vec2_scalar(srcDU, 0.0f));
+  for (; span >= 4; span -= 4) {
+    auto srcpx = textureLinearPackedRGBA8(sampler, ivec2(uv));
+    unaligned_store(dest, srcpx);
+    dest += 4;
+    uv.x += 4 * srcDU;
+  }
+  if (span > 0) {
+    auto srcpx = textureLinearPackedRGBA8(sampler, ivec2(uv));
+    partial_store_span(dest, srcpx, span);
+  }
+}
+
+template <>
+void linear_row_blit<true>(uint32_t* dest, int span, const vec2_scalar& srcUV,
+                           float srcDU, sampler2D sampler) {
+  vec2 uv = init_interp(srcUV, vec2_scalar(srcDU, 0.0f));
+  for (; span >= 4; span -= 4) {
+    WideRGBA8 srcpx = textureLinearUnpackedRGBA8(sampler, ivec2(uv));
+    WideRGBA8 dstpx = unpack(unaligned_load<PackedRGBA8>(dest));
+    PackedRGBA8 r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
+    unaligned_store(dest, r);
+
+    dest += 4;
+    uv.x += 4 * srcDU;
+  }
+  if (span > 0) {
+    WideRGBA8 srcpx = textureLinearUnpackedRGBA8(sampler, ivec2(uv));
+    WideRGBA8 dstpx = unpack(partial_load_span<PackedRGBA8>(dest, span));
+    PackedRGBA8 r = pack(srcpx + dstpx - muldiv255(dstpx, alphas(srcpx)));
+    partial_store_span(dest, r, span);
+  }
+}
+
+template <bool COMPOSITE>
+static void linear_row_blit(uint8_t* dest, int span, const vec2_scalar& srcUV,
+                            float srcDU, sampler2D sampler) {
+  vec2 uv = init_interp(srcUV, vec2_scalar(srcDU, 0.0f));
+  for (; span >= 4; span -= 4) {
+    auto srcpx = textureLinearPackedR8(sampler, ivec2(uv));
+    unaligned_store(dest, srcpx);
+    dest += 4;
+    uv.x += 4 * srcDU;
+  }
+  if (span > 0) {
+    auto srcpx = textureLinearPackedR8(sampler, ivec2(uv));
+    partial_store_span(dest, srcpx, span);
+  }
+}
+
+template <bool COMPOSITE>
+static void linear_row_blit(uint16_t* dest, int span, const vec2_scalar& srcUV,
+                            float srcDU, sampler2D sampler) {
+  vec2 uv = init_interp(srcUV, vec2_scalar(srcDU, 0.0f));
+  for (; span >= 4; span -= 4) {
+    auto srcpx = textureLinearPackedRG8(sampler, ivec2(uv));
+    unaligned_store(dest, srcpx);
+    dest += 4;
+    uv.x += 4 * srcDU;
+  }
+  if (span > 0) {
+    auto srcpx = textureLinearPackedRG8(sampler, ivec2(uv));
+    partial_store_span(dest, srcpx, span);
+  }
+}
+
+template <bool COMPOSITE = false>
+static NO_INLINE void linear_blit(Texture& srctex, const IntRect& srcReq,
+                                  Texture& dsttex, const IntRect& dstReq,
+                                  bool invertX, bool invertY,
+                                  const IntRect& clipRect) {
+  assert(srctex.internal_format == GL_RGBA8 ||
+         srctex.internal_format == GL_R8 || srctex.internal_format == GL_RG8);
+  assert(!COMPOSITE || (srctex.internal_format == GL_RGBA8 &&
+                        dsttex.internal_format == GL_RGBA8));
+  // Compute valid dest bounds
+  IntRect dstBounds = dsttex.sample_bounds(dstReq);
+  dstBounds.intersect(clipRect);
+  // Check if sampling bounds are empty
+  if (dstBounds.is_empty()) {
+    return;
+  }
+  // Initialize sampler for source texture
+  sampler2D_impl sampler;
+  init_sampler(&sampler, srctex);
+  sampler.filter = TextureFilter::LINEAR;
+  // Compute source UVs
+  vec2_scalar srcUV(srcReq.x0, srcReq.y0);
+  vec2_scalar srcDUV(float(srcReq.width()) / dstReq.width(),
+                     float(srcReq.height()) / dstReq.height());
+  if (invertX) {
+    // Advance to the end of the row and flip the step.
+    srcUV.x += srcReq.width();
+    srcDUV.x = -srcDUV.x;
+  }
+  // Inverted Y must step downward along source rows
+  if (invertY) {
+    srcUV.y += srcReq.height();
+    srcDUV.y = -srcDUV.y;
+  }
+  // Skip to clamped source start
+  srcUV += srcDUV * (vec2_scalar(dstBounds.x0, dstBounds.y0) + 0.5f);
+  // Scale UVs by lerp precision
+  srcUV = linearQuantize(srcUV, 128);
+  srcDUV *= 128.0f;
+  // Calculate dest pointer from clamped offsets
+  int bpp = dsttex.bpp();
+  int destStride = dsttex.stride();
+  char* dest = dsttex.sample_ptr(dstReq, dstBounds);
+  int span = dstBounds.width();
+  for (int rows = dstBounds.height(); rows > 0; rows--) {
+    switch (bpp) {
+      case 1:
+        linear_row_blit<COMPOSITE>((uint8_t*)dest, span, srcUV, srcDUV.x,
+                                   &sampler);
+        break;
+      case 2:
+        linear_row_blit<COMPOSITE>((uint16_t*)dest, span, srcUV, srcDUV.x,
+                                   &sampler);
+        break;
+      case 4:
+        linear_row_blit<COMPOSITE>((uint32_t*)dest, span, srcUV, srcDUV.x,
+                                   &sampler);
+        break;
+      default:
+        assert(false);
+        break;
+    }
+    dest += destStride;
+    srcUV.y += srcDUV.y;
+  }
+}
+
+// Whether the blit format is renderable.
+static inline bool is_renderable_format(GLenum format) {
+  switch (format) {
+    case GL_R8:
+    case GL_RG8:
+    case GL_RGBA8:
+      return true;
+    default:
+      return false;
+  }
+}
+
+extern "C" {
+
+void BlitFramebuffer(GLint srcX0, GLint srcY0, GLint srcX1, GLint srcY1,
+                     GLint dstX0, GLint dstY0, GLint dstX1, GLint dstY1,
+                     GLbitfield mask, GLenum filter) {
+  assert(mask == GL_COLOR_BUFFER_BIT);
+  Framebuffer* srcfb = get_framebuffer(GL_READ_FRAMEBUFFER);
+  if (!srcfb) return;
+  Framebuffer* dstfb = get_framebuffer(GL_DRAW_FRAMEBUFFER);
+  if (!dstfb) return;
+  Texture& srctex = ctx->textures[srcfb->color_attachment];
+  if (!srctex.buf) return;
+  Texture& dsttex = ctx->textures[dstfb->color_attachment];
+  if (!dsttex.buf) return;
+  assert(!dsttex.locked);
+  if (srctex.internal_format != dsttex.internal_format &&
+      (!is_renderable_format(srctex.internal_format) ||
+       !is_renderable_format(dsttex.internal_format))) {
+    // If the internal formats don't match, then we may have to convert. Require
+    // that the format is a simple renderable format to limit combinatoric
+    // explosion for now.
+    assert(false);
+    return;
+  }
+  // Force flipped Y onto dest coordinates
+  if (srcY1 < srcY0) {
+    swap(srcY0, srcY1);
+    swap(dstY0, dstY1);
+  }
+  bool invertY = dstY1 < dstY0;
+  if (invertY) {
+    swap(dstY0, dstY1);
+  }
+  IntRect srcReq = IntRect{srcX0, srcY0, srcX1, srcY1} - srctex.offset;
+  IntRect dstReq = IntRect{dstX0, dstY0, dstX1, dstY1} - dsttex.offset;
+  if (srcReq.is_empty() || dstReq.is_empty()) {
+    return;
+  }
+  IntRect clipRect = {0, 0, dstReq.width(), dstReq.height()};
+  prepare_texture(srctex);
+  prepare_texture(dsttex, &dstReq);
+  if (!srcReq.same_size(dstReq) && srctex.width >= 2 && filter == GL_LINEAR &&
+      srctex.internal_format == dsttex.internal_format &&
+      is_renderable_format(srctex.internal_format)) {
+    linear_blit(srctex, srcReq, dsttex, dstReq, false, invertY, dstReq);
+  } else {
+    scale_blit(srctex, srcReq, dsttex, dstReq, invertY, clipRect);
+  }
+}
+
+typedef Texture LockedTexture;
+
+// Lock the given texture to prevent modification.
+LockedTexture* LockTexture(GLuint texId) {
+  Texture& tex = ctx->textures[texId];
+  if (!tex.buf) {
+    assert(tex.buf != nullptr);
+    return nullptr;
+  }
+  if (__sync_fetch_and_add(&tex.locked, 1) == 0) {
+    // If this is the first time locking the texture, flush any delayed clears.
+    prepare_texture(tex);
+  }
+  return (LockedTexture*)&tex;
+}
+
+// Lock the given framebuffer's color attachment to prevent modification.
+LockedTexture* LockFramebuffer(GLuint fboId) {
+  Framebuffer& fb = ctx->framebuffers[fboId];
+  // Only allow locking a framebuffer if it has a valid color attachment.
+  if (!fb.color_attachment) {
+    assert(fb.color_attachment != 0);
+    return nullptr;
+  }
+  return LockTexture(fb.color_attachment);
+}
+
+// Reference an already locked resource
+void LockResource(LockedTexture* resource) {
+  if (!resource) {
+    return;
+  }
+  __sync_fetch_and_add(&resource->locked, 1);
+}
+
+// Remove a lock on a texture that has been previously locked
+void UnlockResource(LockedTexture* resource) {
+  if (!resource) {
+    return;
+  }
+  if (__sync_fetch_and_add(&resource->locked, -1) <= 0) {
+    // The lock should always be non-zero before unlocking.
+    assert(0);
+  }
+}
+
+// Get the underlying buffer for a locked resource
+void* GetResourceBuffer(LockedTexture* resource, int32_t* width,
+                        int32_t* height, int32_t* stride) {
+  *width = resource->width;
+  *height = resource->height;
+  *stride = resource->stride();
+  return resource->buf;
+}
+
+// Extension for optimized compositing of textures or framebuffers that may be
+// safely used across threads. The source and destination must be locked to
+// ensure that they can be safely accessed while the SWGL context might be used
+// by another thread. Band extents along the Y axis may be used to clip the
+// destination rectangle without effecting the integer scaling ratios.
+void Composite(LockedTexture* lockedDst, LockedTexture* lockedSrc, GLint srcX,
+               GLint srcY, GLsizei srcWidth, GLsizei srcHeight, GLint dstX,
+               GLint dstY, GLsizei dstWidth, GLsizei dstHeight,
+               GLboolean opaque, GLboolean flipX, GLboolean flipY,
+               GLenum filter, GLint clipX, GLint clipY, GLsizei clipWidth,
+               GLsizei clipHeight) {
+  if (!lockedDst || !lockedSrc) {
+    return;
+  }
+  Texture& srctex = *lockedSrc;
+  Texture& dsttex = *lockedDst;
+  assert(srctex.bpp() == 4);
+  assert(dsttex.bpp() == 4);
+
+  IntRect srcReq =
+      IntRect{srcX, srcY, srcX + srcWidth, srcY + srcHeight} - srctex.offset;
+  IntRect dstReq =
+      IntRect{dstX, dstY, dstX + dstWidth, dstY + dstHeight} - dsttex.offset;
+  if (srcReq.is_empty() || dstReq.is_empty()) {
+    return;
+  }
+
+  // Compute clip rect as relative to the dstReq, as that's the same coords
+  // as used for the sampling bounds.
+  IntRect clipRect = {clipX - dstX, clipY - dstY, clipX - dstX + clipWidth,
+                      clipY - dstY + clipHeight};
+  // Ensure we have rows of at least 2 pixels when using the linear filter to
+  // avoid overreading the row. Force X flips onto the linear filter for now
+  // until scale_blit supports it.
+  bool useLinear =
+      srctex.width >= 2 &&
+      (flipX || (!srcReq.same_size(dstReq) && filter == GL_LINEAR));
+
+  if (opaque) {
+    if (useLinear) {
+      linear_blit<false>(srctex, srcReq, dsttex, dstReq, flipX, flipY,
+                         clipRect);
+    } else {
+      scale_blit<false>(srctex, srcReq, dsttex, dstReq, flipY, clipRect);
+    }
+  } else {
+    if (useLinear) {
+      linear_blit<true>(srctex, srcReq, dsttex, dstReq, flipX, flipY, clipRect);
+    } else {
+      scale_blit<true>(srctex, srcReq, dsttex, dstReq, flipY, clipRect);
+    }
+  }
+}
+
+}  // extern "C"
+
+// Saturated add helper for YUV conversion. Supported platforms have intrinsics
+// to do this natively, but support a slower generic fallback just in case.
+static inline V8<int16_t> addsat(V8<int16_t> x, V8<int16_t> y) {
+#if USE_SSE2
+  return _mm_adds_epi16(x, y);
+#elif USE_NEON
+  return vqaddq_s16(x, y);
+#else
+  auto r = x + y;
+  // An overflow occurred if the signs of both inputs x and y did not differ
+  // but yet the sign of the result did differ.
+  auto overflow = (~(x ^ y) & (r ^ x)) >> 15;
+  // If there was an overflow, we need to choose the appropriate limit to clamp
+  // to depending on whether or not the inputs are negative.
+  auto limit = (x >> 15) ^ 0x7FFF;
+  // If we didn't overflow, just use the result, and otherwise, use the limit.
+  return (~overflow & r) | (overflow & limit);
+#endif
+}
+
+// Interleave and packing helper for YUV conversion. During transform by the
+// color matrix, the color components are de-interleaved as this format is
+// usually what comes out of the planar YUV textures. The components thus need
+// to be interleaved before finally getting packed to BGRA format. Alpha is
+// forced to be opaque.
+static inline PackedRGBA8 packYUV(V8<int16_t> gg, V8<int16_t> br) {
+  return pack(bit_cast<WideRGBA8>(zip(br, gg))) |
+         PackedRGBA8{0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255, 0, 0, 0, 255};
+}
+
+// clang-format off
+// Supports YUV color matrixes of the form:
+// [R]   [1.1643835616438356,  0.0,  rv ]   [Y -  16]
+// [G] = [1.1643835616438358, -gu,  -gv ] x [U - 128]
+// [B]   [1.1643835616438356,  bu,  0.0 ]   [V - 128]
+// We must be able to multiply a YUV input by a matrix coefficient ranging as
+// high as ~2.2 in the U/V cases, where U/V can be signed values between -128
+// and 127. The largest fixed-point representation we can thus support without
+// overflowing 16 bit integers leaves us 6 bits of fractional precision while
+// also supporting a sign bit. The closest representation of the Y coefficient
+// ~1.164 in this precision is 74.5/2^6 which is common to all color spaces
+// we support. Conversions can still sometimes overflow the precision and
+// require clamping back into range, so we use saturated additions to do this
+// efficiently at no extra cost.
+// clang-format on
+struct YUVMatrix {
+  // These constants are loaded off the "this" pointer via relative addressing
+  // modes and should be about as quick to load as directly addressed SIMD
+  // constant memory.
+
+  V8<int16_t> br_uvCoeffs;  // biased by 6 bits [b_from_u, r_from_v, repeats]
+  V8<int16_t> gg_uvCoeffs;  // biased by 6 bits [g_from_u, g_from_v, repeats]
+  V8<uint16_t> yCoeffs;     // biased by 7 bits
+  V8<int16_t> yBias;        // 0 or 16
+  V8<int16_t> uvBias;       // 128
+  V8<int16_t> br_yMask;
+
+  // E.g. rec709-narrow:
+  // [ 1.16,     0,  1.79, -0.97 ]
+  // [ 1.16, -0.21, -0.53,  0.30 ]
+  // [ 1.16,  2.11,     0, -1.13 ]
+  // =
+  // [ yScale,        0, r_from_v ]   ([Y ]              )
+  // [ yScale, g_from_u, g_from_v ] x ([cb] - ycbcr_bias )
+  // [ yScale, b_from_u,        0 ]   ([cr]              )
+  static YUVMatrix From(const vec3_scalar& ycbcr_bias,
+                        const mat3_scalar& rgb_from_debiased_ycbcr,
+                        int rescale_factor = 0) {
+    assert(ycbcr_bias.z == ycbcr_bias.y);
+
+    const auto rgb_from_y = rgb_from_debiased_ycbcr[0].y;
+    assert(rgb_from_debiased_ycbcr[0].x == rgb_from_debiased_ycbcr[0].z);
+
+    int16_t br_from_y_mask = -1;
+    if (rgb_from_debiased_ycbcr[0].x == 0.0) {
+      // gbr-identity matrix?
+      assert(rgb_from_debiased_ycbcr[0].x == 0);
+      assert(rgb_from_debiased_ycbcr[0].y >= 1);
+      assert(rgb_from_debiased_ycbcr[0].z == 0);
+
+      assert(rgb_from_debiased_ycbcr[1].x == 0);
+      assert(rgb_from_debiased_ycbcr[1].y == 0);
+      assert(rgb_from_debiased_ycbcr[1].z >= 1);
+
+      assert(rgb_from_debiased_ycbcr[2].x >= 1);
+      assert(rgb_from_debiased_ycbcr[2].y == 0);
+      assert(rgb_from_debiased_ycbcr[2].z == 0);
+
+      assert(ycbcr_bias.x == 0);
+      assert(ycbcr_bias.y == 0);
+      assert(ycbcr_bias.z == 0);
+
+      br_from_y_mask = 0;
+    } else {
+      assert(rgb_from_debiased_ycbcr[0].x == rgb_from_y);
+    }
+
+    assert(rgb_from_debiased_ycbcr[1].x == 0.0);
+    const auto g_from_u = rgb_from_debiased_ycbcr[1].y;
+    const auto b_from_u = rgb_from_debiased_ycbcr[1].z;
+
+    const auto r_from_v = rgb_from_debiased_ycbcr[2].x;
+    const auto g_from_v = rgb_from_debiased_ycbcr[2].y;
+    assert(rgb_from_debiased_ycbcr[2].z == 0.0);
+
+    return YUVMatrix({ycbcr_bias.x, ycbcr_bias.y}, rgb_from_y, br_from_y_mask,
+                     r_from_v, g_from_u, g_from_v, b_from_u, rescale_factor);
+  }
+
+  // Convert matrix coefficients to fixed-point representation. If the matrix
+  // has a rescaling applied to it, then we need to take care to undo the
+  // scaling so that we can convert the coefficients to fixed-point range. The
+  // bias still requires shifting to apply the rescaling. The rescaling will be
+  // applied to the actual YCbCr sample data later by manually shifting it
+  // before applying this matrix.
+  YUVMatrix(vec2_scalar yuv_bias, double yCoeff, int16_t br_yMask_, double rv,
+            double gu, double gv, double bu, int rescale_factor = 0)
+      : br_uvCoeffs(zip(I16(int16_t(bu * (1 << (6 - rescale_factor)) + 0.5)),
+                        I16(int16_t(rv * (1 << (6 - rescale_factor)) + 0.5)))),
+        gg_uvCoeffs(
+            zip(I16(-int16_t(-gu * (1 << (6 - rescale_factor)) +
+                             0.5)),  // These are negative coeffs, so
+                                     // round them away from zero
+                I16(-int16_t(-gv * (1 << (6 - rescale_factor)) + 0.5)))),
+        yCoeffs(uint16_t(yCoeff * (1 << (6 + 1 - rescale_factor)) + 0.5)),
+        // We have a +0.5 fudge-factor for -ybias.
+        // Without this, we get white=254 not 255.
+        // This approximates rounding rather than truncation during `gg >>= 6`.
+        yBias(int16_t(((yuv_bias.x * 255 * yCoeff) - 0.5) * (1 << 6))),
+        uvBias(int16_t(yuv_bias.y * (255 << rescale_factor) + 0.5)),
+        br_yMask(br_yMask_) {
+    assert(yuv_bias.x >= 0);
+    assert(yuv_bias.y >= 0);
+    assert(yCoeff > 0);
+    assert(br_yMask_ == 0 || br_yMask_ == -1);
+    assert(bu > 0);
+    assert(rv > 0);
+    assert(gu <= 0);
+    assert(gv <= 0);
+    assert(rescale_factor <= 6);
+  }
+
+  ALWAYS_INLINE PackedRGBA8 convert(V8<int16_t> yy, V8<int16_t> uv) const {
+    // We gave ourselves an extra bit (7 instead of 6) of bias to give us some
+    // extra precision for the more-sensitive y scaling.
+    // Note that we have to use an unsigned multiply with a 2x scale to
+    // represent a fractional scale and to avoid shifting with the sign bit.
+
+    // Note: if you subtract the bias before multiplication, we see more
+    // underflows. This could be fixed by an unsigned subsat.
+    yy = bit_cast<V8<int16_t>>((bit_cast<V8<uint16_t>>(yy) * yCoeffs) >> 1);
+    yy -= yBias;
+
+    // Compute [B] = [yCoeff*Y + bu*U +  0*V]
+    //         [R]   [yCoeff*Y +  0*U + rv*V]
+    uv -= uvBias;
+    auto br = br_uvCoeffs * uv;
+    br = addsat(yy & br_yMask, br);
+    br >>= 6;
+
+    // Compute G = yCoeff*Y + gu*U + gv*V
+    // First calc [gu*U, gv*V, ...]:
+    auto gg = gg_uvCoeffs * uv;
+    // Then cross the streams to get `gu*U + gv*V`:
+    gg = addsat(gg, bit_cast<V8<int16_t>>(bit_cast<V4<uint32_t>>(gg) >> 16));
+    // Add the other parts:
+    gg = addsat(yy, gg);  // This is the part that needs the most headroom
+                          // usually. In particular, ycbcr(255,255,255) hugely
+                          // saturates.
+    gg >>= 6;
+
+    // Interleave B/R and G values. Force alpha (high-gg half) to opaque.
+    return packYUV(gg, br);
+  }
+};
+
+// Helper function for textureLinearRowR8 that samples horizontal taps and
+// combines them based on Y fraction with next row.
+template <typename S>
+static ALWAYS_INLINE V8<int16_t> linearRowTapsR8(S sampler, I32 ix,
+                                                 int32_t offsety,
+                                                 int32_t stridey,
+                                                 int16_t fracy) {
+  uint8_t* buf = (uint8_t*)sampler->buf + offsety;
+  auto a0 = unaligned_load<V2<uint8_t>>(&buf[ix.x]);
+  auto b0 = unaligned_load<V2<uint8_t>>(&buf[ix.y]);
+  auto c0 = unaligned_load<V2<uint8_t>>(&buf[ix.z]);
+  auto d0 = unaligned_load<V2<uint8_t>>(&buf[ix.w]);
+  auto abcd0 = CONVERT(combine(a0, b0, c0, d0), V8<int16_t>);
+  buf += stridey;
+  auto a1 = unaligned_load<V2<uint8_t>>(&buf[ix.x]);
+  auto b1 = unaligned_load<V2<uint8_t>>(&buf[ix.y]);
+  auto c1 = unaligned_load<V2<uint8_t>>(&buf[ix.z]);
+  auto d1 = unaligned_load<V2<uint8_t>>(&buf[ix.w]);
+  auto abcd1 = CONVERT(combine(a1, b1, c1, d1), V8<int16_t>);
+  abcd0 += ((abcd1 - abcd0) * fracy) >> 7;
+  return abcd0;
+}
+
+// Optimized version of textureLinearPackedR8 for Y R8 texture. This assumes
+// constant Y and returns a duplicate of the result interleaved with itself
+// to aid in later YUV transformation.
+template <typename S>
+static inline V8<int16_t> textureLinearRowR8(S sampler, I32 ix, int32_t offsety,
+                                             int32_t stridey, int16_t fracy) {
+  assert(sampler->format == TextureFormat::R8);
+
+  // Calculate X fraction and clamp X offset into range.
+  I32 fracx = ix;
+  ix >>= 7;
+  fracx = ((fracx & (ix >= 0)) | (ix > int32_t(sampler->width) - 2)) & 0x7F;
+  ix = clampCoord(ix, sampler->width - 1);
+
+  // Load the sample taps and combine rows.
+  auto abcd = linearRowTapsR8(sampler, ix, offsety, stridey, fracy);
+
+  // Unzip the result and do final horizontal multiply-add base on X fraction.
+  auto abcdl = SHUFFLE(abcd, abcd, 0, 0, 2, 2, 4, 4, 6, 6);
+  auto abcdh = SHUFFLE(abcd, abcd, 1, 1, 3, 3, 5, 5, 7, 7);
+  abcdl += ((abcdh - abcdl) * CONVERT(fracx, I16).xxyyzzww) >> 7;
+
+  // The final result is the packed values interleaved with a duplicate of
+  // themselves.
+  return abcdl;
+}
+
+// Optimized version of textureLinearPackedR8 for paired U/V R8 textures.
+// Since the two textures have the same dimensions and stride, the addressing
+// math can be shared between both samplers. This also allows a coalesced
+// multiply in the final stage by packing both U/V results into a single
+// operation.
+template <typename S>
+static inline V8<int16_t> textureLinearRowPairedR8(S sampler, S sampler2,
+                                                   I32 ix, int32_t offsety,
+                                                   int32_t stridey,
+                                                   int16_t fracy) {
+  assert(sampler->format == TextureFormat::R8 &&
+         sampler2->format == TextureFormat::R8);
+  assert(sampler->width == sampler2->width &&
+         sampler->height == sampler2->height);
+  assert(sampler->stride == sampler2->stride);
+
+  // Calculate X fraction and clamp X offset into range.
+  I32 fracx = ix;
+  ix >>= 7;
+  fracx = ((fracx & (ix >= 0)) | (ix > int32_t(sampler->width) - 2)) & 0x7F;
+  ix = clampCoord(ix, sampler->width - 1);
+
+  // Load the sample taps for the first sampler and combine rows.
+  auto abcd = linearRowTapsR8(sampler, ix, offsety, stridey, fracy);
+
+  // Load the sample taps for the second sampler and combine rows.
+  auto xyzw = linearRowTapsR8(sampler2, ix, offsety, stridey, fracy);
+
+  // We are left with a result vector for each sampler with values for adjacent
+  // pixels interleaved together in each. We need to unzip these values so that
+  // we can do the final horizontal multiply-add based on the X fraction.
+  auto abcdxyzwl = SHUFFLE(abcd, xyzw, 0, 8, 2, 10, 4, 12, 6, 14);
+  auto abcdxyzwh = SHUFFLE(abcd, xyzw, 1, 9, 3, 11, 5, 13, 7, 15);
+  abcdxyzwl += ((abcdxyzwh - abcdxyzwl) * CONVERT(fracx, I16).xxyyzzww) >> 7;
+
+  // The final result is the packed values for the first sampler interleaved
+  // with the packed values for the second sampler.
+  return abcdxyzwl;
+}
+
+// Casting to int loses some precision while stepping that can offset the
+// image, so shift the values by some extra bits of precision to minimize
+// this. We support up to 16 bits of image size, 7 bits of quantization,
+// and 1 bit for sign, which leaves 8 bits left for extra precision.
+const int STEP_BITS = 8;
+
+// Optimized version of textureLinearPackedR8 for Y R8 texture with
+// half-resolution paired U/V R8 textures. This allows us to more efficiently
+// pack YUV samples into vectors to substantially reduce math operations even
+// further.
+template <bool BLEND>
+static inline void upscaleYUV42R8(uint32_t* dest, int span, uint8_t* yRow,
+                                  I32 yU, int32_t yDU, int32_t yStrideV,
+                                  int16_t yFracV, uint8_t* cRow1,
+                                  uint8_t* cRow2, I32 cU, int32_t cDU,
+                                  int32_t cStrideV, int16_t cFracV,
+                                  const YUVMatrix& colorSpace) {
+  // As much as possible try to utilize the fact that we're only using half
+  // the UV samples to combine Y and UV samples into single vectors. Here we
+  // need to initialize several useful vector quantities for stepping fractional
+  // offsets. For the UV samples, we take the average of the first+second and
+  // third+fourth samples in a chunk which conceptually correspond to offsets
+  // 0.5 and 1.5 (in 0..2 range). This allows us to reconstruct intermediate
+  // samples 0.25, 0.75, 1.25, and 1.75 later. X fraction is shifted over into
+  // the top 7 bits of an unsigned short so that we can mask off the exact
+  // fractional bits we need to blend merely by right shifting them into
+  // position.
+  cU = (cU.xzxz + cU.ywyw) >> 1;
+  auto ycFracX = CONVERT(combine(yU, cU), V8<uint16_t>)
+                 << (16 - (STEP_BITS + 7));
+  auto ycFracDX = combine(I16(yDU), I16(cDU)) << (16 - (STEP_BITS + 7));
+  auto ycFracV = combine(I16(yFracV), I16(cFracV));
+  I32 yI = yU >> (STEP_BITS + 7);
+  I32 cI = cU >> (STEP_BITS + 7);
+  // Load initial combined YUV samples for each row and blend them.
+  auto ycSrc0 =
+      CONVERT(combine(unaligned_load<V4<uint8_t>>(&yRow[yI.x]),
+                      combine(unaligned_load<V2<uint8_t>>(&cRow1[cI.x]),
+                              unaligned_load<V2<uint8_t>>(&cRow2[cI.x]))),
+              V8<int16_t>);
+  auto ycSrc1 = CONVERT(
+      combine(unaligned_load<V4<uint8_t>>(&yRow[yI.x + yStrideV]),
+              combine(unaligned_load<V2<uint8_t>>(&cRow1[cI.x + cStrideV]),
+                      unaligned_load<V2<uint8_t>>(&cRow2[cI.x + cStrideV]))),
+      V8<int16_t>);
+  auto ycSrc = ycSrc0 + (((ycSrc1 - ycSrc0) * ycFracV) >> 7);
+
+  // Here we shift in results from the next sample while caching results from
+  // the previous sample. This allows us to reduce the multiplications in the
+  // inner loop down to only two since we just need to blend the new samples
+  // horizontally and then vertically once each.
+  for (uint32_t* end = dest + span; dest < end; dest += 4) {
+    yU += yDU;
+    I32 yIn = yU >> (STEP_BITS + 7);
+    cU += cDU;
+    I32 cIn = cU >> (STEP_BITS + 7);
+    // Load combined YUV samples for the next chunk on each row and blend them.
+    auto ycSrc0n =
+        CONVERT(combine(unaligned_load<V4<uint8_t>>(&yRow[yIn.x]),
+                        combine(unaligned_load<V2<uint8_t>>(&cRow1[cIn.x]),
+                                unaligned_load<V2<uint8_t>>(&cRow2[cIn.x]))),
+                V8<int16_t>);
+    auto ycSrc1n = CONVERT(
+        combine(unaligned_load<V4<uint8_t>>(&yRow[yIn.x + yStrideV]),
+                combine(unaligned_load<V2<uint8_t>>(&cRow1[cIn.x + cStrideV]),
+                        unaligned_load<V2<uint8_t>>(&cRow2[cIn.x + cStrideV]))),
+        V8<int16_t>);
+    auto ycSrcn = ycSrc0n + (((ycSrc1n - ycSrc0n) * ycFracV) >> 7);
+
+    // The source samples for the chunk may not match the actual tap offsets.
+    // Since we're upscaling, we know the tap offsets fall within all the
+    // samples in a 4-wide chunk. Since we can't rely on PSHUFB or similar,
+    // instead we do laborious shuffling here for the Y samples and then the UV
+    // samples.
+    auto yshuf = lowHalf(ycSrc);
+    auto yshufn =
+        SHUFFLE(yshuf, yIn.x == yI.w ? lowHalf(ycSrcn).yyyy : lowHalf(ycSrcn),
+                1, 2, 3, 4);
+    if (yI.y == yI.x) {
+      yshuf = yshuf.xxyz;
+      yshufn = yshufn.xxyz;
+    }
+    if (yI.z == yI.y) {
+      yshuf = yshuf.xyyz;
+      yshufn = yshufn.xyyz;
+    }
+    if (yI.w == yI.z) {
+      yshuf = yshuf.xyzz;
+      yshufn = yshufn.xyzz;
+    }
+
+    auto cshuf = highHalf(ycSrc);
+    auto cshufn =
+        SHUFFLE(cshuf, cIn.x == cI.y ? highHalf(ycSrcn).yyww : highHalf(ycSrcn),
+                1, 4, 3, 6);
+    if (cI.y == cI.x) {
+      cshuf = cshuf.xxzz;
+      cshufn = cshufn.xxzz;
+    }
+
+    // After shuffling, combine the Y and UV samples back into a single vector
+    // for blending. Shift X fraction into position as unsigned to mask off top
+    // bits and get rid of low bits to avoid multiplication overflow.
+    auto yuvPx = combine(yshuf, cshuf);
+    yuvPx += ((combine(yshufn, cshufn) - yuvPx) *
+              bit_cast<V8<int16_t>>(ycFracX >> (16 - 7))) >>
+             7;
+
+    // Cache the new samples as the current samples on the next iteration.
+    ycSrc = ycSrcn;
+    ycFracX += ycFracDX;
+    yI = yIn;
+    cI = cIn;
+
+    // De-interleave the Y and UV results. We need to average the UV results
+    // to produce values for intermediate samples. Taps for UV were collected at
+    // offsets 0.5 and 1.5, such that if we take a quarter of the difference
+    // (1.5-0.5)/4, subtract it from even samples, and add it to odd samples,
+    // we can estimate samples 0.25, 0.75, 1.25, and 1.75.
+    auto yPx = SHUFFLE(yuvPx, yuvPx, 0, 0, 1, 1, 2, 2, 3, 3);
+    auto uvPx = SHUFFLE(yuvPx, yuvPx, 4, 6, 4, 6, 5, 7, 5, 7) +
+                ((SHUFFLE(yuvPx, yuvPx, 4, 6, 5, 7, 4, 6, 5, 7) -
+                  SHUFFLE(yuvPx, yuvPx, 5, 7, 4, 6, 5, 7, 4, 6)) >>
+                 2);
+
+    commit_blend_span<BLEND>(dest, colorSpace.convert(yPx, uvPx));
+  }
+}
+
+// This is the inner loop driver of CompositeYUV that processes an axis-aligned
+// YUV span, dispatching based on appropriate format and scaling. This is also
+// reused by blendYUV to accelerate some cases of texture sampling in the
+// shader.
+template <bool BLEND = false>
+static void linear_row_yuv(uint32_t* dest, int span, sampler2DRect samplerY,
+                           const vec2_scalar& srcUV, float srcDU,
+                           sampler2DRect samplerU, sampler2DRect samplerV,
+                           const vec2_scalar& chromaUV, float chromaDU,
+                           int colorDepth, const YUVMatrix& colorSpace) {
+  // Calculate varying and constant interp data for Y plane.
+  I32 yU = cast(init_interp(srcUV.x, srcDU) * (1 << STEP_BITS));
+  int32_t yV = int32_t(srcUV.y);
+
+  // Calculate varying and constant interp data for chroma planes.
+  I32 cU = cast(init_interp(chromaUV.x, chromaDU) * (1 << STEP_BITS));
+  int32_t cV = int32_t(chromaUV.y);
+
+  // We need to skip 4 pixels per chunk.
+  int32_t yDU = int32_t((4 << STEP_BITS) * srcDU);
+  int32_t cDU = int32_t((4 << STEP_BITS) * chromaDU);
+
+  if (samplerY->width < 2 || samplerU->width < 2) {
+    // If the source row has less than 2 pixels, it's not safe to use a linear
+    // filter because it may overread the row. Just convert the single pixel
+    // with nearest filtering and fill the row with it.
+    Float yuvF = {texelFetch(samplerY, ivec2(srcUV)).x.x,
+                  texelFetch(samplerU, ivec2(chromaUV)).x.x,
+                  texelFetch(samplerV, ivec2(chromaUV)).x.x, 1.0f};
+    // If this is an HDR LSB format, we need to renormalize the result.
+    if (colorDepth > 8) {
+      int rescaleFactor = 16 - colorDepth;
+      yuvF *= float(1 << rescaleFactor);
+    }
+    I16 yuv = CONVERT(round_pixel(yuvF), I16);
+    commit_solid_span<BLEND>(
+        dest,
+        unpack(colorSpace.convert(V8<int16_t>(yuv.x),
+                                  zip(I16(yuv.y), I16(yuv.z)))),
+        span);
+  } else if (samplerY->format == TextureFormat::R16) {
+    // Sample each YUV plane, rescale it to fit in low 8 bits of word, and
+    // then transform them by the appropriate color space.
+    assert(colorDepth > 8);
+    // Need to right shift the sample by the amount of bits over 8 it
+    // occupies. On output from textureLinearUnpackedR16, we have lost 1 bit
+    // of precision at the low end already, hence 1 is subtracted from the
+    // color depth.
+    int rescaleBits = (colorDepth - 1) - 8;
+    for (; span >= 4; span -= 4) {
+      auto yPx =
+          textureLinearUnpackedR16(samplerY, ivec2(yU >> STEP_BITS, yV)) >>
+          rescaleBits;
+      auto uPx =
+          textureLinearUnpackedR16(samplerU, ivec2(cU >> STEP_BITS, cV)) >>
+          rescaleBits;
+      auto vPx =
+          textureLinearUnpackedR16(samplerV, ivec2(cU >> STEP_BITS, cV)) >>
+          rescaleBits;
+      commit_blend_span<BLEND>(
+          dest, colorSpace.convert(zip(yPx, yPx), zip(uPx, vPx)));
+      dest += 4;
+      yU += yDU;
+      cU += cDU;
+    }
+    if (span > 0) {
+      // Handle any remaining pixels...
+      auto yPx =
+          textureLinearUnpackedR16(samplerY, ivec2(yU >> STEP_BITS, yV)) >>
+          rescaleBits;
+      auto uPx =
+          textureLinearUnpackedR16(samplerU, ivec2(cU >> STEP_BITS, cV)) >>
+          rescaleBits;
+      auto vPx =
+          textureLinearUnpackedR16(samplerV, ivec2(cU >> STEP_BITS, cV)) >>
+          rescaleBits;
+      commit_blend_span<BLEND>(
+          dest, colorSpace.convert(zip(yPx, yPx), zip(uPx, vPx)), span);
+    }
+  } else {
+    assert(samplerY->format == TextureFormat::R8);
+    assert(colorDepth == 8);
+
+    // Calculate varying and constant interp data for Y plane.
+    int16_t yFracV = yV & 0x7F;
+    yV >>= 7;
+    int32_t yOffsetV = clampCoord(yV, samplerY->height) * samplerY->stride;
+    int32_t yStrideV =
+        yV >= 0 && yV < int32_t(samplerY->height) - 1 ? samplerY->stride : 0;
+
+    // Calculate varying and constant interp data for chroma planes.
+    int16_t cFracV = cV & 0x7F;
+    cV >>= 7;
+    int32_t cOffsetV = clampCoord(cV, samplerU->height) * samplerU->stride;
+    int32_t cStrideV =
+        cV >= 0 && cV < int32_t(samplerU->height) - 1 ? samplerU->stride : 0;
+
+    // If we're sampling the UV planes at half the resolution of the Y plane,
+    // then try to use half resolution fast-path.
+    if (yDU >= cDU && cDU > 0 && yDU <= (4 << (STEP_BITS + 7)) &&
+        cDU <= (2 << (STEP_BITS + 7))) {
+      // Ensure that samples don't fall outside of the valid bounds of each
+      // planar texture. Step until the initial X coordinates are positive.
+      for (; (yU.x < 0 || cU.x < 0) && span >= 4; span -= 4) {
+        auto yPx = textureLinearRowR8(samplerY, yU >> STEP_BITS, yOffsetV,
+                                      yStrideV, yFracV);
+        auto uvPx = textureLinearRowPairedR8(
+            samplerU, samplerV, cU >> STEP_BITS, cOffsetV, cStrideV, cFracV);
+        commit_blend_span<BLEND>(dest, colorSpace.convert(yPx, uvPx));
+        dest += 4;
+        yU += yDU;
+        cU += cDU;
+      }
+      // Calculate the number of aligned chunks that we can step inside the
+      // bounds of each planar texture without overreading.
+      int inside = min(
+          min((((int(samplerY->width) - 4) << (STEP_BITS + 7)) - yU.x) / yDU,
+              (((int(samplerU->width) - 4) << (STEP_BITS + 7)) - cU.x) / cDU) *
+              4,
+          span & ~3);
+      if (inside > 0) {
+        uint8_t* yRow = (uint8_t*)samplerY->buf + yOffsetV;
+        uint8_t* cRow1 = (uint8_t*)samplerU->buf + cOffsetV;
+        uint8_t* cRow2 = (uint8_t*)samplerV->buf + cOffsetV;
+        upscaleYUV42R8<BLEND>(dest, inside, yRow, yU, yDU, yStrideV, yFracV,
+                              cRow1, cRow2, cU, cDU, cStrideV, cFracV,
+                              colorSpace);
+        span -= inside;
+        dest += inside;
+        yU += (inside / 4) * yDU;
+        cU += (inside / 4) * cDU;
+      }
+      // If there are any remaining chunks that weren't inside, handle them
+      // below.
+    }
+    for (; span >= 4; span -= 4) {
+      // Sample each YUV plane and then transform them by the appropriate
+      // color space.
+      auto yPx = textureLinearRowR8(samplerY, yU >> STEP_BITS, yOffsetV,
+                                    yStrideV, yFracV);
+      auto uvPx = textureLinearRowPairedR8(samplerU, samplerV, cU >> STEP_BITS,
+                                           cOffsetV, cStrideV, cFracV);
+      commit_blend_span<BLEND>(dest, colorSpace.convert(yPx, uvPx));
+      dest += 4;
+      yU += yDU;
+      cU += cDU;
+    }
+    if (span > 0) {
+      // Handle any remaining pixels...
+      auto yPx = textureLinearRowR8(samplerY, yU >> STEP_BITS, yOffsetV,
+                                    yStrideV, yFracV);
+      auto uvPx = textureLinearRowPairedR8(samplerU, samplerV, cU >> STEP_BITS,
+                                           cOffsetV, cStrideV, cFracV);
+      commit_blend_span<BLEND>(dest, colorSpace.convert(yPx, uvPx), span);
+    }
+  }
+}
+
+static void linear_convert_yuv(Texture& ytex, Texture& utex, Texture& vtex,
+                               const YUVMatrix& rgbFromYcbcr, int colorDepth,
+                               const IntRect& srcReq, Texture& dsttex,
+                               const IntRect& dstReq, bool invertX,
+                               bool invertY, const IntRect& clipRect) {
+  // Compute valid dest bounds
+  IntRect dstBounds = dsttex.sample_bounds(dstReq);
+  dstBounds.intersect(clipRect);
+  // Check if sampling bounds are empty
+  if (dstBounds.is_empty()) {
+    return;
+  }
+  // Initialize samplers for source textures
+  sampler2DRect_impl sampler[3];
+  init_sampler(&sampler[0], ytex);
+  init_sampler(&sampler[1], utex);
+  init_sampler(&sampler[2], vtex);
+
+  // Compute source UVs
+  vec2_scalar srcUV(srcReq.x0, srcReq.y0);
+  vec2_scalar srcDUV(float(srcReq.width()) / dstReq.width(),
+                     float(srcReq.height()) / dstReq.height());
+  if (invertX) {
+    // Advance to the end of the row and flip the step.
+    srcUV.x += srcReq.width();
+    srcDUV.x = -srcDUV.x;
+  }
+  // Inverted Y must step downward along source rows
+  if (invertY) {
+    srcUV.y += srcReq.height();
+    srcDUV.y = -srcDUV.y;
+  }
+  // Skip to clamped source start
+  srcUV += srcDUV * (vec2_scalar(dstBounds.x0, dstBounds.y0) + 0.5f);
+  // Calculate separate chroma UVs for chroma planes with different scale
+  vec2_scalar chromaScale(float(utex.width) / ytex.width,
+                          float(utex.height) / ytex.height);
+  vec2_scalar chromaUV = srcUV * chromaScale;
+  vec2_scalar chromaDUV = srcDUV * chromaScale;
+  // Scale UVs by lerp precision. If the row has only 1 pixel, then don't
+  // quantize so that we can use nearest filtering instead to avoid overreads.
+  if (ytex.width >= 2 && utex.width >= 2) {
+    srcUV = linearQuantize(srcUV, 128);
+    srcDUV *= 128.0f;
+    chromaUV = linearQuantize(chromaUV, 128);
+    chromaDUV *= 128.0f;
+  }
+  // Calculate dest pointer from clamped offsets
+  int destStride = dsttex.stride();
+  char* dest = dsttex.sample_ptr(dstReq, dstBounds);
+  int span = dstBounds.width();
+  for (int rows = dstBounds.height(); rows > 0; rows--) {
+    linear_row_yuv((uint32_t*)dest, span, &sampler[0], srcUV, srcDUV.x,
+                   &sampler[1], &sampler[2], chromaUV, chromaDUV.x, colorDepth,
+                   rgbFromYcbcr);
+    dest += destStride;
+    srcUV.y += srcDUV.y;
+    chromaUV.y += chromaDUV.y;
+  }
+}
+
+// -
+// This section must match gfx/2d/Types.h
+
+enum class YUVRangedColorSpace : uint8_t {
+  BT601_Narrow = 0,
+  BT601_Full,
+  BT709_Narrow,
+  BT709_Full,
+  BT2020_Narrow,
+  BT2020_Full,
+  GbrIdentity,
+};
+
+// -
+// This section must match yuv.glsl
+
+vec4_scalar get_ycbcr_zeros_ones(const YUVRangedColorSpace color_space,
+                                 const GLuint color_depth) {
+  // For SWGL's 8bpc-only pipeline, our extra care here probably doesn't matter.
+  // However, technically e.g. 10-bit achromatic zero for cb and cr is
+  // (128 << 2) / ((1 << 10) - 1) = 512 / 1023, which != 128 / 255, and affects
+  // our matrix values subtly. Maybe not enough to matter? But it's the most
+  // correct thing to do.
+  // Unlike the glsl version, our texture samples are u8([0,255]) not
+  // u16([0,1023]) though.
+  switch (color_space) {
+    case YUVRangedColorSpace::BT601_Narrow:
+    case YUVRangedColorSpace::BT709_Narrow:
+    case YUVRangedColorSpace::BT2020_Narrow: {
+      auto extra_bit_count = color_depth - 8;
+      vec4_scalar zo = {
+          float(16 << extra_bit_count),
+          float(128 << extra_bit_count),
+          float(235 << extra_bit_count),
+          float(240 << extra_bit_count),
+      };
+      float all_bits = (1 << color_depth) - 1;
+      zo /= all_bits;
+      return zo;
+    }
+
+    case YUVRangedColorSpace::BT601_Full:
+    case YUVRangedColorSpace::BT709_Full:
+    case YUVRangedColorSpace::BT2020_Full: {
+      const auto narrow =
+          get_ycbcr_zeros_ones(YUVRangedColorSpace::BT601_Narrow, color_depth);
+      return {0.0, narrow.y, 1.0, 1.0};
+    }
+
+    case YUVRangedColorSpace::GbrIdentity:
+      break;
+  }
+  return {0.0, 0.0, 1.0, 1.0};
+}
+
+constexpr mat3_scalar RgbFromYuv_Rec601 = {
+    {1.00000, 1.00000, 1.00000},
+    {0.00000, -0.17207, 0.88600},
+    {0.70100, -0.35707, 0.00000},
+};
+constexpr mat3_scalar RgbFromYuv_Rec709 = {
+    {1.00000, 1.00000, 1.00000},
+    {0.00000, -0.09366, 0.92780},
+    {0.78740, -0.23406, 0.00000},
+};
+constexpr mat3_scalar RgbFromYuv_Rec2020 = {
+    {1.00000, 1.00000, 1.00000},
+    {0.00000, -0.08228, 0.94070},
+    {0.73730, -0.28568, 0.00000},
+};
+constexpr mat3_scalar RgbFromYuv_GbrIdentity = {
+    {0, 1, 0},
+    {0, 0, 1},
+    {1, 0, 0},
+};
+
+inline mat3_scalar get_rgb_from_yuv(const YUVRangedColorSpace color_space) {
+  switch (color_space) {
+    case YUVRangedColorSpace::BT601_Narrow:
+    case YUVRangedColorSpace::BT601_Full:
+      return RgbFromYuv_Rec601;
+    case YUVRangedColorSpace::BT709_Narrow:
+    case YUVRangedColorSpace::BT709_Full:
+      return RgbFromYuv_Rec709;
+    case YUVRangedColorSpace::BT2020_Narrow:
+    case YUVRangedColorSpace::BT2020_Full:
+      return RgbFromYuv_Rec2020;
+    case YUVRangedColorSpace::GbrIdentity:
+      break;
+  }
+  return RgbFromYuv_GbrIdentity;
+}
+
+struct YcbcrInfo final {
+  vec3_scalar ycbcr_bias;
+  mat3_scalar rgb_from_debiased_ycbcr;
+};
+
+inline YcbcrInfo get_ycbcr_info(const YUVRangedColorSpace color_space,
+                                GLuint color_depth) {
+  // SWGL always does 8bpc math, so don't scale the matrix for 10bpc!
+  color_depth = 8;
+
+  const auto zeros_ones = get_ycbcr_zeros_ones(color_space, color_depth);
+  const auto zeros = vec2_scalar{zeros_ones.x, zeros_ones.y};
+  const auto ones = vec2_scalar{zeros_ones.z, zeros_ones.w};
+  const auto scale = 1.0f / (ones - zeros);
+
+  const auto rgb_from_yuv = get_rgb_from_yuv(color_space);
+  const mat3_scalar yuv_from_debiased_ycbcr = {
+      {scale.x, 0, 0},
+      {0, scale.y, 0},
+      {0, 0, scale.y},
+  };
+
+  YcbcrInfo ret;
+  ret.ycbcr_bias = {zeros.x, zeros.y, zeros.y};
+  ret.rgb_from_debiased_ycbcr = rgb_from_yuv * yuv_from_debiased_ycbcr;
+  return ret;
+}
+
+// -
+
+extern "C" {
+
+// Extension for compositing a YUV surface represented by separate YUV planes
+// to a BGRA destination. The supplied color space is used to determine the
+// transform from YUV to BGRA after sampling.
+void CompositeYUV(LockedTexture* lockedDst, LockedTexture* lockedY,
+                  LockedTexture* lockedU, LockedTexture* lockedV,
+                  YUVRangedColorSpace colorSpace, GLuint colorDepth, GLint srcX,
+                  GLint srcY, GLsizei srcWidth, GLsizei srcHeight, GLint dstX,
+                  GLint dstY, GLsizei dstWidth, GLsizei dstHeight,
+                  GLboolean flipX, GLboolean flipY, GLint clipX, GLint clipY,
+                  GLsizei clipWidth, GLsizei clipHeight) {
+  if (!lockedDst || !lockedY || !lockedU || !lockedV) {
+    return;
+  }
+  if (colorSpace > YUVRangedColorSpace::GbrIdentity) {
+    assert(false);
+    return;
+  }
+  const auto ycbcrInfo = get_ycbcr_info(colorSpace, colorDepth);
+  const auto rgbFromYcbcr =
+      YUVMatrix::From(ycbcrInfo.ycbcr_bias, ycbcrInfo.rgb_from_debiased_ycbcr);
+
+  Texture& ytex = *lockedY;
+  Texture& utex = *lockedU;
+  Texture& vtex = *lockedV;
+  Texture& dsttex = *lockedDst;
+  // All YUV planes must currently be represented by R8 or R16 textures.
+  // The chroma (U/V) planes must have matching dimensions.
+  assert(ytex.bpp() == utex.bpp() && ytex.bpp() == vtex.bpp());
+  assert((ytex.bpp() == 1 && colorDepth == 8) ||
+         (ytex.bpp() == 2 && colorDepth > 8));
+  // assert(ytex.width == utex.width && ytex.height == utex.height);
+  assert(utex.width == vtex.width && utex.height == vtex.height);
+  assert(ytex.offset == utex.offset && ytex.offset == vtex.offset);
+  assert(dsttex.bpp() == 4);
+
+  IntRect srcReq =
+      IntRect{srcX, srcY, srcX + srcWidth, srcY + srcHeight} - ytex.offset;
+  IntRect dstReq =
+      IntRect{dstX, dstY, dstX + dstWidth, dstY + dstHeight} - dsttex.offset;
+  if (srcReq.is_empty() || dstReq.is_empty()) {
+    return;
+  }
+
+  // Compute clip rect as relative to the dstReq, as that's the same coords
+  // as used for the sampling bounds.
+  IntRect clipRect = {clipX - dstX, clipY - dstY, clipX - dstX + clipWidth,
+                      clipY - dstY + clipHeight};
+  // For now, always use a linear filter path that would be required for
+  // scaling. Further fast-paths for non-scaled video might be desirable in the
+  // future.
+  linear_convert_yuv(ytex, utex, vtex, rgbFromYcbcr, colorDepth, srcReq, dsttex,
+                     dstReq, flipX, flipY, clipRect);
+}
+
+}  // extern "C"
diff --git a/gfx/wr/swgl/src/gl.cc b/gfx/wr/swgl/src/gl.cc
new file mode 100644
index 0000000000..0b09e29008
--- /dev/null
+++ b/gfx/wr/swgl/src/gl.cc
@@ -0,0 +1,2851 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#include <stdlib.h>
+#include <stdint.h>
+#include <string.h>
+#include <assert.h>
+#include <stdio.h>
+#include <math.h>
+
+#ifdef __MACH__
+#  include <mach/mach.h>
+#  include <mach/mach_time.h>
+#else
+#  include <time.h>
+#endif
+
+#ifdef NDEBUG
+#  define debugf(...)
+#else
+#  define debugf(...) printf(__VA_ARGS__)
+#endif
+
+// #define PRINT_TIMINGS
+
+#ifdef _WIN32
+#  define ALWAYS_INLINE __forceinline
+#  define NO_INLINE __declspec(noinline)
+
+// Including Windows.h brings a huge amount of namespace polution so just
+// define a couple of things manually
+typedef int BOOL;
+#  define WINAPI __stdcall
+#  define DECLSPEC_IMPORT __declspec(dllimport)
+#  define WINBASEAPI DECLSPEC_IMPORT
+typedef unsigned long DWORD;
+typedef long LONG;
+typedef __int64 LONGLONG;
+#  define DUMMYSTRUCTNAME
+
+typedef union _LARGE_INTEGER {
+  struct {
+    DWORD LowPart;
+    LONG HighPart;
+  } DUMMYSTRUCTNAME;
+  struct {
+    DWORD LowPart;
+    LONG HighPart;
+  } u;
+  LONGLONG QuadPart;
+} LARGE_INTEGER;
+extern "C" {
+WINBASEAPI BOOL WINAPI
+QueryPerformanceCounter(LARGE_INTEGER* lpPerformanceCount);
+
+WINBASEAPI BOOL WINAPI QueryPerformanceFrequency(LARGE_INTEGER* lpFrequency);
+}
+
+#else
+// GCC is slower when dealing with always_inline, especially in debug builds.
+// When using Clang, use always_inline more aggressively.
+#  if defined(__clang__) || defined(NDEBUG)
+#    define ALWAYS_INLINE __attribute__((always_inline)) inline
+#  else
+#    define ALWAYS_INLINE inline
+#  endif
+#  define NO_INLINE __attribute__((noinline))
+#endif
+
+// Some functions may cause excessive binary bloat if inlined in debug or with
+// GCC builds, so use PREFER_INLINE on these instead of ALWAYS_INLINE.
+#if defined(__clang__) && defined(NDEBUG)
+#  define PREFER_INLINE ALWAYS_INLINE
+#else
+#  define PREFER_INLINE inline
+#endif
+
+#define UNREACHABLE __builtin_unreachable()
+
+#define UNUSED [[maybe_unused]]
+
+#define FALLTHROUGH [[fallthrough]]
+
+#if defined(MOZILLA_CLIENT) && defined(MOZ_CLANG_PLUGIN)
+#  define IMPLICIT __attribute__((annotate("moz_implicit")))
+#else
+#  define IMPLICIT
+#endif
+
+#include "gl_defs.h"
+#include "glsl.h"
+#include "program.h"
+#include "texture.h"
+
+using namespace glsl;
+
+typedef ivec2_scalar IntPoint;
+
+struct IntRect {
+  int x0;
+  int y0;
+  int x1;
+  int y1;
+
+  IntRect() : x0(0), y0(0), x1(0), y1(0) {}
+  IntRect(int x0, int y0, int x1, int y1) : x0(x0), y0(y0), x1(x1), y1(y1) {}
+  IntRect(IntPoint origin, IntPoint size)
+      : x0(origin.x),
+        y0(origin.y),
+        x1(origin.x + size.x),
+        y1(origin.y + size.y) {}
+
+  int width() const { return x1 - x0; }
+  int height() const { return y1 - y0; }
+  bool is_empty() const { return width() <= 0 || height() <= 0; }
+
+  IntPoint origin() const { return IntPoint(x0, y0); }
+
+  bool same_size(const IntRect& o) const {
+    return width() == o.width() && height() == o.height();
+  }
+
+  bool contains(const IntRect& o) const {
+    return o.x0 >= x0 && o.y0 >= y0 && o.x1 <= x1 && o.y1 <= y1;
+  }
+
+  IntRect& intersect(const IntRect& o) {
+    x0 = max(x0, o.x0);
+    y0 = max(y0, o.y0);
+    x1 = min(x1, o.x1);
+    y1 = min(y1, o.y1);
+    return *this;
+  }
+
+  IntRect intersection(const IntRect& o) {
+    IntRect result = *this;
+    result.intersect(o);
+    return result;
+  }
+
+  // Scale from source-space to dest-space, optionally rounding inward
+  IntRect& scale(int srcWidth, int srcHeight, int dstWidth, int dstHeight,
+                 bool roundIn = false) {
+    x0 = (x0 * dstWidth + (roundIn ? srcWidth - 1 : 0)) / srcWidth;
+    y0 = (y0 * dstHeight + (roundIn ? srcHeight - 1 : 0)) / srcHeight;
+    x1 = (x1 * dstWidth) / srcWidth;
+    y1 = (y1 * dstHeight) / srcHeight;
+    return *this;
+  }
+
+  // Flip the rect's Y coords around inflection point at Y=offset
+  void invert_y(int offset) {
+    y0 = offset - y0;
+    y1 = offset - y1;
+    swap(y0, y1);
+  }
+
+  IntRect& offset(const IntPoint& o) {
+    x0 += o.x;
+    y0 += o.y;
+    x1 += o.x;
+    y1 += o.y;
+    return *this;
+  }
+
+  IntRect operator+(const IntPoint& o) const {
+    return IntRect(*this).offset(o);
+  }
+  IntRect operator-(const IntPoint& o) const {
+    return IntRect(*this).offset(-o);
+  }
+};
+
+typedef vec2_scalar Point2D;
+typedef vec4_scalar Point3D;
+
+struct IntRange {
+  int start;
+  int end;
+
+  int len() const { return end - start; }
+
+  IntRange intersect(IntRange r) const {
+    return {max(start, r.start), min(end, r.end)};
+  }
+};
+
+struct FloatRange {
+  float start;
+  float end;
+
+  float clip(float x) const { return clamp(x, start, end); }
+
+  FloatRange clip(FloatRange r) const { return {clip(r.start), clip(r.end)}; }
+
+  FloatRange merge(FloatRange r) const {
+    return {min(start, r.start), max(end, r.end)};
+  }
+
+  IntRange round() const {
+    return {int(floor(start + 0.5f)), int(floor(end + 0.5f))};
+  }
+
+  IntRange round_out() const { return {int(floor(start)), int(ceil(end))}; }
+};
+
+template <typename P>
+static inline FloatRange x_range(P p0, P p1) {
+  return {min(p0.x, p1.x), max(p0.x, p1.x)};
+}
+
+struct VertexAttrib {
+  size_t size = 0;  // in bytes
+  GLenum type = 0;
+  bool normalized = false;
+  GLsizei stride = 0;
+  GLuint offset = 0;
+  bool enabled = false;
+  GLuint divisor = 0;
+  int vertex_array = 0;
+  int vertex_buffer = 0;
+  char* buf = nullptr;  // XXX: this can easily dangle
+  size_t buf_size = 0;  // this will let us bounds check
+
+  // Mark the buffer as invalid so we don't accidentally use stale data.
+  void disable() {
+    enabled = false;
+    buf = nullptr;
+    buf_size = 0;
+  }
+};
+
+static int bytes_for_internal_format(GLenum internal_format) {
+  switch (internal_format) {
+    case GL_RGBA32F:
+      return 4 * 4;
+    case GL_RGBA32I:
+      return 4 * 4;
+    case GL_RGBA8:
+    case GL_BGRA8:
+    case GL_RGBA:
+      return 4;
+    case GL_R8:
+    case GL_RED:
+      return 1;
+    case GL_RG8:
+    case GL_RG:
+      return 2;
+    case GL_DEPTH_COMPONENT:
+    case GL_DEPTH_COMPONENT16:
+    case GL_DEPTH_COMPONENT24:
+    case GL_DEPTH_COMPONENT32:
+      return 4;
+    case GL_RGB_RAW_422_APPLE:
+      return 2;
+    case GL_R16:
+      return 2;
+    case GL_RG16:
+      return 4;
+    default:
+      debugf("internal format: %x\n", internal_format);
+      assert(0);
+      return 0;
+  }
+}
+
+static inline int aligned_stride(int row_bytes) { return (row_bytes + 3) & ~3; }
+
+static TextureFormat gl_format_to_texture_format(int type) {
+  switch (type) {
+    case GL_RGBA32F:
+      return TextureFormat::RGBA32F;
+    case GL_RGBA32I:
+      return TextureFormat::RGBA32I;
+    case GL_RGBA8:
+      return TextureFormat::RGBA8;
+    case GL_R8:
+      return TextureFormat::R8;
+    case GL_RG8:
+      return TextureFormat::RG8;
+    case GL_R16:
+      return TextureFormat::R16;
+    case GL_RG16:
+      return TextureFormat::RG16;
+    case GL_RGB_RAW_422_APPLE:
+      return TextureFormat::YUV422;
+    default:
+      assert(0);
+      return TextureFormat::RGBA8;
+  }
+}
+
+struct Query {
+  uint64_t value = 0;
+};
+
+struct Buffer {
+  char* buf = nullptr;
+  size_t size = 0;
+  size_t capacity = 0;
+
+  // Returns true if re-allocation succeeded, false otherwise...
+  bool allocate(size_t new_size) {
+    // If the size remains unchanged, don't allocate anything.
+    if (new_size == size) {
+      return true;
+    }
+    // If the new size is within the existing capacity of the buffer, just
+    // reuse the existing buffer.
+    if (new_size <= capacity) {
+      size = new_size;
+      return true;
+    }
+    // Otherwise we need to reallocate the buffer to hold up to the requested
+    // larger size.
+    char* new_buf = (char*)realloc(buf, new_size);
+    assert(new_buf);
+    if (!new_buf) {
+      // If we fail, null out the buffer rather than leave around the old
+      // allocation state.
+      cleanup();
+      return false;
+    }
+    // The reallocation succeeded, so install the buffer.
+    buf = new_buf;
+    size = new_size;
+    capacity = new_size;
+    return true;
+  }
+
+  void cleanup() {
+    if (buf) {
+      free(buf);
+      buf = nullptr;
+      size = 0;
+      capacity = 0;
+    }
+  }
+
+  ~Buffer() { cleanup(); }
+};
+
+struct Framebuffer {
+  GLuint color_attachment = 0;
+  GLuint depth_attachment = 0;
+};
+
+struct Renderbuffer {
+  GLuint texture = 0;
+
+  void on_erase();
+};
+
+TextureFilter gl_filter_to_texture_filter(int type) {
+  switch (type) {
+    case GL_NEAREST:
+      return TextureFilter::NEAREST;
+    case GL_NEAREST_MIPMAP_LINEAR:
+      return TextureFilter::NEAREST;
+    case GL_NEAREST_MIPMAP_NEAREST:
+      return TextureFilter::NEAREST;
+    case GL_LINEAR:
+      return TextureFilter::LINEAR;
+    case GL_LINEAR_MIPMAP_LINEAR:
+      return TextureFilter::LINEAR;
+    case GL_LINEAR_MIPMAP_NEAREST:
+      return TextureFilter::LINEAR;
+    default:
+      assert(0);
+      return TextureFilter::NEAREST;
+  }
+}
+
+struct Texture {
+  GLenum internal_format = 0;
+  int width = 0;
+  int height = 0;
+  char* buf = nullptr;
+  size_t buf_size = 0;
+  uint32_t buf_stride = 0;
+  uint8_t buf_bpp = 0;
+  GLenum min_filter = GL_NEAREST;
+  GLenum mag_filter = GL_LINEAR;
+  // The number of active locks on this texture. If this texture has any active
+  // locks, we need to disallow modifying or destroying the texture as it may
+  // be accessed by other threads where modifications could lead to races.
+  int32_t locked = 0;
+  // When used as an attachment of a framebuffer, rendering to the texture
+  // behaves as if it is located at the given offset such that the offset is
+  // subtracted from all transformed vertexes after the viewport is applied.
+  IntPoint offset;
+
+  enum FLAGS {
+    // If the buffer is internally-allocated by SWGL
+    SHOULD_FREE = 1 << 1,
+    // If the buffer has been cleared to initialize it. Currently this is only
+    // utilized by depth buffers which need to know when depth runs have reset
+    // to a valid row state. When unset, the depth runs may contain garbage.
+    CLEARED = 1 << 2,
+  };
+  int flags = SHOULD_FREE;
+  bool should_free() const { return bool(flags & SHOULD_FREE); }
+  bool cleared() const { return bool(flags & CLEARED); }
+
+  void set_flag(int flag, bool val) {
+    if (val) {
+      flags |= flag;
+    } else {
+      flags &= ~flag;
+    }
+  }
+  void set_should_free(bool val) {
+    // buf must be null before SHOULD_FREE can be safely toggled. Otherwise, we
+    // might accidentally mistakenly realloc an externally allocated buffer as
+    // if it were an internally allocated one.
+    assert(!buf);
+    set_flag(SHOULD_FREE, val);
+  }
+  void set_cleared(bool val) { set_flag(CLEARED, val); }
+
+  // Delayed-clearing state. When a clear of an FB is requested, we don't
+  // immediately clear each row, as the rows may be subsequently overwritten
+  // by draw calls, allowing us to skip the work of clearing the affected rows
+  // either fully or partially. Instead, we keep a bit vector of rows that need
+  // to be cleared later and save the value they need to be cleared with so
+  // that we can clear these rows individually when they are touched by draws.
+  // This currently only works for 2D textures, but not on texture arrays.
+  int delay_clear = 0;
+  uint32_t clear_val = 0;
+  uint32_t* cleared_rows = nullptr;
+
+  void init_depth_runs(uint32_t z);
+  void fill_depth_runs(uint32_t z, const IntRect& scissor);
+
+  void enable_delayed_clear(uint32_t val) {
+    delay_clear = height;
+    clear_val = val;
+    if (!cleared_rows) {
+      cleared_rows = new uint32_t[(height + 31) / 32];
+    }
+    memset(cleared_rows, 0, ((height + 31) / 32) * sizeof(uint32_t));
+    if (height & 31) {
+      cleared_rows[height / 32] = ~0U << (height & 31);
+    }
+  }
+
+  void disable_delayed_clear() {
+    if (cleared_rows) {
+      delete[] cleared_rows;
+      cleared_rows = nullptr;
+      delay_clear = 0;
+    }
+  }
+
+  int bpp() const { return buf_bpp; }
+  int compute_bpp() const { return bytes_for_internal_format(internal_format); }
+
+  size_t stride() const { return buf_stride; }
+  size_t compute_stride(int bpp, int width) const {
+    return aligned_stride(bpp * width);
+  }
+
+  // Set an external backing buffer of this texture.
+  void set_buffer(void* new_buf, size_t new_stride) {
+    assert(!should_free());
+    // Ensure that the supplied stride is at least as big as the row data and
+    // is aligned to the smaller of either the BPP or word-size. We need to at
+    // least be able to sample data from within a row and sample whole pixels
+    // of smaller formats without risking unaligned access.
+    int new_bpp = compute_bpp();
+    assert(new_stride >= size_t(new_bpp * width) &&
+           new_stride % min(new_bpp, sizeof(uint32_t)) == 0);
+
+    buf = (char*)new_buf;
+    buf_size = 0;
+    buf_bpp = new_bpp;
+    buf_stride = new_stride;
+  }
+
+  // Returns true if re-allocation succeeded, false otherwise...
+  bool allocate(bool force = false, int min_width = 0, int min_height = 0) {
+    assert(!locked);  // Locked textures shouldn't be reallocated
+    // If we get here, some GL API call that invalidates the texture was used.
+    // Mark the buffer as not-cleared to signal this.
+    set_cleared(false);
+    // Check if there is either no buffer currently or if we forced validation
+    // of the buffer size because some dimension might have changed.
+    if ((!buf || force) && should_free()) {
+      // Compute the buffer's BPP and stride, since they may have changed.
+      int new_bpp = compute_bpp();
+      size_t new_stride = compute_stride(new_bpp, width);
+      // Compute new size based on the maximum potential stride, rather than
+      // the current stride, to hopefully avoid reallocations when size would
+      // otherwise change too much...
+      size_t max_stride = compute_stride(new_bpp, max(width, min_width));
+      size_t size = max_stride * max(height, min_height);
+      if ((!buf && size > 0) || size > buf_size) {
+        // Allocate with a SIMD register-sized tail of padding at the end so we
+        // can safely read or write past the end of the texture with SIMD ops.
+        // Currently only the flat Z-buffer texture needs this padding due to
+        // full-register loads and stores in check_depth and discard_depth. In
+        // case some code in the future accidentally uses a linear filter on a
+        // texture with less than 2 pixels per row, we also add this padding
+        // just to be safe. All other texture types and use-cases should be
+        // safe to omit padding.
+        size_t padding =
+            internal_format == GL_DEPTH_COMPONENT24 || max(width, min_width) < 2
+                ? sizeof(Float)
+                : 0;
+        char* new_buf = (char*)realloc(buf, size + padding);
+        assert(new_buf);
+        if (!new_buf) {
+          // Allocation failed, so ensure we don't leave stale buffer state.
+          cleanup();
+          return false;
+        }
+        // Successfully reallocated the buffer, so go ahead and set it.
+        buf = new_buf;
+        buf_size = size;
+      }
+      // Set the BPP and stride in case they changed.
+      buf_bpp = new_bpp;
+      buf_stride = new_stride;
+    }
+    // Allocation succeeded or nothing changed...
+    return true;
+  }
+
+  void cleanup() {
+    assert(!locked);  // Locked textures shouldn't be destroyed
+    if (buf) {
+      // If we need to toggle SHOULD_FREE state, ensure that buf is nulled out,
+      // regardless of whether we internally allocated it. This will prevent us
+      // from wrongly treating buf as having been internally allocated for when
+      // we go to realloc if it actually was externally allocted.
+      if (should_free()) {
+        free(buf);
+      }
+      buf = nullptr;
+      buf_size = 0;
+      buf_bpp = 0;
+      buf_stride = 0;
+    }
+    disable_delayed_clear();
+  }
+
+  ~Texture() { cleanup(); }
+
+  IntRect bounds() const { return IntRect{0, 0, width, height}; }
+  IntRect offset_bounds() const { return bounds() + offset; }
+
+  // Find the valid sampling bounds relative to the requested region
+  IntRect sample_bounds(const IntRect& req, bool invertY = false) const {
+    IntRect bb = bounds().intersect(req) - req.origin();
+    if (invertY) bb.invert_y(req.height());
+    return bb;
+  }
+
+  // Get a pointer for sampling at the given offset
+  char* sample_ptr(int x, int y) const {
+    return buf + y * stride() + x * bpp();
+  }
+
+  // Get a pointer for sampling the requested region and limit to the provided
+  // sampling bounds
+  char* sample_ptr(const IntRect& req, const IntRect& bounds,
+                   bool invertY = false) const {
+    // Offset the sample pointer by the clamped bounds
+    int x = req.x0 + bounds.x0;
+    // Invert the Y offset if necessary
+    int y = invertY ? req.y1 - 1 - bounds.y0 : req.y0 + bounds.y0;
+    return sample_ptr(x, y);
+  }
+};
+
+// The last vertex attribute is reserved as a null attribute in case a vertex
+// attribute is used without being set.
+#define MAX_ATTRIBS 17
+#define NULL_ATTRIB 16
+struct VertexArray {
+  VertexAttrib attribs[MAX_ATTRIBS];
+  int max_attrib = -1;
+  // The GL spec defines element array buffer binding to be part of VAO state.
+  GLuint element_array_buffer_binding = 0;
+
+  void validate();
+};
+
+struct Shader {
+  GLenum type = 0;
+  ProgramLoader loader = nullptr;
+};
+
+struct Program {
+  ProgramImpl* impl = nullptr;
+  VertexShaderImpl* vert_impl = nullptr;
+  FragmentShaderImpl* frag_impl = nullptr;
+  bool deleted = false;
+
+  ~Program() { delete impl; }
+};
+
+// clang-format off
+// Fully-expand GL defines while ignoring more than 4 suffixes
+#define CONCAT_KEY(prefix, x, y, z, w, ...) prefix##x##y##z##w
+// Generate a blend key enum symbol
+#define BLEND_KEY(...) CONCAT_KEY(BLEND_, __VA_ARGS__, 0, 0, 0)
+#define MASK_BLEND_KEY(...) CONCAT_KEY(MASK_BLEND_, __VA_ARGS__, 0, 0, 0)
+#define AA_BLEND_KEY(...) CONCAT_KEY(AA_BLEND_, __VA_ARGS__, 0, 0, 0)
+#define AA_MASK_BLEND_KEY(...) CONCAT_KEY(AA_MASK_BLEND_, __VA_ARGS__, 0, 0, 0)
+
+// Utility macro to easily generate similar code for all implemented blend modes
+#define FOR_EACH_BLEND_KEY(macro)                                              \
+  macro(GL_ONE, GL_ZERO, 0, 0)                                                 \
+  macro(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA, GL_ONE, GL_ONE_MINUS_SRC_ALPHA)  \
+  macro(GL_ONE, GL_ONE_MINUS_SRC_ALPHA, 0, 0)                                  \
+  macro(GL_ZERO, GL_ONE_MINUS_SRC_COLOR, 0, 0)                                 \
+  macro(GL_ZERO, GL_ONE_MINUS_SRC_COLOR, GL_ZERO, GL_ONE)                      \
+  macro(GL_ZERO, GL_ONE_MINUS_SRC_ALPHA, 0, 0)                                 \
+  macro(GL_ZERO, GL_SRC_COLOR, 0, 0)                                           \
+  macro(GL_ONE, GL_ONE, 0, 0)                                                  \
+  macro(GL_ONE, GL_ONE, GL_ONE, GL_ONE_MINUS_SRC_ALPHA)                        \
+  macro(GL_ONE_MINUS_DST_ALPHA, GL_ONE, GL_ZERO, GL_ONE)                       \
+  macro(GL_CONSTANT_COLOR, GL_ONE_MINUS_SRC_COLOR, 0, 0)                       \
+  macro(GL_ONE, GL_ONE_MINUS_SRC1_COLOR, 0, 0)                                 \
+  macro(GL_MIN, 0, 0, 0)                                                       \
+  macro(GL_MAX, 0, 0, 0)                                                       \
+  macro(GL_MULTIPLY_KHR, 0, 0, 0)                                              \
+  macro(GL_SCREEN_KHR, 0, 0, 0)                                                \
+  macro(GL_OVERLAY_KHR, 0, 0, 0)                                               \
+  macro(GL_DARKEN_KHR, 0, 0, 0)                                                \
+  macro(GL_LIGHTEN_KHR, 0, 0, 0)                                               \
+  macro(GL_COLORDODGE_KHR, 0, 0, 0)                                            \
+  macro(GL_COLORBURN_KHR, 0, 0, 0)                                             \
+  macro(GL_HARDLIGHT_KHR, 0, 0, 0)                                             \
+  macro(GL_SOFTLIGHT_KHR, 0, 0, 0)                                             \
+  macro(GL_DIFFERENCE_KHR, 0, 0, 0)                                            \
+  macro(GL_EXCLUSION_KHR, 0, 0, 0)                                             \
+  macro(GL_HSL_HUE_KHR, 0, 0, 0)                                               \
+  macro(GL_HSL_SATURATION_KHR, 0, 0, 0)                                        \
+  macro(GL_HSL_COLOR_KHR, 0, 0, 0)                                             \
+  macro(GL_HSL_LUMINOSITY_KHR, 0, 0, 0)                                        \
+  macro(SWGL_BLEND_DROP_SHADOW, 0, 0, 0)                                       \
+  macro(SWGL_BLEND_SUBPIXEL_TEXT, 0, 0, 0)
+
+#define DEFINE_BLEND_KEY(...) BLEND_KEY(__VA_ARGS__),
+#define DEFINE_MASK_BLEND_KEY(...) MASK_BLEND_KEY(__VA_ARGS__),
+#define DEFINE_AA_BLEND_KEY(...) AA_BLEND_KEY(__VA_ARGS__),
+#define DEFINE_AA_MASK_BLEND_KEY(...) AA_MASK_BLEND_KEY(__VA_ARGS__),
+enum BlendKey : uint8_t {
+  FOR_EACH_BLEND_KEY(DEFINE_BLEND_KEY)
+  FOR_EACH_BLEND_KEY(DEFINE_MASK_BLEND_KEY)
+  FOR_EACH_BLEND_KEY(DEFINE_AA_BLEND_KEY)
+  FOR_EACH_BLEND_KEY(DEFINE_AA_MASK_BLEND_KEY)
+  BLEND_KEY_NONE = BLEND_KEY(GL_ONE, GL_ZERO),
+  MASK_BLEND_KEY_NONE = MASK_BLEND_KEY(GL_ONE, GL_ZERO),
+  AA_BLEND_KEY_NONE = AA_BLEND_KEY(GL_ONE, GL_ZERO),
+  AA_MASK_BLEND_KEY_NONE = AA_MASK_BLEND_KEY(GL_ONE, GL_ZERO),
+};
+// clang-format on
+
+const size_t MAX_TEXTURE_UNITS = 16;
+
+template <typename T>
+static inline bool unlink(T& binding, T n) {
+  if (binding == n) {
+    binding = 0;
+    return true;
+  }
+  return false;
+}
+
+template <typename O>
+struct ObjectStore {
+  O** objects = nullptr;
+  size_t size = 0;
+  // reserve object 0 as null
+  size_t first_free = 1;
+  O invalid;
+
+  ~ObjectStore() {
+    if (objects) {
+      for (size_t i = 0; i < size; i++) delete objects[i];
+      free(objects);
+    }
+  }
+
+  bool grow(size_t i) {
+    size_t new_size = size ? size : 8;
+    while (new_size <= i) new_size += new_size / 2;
+    O** new_objects = (O**)realloc(objects, new_size * sizeof(O*));
+    assert(new_objects);
+    if (!new_objects) return false;
+    while (size < new_size) new_objects[size++] = nullptr;
+    objects = new_objects;
+    return true;
+  }
+
+  void insert(size_t i, const O& o) {
+    if (i >= size && !grow(i)) return;
+    if (!objects[i]) objects[i] = new O(o);
+  }
+
+  size_t next_free() {
+    size_t i = first_free;
+    while (i < size && objects[i]) i++;
+    first_free = i;
+    return i;
+  }
+
+  size_t insert(const O& o = O()) {
+    size_t i = next_free();
+    insert(i, o);
+    return i;
+  }
+
+  O& operator[](size_t i) {
+    insert(i, O());
+    return i < size ? *objects[i] : invalid;
+  }
+
+  O* find(size_t i) const { return i < size ? objects[i] : nullptr; }
+
+  template <typename T>
+  void on_erase(T*, ...) {}
+  template <typename T>
+  void on_erase(T* o, decltype(&T::on_erase)) {
+    o->on_erase();
+  }
+
+  bool erase(size_t i) {
+    if (i < size && objects[i]) {
+      on_erase(objects[i], nullptr);
+      delete objects[i];
+      objects[i] = nullptr;
+      if (i < first_free) first_free = i;
+      return true;
+    }
+    return false;
+  }
+
+  O** begin() const { return objects; }
+  O** end() const { return &objects[size]; }
+};
+
+struct Context {
+  int32_t references = 1;
+
+  ObjectStore<Query> queries;
+  ObjectStore<Buffer> buffers;
+  ObjectStore<Texture> textures;
+  ObjectStore<VertexArray> vertex_arrays;
+  ObjectStore<Framebuffer> framebuffers;
+  ObjectStore<Renderbuffer> renderbuffers;
+  ObjectStore<Shader> shaders;
+  ObjectStore<Program> programs;
+
+  GLenum last_error = GL_NO_ERROR;
+
+  IntRect viewport = {0, 0, 0, 0};
+
+  bool blend = false;
+  GLenum blendfunc_srgb = GL_ONE;
+  GLenum blendfunc_drgb = GL_ZERO;
+  GLenum blendfunc_sa = GL_ONE;
+  GLenum blendfunc_da = GL_ZERO;
+  GLenum blend_equation = GL_FUNC_ADD;
+  V8<uint16_t> blendcolor = 0;
+  BlendKey blend_key = BLEND_KEY_NONE;
+
+  bool depthtest = false;
+  bool depthmask = true;
+  GLenum depthfunc = GL_LESS;
+
+  bool scissortest = false;
+  IntRect scissor = {0, 0, 0, 0};
+
+  GLfloat clearcolor[4] = {0, 0, 0, 0};
+  GLdouble cleardepth = 1;
+
+  int unpack_row_length = 0;
+
+  int shaded_rows = 0;
+  int shaded_pixels = 0;
+
+  struct TextureUnit {
+    GLuint texture_2d_binding = 0;
+    GLuint texture_rectangle_binding = 0;
+
+    void unlink(GLuint n) {
+      ::unlink(texture_2d_binding, n);
+      ::unlink(texture_rectangle_binding, n);
+    }
+  };
+  TextureUnit texture_units[MAX_TEXTURE_UNITS];
+  int active_texture_unit = 0;
+
+  GLuint current_program = 0;
+
+  GLuint current_vertex_array = 0;
+  bool validate_vertex_array = true;
+
+  GLuint pixel_pack_buffer_binding = 0;
+  GLuint pixel_unpack_buffer_binding = 0;
+  GLuint array_buffer_binding = 0;
+  GLuint time_elapsed_query = 0;
+  GLuint samples_passed_query = 0;
+  GLuint renderbuffer_binding = 0;
+  GLuint draw_framebuffer_binding = 0;
+  GLuint read_framebuffer_binding = 0;
+  GLuint unknown_binding = 0;
+
+  GLuint& get_binding(GLenum name) {
+    switch (name) {
+      case GL_PIXEL_PACK_BUFFER:
+        return pixel_pack_buffer_binding;
+      case GL_PIXEL_UNPACK_BUFFER:
+        return pixel_unpack_buffer_binding;
+      case GL_ARRAY_BUFFER:
+        return array_buffer_binding;
+      case GL_ELEMENT_ARRAY_BUFFER:
+        return vertex_arrays[current_vertex_array].element_array_buffer_binding;
+      case GL_TEXTURE_2D:
+        return texture_units[active_texture_unit].texture_2d_binding;
+      case GL_TEXTURE_RECTANGLE:
+        return texture_units[active_texture_unit].texture_rectangle_binding;
+      case GL_TIME_ELAPSED:
+        return time_elapsed_query;
+      case GL_SAMPLES_PASSED:
+        return samples_passed_query;
+      case GL_RENDERBUFFER:
+        return renderbuffer_binding;
+      case GL_DRAW_FRAMEBUFFER:
+        return draw_framebuffer_binding;
+      case GL_READ_FRAMEBUFFER:
+        return read_framebuffer_binding;
+      default:
+        debugf("unknown binding %x\n", name);
+        assert(false);
+        return unknown_binding;
+    }
+  }
+
+  Texture& get_texture(sampler2D, int unit) {
+    return textures[texture_units[unit].texture_2d_binding];
+  }
+
+  Texture& get_texture(isampler2D, int unit) {
+    return textures[texture_units[unit].texture_2d_binding];
+  }
+
+  Texture& get_texture(sampler2DRect, int unit) {
+    return textures[texture_units[unit].texture_rectangle_binding];
+  }
+
+  IntRect apply_scissor(IntRect bb,
+                        const IntPoint& origin = IntPoint(0, 0)) const {
+    return scissortest ? bb.intersect(scissor - origin) : bb;
+  }
+
+  IntRect apply_scissor(const Texture& t) const {
+    return apply_scissor(t.bounds(), t.offset);
+  }
+};
+static Context* ctx = nullptr;
+static VertexShaderImpl* vertex_shader = nullptr;
+static FragmentShaderImpl* fragment_shader = nullptr;
+static BlendKey blend_key = BLEND_KEY_NONE;
+
+static void prepare_texture(Texture& t, const IntRect* skip = nullptr);
+
+template <typename S>
+static inline void init_filter(S* s, Texture& t) {
+  // If the width is not at least 2 pixels, then we can't safely sample the end
+  // of the row with a linear filter. In that case, just punt to using nearest
+  // filtering instead.
+  s->filter = t.width >= 2 ? gl_filter_to_texture_filter(t.mag_filter)
+                           : TextureFilter::NEAREST;
+}
+
+template <typename S>
+static inline void init_sampler(S* s, Texture& t) {
+  prepare_texture(t);
+  s->width = t.width;
+  s->height = t.height;
+  s->stride = t.stride();
+  int bpp = t.bpp();
+  if (bpp >= 4)
+    s->stride /= 4;
+  else if (bpp == 2)
+    s->stride /= 2;
+  else
+    assert(bpp == 1);
+  // Use uint32_t* for easier sampling, but need to cast to uint8_t* or
+  // uint16_t* for formats with bpp < 4.
+  s->buf = (uint32_t*)t.buf;
+  s->format = gl_format_to_texture_format(t.internal_format);
+}
+
+template <typename S>
+static inline void null_sampler(S* s) {
+  // For null texture data, just make the sampler provide a 1x1 buffer that is
+  // transparent black. Ensure buffer holds at least a SIMD vector of zero data
+  // for SIMD padding of unaligned loads.
+  static const uint32_t zeroBuf[sizeof(Float) / sizeof(uint32_t)] = {0};
+  s->width = 1;
+  s->height = 1;
+  s->stride = s->width;
+  s->buf = (uint32_t*)zeroBuf;
+  s->format = TextureFormat::RGBA8;
+}
+
+template <typename S>
+static inline void null_filter(S* s) {
+  s->filter = TextureFilter::NEAREST;
+}
+
+template <typename S>
+S* lookup_sampler(S* s, int texture) {
+  Texture& t = ctx->get_texture(s, texture);
+  if (!t.buf) {
+    null_sampler(s);
+    null_filter(s);
+  } else {
+    init_sampler(s, t);
+    init_filter(s, t);
+  }
+  return s;
+}
+
+template <typename S>
+S* lookup_isampler(S* s, int texture) {
+  Texture& t = ctx->get_texture(s, texture);
+  if (!t.buf) {
+    null_sampler(s);
+  } else {
+    init_sampler(s, t);
+  }
+  return s;
+}
+
+int bytes_per_type(GLenum type) {
+  switch (type) {
+    case GL_INT:
+      return 4;
+    case GL_FLOAT:
+      return 4;
+    case GL_UNSIGNED_SHORT:
+      return 2;
+    case GL_UNSIGNED_BYTE:
+      return 1;
+    default:
+      assert(0);
+      return 0;
+  }
+}
+
+template <typename S, typename C>
+static inline S expand_attrib(const char* buf, size_t size, bool normalized) {
+  typedef typename ElementType<S>::ty elem_type;
+  S scalar = {0};
+  const C* src = reinterpret_cast<const C*>(buf);
+  if (normalized) {
+    const float scale = 1.0f / ((1 << (8 * sizeof(C))) - 1);
+    for (size_t i = 0; i < size / sizeof(C); i++) {
+      put_nth_component(scalar, i, elem_type(src[i]) * scale);
+    }
+  } else {
+    for (size_t i = 0; i < size / sizeof(C); i++) {
+      put_nth_component(scalar, i, elem_type(src[i]));
+    }
+  }
+  return scalar;
+}
+
+template <typename S>
+static inline S load_attrib_scalar(VertexAttrib& va, const char* src) {
+  if (sizeof(S) <= va.size) {
+    return *reinterpret_cast<const S*>(src);
+  }
+  if (va.type == GL_UNSIGNED_SHORT) {
+    return expand_attrib<S, uint16_t>(src, va.size, va.normalized);
+  }
+  if (va.type == GL_UNSIGNED_BYTE) {
+    return expand_attrib<S, uint8_t>(src, va.size, va.normalized);
+  }
+  assert(sizeof(typename ElementType<S>::ty) == bytes_per_type(va.type));
+  S scalar = {0};
+  memcpy(&scalar, src, va.size);
+  return scalar;
+}
+
+template <typename T>
+void load_attrib(T& attrib, VertexAttrib& va, uint32_t start, int instance,
+                 int count) {
+  typedef decltype(force_scalar(attrib)) scalar_type;
+  // If no buffer is available, just use a zero default.
+  if (!va.buf_size) {
+    attrib = T(scalar_type{0});
+  } else if (va.divisor != 0) {
+    char* src = (char*)va.buf + va.stride * instance + va.offset;
+    assert(src + va.size <= va.buf + va.buf_size);
+    attrib = T(load_attrib_scalar<scalar_type>(va, src));
+  } else {
+    // Specialized for WR's primitive vertex order/winding.
+    if (!count) return;
+    assert(count >= 2 && count <= 4);
+    char* src = (char*)va.buf + va.stride * start + va.offset;
+    switch (count) {
+      case 2: {
+        // Lines must be indexed at offsets 0, 1.
+        // Line vertexes fill vertex shader SIMD lanes as 0, 1, 1, 0.
+        scalar_type lanes[2] = {
+            load_attrib_scalar<scalar_type>(va, src),
+            load_attrib_scalar<scalar_type>(va, src + va.stride)};
+        attrib = (T){lanes[0], lanes[1], lanes[1], lanes[0]};
+        break;
+      }
+      case 3: {
+        // Triangles must be indexed at offsets 0, 1, 2.
+        // Triangle vertexes fill vertex shader SIMD lanes as 0, 1, 2, 2.
+        scalar_type lanes[3] = {
+            load_attrib_scalar<scalar_type>(va, src),
+            load_attrib_scalar<scalar_type>(va, src + va.stride),
+            load_attrib_scalar<scalar_type>(va, src + va.stride * 2)};
+        attrib = (T){lanes[0], lanes[1], lanes[2], lanes[2]};
+        break;
+      }
+      default:
+        // Quads must be successive triangles indexed at offsets 0, 1, 2, 2,
+        // 1, 3. Quad vertexes fill vertex shader SIMD lanes as 0, 1, 3, 2, so
+        // that the points form a convex path that can be traversed by the
+        // rasterizer.
+        attrib = (T){load_attrib_scalar<scalar_type>(va, src),
+                     load_attrib_scalar<scalar_type>(va, src + va.stride),
+                     load_attrib_scalar<scalar_type>(va, src + va.stride * 3),
+                     load_attrib_scalar<scalar_type>(va, src + va.stride * 2)};
+        break;
+    }
+  }
+}
+
+template <typename T>
+void load_flat_attrib(T& attrib, VertexAttrib& va, uint32_t start, int instance,
+                      int count) {
+  typedef decltype(force_scalar(attrib)) scalar_type;
+  // If no buffer is available, just use a zero default.
+  if (!va.buf_size) {
+    attrib = T{0};
+    return;
+  }
+  char* src = nullptr;
+  if (va.divisor != 0) {
+    src = (char*)va.buf + va.stride * instance + va.offset;
+  } else {
+    if (!count) return;
+    src = (char*)va.buf + va.stride * start + va.offset;
+  }
+  assert(src + va.size <= va.buf + va.buf_size);
+  attrib = T(load_attrib_scalar<scalar_type>(va, src));
+}
+
+void setup_program(GLuint program) {
+  if (!program) {
+    vertex_shader = nullptr;
+    fragment_shader = nullptr;
+    return;
+  }
+  Program& p = ctx->programs[program];
+  assert(p.impl);
+  assert(p.vert_impl);
+  assert(p.frag_impl);
+  vertex_shader = p.vert_impl;
+  fragment_shader = p.frag_impl;
+}
+
+extern ProgramLoader load_shader(const char* name);
+
+extern "C" {
+
+void UseProgram(GLuint program) {
+  if (ctx->current_program && program != ctx->current_program) {
+    auto* p = ctx->programs.find(ctx->current_program);
+    if (p && p->deleted) {
+      ctx->programs.erase(ctx->current_program);
+    }
+  }
+  ctx->current_program = program;
+  setup_program(program);
+}
+
+void SetViewport(GLint x, GLint y, GLsizei width, GLsizei height) {
+  ctx->viewport = IntRect{x, y, x + width, y + height};
+}
+
+void Enable(GLenum cap) {
+  switch (cap) {
+    case GL_BLEND:
+      ctx->blend = true;
+      break;
+    case GL_DEPTH_TEST:
+      ctx->depthtest = true;
+      break;
+    case GL_SCISSOR_TEST:
+      ctx->scissortest = true;
+      break;
+  }
+}
+
+void Disable(GLenum cap) {
+  switch (cap) {
+    case GL_BLEND:
+      ctx->blend = false;
+      break;
+    case GL_DEPTH_TEST:
+      ctx->depthtest = false;
+      break;
+    case GL_SCISSOR_TEST:
+      ctx->scissortest = false;
+      break;
+  }
+}
+
+// Report the last error generated and clear the error status.
+GLenum GetError() {
+  GLenum error = ctx->last_error;
+  ctx->last_error = GL_NO_ERROR;
+  return error;
+}
+
+// Sets the error status to out-of-memory to indicate that a buffer
+// or texture re-allocation failed.
+static void out_of_memory() { ctx->last_error = GL_OUT_OF_MEMORY; }
+
+static const char* const extensions[] = {
+    "GL_ARB_blend_func_extended",
+    "GL_ARB_clear_texture",
+    "GL_ARB_copy_image",
+    "GL_ARB_draw_instanced",
+    "GL_ARB_explicit_attrib_location",
+    "GL_ARB_instanced_arrays",
+    "GL_ARB_invalidate_subdata",
+    "GL_ARB_texture_storage",
+    "GL_EXT_timer_query",
+    "GL_KHR_blend_equation_advanced",
+    "GL_KHR_blend_equation_advanced_coherent",
+    "GL_APPLE_rgb_422",
+};
+
+void GetIntegerv(GLenum pname, GLint* params) {
+  assert(params);
+  switch (pname) {
+    case GL_MAX_TEXTURE_UNITS:
+    case GL_MAX_TEXTURE_IMAGE_UNITS:
+      params[0] = MAX_TEXTURE_UNITS;
+      break;
+    case GL_MAX_TEXTURE_SIZE:
+      params[0] = 1 << 15;
+      break;
+    case GL_MAX_ARRAY_TEXTURE_LAYERS:
+      params[0] = 0;
+      break;
+    case GL_READ_FRAMEBUFFER_BINDING:
+      params[0] = ctx->read_framebuffer_binding;
+      break;
+    case GL_DRAW_FRAMEBUFFER_BINDING:
+      params[0] = ctx->draw_framebuffer_binding;
+      break;
+    case GL_PIXEL_PACK_BUFFER_BINDING:
+      params[0] = ctx->pixel_pack_buffer_binding;
+      break;
+    case GL_PIXEL_UNPACK_BUFFER_BINDING:
+      params[0] = ctx->pixel_unpack_buffer_binding;
+      break;
+    case GL_NUM_EXTENSIONS:
+      params[0] = sizeof(extensions) / sizeof(extensions[0]);
+      break;
+    case GL_MAJOR_VERSION:
+      params[0] = 3;
+      break;
+    case GL_MINOR_VERSION:
+      params[0] = 2;
+      break;
+    case GL_MIN_PROGRAM_TEXEL_OFFSET:
+      params[0] = 0;
+      break;
+    case GL_MAX_PROGRAM_TEXEL_OFFSET:
+      params[0] = MAX_TEXEL_OFFSET;
+      break;
+    default:
+      debugf("unhandled glGetIntegerv parameter %x\n", pname);
+      assert(false);
+  }
+}
+
+void GetBooleanv(GLenum pname, GLboolean* params) {
+  assert(params);
+  switch (pname) {
+    case GL_DEPTH_WRITEMASK:
+      params[0] = ctx->depthmask;
+      break;
+    default:
+      debugf("unhandled glGetBooleanv parameter %x\n", pname);
+      assert(false);
+  }
+}
+
+const char* GetString(GLenum name) {
+  switch (name) {
+    case GL_VENDOR:
+      return "Mozilla Gfx";
+    case GL_RENDERER:
+      return "Software WebRender";
+    case GL_VERSION:
+      return "3.2";
+    case GL_SHADING_LANGUAGE_VERSION:
+      return "1.50";
+    default:
+      debugf("unhandled glGetString parameter %x\n", name);
+      assert(false);
+      return nullptr;
+  }
+}
+
+const char* GetStringi(GLenum name, GLuint index) {
+  switch (name) {
+    case GL_EXTENSIONS:
+      if (index >= sizeof(extensions) / sizeof(extensions[0])) {
+        return nullptr;
+      }
+      return extensions[index];
+    default:
+      debugf("unhandled glGetStringi parameter %x\n", name);
+      assert(false);
+      return nullptr;
+  }
+}
+
+GLenum remap_blendfunc(GLenum rgb, GLenum a) {
+  switch (a) {
+    case GL_SRC_ALPHA:
+      if (rgb == GL_SRC_COLOR) a = GL_SRC_COLOR;
+      break;
+    case GL_ONE_MINUS_SRC_ALPHA:
+      if (rgb == GL_ONE_MINUS_SRC_COLOR) a = GL_ONE_MINUS_SRC_COLOR;
+      break;
+    case GL_DST_ALPHA:
+      if (rgb == GL_DST_COLOR) a = GL_DST_COLOR;
+      break;
+    case GL_ONE_MINUS_DST_ALPHA:
+      if (rgb == GL_ONE_MINUS_DST_COLOR) a = GL_ONE_MINUS_DST_COLOR;
+      break;
+    case GL_CONSTANT_ALPHA:
+      if (rgb == GL_CONSTANT_COLOR) a = GL_CONSTANT_COLOR;
+      break;
+    case GL_ONE_MINUS_CONSTANT_ALPHA:
+      if (rgb == GL_ONE_MINUS_CONSTANT_COLOR) a = GL_ONE_MINUS_CONSTANT_COLOR;
+      break;
+    case GL_SRC_COLOR:
+      if (rgb == GL_SRC_ALPHA) a = GL_SRC_ALPHA;
+      break;
+    case GL_ONE_MINUS_SRC_COLOR:
+      if (rgb == GL_ONE_MINUS_SRC_ALPHA) a = GL_ONE_MINUS_SRC_ALPHA;
+      break;
+    case GL_DST_COLOR:
+      if (rgb == GL_DST_ALPHA) a = GL_DST_ALPHA;
+      break;
+    case GL_ONE_MINUS_DST_COLOR:
+      if (rgb == GL_ONE_MINUS_DST_ALPHA) a = GL_ONE_MINUS_DST_ALPHA;
+      break;
+    case GL_CONSTANT_COLOR:
+      if (rgb == GL_CONSTANT_ALPHA) a = GL_CONSTANT_ALPHA;
+      break;
+    case GL_ONE_MINUS_CONSTANT_COLOR:
+      if (rgb == GL_ONE_MINUS_CONSTANT_ALPHA) a = GL_ONE_MINUS_CONSTANT_ALPHA;
+      break;
+    case GL_SRC1_ALPHA:
+      if (rgb == GL_SRC1_COLOR) a = GL_SRC1_COLOR;
+      break;
+    case GL_ONE_MINUS_SRC1_ALPHA:
+      if (rgb == GL_ONE_MINUS_SRC1_COLOR) a = GL_ONE_MINUS_SRC1_COLOR;
+      break;
+    case GL_SRC1_COLOR:
+      if (rgb == GL_SRC1_ALPHA) a = GL_SRC1_ALPHA;
+      break;
+    case GL_ONE_MINUS_SRC1_COLOR:
+      if (rgb == GL_ONE_MINUS_SRC1_ALPHA) a = GL_ONE_MINUS_SRC1_ALPHA;
+      break;
+  }
+  return a;
+}
+
+// Generate a hashed blend key based on blend func and equation state. This
+// allows all the blend state to be processed down to a blend key that can be
+// dealt with inside a single switch statement.
+static void hash_blend_key() {
+  GLenum srgb = ctx->blendfunc_srgb;
+  GLenum drgb = ctx->blendfunc_drgb;
+  GLenum sa = ctx->blendfunc_sa;
+  GLenum da = ctx->blendfunc_da;
+  GLenum equation = ctx->blend_equation;
+#define HASH_BLEND_KEY(x, y, z, w) ((x << 4) | (y) | (z << 24) | (w << 20))
+  // Basic non-separate blend funcs used the two argument form
+  int hash = HASH_BLEND_KEY(srgb, drgb, 0, 0);
+  // Separate alpha blend funcs use the 4 argument hash
+  if (srgb != sa || drgb != da) hash |= HASH_BLEND_KEY(0, 0, sa, da);
+  // Any other blend equation than the default func_add ignores the func and
+  // instead generates a one-argument hash based on the equation
+  if (equation != GL_FUNC_ADD) hash = HASH_BLEND_KEY(equation, 0, 0, 0);
+  switch (hash) {
+#define MAP_BLEND_KEY(...)                   \
+  case HASH_BLEND_KEY(__VA_ARGS__):          \
+    ctx->blend_key = BLEND_KEY(__VA_ARGS__); \
+    break;
+    FOR_EACH_BLEND_KEY(MAP_BLEND_KEY)
+    default:
+      debugf("blendfunc: %x, %x, separate: %x, %x, equation: %x\n", srgb, drgb,
+             sa, da, equation);
+      assert(false);
+      break;
+  }
+}
+
+void BlendFunc(GLenum srgb, GLenum drgb, GLenum sa, GLenum da) {
+  ctx->blendfunc_srgb = srgb;
+  ctx->blendfunc_drgb = drgb;
+  sa = remap_blendfunc(srgb, sa);
+  da = remap_blendfunc(drgb, da);
+  ctx->blendfunc_sa = sa;
+  ctx->blendfunc_da = da;
+
+  hash_blend_key();
+}
+
+void BlendColor(GLfloat r, GLfloat g, GLfloat b, GLfloat a) {
+  I32 c = round_pixel((Float){b, g, r, a});
+  ctx->blendcolor = CONVERT(c, U16).xyzwxyzw;
+}
+
+void BlendEquation(GLenum mode) {
+  assert(mode == GL_FUNC_ADD || mode == GL_MIN || mode == GL_MAX ||
+         (mode >= GL_MULTIPLY_KHR && mode <= GL_HSL_LUMINOSITY_KHR));
+  if (mode != ctx->blend_equation) {
+    ctx->blend_equation = mode;
+    hash_blend_key();
+  }
+}
+
+void DepthMask(GLboolean flag) { ctx->depthmask = flag; }
+
+void DepthFunc(GLenum func) {
+  switch (func) {
+    case GL_LESS:
+    case GL_LEQUAL:
+      break;
+    default:
+      assert(false);
+  }
+  ctx->depthfunc = func;
+}
+
+void SetScissor(GLint x, GLint y, GLsizei width, GLsizei height) {
+  ctx->scissor = IntRect{x, y, x + width, y + height};
+}
+
+void ClearColor(GLfloat r, GLfloat g, GLfloat b, GLfloat a) {
+  ctx->clearcolor[0] = r;
+  ctx->clearcolor[1] = g;
+  ctx->clearcolor[2] = b;
+  ctx->clearcolor[3] = a;
+}
+
+void ClearDepth(GLdouble depth) { ctx->cleardepth = depth; }
+
+void ActiveTexture(GLenum texture) {
+  assert(texture >= GL_TEXTURE0);
+  assert(texture < GL_TEXTURE0 + MAX_TEXTURE_UNITS);
+  ctx->active_texture_unit =
+      clamp(int(texture - GL_TEXTURE0), 0, int(MAX_TEXTURE_UNITS - 1));
+}
+
+void GenQueries(GLsizei n, GLuint* result) {
+  for (int i = 0; i < n; i++) {
+    Query q;
+    result[i] = ctx->queries.insert(q);
+  }
+}
+
+void DeleteQuery(GLuint n) {
+  if (n && ctx->queries.erase(n)) {
+    unlink(ctx->time_elapsed_query, n);
+    unlink(ctx->samples_passed_query, n);
+  }
+}
+
+void GenBuffers(int n, GLuint* result) {
+  for (int i = 0; i < n; i++) {
+    Buffer b;
+    result[i] = ctx->buffers.insert(b);
+  }
+}
+
+void DeleteBuffer(GLuint n) {
+  if (n && ctx->buffers.erase(n)) {
+    unlink(ctx->pixel_pack_buffer_binding, n);
+    unlink(ctx->pixel_unpack_buffer_binding, n);
+    unlink(ctx->array_buffer_binding, n);
+  }
+}
+
+void GenVertexArrays(int n, GLuint* result) {
+  for (int i = 0; i < n; i++) {
+    VertexArray v;
+    result[i] = ctx->vertex_arrays.insert(v);
+  }
+}
+
+void DeleteVertexArray(GLuint n) {
+  if (n && ctx->vertex_arrays.erase(n)) {
+    unlink(ctx->current_vertex_array, n);
+  }
+}
+
+GLuint CreateShader(GLenum type) {
+  Shader s;
+  s.type = type;
+  return ctx->shaders.insert(s);
+}
+
+void ShaderSourceByName(GLuint shader, char* name) {
+  Shader& s = ctx->shaders[shader];
+  s.loader = load_shader(name);
+  if (!s.loader) {
+    debugf("unknown shader %s\n", name);
+  }
+}
+
+void AttachShader(GLuint program, GLuint shader) {
+  Program& p = ctx->programs[program];
+  Shader& s = ctx->shaders[shader];
+  if (s.type == GL_VERTEX_SHADER) {
+    if (!p.impl && s.loader) p.impl = s.loader();
+  } else if (s.type == GL_FRAGMENT_SHADER) {
+    if (!p.impl && s.loader) p.impl = s.loader();
+  } else {
+    assert(0);
+  }
+}
+
+void DeleteShader(GLuint n) {
+  if (n) ctx->shaders.erase(n);
+}
+
+GLuint CreateProgram() {
+  Program p;
+  return ctx->programs.insert(p);
+}
+
+void DeleteProgram(GLuint n) {
+  if (!n) return;
+  if (ctx->current_program == n) {
+    if (auto* p = ctx->programs.find(n)) {
+      p->deleted = true;
+    }
+  } else {
+    ctx->programs.erase(n);
+  }
+}
+
+void LinkProgram(GLuint program) {
+  Program& p = ctx->programs[program];
+  assert(p.impl);
+  if (!p.impl) {
+    return;
+  }
+  assert(p.impl->interpolants_size() <= sizeof(Interpolants));
+  if (!p.vert_impl) p.vert_impl = p.impl->get_vertex_shader();
+  if (!p.frag_impl) p.frag_impl = p.impl->get_fragment_shader();
+}
+
+GLint GetLinkStatus(GLuint program) {
+  if (auto* p = ctx->programs.find(program)) {
+    return p->impl ? 1 : 0;
+  }
+  return 0;
+}
+
+void BindAttribLocation(GLuint program, GLuint index, char* name) {
+  Program& p = ctx->programs[program];
+  assert(p.impl);
+  if (!p.impl) {
+    return;
+  }
+  p.impl->bind_attrib(name, index);
+}
+
+GLint GetAttribLocation(GLuint program, char* name) {
+  Program& p = ctx->programs[program];
+  assert(p.impl);
+  if (!p.impl) {
+    return -1;
+  }
+  return p.impl->get_attrib(name);
+}
+
+GLint GetUniformLocation(GLuint program, char* name) {
+  Program& p = ctx->programs[program];
+  assert(p.impl);
+  if (!p.impl) {
+    return -1;
+  }
+  GLint loc = p.impl->get_uniform(name);
+  // debugf("location: %d\n", loc);
+  return loc;
+}
+
+static uint64_t get_time_value() {
+#ifdef __MACH__
+  return mach_absolute_time();
+#elif defined(_WIN32)
+  LARGE_INTEGER time;
+  static bool have_frequency = false;
+  static LARGE_INTEGER frequency;
+  if (!have_frequency) {
+    QueryPerformanceFrequency(&frequency);
+    have_frequency = true;
+  }
+  QueryPerformanceCounter(&time);
+  return time.QuadPart * 1000000000ULL / frequency.QuadPart;
+#else
+  return ({
+    struct timespec tp;
+    clock_gettime(CLOCK_MONOTONIC, &tp);
+    tp.tv_sec * 1000000000ULL + tp.tv_nsec;
+  });
+#endif
+}
+
+void BeginQuery(GLenum target, GLuint id) {
+  ctx->get_binding(target) = id;
+  Query& q = ctx->queries[id];
+  switch (target) {
+    case GL_SAMPLES_PASSED:
+      q.value = 0;
+      break;
+    case GL_TIME_ELAPSED:
+      q.value = get_time_value();
+      break;
+    default:
+      debugf("unknown query target %x for query %d\n", target, id);
+      assert(false);
+  }
+}
+
+void EndQuery(GLenum target) {
+  Query& q = ctx->queries[ctx->get_binding(target)];
+  switch (target) {
+    case GL_SAMPLES_PASSED:
+      break;
+    case GL_TIME_ELAPSED:
+      q.value = get_time_value() - q.value;
+      break;
+    default:
+      debugf("unknown query target %x\n", target);
+      assert(false);
+  }
+  ctx->get_binding(target) = 0;
+}
+
+void GetQueryObjectui64v(GLuint id, GLenum pname, GLuint64* params) {
+  Query& q = ctx->queries[id];
+  switch (pname) {
+    case GL_QUERY_RESULT:
+      assert(params);
+      params[0] = q.value;
+      break;
+    default:
+      assert(false);
+  }
+}
+
+void BindVertexArray(GLuint vertex_array) {
+  if (vertex_array != ctx->current_vertex_array) {
+    ctx->validate_vertex_array = true;
+  }
+  ctx->current_vertex_array = vertex_array;
+}
+
+void BindTexture(GLenum target, GLuint texture) {
+  ctx->get_binding(target) = texture;
+}
+
+void BindBuffer(GLenum target, GLuint buffer) {
+  ctx->get_binding(target) = buffer;
+}
+
+void BindFramebuffer(GLenum target, GLuint fb) {
+  if (target == GL_FRAMEBUFFER) {
+    ctx->read_framebuffer_binding = fb;
+    ctx->draw_framebuffer_binding = fb;
+  } else {
+    assert(target == GL_READ_FRAMEBUFFER || target == GL_DRAW_FRAMEBUFFER);
+    ctx->get_binding(target) = fb;
+  }
+}
+
+void BindRenderbuffer(GLenum target, GLuint rb) {
+  ctx->get_binding(target) = rb;
+}
+
+void PixelStorei(GLenum name, GLint param) {
+  if (name == GL_UNPACK_ALIGNMENT) {
+    assert(param == 1);
+  } else if (name == GL_UNPACK_ROW_LENGTH) {
+    ctx->unpack_row_length = param;
+  }
+}
+
+static GLenum remap_internal_format(GLenum format) {
+  switch (format) {
+    case GL_DEPTH_COMPONENT:
+      return GL_DEPTH_COMPONENT24;
+    case GL_RGBA:
+      return GL_RGBA8;
+    case GL_RED:
+      return GL_R8;
+    case GL_RG:
+      return GL_RG8;
+    case GL_RGB_422_APPLE:
+      return GL_RGB_RAW_422_APPLE;
+    default:
+      return format;
+  }
+}
+
+}  // extern "C"
+
+static bool format_requires_conversion(GLenum external_format,
+                                       GLenum internal_format) {
+  switch (external_format) {
+    case GL_RGBA:
+      return internal_format == GL_RGBA8;
+    default:
+      return false;
+  }
+}
+
+static inline void copy_bgra8_to_rgba8(uint32_t* dest, const uint32_t* src,
+                                       int width) {
+  for (; width >= 4; width -= 4, dest += 4, src += 4) {
+    U32 p = unaligned_load<U32>(src);
+    U32 rb = p & 0x00FF00FF;
+    unaligned_store(dest, (p & 0xFF00FF00) | (rb << 16) | (rb >> 16));
+  }
+  for (; width > 0; width--, dest++, src++) {
+    uint32_t p = *src;
+    uint32_t rb = p & 0x00FF00FF;
+    *dest = (p & 0xFF00FF00) | (rb << 16) | (rb >> 16);
+  }
+}
+
+static void convert_copy(GLenum external_format, GLenum internal_format,
+                         uint8_t* dst_buf, size_t dst_stride,
+                         const uint8_t* src_buf, size_t src_stride,
+                         size_t width, size_t height) {
+  switch (external_format) {
+    case GL_RGBA:
+      if (internal_format == GL_RGBA8) {
+        for (; height; height--) {
+          copy_bgra8_to_rgba8((uint32_t*)dst_buf, (const uint32_t*)src_buf,
+                              width);
+          dst_buf += dst_stride;
+          src_buf += src_stride;
+        }
+        return;
+      }
+      break;
+    default:
+      break;
+  }
+  size_t row_bytes = width * bytes_for_internal_format(internal_format);
+  for (; height; height--) {
+    memcpy(dst_buf, src_buf, row_bytes);
+    dst_buf += dst_stride;
+    src_buf += src_stride;
+  }
+}
+
+static void set_tex_storage(Texture& t, GLenum external_format, GLsizei width,
+                            GLsizei height, void* buf = nullptr,
+                            GLsizei stride = 0, GLsizei min_width = 0,
+                            GLsizei min_height = 0) {
+  GLenum internal_format = remap_internal_format(external_format);
+  bool changed = false;
+  if (t.width != width || t.height != height ||
+      t.internal_format != internal_format) {
+    changed = true;
+    t.internal_format = internal_format;
+    t.width = width;
+    t.height = height;
+  }
+  // If we are changed from an internally managed buffer to an externally
+  // supplied one or vice versa, ensure that we clean up old buffer state.
+  // However, if we have to convert the data from a non-native format, then
+  // always treat it as internally managed since we will need to copy to an
+  // internally managed native format buffer.
+  bool should_free = buf == nullptr || format_requires_conversion(
+                                           external_format, internal_format);
+  if (t.should_free() != should_free) {
+    changed = true;
+    t.cleanup();
+    t.set_should_free(should_free);
+  }
+  // If now an external buffer, explicitly set it...
+  if (!should_free) {
+    t.set_buffer(buf, stride);
+  }
+  t.disable_delayed_clear();
+  if (!t.allocate(changed, min_width, min_height)) {
+    out_of_memory();
+  }
+  // If we have a buffer that needs format conversion, then do that now.
+  if (buf && should_free) {
+    convert_copy(external_format, internal_format, (uint8_t*)t.buf, t.stride(),
+                 (const uint8_t*)buf, stride, width, height);
+  }
+}
+
+extern "C" {
+
+void TexStorage2D(GLenum target, GLint levels, GLenum internal_format,
+                  GLsizei width, GLsizei height) {
+  assert(levels == 1);
+  Texture& t = ctx->textures[ctx->get_binding(target)];
+  set_tex_storage(t, internal_format, width, height);
+}
+
+GLenum internal_format_for_data(GLenum format, GLenum ty) {
+  if (format == GL_RED && ty == GL_UNSIGNED_BYTE) {
+    return GL_R8;
+  } else if ((format == GL_RGBA || format == GL_BGRA) &&
+             (ty == GL_UNSIGNED_BYTE || ty == GL_UNSIGNED_INT_8_8_8_8_REV)) {
+    return GL_RGBA8;
+  } else if (format == GL_RGBA && ty == GL_FLOAT) {
+    return GL_RGBA32F;
+  } else if (format == GL_RGBA_INTEGER && ty == GL_INT) {
+    return GL_RGBA32I;
+  } else if (format == GL_RG && ty == GL_UNSIGNED_BYTE) {
+    return GL_RG8;
+  } else if (format == GL_RGB_422_APPLE &&
+             ty == GL_UNSIGNED_SHORT_8_8_REV_APPLE) {
+    return GL_RGB_RAW_422_APPLE;
+  } else if (format == GL_RED && ty == GL_UNSIGNED_SHORT) {
+    return GL_R16;
+  } else if (format == GL_RG && ty == GL_UNSIGNED_SHORT) {
+    return GL_RG16;
+  } else {
+    debugf("unknown internal format for format %x, type %x\n", format, ty);
+    assert(false);
+    return 0;
+  }
+}
+
+static Buffer* get_pixel_pack_buffer() {
+  return ctx->pixel_pack_buffer_binding
+             ? &ctx->buffers[ctx->pixel_pack_buffer_binding]
+             : nullptr;
+}
+
+static void* get_pixel_pack_buffer_data(void* data) {
+  if (Buffer* b = get_pixel_pack_buffer()) {
+    return b->buf ? b->buf + (size_t)data : nullptr;
+  }
+  return data;
+}
+
+static Buffer* get_pixel_unpack_buffer() {
+  return ctx->pixel_unpack_buffer_binding
+             ? &ctx->buffers[ctx->pixel_unpack_buffer_binding]
+             : nullptr;
+}
+
+static void* get_pixel_unpack_buffer_data(void* data) {
+  if (Buffer* b = get_pixel_unpack_buffer()) {
+    return b->buf ? b->buf + (size_t)data : nullptr;
+  }
+  return data;
+}
+
+void TexSubImage2D(GLenum target, GLint level, GLint xoffset, GLint yoffset,
+                   GLsizei width, GLsizei height, GLenum format, GLenum ty,
+                   void* data) {
+  if (level != 0) {
+    assert(false);
+    return;
+  }
+  data = get_pixel_unpack_buffer_data(data);
+  if (!data) return;
+  Texture& t = ctx->textures[ctx->get_binding(target)];
+  IntRect skip = {xoffset, yoffset, xoffset + width, yoffset + height};
+  prepare_texture(t, &skip);
+  assert(xoffset + width <= t.width);
+  assert(yoffset + height <= t.height);
+  assert(ctx->unpack_row_length == 0 || ctx->unpack_row_length >= width);
+  GLsizei row_length =
+      ctx->unpack_row_length != 0 ? ctx->unpack_row_length : width;
+  assert(t.internal_format == internal_format_for_data(format, ty));
+  int src_bpp = format_requires_conversion(format, t.internal_format)
+                    ? bytes_for_internal_format(format)
+                    : t.bpp();
+  if (!src_bpp || !t.buf) return;
+  convert_copy(format, t.internal_format,
+               (uint8_t*)t.sample_ptr(xoffset, yoffset), t.stride(),
+               (const uint8_t*)data, row_length * src_bpp, width, height);
+}
+
+void TexImage2D(GLenum target, GLint level, GLint internal_format,
+                GLsizei width, GLsizei height, GLint border, GLenum format,
+                GLenum ty, void* data) {
+  if (level != 0) {
+    assert(false);
+    return;
+  }
+  assert(border == 0);
+  TexStorage2D(target, 1, internal_format, width, height);
+  TexSubImage2D(target, 0, 0, 0, width, height, format, ty, data);
+}
+
+void GenerateMipmap(UNUSED GLenum target) {
+  // TODO: support mipmaps
+}
+
+void SetTextureParameter(GLuint texid, GLenum pname, GLint param) {
+  Texture& t = ctx->textures[texid];
+  switch (pname) {
+    case GL_TEXTURE_WRAP_S:
+      assert(param == GL_CLAMP_TO_EDGE);
+      break;
+    case GL_TEXTURE_WRAP_T:
+      assert(param == GL_CLAMP_TO_EDGE);
+      break;
+    case GL_TEXTURE_MIN_FILTER:
+      t.min_filter = param;
+      break;
+    case GL_TEXTURE_MAG_FILTER:
+      t.mag_filter = param;
+      break;
+    default:
+      break;
+  }
+}
+
+void TexParameteri(GLenum target, GLenum pname, GLint param) {
+  SetTextureParameter(ctx->get_binding(target), pname, param);
+}
+
+void GenTextures(int n, GLuint* result) {
+  for (int i = 0; i < n; i++) {
+    Texture t;
+    result[i] = ctx->textures.insert(t);
+  }
+}
+
+void DeleteTexture(GLuint n) {
+  if (n && ctx->textures.erase(n)) {
+    for (size_t i = 0; i < MAX_TEXTURE_UNITS; i++) {
+      ctx->texture_units[i].unlink(n);
+    }
+  }
+}
+
+void GenRenderbuffers(int n, GLuint* result) {
+  for (int i = 0; i < n; i++) {
+    Renderbuffer r;
+    result[i] = ctx->renderbuffers.insert(r);
+  }
+}
+
+void Renderbuffer::on_erase() {
+  for (auto* fb : ctx->framebuffers) {
+    if (fb) {
+      unlink(fb->color_attachment, texture);
+      unlink(fb->depth_attachment, texture);
+    }
+  }
+  DeleteTexture(texture);
+}
+
+void DeleteRenderbuffer(GLuint n) {
+  if (n && ctx->renderbuffers.erase(n)) {
+    unlink(ctx->renderbuffer_binding, n);
+  }
+}
+
+void GenFramebuffers(int n, GLuint* result) {
+  for (int i = 0; i < n; i++) {
+    Framebuffer f;
+    result[i] = ctx->framebuffers.insert(f);
+  }
+}
+
+void DeleteFramebuffer(GLuint n) {
+  if (n && ctx->framebuffers.erase(n)) {
+    unlink(ctx->read_framebuffer_binding, n);
+    unlink(ctx->draw_framebuffer_binding, n);
+  }
+}
+
+void RenderbufferStorage(GLenum target, GLenum internal_format, GLsizei width,
+                         GLsizei height) {
+  // Just refer a renderbuffer to a texture to simplify things for now...
+  Renderbuffer& r = ctx->renderbuffers[ctx->get_binding(target)];
+  if (!r.texture) {
+    GenTextures(1, &r.texture);
+  }
+  switch (internal_format) {
+    case GL_DEPTH_COMPONENT:
+    case GL_DEPTH_COMPONENT16:
+    case GL_DEPTH_COMPONENT24:
+    case GL_DEPTH_COMPONENT32:
+      // Force depth format to 24 bits...
+      internal_format = GL_DEPTH_COMPONENT24;
+      break;
+  }
+  set_tex_storage(ctx->textures[r.texture], internal_format, width, height);
+}
+
+void VertexAttribPointer(GLuint index, GLint size, GLenum type, bool normalized,
+                         GLsizei stride, GLuint offset) {
+  // debugf("cva: %d\n", ctx->current_vertex_array);
+  VertexArray& v = ctx->vertex_arrays[ctx->current_vertex_array];
+  if (index >= NULL_ATTRIB) {
+    assert(0);
+    return;
+  }
+  VertexAttrib& va = v.attribs[index];
+  va.size = size * bytes_per_type(type);
+  va.type = type;
+  va.normalized = normalized;
+  va.stride = stride;
+  va.offset = offset;
+  // Buffer &vertex_buf = ctx->buffers[ctx->array_buffer_binding];
+  va.vertex_buffer = ctx->array_buffer_binding;
+  va.vertex_array = ctx->current_vertex_array;
+  ctx->validate_vertex_array = true;
+}
+
+void VertexAttribIPointer(GLuint index, GLint size, GLenum type, GLsizei stride,
+                          GLuint offset) {
+  // debugf("cva: %d\n", ctx->current_vertex_array);
+  VertexArray& v = ctx->vertex_arrays[ctx->current_vertex_array];
+  if (index >= NULL_ATTRIB) {
+    assert(0);
+    return;
+  }
+  VertexAttrib& va = v.attribs[index];
+  va.size = size * bytes_per_type(type);
+  va.type = type;
+  va.normalized = false;
+  va.stride = stride;
+  va.offset = offset;
+  // Buffer &vertex_buf = ctx->buffers[ctx->array_buffer_binding];
+  va.vertex_buffer = ctx->array_buffer_binding;
+  va.vertex_array = ctx->current_vertex_array;
+  ctx->validate_vertex_array = true;
+}
+
+void EnableVertexAttribArray(GLuint index) {
+  VertexArray& v = ctx->vertex_arrays[ctx->current_vertex_array];
+  if (index >= NULL_ATTRIB) {
+    assert(0);
+    return;
+  }
+  VertexAttrib& va = v.attribs[index];
+  if (!va.enabled) {
+    ctx->validate_vertex_array = true;
+  }
+  va.enabled = true;
+  v.max_attrib = max(v.max_attrib, (int)index);
+}
+
+void DisableVertexAttribArray(GLuint index) {
+  VertexArray& v = ctx->vertex_arrays[ctx->current_vertex_array];
+  if (index >= NULL_ATTRIB) {
+    assert(0);
+    return;
+  }
+  VertexAttrib& va = v.attribs[index];
+  if (va.enabled) {
+    ctx->validate_vertex_array = true;
+  }
+  va.disable();
+}
+
+void VertexAttribDivisor(GLuint index, GLuint divisor) {
+  VertexArray& v = ctx->vertex_arrays[ctx->current_vertex_array];
+  // Only support divisor being 0 (per-vertex) or 1 (per-instance).
+  if (index >= NULL_ATTRIB || divisor > 1) {
+    assert(0);
+    return;
+  }
+  VertexAttrib& va = v.attribs[index];
+  va.divisor = divisor;
+}
+
+void BufferData(GLenum target, GLsizeiptr size, void* data,
+                UNUSED GLenum usage) {
+  Buffer& b = ctx->buffers[ctx->get_binding(target)];
+  if (size != b.size) {
+    if (!b.allocate(size)) {
+      out_of_memory();
+    }
+    ctx->validate_vertex_array = true;
+  }
+  if (data && b.buf && size <= b.size) {
+    memcpy(b.buf, data, size);
+  }
+}
+
+void BufferSubData(GLenum target, GLintptr offset, GLsizeiptr size,
+                   void* data) {
+  Buffer& b = ctx->buffers[ctx->get_binding(target)];
+  assert(offset + size <= b.size);
+  if (data && b.buf && offset + size <= b.size) {
+    memcpy(&b.buf[offset], data, size);
+  }
+}
+
+void* MapBuffer(GLenum target, UNUSED GLbitfield access) {
+  Buffer& b = ctx->buffers[ctx->get_binding(target)];
+  return b.buf;
+}
+
+void* MapBufferRange(GLenum target, GLintptr offset, GLsizeiptr length,
+                     UNUSED GLbitfield access) {
+  Buffer& b = ctx->buffers[ctx->get_binding(target)];
+  if (b.buf && offset >= 0 && length > 0 && offset + length <= b.size) {
+    return b.buf + offset;
+  }
+  return nullptr;
+}
+
+GLboolean UnmapBuffer(GLenum target) {
+  Buffer& b = ctx->buffers[ctx->get_binding(target)];
+  return b.buf != nullptr;
+}
+
+void Uniform1i(GLint location, GLint V0) {
+  // debugf("tex: %d\n", (int)ctx->textures.size);
+  if (vertex_shader) {
+    vertex_shader->set_uniform_1i(location, V0);
+  }
+}
+void Uniform4fv(GLint location, GLsizei count, const GLfloat* v) {
+  assert(count == 1);
+  if (vertex_shader) {
+    vertex_shader->set_uniform_4fv(location, v);
+  }
+}
+void UniformMatrix4fv(GLint location, GLsizei count, GLboolean transpose,
+                      const GLfloat* value) {
+  assert(count == 1);
+  assert(!transpose);
+  if (vertex_shader) {
+    vertex_shader->set_uniform_matrix4fv(location, value);
+  }
+}
+
+void FramebufferTexture2D(GLenum target, GLenum attachment, GLenum textarget,
+                          GLuint texture, GLint level) {
+  assert(target == GL_READ_FRAMEBUFFER || target == GL_DRAW_FRAMEBUFFER);
+  assert(textarget == GL_TEXTURE_2D || textarget == GL_TEXTURE_RECTANGLE);
+  assert(level == 0);
+  Framebuffer& fb = ctx->framebuffers[ctx->get_binding(target)];
+  if (attachment == GL_COLOR_ATTACHMENT0) {
+    fb.color_attachment = texture;
+  } else if (attachment == GL_DEPTH_ATTACHMENT) {
+    fb.depth_attachment = texture;
+  } else {
+    assert(0);
+  }
+}
+
+void FramebufferRenderbuffer(GLenum target, GLenum attachment,
+                             GLenum renderbuffertarget, GLuint renderbuffer) {
+  assert(target == GL_READ_FRAMEBUFFER || target == GL_DRAW_FRAMEBUFFER);
+  assert(renderbuffertarget == GL_RENDERBUFFER);
+  Framebuffer& fb = ctx->framebuffers[ctx->get_binding(target)];
+  Renderbuffer& rb = ctx->renderbuffers[renderbuffer];
+  if (attachment == GL_COLOR_ATTACHMENT0) {
+    fb.color_attachment = rb.texture;
+  } else if (attachment == GL_DEPTH_ATTACHMENT) {
+    fb.depth_attachment = rb.texture;
+  } else {
+    assert(0);
+  }
+}
+
+}  // extern "C"
+
+static inline Framebuffer* get_framebuffer(GLenum target,
+                                           bool fallback = false) {
+  if (target == GL_FRAMEBUFFER) {
+    target = GL_DRAW_FRAMEBUFFER;
+  }
+  Framebuffer* fb = ctx->framebuffers.find(ctx->get_binding(target));
+  if (fallback && !fb) {
+    // If the specified framebuffer isn't found and a fallback is requested,
+    // use the default framebuffer.
+    fb = &ctx->framebuffers[0];
+  }
+  return fb;
+}
+
+template <typename T>
+static inline void fill_n(T* dst, size_t n, T val) {
+  for (T* end = &dst[n]; dst < end; dst++) *dst = val;
+}
+
+#if USE_SSE2
+template <>
+inline void fill_n<uint32_t>(uint32_t* dst, size_t n, uint32_t val) {
+  __asm__ __volatile__("rep stosl\n"
+                       : "+D"(dst), "+c"(n)
+                       : "a"(val)
+                       : "memory", "cc");
+}
+#endif
+
+static inline uint32_t clear_chunk(uint8_t value) {
+  return uint32_t(value) * 0x01010101U;
+}
+
+static inline uint32_t clear_chunk(uint16_t value) {
+  return uint32_t(value) | (uint32_t(value) << 16);
+}
+
+static inline uint32_t clear_chunk(uint32_t value) { return value; }
+
+template <typename T>
+static inline void clear_row(T* buf, size_t len, T value, uint32_t chunk) {
+  const size_t N = sizeof(uint32_t) / sizeof(T);
+  // fill any leading unaligned values
+  if (N > 1) {
+    size_t align = (-(intptr_t)buf & (sizeof(uint32_t) - 1)) / sizeof(T);
+    if (align <= len) {
+      fill_n(buf, align, value);
+      len -= align;
+      buf += align;
+    }
+  }
+  // fill as many aligned chunks as possible
+  fill_n((uint32_t*)buf, len / N, chunk);
+  // fill any remaining values
+  if (N > 1) {
+    fill_n(buf + (len & ~(N - 1)), len & (N - 1), value);
+  }
+}
+
+template <typename T>
+static void clear_buffer(Texture& t, T value, IntRect bb, int skip_start = 0,
+                         int skip_end = 0) {
+  if (!t.buf) return;
+  skip_start = max(skip_start, bb.x0);
+  skip_end = max(skip_end, skip_start);
+  assert(sizeof(T) == t.bpp());
+  size_t stride = t.stride();
+  // When clearing multiple full-width rows, collapse them into a single large
+  // "row" to avoid redundant setup from clearing each row individually. Note
+  // that we can only safely do this if the stride is tightly packed.
+  if (bb.width() == t.width && bb.height() > 1 && skip_start >= skip_end &&
+      (t.should_free() || stride == t.width * sizeof(T))) {
+    bb.x1 += (stride / sizeof(T)) * (bb.height() - 1);
+    bb.y1 = bb.y0 + 1;
+  }
+  T* buf = (T*)t.sample_ptr(bb.x0, bb.y0);
+  uint32_t chunk = clear_chunk(value);
+  for (int rows = bb.height(); rows > 0; rows--) {
+    if (bb.x0 < skip_start) {
+      clear_row(buf, skip_start - bb.x0, value, chunk);
+    }
+    if (skip_end < bb.x1) {
+      clear_row(buf + (skip_end - bb.x0), bb.x1 - skip_end, value, chunk);
+    }
+    buf += stride / sizeof(T);
+  }
+}
+
+template <typename T>
+static inline void force_clear_row(Texture& t, int y, int skip_start = 0,
+                                   int skip_end = 0) {
+  assert(t.buf != nullptr);
+  assert(sizeof(T) == t.bpp());
+  assert(skip_start <= skip_end);
+  T* buf = (T*)t.sample_ptr(0, y);
+  uint32_t chunk = clear_chunk((T)t.clear_val);
+  if (skip_start > 0) {
+    clear_row<T>(buf, skip_start, t.clear_val, chunk);
+  }
+  if (skip_end < t.width) {
+    clear_row<T>(buf + skip_end, t.width - skip_end, t.clear_val, chunk);
+  }
+}
+
+template <typename T>
+static void force_clear(Texture& t, const IntRect* skip = nullptr) {
+  if (!t.delay_clear || !t.cleared_rows) {
+    return;
+  }
+  int y0 = 0;
+  int y1 = t.height;
+  int skip_start = 0;
+  int skip_end = 0;
+  if (skip) {
+    y0 = clamp(skip->y0, 0, t.height);
+    y1 = clamp(skip->y1, y0, t.height);
+    skip_start = clamp(skip->x0, 0, t.width);
+    skip_end = clamp(skip->x1, skip_start, t.width);
+    if (skip_start <= 0 && skip_end >= t.width && y0 <= 0 && y1 >= t.height) {
+      t.disable_delayed_clear();
+      return;
+    }
+  }
+  int num_masks = (y1 + 31) / 32;
+  uint32_t* rows = t.cleared_rows;
+  for (int i = y0 / 32; i < num_masks; i++) {
+    uint32_t mask = rows[i];
+    if (mask != ~0U) {
+      rows[i] = ~0U;
+      int start = i * 32;
+      while (mask) {
+        int count = __builtin_ctz(mask);
+        if (count > 0) {
+          clear_buffer<T>(t, t.clear_val,
+                          IntRect{0, start, t.width, start + count}, skip_start,
+                          skip_end);
+          t.delay_clear -= count;
+          start += count;
+          mask >>= count;
+        }
+        count = __builtin_ctz(mask + 1);
+        start += count;
+        mask >>= count;
+      }
+      int count = (i + 1) * 32 - start;
+      if (count > 0) {
+        clear_buffer<T>(t, t.clear_val,
+                        IntRect{0, start, t.width, start + count}, skip_start,
+                        skip_end);
+        t.delay_clear -= count;
+      }
+    }
+  }
+  if (t.delay_clear <= 0) t.disable_delayed_clear();
+}
+
+static void prepare_texture(Texture& t, const IntRect* skip) {
+  if (t.delay_clear) {
+    switch (t.internal_format) {
+      case GL_RGBA8:
+        force_clear<uint32_t>(t, skip);
+        break;
+      case GL_R8:
+        force_clear<uint8_t>(t, skip);
+        break;
+      case GL_RG8:
+        force_clear<uint16_t>(t, skip);
+        break;
+      default:
+        assert(false);
+        break;
+    }
+  }
+}
+
+// Setup a clear on a texture. This may either force an immediate clear or
+// potentially punt to a delayed clear, if applicable.
+template <typename T>
+static void request_clear(Texture& t, T value, const IntRect& scissor) {
+  // If the clear would require a scissor, force clear anything outside
+  // the scissor, and then immediately clear anything inside the scissor.
+  if (!scissor.contains(t.offset_bounds())) {
+    IntRect skip = scissor - t.offset;
+    force_clear<T>(t, &skip);
+    clear_buffer<T>(t, value, skip.intersection(t.bounds()));
+  } else {
+    // Do delayed clear for 2D texture without scissor.
+    t.enable_delayed_clear(value);
+  }
+}
+
+template <typename T>
+static inline void request_clear(Texture& t, T value) {
+  // If scissoring is enabled, use the scissor rect. Otherwise, just scissor to
+  // the entire texture bounds.
+  request_clear(t, value, ctx->scissortest ? ctx->scissor : t.offset_bounds());
+}
+
+extern "C" {
+
+void InitDefaultFramebuffer(int x, int y, int width, int height, int stride,
+                            void* buf) {
+  Framebuffer& fb = ctx->framebuffers[0];
+  if (!fb.color_attachment) {
+    GenTextures(1, &fb.color_attachment);
+  }
+  // If the dimensions or buffer properties changed, we need to reallocate
+  // the underlying storage for the color buffer texture.
+  Texture& colortex = ctx->textures[fb.color_attachment];
+  set_tex_storage(colortex, GL_RGBA8, width, height, buf, stride);
+  colortex.offset = IntPoint(x, y);
+  if (!fb.depth_attachment) {
+    GenTextures(1, &fb.depth_attachment);
+  }
+  // Ensure dimensions of the depth buffer match the color buffer.
+  Texture& depthtex = ctx->textures[fb.depth_attachment];
+  set_tex_storage(depthtex, GL_DEPTH_COMPONENT24, width, height);
+  depthtex.offset = IntPoint(x, y);
+}
+
+void* GetColorBuffer(GLuint fbo, GLboolean flush, int32_t* width,
+                     int32_t* height, int32_t* stride) {
+  Framebuffer* fb = ctx->framebuffers.find(fbo);
+  if (!fb || !fb->color_attachment) {
+    return nullptr;
+  }
+  Texture& colortex = ctx->textures[fb->color_attachment];
+  if (flush) {
+    prepare_texture(colortex);
+  }
+  assert(colortex.offset == IntPoint(0, 0));
+  if (width) {
+    *width = colortex.width;
+  }
+  if (height) {
+    *height = colortex.height;
+  }
+  if (stride) {
+    *stride = colortex.stride();
+  }
+  return colortex.buf ? colortex.sample_ptr(0, 0) : nullptr;
+}
+
+void ResolveFramebuffer(GLuint fbo) {
+  Framebuffer* fb = ctx->framebuffers.find(fbo);
+  if (!fb || !fb->color_attachment) {
+    return;
+  }
+  Texture& colortex = ctx->textures[fb->color_attachment];
+  prepare_texture(colortex);
+}
+
+void SetTextureBuffer(GLuint texid, GLenum internal_format, GLsizei width,
+                      GLsizei height, GLsizei stride, void* buf,
+                      GLsizei min_width, GLsizei min_height) {
+  Texture& t = ctx->textures[texid];
+  set_tex_storage(t, internal_format, width, height, buf, stride, min_width,
+                  min_height);
+}
+
+GLenum CheckFramebufferStatus(GLenum target) {
+  Framebuffer* fb = get_framebuffer(target);
+  if (!fb || !fb->color_attachment) {
+    return GL_FRAMEBUFFER_UNSUPPORTED;
+  }
+  return GL_FRAMEBUFFER_COMPLETE;
+}
+
+void ClearTexSubImage(GLuint texture, GLint level, GLint xoffset, GLint yoffset,
+                      GLint zoffset, GLsizei width, GLsizei height,
+                      GLsizei depth, GLenum format, GLenum type,
+                      const void* data) {
+  if (level != 0) {
+    assert(false);
+    return;
+  }
+  Texture& t = ctx->textures[texture];
+  assert(!t.locked);
+  if (width <= 0 || height <= 0 || depth <= 0) {
+    return;
+  }
+  assert(zoffset == 0 && depth == 1);
+  IntRect scissor = {xoffset, yoffset, xoffset + width, yoffset + height};
+  if (t.internal_format == GL_DEPTH_COMPONENT24) {
+    uint32_t value = 0xFFFFFF;
+    switch (format) {
+      case GL_DEPTH_COMPONENT:
+        switch (type) {
+          case GL_DOUBLE:
+            value = uint32_t(*(const GLdouble*)data * 0xFFFFFF);
+            break;
+          case GL_FLOAT:
+            value = uint32_t(*(const GLfloat*)data * 0xFFFFFF);
+            break;
+          default:
+            assert(false);
+            break;
+        }
+        break;
+      default:
+        assert(false);
+        break;
+    }
+    if (t.cleared() && !scissor.contains(t.offset_bounds())) {
+      // If we need to scissor the clear and the depth buffer was already
+      // initialized, then just fill runs for that scissor area.
+      t.fill_depth_runs(value, scissor);
+    } else {
+      // Otherwise, the buffer is either uninitialized or the clear would
+      // encompass the entire buffer. If uninitialized, we can safely fill
+      // the entire buffer with any value and thus ignore any scissoring.
+      t.init_depth_runs(value);
+    }
+    return;
+  }
+
+  uint32_t color = 0xFF000000;
+  switch (type) {
+    case GL_FLOAT: {
+      const GLfloat* f = (const GLfloat*)data;
+      Float v = {0.0f, 0.0f, 0.0f, 1.0f};
+      switch (format) {
+        case GL_RGBA:
+          v.w = f[3];  // alpha
+          FALLTHROUGH;
+        case GL_RGB:
+          v.z = f[2];  // blue
+          FALLTHROUGH;
+        case GL_RG:
+          v.y = f[1];  // green
+          FALLTHROUGH;
+        case GL_RED:
+          v.x = f[0];  // red
+          break;
+        default:
+          assert(false);
+          break;
+      }
+      color = bit_cast<uint32_t>(CONVERT(round_pixel(v), U8));
+      break;
+    }
+    case GL_UNSIGNED_BYTE: {
+      const GLubyte* b = (const GLubyte*)data;
+      switch (format) {
+        case GL_RGBA:
+          color = (color & ~0xFF000000) | (uint32_t(b[3]) << 24);  // alpha
+          FALLTHROUGH;
+        case GL_RGB:
+          color = (color & ~0x00FF0000) | (uint32_t(b[2]) << 16);  // blue
+          FALLTHROUGH;
+        case GL_RG:
+          color = (color & ~0x0000FF00) | (uint32_t(b[1]) << 8);  // green
+          FALLTHROUGH;
+        case GL_RED:
+          color = (color & ~0x000000FF) | uint32_t(b[0]);  // red
+          break;
+        default:
+          assert(false);
+          break;
+      }
+      break;
+    }
+    default:
+      assert(false);
+      break;
+  }
+
+  switch (t.internal_format) {
+    case GL_RGBA8:
+      // Clear color needs to swizzle to BGRA.
+      request_clear<uint32_t>(t,
+                              (color & 0xFF00FF00) |
+                                  ((color << 16) & 0xFF0000) |
+                                  ((color >> 16) & 0xFF),
+                              scissor);
+      break;
+    case GL_R8:
+      request_clear<uint8_t>(t, uint8_t(color & 0xFF), scissor);
+      break;
+    case GL_RG8:
+      request_clear<uint16_t>(t, uint16_t(color & 0xFFFF), scissor);
+      break;
+    default:
+      assert(false);
+      break;
+  }
+}
+
+void ClearTexImage(GLuint texture, GLint level, GLenum format, GLenum type,
+                   const void* data) {
+  Texture& t = ctx->textures[texture];
+  IntRect scissor = t.offset_bounds();
+  ClearTexSubImage(texture, level, scissor.x0, scissor.y0, 0, scissor.width(),
+                   scissor.height(), 1, format, type, data);
+}
+
+void Clear(GLbitfield mask) {
+  Framebuffer& fb = *get_framebuffer(GL_DRAW_FRAMEBUFFER, true);
+  if ((mask & GL_COLOR_BUFFER_BIT) && fb.color_attachment) {
+    Texture& t = ctx->textures[fb.color_attachment];
+    IntRect scissor = ctx->scissortest
+                          ? ctx->scissor.intersection(t.offset_bounds())
+                          : t.offset_bounds();
+    ClearTexSubImage(fb.color_attachment, 0, scissor.x0, scissor.y0, 0,
+                     scissor.width(), scissor.height(), 1, GL_RGBA, GL_FLOAT,
+                     ctx->clearcolor);
+  }
+  if ((mask & GL_DEPTH_BUFFER_BIT) && fb.depth_attachment) {
+    Texture& t = ctx->textures[fb.depth_attachment];
+    IntRect scissor = ctx->scissortest
+                          ? ctx->scissor.intersection(t.offset_bounds())
+                          : t.offset_bounds();
+    ClearTexSubImage(fb.depth_attachment, 0, scissor.x0, scissor.y0, 0,
+                     scissor.width(), scissor.height(), 1, GL_DEPTH_COMPONENT,
+                     GL_DOUBLE, &ctx->cleardepth);
+  }
+}
+
+void ClearColorRect(GLuint fbo, GLint xoffset, GLint yoffset, GLsizei width,
+                    GLsizei height, GLfloat r, GLfloat g, GLfloat b,
+                    GLfloat a) {
+  GLfloat color[] = {r, g, b, a};
+  Framebuffer& fb = ctx->framebuffers[fbo];
+  Texture& t = ctx->textures[fb.color_attachment];
+  IntRect scissor =
+      IntRect{xoffset, yoffset, xoffset + width, yoffset + height}.intersection(
+          t.offset_bounds());
+  ClearTexSubImage(fb.color_attachment, 0, scissor.x0, scissor.y0, 0,
+                   scissor.width(), scissor.height(), 1, GL_RGBA, GL_FLOAT,
+                   color);
+}
+
+void InvalidateFramebuffer(GLenum target, GLsizei num_attachments,
+                           const GLenum* attachments) {
+  Framebuffer* fb = get_framebuffer(target);
+  if (!fb || num_attachments <= 0 || !attachments) {
+    return;
+  }
+  for (GLsizei i = 0; i < num_attachments; i++) {
+    switch (attachments[i]) {
+      case GL_DEPTH_ATTACHMENT: {
+        Texture& t = ctx->textures[fb->depth_attachment];
+        t.set_cleared(false);
+        break;
+      }
+      case GL_COLOR_ATTACHMENT0: {
+        Texture& t = ctx->textures[fb->color_attachment];
+        t.disable_delayed_clear();
+        break;
+      }
+    }
+  }
+}
+
+void ReadPixels(GLint x, GLint y, GLsizei width, GLsizei height, GLenum format,
+                GLenum type, void* data) {
+  data = get_pixel_pack_buffer_data(data);
+  if (!data) return;
+  Framebuffer* fb = get_framebuffer(GL_READ_FRAMEBUFFER);
+  if (!fb) return;
+  assert(format == GL_RED || format == GL_RGBA || format == GL_RGBA_INTEGER ||
+         format == GL_BGRA || format == GL_RG);
+  Texture& t = ctx->textures[fb->color_attachment];
+  if (!t.buf) return;
+  prepare_texture(t);
+  // debugf("read pixels %d, %d, %d, %d from fb %d with format %x\n", x, y,
+  // width, height, ctx->read_framebuffer_binding, t.internal_format);
+  x -= t.offset.x;
+  y -= t.offset.y;
+  assert(x >= 0 && y >= 0);
+  assert(x + width <= t.width);
+  assert(y + height <= t.height);
+  if (internal_format_for_data(format, type) != t.internal_format) {
+    debugf("mismatched format for read pixels: %x vs %x\n", t.internal_format,
+           internal_format_for_data(format, type));
+    assert(false);
+    return;
+  }
+  // Only support readback conversions that are reversible
+  assert(!format_requires_conversion(format, t.internal_format) ||
+         bytes_for_internal_format(format) == t.bpp());
+  uint8_t* dest = (uint8_t*)data;
+  size_t destStride = width * t.bpp();
+  if (y < 0) {
+    dest += -y * destStride;
+    height += y;
+    y = 0;
+  }
+  if (y + height > t.height) {
+    height = t.height - y;
+  }
+  if (x < 0) {
+    dest += -x * t.bpp();
+    width += x;
+    x = 0;
+  }
+  if (x + width > t.width) {
+    width = t.width - x;
+  }
+  if (width <= 0 || height <= 0) {
+    return;
+  }
+  convert_copy(format, t.internal_format, dest, destStride,
+               (const uint8_t*)t.sample_ptr(x, y), t.stride(), width, height);
+}
+
+void CopyImageSubData(GLuint srcName, GLenum srcTarget, UNUSED GLint srcLevel,
+                      GLint srcX, GLint srcY, GLint srcZ, GLuint dstName,
+                      GLenum dstTarget, UNUSED GLint dstLevel, GLint dstX,
+                      GLint dstY, GLint dstZ, GLsizei srcWidth,
+                      GLsizei srcHeight, GLsizei srcDepth) {
+  assert(srcLevel == 0 && dstLevel == 0);
+  assert(srcZ == 0 && srcDepth == 1 && dstZ == 0);
+  if (srcTarget == GL_RENDERBUFFER) {
+    Renderbuffer& rb = ctx->renderbuffers[srcName];
+    srcName = rb.texture;
+  }
+  if (dstTarget == GL_RENDERBUFFER) {
+    Renderbuffer& rb = ctx->renderbuffers[dstName];
+    dstName = rb.texture;
+  }
+  Texture& srctex = ctx->textures[srcName];
+  if (!srctex.buf) return;
+  prepare_texture(srctex);
+  Texture& dsttex = ctx->textures[dstName];
+  if (!dsttex.buf) return;
+  assert(!dsttex.locked);
+  IntRect skip = {dstX, dstY, dstX + srcWidth, dstY + srcHeight};
+  prepare_texture(dsttex, &skip);
+  assert(srctex.internal_format == dsttex.internal_format);
+  assert(srcWidth >= 0);
+  assert(srcHeight >= 0);
+  assert(srcX + srcWidth <= srctex.width);
+  assert(srcY + srcHeight <= srctex.height);
+  assert(dstX + srcWidth <= dsttex.width);
+  assert(dstY + srcHeight <= dsttex.height);
+  int bpp = srctex.bpp();
+  int src_stride = srctex.stride();
+  int dest_stride = dsttex.stride();
+  char* dest = dsttex.sample_ptr(dstX, dstY);
+  char* src = srctex.sample_ptr(srcX, srcY);
+  for (int y = 0; y < srcHeight; y++) {
+    memcpy(dest, src, srcWidth * bpp);
+    dest += dest_stride;
+    src += src_stride;
+  }
+}
+
+void CopyTexSubImage2D(GLenum target, UNUSED GLint level, GLint xoffset,
+                       GLint yoffset, GLint x, GLint y, GLsizei width,
+                       GLsizei height) {
+  assert(level == 0);
+  Framebuffer* fb = get_framebuffer(GL_READ_FRAMEBUFFER);
+  if (!fb) return;
+  CopyImageSubData(fb->color_attachment, GL_TEXTURE_2D, 0, x, y, 0,
+                   ctx->get_binding(target), GL_TEXTURE_2D, 0, xoffset, yoffset,
+                   0, width, height, 1);
+}
+
+}  // extern "C"
+
+#include "blend.h"
+#include "composite.h"
+#include "swgl_ext.h"
+
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wuninitialized"
+#pragma GCC diagnostic ignored "-Wunused-function"
+#pragma GCC diagnostic ignored "-Wunused-parameter"
+#pragma GCC diagnostic ignored "-Wunused-variable"
+#pragma GCC diagnostic ignored "-Wimplicit-fallthrough"
+#ifdef __clang__
+#  pragma GCC diagnostic ignored "-Wunused-private-field"
+#else
+#  pragma GCC diagnostic ignored "-Wunused-but-set-variable"
+#endif
+#include "load_shader.h"
+#pragma GCC diagnostic pop
+
+#include "rasterize.h"
+
+void VertexArray::validate() {
+  int last_enabled = -1;
+  for (int i = 0; i <= max_attrib; i++) {
+    VertexAttrib& attr = attribs[i];
+    if (attr.enabled) {
+      // VertexArray &v = ctx->vertex_arrays[attr.vertex_array];
+      Buffer& vertex_buf = ctx->buffers[attr.vertex_buffer];
+      attr.buf = vertex_buf.buf;
+      attr.buf_size = vertex_buf.size;
+      // debugf("%d %x %d %d %d %d\n", i, attr.type, attr.size, attr.stride,
+      // attr.offset, attr.divisor);
+      last_enabled = i;
+    }
+  }
+  max_attrib = last_enabled;
+}
+
+extern "C" {
+
+void DrawElementsInstanced(GLenum mode, GLsizei count, GLenum type,
+                           GLintptr offset, GLsizei instancecount) {
+  if (offset < 0 || count <= 0 || instancecount <= 0 || !vertex_shader ||
+      !fragment_shader) {
+    return;
+  }
+
+  Framebuffer& fb = *get_framebuffer(GL_DRAW_FRAMEBUFFER, true);
+  if (!fb.color_attachment) {
+    return;
+  }
+  Texture& colortex = ctx->textures[fb.color_attachment];
+  if (!colortex.buf) {
+    return;
+  }
+  assert(!colortex.locked);
+  assert(colortex.internal_format == GL_RGBA8 ||
+         colortex.internal_format == GL_R8);
+  Texture& depthtex = ctx->textures[ctx->depthtest ? fb.depth_attachment : 0];
+  if (depthtex.buf) {
+    assert(depthtex.internal_format == GL_DEPTH_COMPONENT24);
+    assert(colortex.width == depthtex.width &&
+           colortex.height == depthtex.height);
+    assert(colortex.offset == depthtex.offset);
+  }
+
+  // debugf("current_vertex_array %d\n", ctx->current_vertex_array);
+  // debugf("indices size: %d\n", indices_buf.size);
+  VertexArray& v = ctx->vertex_arrays[ctx->current_vertex_array];
+  if (ctx->validate_vertex_array) {
+    ctx->validate_vertex_array = false;
+    v.validate();
+  }
+
+#ifdef PRINT_TIMINGS
+  uint64_t start = get_time_value();
+#endif
+
+  ctx->shaded_rows = 0;
+  ctx->shaded_pixels = 0;
+
+  vertex_shader->init_batch();
+
+  switch (type) {
+    case GL_UNSIGNED_SHORT:
+      assert(mode == GL_TRIANGLES);
+      draw_elements<uint16_t>(count, instancecount, offset, v, colortex,
+                              depthtex);
+      break;
+    case GL_UNSIGNED_INT:
+      assert(mode == GL_TRIANGLES);
+      draw_elements<uint32_t>(count, instancecount, offset, v, colortex,
+                              depthtex);
+      break;
+    case GL_NONE:
+      // Non-standard GL extension - if element type is GL_NONE, then we don't
+      // use any element buffer and behave as if DrawArrays was called instead.
+      for (GLsizei instance = 0; instance < instancecount; instance++) {
+        switch (mode) {
+          case GL_LINES:
+            for (GLsizei i = 0; i + 2 <= count; i += 2) {
+              vertex_shader->load_attribs(v.attribs, offset + i, instance, 2);
+              draw_quad(2, colortex, depthtex);
+            }
+            break;
+          case GL_TRIANGLES:
+            for (GLsizei i = 0; i + 3 <= count; i += 3) {
+              vertex_shader->load_attribs(v.attribs, offset + i, instance, 3);
+              draw_quad(3, colortex, depthtex);
+            }
+            break;
+          default:
+            assert(false);
+            break;
+        }
+      }
+      break;
+    default:
+      assert(false);
+      break;
+  }
+
+  if (ctx->samples_passed_query) {
+    Query& q = ctx->queries[ctx->samples_passed_query];
+    q.value += ctx->shaded_pixels;
+  }
+
+#ifdef PRINT_TIMINGS
+  uint64_t end = get_time_value();
+  printf(
+      "%7.3fms draw(%s, %d): %d pixels in %d rows (avg %f pixels/row, "
+      "%fns/pixel)\n",
+      double(end - start) / (1000. * 1000.),
+      ctx->programs[ctx->current_program].impl->get_name(), instancecount,
+      ctx->shaded_pixels, ctx->shaded_rows,
+      double(ctx->shaded_pixels) / ctx->shaded_rows,
+      double(end - start) / max(ctx->shaded_pixels, 1));
+#endif
+}
+
+void Finish() {
+#ifdef PRINT_TIMINGS
+  printf("Finish\n");
+#endif
+}
+
+void MakeCurrent(Context* c) {
+  if (ctx == c) {
+    return;
+  }
+  ctx = c;
+  setup_program(ctx ? ctx->current_program : 0);
+}
+
+Context* CreateContext() { return new Context; }
+
+void ReferenceContext(Context* c) {
+  if (!c) {
+    return;
+  }
+  ++c->references;
+}
+
+void DestroyContext(Context* c) {
+  if (!c) {
+    return;
+  }
+  assert(c->references > 0);
+  --c->references;
+  if (c->references > 0) {
+    return;
+  }
+  if (ctx == c) {
+    MakeCurrent(nullptr);
+  }
+  delete c;
+}
+
+size_t ReportMemory(Context* ctx, size_t (*size_of_op)(const void*)) {
+  size_t size = 0;
+  if (ctx) {
+    for (auto& t : ctx->textures) {
+      if (t && t->should_free()) {
+        size += size_of_op(t->buf);
+      }
+    }
+  }
+  return size;
+}
+}  // extern "C"
diff --git a/gfx/wr/swgl/src/gl_defs.h b/gfx/wr/swgl/src/gl_defs.h
new file mode 100644
index 0000000000..b60eaad0ec
--- /dev/null
+++ b/gfx/wr/swgl/src/gl_defs.h
@@ -0,0 +1,220 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+typedef int8_t GLbyte;
+typedef uint8_t GLubyte;
+typedef int16_t GLshort;
+typedef uint16_t GLushort;
+typedef int32_t GLint;
+typedef uint32_t GLuint;
+typedef int64_t GLint64;
+typedef uint64_t GLuint64;
+
+typedef float GLfloat;
+typedef double GLdouble;
+
+typedef uint32_t GLenum;
+typedef uint8_t GLboolean;
+typedef uint32_t GLbitfield;
+
+typedef int32_t GLsizei;
+typedef size_t GLsizeiptr;
+typedef intptr_t GLintptr;
+
+#define GL_FALSE 0
+#define GL_TRUE 1
+
+#define GL_NONE 0
+
+#define GL_NO_ERROR 0
+#define GL_OUT_OF_MEMORY 0x0505
+
+#define GL_RGBA32F 0x8814
+#define GL_RGBA8 0x8058
+#define GL_R8 0x8229
+#define GL_R16 0x822A
+#define GL_RG16 0x822C
+#define GL_RGBA32I 0x8D82
+#define GL_BGRA8 0x93A1
+#define GL_RG8 0x822B
+
+#define GL_BYTE 0x1400
+#define GL_UNSIGNED_BYTE 0x1401
+#define GL_SHORT 0x1402
+#define GL_UNSIGNED_SHORT 0x1403
+#define GL_INT 0x1404
+#define GL_UNSIGNED_INT 0x1405
+#define GL_FLOAT 0x1406
+#define GL_DOUBLE 0x1408
+
+#define GL_RED 0x1903
+#define GL_GREEN 0x1904
+#define GL_BLUE 0x1905
+#define GL_ALPHA 0x1906
+#define GL_RGB 0x1907
+#define GL_RGBA 0x1908
+#define GL_RGBA_INTEGER 0x8D99
+#define GL_BGRA 0x80E1
+#define GL_RG 0x8227
+
+#define GL_DEPTH_COMPONENT 0x1902
+#define GL_DEPTH_COMPONENT16 0x81A5
+#define GL_DEPTH_COMPONENT24 0x81A6
+#define GL_DEPTH_COMPONENT32 0x81A7
+
+#define GL_ARRAY_BUFFER 0x8892
+#define GL_ELEMENT_ARRAY_BUFFER 0x8893
+
+#define GL_READ_FRAMEBUFFER 0x8CA8
+#define GL_DRAW_FRAMEBUFFER 0x8CA9
+#define GL_FRAMEBUFFER 0x8D40
+#define GL_DRAW_FRAMEBUFFER_BINDING 0x8CA6
+#define GL_READ_FRAMEBUFFER_BINDING 0x8CAA
+#define GL_RENDERBUFFER 0x8D41
+#define GL_COLOR_ATTACHMENT0 0x8CE0
+#define GL_DEPTH_ATTACHMENT 0x8D00
+#define GL_STENCIL_ATTACHMENT 0x8D20
+#define GL_FRAMEBUFFER_COMPLETE 0x8CD5
+#define GL_FRAMEBUFFER_INCOMPLETE_ATTACHMENT 0x8CD6
+#define GL_FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT 0x8CD7
+#define GL_FRAMEBUFFER_INCOMPLETE_DRAW_BUFFER 0x8CDB
+#define GL_FRAMEBUFFER_INCOMPLETE_READ_BUFFER 0x8CDC
+#define GL_FRAMEBUFFER_UNSUPPORTED 0x8CDD
+#define GL_COLOR_BUFFER_BIT 0x00004000
+#define GL_DEPTH_BUFFER_BIT 0x00000100
+#define GL_STENCIL_BUFFER_BIT 0x00000400
+
+#define GL_PIXEL_PACK_BUFFER 0x88EB
+#define GL_PIXEL_UNPACK_BUFFER 0x88EC
+#define GL_PIXEL_PACK_BUFFER_BINDING 0x88ED
+#define GL_PIXEL_UNPACK_BUFFER_BINDING 0x88EF
+#define GL_UNPACK_ROW_LENGTH 0x0CF2
+#define GL_UNPACK_ALIGNMENT 0x0CF5
+
+#define GL_QUERY_RESULT 0x8866
+#define GL_QUERY_RESULT_AVAILABLE 0x8867
+#define GL_TIME_ELAPSED 0x88BF
+#define GL_SAMPLES_PASSED 0x8914
+
+#define GL_NEAREST 0x2600
+#define GL_LINEAR 0x2601
+#define GL_NEAREST_MIPMAP_NEAREST 0x2700
+#define GL_NEAREST_MIPMAP_LINEAR 0x2702
+#define GL_LINEAR_MIPMAP_NEAREST 0x2701
+#define GL_LINEAR_MIPMAP_LINEAR 0x2703
+#define GL_TEXTURE_WRAP_S 0x2802
+#define GL_TEXTURE_WRAP_T 0x2803
+#define GL_TEXTURE_MAG_FILTER 0x2800
+#define GL_TEXTURE_MIN_FILTER 0x2801
+#define GL_CLAMP_TO_EDGE 0x812F
+#define GL_TEXTURE_2D 0x0DE1
+#define GL_TEXTURE_3D 0x806F
+#define GL_TEXTURE_2D_ARRAY 0x8C1A
+#define GL_TEXTURE_RECTANGLE 0x84F5
+#define GL_TEXTURE0 0x84C0
+#define GL_TEXTURE1 0x84C1
+#define GL_TEXTURE2 0x84C2
+#define GL_TEXTURE3 0x84C3
+#define GL_TEXTURE4 0x84C4
+#define GL_TEXTURE5 0x84C5
+#define GL_TEXTURE6 0x84C6
+#define GL_TEXTURE7 0x84C7
+#define GL_TEXTURE8 0x84C8
+#define GL_TEXTURE9 0x84C9
+#define GL_TEXTURE10 0x84CA
+#define GL_TEXTURE11 0x84CB
+#define GL_TEXTURE12 0x84CC
+#define GL_TEXTURE13 0x84CD
+#define GL_TEXTURE14 0x84CE
+#define GL_TEXTURE15 0x84CF
+#define GL_MAX_TEXTURE_UNITS 0x84E2
+#define GL_MAX_TEXTURE_IMAGE_UNITS 0x8872
+#define GL_MAX_TEXTURE_SIZE 0x0D33
+#define GL_MAX_ARRAY_TEXTURE_LAYERS 0x88FF
+#define GL_MIN_PROGRAM_TEXEL_OFFSET 0x8904
+#define GL_MAX_PROGRAM_TEXEL_OFFSET 0x8905
+
+#define GL_VERTEX_SHADER 0x8B31
+#define GL_FRAGMENT_SHADER 0x8B30
+
+#define GL_BLEND 0x0BE2
+#define GL_ZERO 0
+#define GL_ONE 1
+#define GL_SRC_COLOR 0x0300
+#define GL_ONE_MINUS_SRC_COLOR 0x0301
+#define GL_SRC_ALPHA 0x0302
+#define GL_ONE_MINUS_SRC_ALPHA 0x0303
+#define GL_DST_ALPHA 0x0304
+#define GL_ONE_MINUS_DST_ALPHA 0x0305
+#define GL_DST_COLOR 0x0306
+#define GL_ONE_MINUS_DST_COLOR 0x0307
+#define GL_CONSTANT_COLOR 0x8001
+#define GL_ONE_MINUS_CONSTANT_COLOR 0x8002
+#define GL_CONSTANT_ALPHA 0x8003
+#define GL_ONE_MINUS_CONSTANT_ALPHA 0x8004
+#define GL_SRC1_ALPHA 0x8589
+#define GL_SRC1_COLOR 0x88F9
+#define GL_ONE_MINUS_SRC1_COLOR 0x88FA
+#define GL_ONE_MINUS_SRC1_ALPHA 0x88FB
+
+#define GL_FUNC_ADD 0x8006
+#define GL_MIN 0x8007
+#define GL_MAX 0x8008
+
+#define GL_NEVER 0x0200
+#define GL_LESS 0x0201
+#define GL_EQUAL 0x0202
+#define GL_LEQUAL 0x0203
+#define GL_GREATER 0x0204
+#define GL_NOTEQUAL 0x0205
+#define GL_GEQUAL 0x0206
+#define GL_ALWAYS 0x0207
+#define GL_DEPTH_TEST 0x0B71
+#define GL_DEPTH_WRITEMASK 0x0B72
+
+#define GL_SCISSOR_TEST 0x0C11
+
+#define GL_VENDOR 0x1F00
+#define GL_RENDERER 0x1F01
+#define GL_VERSION 0x1F02
+#define GL_EXTENSIONS 0x1F03
+#define GL_NUM_EXTENSIONS 0x821D
+#define GL_MINOR_VERSION 0x821C
+#define GL_MAJOR_VERSION 0x821B
+#define GL_SHADING_LANGUAGE_VERSION 0x8B8C
+
+#define GL_POINTS 0x0000
+#define GL_LINES 0x0001
+#define GL_LINE_LOOP 0x0002
+#define GL_LINE_STRIP 0x0003
+#define GL_TRIANGLES 0x0004
+#define GL_TRIANGLE_STRIP 0x0005
+#define GL_TRIANGLE_FAN 0x0006
+#define GL_QUADS 0x0007
+
+#define GL_UNSIGNED_INT_8_8_8_8_REV 0x8367
+
+#define GL_RGB_422_APPLE 0x8A1F
+#define GL_UNSIGNED_SHORT_8_8_APPLE 0x85BA
+#define GL_UNSIGNED_SHORT_8_8_REV_APPLE 0x85BB
+#define GL_RGB_RAW_422_APPLE 0x8A51
+
+#define GL_MULTIPLY_KHR 0x9294
+#define GL_SCREEN_KHR 0x9295
+#define GL_OVERLAY_KHR 0x9296
+#define GL_DARKEN_KHR 0x9297
+#define GL_LIGHTEN_KHR 0x9298
+#define GL_COLORDODGE_KHR 0x9299
+#define GL_COLORBURN_KHR 0x929A
+#define GL_HARDLIGHT_KHR 0x929B
+#define GL_SOFTLIGHT_KHR 0x929C
+#define GL_DIFFERENCE_KHR 0x929E
+#define GL_EXCLUSION_KHR 0x92A0
+#define GL_HSL_HUE_KHR 0x92AD
+#define GL_HSL_SATURATION_KHR 0x92AE
+#define GL_HSL_COLOR_KHR 0x92AF
+#define GL_HSL_LUMINOSITY_KHR 0x92B0
+
+#define SWGL_BLEND_DROP_SHADOW 0xB001
+#define SWGL_BLEND_SUBPIXEL_TEXT 0xB002
diff --git a/gfx/wr/swgl/src/glsl.h b/gfx/wr/swgl/src/glsl.h
new file mode 100644
index 0000000000..64743b78c8
--- /dev/null
+++ b/gfx/wr/swgl/src/glsl.h
@@ -0,0 +1,3117 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#define SI ALWAYS_INLINE static
+
+#include "vector_type.h"
+
+namespace glsl {
+
+enum TextureFormat { RGBA32F, RGBA32I, RGBA8, R8, RG8, R16, RG16, YUV422 };
+
+enum TextureFilter { NEAREST, LINEAR };
+
+struct samplerCommon {
+  uint32_t* buf = nullptr;
+  uint32_t stride = 0;  // in units of BPP if < 4, or dwords if BPP >= 4
+  uint32_t height = 0;
+  uint32_t width = 0;
+  TextureFormat format = TextureFormat::RGBA8;
+};
+
+struct samplerFilter {
+  TextureFilter filter = TextureFilter::NEAREST;
+};
+
+struct sampler2D_impl : samplerCommon, samplerFilter {};
+typedef sampler2D_impl* sampler2D;
+
+typedef struct sampler2DR8_impl : sampler2D_impl{} * sampler2DR8;
+typedef struct sampler2DRG8_impl : sampler2D_impl{} * sampler2DRG8;
+typedef struct sampler2DRGBA8_impl : sampler2D_impl{} * sampler2DRGBA8;
+typedef struct sampler2DRGBA32F_impl : sampler2D_impl{} * sampler2DRGBA32F;
+
+struct isampler2D_impl : samplerCommon {};
+typedef isampler2D_impl* isampler2D;
+
+struct isampler2DRGBA32I_impl : isampler2D_impl {};
+typedef isampler2DRGBA32I_impl* isampler2DRGBA32I;
+
+struct sampler2DRect_impl : samplerCommon, samplerFilter {};
+typedef sampler2DRect_impl* sampler2DRect;
+
+#if USE_SSE2
+SI bool test_all(Bool cond) { return _mm_movemask_ps(cond) == 0xF; }
+SI bool test_any(Bool cond) { return _mm_movemask_ps(cond) != 0; }
+SI bool test_none(Bool cond) { return _mm_movemask_ps(cond) == 0; }
+#else
+SI bool test_all(Bool cond) {
+  return bit_cast<uint32_t>(CONVERT(cond, U8)) == 0xFFFFFFFFU;
+}
+SI bool test_any(Bool cond) {
+  return bit_cast<uint32_t>(CONVERT(cond, U8)) != 0;
+}
+SI bool test_none(Bool cond) {
+  return bit_cast<uint32_t>(CONVERT(cond, U8)) == 0;
+}
+#endif
+SI bool test_equal(Bool cond) { return test_none(cond != cond.x); }
+
+float make_float(float n) { return n; }
+
+float make_float(int32_t n) { return float(n); }
+
+float make_float(uint32_t n) { return float(n); }
+
+float make_float(bool n) { return float(n); }
+
+template <typename T>
+Float make_float(T v) {
+  return CONVERT(v, Float);
+}
+
+int32_t make_int(uint32_t n) { return n; }
+
+int32_t make_int(int32_t n) { return n; }
+
+int32_t make_int(float n) { return int32_t(n); }
+
+int32_t make_int(bool n) { return int32_t(n); }
+
+template <typename T>
+I32 make_int(T v) {
+  return CONVERT(v, I32);
+}
+
+uint32_t make_uint(uint32_t n) { return n; }
+
+uint32_t make_uint(int32_t n) { return n; }
+
+uint32_t make_uint(float n) { return uint32_t(n); }
+
+uint32_t make_uint(bool n) { return uint32_t(n); }
+
+template <typename T>
+U32 make_uint(T v) {
+  return CONVERT(v, U32);
+}
+
+template <typename T>
+T force_scalar(T n) {
+  return n;
+}
+
+float force_scalar(Float f) { return f[0]; }
+
+int32_t force_scalar(I32 i) { return i[0]; }
+
+struct vec4;
+struct ivec2;
+
+SI int32_t if_then_else(int32_t c, int32_t t, int32_t e) { return c ? t : e; }
+SI int32_t if_then_else(bool c, int32_t t, int32_t e) { return c ? t : e; }
+
+SI float if_then_else(int32_t c, float t, float e) { return c ? t : e; }
+
+SI Float if_then_else(I32 c, float t, float e) {
+  return bit_cast<Float>((c & bit_cast<I32>(Float(t))) |
+                         (~c & bit_cast<I32>(Float(e))));
+}
+
+SI I32 if_then_else(I32 c, int32_t t, int32_t e) {
+  return (c & I32(t)) | (~c & I32(e));
+}
+
+SI U32 if_then_else(I32 c, U32 t, U32 e) {
+  return bit_cast<U32>((c & bit_cast<I32>(t)) | (~c & bit_cast<I32>(e)));
+}
+
+// Cheaper version of if_then_else that returns Float(0) if condition is false.
+SI Float if_then(I32 c, Float t) {
+  return bit_cast<Float>(c & bit_cast<I32>(t));
+}
+
+SI Float if_then_else(I32 c, Float t, Float e) {
+  return bit_cast<Float>((c & bit_cast<I32>(t)) | (~c & bit_cast<I32>(e)));
+}
+
+SI Float if_then_else(int32_t c, Float t, Float e) { return c ? t : e; }
+
+SI Bool if_then_else(I32 c, Bool t, Bool e) { return (c & t) | (~c & e); }
+
+SI Bool if_then_else(int32_t c, Bool t, Bool e) { return c ? t : e; }
+
+SI I16 if_then_else(I16 c, I16 t, I16 e) { return (c & t) | (~c & e); }
+
+template <typename T>
+SI void swap(T& a, T& b) {
+  T t(a);
+  a = b;
+  b = t;
+}
+
+SI int32_t min(int32_t a, int32_t b) { return a < b ? a : b; }
+SI int32_t max(int32_t a, int32_t b) { return a > b ? a : b; }
+
+SI int32_t clamp(int32_t a, int32_t minVal, int32_t maxVal) {
+  return min(max(a, minVal), maxVal);
+}
+
+SI float min(float a, float b) { return a < b ? a : b; }
+SI float max(float a, float b) { return a > b ? a : b; }
+
+SI float clamp(float a, float minVal, float maxVal) {
+  return min(max(a, minVal), maxVal);
+}
+
+SI Float min(Float a, Float b) {
+#if USE_SSE2
+  return _mm_min_ps(a, b);
+#elif USE_NEON
+  return vminq_f32(a, b);
+#else
+  return if_then_else(a < b, a, b);
+#endif
+}
+
+SI Float max(Float a, Float b) {
+#if USE_SSE2
+  return _mm_max_ps(a, b);
+#elif USE_NEON
+  return vmaxq_f32(a, b);
+#else
+  return if_then_else(a > b, a, b);
+#endif
+}
+
+SI Float clamp(Float a, Float minVal, Float maxVal) {
+  return min(max(a, minVal), maxVal);
+}
+
+#define sqrt __glsl_sqrt
+
+SI float sqrt(float x) { return sqrtf(x); }
+
+SI Float sqrt(Float v) {
+#if USE_SSE2
+  return _mm_sqrt_ps(v);
+#elif USE_NEON
+  Float e = vrsqrteq_f32(v);
+  e *= vrsqrtsq_f32(v, e * e);
+  e *= vrsqrtsq_f32(v, e * e);
+  return if_then(v != Float(0.0f), v * e);
+#else
+  return (Float){sqrtf(v.x), sqrtf(v.y), sqrtf(v.z), sqrtf(v.w)};
+#endif
+}
+
+SI float recip(float x) {
+#if USE_SSE2
+  return _mm_cvtss_f32(_mm_rcp_ss(_mm_set_ss(x)));
+#else
+  return 1.0f / x;
+#endif
+}
+
+// Use a fast vector reciprocal approximation when available. This should only
+// be used in cases where it is okay that the approximation is imprecise -
+// essentially visually correct but numerically wrong. Otherwise just rely on
+// however the compiler would implement slower division if the platform doesn't
+// provide a convenient intrinsic.
+SI Float recip(Float v) {
+#if USE_SSE2
+  return _mm_rcp_ps(v);
+#elif USE_NEON
+  Float e = vrecpeq_f32(v);
+  return vrecpsq_f32(v, e) * e;
+#else
+  return 1.0f / v;
+#endif
+}
+
+SI float inversesqrt(float x) {
+#if USE_SSE2
+  return _mm_cvtss_f32(_mm_rsqrt_ss(_mm_set_ss(x)));
+#else
+  return 1.0f / sqrtf(x);
+#endif
+}
+
+SI Float inversesqrt(Float v) {
+#if USE_SSE2
+  return _mm_rsqrt_ps(v);
+#elif USE_NEON
+  Float e = vrsqrteq_f32(v);
+  return vrsqrtsq_f32(v, e * e) * e;
+#else
+  return 1.0f / sqrt(v);
+#endif
+}
+
+SI float step(float edge, float x) { return float(x >= edge); }
+
+SI Float step(Float edge, Float x) { return if_then(x >= edge, Float(1)); }
+
+/*
+enum RGBA {
+        R,
+        G,
+        B,
+        A
+};*/
+
+enum XYZW {
+  X = 0,
+  Y = 1,
+  Z = 2,
+  W = 3,
+  R = 0,
+  G = 1,
+  B = 2,
+  A = 3,
+};
+
+struct bvec4_scalar;
+
+struct bvec2_scalar {
+  bool x;
+  bool y;
+
+  bvec2_scalar() : bvec2_scalar(false) {}
+  IMPLICIT constexpr bvec2_scalar(bool a) : x(a), y(a) {}
+  constexpr bvec2_scalar(bool x, bool y) : x(x), y(y) {}
+
+  bool& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      default:
+        UNREACHABLE;
+    }
+  }
+  bool sel(XYZW c1) { return select(c1); }
+
+  bvec2_scalar sel(XYZW c1, XYZW c2) {
+    return bvec2_scalar(select(c1), select(c2));
+  }
+  bvec4_scalar sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4);
+};
+
+struct bvec2_scalar1 {
+  bool x;
+
+  IMPLICIT constexpr bvec2_scalar1(bool a) : x(a) {}
+
+  operator bvec2_scalar() const { return bvec2_scalar(x); }
+};
+
+struct bvec2 {
+  bvec2() : bvec2(0) {}
+  IMPLICIT bvec2(Bool a) : x(a), y(a) {}
+  bvec2(Bool x, Bool y) : x(x), y(y) {}
+  Bool& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      default:
+        UNREACHABLE;
+    }
+  }
+  Bool sel(XYZW c1) { return select(c1); }
+
+  bvec2 sel(XYZW c1, XYZW c2) { return bvec2(select(c1), select(c2)); }
+
+  bvec2 operator~() { return bvec2(~x, ~y); }
+
+  Bool x;
+  Bool y;
+};
+
+bvec2_scalar1 make_bvec2(bool n) { return bvec2_scalar1(n); }
+
+bvec2_scalar make_bvec2(bool x, bool y) { return bvec2_scalar{x, y}; }
+
+template <typename N>
+bvec2 make_bvec2(const N& n) {
+  return bvec2(n);
+}
+
+template <typename X, typename Y>
+bvec2 make_bvec2(const X& x, const Y& y) {
+  return bvec2(x, y);
+}
+
+struct vec3_scalar;
+struct vec4_scalar;
+
+struct vec2_scalar {
+  typedef struct vec2 vector_type;
+  typedef float element_type;
+
+  float x;
+  float y;
+
+  constexpr vec2_scalar() : vec2_scalar(0.0f) {}
+  IMPLICIT constexpr vec2_scalar(float a) : x(a), y(a) {}
+  IMPLICIT constexpr vec2_scalar(int a) : x(a), y(a) {}
+  constexpr vec2_scalar(float x, float y) : x(x), y(y) {}
+
+  float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      default:
+        UNREACHABLE;
+    }
+  }
+  float& sel(XYZW c1) { return select(c1); }
+  vec2_scalar sel(XYZW c1, XYZW c2) {
+    return vec2_scalar(select(c1), select(c2));
+  }
+  vec3_scalar sel(XYZW c1, XYZW c2, XYZW c3);
+  vec4_scalar sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4);
+
+  friend bool operator==(const vec2_scalar& l, const vec2_scalar& r) {
+    return l.x == r.x && l.y == r.y;
+  }
+
+  friend bool operator!=(const vec2_scalar& l, const vec2_scalar& r) {
+    return l.x != r.x || l.y != r.y;
+  }
+
+  friend vec2_scalar operator*(float a, vec2_scalar b) {
+    return vec2_scalar(a * b.x, a * b.y);
+  }
+  friend vec2_scalar operator*(vec2_scalar a, float b) {
+    return vec2_scalar(a.x * b, a.y * b);
+  }
+  friend vec2_scalar operator*(vec2_scalar a, vec2_scalar b) {
+    return vec2_scalar(a.x * b.x, a.y * b.y);
+  }
+  friend vec2_scalar operator/(vec2_scalar a, float b) {
+    return vec2_scalar(a.x / b, a.y / b);
+  }
+  friend vec2_scalar operator/(vec2_scalar a, vec2_scalar b) {
+    return vec2_scalar(a.x / b.x, a.y / b.y);
+  }
+
+  friend vec2_scalar operator-(vec2_scalar a, vec2_scalar b) {
+    return vec2_scalar(a.x - b.x, a.y - b.y);
+  }
+  friend vec2_scalar operator-(vec2_scalar a, float b) {
+    return vec2_scalar(a.x - b, a.y - b);
+  }
+  friend vec2_scalar operator-(float a, vec2_scalar b) {
+    return vec2_scalar(a - b.x, a - b.y);
+  }
+  friend vec2_scalar operator+(vec2_scalar a, vec2_scalar b) {
+    return vec2_scalar(a.x + b.x, a.y + b.y);
+  }
+  friend vec2_scalar operator+(vec2_scalar a, float b) {
+    return vec2_scalar(a.x + b, a.y + b);
+  }
+
+  vec2_scalar operator-() { return vec2_scalar(-x, -y); }
+
+  vec2_scalar operator*=(vec2_scalar a) {
+    x *= a.x;
+    y *= a.y;
+    return *this;
+  }
+
+  vec2_scalar operator/=(vec2_scalar a) {
+    x /= a.x;
+    y /= a.y;
+    return *this;
+  }
+
+  vec2_scalar operator+=(vec2_scalar a) {
+    x += a.x;
+    y += a.y;
+    return *this;
+  }
+
+  vec2_scalar operator-=(vec2_scalar a) {
+    x -= a.x;
+    y -= a.y;
+    return *this;
+  }
+};
+
+struct vec2_scalar_ref {
+  vec2_scalar_ref(float& x, float& y) : x(x), y(y) {}
+  float& x;
+  float& y;
+
+  float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      default:
+        UNREACHABLE;
+    }
+  }
+  float& sel(XYZW c1) { return select(c1); }
+
+  vec2_scalar_ref& operator=(const vec2_scalar& a) {
+    x = a.x;
+    y = a.y;
+    return *this;
+  }
+  vec2_scalar_ref& operator*=(vec2_scalar a) {
+    x *= a.x;
+    y *= a.y;
+    return *this;
+  }
+  operator vec2_scalar() const { return vec2_scalar{x, y}; }
+};
+
+struct vec2 {
+  typedef struct vec2 vector_type;
+  typedef float element_type;
+
+  constexpr vec2() : vec2(Float(0.0f)) {}
+  IMPLICIT constexpr vec2(Float a) : x(a), y(a) {}
+  vec2(Float x, Float y) : x(x), y(y) {}
+  IMPLICIT constexpr vec2(vec2_scalar s) : x(s.x), y(s.y) {}
+  constexpr vec2(vec2_scalar s0, vec2_scalar s1, vec2_scalar s2, vec2_scalar s3)
+      : x(Float{s0.x, s1.x, s2.x, s3.x}), y(Float{s0.y, s1.y, s2.y, s3.y}) {}
+  explicit vec2(ivec2 a);
+  Float x;
+  Float y;
+
+  Float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      default:
+        UNREACHABLE;
+    }
+  }
+  Float& sel(XYZW c1) { return select(c1); }
+  vec2 sel(XYZW c1, XYZW c2) { return vec2(select(c1), select(c2)); }
+
+  vec4 sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4);
+
+  vec2 operator*=(Float a) {
+    x *= a;
+    y *= a;
+    return *this;
+  }
+  vec2 operator*=(vec2 a) {
+    x *= a.x;
+    y *= a.y;
+    return *this;
+  }
+
+  vec2 operator/=(Float a) {
+    x /= a;
+    y /= a;
+    return *this;
+  }
+  vec2 operator/=(vec2 a) {
+    x /= a.x;
+    y /= a.y;
+    return *this;
+  }
+
+  vec2 operator+=(vec2 a) {
+    x += a.x;
+    y += a.y;
+    return *this;
+  }
+  vec2 operator-=(vec2 a) {
+    x -= a.x;
+    y -= a.y;
+    return *this;
+  }
+  vec2 operator-=(Float a) {
+    x -= a;
+    y -= a;
+    return *this;
+  }
+
+  vec2 operator-() { return vec2(-x, -y); }
+
+  friend I32 operator==(const vec2& l, const vec2& r) {
+    return l.x == r.x && l.y == r.y;
+  }
+
+  friend I32 operator!=(const vec2& l, const vec2& r) {
+    return l.x != r.x || l.y != r.y;
+  }
+
+  friend vec2 operator*(vec2 a, Float b) { return vec2(a.x * b, a.y * b); }
+  friend vec2 operator*(vec2 a, vec2 b) { return vec2(a.x * b.x, a.y * b.y); }
+  friend vec2 operator*(Float a, vec2 b) { return vec2(a * b.x, a * b.y); }
+
+  friend vec2 operator/(vec2 a, vec2 b) { return vec2(a.x / b.x, a.y / b.y); }
+  friend vec2 operator/(vec2 a, Float b) { return vec2(a.x / b, a.y / b); }
+
+  friend vec2 operator-(vec2 a, vec2 b) { return vec2(a.x - b.x, a.y - b.y); }
+  friend vec2 operator-(vec2 a, Float b) { return vec2(a.x - b, a.y - b); }
+  friend vec2 operator-(Float a, vec2 b) { return vec2(a - b.x, a - b.y); }
+  friend vec2 operator+(vec2 a, vec2 b) { return vec2(a.x + b.x, a.y + b.y); }
+  friend vec2 operator+(vec2 a, Float b) { return vec2(a.x + b, a.y + b); }
+  friend vec2 operator+(Float a, vec2 b) { return vec2(a + b.x, a + b.y); }
+};
+
+vec2_scalar force_scalar(const vec2& v) {
+  return vec2_scalar{force_scalar(v.x), force_scalar(v.y)};
+}
+
+vec2_scalar make_vec2(float n) { return vec2_scalar{n, n}; }
+
+vec2_scalar make_vec2(float x, float y) { return vec2_scalar{x, y}; }
+
+vec2_scalar make_vec2(int32_t x, int32_t y) {
+  return vec2_scalar{float(x), float(y)};
+}
+
+template <typename N>
+vec2 make_vec2(const N& n) {
+  return vec2(n);
+}
+
+template <typename X, typename Y>
+vec2 make_vec2(const X& x, const Y& y) {
+  return vec2(x, y);
+}
+
+vec2 operator*(vec2_scalar a, Float b) { return vec2(a.x * b, a.y * b); }
+
+vec2 operator*(Float a, vec2_scalar b) { return vec2(a * b.x, a * b.y); }
+
+SI vec2 min(vec2 a, vec2 b) { return vec2(min(a.x, b.x), min(a.y, b.y)); }
+SI vec2 min(vec2 a, Float b) { return vec2(min(a.x, b), min(a.y, b)); }
+
+SI vec2_scalar min(vec2_scalar a, vec2_scalar b) {
+  return vec2_scalar{min(a.x, b.x), min(a.y, b.y)};
+}
+
+SI vec2 if_then_else(I32 c, vec2 t, vec2 e) {
+  return vec2(if_then_else(c, t.x, e.x), if_then_else(c, t.y, e.y));
+}
+
+SI vec2 if_then_else(int32_t c, vec2 t, vec2 e) { return c ? t : e; }
+
+vec2 step(vec2 edge, vec2 x) {
+  return vec2(step(edge.x, x.x), step(edge.y, x.y));
+}
+
+vec2_scalar step(vec2_scalar edge, vec2_scalar x) {
+  return vec2_scalar(step(edge.x, x.x), step(edge.y, x.y));
+}
+
+SI vec2 max(vec2 a, vec2 b) { return vec2(max(a.x, b.x), max(a.y, b.y)); }
+SI vec2 max(vec2 a, Float b) { return vec2(max(a.x, b), max(a.y, b)); }
+
+SI vec2_scalar max(vec2_scalar a, vec2_scalar b) {
+  return vec2_scalar{max(a.x, b.x), max(a.y, b.y)};
+}
+SI vec2_scalar max(vec2_scalar a, float b) {
+  return vec2_scalar{max(a.x, b), max(a.y, b)};
+}
+
+Float length(vec2 a) { return sqrt(a.x * a.x + a.y * a.y); }
+
+float length(vec2_scalar a) { return hypotf(a.x, a.y); }
+
+template <typename A, typename B>
+SI auto distance(A a, B b) {
+  return length(a - b);
+}
+
+template <typename T>
+SI T normalize(T a) {
+  return a / length(a);
+}
+
+SI vec2 sqrt(vec2 a) { return vec2(sqrt(a.x), sqrt(a.y)); }
+
+SI vec2_scalar sqrt(vec2_scalar a) { return vec2_scalar(sqrt(a.x), sqrt(a.y)); }
+
+SI vec2 recip(vec2 a) { return vec2(recip(a.x), recip(a.y)); }
+
+SI vec2_scalar recip(vec2_scalar a) {
+  return vec2_scalar(recip(a.x), recip(a.y));
+}
+
+SI vec2 inversesqrt(vec2 a) { return vec2(inversesqrt(a.x), inversesqrt(a.y)); }
+
+SI vec2_scalar inversesqrt(vec2_scalar a) {
+  return vec2_scalar(inversesqrt(a.x), inversesqrt(a.y));
+}
+
+#define abs __glsl_abs
+
+int32_t abs(int32_t a) { return a < 0 ? -a : a; }
+
+float abs(float a) { return fabsf(a); }
+
+Float abs(Float v) {
+#if USE_NEON
+  return vabsq_f32(v);
+#else
+  return bit_cast<Float>(bit_cast<I32>(v) & bit_cast<I32>(0.0f - v));
+#endif
+}
+
+float sign(float a) { return copysignf(1.0f, a); }
+
+Float sign(Float v) {
+  return bit_cast<Float>((bit_cast<I32>(v) & 0x80000000) |
+                         bit_cast<I32>(Float(1.0f)));
+}
+
+Float cast(U32 v) { return CONVERT((I32)v, Float); }
+Float cast(I32 v) { return CONVERT((I32)v, Float); }
+I32 cast(Float v) { return CONVERT(v, I32); }
+
+#define floor __glsl_floor
+
+float floor(float a) { return floorf(a); }
+
+Float floor(Float v) {
+  Float roundtrip = cast(cast(v));
+  return roundtrip - if_then(roundtrip > v, Float(1));
+}
+
+vec2 floor(vec2 v) { return vec2(floor(v.x), floor(v.y)); }
+
+vec2_scalar floor(vec2_scalar v) {
+  return vec2_scalar{floorf(v.x), floorf(v.y)};
+}
+
+#define ceil __glsl_ceil
+
+float ceil(float a) { return ceilf(a); }
+
+Float ceil(Float v) {
+  Float roundtrip = cast(cast(v));
+  return roundtrip + if_then(roundtrip < v, Float(1));
+}
+
+// Round to nearest even
+SI int32_t roundeven(float v, float scale) {
+#if USE_SSE2
+  return _mm_cvtss_si32(_mm_set_ss(v * scale));
+#else
+  return bit_cast<int32_t>(v * scale + float(0xC00000)) - 0x4B400000;
+#endif
+}
+
+SI I32 roundeven(Float v, Float scale) {
+#if USE_SSE2
+  return _mm_cvtps_epi32(v * scale);
+#else
+  // Magic number implementation of round-to-nearest-even
+  // see http://stereopsis.com/sree/fpu2006.html
+  return bit_cast<I32>(v * scale + Float(0xC00000)) - 0x4B400000;
+#endif
+}
+
+// Round towards zero
+SI int32_t roundzero(float v, float scale) { return int32_t(v * scale); }
+
+SI I32 roundzero(Float v, Float scale) { return cast(v * scale); }
+
+// Round whichever direction is fastest for positive numbers
+SI I32 roundfast(Float v, Float scale) {
+#if USE_SSE2
+  return _mm_cvtps_epi32(v * scale);
+#else
+  return cast(v * scale + 0.5f);
+#endif
+}
+
+template <typename T>
+SI auto round_pixel(T v, float scale = 255.0f) {
+  return roundfast(v, scale);
+}
+
+#define round __glsl_round
+
+float round(float a) { return roundf(a); }
+
+Float round(Float v) { return floor(v + 0.5f); }
+
+float fract(float a) { return a - floor(a); }
+
+Float fract(Float v) { return v - floor(v); }
+
+vec2 fract(vec2 v) { return vec2(fract(v.x), fract(v.y)); }
+
+// X derivatives can be approximated by dFdx(x) = x[1] - x[0].
+// Y derivatives are not easily available since we operate in terms of X spans
+// only. To work around, assume dFdy(p.x) = dFdx(p.y), which only holds for
+// uniform scaling, and thus abs(dFdx(p.x)) + abs(dFdy(p.x)) = abs(dFdx(p.x)) +
+// abs(dFdx(p.y)) which mirrors abs(dFdx(p.y)) + abs(dFdy(p.y)) = abs(dFdx(p.y))
+// + abs(dFdx(p.x)).
+vec2_scalar fwidth(vec2 p) {
+  Float d = abs(SHUFFLE(p.x, p.y, 1, 1, 5, 5) - SHUFFLE(p.x, p.y, 0, 0, 4, 4));
+  return vec2_scalar(d.x + d.z);
+}
+
+float dFdx(Float x) { return x.y - x.x; }
+
+vec2_scalar dFdx(vec2 p) { return vec2_scalar(dFdx(p.x), dFdx(p.y)); }
+
+// See
+// http://www.machinedlearnings.com/2011/06/fast-approximate-logarithm-exponential.html.
+Float approx_log2(Float x) {
+  // e - 127 is a fair approximation of log2(x) in its own right...
+  Float e = cast(bit_cast<U32>(x)) * (1.0f / (1 << 23));
+
+  // ... but using the mantissa to refine its error is _much_ better.
+  Float m = bit_cast<Float>((bit_cast<U32>(x) & 0x007fffff) | 0x3f000000);
+  return e - 124.225514990f - 1.498030302f * m -
+         1.725879990f / (0.3520887068f + m);
+}
+
+Float approx_pow2(Float x) {
+  Float f = fract(x);
+  return bit_cast<Float>(
+      roundfast(1.0f * (1 << 23), x + 121.274057500f - 1.490129070f * f +
+                                      27.728023300f / (4.84252568f - f)));
+}
+
+#define pow __glsl_pow
+
+SI float pow(float x, float y) { return powf(x, y); }
+
+Float pow(Float x, Float y) {
+  return if_then_else((x == 0) | (x == 1), x, approx_pow2(approx_log2(x) * y));
+}
+
+#define exp __glsl_exp
+
+SI float exp(float x) { return expf(x); }
+
+Float exp(Float y) {
+  float l2e = 1.4426950408889634074f;
+  return approx_pow2(l2e * y);
+}
+
+#define exp2 __glsl_exp2
+
+SI float exp2(float x) { return exp2f(x); }
+
+Float exp2(Float x) { return approx_pow2(x); }
+
+#define log __glsl_log
+
+SI float log(float x) { return logf(x); }
+
+Float log(Float x) { return approx_log2(x) * 0.69314718f; }
+
+#define log2 __glsl_log2
+
+SI float log2(float x) { return log2f(x); }
+
+Float log2(Float x) { return approx_log2(x); }
+
+struct ivec4;
+
+struct ivec2_scalar {
+  typedef int32_t element_type;
+
+  int32_t x;
+  int32_t y;
+
+  ivec2_scalar() : ivec2_scalar(0) {}
+  IMPLICIT constexpr ivec2_scalar(int32_t a) : x(a), y(a) {}
+  constexpr ivec2_scalar(int32_t x, int32_t y) : x(x), y(y) {}
+
+  int32_t& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      default:
+        UNREACHABLE;
+    }
+  }
+  int32_t& sel(XYZW c1) { return select(c1); }
+  ivec2_scalar sel(XYZW c1, XYZW c2) {
+    return ivec2_scalar{select(c1), select(c2)};
+  }
+
+  ivec2_scalar operator-() const { return ivec2_scalar{-x, -y}; }
+
+  ivec2_scalar& operator+=(ivec2_scalar a) {
+    x += a.x;
+    y += a.y;
+    return *this;
+  }
+  ivec2_scalar& operator+=(int n) {
+    x += n;
+    y += n;
+    return *this;
+  }
+
+  ivec2_scalar& operator>>=(int shift) {
+    x >>= shift;
+    y >>= shift;
+    return *this;
+  }
+
+  friend ivec2_scalar operator&(ivec2_scalar a, int b) {
+    return ivec2_scalar{a.x & b, a.y & b};
+  }
+
+  friend ivec2_scalar operator+(ivec2_scalar a, ivec2_scalar b) {
+    return ivec2_scalar{a.x + b.x, a.y + b.y};
+  }
+  friend ivec2_scalar operator+(ivec2_scalar a, int b) {
+    return ivec2_scalar{a.x + b, a.y + b};
+  }
+
+  friend ivec2_scalar operator-(ivec2_scalar a, ivec2_scalar b) {
+    return ivec2_scalar{a.x - b.x, a.y - b.y};
+  }
+  friend ivec2_scalar operator-(ivec2_scalar a, int b) {
+    return ivec2_scalar{a.x - b, a.y - b};
+  }
+
+  friend bool operator==(const ivec2_scalar& l, const ivec2_scalar& r) {
+    return l.x == r.x && l.y == r.y;
+  }
+};
+
+struct ivec2 {
+  typedef int32_t element_type;
+
+  ivec2() : ivec2(I32(0)) {}
+  IMPLICIT ivec2(I32 a) : x(a), y(a) {}
+  ivec2(I32 x, I32 y) : x(x), y(y) {}
+  IMPLICIT ivec2(vec2 a) : x(cast(a.x)), y(cast(a.y)) {}
+  ivec2(U32 x, U32 y) : x(CONVERT(x, I32)), y(CONVERT(y, I32)) {}
+  IMPLICIT constexpr ivec2(ivec2_scalar s) : x(s.x), y(s.y) {}
+  constexpr ivec2(ivec2_scalar s0, ivec2_scalar s1, ivec2_scalar s2,
+                  ivec2_scalar s3)
+      : x(I32{s0.x, s1.x, s2.x, s3.x}), y(I32{s0.y, s1.y, s2.y, s3.y}) {}
+  I32 x;
+  I32 y;
+
+  I32& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      default:
+        UNREACHABLE;
+    }
+  }
+  I32& sel(XYZW c1) { return select(c1); }
+
+  ivec2 sel(XYZW c1, XYZW c2) { return ivec2(select(c1), select(c2)); }
+
+  ivec4 sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4);
+
+  ivec2& operator*=(I32 a) {
+    x *= a;
+    y *= a;
+    return *this;
+  }
+  ivec2& operator+=(ivec2 a) {
+    x += a.x;
+    y += a.y;
+    return *this;
+  }
+  ivec2& operator>>=(int shift) {
+    x >>= shift;
+    y >>= shift;
+    return *this;
+  }
+
+  friend ivec2 operator*(ivec2 a, I32 b) { return ivec2(a.x * b, a.y * b); }
+  friend ivec2 operator&(ivec2 a, ivec2 b) {
+    return ivec2(a.x & b.x, a.y & b.y);
+  }
+  friend ivec2 operator&(ivec2 a, I32 b) { return ivec2(a.x & b, a.y & b); }
+  friend ivec2 operator+(ivec2 a, ivec2 b) {
+    return ivec2(a.x + b.x, a.y + b.y);
+  }
+};
+
+vec2::vec2(ivec2 a) : x(cast(a.x)), y(cast(a.y)) {}
+
+ivec2_scalar make_ivec2(int32_t n) { return ivec2_scalar{n, n}; }
+
+ivec2_scalar make_ivec2(uint32_t n) {
+  return ivec2_scalar{int32_t(n), int32_t(n)};
+}
+
+ivec2_scalar make_ivec2(int32_t x, int32_t y) { return ivec2_scalar{x, y}; }
+
+ivec2_scalar make_ivec2(uint32_t x, uint32_t y) {
+  return ivec2_scalar{int32_t(x), int32_t(y)};
+}
+
+vec2_scalar make_vec2(const ivec2_scalar& v) {
+  return vec2_scalar{float(v.x), float(v.y)};
+}
+
+ivec2_scalar make_ivec2(const vec2_scalar& v) {
+  return ivec2_scalar{int32_t(v.x), int32_t(v.y)};
+}
+
+template <typename N>
+ivec2 make_ivec2(const N& n) {
+  return ivec2(n);
+}
+
+template <typename X, typename Y>
+ivec2 make_ivec2(const X& x, const Y& y) {
+  return ivec2(x, y);
+}
+
+ivec2_scalar force_scalar(const ivec2& v) {
+  return ivec2_scalar{force_scalar(v.x), force_scalar(v.y)};
+}
+
+struct ivec3_scalar {
+  int32_t x;
+  int32_t y;
+  int32_t z;
+
+  ivec3_scalar() : ivec3_scalar(0) {}
+  IMPLICIT constexpr ivec3_scalar(int32_t a) : x(a), y(a), z(a) {}
+  constexpr ivec3_scalar(int32_t x, int32_t y, int32_t z) : x(x), y(y), z(z) {}
+
+  int32_t& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      default:
+        UNREACHABLE;
+    }
+  }
+  int32_t& sel(XYZW c1) { return select(c1); }
+  ivec2_scalar sel(XYZW c1, XYZW c2) {
+    return ivec2_scalar{select(c1), select(c2)};
+  }
+};
+
+struct ivec3 {
+  ivec3() : ivec3(0) {}
+  IMPLICIT ivec3(I32 a) : x(a), y(a), z(a) {}
+  ivec3(I32 x, I32 y, I32 z) : x(x), y(y), z(z) {}
+  ivec3(ivec2 a, I32 b) : x(a.x), y(a.y), z(b) {}
+  ivec3(vec2 a, Float b) : x(cast(a.x)), y(cast(a.y)), z(cast(b)) {}
+  I32 x;
+  I32 y;
+  I32 z;
+
+  friend ivec3 operator+(ivec3 a, ivec3 b) {
+    return ivec3(a.x + b.x, a.y + b.y, a.z + b.z);
+  }
+};
+
+vec2_scalar make_vec2(ivec3_scalar s) {
+  return vec2_scalar{float(s.x), float(s.y)};
+}
+
+ivec3_scalar make_ivec3(int32_t n) { return ivec3_scalar{n, n, n}; }
+
+ivec3_scalar make_ivec3(const ivec2_scalar& v, int32_t z) {
+  return ivec3_scalar{v.x, v.y, z};
+}
+
+ivec3_scalar make_ivec3(int32_t x, int32_t y, int32_t z) {
+  return ivec3_scalar{x, y, z};
+}
+
+template <typename N>
+ivec3 make_ivec3(const N& n) {
+  return ivec3(n);
+}
+
+template <typename X, typename Y>
+ivec3 make_ivec3(const X& x, const Y& y) {
+  return ivec3(x, y);
+}
+
+template <typename X, typename Y, typename Z>
+ivec3 make_ivec3(const X& x, const Y& y, const Z& z) {
+  return ivec3(x, y, z);
+}
+
+struct ivec4_scalar {
+  typedef int32_t element_type;
+
+  int32_t x;
+  int32_t y;
+  int32_t z;
+  int32_t w;
+
+  ivec4_scalar() : ivec4_scalar(0) {}
+  IMPLICIT constexpr ivec4_scalar(int32_t a) : x(a), y(a), z(a), w(a) {}
+  constexpr ivec4_scalar(int32_t x, int32_t y, int32_t z, int32_t w)
+      : x(x), y(y), z(z), w(w) {}
+
+  int32_t& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      case W:
+        return w;
+      default:
+        UNREACHABLE;
+    }
+  }
+  int32_t& sel(XYZW c1) { return select(c1); }
+  ivec2_scalar sel(XYZW c1, XYZW c2) {
+    return ivec2_scalar{select(c1), select(c2)};
+  }
+
+  friend ivec4_scalar operator&(int32_t a, ivec4_scalar b) {
+    return ivec4_scalar{a & b.x, a & b.y, a & b.z, a & b.w};
+  }
+  friend ivec4_scalar operator<<(ivec4_scalar a, int32_t b) {
+    return ivec4_scalar{a.x << b, a.y << b, a.z << b, a.w << b};
+  }
+
+  int32_t& operator[](int index) {
+    switch (index) {
+      case 0:
+        return x;
+      case 1:
+        return y;
+      case 2:
+        return z;
+      case 3:
+        return w;
+      default:
+        UNREACHABLE;
+    }
+  }
+};
+
+struct ivec4 {
+  typedef int32_t element_type;
+
+  ivec4() : ivec4(I32(0)) {}
+  IMPLICIT ivec4(I32 a) : x(a), y(a), z(a), w(a) {}
+  ivec4(I32 x, I32 y, I32 z, I32 w) : x(x), y(y), z(z), w(w) {}
+  ivec4(ivec2 a, I32 b, I32 c) : x(a.x), y(a.y), z(b), w(c) {}
+  IMPLICIT constexpr ivec4(ivec4_scalar s) : x(s.x), y(s.y), z(s.z), w(s.w) {}
+  constexpr ivec4(ivec4_scalar s0, ivec4_scalar s1, ivec4_scalar s2,
+                  ivec4_scalar s3)
+      : x(I32{s0.x, s1.x, s2.x, s3.x}),
+        y(I32{s0.y, s1.y, s2.y, s3.y}),
+        z(I32{s0.z, s1.z, s2.z, s3.z}),
+        w(I32{s0.w, s1.w, s2.w, s3.w}) {}
+
+  I32& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      case W:
+        return w;
+      default:
+        UNREACHABLE;
+    }
+  }
+  I32 sel(XYZW c1) { return select(c1); }
+
+  ivec2 sel(XYZW c1, XYZW c2) { return ivec2(select(c1), select(c2)); }
+
+  ivec3 sel(XYZW c1, XYZW c2, XYZW c3) {
+    return ivec3(select(c1), select(c2), select(c3));
+  }
+
+  friend ivec4 operator&(I32 a, ivec4 b) {
+    return ivec4(a & b.x, a & b.y, a & b.z, a & b.w);
+  }
+
+  I32 x;
+  I32 y;
+  I32 z;
+  I32 w;
+};
+
+ivec4_scalar force_scalar(const ivec4& v) {
+  return ivec4_scalar{force_scalar(v.x), force_scalar(v.y), force_scalar(v.z),
+                      force_scalar(v.w)};
+}
+
+ivec4_scalar make_ivec4(int32_t n) { return ivec4_scalar{n, n, n, n}; }
+
+ivec4_scalar make_ivec4(const ivec2_scalar& xy, int32_t z, int32_t w) {
+  return ivec4_scalar{xy.x, xy.y, z, w};
+}
+
+ivec4_scalar make_ivec4(int32_t x, int32_t y, int32_t z, int32_t w) {
+  return ivec4_scalar{x, y, z, w};
+}
+
+template <typename N>
+ivec4 make_ivec4(const N& n) {
+  return ivec4(n);
+}
+
+template <typename X, typename Y, typename Z>
+ivec4 make_ivec4(const X& x, const Y& y, const Z& z) {
+  return ivec4(x, y, z);
+}
+
+template <typename X, typename Y, typename Z, typename W>
+ivec4 make_ivec4(const X& x, const Y& y, const Z& z, const W& w) {
+  return ivec4(x, y, z, w);
+}
+
+SI ivec2 if_then_else(I32 c, ivec2 t, ivec2 e) {
+  return ivec2(if_then_else(c, t.x, e.x), if_then_else(c, t.y, e.y));
+}
+
+SI ivec2 if_then_else(int32_t c, ivec2 t, ivec2 e) { return c ? t : e; }
+
+SI ivec4 if_then_else(I32 c, ivec4 t, ivec4 e) {
+  return ivec4(if_then_else(c, t.x, e.x), if_then_else(c, t.y, e.y),
+               if_then_else(c, t.z, e.z), if_then_else(c, t.w, e.w));
+}
+
+SI ivec4 if_then_else(int32_t c, ivec4 t, ivec4 e) { return c ? t : e; }
+
+ivec4 operator&(I32 a, ivec4_scalar b) {
+  return ivec4(a & b.x, a & b.y, a & b.z, a & b.w);
+}
+
+struct bvec3_scalar {
+  bool x;
+  bool y;
+  bool z;
+
+  bvec3_scalar() : bvec3_scalar(false) {}
+  IMPLICIT constexpr bvec3_scalar(bool a) : x(a), y(a), z(a) {}
+  constexpr bvec3_scalar(bool x, bool y, bool z) : x(x), y(y), z(z) {}
+};
+
+struct bvec3_scalar1 {
+  bool x;
+
+  IMPLICIT constexpr bvec3_scalar1(bool a) : x(a) {}
+
+  operator bvec3_scalar() const { return bvec3_scalar(x); }
+};
+
+struct bvec3 {
+  bvec3() : bvec3(0) {}
+  IMPLICIT bvec3(Bool a) : x(a), y(a), z(a) {}
+  bvec3(Bool x, Bool y, Bool z) : x(x), y(y), z(z) {}
+  Bool& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      default:
+        UNREACHABLE;
+    }
+  }
+  Bool sel(XYZW c1) { return select(c1); }
+
+  Bool x;
+  Bool y;
+  Bool z;
+};
+
+bvec3_scalar1 make_bvec3(bool n) { return bvec3_scalar1(n); }
+
+struct bvec4_scalar {
+  bool x;
+  bool y;
+  bool z;
+  bool w;
+
+  bvec4_scalar() : bvec4_scalar(false) {}
+  IMPLICIT constexpr bvec4_scalar(bool a) : x(a), y(a), z(a), w(a) {}
+  constexpr bvec4_scalar(bool x, bool y, bool z, bool w)
+      : x(x), y(y), z(z), w(w) {}
+
+  bool& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      case W:
+        return w;
+      default:
+        UNREACHABLE;
+    }
+  }
+  bool sel(XYZW c1) { return select(c1); }
+  bvec2_scalar sel(XYZW c1, XYZW c2) {
+    return bvec2_scalar(select(c1), select(c2));
+  }
+};
+
+bvec4_scalar bvec2_scalar::sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4) {
+  return bvec4_scalar{select(c1), select(c2), select(c3), select(c4)};
+}
+
+struct bvec4_scalar1 {
+  bool x;
+
+  IMPLICIT constexpr bvec4_scalar1(bool a) : x(a) {}
+
+  operator bvec4_scalar() const { return bvec4_scalar(x); }
+};
+
+struct bvec4 {
+  bvec4() : bvec4(0) {}
+  IMPLICIT bvec4(Bool a) : x(a), y(a), z(a), w(a) {}
+  bvec4(Bool x, Bool y, Bool z, Bool w) : x(x), y(y), z(z), w(w) {}
+  bvec4(bvec2 x, bvec2 y) : x(x.x), y(x.y), z(y.x), w(y.y) {}
+  Bool& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      case W:
+        return w;
+      default:
+        UNREACHABLE;
+    }
+  }
+  Bool sel(XYZW c1) { return select(c1); }
+
+  Bool x;
+  Bool y;
+  Bool z;
+  Bool w;
+};
+
+bvec4_scalar1 make_bvec4(bool n) { return bvec4_scalar1(n); }
+
+bvec4_scalar make_bvec4(bool x, bool y, bool z, bool w) {
+  return bvec4_scalar{x, y, z, w};
+}
+
+bvec4_scalar make_bvec4(bvec2_scalar a, bvec2_scalar b) {
+  return bvec4_scalar{a.x, a.y, b.x, b.y};
+}
+
+template <typename N>
+bvec4 make_bvec4(const N& n) {
+  return bvec4(n);
+}
+
+template <typename X, typename Y>
+bvec4 make_bvec4(const X& x, const Y& y) {
+  return bvec4(x, y);
+}
+
+template <typename X, typename Y, typename Z, typename W>
+bvec4 make_bvec4(const X& x, const Y& y, const Z& z, const W& w) {
+  return bvec4(x, y, z, w);
+}
+
+struct vec2_ref {
+  vec2_ref(Float& x, Float& y) : x(x), y(y) {}
+  Float& x;
+  Float& y;
+
+  Float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      default:
+        UNREACHABLE;
+    }
+  }
+  Float& sel(XYZW c1) { return select(c1); }
+
+  vec2_ref& operator=(const vec2& a) {
+    x = a.x;
+    y = a.y;
+    return *this;
+  }
+
+  vec2_ref& operator/=(Float a) {
+    x /= a;
+    y /= a;
+    return *this;
+  }
+
+  vec2_ref& operator/=(vec2 a) {
+    x /= a.x;
+    y /= a.y;
+    return *this;
+  }
+
+  vec2_ref& operator+=(vec2 a) {
+    x += a.x;
+    y += a.y;
+    return *this;
+  }
+  vec2_ref& operator-=(vec2 a) {
+    x -= a.x;
+    y -= a.y;
+    return *this;
+  }
+  vec2_ref& operator*=(vec2 a) {
+    x *= a.x;
+    y *= a.y;
+    return *this;
+  }
+};
+
+struct vec3_scalar {
+  typedef struct vec3 vector_type;
+  typedef float element_type;
+
+  float x;
+  float y;
+  float z;
+
+  constexpr vec3_scalar() : vec3_scalar(0.0f) {}
+  IMPLICIT constexpr vec3_scalar(float a) : x(a), y(a), z(a) {}
+  constexpr vec3_scalar(float x, float y, float z) : x(x), y(y), z(z) {}
+
+  float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      default:
+        UNREACHABLE;
+    }
+  }
+  float& sel(XYZW c1) { return select(c1); }
+  vec2_scalar sel(XYZW c1, XYZW c2) {
+    return vec2_scalar(select(c1), select(c2));
+  }
+  vec3_scalar sel(XYZW c1, XYZW c2, XYZW c3) {
+    return vec3_scalar(select(c1), select(c2), select(c3));
+  }
+  vec2_scalar_ref lsel(XYZW c1, XYZW c2) {
+    return vec2_scalar_ref(select(c1), select(c2));
+  }
+
+  friend vec3_scalar operator*(vec3_scalar a, vec3_scalar b) {
+    return vec3_scalar{a.x * b.x, a.y * b.y, a.z * b.z};
+  }
+  friend vec3_scalar operator*(vec3_scalar a, float b) {
+    return vec3_scalar{a.x * b, a.y * b, a.z * b};
+  }
+
+  friend vec3_scalar operator-(vec3_scalar a, vec3_scalar b) {
+    return vec3_scalar{a.x - b.x, a.y - b.y, a.z - b.z};
+  }
+  friend vec3_scalar operator-(vec3_scalar a, float b) {
+    return vec3_scalar{a.x - b, a.y - b, a.z - b};
+  }
+  friend vec3_scalar operator+(vec3_scalar a, vec3_scalar b) {
+    return vec3_scalar{a.x + b.x, a.y + b.y, a.z + b.z};
+  }
+  friend vec3_scalar operator+(vec3_scalar a, float b) {
+    return vec3_scalar{a.x + b, a.y + b, a.z + b};
+  }
+
+  friend vec3_scalar operator/(vec3_scalar a, vec3_scalar b) {
+    return vec3_scalar{a.x / b.x, a.y / b.y, a.z / b.z};
+  }
+  friend vec3_scalar operator/(vec3_scalar a, float b) {
+    return vec3_scalar{a.x / b, a.y / b, a.z / b};
+  }
+
+  vec3_scalar operator+=(vec3_scalar a) {
+    x += a.x;
+    y += a.y;
+    z += a.z;
+    return *this;
+  }
+
+  friend bool operator==(const vec3_scalar& l, const vec3_scalar& r) {
+    return l.x == r.x && l.y == r.y && l.z == r.z;
+  }
+};
+
+struct vec3_scalar_ref {
+  vec3_scalar_ref(float& x, float& y, float& z) : x(x), y(y), z(z) {}
+  float& x;
+  float& y;
+  float& z;
+
+  float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      default:
+        UNREACHABLE;
+    }
+  }
+  float& sel(XYZW c1) { return select(c1); }
+
+  vec3_scalar_ref& operator=(const vec3_scalar& a) {
+    x = a.x;
+    y = a.y;
+    z = a.z;
+    return *this;
+  }
+
+  operator vec3_scalar() const { return vec3_scalar{x, y, z}; }
+};
+
+struct vec3 {
+  typedef struct vec3 vector_type;
+  typedef float element_type;
+
+  constexpr vec3() : vec3(Float(0.0f)) {}
+  IMPLICIT constexpr vec3(Float a) : x(a), y(a), z(a) {}
+  constexpr vec3(Float x, Float y, Float z) : x(x), y(y), z(z) {}
+  vec3(vec2 a, Float z) : x(a.x), y(a.y), z(z) {}
+  explicit vec3(vec4);
+  IMPLICIT constexpr vec3(vec3_scalar s) : x(s.x), y(s.y), z(s.z) {}
+  constexpr vec3(vec3_scalar s0, vec3_scalar s1, vec3_scalar s2, vec3_scalar s3)
+      : x(Float{s0.x, s1.x, s2.x, s3.x}),
+        y(Float{s0.y, s1.y, s2.y, s3.y}),
+        z(Float{s0.z, s1.z, s2.z, s3.z}) {}
+  Float x;
+  Float y;
+  Float z;
+
+  Float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      default:
+        UNREACHABLE;
+    }
+  }
+  Float& sel(XYZW c1) { return select(c1); }
+
+  vec2 sel(XYZW c1, XYZW c2) { return vec2(select(c1), select(c2)); }
+
+  vec3 sel(XYZW c1, XYZW c2, XYZW c3) {
+    return vec3(select(c1), select(c2), select(c3));
+  }
+
+  vec4 sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4);
+
+  vec2_ref lsel(XYZW c1, XYZW c2) { return vec2_ref(select(c1), select(c2)); }
+
+  friend vec3 operator*(vec3 a, Float b) {
+    return vec3(a.x * b, a.y * b, a.z * b);
+  }
+  friend vec3 operator*(vec3 a, vec3 b) {
+    return vec3(a.x * b.x, a.y * b.y, a.z * b.z);
+  }
+  friend vec3 operator*(Float a, vec3 b) {
+    return vec3(a * b.x, a * b.y, a * b.z);
+  }
+
+  friend vec3 operator/(vec3 a, Float b) {
+    return vec3(a.x / b, a.y / b, a.z / b);
+  }
+  friend vec3 operator/(vec3 a, vec3 b) {
+    return vec3(a.x / b.x, a.y / b.y, a.z / b.z);
+  }
+
+  friend I32 operator==(const vec3& l, const vec3& r) {
+    return l.x == r.x && l.y == r.y && l.z == r.z;
+  }
+
+  friend vec3 operator-(vec3 a, Float b) {
+    return vec3(a.x - b, a.y - b, a.z - b);
+  }
+  friend vec3 operator-(vec3 a, vec3 b) {
+    return vec3(a.x - b.x, a.y - b.y, a.z - b.z);
+  }
+  friend vec3 operator+(vec3 a, Float b) {
+    return vec3(a.x + b, a.y + b, a.z + b);
+  }
+  friend vec3 operator+(vec3 a, vec3 b) {
+    return vec3(a.x + b.x, a.y + b.y, a.z + b.z);
+  }
+
+  vec3 operator+=(vec3_scalar a) {
+    x += a.x;
+    y += a.y;
+    z += a.z;
+    return *this;
+  }
+  vec3& operator+=(vec3 a) {
+    x += a.x;
+    y += a.y;
+    z += a.z;
+    return *this;
+  }
+};
+
+vec3_scalar force_scalar(const vec3& v) {
+  return vec3_scalar{force_scalar(v.x), force_scalar(v.y), force_scalar(v.z)};
+}
+
+vec3_scalar make_vec3(float n) { return vec3_scalar{n, n, n}; }
+
+vec3_scalar make_vec3(const vec2_scalar& v, float z) {
+  return vec3_scalar{v.x, v.y, z};
+}
+
+vec3_scalar make_vec3(float x, float y, float z) {
+  return vec3_scalar{x, y, z};
+}
+
+vec3_scalar make_vec3(int32_t x, int32_t y, float z) {
+  return vec3_scalar{float(x), float(y), z};
+}
+
+template <typename N>
+vec3 make_vec3(const N& n) {
+  return vec3(n);
+}
+
+template <typename X, typename Y>
+vec3 make_vec3(const X& x, const Y& y) {
+  return vec3(x, y);
+}
+
+template <typename X, typename Y, typename Z>
+vec3 make_vec3(const X& x, const Y& y, const Z& z) {
+  return vec3(x, y, z);
+}
+
+SI vec3 if_then_else(I32 c, vec3 t, vec3 e) {
+  return vec3(if_then_else(c, t.x, e.x), if_then_else(c, t.y, e.y),
+              if_then_else(c, t.z, e.z));
+}
+
+SI vec3 if_then_else(int32_t c, vec3 t, vec3 e) { return c ? t : e; }
+
+SI vec3 if_then_else(ivec3 c, vec3 t, vec3 e) {
+  return vec3(if_then_else(c.x, t.x, e.x), if_then_else(c.y, t.y, e.y),
+              if_then_else(c.z, t.z, e.z));
+}
+
+vec3 step(vec3 edge, vec3 x) {
+  return vec3(step(edge.x, x.x), step(edge.y, x.y), step(edge.z, x.z));
+}
+
+vec3_scalar step(vec3_scalar edge, vec3_scalar x) {
+  return vec3_scalar(step(edge.x, x.x), step(edge.y, x.y), step(edge.z, x.z));
+}
+
+SI vec3 min(vec3 a, vec3 b) {
+  return vec3(min(a.x, b.x), min(a.y, b.y), min(a.z, b.z));
+}
+SI vec3 min(vec3 a, Float b) {
+  return vec3(min(a.x, b), min(a.y, b), min(a.z, b));
+}
+SI vec3_scalar min(vec3_scalar a, vec3_scalar b) {
+  return vec3_scalar{min(a.x, b.x), min(a.y, b.y), min(a.z, b.z)};
+}
+
+SI vec3 max(vec3 a, vec3 b) {
+  return vec3(max(a.x, b.x), max(a.y, b.y), max(a.z, b.z));
+}
+SI vec3 max(vec3 a, Float b) {
+  return vec3(max(a.x, b), max(a.y, b), max(a.z, b));
+}
+SI vec3_scalar max(vec3_scalar a, vec3_scalar b) {
+  return vec3_scalar{max(a.x, b.x), max(a.y, b.y), max(a.z, b.z)};
+}
+
+vec3 pow(vec3 x, vec3 y) {
+  return vec3(pow(x.x, y.x), pow(x.y, y.y), pow(x.z, y.z));
+}
+
+struct vec3_ref {
+  vec3_ref(Float& x, Float& y, Float& z) : x(x), y(y), z(z) {}
+  Float& x;
+  Float& y;
+  Float& z;
+  vec3_ref& operator=(const vec3& a) {
+    x = a.x;
+    y = a.y;
+    z = a.z;
+    return *this;
+  }
+
+  vec3_ref& operator/=(Float a) {
+    x /= a;
+    y /= a;
+    z /= a;
+    return *this;
+  }
+
+  vec3_ref& operator*=(Float a) {
+    x *= a;
+    y *= a;
+    z *= a;
+    return *this;
+  }
+};
+
+struct vec4_scalar {
+  typedef struct vec4 vector_type;
+  typedef float element_type;
+
+  float x;
+  float y;
+  float z;
+  float w;
+
+  constexpr vec4_scalar() : vec4_scalar(0.0f) {}
+  IMPLICIT constexpr vec4_scalar(float a) : x(a), y(a), z(a), w(a) {}
+  constexpr vec4_scalar(float x, float y, float z, float w)
+      : x(x), y(y), z(z), w(w) {}
+  vec4_scalar(vec3_scalar xyz, float w) : x(xyz.x), y(xyz.y), z(xyz.z), w(w) {}
+
+  static vec4_scalar load_from_ptr(const float* f) {
+    return vec4_scalar(f[0], f[1], f[2], f[3]);
+  }
+
+  ALWAYS_INLINE float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      case W:
+        return w;
+      default:
+        UNREACHABLE;
+    }
+  }
+  float& sel(XYZW c1) { return select(c1); }
+  vec2_scalar sel(XYZW c1, XYZW c2) {
+    return vec2_scalar{select(c1), select(c2)};
+  }
+  vec3_scalar sel(XYZW c1, XYZW c2, XYZW c3) {
+    return vec3_scalar{select(c1), select(c2), select(c3)};
+  }
+  vec4_scalar sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4) {
+    return vec4_scalar{select(c1), select(c2), select(c3), select(c4)};
+  }
+  vec2_scalar_ref lsel(XYZW c1, XYZW c2) {
+    return vec2_scalar_ref(select(c1), select(c2));
+  }
+  vec3_scalar_ref lsel(XYZW c1, XYZW c2, XYZW c3) {
+    return vec3_scalar_ref(select(c1), select(c2), select(c3));
+  }
+
+  friend vec4_scalar operator*(vec4_scalar a, vec4_scalar b) {
+    return vec4_scalar{a.x * b.x, a.y * b.y, a.z * b.z, a.w * b.w};
+  }
+  friend vec4_scalar operator*(vec4_scalar a, float b) {
+    return vec4_scalar{a.x * b, a.y * b, a.z * b, a.w * b};
+  }
+  friend vec4_scalar operator*(float a, vec4_scalar b) {
+    return vec4_scalar{a * b.x, a * b.y, a * b.z, a * b.w};
+  }
+  vec4_scalar& operator*=(float a) {
+    x *= a;
+    y *= a;
+    z *= a;
+    w *= a;
+    return *this;
+  }
+
+  friend vec4_scalar operator-(vec4_scalar a, vec4_scalar b) {
+    return vec4_scalar{a.x - b.x, a.y - b.y, a.z - b.z, a.w - b.w};
+  }
+  friend vec4_scalar operator-(vec4_scalar a, float b) {
+    return vec4_scalar{a.x - b, a.y - b, a.z - b, a.w - b};
+  }
+  friend vec4_scalar operator+(vec4_scalar a, vec4_scalar b) {
+    return vec4_scalar{a.x + b.x, a.y + b.y, a.z + b.z, a.w + b.w};
+  }
+  friend vec4_scalar operator+(vec4_scalar a, float b) {
+    return vec4_scalar{a.x + b, a.y + b, a.z + b, a.w + b};
+  }
+
+  friend vec4_scalar operator/(vec4_scalar a, vec4_scalar b) {
+    return vec4_scalar{a.x / b.x, a.y / b.y, a.z / b.z, a.w / b.w};
+  }
+  friend vec4_scalar operator/(vec4_scalar a, float b) {
+    return vec4_scalar{a.x / b, a.y / b, a.z / b, a.w / b};
+  }
+
+  vec4_scalar& operator+=(vec4_scalar a) {
+    x += a.x;
+    y += a.y;
+    z += a.z;
+    w += a.w;
+    return *this;
+  }
+
+  vec4_scalar& operator/=(vec4_scalar a) {
+    x /= a.x;
+    y /= a.y;
+    z /= a.z;
+    w /= a.w;
+    return *this;
+  }
+
+  vec4_scalar& operator*=(vec4_scalar a) {
+    x *= a.x;
+    y *= a.y;
+    z *= a.z;
+    w *= a.w;
+    return *this;
+  }
+
+  friend bool operator==(const vec4_scalar& l, const vec4_scalar& r) {
+    return l.x == r.x && l.y == r.y && l.z == r.z && l.w == r.w;
+  }
+
+  friend bool operator!=(const vec4_scalar& l, const vec4_scalar& r) {
+    return l.x != r.x || l.y != r.y || l.z != r.z || l.w != r.w;
+  }
+};
+
+vec3_scalar vec2_scalar::sel(XYZW c1, XYZW c2, XYZW c3) {
+  return {select(c1), select(c2), select(c3)};
+}
+vec4_scalar vec2_scalar::sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4) {
+  return vec4_scalar{select(c1), select(c2), select(c3), select(c4)};
+}
+
+struct vec4_ref {
+  vec4_ref(Float& x, Float& y, Float& z, Float& w) : x(x), y(y), z(z), w(w) {}
+  Float& x;
+  Float& y;
+  Float& z;
+  Float& w;
+
+  vec4_ref& operator=(const vec4& a);
+};
+
+struct vec4 {
+  typedef struct vec4 vector_type;
+  typedef float element_type;
+
+  constexpr vec4() : vec4(Float(0.0f)) {}
+  IMPLICIT constexpr vec4(Float a) : x(a), y(a), z(a), w(a) {}
+  vec4(Float x, Float y, Float z, Float w) : x(x), y(y), z(z), w(w) {}
+  vec4(vec3 xyz, Float w) : x(xyz.x), y(xyz.y), z(xyz.z), w(w) {}
+  vec4(vec2 xy, vec2 zw) : x(xy.x), y(xy.y), z(zw.x), w(zw.y) {}
+  vec4(vec2 xy, Float z, Float w) : x(xy.x), y(xy.y), z(z), w(w) {}
+  vec4(Float x, Float y, vec2 zw) : x(x), y(y), z(zw.x), w(zw.y) {}
+  IMPLICIT constexpr vec4(vec4_scalar s) : x(s.x), y(s.y), z(s.z), w(s.w) {}
+  constexpr vec4(vec4_scalar s0, vec4_scalar s1, vec4_scalar s2, vec4_scalar s3)
+      : x(Float{s0.x, s1.x, s2.x, s3.x}),
+        y(Float{s0.y, s1.y, s2.y, s3.y}),
+        z(Float{s0.z, s1.z, s2.z, s3.z}),
+        w(Float{s0.w, s1.w, s2.w, s3.w}) {}
+  ALWAYS_INLINE Float& select(XYZW c) {
+    switch (c) {
+      case X:
+        return x;
+      case Y:
+        return y;
+      case Z:
+        return z;
+      case W:
+        return w;
+      default:
+        UNREACHABLE;
+    }
+  }
+  ALWAYS_INLINE Float& sel(XYZW c1) { return select(c1); }
+
+  ALWAYS_INLINE vec2 sel(XYZW c1, XYZW c2) {
+    return vec2(select(c1), select(c2));
+  }
+
+  ALWAYS_INLINE vec3 sel(XYZW c1, XYZW c2, XYZW c3) {
+    return vec3(select(c1), select(c2), select(c3));
+  }
+  ALWAYS_INLINE vec3_ref lsel(XYZW c1, XYZW c2, XYZW c3) {
+    return vec3_ref(select(c1), select(c2), select(c3));
+  }
+
+  ALWAYS_INLINE vec2_ref lsel(XYZW c1, XYZW c2) {
+    return vec2_ref(select(c1), select(c2));
+  }
+
+  ALWAYS_INLINE vec4 sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4) {
+    return vec4(select(c1), select(c2), select(c3), select(c4));
+  }
+  ALWAYS_INLINE vec4_ref lsel(XYZW c1, XYZW c2, XYZW c3, XYZW c4) {
+    return vec4_ref(select(c1), select(c2), select(c3), select(c4));
+  }
+
+  Float& operator[](int index) {
+    switch (index) {
+      case 0:
+        return x;
+      case 1:
+        return y;
+      case 2:
+        return z;
+      case 3:
+        return w;
+      default:
+        UNREACHABLE;
+    }
+  }
+
+  // glsl supports non-const indexing of vecs.
+  // hlsl doesn't. The code it generates is probably not wonderful.
+  Float operator[](I32 index) {
+    float sel_x = 0;
+    switch (index.x) {
+      case 0:
+        sel_x = x.x;
+        break;
+      case 1:
+        sel_x = y.x;
+        break;
+      case 2:
+        sel_x = z.x;
+        break;
+      case 3:
+        sel_x = w.x;
+        break;
+    }
+    float sel_y = 0;
+    switch (index.y) {
+      case 0:
+        sel_y = x.y;
+        break;
+      case 1:
+        sel_y = y.y;
+        break;
+      case 2:
+        sel_y = z.y;
+        break;
+      case 3:
+        sel_y = w.y;
+        break;
+    }
+    float sel_z = 0;
+    switch (index.z) {
+      case 0:
+        sel_z = x.z;
+        break;
+      case 1:
+        sel_z = y.z;
+        break;
+      case 2:
+        sel_z = z.z;
+        break;
+      case 3:
+        sel_z = w.z;
+        break;
+    }
+    float sel_w = 0;
+    switch (index.w) {
+      case 0:
+        sel_w = x.w;
+        break;
+      case 1:
+        sel_w = y.w;
+        break;
+      case 2:
+        sel_w = z.w;
+        break;
+      case 3:
+        sel_w = w.w;
+        break;
+    }
+    Float ret = {sel_x, sel_y, sel_z, sel_w};
+    return ret;
+  }
+
+  friend vec4 operator/(vec4 a, Float b) {
+    return vec4(a.x / b, a.y / b, a.z / b, a.w / b);
+  }
+  friend vec4 operator/(vec4 a, vec4 b) {
+    return vec4(a.x / b.x, a.y / b.y, a.z / b.z, a.w / b.w);
+  }
+
+  friend vec4 operator*(vec4 a, Float b) {
+    return vec4(a.x * b, a.y * b, a.z * b, a.w * b);
+  }
+
+  friend vec4 operator*(Float b, vec4 a) {
+    return vec4(a.x * b, a.y * b, a.z * b, a.w * b);
+  }
+  friend vec4 operator*(vec4 a, vec4 b) {
+    return vec4(a.x * b.x, a.y * b.y, a.z * b.z, a.w * b.w);
+  }
+
+  friend vec4 operator-(vec4 a, vec4 b) {
+    return vec4(a.x - b.x, a.y - b.y, a.z - b.z, a.w - b.w);
+  }
+  friend vec4 operator+(vec4 a, vec4 b) {
+    return vec4(a.x + b.x, a.y + b.y, a.z + b.z, a.w + b.w);
+  }
+  vec4& operator+=(vec4 a) {
+    x += a.x;
+    y += a.y;
+    z += a.z;
+    w += a.w;
+    return *this;
+  }
+  vec4& operator/=(vec4 a) {
+    x /= a.x;
+    y /= a.y;
+    z /= a.z;
+    w /= a.w;
+    return *this;
+  }
+  vec4& operator*=(vec4 a) {
+    x *= a.x;
+    y *= a.y;
+    z *= a.z;
+    w *= a.w;
+    return *this;
+  }
+  vec4& operator*=(Float a) {
+    x *= a;
+    y *= a;
+    z *= a;
+    w *= a;
+    return *this;
+  }
+
+  Float x;
+  Float y;
+  Float z;
+  Float w;
+};
+
+inline vec4_ref& vec4_ref::operator=(const vec4& a) {
+  x = a.x;
+  y = a.y;
+  z = a.z;
+  w = a.w;
+  return *this;
+}
+
+inline vec4 vec3::sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4) {
+  return vec4(select(c1), select(c2), select(c3), select(c4));
+}
+
+vec4_scalar force_scalar(const vec4& v) {
+  return vec4_scalar{force_scalar(v.x), force_scalar(v.y), force_scalar(v.z),
+                     force_scalar(v.w)};
+}
+
+vec4_scalar make_vec4(float n) { return vec4_scalar{n, n, n, n}; }
+
+vec4_scalar make_vec4(const vec2_scalar& v, float z, float w) {
+  return vec4_scalar{v.x, v.y, z, w};
+}
+
+vec4_scalar make_vec4(const vec2_scalar& a, const vec2_scalar& b) {
+  return vec4_scalar{a.x, a.y, b.x, b.y};
+}
+
+vec4_scalar make_vec4(const vec3_scalar& v, float w) {
+  return vec4_scalar{v.x, v.y, v.z, w};
+}
+
+vec4_scalar make_vec4(float x, float y, float z, float w) {
+  return vec4_scalar{x, y, z, w};
+}
+
+vec4_scalar make_vec4(float x, float y, const vec2_scalar& v) {
+  return vec4_scalar{x, y, v.x, v.y};
+}
+
+ivec4_scalar make_ivec4(const vec4_scalar& v) {
+  return ivec4_scalar{int32_t(v.x), int32_t(v.y), int32_t(v.z), int32_t(v.w)};
+}
+
+template <typename N>
+vec4 make_vec4(const N& n) {
+  return vec4(n);
+}
+
+template <typename X, typename Y>
+vec4 make_vec4(const X& x, const Y& y) {
+  return vec4(x, y);
+}
+
+template <typename X, typename Y, typename Z>
+vec4 make_vec4(const X& x, const Y& y, const Z& z) {
+  return vec4(x, y, z);
+}
+
+template <typename X, typename Y, typename Z, typename W>
+vec4 make_vec4(const X& x, const Y& y, const Z& z, const W& w) {
+  return vec4(x, y, z, w);
+}
+
+vec4_scalar make_vec4(const ivec4_scalar& v) {
+  return vec4_scalar{float(v.x), float(v.y), float(v.z), float(v.w)};
+}
+
+ALWAYS_INLINE vec3::vec3(vec4 v) : x(v.x), y(v.y), z(v.z) {}
+
+SI ivec4 roundfast(vec4 v, Float scale) {
+  return ivec4(roundfast(v.x, scale), roundfast(v.y, scale),
+               roundfast(v.z, scale), roundfast(v.w, scale));
+}
+
+vec4 operator*(vec4_scalar a, Float b) {
+  return vec4(a.x * b, a.y * b, a.z * b, a.w * b);
+}
+
+SI vec4 if_then_else(I32 c, vec4 t, vec4 e) {
+  return vec4(if_then_else(c, t.x, e.x), if_then_else(c, t.y, e.y),
+              if_then_else(c, t.z, e.z), if_then_else(c, t.w, e.w));
+}
+
+SI vec4 if_then_else(int32_t c, vec4 t, vec4 e) { return c ? t : e; }
+
+SI vec4_scalar if_then_else(int32_t c, vec4_scalar t, vec4_scalar e) {
+  return c ? t : e;
+}
+
+SI vec2 clamp(vec2 a, Float minVal, Float maxVal) {
+  return vec2(clamp(a.x, minVal, maxVal), clamp(a.y, minVal, maxVal));
+}
+
+SI vec2 clamp(vec2 a, vec2 minVal, vec2 maxVal) {
+  return vec2(clamp(a.x, minVal.x, maxVal.x), clamp(a.y, minVal.y, maxVal.y));
+}
+
+SI vec2_scalar clamp(vec2_scalar a, vec2_scalar minVal, vec2_scalar maxVal) {
+  return vec2_scalar{clamp(a.x, minVal.x, maxVal.x),
+                     clamp(a.y, minVal.y, maxVal.y)};
+}
+
+SI vec2_scalar clamp(vec2_scalar a, float minVal, float maxVal) {
+  return vec2_scalar{clamp(a.x, minVal, maxVal), clamp(a.y, minVal, maxVal)};
+}
+
+SI I32 clamp(I32 a, I32 minVal, I32 maxVal) {
+  a = if_then_else(a < minVal, minVal, a);
+  return if_then_else(a > maxVal, maxVal, a);
+}
+
+SI vec3 clamp(vec3 a, Float minVal, Float maxVal) {
+  return vec3(clamp(a.x, minVal, maxVal), clamp(a.y, minVal, maxVal),
+              clamp(a.z, minVal, maxVal));
+}
+
+SI vec3 clamp(vec3 a, vec3 minVal, vec3 maxVal) {
+  return vec3(clamp(a.x, minVal.x, maxVal.x), clamp(a.y, minVal.y, maxVal.y),
+              clamp(a.z, minVal.z, maxVal.z));
+}
+
+SI vec4 clamp(vec4 a, Float minVal, Float maxVal) {
+  return vec4(clamp(a.x, minVal, maxVal), clamp(a.y, minVal, maxVal),
+              clamp(a.z, minVal, maxVal), clamp(a.w, minVal, maxVal));
+}
+
+SI vec4 clamp(vec4 a, vec4 minVal, vec4 maxVal) {
+  return vec4(clamp(a.x, minVal.x, maxVal.x), clamp(a.y, minVal.y, maxVal.y),
+              clamp(a.z, minVal.z, maxVal.z), clamp(a.w, minVal.w, maxVal.w));
+}
+
+SI vec4_scalar clamp(vec4_scalar a, vec4_scalar minVal, vec4_scalar maxVal) {
+  return vec4_scalar{
+      clamp(a.x, minVal.x, maxVal.x), clamp(a.y, minVal.y, maxVal.y),
+      clamp(a.z, minVal.z, maxVal.z), clamp(a.w, minVal.w, maxVal.w)};
+}
+
+SI vec4_scalar clamp(vec4_scalar a, float minVal, float maxVal) {
+  return vec4_scalar{clamp(a.x, minVal, maxVal), clamp(a.y, minVal, maxVal),
+                     clamp(a.z, minVal, maxVal), clamp(a.w, minVal, maxVal)};
+}
+
+vec4 step(vec4 edge, vec4 x) {
+  return vec4(step(edge.x, x.x), step(edge.y, x.y), step(edge.z, x.z),
+              step(edge.w, x.w));
+}
+
+vec4_scalar step(vec4_scalar edge, vec4_scalar x) {
+  return vec4_scalar(step(edge.x, x.x), step(edge.y, x.y), step(edge.z, x.z),
+                     step(edge.w, x.w));
+}
+
+template <typename T>
+auto lessThanEqual(T x, T y) -> decltype(x <= y) {
+  return x <= y;
+}
+
+template <typename T>
+auto lessThan(T x, T y) -> decltype(x < y) {
+  return x < y;
+}
+
+SI bvec3 lessThanEqual(vec3 x, vec3 y) {
+  return bvec3(lessThanEqual(x.x, y.x), lessThanEqual(x.y, y.y),
+               lessThanEqual(x.z, y.z));
+}
+
+SI bvec2 lessThanEqual(vec2 x, vec2 y) {
+  return bvec2(lessThanEqual(x.x, y.x), lessThanEqual(x.y, y.y));
+}
+
+SI bvec2_scalar lessThanEqual(vec2_scalar x, vec2_scalar y) {
+  return bvec2_scalar{lessThanEqual(x.x, y.x), lessThanEqual(x.y, y.y)};
+}
+
+SI bvec4 lessThanEqual(vec4 x, vec4 y) {
+  return bvec4(lessThanEqual(x.x, y.x), lessThanEqual(x.y, y.y),
+               lessThanEqual(x.z, y.z), lessThanEqual(x.w, y.w));
+}
+
+SI bvec4_scalar lessThanEqual(vec4_scalar x, vec4_scalar y) {
+  return bvec4_scalar{lessThanEqual(x.x, y.x), lessThanEqual(x.y, y.y),
+                      lessThanEqual(x.z, y.z), lessThanEqual(x.w, y.w)};
+}
+
+SI bvec2 lessThan(vec2 x, vec2 y) {
+  return bvec2(lessThan(x.x, y.x), lessThan(x.y, y.y));
+}
+
+SI bvec2_scalar lessThan(vec2_scalar x, vec2_scalar y) {
+  return bvec2_scalar(lessThan(x.x, y.x), lessThan(x.y, y.y));
+}
+
+SI bvec4 lessThan(vec4 x, vec4 y) {
+  return bvec4(lessThan(x.x, y.x), lessThan(x.y, y.y), lessThan(x.z, y.z),
+               lessThan(x.w, y.w));
+}
+
+SI bvec4_scalar lessThan(vec4_scalar x, vec4_scalar y) {
+  return bvec4_scalar{lessThan(x.x, y.x), lessThan(x.y, y.y),
+                      lessThan(x.z, y.z), lessThan(x.w, y.w)};
+}
+
+template <typename T>
+auto greaterThan(T x, T y) -> decltype(x > y) {
+  return x > y;
+}
+
+bvec2 greaterThan(vec2 x, vec2 y) {
+  return bvec2(greaterThan(x.x, y.x), greaterThan(x.y, y.y));
+}
+
+bvec2_scalar greaterThan(vec2_scalar x, vec2_scalar y) {
+  return bvec2_scalar(greaterThan(x.x, y.x), greaterThan(x.y, y.y));
+}
+
+SI bvec4 greaterThan(vec4 x, vec4 y) {
+  return bvec4(greaterThan(x.x, y.x), greaterThan(x.y, y.y),
+               greaterThan(x.z, y.z), greaterThan(x.w, y.w));
+}
+
+SI bvec4_scalar greaterThan(vec4_scalar x, vec4_scalar y) {
+  return bvec4_scalar{greaterThan(x.x, y.x), greaterThan(x.y, y.y),
+                      greaterThan(x.z, y.z), greaterThan(x.w, y.w)};
+}
+
+template <typename T>
+auto greaterThanEqual(T x, T y) -> decltype(x >= y) {
+  return x >= y;
+}
+
+bvec4 greaterThanEqual(vec4 x, vec4 y) {
+  return bvec4(greaterThanEqual(x.x, y.x), greaterThanEqual(x.y, y.y),
+               greaterThanEqual(x.z, y.z), greaterThanEqual(x.w, y.w));
+}
+
+template <typename T>
+auto equal(T x, T y) -> decltype(x > y) {
+  return x == y;
+}
+
+bvec2 equal(vec2 x, vec2 y) { return bvec2(equal(x.x, y.x), equal(x.y, y.y)); }
+
+bvec2_scalar equal(vec2_scalar x, vec2_scalar y) {
+  return bvec2_scalar(equal(x.x, y.x), equal(x.y, y.y));
+}
+
+template <typename T>
+auto notEqual(T x, T y) -> decltype(x > y) {
+  return x != y;
+}
+
+bvec2 notEqual(vec2 x, vec2 y) {
+  return bvec2(notEqual(x.x, y.x), notEqual(x.y, y.y));
+}
+
+bvec2_scalar notEqual(vec2_scalar x, vec2_scalar y) {
+  return bvec2_scalar(notEqual(x.x, y.x), notEqual(x.y, y.y));
+}
+
+struct mat4_scalar;
+
+struct mat2_scalar {
+  vec2_scalar data[2];
+
+  mat2_scalar() = default;
+  IMPLICIT constexpr mat2_scalar(float a)
+      : data{vec2_scalar(a), vec2_scalar(a)} {}
+  constexpr mat2_scalar(vec2_scalar a, vec2_scalar b) : data{a, b} {}
+  IMPLICIT mat2_scalar(const mat4_scalar& mat);
+
+  vec2_scalar& operator[](int index) { return data[index]; }
+  const vec2_scalar& operator[](int index) const { return data[index]; }
+
+  friend vec2_scalar operator*(mat2_scalar m, vec2_scalar v) {
+    vec2_scalar u;
+    u.x = m[0].x * v.x + m[1].x * v.y;
+    u.y = m[0].y * v.x + m[1].y * v.y;
+    return u;
+  }
+
+  friend vec2 operator*(mat2_scalar m, vec2 v) {
+    vec2 u;
+    u.x = m[0].x * v.x + m[1].x * v.y;
+    u.y = m[0].y * v.x + m[1].y * v.y;
+    return u;
+  }
+
+  friend mat2_scalar operator*(mat2_scalar m, float f) {
+    mat2_scalar u = m;
+    u[0].x *= f;
+    u[0].y *= f;
+    u[1].x *= f;
+    u[1].y *= f;
+    return u;
+  }
+};
+
+struct mat4;
+
+struct mat2 {
+  vec2 data[2];
+
+  vec2& operator[](int index) { return data[index]; }
+  const vec2& operator[](int index) const { return data[index]; }
+  mat2() = default;
+
+  IMPLICIT constexpr mat2(Float a) : data{vec2(a), vec2(a)} {}
+
+  constexpr mat2(vec2 a, vec2 b) : data{a, b} {}
+  IMPLICIT mat2(const mat4& mat);
+  IMPLICIT constexpr mat2(mat2_scalar s)
+      : data{vec2(s.data[0]), vec2(s.data[1])} {}
+
+  friend vec2 operator*(mat2 m, vec2 v) {
+    vec2 u;
+    u.x = m[0].x * v.x + m[1].x * v.y;
+    u.y = m[0].y * v.x + m[1].y * v.y;
+    return u;
+  }
+  friend mat2 operator*(mat2 m, Float f) {
+    mat2 u = m;
+    u[0].x *= f;
+    u[0].y *= f;
+    u[1].x *= f;
+    u[1].y *= f;
+    return u;
+  }
+};
+
+mat2_scalar make_mat2(float n) { return mat2_scalar{{n, n}, {n, n}}; }
+
+mat2_scalar make_mat2(const mat2_scalar& m) { return m; }
+
+mat2_scalar make_mat2(const vec2_scalar& x, const vec2_scalar& y) {
+  return mat2_scalar{x, y};
+}
+
+template <typename N>
+mat2 make_mat2(const N& n) {
+  return mat2(n);
+}
+
+template <typename X, typename Y>
+mat2 make_mat2(const X& x, const Y& y) {
+  return mat2(x, y);
+}
+
+SI mat2 if_then_else(I32 c, mat2 t, mat2 e) {
+  return mat2(if_then_else(c, t[0], e[0]), if_then_else(c, t[0], e[1]));
+}
+
+SI mat2 if_then_else(int32_t c, mat2 t, mat2 e) { return c ? t : e; }
+
+struct mat3_scalar {
+  vec3_scalar data[3];
+
+  mat3_scalar() = default;
+  constexpr mat3_scalar(vec3_scalar a, vec3_scalar b, vec3_scalar c)
+      : data{a, b, c} {}
+  IMPLICIT mat3_scalar(const mat4_scalar& mat);
+
+  vec3_scalar& operator[](int index) { return data[index]; }
+  const vec3_scalar& operator[](int index) const { return data[index]; }
+
+  friend vec3_scalar operator*(mat3_scalar m, vec3_scalar v) {
+    vec3_scalar u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z;
+    return u;
+  }
+
+  friend vec3 operator*(mat3_scalar m, vec3 v) {
+    vec3 u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z;
+    return u;
+  }
+
+  friend auto operator*(mat3_scalar a, mat3_scalar b) {
+    mat3_scalar r;
+    for (int c = 0; c < 3; c++) {
+      const auto& v = b[c];
+      r[c].x = a[0].x * v.x + a[1].x * v.y + a[2].x * v.z;
+      r[c].y = a[0].y * v.x + a[1].y * v.y + a[2].y * v.z;
+      r[c].z = a[0].z * v.x + a[1].z * v.y + a[2].z * v.z;
+    }
+    return r;
+  }
+};
+
+struct mat3 {
+  vec3 data[3];
+
+  vec3& operator[](int index) { return data[index]; }
+  const vec3& operator[](int index) const { return data[index]; }
+  mat3() = default;
+  constexpr mat3(vec3 a, vec3 b, vec3 c) : data{a, b, c} {}
+
+  IMPLICIT constexpr mat3(mat3_scalar s)
+      : data{vec3(s.data[0]), vec3(s.data[1]), vec3(s.data[2])} {}
+
+  constexpr mat3(mat3_scalar s0, mat3_scalar s1, mat3_scalar s2, mat3_scalar s3)
+      : data{vec3(s0.data[0], s1.data[0], s2.data[0], s3.data[0]),
+             vec3(s0.data[1], s1.data[1], s2.data[1], s3.data[1]),
+             vec3(s0.data[2], s1.data[2], s2.data[2], s3.data[2])} {}
+
+  constexpr mat3(Float d1, Float d2, Float d3, Float d4, Float d5, Float d6,
+                 Float d7, Float d8, Float d9)
+      : data{vec3(d1, d2, d3), vec3(d4, d5, d6), vec3(d7, d8, d9)} {}
+
+  IMPLICIT mat3(const mat4& mat);
+
+  friend vec3 operator*(mat3 m, vec3 v) {
+    vec3 u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z;
+    return u;
+  }
+};
+
+mat3_scalar force_scalar(const mat3& v) {
+  return mat3_scalar{force_scalar(v[0]), force_scalar(v[1]),
+                     force_scalar(v[2])};
+}
+
+mat3_scalar make_mat3(const mat3_scalar& m) { return m; }
+
+mat3_scalar make_mat3(const vec3_scalar& x, const vec3_scalar& y,
+                      const vec3_scalar& z) {
+  return mat3_scalar{x, y, z};
+}
+
+constexpr mat3_scalar make_mat3(float m0, float m1, float m2, float m3,
+                                float m4, float m5, float m6, float m7,
+                                float m8) {
+  return mat3_scalar{{m0, m1, m2}, {m3, m4, m5}, {m6, m7, m8}};
+}
+
+template <typename N>
+mat3 make_mat3(const N& n) {
+  return mat3(n);
+}
+
+template <typename X, typename Y, typename Z>
+mat3 make_mat3(const X& x, const Y& y, const Z& z) {
+  return mat3(x, y, z);
+}
+
+struct mat3x4_scalar {
+  vec4_scalar data[3];
+
+  mat3x4_scalar() = default;
+  constexpr mat3x4_scalar(vec4_scalar a, vec4_scalar b, vec4_scalar c)
+      : data{a, b, c} {}
+
+  auto& operator[](int index) { return data[index]; }
+  constexpr auto operator[](int index) const { return data[index]; }
+
+  friend auto operator*(mat3x4_scalar m, vec3_scalar v) {
+    vec4_scalar u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z;
+    u.w = m[0].w * v.x + m[1].w * v.y + m[2].w * v.z;
+    return u;
+  }
+
+  friend auto operator*(mat3x4_scalar m, vec3 v) {
+    vec4 u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z;
+    u.w = m[0].w * v.x + m[1].w * v.y + m[2].w * v.z;
+    return u;
+  }
+};
+
+constexpr mat3x4_scalar make_mat3x4(float m0, float m1, float m2, float m3,
+                                    float m4, float m5, float m6, float m7,
+                                    float m8, float m9, float m10, float m11) {
+  return mat3x4_scalar{
+      {m0, m1, m2, m3},
+      {m4, m5, m6, m7},
+      {m8, m9, m10, m11},
+  };
+}
+
+struct mat4x3_scalar {
+  vec3_scalar data[4];
+
+  mat4x3_scalar() = default;
+  constexpr mat4x3_scalar(vec3_scalar a, vec3_scalar b, vec3_scalar c,
+                          vec3_scalar d)
+      : data{a, b, c, d} {}
+
+  auto& operator[](int index) { return data[index]; }
+  constexpr auto operator[](int index) const { return data[index]; }
+
+  friend auto operator*(mat4x3_scalar m, vec4_scalar v) {
+    vec3_scalar u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z + m[3].x * v.w;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z + m[3].y * v.w;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z + m[3].z * v.w;
+    return u;
+  }
+
+  friend auto operator*(mat4x3_scalar m, vec4 v) {
+    vec3 u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z + m[3].x * v.w;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z + m[3].y * v.w;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z + m[3].z * v.w;
+    return u;
+  }
+};
+
+constexpr mat4x3_scalar transpose(const mat3x4_scalar m) {
+  return {{m[0].x, m[1].x, m[2].x},
+          {m[0].y, m[1].y, m[2].y},
+          {m[0].z, m[1].z, m[2].z},
+          {m[0].w, m[1].w, m[2].w}};
+}
+
+struct mat4_scalar {
+  vec4_scalar data[4];
+
+  mat4_scalar() = default;
+  constexpr mat4_scalar(vec4_scalar a, vec4_scalar b, vec4_scalar c,
+                        vec4_scalar d)
+      : data{a, b, c, d} {}
+
+  vec4_scalar& operator[](int index) { return data[index]; }
+  const vec4_scalar& operator[](int index) const { return data[index]; }
+
+  static mat4_scalar load_from_ptr(const float* f) {
+    return mat4_scalar(
+        vec4_scalar::load_from_ptr(&f[0]), vec4_scalar::load_from_ptr(&f[4]),
+        vec4_scalar::load_from_ptr(&f[8]), vec4_scalar::load_from_ptr(&f[12]));
+  }
+
+  friend vec4_scalar operator*(mat4_scalar m, vec4_scalar v) {
+    vec4_scalar u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z + m[3].x * v.w;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z + m[3].y * v.w;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z + m[3].z * v.w;
+    u.w = m[0].w * v.x + m[1].w * v.y + m[2].w * v.z + m[3].w * v.w;
+    return u;
+  }
+
+  friend vec4 operator*(mat4_scalar m, vec4 v) {
+    vec4 u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z + m[3].x * v.w;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z + m[3].y * v.w;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z + m[3].z * v.w;
+    u.w = m[0].w * v.x + m[1].w * v.y + m[2].w * v.z + m[3].w * v.w;
+    return u;
+  }
+};
+
+struct mat4 {
+  vec4 data[4];
+
+  mat4() = default;
+  IMPLICIT constexpr mat4(mat4_scalar s)
+      : data{vec4(s.data[0]), vec4(s.data[1]), vec4(s.data[2]),
+             vec4(s.data[3])} {}
+
+  constexpr mat4(vec4 a, vec4 b, vec4 c, vec4 d) : data{a, b, c, d} {}
+
+  vec4& operator[](int index) { return data[index]; }
+  const vec4& operator[](int index) const { return data[index]; }
+
+  friend vec4 operator*(mat4 m, vec4 v) {
+    vec4 u;
+    u.x = m[0].x * v.x + m[1].x * v.y + m[2].x * v.z + m[3].x * v.w;
+    u.y = m[0].y * v.x + m[1].y * v.y + m[2].y * v.z + m[3].y * v.w;
+    u.z = m[0].z * v.x + m[1].z * v.y + m[2].z * v.z + m[3].z * v.w;
+    u.w = m[0].w * v.x + m[1].w * v.y + m[2].w * v.z + m[3].w * v.w;
+    return u;
+  }
+};
+
+mat3::mat3(const mat4& mat)
+    : mat3(vec3(mat[0].x, mat[0].y, mat[0].z),
+           vec3(mat[1].x, mat[1].y, mat[1].z),
+           vec3(mat[2].x, mat[2].y, mat[2].z)) {}
+
+IMPLICIT mat3_scalar::mat3_scalar(const mat4_scalar& mat)
+    : mat3_scalar(vec3_scalar(mat[0].x, mat[0].y, mat[0].z),
+                  vec3_scalar(mat[1].x, mat[1].y, mat[1].z),
+                  vec3_scalar(mat[2].x, mat[2].y, mat[2].z)) {}
+
+IMPLICIT mat2::mat2(const mat4& mat)
+    : mat2(vec2(mat[0].x, mat[0].y), vec2(mat[1].x, mat[1].y)) {}
+
+IMPLICIT mat2_scalar::mat2_scalar(const mat4_scalar& mat)
+    : mat2_scalar(vec2_scalar(mat[0].x, mat[0].y),
+                  vec2_scalar(mat[1].x, mat[1].y)) {}
+
+mat2_scalar make_mat2(const mat4_scalar& m) { return mat2_scalar(m); }
+
+mat3_scalar make_mat3(const mat4_scalar& m) { return mat3_scalar(m); }
+
+mat4_scalar force_scalar(const mat4& v) {
+  return mat4_scalar(force_scalar(v[0]), force_scalar(v[1]), force_scalar(v[2]),
+                     force_scalar(v[3]));
+}
+
+mat4_scalar make_mat4(const mat4_scalar& m) { return m; }
+
+mat4_scalar make_mat4(const vec4_scalar& x, const vec4_scalar& y,
+                      const vec4_scalar& z, const vec4_scalar& w) {
+  return mat4_scalar{x, y, z, w};
+}
+
+constexpr mat4_scalar make_mat4(float m0, float m1, float m2, float m3,
+                                float m4, float m5, float m6, float m7,
+                                float m8, float m9, float m10, float m11,
+                                float m12, float m13, float m14, float m15) {
+  return mat4_scalar{{m0, m1, m2, m3},
+                     {m4, m5, m6, m7},
+                     {m8, m9, m10, m11},
+                     {m12, m13, m14, m15}};
+}
+
+template <typename N>
+mat4 make_mat4(const N& n) {
+  return mat4(n);
+}
+
+template <typename X, typename Y, typename Z, typename W>
+mat4 make_mat4(const X& x, const Y& y, const Z& z, const W& w) {
+  return mat4(x, y, z, w);
+}
+
+SI mat3 if_then_else(I32 c, mat3 t, mat3 e) {
+  return mat3{if_then_else(c, t[0], e[0]), if_then_else(c, t[1], e[1]),
+              if_then_else(c, t[2], e[2])};
+}
+
+SI mat3 if_then_else(int32_t c, mat3 t, mat3 e) { return c ? t : e; }
+
+SI mat4 if_then_else(I32 c, mat4 t, mat4 e) {
+  return mat4{if_then_else(c, t[0], e[0]), if_then_else(c, t[1], e[1]),
+              if_then_else(c, t[2], e[2]), if_then_else(c, t[3], e[3])};
+}
+
+SI mat4 if_then_else(int32_t c, mat4 t, mat4 e) { return c ? t : e; }
+
+template <typename T, typename U, typename A,
+          typename R = typename T::vector_type>
+SI R mix(T x, U y, A a) {
+  return (y - x) * a + x;
+}
+
+SI Float mix(Float x, Float y, Float a) { return (y - x) * a + x; }
+
+template <typename T>
+SI T mix(T x, T y, float a) {
+  return (y - x) * a + x;
+}
+
+template <typename T>
+SI T mix(T x, T y, vec2_scalar a) {
+  return T{mix(x.x, y.x, a.x), mix(x.y, y.y, a.y)};
+}
+
+template <typename T>
+SI T mix(T x, T y, vec3_scalar a) {
+  return T{mix(x.x, y.x, a.x), mix(x.y, y.y, a.y), mix(x.z, y.z, a.z)};
+}
+
+template <typename T>
+SI T mix(T x, T y, vec4_scalar a) {
+  return T{mix(x.x, y.x, a.x), mix(x.y, y.y, a.y), mix(x.z, y.z, a.z),
+           mix(x.w, y.w, a.w)};
+}
+
+ivec4 ivec2::sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4) {
+  return ivec4(select(c1), select(c2), select(c3), select(c4));
+}
+
+vec4 vec2::sel(XYZW c1, XYZW c2, XYZW c3, XYZW c4) {
+  return vec4(select(c1), select(c2), select(c3), select(c4));
+}
+
+bool any(bool x) { return x; }
+
+Bool any(bvec4 x) { return x.x | x.y | x.z | x.w; }
+
+bool any(bvec4_scalar x) { return x.x | x.y | x.z | x.w; }
+
+Bool any(bvec2 x) { return x.x | x.y; }
+
+bool any(bvec2_scalar x) { return x.x | x.y; }
+
+bool all(bool x) { return x; }
+
+Bool all(bvec2 x) { return x.x & x.y; }
+
+bool all(bvec2_scalar x) { return x.x & x.y; }
+
+Bool all(bvec4 x) { return x.x & x.y & x.z & x.w; }
+
+bool all(bvec4_scalar x) { return x.x & x.y & x.z & x.w; }
+
+SI vec4 if_then_else(bvec4 c, vec4 t, vec4 e) {
+  return vec4(if_then_else(c.x, t.x, e.x), if_then_else(c.y, t.y, e.y),
+              if_then_else(c.z, t.z, e.z), if_then_else(c.w, t.w, e.w));
+}
+SI vec3 if_then_else(bvec3 c, vec3 t, vec3 e) {
+  return vec3(if_then_else(c.x, t.x, e.x), if_then_else(c.y, t.y, e.y),
+              if_then_else(c.z, t.z, e.z));
+}
+
+SI vec2 if_then_else(bvec2 c, vec2 t, vec2 e) {
+  return vec2(if_then_else(c.x, t.x, e.x), if_then_else(c.y, t.y, e.y));
+}
+
+template <typename T, typename R = typename T::vector_type>
+SI R mix(T x, T y, bvec4 a) {
+  return if_then_else(a, y, x);
+}
+
+template <typename T, typename R = typename T::vector_type>
+SI R mix(T x, T y, bvec3 a) {
+  return if_then_else(a, y, x);
+}
+
+template <typename T, typename R = typename T::vector_type>
+SI R mix(T x, T y, bvec2 a) {
+  return if_then_else(a, y, x);
+}
+
+template <typename T>
+SI T mix(T x, T y, bvec4_scalar a) {
+  return T{a.x ? y.x : x.x, a.y ? y.y : x.y, a.z ? y.z : x.z, a.w ? y.w : x.w};
+}
+
+template <typename T>
+SI T mix(T x, T y, bvec4_scalar1 a) {
+  return a.x ? y : x;
+}
+
+template <typename T>
+SI T mix(T x, T y, bvec3_scalar a) {
+  return T{a.x ? y.x : x.x, a.y ? y.y : x.y, a.z ? y.z : x.z};
+}
+
+template <typename T>
+SI T mix(T x, T y, bvec3_scalar1 a) {
+  return a.x ? y : x;
+}
+
+template <typename T>
+SI T mix(T x, T y, bvec2_scalar a) {
+  return T{a.x ? y.x : x.x, a.y ? y.y : x.y};
+}
+
+template <typename T>
+SI T mix(T x, T y, bvec2_scalar1 a) {
+  return a.x ? y : x;
+}
+
+float dot(vec3_scalar a, vec3_scalar b) {
+  return a.x * b.x + a.y * b.y + a.z * b.z;
+}
+
+Float dot(vec3 a, vec3 b) { return a.x * b.x + a.y * b.y + a.z * b.z; }
+
+float dot(vec2_scalar a, vec2_scalar b) { return a.x * b.x + a.y * b.y; }
+
+Float dot(vec2 a, vec2 b) { return a.x * b.x + a.y * b.y; }
+
+#define sin __glsl_sin
+
+float sin(float x) { return sinf(x); }
+
+Float sin(Float v) { return {sinf(v.x), sinf(v.y), sinf(v.z), sinf(v.w)}; }
+
+#define cos __glsl_cos
+
+float cos(float x) { return cosf(x); }
+
+Float cos(Float v) { return {cosf(v.x), cosf(v.y), cosf(v.z), cosf(v.w)}; }
+
+#define tan __glsl_tan
+
+float tan(float x) { return tanf(x); }
+
+Float tan(Float v) { return {tanf(v.x), tanf(v.y), tanf(v.z), tanf(v.w)}; }
+
+#define atan __glsl_atan
+
+float atan(float x) { return atanf(x); }
+
+Float atan(Float v) { return {atanf(v.x), atanf(v.y), atanf(v.z), atanf(v.w)}; }
+
+float atan(float a, float b) { return atan2f(a, b); }
+
+Float atan(Float a, Float b) {
+  return {atan2f(a.x, b.x), atan2f(a.y, b.y), atan2f(a.z, b.z),
+          atan2f(a.w, b.w)};
+}
+
+bvec4 equal(vec4 x, vec4 y) {
+  return bvec4(equal(x.x, y.x), equal(x.y, y.y), equal(x.z, y.z),
+               equal(x.w, y.w));
+}
+
+bvec4_scalar equal(vec4_scalar x, vec4_scalar y) {
+  return bvec4_scalar(equal(x.x, y.x), equal(x.y, y.y), equal(x.z, y.z),
+                      equal(x.w, y.w));
+}
+
+bvec4 notEqual(vec4 x, vec4 y) {
+  return bvec4(notEqual(x.x, y.x), notEqual(x.y, y.y), notEqual(x.z, y.z),
+               notEqual(x.w, y.w));
+}
+
+bvec4_scalar notEqual(vec4_scalar x, vec4_scalar y) {
+  return bvec4_scalar(notEqual(x.x, y.x), notEqual(x.y, y.y),
+                      notEqual(x.z, y.z), notEqual(x.w, y.w));
+}
+
+bvec4 notEqual(ivec4 a, ivec4 b) {
+  return bvec4(a.x != b.x, a.y != b.y, a.z != b.z, a.w != b.w);
+}
+
+bvec4_scalar notEqual(ivec4_scalar a, ivec4_scalar b) {
+  return bvec4_scalar{a.x != b.x, a.y != b.y, a.z != b.z, a.w != b.w};
+}
+
+mat3 transpose(mat3 m) {
+  return mat3(vec3(m[0].x, m[1].x, m[2].x), vec3(m[0].y, m[1].y, m[2].y),
+              vec3(m[0].z, m[1].z, m[2].z));
+}
+
+mat3_scalar transpose(mat3_scalar m) {
+  return mat3_scalar{vec3_scalar(m[0].x, m[1].x, m[2].x),
+                     vec3_scalar(m[0].y, m[1].y, m[2].y),
+                     vec3_scalar(m[0].z, m[1].z, m[2].z)};
+}
+
+vec2 abs(vec2 v) { return vec2(abs(v.x), abs(v.y)); }
+
+vec2_scalar abs(vec2_scalar v) { return vec2_scalar{fabsf(v.x), fabsf(v.y)}; }
+
+vec2 sign(vec2 v) { return vec2(sign(v.x), sign(v.y)); }
+
+vec2_scalar sign(vec2_scalar v) { return vec2_scalar{sign(v.x), sign(v.y)}; }
+
+Float mod(Float a, Float b) { return a - b * floor(a / b); }
+
+vec2 mod(vec2 a, vec2 b) { return vec2(mod(a.x, b.x), mod(a.y, b.y)); }
+
+vec3 abs(vec3 v) { return vec3(abs(v.x), abs(v.y), abs(v.z)); }
+
+vec3 sign(vec3 v) { return vec3(sign(v.x), sign(v.y), sign(v.z)); }
+
+mat2 inverse(mat2 v) {
+  Float det = v[0].x * v[1].y - v[0].y * v[1].x;
+  return mat2(vec2(v[1].y, -v[0].y), vec2(-v[1].x, v[0].x)) * (1. / det);
+}
+
+mat2_scalar inverse(mat2_scalar v) {
+  float det = v[0].x * v[1].y - v[0].y * v[1].x;
+  return mat2_scalar{{v[1].y, -v[0].y}, {-v[1].x, v[0].x}} * (1. / det);
+}
+
+int32_t get_nth(I32 a, int n) { return a[n]; }
+
+float get_nth(Float a, int n) { return a[n]; }
+
+float get_nth(float a, int) { return a; }
+
+ivec2_scalar get_nth(ivec2 a, int n) { return ivec2_scalar{a.x[n], a.y[n]}; }
+
+vec2_scalar get_nth(vec2 a, int n) { return vec2_scalar{a.x[n], a.y[n]}; }
+
+vec3_scalar get_nth(vec3 a, int n) {
+  return vec3_scalar{a.x[n], a.y[n], a.z[n]};
+}
+
+vec4_scalar get_nth(vec4 a, int n) {
+  return vec4_scalar{a.x[n], a.y[n], a.z[n], a.w[n]};
+}
+
+ivec4_scalar get_nth(ivec4 a, int n) {
+  return ivec4_scalar{a.x[n], a.y[n], a.z[n], a.w[n]};
+}
+
+mat3_scalar get_nth(mat3 a, int n) {
+  return make_mat3(get_nth(a[0], n), get_nth(a[1], n), get_nth(a[2], n));
+}
+
+void put_nth(Float& dst, int n, float src) { dst[n] = src; }
+
+void put_nth(I32& dst, int n, int32_t src) { dst[n] = src; }
+
+void put_nth(ivec2& dst, int n, ivec2_scalar src) {
+  dst.x[n] = src.x;
+  dst.y[n] = src.y;
+}
+
+void put_nth(vec2& dst, int n, vec2_scalar src) {
+  dst.x[n] = src.x;
+  dst.y[n] = src.y;
+}
+
+void put_nth(vec3& dst, int n, vec3_scalar src) {
+  dst.x[n] = src.x;
+  dst.y[n] = src.y;
+  dst.z[n] = src.z;
+}
+
+void put_nth(ivec4& dst, int n, ivec4_scalar src) {
+  dst.x[n] = src.x;
+  dst.y[n] = src.y;
+  dst.z[n] = src.z;
+  dst.w[n] = src.w;
+}
+
+void put_nth(vec4& dst, int n, vec4_scalar src) {
+  dst.x[n] = src.x;
+  dst.y[n] = src.y;
+  dst.z[n] = src.z;
+  dst.w[n] = src.w;
+}
+
+// Use an ElementType type constructor
+// so that we can implement element_type for
+// Int and Float
+template <typename V>
+struct ElementType {
+  typedef typename V::element_type ty;
+};
+
+template <>
+struct ElementType<float> {
+  typedef float ty;
+};
+
+template <>
+struct ElementType<int> {
+  typedef float ty;
+};
+
+template <>
+struct ElementType<Float> {
+  typedef float ty;
+};
+
+template <>
+struct ElementType<I32> {
+  typedef int32_t ty;
+};
+
+void put_nth_component(ivec2_scalar& dst, int n, int32_t src) {
+  switch (n) {
+    case 0:
+      dst.x = src;
+      break;
+    case 1:
+      dst.y = src;
+      break;
+  }
+}
+
+void put_nth_component(ivec4_scalar& dst, int n, int32_t src) {
+  switch (n) {
+    case 0:
+      dst.x = src;
+      break;
+    case 1:
+      dst.y = src;
+      break;
+    case 2:
+      dst.z = src;
+      break;
+    case 3:
+      dst.w = src;
+      break;
+  }
+}
+
+void put_nth_component(int& dst, int n, int src) {
+  switch (n) {
+    case 0:
+      dst = src;
+      break;
+  }
+}
+
+void put_nth_component(float& dst, int n, float src) {
+  switch (n) {
+    case 0:
+      dst = src;
+      break;
+  }
+}
+
+void put_nth_component(vec2_scalar& dst, int n, float src) {
+  switch (n) {
+    case 0:
+      dst.x = src;
+      break;
+    case 1:
+      dst.y = src;
+      break;
+  }
+}
+
+void put_nth_component(vec3_scalar& dst, int n, float src) {
+  switch (n) {
+    case 0:
+      dst.x = src;
+      break;
+    case 1:
+      dst.y = src;
+      break;
+    case 2:
+      dst.z = src;
+      break;
+  }
+}
+
+void put_nth_component(vec4_scalar& dst, int n, float src) {
+  switch (n) {
+    case 0:
+      dst.x = src;
+      break;
+    case 1:
+      dst.y = src;
+      break;
+    case 2:
+      dst.z = src;
+      break;
+    case 3:
+      dst.w = src;
+      break;
+  }
+}
+
+Float init_interp(float init0, float step) {
+  float init1 = init0 + step;
+  float init2 = init1 + step;
+  float init3 = init2 + step;
+  return {init0, init1, init2, init3};
+}
+
+vec2 init_interp(vec2_scalar init, vec2_scalar step) {
+  return vec2(init_interp(init.x, step.x), init_interp(init.y, step.y));
+}
+
+vec3 init_interp(vec3_scalar init, vec3_scalar step) {
+  return vec3(init_interp(init.x, step.x), init_interp(init.y, step.y),
+              init_interp(init.z, step.z));
+}
+
+vec4 init_interp(vec4_scalar init, vec4_scalar step) {
+  return vec4(init_interp(init.x, step.x), init_interp(init.y, step.y),
+              init_interp(init.z, step.z), init_interp(init.w, step.w));
+}
+
+template <typename T, size_t N>
+struct Array {
+  T elements[N];
+  T& operator[](size_t i) { return elements[i]; }
+  const T& operator[](size_t i) const { return elements[i]; }
+  template <typename S>
+  void convert(const Array<S, N>& s) {
+    for (size_t i = 0; i < N; ++i) elements[i] = T(s[i]);
+  }
+};
+
+template <size_t SIZE>
+Array<vec2, SIZE> if_then_else(I32 c, Array<vec2, SIZE> t,
+                               Array<vec2, SIZE> e) {
+  Array<vec2, SIZE> r;
+  for (size_t i = 0; i < SIZE; i++) {
+    r[i] = if_then_else(c, t[i], e[i]);
+  }
+  return r;
+}
+
+}  // namespace glsl
diff --git a/gfx/wr/swgl/src/lib.rs b/gfx/wr/swgl/src/lib.rs
new file mode 100644
index 0000000000..e8fc030e0c
--- /dev/null
+++ b/gfx/wr/swgl/src/lib.rs
@@ -0,0 +1,12 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#![crate_name = "swgl"]
+#![crate_type = "lib"]
+
+extern crate gleam;
+
+mod swgl_fns;
+
+pub use crate::swgl_fns::*;
diff --git a/gfx/wr/swgl/src/program.h b/gfx/wr/swgl/src/program.h
new file mode 100644
index 0000000000..9ea7c6dd6e
--- /dev/null
+++ b/gfx/wr/swgl/src/program.h
@@ -0,0 +1,186 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+struct VertexAttrib;
+
+namespace glsl {
+
+// Type holding group of scalars interpolated across rasterized rows and spans,
+// shuttling values between vertex shaders and fragment shaders.
+// GCC requires power-of-two vector sizes, so must use glsl type as workaround
+// to operate in Float-sized chunks.
+typedef vec3 Interpolants;
+
+// Clip distances, if enabled, are always stored in the first SIMD chunk of the
+// interpolants.
+static ALWAYS_INLINE Float get_clip_distances(const Interpolants& interp) {
+  return interp.x;
+}
+
+struct VertexShaderImpl;
+struct FragmentShaderImpl;
+
+struct ProgramImpl {
+  virtual ~ProgramImpl() {}
+  virtual int get_uniform(const char* name) const = 0;
+  virtual void bind_attrib(const char* name, int index) = 0;
+  virtual int get_attrib(const char* name) const = 0;
+  virtual size_t interpolants_size() const = 0;
+  virtual VertexShaderImpl* get_vertex_shader() = 0;
+  virtual FragmentShaderImpl* get_fragment_shader() = 0;
+  virtual const char* get_name() const = 0;
+};
+
+typedef ProgramImpl* (*ProgramLoader)();
+
+// The maximum size of the gl_ClipDistance array.
+constexpr int32_t gl_MaxClipDistances = 4;
+
+struct VertexShaderImpl {
+  typedef void (*SetUniform1iFunc)(VertexShaderImpl*, int index, int value);
+  typedef void (*SetUniform4fvFunc)(VertexShaderImpl*, int index,
+                                    const float* value);
+  typedef void (*SetUniformMatrix4fvFunc)(VertexShaderImpl*, int index,
+                                          const float* value);
+  typedef void (*InitBatchFunc)(VertexShaderImpl*);
+  typedef void (*LoadAttribsFunc)(VertexShaderImpl*, VertexAttrib* attribs,
+                                  uint32_t start, int instance, int count);
+  typedef void (*RunPrimitiveFunc)(VertexShaderImpl*, char* interps,
+                                   size_t interp_stride);
+
+  SetUniform1iFunc set_uniform_1i_func = nullptr;
+  SetUniform4fvFunc set_uniform_4fv_func = nullptr;
+  SetUniformMatrix4fvFunc set_uniform_matrix4fv_func = nullptr;
+  InitBatchFunc init_batch_func = nullptr;
+  LoadAttribsFunc load_attribs_func = nullptr;
+  RunPrimitiveFunc run_primitive_func = nullptr;
+
+  enum FLAGS {
+    CLIP_DISTANCE = 1 << 0,
+  };
+  int flags = 0;
+  void enable_clip_distance() { flags |= CLIP_DISTANCE; }
+  ALWAYS_INLINE bool use_clip_distance() const {
+    return (flags & CLIP_DISTANCE) != 0;
+  }
+
+  vec4 gl_Position;
+  Float gl_ClipDistance[gl_MaxClipDistances];
+
+  void set_uniform_1i(int index, int value) {
+    (*set_uniform_1i_func)(this, index, value);
+  }
+
+  void set_uniform_4fv(int index, const float* value) {
+    (*set_uniform_4fv_func)(this, index, value);
+  }
+
+  void set_uniform_matrix4fv(int index, const float* value) {
+    (*set_uniform_matrix4fv_func)(this, index, value);
+  }
+
+  void init_batch() { (*init_batch_func)(this); }
+
+  ALWAYS_INLINE void load_attribs(VertexAttrib* attribs, uint32_t start,
+                                  int instance, int count) {
+    (*load_attribs_func)(this, attribs, start, instance, count);
+  }
+
+  ALWAYS_INLINE void run_primitive(char* interps, size_t interp_stride) {
+    (*run_primitive_func)(this, interps, interp_stride);
+  }
+};
+
+// The number of pixels in a step.
+constexpr int32_t swgl_StepSize = 4;
+
+struct FragmentShaderImpl {
+  typedef void (*InitSpanFunc)(FragmentShaderImpl*, const void* interps,
+                               const void* step);
+  typedef void (*RunFunc)(FragmentShaderImpl*);
+  typedef void (*SkipFunc)(FragmentShaderImpl*, int steps);
+  typedef void (*InitSpanWFunc)(FragmentShaderImpl*, const void* interps,
+                                const void* step);
+  typedef void (*RunWFunc)(FragmentShaderImpl*);
+  typedef void (*SkipWFunc)(FragmentShaderImpl*, int steps);
+  typedef int (*DrawSpanRGBA8Func)(FragmentShaderImpl*);
+  typedef int (*DrawSpanR8Func)(FragmentShaderImpl*);
+
+  InitSpanFunc init_span_func = nullptr;
+  RunFunc run_func = nullptr;
+  SkipFunc skip_func = nullptr;
+  InitSpanWFunc init_span_w_func = nullptr;
+  RunWFunc run_w_func = nullptr;
+  SkipWFunc skip_w_func = nullptr;
+  DrawSpanRGBA8Func draw_span_RGBA8_func = nullptr;
+  DrawSpanR8Func draw_span_R8_func = nullptr;
+
+  enum FLAGS {
+    DISCARD = 1 << 0,
+    PERSPECTIVE = 1 << 1,
+  };
+  int flags = 0;
+  void enable_discard() { flags |= DISCARD; }
+  void enable_perspective() { flags |= PERSPECTIVE; }
+  ALWAYS_INLINE bool use_discard() const { return (flags & DISCARD) != 0; }
+  ALWAYS_INLINE bool use_perspective() const {
+    return (flags & PERSPECTIVE) != 0;
+  }
+
+  vec4 gl_FragCoord;
+  vec4 gl_FragColor;
+  vec4 gl_SecondaryFragColor;
+
+  vec2_scalar swgl_StepZW;
+  Bool swgl_IsPixelDiscarded = false;
+  // The current buffer position for committing span output.
+  uint32_t* swgl_OutRGBA8 = nullptr;
+  uint8_t* swgl_OutR8 = nullptr;
+  // The remaining number of pixels in the span.
+  int32_t swgl_SpanLength = 0;
+
+  ALWAYS_INLINE void step_fragcoord(int steps = 4) { gl_FragCoord.x += steps; }
+
+  ALWAYS_INLINE void step_perspective(int steps = 4) {
+    gl_FragCoord.z += swgl_StepZW.x * steps;
+    gl_FragCoord.w += swgl_StepZW.y * steps;
+  }
+
+  template <bool W = false>
+  ALWAYS_INLINE void init_span(const void* interps, const void* step) {
+    (*(W ? init_span_w_func : init_span_func))(this, interps, step);
+  }
+
+  template <bool W = false>
+  ALWAYS_INLINE void run() {
+    (*(W ? run_w_func : run_func))(this);
+  }
+
+  template <bool W = false>
+  ALWAYS_INLINE void skip(int steps = 4) {
+    (*(W ? skip_w_func : skip_func))(this, steps);
+  }
+
+  ALWAYS_INLINE int draw_span(uint32_t* buf, int len) {
+    swgl_OutRGBA8 = buf;
+    swgl_SpanLength = len;
+    return (*draw_span_RGBA8_func)(this);
+  }
+
+  ALWAYS_INLINE bool has_draw_span(uint32_t*) {
+    return draw_span_RGBA8_func != nullptr;
+  }
+
+  ALWAYS_INLINE int draw_span(uint8_t* buf, int len) {
+    swgl_OutR8 = buf;
+    swgl_SpanLength = len;
+    return (*draw_span_R8_func)(this);
+  }
+
+  ALWAYS_INLINE bool has_draw_span(uint8_t*) {
+    return draw_span_R8_func != nullptr;
+  }
+};
+
+}  // namespace glsl
diff --git a/gfx/wr/swgl/src/rasterize.h b/gfx/wr/swgl/src/rasterize.h
new file mode 100644
index 0000000000..9b49f25315
--- /dev/null
+++ b/gfx/wr/swgl/src/rasterize.h
@@ -0,0 +1,1680 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// The SWGL depth buffer is roughly organized as a span buffer where each row
+// of the depth buffer is a list of spans, and each span has a constant depth
+// and a run length (represented by DepthRun). The span from start..start+count
+// is placed directly at that start index in the row's array of runs, so that
+// there is no need to explicitly record the start index at all. This also
+// avoids the need to move items around in the run array to manage insertions
+// since space is implicitly always available for a run between any two
+// pre-existing runs. Linkage from one run to the next is implicitly defined by
+// the count, so if a run exists from start..start+count, the next run will
+// implicitly pick up right at index start+count where that preceding run left
+// off. All of the DepthRun items that are after the head of the run can remain
+// uninitialized until the run needs to be split and a new run needs to start
+// somewhere in between.
+// For uses like perspective-correct rasterization or with a discard mask, a
+// run is not an efficient representation, and it is more beneficial to have
+// a flattened array of individual depth samples that can be masked off easily.
+// To support this case, the first run in a given row's run array may have a
+// zero count, signaling that this entire row is flattened. Critically, the
+// depth and count fields in DepthRun are ordered (endian-dependently) so that
+// the DepthRun struct can be interpreted as a sign-extended int32_t depth. It
+// is then possible to just treat the entire row as an array of int32_t depth
+// samples that can be processed with SIMD comparisons, since the count field
+// behaves as just the sign-extension of the depth field. The count field is
+// limited to 8 bits so that we can support depth values up to 24 bits.
+// When a depth buffer is cleared, each row is initialized to a maximal runs
+// spanning the entire row. In the normal case, the depth buffer will continue
+// to manage itself as a list of runs. If perspective or discard is used for
+// a given row, the row will be converted to the flattened representation to
+// support it, after which it will only ever revert back to runs if the depth
+// buffer is cleared.
+
+// The largest 24-bit depth value supported.
+constexpr uint32_t MAX_DEPTH_VALUE = 0xFFFFFF;
+// The longest 8-bit depth run that is supported, aligned to SIMD chunk size.
+constexpr uint32_t MAX_DEPTH_RUN = 255 & ~3;
+
+struct DepthRun {
+  // Ensure that depth always occupies the LSB and count the MSB so that we
+  // can sign-extend depth just by setting count to zero, marking it flat.
+  // When count is non-zero, then this is interpreted as an actual run and
+  // depth is read in isolation.
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+  uint32_t depth : 24;
+  uint32_t count : 8;
+#else
+  uint32_t count : 8;
+  uint32_t depth : 24;
+#endif
+
+  DepthRun() = default;
+  DepthRun(uint32_t depth, uint8_t count) : depth(depth), count(count) {}
+
+  // If count is zero, this is actually a flat depth sample rather than a run.
+  bool is_flat() const { return !count; }
+
+  // Compare a source depth from rasterization with a stored depth value.
+  template <int FUNC>
+  ALWAYS_INLINE bool compare(uint32_t src) const {
+    switch (FUNC) {
+      case GL_LEQUAL:
+        return src <= depth;
+      case GL_LESS:
+        return src < depth;
+      case GL_ALWAYS:
+        return true;
+      default:
+        assert(false);
+        return false;
+    }
+  }
+};
+
+// Fills runs at the given position with the given depth up to the span width.
+static ALWAYS_INLINE void set_depth_runs(DepthRun* runs, uint32_t depth,
+                                         uint32_t width) {
+  // If the width exceeds the maximum run size, then we need to output clamped
+  // runs first.
+  for (; width >= MAX_DEPTH_RUN;
+       runs += MAX_DEPTH_RUN, width -= MAX_DEPTH_RUN) {
+    *runs = DepthRun(depth, MAX_DEPTH_RUN);
+  }
+  // If there are still any left over samples to fill under the maximum run
+  // size, then output one last run for them.
+  if (width > 0) {
+    *runs = DepthRun(depth, width);
+  }
+}
+
+// A cursor for reading and modifying a row's depth run array. It locates
+// and iterates through a desired span within all the runs, testing if
+// the depth of this span passes or fails the depth test against existing
+// runs. If desired, new runs may be inserted to represent depth occlusion
+// from this span in the run array.
+struct DepthCursor {
+  // Current position of run the cursor has advanced to.
+  DepthRun* cur = nullptr;
+  // The start of the remaining potential samples in the desired span.
+  DepthRun* start = nullptr;
+  // The end of the potential samples in the desired span.
+  DepthRun* end = nullptr;
+
+  DepthCursor() = default;
+
+  // Construct a cursor with runs for a given row's run array and the bounds
+  // of the span we wish to iterate within it.
+  DepthCursor(DepthRun* runs, int num_runs, int span_offset, int span_count)
+      : cur(runs), start(&runs[span_offset]), end(start + span_count) {
+    // This cursor should never iterate over flat runs
+    assert(!runs->is_flat());
+    DepthRun* end_runs = &runs[num_runs];
+    // Clamp end of span to end of row
+    if (end > end_runs) {
+      end = end_runs;
+    }
+    // If the span starts past the end of the row, just advance immediately
+    // to it to signal that we're done.
+    if (start >= end_runs) {
+      cur = end_runs;
+      start = end_runs;
+      return;
+    }
+    // Otherwise, find the first depth run that contains the start of the span.
+    // If the span starts after the given run, then we need to keep searching
+    // through the row to find an appropriate run. The check above already
+    // guaranteed that the span starts within the row's runs, and the search
+    // won't fall off the end.
+    for (;;) {
+      assert(cur < end);
+      DepthRun* next = cur + cur->count;
+      if (start < next) {
+        break;
+      }
+      cur = next;
+    }
+  }
+
+  // The cursor is valid if the current position is at the end or if the run
+  // contains the start position.
+  bool valid() const {
+    return cur >= end || (cur <= start && start < cur + cur->count);
+  }
+
+  // Skip past any initial runs that fail the depth test. If we find a run that
+  // would pass, then return the accumulated length between where we started
+  // and that position. Otherwise, if we fall off the end, return -1 to signal
+  // that there are no more passed runs at the end of this failed region and
+  // so it is safe for the caller to stop processing any more regions in this
+  // row.
+  template <int FUNC>
+  int skip_failed(uint32_t val) {
+    assert(valid());
+    DepthRun* prev = start;
+    while (cur < end) {
+      if (cur->compare<FUNC>(val)) {
+        return start - prev;
+      }
+      cur += cur->count;
+      start = cur;
+    }
+    return -1;
+  }
+
+  // Helper to convert function parameters into template parameters to hoist
+  // some checks out of inner loops.
+  ALWAYS_INLINE int skip_failed(uint32_t val, GLenum func) {
+    switch (func) {
+      case GL_LEQUAL:
+        return skip_failed<GL_LEQUAL>(val);
+      case GL_LESS:
+        return skip_failed<GL_LESS>(val);
+      default:
+        assert(false);
+        return -1;
+    }
+  }
+
+  // Find a region of runs that passes the depth test. It is assumed the caller
+  // has called skip_failed first to skip past any runs that failed the depth
+  // test. This stops when it finds a run that fails the depth test or we fall
+  // off the end of the row. If the write mask is enabled, this will insert runs
+  // to represent this new region that passed the depth test. The length of the
+  // region is returned.
+  template <int FUNC, bool MASK>
+  int check_passed(uint32_t val) {
+    assert(valid());
+    DepthRun* prev = cur;
+    while (cur < end) {
+      if (!cur->compare<FUNC>(val)) {
+        break;
+      }
+      DepthRun* next = cur + cur->count;
+      if (next > end) {
+        if (MASK) {
+          // Chop the current run where the end of the span falls, making a new
+          // run from the end of the span till the next run. The beginning of
+          // the current run will be folded into the run from the start of the
+          // passed region before returning below.
+          *end = DepthRun(cur->depth, next - end);
+        }
+        // If the next run starts past the end, then just advance the current
+        // run to the end to signal that we're now at the end of the row.
+        next = end;
+      }
+      cur = next;
+    }
+    // If we haven't advanced past the start of the span region, then we found
+    // nothing that passed.
+    if (cur <= start) {
+      return 0;
+    }
+    // If 'end' fell within the middle of a passing run, then 'cur' will end up
+    // pointing at the new partial run created at 'end' where the passing run
+    // was split to accommodate starting in the middle. The preceding runs will
+    // be fixed below to properly join with this new split.
+    int passed = cur - start;
+    if (MASK) {
+      // If the search started from a run before the start of the span, then
+      // edit that run to meet up with the start.
+      if (prev < start) {
+        prev->count = start - prev;
+      }
+      // Create a new run for the entirety of the passed samples.
+      set_depth_runs(start, val, passed);
+    }
+    start = cur;
+    return passed;
+  }
+
+  // Helper to convert function parameters into template parameters to hoist
+  // some checks out of inner loops.
+  template <bool MASK>
+  ALWAYS_INLINE int check_passed(uint32_t val, GLenum func) {
+    switch (func) {
+      case GL_LEQUAL:
+        return check_passed<GL_LEQUAL, MASK>(val);
+      case GL_LESS:
+        return check_passed<GL_LESS, MASK>(val);
+      default:
+        assert(false);
+        return 0;
+    }
+  }
+
+  ALWAYS_INLINE int check_passed(uint32_t val, GLenum func, bool mask) {
+    return mask ? check_passed<true>(val, func)
+                : check_passed<false>(val, func);
+  }
+
+  // Fill a region of runs with a given depth value, bypassing any depth test.
+  ALWAYS_INLINE void fill(uint32_t depth) {
+    check_passed<GL_ALWAYS, true>(depth);
+  }
+};
+
+// Initialize a depth texture by setting the first run in each row to encompass
+// the entire row.
+void Texture::init_depth_runs(uint32_t depth) {
+  if (!buf) return;
+  DepthRun* runs = (DepthRun*)buf;
+  for (int y = 0; y < height; y++) {
+    set_depth_runs(runs, depth, width);
+    runs += stride() / sizeof(DepthRun);
+  }
+  set_cleared(true);
+}
+
+// Fill a portion of the run array with flattened depth samples.
+static ALWAYS_INLINE void fill_flat_depth(DepthRun* dst, size_t n,
+                                          uint32_t depth) {
+  fill_n((uint32_t*)dst, n, depth);
+}
+
+// Fills a scissored region of a depth texture with a given depth.
+void Texture::fill_depth_runs(uint32_t depth, const IntRect& scissor) {
+  if (!buf) return;
+  assert(cleared());
+  IntRect bb = bounds().intersection(scissor - offset);
+  DepthRun* runs = (DepthRun*)sample_ptr(0, bb.y0);
+  for (int rows = bb.height(); rows > 0; rows--) {
+    if (bb.width() >= width) {
+      // If the scissor region encompasses the entire row, reset the row to a
+      // single run encompassing the entire row.
+      set_depth_runs(runs, depth, width);
+    } else if (runs->is_flat()) {
+      // If the row is flattened, just directly fill the portion of the row.
+      fill_flat_depth(&runs[bb.x0], bb.width(), depth);
+    } else {
+      // Otherwise, if we are still using runs, then set up a cursor to fill
+      // it with depth runs.
+      DepthCursor(runs, width, bb.x0, bb.width()).fill(depth);
+    }
+    runs += stride() / sizeof(DepthRun);
+  }
+}
+
+using ZMask = I32;
+
+#if USE_SSE2
+#  define ZMASK_NONE_PASSED 0xFFFF
+#  define ZMASK_ALL_PASSED 0
+static inline uint32_t zmask_code(ZMask mask) {
+  return _mm_movemask_epi8(mask);
+}
+#else
+#  define ZMASK_NONE_PASSED 0xFFFFFFFFU
+#  define ZMASK_ALL_PASSED 0
+static inline uint32_t zmask_code(ZMask mask) {
+  return bit_cast<uint32_t>(CONVERT(mask, U8));
+}
+#endif
+
+// Interprets items in the depth buffer as sign-extended 32-bit depth values
+// instead of as runs. Returns a mask that signals which samples in the given
+// chunk passed or failed the depth test with given Z value.
+template <bool DISCARD>
+static ALWAYS_INLINE bool check_depth(I32 src, DepthRun* zbuf, ZMask& outmask,
+                                      int span = 4) {
+  // SSE2 does not support unsigned comparison. So ensure Z value is
+  // sign-extended to int32_t.
+  I32 dest = unaligned_load<I32>(zbuf);
+  // Invert the depth test to check which pixels failed and should be discarded.
+  ZMask mask = ctx->depthfunc == GL_LEQUAL
+                   ?
+                   // GL_LEQUAL: Not(LessEqual) = Greater
+                   ZMask(src > dest)
+                   :
+                   // GL_LESS: Not(Less) = GreaterEqual
+                   ZMask(src >= dest);
+  // Mask off any unused lanes in the span.
+  mask |= ZMask(span) < ZMask{1, 2, 3, 4};
+  if (zmask_code(mask) == ZMASK_NONE_PASSED) {
+    return false;
+  }
+  if (!DISCARD && ctx->depthmask) {
+    unaligned_store(zbuf, (mask & dest) | (~mask & src));
+  }
+  outmask = mask;
+  return true;
+}
+
+static ALWAYS_INLINE I32 packDepth() {
+  return cast(fragment_shader->gl_FragCoord.z * MAX_DEPTH_VALUE);
+}
+
+static ALWAYS_INLINE void discard_depth(I32 src, DepthRun* zbuf, I32 mask) {
+  if (ctx->depthmask) {
+    I32 dest = unaligned_load<I32>(zbuf);
+    mask |= fragment_shader->swgl_IsPixelDiscarded;
+    unaligned_store(zbuf, (mask & dest) | (~mask & src));
+  }
+}
+
+static ALWAYS_INLINE void mask_output(uint32_t* buf, ZMask zmask,
+                                      int span = 4) {
+  WideRGBA8 r = pack_pixels_RGBA8();
+  PackedRGBA8 dst = load_span<PackedRGBA8>(buf, span);
+  if (blend_key) r = blend_pixels(buf, dst, r, span);
+  PackedRGBA8 mask = bit_cast<PackedRGBA8>(zmask);
+  store_span(buf, (mask & dst) | (~mask & pack(r)), span);
+}
+
+template <bool DISCARD>
+static ALWAYS_INLINE void discard_output(uint32_t* buf, int span = 4) {
+  mask_output(buf, fragment_shader->swgl_IsPixelDiscarded, span);
+}
+
+template <>
+ALWAYS_INLINE void discard_output<false>(uint32_t* buf, int span) {
+  WideRGBA8 r = pack_pixels_RGBA8();
+  if (blend_key)
+    r = blend_pixels(buf, load_span<PackedRGBA8>(buf, span), r, span);
+  store_span(buf, pack(r), span);
+}
+
+static ALWAYS_INLINE void mask_output(uint8_t* buf, ZMask zmask, int span = 4) {
+  WideR8 r = pack_pixels_R8();
+  WideR8 dst = unpack(load_span<PackedR8>(buf, span));
+  if (blend_key) r = blend_pixels(buf, dst, r, span);
+  WideR8 mask = packR8(zmask);
+  store_span(buf, pack((mask & dst) | (~mask & r)), span);
+}
+
+template <bool DISCARD>
+static ALWAYS_INLINE void discard_output(uint8_t* buf, int span = 4) {
+  mask_output(buf, fragment_shader->swgl_IsPixelDiscarded, span);
+}
+
+template <>
+ALWAYS_INLINE void discard_output<false>(uint8_t* buf, int span) {
+  WideR8 r = pack_pixels_R8();
+  if (blend_key)
+    r = blend_pixels(buf, unpack(load_span<PackedR8>(buf, span)), r, span);
+  store_span(buf, pack(r), span);
+}
+
+struct ClipRect {
+  float x0;
+  float y0;
+  float x1;
+  float y1;
+
+  explicit ClipRect(const IntRect& i)
+      : x0(i.x0), y0(i.y0), x1(i.x1), y1(i.y1) {}
+  explicit ClipRect(const Texture& t) : ClipRect(ctx->apply_scissor(t)) {
+    // If blending is enabled, set blend_key to reflect the resolved blend
+    // state for the currently drawn primitive.
+    if (ctx->blend) {
+      blend_key = ctx->blend_key;
+      if (swgl_ClipFlags) {
+        // If there is a blend override set, replace the blend key with it.
+        if (swgl_ClipFlags & SWGL_CLIP_FLAG_BLEND_OVERRIDE) {
+          blend_key = swgl_BlendOverride;
+        }
+        // If a clip mask is available, set up blending state to use the clip
+        // mask.
+        if (swgl_ClipFlags & SWGL_CLIP_FLAG_MASK) {
+          assert(swgl_ClipMask->format == TextureFormat::R8);
+          // Constrain the clip mask bounds to always fall within the clip mask.
+          swgl_ClipMaskBounds.intersect(IntRect{0, 0, int(swgl_ClipMask->width),
+                                                int(swgl_ClipMask->height)});
+          // The clip mask offset is relative to the viewport.
+          swgl_ClipMaskOffset += ctx->viewport.origin() - t.offset;
+          // The clip mask bounds are relative to the clip mask offset.
+          swgl_ClipMaskBounds.offset(swgl_ClipMaskOffset);
+          // Finally, constrain the clip rectangle by the clip mask bounds.
+          intersect(swgl_ClipMaskBounds);
+          // Modify the blend key so that it will use the clip mask while
+          // blending.
+          restore_clip_mask();
+        }
+        if (swgl_ClipFlags & SWGL_CLIP_FLAG_AA) {
+          // Modify the blend key so that it will use AA while blending.
+          restore_aa();
+        }
+      }
+    } else {
+      blend_key = BLEND_KEY_NONE;
+      swgl_ClipFlags = 0;
+    }
+  }
+
+  FloatRange x_range() const { return {x0, x1}; }
+
+  void intersect(const IntRect& c) {
+    x0 = max(x0, float(c.x0));
+    y0 = max(y0, float(c.y0));
+    x1 = min(x1, float(c.x1));
+    y1 = min(y1, float(c.y1));
+  }
+
+  template <typename P>
+  void set_clip_mask(int x, int y, P* buf) const {
+    if (swgl_ClipFlags & SWGL_CLIP_FLAG_MASK) {
+      swgl_SpanBuf = buf;
+      swgl_ClipMaskBuf = (uint8_t*)swgl_ClipMask->buf +
+                         (y - swgl_ClipMaskOffset.y) * swgl_ClipMask->stride +
+                         (x - swgl_ClipMaskOffset.x);
+    }
+  }
+
+  template <typename P>
+  bool overlaps(int nump, const P* p) const {
+    // Generate a mask of which side of the clip rect all of a polygon's points
+    // fall inside of. This is a cheap conservative estimate of whether the
+    // bounding box of the polygon might overlap the clip rect, rather than an
+    // exact test that would require multiple slower line intersections.
+    int sides = 0;
+    for (int i = 0; i < nump; i++) {
+      sides |= p[i].x < x1 ? (p[i].x > x0 ? 1 | 2 : 1) : 2;
+      sides |= p[i].y < y1 ? (p[i].y > y0 ? 4 | 8 : 4) : 8;
+    }
+    return sides == 0xF;
+  }
+};
+
+// Given a current X position at the center Y position of a row, return the X
+// position of the left and right intercepts of the row top and bottom.
+template <typename E>
+static ALWAYS_INLINE FloatRange x_intercepts(const E& e) {
+  float rad = 0.5f * abs(e.x_slope());
+  return {e.cur_x() - rad, e.cur_x() + rad};
+}
+
+// Return the AA sub-span corresponding to a given edge. If AA is requested,
+// then this finds the X intercepts with the row clipped into range of the
+// edge and finally conservatively rounds them out. If there is no AA, then
+// it just returns the current rounded X position clipped within bounds.
+template <typename E>
+static ALWAYS_INLINE IntRange aa_edge(const E& e, const FloatRange& bounds) {
+  return e.edgeMask ? bounds.clip(x_intercepts(e)).round_out()
+                    : bounds.clip({e.cur_x(), e.cur_x()}).round();
+}
+
+// Calculate the initial AA coverage as an approximation of the distance from
+// the center of the pixel in the direction of the edge slope. Given an edge
+// (x,y)..(x+dx,y+dy), then the normalized tangent vector along the edge is
+// (dx,dy)/sqrt(dx^2+dy^2). We know that for dy=1 then dx=e.x_slope. We rotate
+// the tangent vector either -90 or +90 degrees to get the edge normal vector,
+// where 'dx=-dy and 'dy=dx. Once normalized by 1/sqrt(dx^2+dy^2), scale into
+// the range of 0..256 so that we can cheaply convert to a fixed-point scale
+// factor. It is assumed that at exactly the pixel center the opacity is half
+// (128) and linearly decreases along the normal vector at 1:1 scale with the
+// slope. While not entirely accurate, this gives a reasonably agreeable looking
+// approximation of AA. For edges on which there is no AA, just force the
+// opacity to maximum (256) with no slope, relying on the span clipping to trim
+// pixels outside the span.
+template <typename E>
+static ALWAYS_INLINE FloatRange aa_dist(const E& e, float dir) {
+  if (e.edgeMask) {
+    float dx = (dir * 256.0f) * inversesqrt(1.0f + e.x_slope() * e.x_slope());
+    return {128.0f + dx * (e.cur_x() - 0.5f), -dx};
+  } else {
+    return {256.0f, 0.0f};
+  }
+}
+
+template <typename P, typename E>
+static ALWAYS_INLINE IntRange aa_span(P* buf, const E& left, const E& right,
+                                      const FloatRange& bounds) {
+  // If there is no AA, just return the span from the rounded left edge X
+  // position to the rounded right edge X position. Clip the span to be within
+  // the valid bounds.
+  if (!(swgl_ClipFlags & SWGL_CLIP_FLAG_AA)) {
+    return bounds.clip({left.cur_x(), right.cur_x()}).round();
+  }
+
+  // Calculate the left and right AA spans along with the coverage distances
+  // and slopes necessary to do blending.
+  IntRange leftAA = aa_edge(left, bounds);
+  FloatRange leftDist = aa_dist(left, -1.0f);
+  IntRange rightAA = aa_edge(right, bounds);
+  FloatRange rightDist = aa_dist(right, 1.0f);
+
+  // Use the pointer into the destination buffer as a status indicator of the
+  // coverage offset. The pointer is calculated so that subtracting it with
+  // the current destination pointer will yield a negative value if the span
+  // is outside the opaque area and otherwise will yield a positive value
+  // above the opaque size. This pointer is stored as a uint8 pointer so that
+  // there are no hidden multiplication instructions and will just return a
+  // 1:1 linear memory address. Thus the size of the opaque region must also
+  // be scaled by the pixel size in bytes.
+  swgl_OpaqueStart = (const uint8_t*)(buf + leftAA.end);
+  swgl_OpaqueSize = max(rightAA.start - leftAA.end - 3, 0) * sizeof(P);
+
+  // Offset the coverage distances by the end of the left AA span, which
+  // corresponds to the opaque start pointer, so that pixels become opaque
+  // immediately after. The distances are also offset for each lane in the
+  // chunk.
+  Float offset = cast(leftAA.end + (I32){0, 1, 2, 3});
+  swgl_LeftAADist = leftDist.start + offset * leftDist.end;
+  swgl_RightAADist = rightDist.start + offset * rightDist.end;
+  swgl_AASlope =
+      (Float){leftDist.end, rightDist.end, 0.0f, 0.0f} / float(sizeof(P));
+
+  // Return the full span width from the start of the left span to the end of
+  // the right span.
+  return {leftAA.start, rightAA.end};
+}
+
+// Calculate the span the user clip distances occupy from the left and right
+// edges at the current row.
+template <typename E>
+static ALWAYS_INLINE IntRange clip_distance_range(const E& left,
+                                                  const E& right) {
+  Float leftClip = get_clip_distances(left.interp);
+  Float rightClip = get_clip_distances(right.interp);
+  // Get the change in clip dist per X step.
+  Float clipStep = (rightClip - leftClip) / (right.cur_x() - left.cur_x());
+  // Find the zero intercepts starting from the left edge.
+  Float clipDist =
+      clamp(left.cur_x() - leftClip * recip(clipStep), 0.0f, 1.0e6f);
+  // Find the distance to the start of the span for any clip distances that
+  // are increasing in value. If the clip distance is constant or decreasing
+  // in value, then check if it starts outside the clip volume.
+  Float start = if_then_else(clipStep > 0.0f, clipDist,
+                             if_then_else(leftClip < 0.0f, 1.0e6f, 0.0f));
+  // Find the distance to the end of the span for any clip distances that are
+  // decreasing in value. If the clip distance is constant or increasing in
+  // value, then check if it ends inside the clip volume.
+  Float end = if_then_else(clipStep < 0.0f, clipDist,
+                           if_then_else(rightClip >= 0.0f, 1.0e6f, 0.0f));
+  // Find the furthest start offset.
+  start = max(start, start.zwxy);
+  // Find the closest end offset.
+  end = min(end, end.zwxy);
+  // Finally, round the offsets to an integer span that can be used to bound
+  // the current span.
+  return FloatRange{max(start.x, start.y), min(end.x, end.y)}.round();
+}
+
+// Converts a run array into a flattened array of depth samples. This just
+// walks through every run and fills the samples with the depth value from
+// the run.
+static void flatten_depth_runs(DepthRun* runs, size_t width) {
+  if (runs->is_flat()) {
+    return;
+  }
+  while (width > 0) {
+    size_t n = runs->count;
+    fill_flat_depth(runs, n, runs->depth);
+    runs += n;
+    width -= n;
+  }
+}
+
+// Helper function for drawing passed depth runs within the depth buffer.
+// Flattened depth (perspective or discard) is not supported.
+template <typename P>
+static ALWAYS_INLINE void draw_depth_span(uint32_t z, P* buf,
+                                          DepthCursor& cursor) {
+  for (;;) {
+    // Get the span that passes the depth test. Assume on entry that
+    // any failed runs have already been skipped.
+    int span = cursor.check_passed(z, ctx->depthfunc, ctx->depthmask);
+    // If nothing passed, since we already skipped passed failed runs
+    // previously, we must have hit the end of the row. Bail out.
+    if (span <= 0) {
+      break;
+    }
+    if (span >= 4) {
+      // If we have a draw specialization, try to process as many 4-pixel
+      // chunks as possible using it.
+      if (fragment_shader->has_draw_span(buf)) {
+        int drawn = fragment_shader->draw_span(buf, span & ~3);
+        buf += drawn;
+        span -= drawn;
+      }
+      // Otherwise, just process each chunk individually.
+      while (span >= 4) {
+        fragment_shader->run();
+        discard_output<false>(buf);
+        buf += 4;
+        span -= 4;
+      }
+    }
+    // If we have a partial chunk left over, we still have to process it as if
+    // it were a full chunk. Mask off only the part of the chunk we want to
+    // use.
+    if (span > 0) {
+      fragment_shader->run();
+      discard_output<false>(buf, span);
+      buf += span;
+    }
+    // Skip past any runs that fail the depth test.
+    int skip = cursor.skip_failed(z, ctx->depthfunc);
+    // If there aren't any, that means we won't encounter any more passing runs
+    // and so it's safe to bail out.
+    if (skip <= 0) {
+      break;
+    }
+    // Advance interpolants for the fragment shader past the skipped region.
+    // If we processed a partial chunk above, we actually advanced the
+    // interpolants a full chunk in the fragment shader's run function. Thus,
+    // we need to first subtract off that 4-pixel chunk and only partially
+    // advance them to that partial chunk before we can add on the rest of the
+    // skips. This is combined with the skip here for efficiency's sake.
+    fragment_shader->skip(skip - (span > 0 ? 4 - span : 0));
+    buf += skip;
+  }
+}
+
+// Draw a simple span in 4-pixel wide chunks, optionally using depth.
+template <bool DISCARD, bool W, typename P, typename Z>
+static ALWAYS_INLINE void draw_span(P* buf, DepthRun* depth, int span, Z z) {
+  if (depth) {
+    // Depth testing is enabled. If perspective is used, Z values will vary
+    // across the span, we use packDepth to generate packed Z values suitable
+    // for depth testing based on current values from gl_FragCoord.z.
+    // Otherwise, for the no-perspective case, we just use the provided Z.
+    // Process 4-pixel chunks first.
+    for (; span >= 4; span -= 4, buf += 4, depth += 4) {
+      I32 zsrc = z();
+      ZMask zmask;
+      if (check_depth<DISCARD>(zsrc, depth, zmask)) {
+        fragment_shader->run<W>();
+        mask_output(buf, zmask);
+        if (DISCARD) discard_depth(zsrc, depth, zmask);
+      } else {
+        fragment_shader->skip<W>();
+      }
+    }
+    // If there are any remaining pixels, do a partial chunk.
+    if (span > 0) {
+      I32 zsrc = z();
+      ZMask zmask;
+      if (check_depth<DISCARD>(zsrc, depth, zmask, span)) {
+        fragment_shader->run<W>();
+        mask_output(buf, zmask, span);
+        if (DISCARD) discard_depth(zsrc, depth, zmask);
+      }
+    }
+  } else {
+    // Process 4-pixel chunks first.
+    for (; span >= 4; span -= 4, buf += 4) {
+      fragment_shader->run<W>();
+      discard_output<DISCARD>(buf);
+    }
+    // If there are any remaining pixels, do a partial chunk.
+    if (span > 0) {
+      fragment_shader->run<W>();
+      discard_output<DISCARD>(buf, span);
+    }
+  }
+}
+
+// Called during rasterization to forcefully clear a row on which delayed clear
+// has been enabled. If we know that we are going to completely overwrite a part
+// of the row, then we only need to clear the row outside of that part. However,
+// if blending or discard is enabled, the values of that underlying part of the
+// row may be used regardless to produce the final rasterization result, so we
+// have to then clear the entire underlying row to prepare it.
+template <typename P>
+static inline void prepare_row(Texture& colortex, int y, int startx, int endx,
+                               bool use_discard, DepthRun* depth,
+                               uint32_t z = 0, DepthCursor* cursor = nullptr) {
+  assert(colortex.delay_clear > 0);
+  // Delayed clear is enabled for the color buffer. Check if needs clear.
+  uint32_t& mask = colortex.cleared_rows[y / 32];
+  if ((mask & (1 << (y & 31))) == 0) {
+    mask |= 1 << (y & 31);
+    colortex.delay_clear--;
+    if (blend_key || use_discard) {
+      // If depth test, blending, or discard is used, old color values
+      // might be sampled, so we need to clear the entire row to fill it.
+      force_clear_row<P>(colortex, y);
+    } else if (depth) {
+      if (depth->is_flat() || !cursor) {
+        // If flat depth is used, we can't cheaply predict if which samples will
+        // pass.
+        force_clear_row<P>(colortex, y);
+      } else {
+        // Otherwise if depth runs are used, see how many samples initially pass
+        // the depth test and only fill the row outside those. The fragment
+        // shader will fill the row within the passed samples.
+        int passed =
+            DepthCursor(*cursor).check_passed<false>(z, ctx->depthfunc);
+        if (startx > 0 || startx + passed < colortex.width) {
+          force_clear_row<P>(colortex, y, startx, startx + passed);
+        }
+      }
+    } else if (startx > 0 || endx < colortex.width) {
+      // Otherwise, we only need to clear the row outside of the span.
+      // The fragment shader will fill the row within the span itself.
+      force_clear_row<P>(colortex, y, startx, endx);
+    }
+  }
+}
+
+// Perpendicular dot-product is the dot-product of a vector with the
+// perpendicular vector of the other, i.e. dot(a, {-b.y, b.x})
+template <typename T>
+static ALWAYS_INLINE auto perpDot(T a, T b) {
+  return a.x * b.y - a.y * b.x;
+}
+
+// Check if the winding of the initial edges is flipped, requiring us to swap
+// the edges to avoid spans having negative lengths. Assume that l0.y == r0.y
+// due to the initial edge scan in draw_quad/perspective_spans.
+template <typename T>
+static ALWAYS_INLINE bool checkIfEdgesFlipped(T l0, T l1, T r0, T r1) {
+  // If the starting point of the left edge is to the right of the starting
+  // point of the right edge, then just assume the edges are flipped. If the
+  // left and right starting points are the same, then check the sign of the
+  // cross-product of the edges to see if the edges are flipped. Otherwise,
+  // if the left starting point is actually just to the left of the right
+  // starting point, then assume no edge flip.
+  return l0.x > r0.x || (l0.x == r0.x && perpDot(l1 - l0, r1 - r0) > 0.0f);
+}
+
+// Draw spans for each row of a given quad (or triangle) with a constant Z
+// value. The quad is assumed convex. It is clipped to fall within the given
+// clip rect. In short, this function rasterizes a quad by first finding a
+// top most starting point and then from there tracing down the left and right
+// sides of this quad until it hits the bottom, outputting a span between the
+// current left and right positions at each row along the way. Points are
+// assumed to be ordered in either CW or CCW to support this, but currently
+// both orders (CW and CCW) are supported and equivalent.
+template <typename P>
+static inline void draw_quad_spans(int nump, Point2D p[4], uint32_t z,
+                                   Interpolants interp_outs[4],
+                                   Texture& colortex, Texture& depthtex,
+                                   const ClipRect& clipRect) {
+  // Only triangles and convex quads supported.
+  assert(nump == 3 || nump == 4);
+
+  Point2D l0, r0, l1, r1;
+  int l0i, r0i, l1i, r1i;
+  {
+    // Find the index of the top-most (smallest Y) point from which
+    // rasterization can start.
+    int top = nump > 3 && p[3].y < p[2].y
+                  ? (p[0].y < p[1].y ? (p[0].y < p[3].y ? 0 : 3)
+                                     : (p[1].y < p[3].y ? 1 : 3))
+                  : (p[0].y < p[1].y ? (p[0].y < p[2].y ? 0 : 2)
+                                     : (p[1].y < p[2].y ? 1 : 2));
+    // Helper to find next index in the points array, walking forward.
+#define NEXT_POINT(idx)   \
+  ({                      \
+    int cur = (idx) + 1;  \
+    cur < nump ? cur : 0; \
+  })
+    // Helper to find the previous index in the points array, walking backward.
+#define PREV_POINT(idx)        \
+  ({                           \
+    int cur = (idx)-1;         \
+    cur >= 0 ? cur : nump - 1; \
+  })
+    // Start looking for "left"-side and "right"-side descending edges starting
+    // from the determined top point.
+    int next = NEXT_POINT(top);
+    int prev = PREV_POINT(top);
+    if (p[top].y == p[next].y) {
+      // If the next point is on the same row as the top, then advance one more
+      // time to the next point and use that as the "left" descending edge.
+      l0i = next;
+      l1i = NEXT_POINT(next);
+      // Assume top and prev form a descending "right" edge, as otherwise this
+      // will be a collapsed polygon and harmlessly bail out down below.
+      r0i = top;
+      r1i = prev;
+    } else if (p[top].y == p[prev].y) {
+      // If the prev point is on the same row as the top, then advance to the
+      // prev again and use that as the "right" descending edge.
+      // Assume top and next form a non-empty descending "left" edge.
+      l0i = top;
+      l1i = next;
+      r0i = prev;
+      r1i = PREV_POINT(prev);
+    } else {
+      // Both next and prev are on distinct rows from top, so both "left" and
+      // "right" edges are non-empty/descending.
+      l0i = r0i = top;
+      l1i = next;
+      r1i = prev;
+    }
+    // Load the points from the indices.
+    l0 = p[l0i];  // Start of left edge
+    r0 = p[r0i];  // End of left edge
+    l1 = p[l1i];  // Start of right edge
+    r1 = p[r1i];  // End of right edge
+    //    debugf("l0: %d(%f,%f), r0: %d(%f,%f) -> l1: %d(%f,%f), r1:
+    //    %d(%f,%f)\n", l0i, l0.x, l0.y, r0i, r0.x, r0.y, l1i, l1.x, l1.y, r1i,
+    //    r1.x, r1.y);
+  }
+
+  struct Edge {
+    float yScale;
+    float xSlope;
+    float x;
+    Interpolants interpSlope;
+    Interpolants interp;
+    bool edgeMask;
+
+    Edge(float y, const Point2D& p0, const Point2D& p1, const Interpolants& i0,
+         const Interpolants& i1, int edgeIndex)
+        :  // Inverse Y scale for slope calculations. Avoid divide on 0-length
+           // edge. Later checks below ensure that Y <= p1.y, or otherwise we
+           // don't use this edge. We just need to guard against Y == p1.y ==
+           // p0.y. In that case, Y - p0.y == 0 and will cancel out the slopes
+           // below, except if yScale is Inf for some reason (or worse, NaN),
+           // which 1/(p1.y-p0.y) might produce if we don't bound it.
+          yScale(1.0f / max(p1.y - p0.y, 1.0f / 256)),
+          // Calculate dX/dY slope
+          xSlope((p1.x - p0.x) * yScale),
+          // Initialize current X based on Y and slope
+          x(p0.x + (y - p0.y) * xSlope),
+          // Calculate change in interpolants per change in Y
+          interpSlope((i1 - i0) * yScale),
+          // Initialize current interpolants based on Y and slope
+          interp(i0 + (y - p0.y) * interpSlope),
+          // Extract the edge mask status for this edge
+          edgeMask((swgl_AAEdgeMask >> edgeIndex) & 1) {}
+
+    void nextRow() {
+      // step current X and interpolants to next row from slope
+      x += xSlope;
+      interp += interpSlope;
+    }
+
+    float cur_x() const { return x; }
+    float x_slope() const { return xSlope; }
+  };
+
+  // Vertex selection above should result in equal left and right start rows
+  assert(l0.y == r0.y);
+  // Find the start y, clip to within the clip rect, and round to row center.
+  // If AA is enabled, round out conservatively rather than round to nearest.
+  float aaRound = swgl_ClipFlags & SWGL_CLIP_FLAG_AA ? 0.0f : 0.5f;
+  float y = floor(max(min(l0.y, clipRect.y1), clipRect.y0) + aaRound) + 0.5f;
+  // Initialize left and right edges from end points and start Y
+  Edge left(y, l0, l1, interp_outs[l0i], interp_outs[l1i], l1i);
+  Edge right(y, r0, r1, interp_outs[r0i], interp_outs[r1i], r0i);
+  // WR does not use backface culling, so check if edges are flipped.
+  bool flipped = checkIfEdgesFlipped(l0, l1, r0, r1);
+  if (flipped) swap(left, right);
+  // Get pointer to color buffer and depth buffer at current Y
+  P* fbuf = (P*)colortex.sample_ptr(0, int(y));
+  DepthRun* fdepth = depthtex.buf != nullptr
+                         ? (DepthRun*)depthtex.sample_ptr(0, int(y))
+                         : nullptr;
+  // Loop along advancing Ys, rasterizing spans at each row
+  float checkY = min(min(l1.y, r1.y), clipRect.y1);
+  // Ensure we don't rasterize out edge bounds
+  FloatRange clipSpan =
+      clipRect.x_range().clip(x_range(l0, l1).merge(x_range(r0, r1)));
+  for (;;) {
+    // Check if we maybe passed edge ends or outside clip rect...
+    if (y > checkY) {
+      // If we're outside the clip rect, we're done.
+      if (y > clipRect.y1) break;
+        // Helper to find the next non-duplicate vertex that doesn't loop back.
+#define STEP_EDGE(y, e0i, e0, e1i, e1, STEP_POINT, end)     \
+  do {                                                      \
+    /* Set new start of edge to be end of old edge */       \
+    e0i = e1i;                                              \
+    e0 = e1;                                                \
+    /* Set new end of edge to next point */                 \
+    e1i = STEP_POINT(e1i);                                  \
+    e1 = p[e1i];                                            \
+    /* If the edge crossed the end, we're done. */          \
+    if (e0i == end) return;                                 \
+    /* Otherwise, it doesn't advance, so keep searching. */ \
+  } while (y > e1.y)
+      // Check if Y advanced past the end of the left edge
+      if (y > l1.y) {
+        // Step to next left edge past Y and reset edge interpolants.
+        STEP_EDGE(y, l0i, l0, l1i, l1, NEXT_POINT, r1i);
+        (flipped ? right : left) =
+            Edge(y, l0, l1, interp_outs[l0i], interp_outs[l1i], l1i);
+      }
+      // Check if Y advanced past the end of the right edge
+      if (y > r1.y) {
+        // Step to next right edge past Y and reset edge interpolants.
+        STEP_EDGE(y, r0i, r0, r1i, r1, PREV_POINT, l1i);
+        (flipped ? left : right) =
+            Edge(y, r0, r1, interp_outs[r0i], interp_outs[r1i], r0i);
+      }
+      // Reset the clip bounds for the new edges
+      clipSpan =
+          clipRect.x_range().clip(x_range(l0, l1).merge(x_range(r0, r1)));
+      // Reset check condition for next time around.
+      checkY = min(ceil(min(l1.y, r1.y) - aaRound), clipRect.y1);
+    }
+
+    // Calculate a potentially AA'd span and check if it is non-empty.
+    IntRange span = aa_span(fbuf, left, right, clipSpan);
+    if (span.len() > 0) {
+      // If user clip planes are enabled, use them to bound the current span.
+      if (vertex_shader->use_clip_distance()) {
+        span = span.intersect(clip_distance_range(left, right));
+        if (span.len() <= 0) goto next_span;
+      }
+      ctx->shaded_rows++;
+      ctx->shaded_pixels += span.len();
+      // Advance color/depth buffer pointers to the start of the span.
+      P* buf = fbuf + span.start;
+      // Check if we will need to use depth-buffer or discard on this span.
+      DepthRun* depth =
+          depthtex.buf != nullptr && depthtex.cleared() ? fdepth : nullptr;
+      DepthCursor cursor;
+      bool use_discard = fragment_shader->use_discard();
+      if (use_discard) {
+        if (depth) {
+          // If we're using discard, we may have to unpredictably drop out some
+          // samples. Flatten the depth run array here to allow this.
+          if (!depth->is_flat()) {
+            flatten_depth_runs(depth, depthtex.width);
+          }
+          // Advance to the depth sample at the start of the span.
+          depth += span.start;
+        }
+      } else if (depth) {
+        if (!depth->is_flat()) {
+          // We're not using discard and the depth row is still organized into
+          // runs. Skip past any runs that would fail the depth test so we
+          // don't have to do any extra work to process them with the rest of
+          // the span.
+          cursor = DepthCursor(depth, depthtex.width, span.start, span.len());
+          int skipped = cursor.skip_failed(z, ctx->depthfunc);
+          // If we fell off the row, that means we couldn't find any passing
+          // runs. We can just skip the entire span.
+          if (skipped < 0) {
+            goto next_span;
+          }
+          buf += skipped;
+          span.start += skipped;
+        } else {
+          // The row is already flattened, so just advance to the span start.
+          depth += span.start;
+        }
+      }
+
+      if (colortex.delay_clear) {
+        // Delayed clear is enabled for the color buffer. Check if needs clear.
+        prepare_row<P>(colortex, int(y), span.start, span.end, use_discard,
+                       depth, z, &cursor);
+      }
+
+      // Initialize fragment shader interpolants to current span position.
+      fragment_shader->gl_FragCoord.x = init_interp(span.start + 0.5f, 1);
+      fragment_shader->gl_FragCoord.y = y;
+      {
+        // Change in interpolants is difference between current right and left
+        // edges per the change in right and left X. If the left and right X
+        // positions are extremely close together, then avoid stepping the
+        // interpolants.
+        float stepScale = 1.0f / (right.x - left.x);
+        if (!isfinite(stepScale)) stepScale = 0.0f;
+        Interpolants step = (right.interp - left.interp) * stepScale;
+        // Advance current interpolants to X at start of span.
+        Interpolants o = left.interp + step * (span.start + 0.5f - left.x);
+        fragment_shader->init_span(&o, &step);
+      }
+      clipRect.set_clip_mask(span.start, y, buf);
+      if (!use_discard) {
+        // Fast paths for the case where fragment discard is not used.
+        if (depth) {
+          // If depth is used, we want to process entire depth runs if depth is
+          // not flattened.
+          if (!depth->is_flat()) {
+            draw_depth_span(z, buf, cursor);
+            goto next_span;
+          }
+          // Otherwise, flattened depth must fall back to the slightly slower
+          // per-chunk depth test path in draw_span below.
+        } else {
+          // Check if the fragment shader has an optimized draw specialization.
+          if (span.len() >= 4 && fragment_shader->has_draw_span(buf)) {
+            // Draw specialization expects 4-pixel chunks.
+            int drawn = fragment_shader->draw_span(buf, span.len() & ~3);
+            buf += drawn;
+            span.start += drawn;
+          }
+        }
+        draw_span<false, false>(buf, depth, span.len(), [=] { return z; });
+      } else {
+        // If discard is used, then use slower fallbacks. This should be rare.
+        // Just needs to work, doesn't need to be too fast yet...
+        draw_span<true, false>(buf, depth, span.len(), [=] { return z; });
+      }
+    }
+  next_span:
+    // Advance Y and edge interpolants to next row.
+    y++;
+    left.nextRow();
+    right.nextRow();
+    // Advance buffers to next row.
+    fbuf += colortex.stride() / sizeof(P);
+    fdepth += depthtex.stride() / sizeof(DepthRun);
+  }
+}
+
+// Draw perspective-correct spans for a convex quad that has been clipped to
+// the near and far Z planes, possibly producing a clipped convex polygon with
+// more than 4 sides. This assumes the Z value will vary across the spans and
+// requires interpolants to factor in W values. This tends to be slower than
+// the simpler 2D draw_quad_spans above, especially since we can't optimize the
+// depth test easily when Z values, and should be used only rarely if possible.
+template <typename P>
+static inline void draw_perspective_spans(int nump, Point3D* p,
+                                          Interpolants* interp_outs,
+                                          Texture& colortex, Texture& depthtex,
+                                          const ClipRect& clipRect) {
+  Point3D l0, r0, l1, r1;
+  int l0i, r0i, l1i, r1i;
+  {
+    // Find the index of the top-most point (smallest Y) from which
+    // rasterization can start.
+    int top = 0;
+    for (int i = 1; i < nump; i++) {
+      if (p[i].y < p[top].y) {
+        top = i;
+      }
+    }
+    // Find left-most top point, the start of the left descending edge.
+    // Advance forward in the points array, searching at most nump points
+    // in case the polygon is flat.
+    l0i = top;
+    for (int i = top + 1; i < nump && p[i].y == p[top].y; i++) {
+      l0i = i;
+    }
+    if (l0i == nump - 1) {
+      for (int i = 0; i <= top && p[i].y == p[top].y; i++) {
+        l0i = i;
+      }
+    }
+    // Find right-most top point, the start of the right descending edge.
+    // Advance backward in the points array, searching at most nump points.
+    r0i = top;
+    for (int i = top - 1; i >= 0 && p[i].y == p[top].y; i--) {
+      r0i = i;
+    }
+    if (r0i == 0) {
+      for (int i = nump - 1; i >= top && p[i].y == p[top].y; i--) {
+        r0i = i;
+      }
+    }
+    // End of left edge is next point after left edge start.
+    l1i = NEXT_POINT(l0i);
+    // End of right edge is prev point after right edge start.
+    r1i = PREV_POINT(r0i);
+    l0 = p[l0i];  // Start of left edge
+    r0 = p[r0i];  // End of left edge
+    l1 = p[l1i];  // Start of right edge
+    r1 = p[r1i];  // End of right edge
+  }
+
+  struct Edge {
+    float yScale;
+    // Current coordinates for edge. Where in the 2D case of draw_quad_spans,
+    // it is enough to just track the X coordinate as we advance along the rows,
+    // for the perspective case we also need to keep track of Z and W. For
+    // simplicity, we just use the full 3D point to track all these coordinates.
+    Point3D pSlope;
+    Point3D p;
+    Interpolants interpSlope;
+    Interpolants interp;
+    bool edgeMask;
+
+    Edge(float y, const Point3D& p0, const Point3D& p1, const Interpolants& i0,
+         const Interpolants& i1, int edgeIndex)
+        :  // Inverse Y scale for slope calculations. Avoid divide on 0-length
+           // edge.
+          yScale(1.0f / max(p1.y - p0.y, 1.0f / 256)),
+          // Calculate dX/dY slope
+          pSlope((p1 - p0) * yScale),
+          // Initialize current coords based on Y and slope
+          p(p0 + (y - p0.y) * pSlope),
+          // Crucially, these interpolants must be scaled by the point's 1/w
+          // value, which allows linear interpolation in a perspective-correct
+          // manner. This will be canceled out inside the fragment shader later.
+          // Calculate change in interpolants per change in Y
+          interpSlope((i1 * p1.w - i0 * p0.w) * yScale),
+          // Initialize current interpolants based on Y and slope
+          interp(i0 * p0.w + (y - p0.y) * interpSlope),
+          // Extract the edge mask status for this edge
+          edgeMask((swgl_AAEdgeMask >> edgeIndex) & 1) {}
+
+    float x() const { return p.x; }
+    vec2_scalar zw() const { return {p.z, p.w}; }
+
+    void nextRow() {
+      // step current coords and interpolants to next row from slope
+      p += pSlope;
+      interp += interpSlope;
+    }
+
+    float cur_x() const { return p.x; }
+    float x_slope() const { return pSlope.x; }
+  };
+
+  // Vertex selection above should result in equal left and right start rows
+  assert(l0.y == r0.y);
+  // Find the start y, clip to within the clip rect, and round to row center.
+  // If AA is enabled, round out conservatively rather than round to nearest.
+  float aaRound = swgl_ClipFlags & SWGL_CLIP_FLAG_AA ? 0.0f : 0.5f;
+  float y = floor(max(min(l0.y, clipRect.y1), clipRect.y0) + aaRound) + 0.5f;
+  // Initialize left and right edges from end points and start Y
+  Edge left(y, l0, l1, interp_outs[l0i], interp_outs[l1i], l1i);
+  Edge right(y, r0, r1, interp_outs[r0i], interp_outs[r1i], r0i);
+  // WR does not use backface culling, so check if edges are flipped.
+  bool flipped = checkIfEdgesFlipped(l0, l1, r0, r1);
+  if (flipped) swap(left, right);
+  // Get pointer to color buffer and depth buffer at current Y
+  P* fbuf = (P*)colortex.sample_ptr(0, int(y));
+  DepthRun* fdepth = depthtex.buf != nullptr
+                         ? (DepthRun*)depthtex.sample_ptr(0, int(y))
+                         : nullptr;
+  // Loop along advancing Ys, rasterizing spans at each row
+  float checkY = min(min(l1.y, r1.y), clipRect.y1);
+  // Ensure we don't rasterize out edge bounds
+  FloatRange clipSpan =
+      clipRect.x_range().clip(x_range(l0, l1).merge(x_range(r0, r1)));
+  for (;;) {
+    // Check if we maybe passed edge ends or outside clip rect...
+    if (y > checkY) {
+      // If we're outside the clip rect, we're done.
+      if (y > clipRect.y1) break;
+      // Check if Y advanced past the end of the left edge
+      if (y > l1.y) {
+        // Step to next left edge past Y and reset edge interpolants.
+        STEP_EDGE(y, l0i, l0, l1i, l1, NEXT_POINT, r1i);
+        (flipped ? right : left) =
+            Edge(y, l0, l1, interp_outs[l0i], interp_outs[l1i], l1i);
+      }
+      // Check if Y advanced past the end of the right edge
+      if (y > r1.y) {
+        // Step to next right edge past Y and reset edge interpolants.
+        STEP_EDGE(y, r0i, r0, r1i, r1, PREV_POINT, l1i);
+        (flipped ? left : right) =
+            Edge(y, r0, r1, interp_outs[r0i], interp_outs[r1i], r0i);
+      }
+      // Reset the clip bounds for the new edges
+      clipSpan =
+          clipRect.x_range().clip(x_range(l0, l1).merge(x_range(r0, r1)));
+      // Reset check condition for next time around.
+      checkY = min(ceil(min(l1.y, r1.y) - aaRound), clipRect.y1);
+    }
+
+    // Calculate a potentially AA'd span and check if it is non-empty.
+    IntRange span = aa_span(fbuf, left, right, clipSpan);
+    if (span.len() > 0) {
+      // If user clip planes are enabled, use them to bound the current span.
+      if (vertex_shader->use_clip_distance()) {
+        span = span.intersect(clip_distance_range(left, right));
+        if (span.len() <= 0) goto next_span;
+      }
+      ctx->shaded_rows++;
+      ctx->shaded_pixels += span.len();
+      // Advance color/depth buffer pointers to the start of the span.
+      P* buf = fbuf + span.start;
+      // Check if the we will need to use depth-buffer or discard on this span.
+      DepthRun* depth =
+          depthtex.buf != nullptr && depthtex.cleared() ? fdepth : nullptr;
+      bool use_discard = fragment_shader->use_discard();
+      if (depth) {
+        // Perspective may cause the depth value to vary on a per sample basis.
+        // Ensure the depth row is flattened to allow testing of individual
+        // samples
+        if (!depth->is_flat()) {
+          flatten_depth_runs(depth, depthtex.width);
+        }
+        // Advance to the depth sample at the start of the span.
+        depth += span.start;
+      }
+      if (colortex.delay_clear) {
+        // Delayed clear is enabled for the color buffer. Check if needs clear.
+        prepare_row<P>(colortex, int(y), span.start, span.end, use_discard,
+                       depth);
+      }
+      // Initialize fragment shader interpolants to current span position.
+      fragment_shader->gl_FragCoord.x = init_interp(span.start + 0.5f, 1);
+      fragment_shader->gl_FragCoord.y = y;
+      {
+        // Calculate the fragment Z and W change per change in fragment X step.
+        // If the left and right X positions are extremely close together, then
+        // avoid stepping.
+        float stepScale = 1.0f / (right.x() - left.x());
+        if (!isfinite(stepScale)) stepScale = 0.0f;
+        vec2_scalar stepZW = (right.zw() - left.zw()) * stepScale;
+        // Calculate initial Z and W values for span start.
+        vec2_scalar zw = left.zw() + stepZW * (span.start + 0.5f - left.x());
+        // Set fragment shader's Z and W values so that it can use them to
+        // cancel out the 1/w baked into the interpolants.
+        fragment_shader->gl_FragCoord.z = init_interp(zw.x, stepZW.x);
+        fragment_shader->gl_FragCoord.w = init_interp(zw.y, stepZW.y);
+        fragment_shader->swgl_StepZW = stepZW;
+        // Change in interpolants is difference between current right and left
+        // edges per the change in right and left X. The left and right
+        // interpolant values were previously multipled by 1/w, so the step and
+        // initial span values take this into account.
+        Interpolants step = (right.interp - left.interp) * stepScale;
+        // Advance current interpolants to X at start of span.
+        Interpolants o = left.interp + step * (span.start + 0.5f - left.x());
+        fragment_shader->init_span<true>(&o, &step);
+      }
+      clipRect.set_clip_mask(span.start, y, buf);
+      if (!use_discard) {
+        // No discard is used. Common case.
+        draw_span<false, true>(buf, depth, span.len(), packDepth);
+      } else {
+        // Discard is used. Rare.
+        draw_span<true, true>(buf, depth, span.len(), packDepth);
+      }
+    }
+  next_span:
+    // Advance Y and edge interpolants to next row.
+    y++;
+    left.nextRow();
+    right.nextRow();
+    // Advance buffers to next row.
+    fbuf += colortex.stride() / sizeof(P);
+    fdepth += depthtex.stride() / sizeof(DepthRun);
+  }
+}
+
+// Clip a primitive against both sides of a view-frustum axis, producing
+// intermediate vertexes with interpolated attributes that will no longer
+// intersect the selected axis planes. This assumes the primitive is convex
+// and should produce at most N+2 vertexes for each invocation (only in the
+// worst case where one point falls outside on each of the opposite sides
+// with the rest of the points inside). The supplied AA edge mask will be
+// modified such that it corresponds to the clipped polygon edges.
+template <XYZW AXIS>
+static int clip_side(int nump, Point3D* p, Interpolants* interp, Point3D* outP,
+                     Interpolants* outInterp, int& outEdgeMask) {
+  // Potential mask bits of which side of a plane a coordinate falls on.
+  enum SIDE { POSITIVE = 1, NEGATIVE = 2 };
+  int numClip = 0;
+  int edgeMask = outEdgeMask;
+  Point3D prev = p[nump - 1];
+  Interpolants prevInterp = interp[nump - 1];
+  float prevCoord = prev.select(AXIS);
+  // Coordinate must satisfy -W <= C <= W. Determine if it is outside, and
+  // if so, remember which side it is outside of. In the special case that W is
+  // negative and |C| < |W|, both -W <= C and C <= W will be false, such that
+  // we must consider the coordinate as falling outside of both plane sides
+  // simultaneously. We test each condition separately and combine them to form
+  // a mask of which plane sides we exceeded. If we neglect to consider both
+  // sides simultaneously, points can erroneously oscillate from one plane side
+  // to the other and exceed the supported maximum number of clip outputs.
+  int prevMask = (prevCoord < -prev.w ? NEGATIVE : 0) |
+                 (prevCoord > prev.w ? POSITIVE : 0);
+  // Loop through points, finding edges that cross the planes by evaluating
+  // the side at each point.
+  outEdgeMask = 0;
+  for (int i = 0; i < nump; i++, edgeMask >>= 1) {
+    Point3D cur = p[i];
+    Interpolants curInterp = interp[i];
+    float curCoord = cur.select(AXIS);
+    int curMask =
+        (curCoord < -cur.w ? NEGATIVE : 0) | (curCoord > cur.w ? POSITIVE : 0);
+    // Check if the previous and current end points are on different sides. If
+    // the masks of sides intersect, then we consider them to be on the same
+    // side. So in the case the masks do not intersect, we then consider them
+    // to fall on different sides.
+    if (!(curMask & prevMask)) {
+      // One of the edge's end points is outside the plane with the other
+      // inside the plane. Find the offset where it crosses the plane and
+      // adjust the point and interpolants to there.
+      if (prevMask) {
+        // Edge that was previously outside crosses inside.
+        // Evaluate plane equation for previous and current end-point
+        // based on previous side and calculate relative offset.
+        if (numClip >= nump + 2) {
+          // If for some reason we produced more vertexes than we support, just
+          // bail out.
+          assert(false);
+          return 0;
+        }
+        // The positive plane is assigned the sign 1, and the negative plane is
+        // assigned -1. If the point falls outside both planes, that means W is
+        // negative. To compensate for this, we must interpolate the coordinate
+        // till W=0, at which point we can choose a single plane side for the
+        // coordinate to fall on since W will no longer be negative. To compute
+        // the coordinate where W=0, we compute K = prev.w / (prev.w-cur.w) and
+        // interpolate C = prev.C + K*(cur.C - prev.C). The sign of C will be
+        // the side of the plane we need to consider. Substituting K into the
+        // comparison C < 0, we can then avoid the division in K with a
+        // cross-multiplication.
+        float prevSide =
+            (prevMask & NEGATIVE) && (!(prevMask & POSITIVE) ||
+                                      prevCoord * (cur.w - prev.w) <
+                                          prev.w * (curCoord - prevCoord))
+                ? -1
+                : 1;
+        float prevDist = prevCoord - prevSide * prev.w;
+        float curDist = curCoord - prevSide * cur.w;
+        // It may happen that after we interpolate by the weight k that due to
+        // floating point rounding we've underestimated the value necessary to
+        // push it over the clipping boundary. Just in case, nudge the mantissa
+        // by a single increment so that we essentially round it up and move it
+        // further inside the clipping boundary. We use nextafter to do this in
+        // a portable fashion.
+        float k = prevDist / (prevDist - curDist);
+        Point3D clipped = prev + (cur - prev) * k;
+        if (prevSide * clipped.select(AXIS) > clipped.w) {
+          k = nextafterf(k, 1.0f);
+          clipped = prev + (cur - prev) * k;
+        }
+        outP[numClip] = clipped;
+        outInterp[numClip] = prevInterp + (curInterp - prevInterp) * k;
+        // Don't output the current edge mask since start point was outside.
+        numClip++;
+      }
+      if (curMask) {
+        // Edge that was previously inside crosses outside.
+        // Evaluate plane equation for previous and current end-point
+        // based on current side and calculate relative offset.
+        if (numClip >= nump + 2) {
+          assert(false);
+          return 0;
+        }
+        // In the case the coordinate falls on both plane sides, the computation
+        // here is much the same as for prevSide, but since we are going from a
+        // previous W that is positive to current W that is negative, then the
+        // sign of cur.w - prev.w will flip in the equation. The resulting sign
+        // is negated to compensate for this.
+        float curSide =
+            (curMask & POSITIVE) && (!(curMask & NEGATIVE) ||
+                                     prevCoord * (cur.w - prev.w) <
+                                         prev.w * (curCoord - prevCoord))
+                ? 1
+                : -1;
+        float prevDist = prevCoord - curSide * prev.w;
+        float curDist = curCoord - curSide * cur.w;
+        // Calculate interpolation weight k and the nudge it inside clipping
+        // boundary with nextafter. Note that since we were previously inside
+        // and now crossing outside, we have to flip the nudge direction for
+        // the weight towards 0 instead of 1.
+        float k = prevDist / (prevDist - curDist);
+        Point3D clipped = prev + (cur - prev) * k;
+        if (curSide * clipped.select(AXIS) > clipped.w) {
+          k = nextafterf(k, 0.0f);
+          clipped = prev + (cur - prev) * k;
+        }
+        outP[numClip] = clipped;
+        outInterp[numClip] = prevInterp + (curInterp - prevInterp) * k;
+        // Output the current edge mask since the end point is inside.
+        outEdgeMask |= (edgeMask & 1) << numClip;
+        numClip++;
+      }
+    }
+    if (!curMask) {
+      // The current end point is inside the plane, so output point unmodified.
+      if (numClip >= nump + 2) {
+        assert(false);
+        return 0;
+      }
+      outP[numClip] = cur;
+      outInterp[numClip] = curInterp;
+      // Output the current edge mask since the end point is inside.
+      outEdgeMask |= (edgeMask & 1) << numClip;
+      numClip++;
+    }
+    prev = cur;
+    prevInterp = curInterp;
+    prevCoord = curCoord;
+    prevMask = curMask;
+  }
+  return numClip;
+}
+
+// Helper function to dispatch to perspective span drawing with points that
+// have already been transformed and clipped.
+static inline void draw_perspective_clipped(int nump, Point3D* p_clip,
+                                            Interpolants* interp_clip,
+                                            Texture& colortex,
+                                            Texture& depthtex) {
+  // If polygon is ouside clip rect, nothing to draw.
+  ClipRect clipRect(colortex);
+  if (!clipRect.overlaps(nump, p_clip)) {
+    return;
+  }
+
+  // Finally draw perspective-correct spans for the polygon.
+  if (colortex.internal_format == GL_RGBA8) {
+    draw_perspective_spans<uint32_t>(nump, p_clip, interp_clip, colortex,
+                                     depthtex, clipRect);
+  } else if (colortex.internal_format == GL_R8) {
+    draw_perspective_spans<uint8_t>(nump, p_clip, interp_clip, colortex,
+                                    depthtex, clipRect);
+  } else {
+    assert(false);
+  }
+}
+
+// Draws a perspective-correct 3D primitive with varying Z value, as opposed
+// to a simple 2D planar primitive with a constant Z value that could be
+// trivially Z rejected. This requires clipping the primitive against the near
+// and far planes to ensure it stays within the valid Z-buffer range. The Z
+// and W of each fragment of the primitives are interpolated across the
+// generated spans and then depth-tested as appropriate.
+// Additionally, vertex attributes must be interpolated with perspective-
+// correction by dividing by W before interpolation, and then later multiplied
+// by W again to produce the final correct attribute value for each fragment.
+// This process is expensive and should be avoided if possible for primitive
+// batches that are known ahead of time to not need perspective-correction.
+static void draw_perspective(int nump, Interpolants interp_outs[4],
+                             Texture& colortex, Texture& depthtex) {
+  // Lines are not supported with perspective.
+  assert(nump >= 3);
+  // Convert output of vertex shader to screen space.
+  vec4 pos = vertex_shader->gl_Position;
+  vec3_scalar scale =
+      vec3_scalar(ctx->viewport.width(), ctx->viewport.height(), 1) * 0.5f;
+  vec3_scalar offset =
+      make_vec3(make_vec2(ctx->viewport.origin() - colortex.offset), 0.0f) +
+      scale;
+  // Verify if point is between near and far planes, rejecting NaN.
+  if (test_all(pos.z > -pos.w && pos.z < pos.w)) {
+    // No points cross the near or far planes, so no clipping required.
+    // Just divide coords by W and convert to viewport. We assume the W
+    // coordinate is non-zero and the reciprocal is finite since it would
+    // otherwise fail the test_none condition.
+    Float w = 1.0f / pos.w;
+    vec3 screen = pos.sel(X, Y, Z) * w * scale + offset;
+    Point3D p[4] = {{screen.x.x, screen.y.x, screen.z.x, w.x},
+                    {screen.x.y, screen.y.y, screen.z.y, w.y},
+                    {screen.x.z, screen.y.z, screen.z.z, w.z},
+                    {screen.x.w, screen.y.w, screen.z.w, w.w}};
+    draw_perspective_clipped(nump, p, interp_outs, colortex, depthtex);
+  } else {
+    // Points cross the near or far planes, so we need to clip.
+    // Start with the original 3 or 4 points...
+    Point3D p[4] = {{pos.x.x, pos.y.x, pos.z.x, pos.w.x},
+                    {pos.x.y, pos.y.y, pos.z.y, pos.w.y},
+                    {pos.x.z, pos.y.z, pos.z.z, pos.w.z},
+                    {pos.x.w, pos.y.w, pos.z.w, pos.w.w}};
+    // Clipping can expand the points by 1 for each of 6 view frustum planes.
+    Point3D p_clip[4 + 6];
+    Interpolants interp_clip[4 + 6];
+    // Clip against near and far Z planes.
+    nump = clip_side<Z>(nump, p, interp_outs, p_clip, interp_clip,
+                        swgl_AAEdgeMask);
+    // If no points are left inside the view frustum, there's nothing to draw.
+    if (nump < 3) {
+      return;
+    }
+    // After clipping against only the near and far planes, we might still
+    // produce points where W = 0, exactly at the camera plane. OpenGL specifies
+    // that for clip coordinates, points must satisfy:
+    //   -W <= X <= W
+    //   -W <= Y <= W
+    //   -W <= Z <= W
+    // When Z = W = 0, this is trivially satisfied, but when we transform and
+    // divide by W below it will produce a divide by 0. Usually we want to only
+    // clip Z to avoid the extra work of clipping X and Y. We can still project
+    // points that fall outside the view frustum X and Y so long as Z is valid.
+    // The span drawing code will then ensure X and Y are clamped to viewport
+    // boundaries. However, in the Z = W = 0 case, sometimes clipping X and Y,
+    // will push W further inside the view frustum so that it is no longer 0,
+    // allowing us to finally proceed to projecting the points to the screen.
+    for (int i = 0; i < nump; i++) {
+      // Found an invalid W, so need to clip against X and Y...
+      if (p_clip[i].w <= 0.0f) {
+        // Ping-pong p_clip -> p_tmp -> p_clip.
+        Point3D p_tmp[4 + 6];
+        Interpolants interp_tmp[4 + 6];
+        nump = clip_side<X>(nump, p_clip, interp_clip, p_tmp, interp_tmp,
+                            swgl_AAEdgeMask);
+        if (nump < 3) return;
+        nump = clip_side<Y>(nump, p_tmp, interp_tmp, p_clip, interp_clip,
+                            swgl_AAEdgeMask);
+        if (nump < 3) return;
+        // After clipping against X and Y planes, there's still points left
+        // to draw, so proceed to trying projection now...
+        break;
+      }
+    }
+    // Divide coords by W and convert to viewport.
+    for (int i = 0; i < nump; i++) {
+      float w = 1.0f / p_clip[i].w;
+      // If the W coord is essentially zero, small enough that division would
+      // result in Inf/NaN, then just set the reciprocal itself to zero so that
+      // the coordinates becomes zeroed out, as the only valid point that
+      // satisfies -W <= X/Y/Z <= W is all zeroes.
+      if (!isfinite(w)) w = 0.0f;
+      p_clip[i] = Point3D(p_clip[i].sel(X, Y, Z) * w * scale + offset, w);
+    }
+    draw_perspective_clipped(nump, p_clip, interp_clip, colortex, depthtex);
+  }
+}
+
+static void draw_quad(int nump, Texture& colortex, Texture& depthtex) {
+  // Run vertex shader once for the primitive's vertices.
+  // Reserve space for 6 sets of interpolants, in case we need to clip against
+  // near and far planes in the perspective case.
+  Interpolants interp_outs[4];
+  swgl_ClipFlags = 0;
+  vertex_shader->run_primitive((char*)interp_outs, sizeof(Interpolants));
+  vec4 pos = vertex_shader->gl_Position;
+  // Check if any vertex W is different from another. If so, use perspective.
+  if (test_any(pos.w != pos.w.x)) {
+    draw_perspective(nump, interp_outs, colortex, depthtex);
+    return;
+  }
+
+  // Convert output of vertex shader to screen space.
+  // Divide coords by W and convert to viewport.
+  float w = 1.0f / pos.w.x;
+  // If the W coord is essentially zero, small enough that division would
+  // result in Inf/NaN, then just set the reciprocal itself to zero so that
+  // the coordinates becomes zeroed out, as the only valid point that
+  // satisfies -W <= X/Y/Z <= W is all zeroes.
+  if (!isfinite(w)) w = 0.0f;
+  vec2 screen = (pos.sel(X, Y) * w + 1) * 0.5f *
+                    vec2_scalar(ctx->viewport.width(), ctx->viewport.height()) +
+                make_vec2(ctx->viewport.origin() - colortex.offset);
+  Point2D p[4] = {{screen.x.x, screen.y.x},
+                  {screen.x.y, screen.y.y},
+                  {screen.x.z, screen.y.z},
+                  {screen.x.w, screen.y.w}};
+
+  // If quad is ouside clip rect, nothing to draw.
+  ClipRect clipRect(colortex);
+  if (!clipRect.overlaps(nump, p)) {
+    return;
+  }
+
+  // Since the quad is assumed 2D, Z is constant across the quad.
+  float screenZ = (pos.z.x * w + 1) * 0.5f;
+  if (screenZ < 0 || screenZ > 1) {
+    // Z values would cross the near or far plane, so just bail.
+    return;
+  }
+  // Since Z doesn't need to be interpolated, just set the fragment shader's
+  // Z and W values here, once and for all fragment shader invocations.
+  uint32_t z = uint32_t(MAX_DEPTH_VALUE * screenZ);
+  fragment_shader->gl_FragCoord.z = screenZ;
+  fragment_shader->gl_FragCoord.w = w;
+
+  // If supplied a line, adjust it so that it is a quad at least 1 pixel thick.
+  // Assume that for a line that all 4 SIMD lanes were actually filled with
+  // vertexes 0, 1, 1, 0.
+  if (nump == 2) {
+    // Nudge Y height to span at least 1 pixel by advancing to next pixel
+    // boundary so that we step at least 1 row when drawing spans.
+    if (int(p[0].y + 0.5f) == int(p[1].y + 0.5f)) {
+      p[2].y = 1 + int(p[1].y + 0.5f);
+      p[3].y = p[2].y;
+      // Nudge X width to span at least 1 pixel so that rounded coords fall on
+      // separate pixels.
+      if (int(p[0].x + 0.5f) == int(p[1].x + 0.5f)) {
+        p[1].x += 1.0f;
+        p[2].x += 1.0f;
+      }
+    } else {
+      // If the line already spans at least 1 row, then assume line is vertical
+      // or diagonal and just needs to be dilated horizontally.
+      p[2].x += 1.0f;
+      p[3].x += 1.0f;
+    }
+    // Pretend that it's a quad now...
+    nump = 4;
+  }
+
+  // Finally draw 2D spans for the quad. Currently only supports drawing to
+  // RGBA8 and R8 color buffers.
+  if (colortex.internal_format == GL_RGBA8) {
+    draw_quad_spans<uint32_t>(nump, p, z, interp_outs, colortex, depthtex,
+                              clipRect);
+  } else if (colortex.internal_format == GL_R8) {
+    draw_quad_spans<uint8_t>(nump, p, z, interp_outs, colortex, depthtex,
+                             clipRect);
+  } else {
+    assert(false);
+  }
+}
+
+template <typename INDEX>
+static inline void draw_elements(GLsizei count, GLsizei instancecount,
+                                 size_t offset, VertexArray& v,
+                                 Texture& colortex, Texture& depthtex) {
+  Buffer& indices_buf = ctx->buffers[v.element_array_buffer_binding];
+  if (!indices_buf.buf || offset >= indices_buf.size) {
+    return;
+  }
+  assert((offset & (sizeof(INDEX) - 1)) == 0);
+  INDEX* indices = (INDEX*)(indices_buf.buf + offset);
+  count = min(count, (GLsizei)((indices_buf.size - offset) / sizeof(INDEX)));
+  // Triangles must be indexed at offsets 0, 1, 2.
+  // Quads must be successive triangles indexed at offsets 0, 1, 2, 2, 1, 3.
+  if (count == 6 && indices[1] == indices[0] + 1 &&
+      indices[2] == indices[0] + 2 && indices[5] == indices[0] + 3) {
+    assert(indices[3] == indices[0] + 2 && indices[4] == indices[0] + 1);
+    // Fast path - since there is only a single quad, we only load per-vertex
+    // attribs once for all instances, as they won't change across instances
+    // or within an instance.
+    vertex_shader->load_attribs(v.attribs, indices[0], 0, 4);
+    draw_quad(4, colortex, depthtex);
+    for (GLsizei instance = 1; instance < instancecount; instance++) {
+      vertex_shader->load_attribs(v.attribs, indices[0], instance, 0);
+      draw_quad(4, colortex, depthtex);
+    }
+  } else {
+    for (GLsizei instance = 0; instance < instancecount; instance++) {
+      for (GLsizei i = 0; i + 3 <= count; i += 3) {
+        if (indices[i + 1] != indices[i] + 1 ||
+            indices[i + 2] != indices[i] + 2) {
+          continue;
+        }
+        if (i + 6 <= count && indices[i + 5] == indices[i] + 3) {
+          assert(indices[i + 3] == indices[i] + 2 &&
+                 indices[i + 4] == indices[i] + 1);
+          vertex_shader->load_attribs(v.attribs, indices[i], instance, 4);
+          draw_quad(4, colortex, depthtex);
+          i += 3;
+        } else {
+          vertex_shader->load_attribs(v.attribs, indices[i], instance, 3);
+          draw_quad(3, colortex, depthtex);
+        }
+      }
+    }
+  }
+}
diff --git a/gfx/wr/swgl/src/swgl_ext.h b/gfx/wr/swgl/src/swgl_ext.h
new file mode 100644
index 0000000000..164ca8d7b1
--- /dev/null
+++ b/gfx/wr/swgl/src/swgl_ext.h
@@ -0,0 +1,1906 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+// When using a solid color with clip masking, the cost of loading the clip mask
+// in the blend stage exceeds the cost of processing the color. Here we handle
+// the entire span of clip mask texture before the blend stage to more
+// efficiently process it and modulate it with color without incurring blend
+// stage overheads.
+template <typename P, typename C>
+static void commit_masked_solid_span(P* buf, C color, int len) {
+  override_clip_mask();
+  uint8_t* mask = get_clip_mask(buf);
+  for (P* end = &buf[len]; buf < end; buf += 4, mask += 4) {
+    commit_span(
+        buf,
+        blend_span(
+            buf,
+            applyColor(expand_mask(buf, unpack(unaligned_load<PackedR8>(mask))),
+                       color)));
+  }
+  restore_clip_mask();
+}
+
+// When using a solid color with anti-aliasing, most of the solid span will not
+// benefit from anti-aliasing in the opaque region. We only want to apply the AA
+// blend stage in the non-opaque start and end of the span where AA is needed.
+template <typename P, typename R>
+static ALWAYS_INLINE void commit_aa_solid_span(P* buf, R r, int len) {
+  if (int start = min((get_aa_opaque_start(buf) + 3) & ~3, len)) {
+    commit_solid_span<true>(buf, r, start);
+    buf += start;
+    len -= start;
+  }
+  if (int opaque = min((get_aa_opaque_size(buf) + 3) & ~3, len)) {
+    override_aa();
+    commit_solid_span<true>(buf, r, opaque);
+    restore_aa();
+    buf += opaque;
+    len -= opaque;
+  }
+  if (len > 0) {
+    commit_solid_span<true>(buf, r, len);
+  }
+}
+
+// Forces a value with vector run-class to have scalar run-class.
+template <typename T>
+static ALWAYS_INLINE auto swgl_forceScalar(T v) -> decltype(force_scalar(v)) {
+  return force_scalar(v);
+}
+
+// Advance all varying inperpolants by a single chunk
+#define swgl_stepInterp() step_interp_inputs()
+
+// Pseudo-intrinsic that accesses the interpolation step for a given varying
+#define swgl_interpStep(v) (interp_step.v)
+
+// Commit an entire span of a solid color. This dispatches to clip-masked and
+// anti-aliased fast-paths as appropriate.
+#define swgl_commitSolid(format, v, n)                                   \
+  do {                                                                   \
+    int len = (n);                                                       \
+    if (blend_key) {                                                     \
+      if (swgl_ClipFlags & SWGL_CLIP_FLAG_MASK) {                        \
+        commit_masked_solid_span(swgl_Out##format,                       \
+                                 packColor(swgl_Out##format, (v)), len); \
+      } else if (swgl_ClipFlags & SWGL_CLIP_FLAG_AA) {                   \
+        commit_aa_solid_span(swgl_Out##format,                           \
+                             pack_span(swgl_Out##format, (v)), len);     \
+      } else {                                                           \
+        commit_solid_span<true>(swgl_Out##format,                        \
+                                pack_span(swgl_Out##format, (v)), len);  \
+      }                                                                  \
+    } else {                                                             \
+      commit_solid_span<false>(swgl_Out##format,                         \
+                               pack_span(swgl_Out##format, (v)), len);   \
+    }                                                                    \
+    swgl_Out##format += len;                                             \
+    swgl_SpanLength -= len;                                              \
+  } while (0)
+#define swgl_commitSolidRGBA8(v) swgl_commitSolid(RGBA8, v, swgl_SpanLength)
+#define swgl_commitSolidR8(v) swgl_commitSolid(R8, v, swgl_SpanLength)
+#define swgl_commitPartialSolidRGBA8(len, v) \
+  swgl_commitSolid(RGBA8, v, min(int(len), swgl_SpanLength))
+#define swgl_commitPartialSolidR8(len, v) \
+  swgl_commitSolid(R8, v, min(int(len), swgl_SpanLength))
+
+#define swgl_commitChunk(format, chunk)                 \
+  do {                                                  \
+    auto r = chunk;                                     \
+    if (blend_key) r = blend_span(swgl_Out##format, r); \
+    commit_span(swgl_Out##format, r);                   \
+    swgl_Out##format += swgl_StepSize;                  \
+    swgl_SpanLength -= swgl_StepSize;                   \
+  } while (0)
+
+// Commit a single chunk of a color
+#define swgl_commitColor(format, color) \
+  swgl_commitChunk(format, pack_pixels_##format(color))
+#define swgl_commitColorRGBA8(color) swgl_commitColor(RGBA8, color)
+#define swgl_commitColorR8(color) swgl_commitColor(R8, color)
+
+template <typename S>
+static ALWAYS_INLINE bool swgl_isTextureLinear(S s) {
+  return s->filter == TextureFilter::LINEAR;
+}
+
+template <typename S>
+static ALWAYS_INLINE bool swgl_isTextureRGBA8(S s) {
+  return s->format == TextureFormat::RGBA8;
+}
+
+template <typename S>
+static ALWAYS_INLINE bool swgl_isTextureR8(S s) {
+  return s->format == TextureFormat::R8;
+}
+
+// Use the default linear quantization scale of 128. This gives 7 bits of
+// fractional precision, which when multiplied with a signed 9 bit value
+// still fits in a 16 bit integer.
+const int swgl_LinearQuantizeScale = 128;
+
+// Quantizes UVs for access into a linear texture.
+template <typename S, typename T>
+static ALWAYS_INLINE T swgl_linearQuantize(S s, T p) {
+  return linearQuantize(p, swgl_LinearQuantizeScale, s);
+}
+
+// Quantizes an interpolation step for UVs for access into a linear texture.
+template <typename S, typename T>
+static ALWAYS_INLINE T swgl_linearQuantizeStep(S s, T p) {
+  return samplerScale(s, p) * swgl_LinearQuantizeScale;
+}
+
+template <typename S>
+static ALWAYS_INLINE WideRGBA8 textureLinearUnpacked(UNUSED uint32_t* buf,
+                                                     S sampler, ivec2 i) {
+  return textureLinearUnpackedRGBA8(sampler, i);
+}
+
+template <typename S>
+static ALWAYS_INLINE WideR8 textureLinearUnpacked(UNUSED uint8_t* buf,
+                                                  S sampler, ivec2 i) {
+  return textureLinearUnpackedR8(sampler, i);
+}
+
+template <typename S>
+static ALWAYS_INLINE bool matchTextureFormat(S s, UNUSED uint32_t* buf) {
+  return swgl_isTextureRGBA8(s);
+}
+
+template <typename S>
+static ALWAYS_INLINE bool matchTextureFormat(S s, UNUSED uint8_t* buf) {
+  return swgl_isTextureR8(s);
+}
+
+// Quantizes the UVs to the 2^7 scale needed for calculating fractional offsets
+// for linear sampling.
+#define LINEAR_QUANTIZE_UV(sampler, uv, uv_step, uv_rect, min_uv, max_uv)     \
+  uv = swgl_linearQuantize(sampler, uv);                                      \
+  vec2_scalar uv_step =                                                       \
+      float(swgl_StepSize) * vec2_scalar{uv.x.y - uv.x.x, uv.y.y - uv.y.x};   \
+  vec2_scalar min_uv = max(                                                   \
+      swgl_linearQuantize(sampler, vec2_scalar{uv_rect.x, uv_rect.y}), 0.0f); \
+  vec2_scalar max_uv =                                                        \
+      max(swgl_linearQuantize(sampler, vec2_scalar{uv_rect.z, uv_rect.w}),    \
+          min_uv);
+
+// Implements the fallback linear filter that can deal with clamping and
+// arbitrary scales.
+template <bool BLEND, typename S, typename C, typename P>
+static P* blendTextureLinearFallback(S sampler, vec2 uv, int span,
+                                     vec2_scalar uv_step, vec2_scalar min_uv,
+                                     vec2_scalar max_uv, C color, P* buf) {
+  for (P* end = buf + span; buf < end; buf += swgl_StepSize, uv += uv_step) {
+    commit_blend_span<BLEND>(
+        buf, applyColor(textureLinearUnpacked(buf, sampler,
+                                              ivec2(clamp(uv, min_uv, max_uv))),
+                        color));
+  }
+  return buf;
+}
+
+static ALWAYS_INLINE U64 castForShuffle(V16<int16_t> r) {
+  return bit_cast<U64>(r);
+}
+static ALWAYS_INLINE U16 castForShuffle(V4<int16_t> r) {
+  return bit_cast<U16>(r);
+}
+
+static ALWAYS_INLINE V16<int16_t> applyFracX(V16<int16_t> r, I16 fracx) {
+  return r * fracx.xxxxyyyyzzzzwwww;
+}
+static ALWAYS_INLINE V4<int16_t> applyFracX(V4<int16_t> r, I16 fracx) {
+  return r * fracx;
+}
+
+// Implements a faster linear filter that works with axis-aligned constant Y but
+// scales less than 1, i.e. upscaling. In this case we can optimize for the
+// constant Y fraction as well as load all chunks from memory in a single tap
+// for each row.
+template <bool BLEND, typename S, typename C, typename P>
+static void blendTextureLinearUpscale(S sampler, vec2 uv, int span,
+                                      vec2_scalar uv_step, vec2_scalar min_uv,
+                                      vec2_scalar max_uv, C color, P* buf) {
+  typedef VectorType<uint8_t, 4 * sizeof(P)> packed_type;
+  typedef VectorType<uint16_t, 4 * sizeof(P)> unpacked_type;
+  typedef VectorType<int16_t, 4 * sizeof(P)> signed_unpacked_type;
+
+  ivec2 i(clamp(uv, min_uv, max_uv));
+  ivec2 frac = i;
+  i >>= 7;
+  P* row0 = (P*)sampler->buf + computeRow(sampler, ivec2_scalar(0, i.y.x));
+  P* row1 = row0 + computeNextRowOffset(sampler, ivec2_scalar(0, i.y.x));
+  I16 fracx = computeFracX(sampler, i, frac);
+  int16_t fracy = computeFracY(frac).x;
+  auto src0 =
+      CONVERT(unaligned_load<packed_type>(&row0[i.x.x]), signed_unpacked_type);
+  auto src1 =
+      CONVERT(unaligned_load<packed_type>(&row1[i.x.x]), signed_unpacked_type);
+  auto src = castForShuffle(src0 + (((src1 - src0) * fracy) >> 7));
+
+  // We attempt to sample ahead by one chunk and interpolate it with the current
+  // one. However, due to the complication of upscaling, we may not necessarily
+  // shift in all the next set of samples.
+  for (P* end = buf + span; buf < end; buf += 4) {
+    uv.x += uv_step.x;
+    I32 ixn = cast(uv.x);
+    I16 fracn = computeFracNoClamp(ixn);
+    ixn >>= 7;
+    auto src0n = CONVERT(unaligned_load<packed_type>(&row0[ixn.x]),
+                         signed_unpacked_type);
+    auto src1n = CONVERT(unaligned_load<packed_type>(&row1[ixn.x]),
+                         signed_unpacked_type);
+    auto srcn = castForShuffle(src0n + (((src1n - src0n) * fracy) >> 7));
+
+    // Since we're upscaling, we know that a source pixel has a larger footprint
+    // than the destination pixel, and thus all the source pixels needed for
+    // this chunk will fall within a single chunk of texture data. However,
+    // since the source pixels don't map 1:1 with destination pixels, we need to
+    // shift the source pixels over based on their offset from the start of the
+    // chunk. This could conceivably be optimized better with usage of PSHUFB or
+    // VTBL instructions However, since PSHUFB requires SSSE3, instead we resort
+    // to masking in the correct pixels to avoid having to index into memory.
+    // For the last sample to interpolate with, we need to potentially shift in
+    // a sample from the next chunk over in the case the samples fill out an
+    // entire chunk.
+    auto shuf = src;
+    auto shufn = SHUFFLE(src, ixn.x == i.x.w ? srcn.yyyy : srcn, 1, 2, 3, 4);
+    if (i.x.y == i.x.x) {
+      shuf = shuf.xxyz;
+      shufn = shufn.xxyz;
+    }
+    if (i.x.z == i.x.y) {
+      shuf = shuf.xyyz;
+      shufn = shufn.xyyz;
+    }
+    if (i.x.w == i.x.z) {
+      shuf = shuf.xyzz;
+      shufn = shufn.xyzz;
+    }
+
+    // Convert back to a signed unpacked type so that we can interpolate the
+    // final result.
+    auto interp = bit_cast<signed_unpacked_type>(shuf);
+    auto interpn = bit_cast<signed_unpacked_type>(shufn);
+    interp += applyFracX(interpn - interp, fracx) >> 7;
+
+    commit_blend_span<BLEND>(
+        buf, applyColor(bit_cast<unpacked_type>(interp), color));
+
+    i.x = ixn;
+    fracx = fracn;
+    src = srcn;
+  }
+}
+
+// This is the fastest variant of the linear filter that still provides
+// filtering. In cases where there is no scaling required, but we have a
+// subpixel offset that forces us to blend in neighboring pixels, we can
+// optimize away most of the memory loads and shuffling that is required by the
+// fallback filter.
+template <bool BLEND, typename S, typename C, typename P>
+static void blendTextureLinearFast(S sampler, vec2 uv, int span,
+                                   vec2_scalar min_uv, vec2_scalar max_uv,
+                                   C color, P* buf) {
+  typedef VectorType<uint8_t, 4 * sizeof(P)> packed_type;
+  typedef VectorType<uint16_t, 4 * sizeof(P)> unpacked_type;
+  typedef VectorType<int16_t, 4 * sizeof(P)> signed_unpacked_type;
+
+  ivec2 i(clamp(uv, min_uv, max_uv));
+  ivec2 frac = i;
+  i >>= 7;
+  P* row0 = (P*)sampler->buf + computeRow(sampler, force_scalar(i));
+  P* row1 = row0 + computeNextRowOffset(sampler, force_scalar(i));
+  int16_t fracx = computeFracX(sampler, i, frac).x;
+  int16_t fracy = computeFracY(frac).x;
+  auto src0 = CONVERT(unaligned_load<packed_type>(row0), signed_unpacked_type);
+  auto src1 = CONVERT(unaligned_load<packed_type>(row1), signed_unpacked_type);
+  auto src = castForShuffle(src0 + (((src1 - src0) * fracy) >> 7));
+
+  // Since there is no scaling, we sample ahead by one chunk and interpolate it
+  // with the current one. We can then reuse this value on the next iteration.
+  for (P* end = buf + span; buf < end; buf += 4) {
+    row0 += 4;
+    row1 += 4;
+    auto src0n =
+        CONVERT(unaligned_load<packed_type>(row0), signed_unpacked_type);
+    auto src1n =
+        CONVERT(unaligned_load<packed_type>(row1), signed_unpacked_type);
+    auto srcn = castForShuffle(src0n + (((src1n - src0n) * fracy) >> 7));
+
+    // For the last sample to interpolate with, we need to potentially shift in
+    // a sample from the next chunk over since the samples fill out an entire
+    // chunk.
+    auto interp = bit_cast<signed_unpacked_type>(src);
+    auto interpn =
+        bit_cast<signed_unpacked_type>(SHUFFLE(src, srcn, 1, 2, 3, 4));
+    interp += ((interpn - interp) * fracx) >> 7;
+
+    commit_blend_span<BLEND>(
+        buf, applyColor(bit_cast<unpacked_type>(interp), color));
+
+    src = srcn;
+  }
+}
+
+// Implements a faster linear filter that works with axis-aligned constant Y but
+// downscaling the texture by half. In this case we can optimize for the
+// constant X/Y fractions and reduction factor while minimizing shuffling.
+template <bool BLEND, typename S, typename C, typename P>
+static NO_INLINE void blendTextureLinearDownscale(S sampler, vec2 uv, int span,
+                                                  vec2_scalar min_uv,
+                                                  vec2_scalar max_uv, C color,
+                                                  P* buf) {
+  typedef VectorType<uint8_t, 4 * sizeof(P)> packed_type;
+  typedef VectorType<uint16_t, 4 * sizeof(P)> unpacked_type;
+  typedef VectorType<int16_t, 4 * sizeof(P)> signed_unpacked_type;
+
+  ivec2 i(clamp(uv, min_uv, max_uv));
+  ivec2 frac = i;
+  i >>= 7;
+  P* row0 = (P*)sampler->buf + computeRow(sampler, force_scalar(i));
+  P* row1 = row0 + computeNextRowOffset(sampler, force_scalar(i));
+  int16_t fracx = computeFracX(sampler, i, frac).x;
+  int16_t fracy = computeFracY(frac).x;
+
+  for (P* end = buf + span; buf < end; buf += 4) {
+    auto src0 =
+        CONVERT(unaligned_load<packed_type>(row0), signed_unpacked_type);
+    auto src1 =
+        CONVERT(unaligned_load<packed_type>(row1), signed_unpacked_type);
+    auto src = castForShuffle(src0 + (((src1 - src0) * fracy) >> 7));
+    row0 += 4;
+    row1 += 4;
+    auto src0n =
+        CONVERT(unaligned_load<packed_type>(row0), signed_unpacked_type);
+    auto src1n =
+        CONVERT(unaligned_load<packed_type>(row1), signed_unpacked_type);
+    auto srcn = castForShuffle(src0n + (((src1n - src0n) * fracy) >> 7));
+    row0 += 4;
+    row1 += 4;
+
+    auto interp =
+        bit_cast<signed_unpacked_type>(SHUFFLE(src, srcn, 0, 2, 4, 6));
+    auto interpn =
+        bit_cast<signed_unpacked_type>(SHUFFLE(src, srcn, 1, 3, 5, 7));
+    interp += ((interpn - interp) * fracx) >> 7;
+
+    commit_blend_span<BLEND>(
+        buf, applyColor(bit_cast<unpacked_type>(interp), color));
+  }
+}
+
+enum LinearFilter {
+  // No linear filter is needed.
+  LINEAR_FILTER_NEAREST = 0,
+  // The most general linear filter that handles clamping and varying scales.
+  LINEAR_FILTER_FALLBACK,
+  // A linear filter optimized for axis-aligned upscaling.
+  LINEAR_FILTER_UPSCALE,
+  // A linear filter with no scaling but with subpixel offset.
+  LINEAR_FILTER_FAST,
+  // A linear filter optimized for 2x axis-aligned downscaling.
+  LINEAR_FILTER_DOWNSCALE
+};
+
+// Dispatches to an appropriate linear filter depending on the selected filter.
+template <bool BLEND, typename S, typename C, typename P>
+static P* blendTextureLinearDispatch(S sampler, vec2 uv, int span,
+                                     vec2_scalar uv_step, vec2_scalar min_uv,
+                                     vec2_scalar max_uv, C color, P* buf,
+                                     LinearFilter filter) {
+  P* end = buf + span;
+  if (filter != LINEAR_FILTER_FALLBACK) {
+    // If we're not using the fallback, then Y is constant across the entire
+    // row. We just need to ensure that we handle any samples that might pull
+    // data from before the start of the row and require clamping.
+    float beforeDist = max(0.0f, min_uv.x) - uv.x.x;
+    if (beforeDist > 0) {
+      int before = clamp(int(ceil(beforeDist / uv_step.x)) * swgl_StepSize, 0,
+                         int(end - buf));
+      buf = blendTextureLinearFallback<BLEND>(sampler, uv, before, uv_step,
+                                              min_uv, max_uv, color, buf);
+      uv.x += (before / swgl_StepSize) * uv_step.x;
+    }
+    // We need to check how many samples we can take from inside the row without
+    // requiring clamping. In case the filter oversamples the row by a step, we
+    // subtract off a step from the width to leave some room.
+    float insideDist =
+        min(max_uv.x, float((int(sampler->width) - swgl_StepSize) *
+                            swgl_LinearQuantizeScale)) -
+        uv.x.x;
+    if (uv_step.x > 0.0f && insideDist >= uv_step.x) {
+      int32_t inside = int(end - buf);
+      if (filter == LINEAR_FILTER_DOWNSCALE) {
+        inside = min(int(insideDist * (0.5f / swgl_LinearQuantizeScale)) &
+                         ~(swgl_StepSize - 1),
+                     inside);
+        if (inside > 0) {
+          blendTextureLinearDownscale<BLEND>(sampler, uv, inside, min_uv,
+                                             max_uv, color, buf);
+          buf += inside;
+          uv.x += (inside / swgl_StepSize) * uv_step.x;
+        }
+      } else if (filter == LINEAR_FILTER_UPSCALE) {
+        inside = min(int(insideDist / uv_step.x) * swgl_StepSize, inside);
+        if (inside > 0) {
+          blendTextureLinearUpscale<BLEND>(sampler, uv, inside, uv_step, min_uv,
+                                           max_uv, color, buf);
+          buf += inside;
+          uv.x += (inside / swgl_StepSize) * uv_step.x;
+        }
+      } else {
+        inside = min(int(insideDist * (1.0f / swgl_LinearQuantizeScale)) &
+                         ~(swgl_StepSize - 1),
+                     inside);
+        if (inside > 0) {
+          blendTextureLinearFast<BLEND>(sampler, uv, inside, min_uv, max_uv,
+                                        color, buf);
+          buf += inside;
+          uv.x += (inside / swgl_StepSize) * uv_step.x;
+        }
+      }
+    }
+  }
+  // If the fallback filter was requested, or if there are any samples left that
+  // may be outside the row and require clamping, then handle that with here.
+  if (buf < end) {
+    buf = blendTextureLinearFallback<BLEND>(
+        sampler, uv, int(end - buf), uv_step, min_uv, max_uv, color, buf);
+  }
+  return buf;
+}
+
+// Helper function to quantize UVs for linear filtering before dispatch
+template <bool BLEND, typename S, typename C, typename P>
+static inline int blendTextureLinear(S sampler, vec2 uv, int span,
+                                     const vec4_scalar& uv_rect, C color,
+                                     P* buf, LinearFilter filter) {
+  if (!matchTextureFormat(sampler, buf)) {
+    return 0;
+  }
+  LINEAR_QUANTIZE_UV(sampler, uv, uv_step, uv_rect, min_uv, max_uv);
+  blendTextureLinearDispatch<BLEND>(sampler, uv, span, uv_step, min_uv, max_uv,
+                                    color, buf, filter);
+  return span;
+}
+
+// Samples an axis-aligned span of on a single row of a texture using 1:1
+// nearest filtering. Sampling is constrained to only fall within the given UV
+// bounds. This requires a pointer to the destination buffer. An optional color
+// modulus can be supplied.
+template <bool BLEND, typename S, typename C, typename P>
+static int blendTextureNearestFast(S sampler, vec2 uv, int span,
+                                   const vec4_scalar& uv_rect, C color,
+                                   P* buf) {
+  if (!matchTextureFormat(sampler, buf)) {
+    return 0;
+  }
+
+  typedef VectorType<uint8_t, 4 * sizeof(P)> packed_type;
+
+  ivec2_scalar i = make_ivec2(samplerScale(sampler, force_scalar(uv)));
+  ivec2_scalar minUV =
+      make_ivec2(samplerScale(sampler, vec2_scalar{uv_rect.x, uv_rect.y}));
+  ivec2_scalar maxUV =
+      make_ivec2(samplerScale(sampler, vec2_scalar{uv_rect.z, uv_rect.w}));
+
+  // Calculate the row pointer within the buffer, clamping to within valid row
+  // bounds.
+  P* row =
+      &((P*)sampler
+            ->buf)[clampCoord(clamp(i.y, minUV.y, maxUV.y), sampler->height) *
+                   sampler->stride];
+  // Find clamped X bounds within the row.
+  int minX = clamp(minUV.x, 0, sampler->width - 1);
+  int maxX = clamp(maxUV.x, minX, sampler->width - 1);
+  int curX = i.x;
+  int endX = i.x + span;
+  // If we need to start sampling below the valid sample bounds, then we need to
+  // fill this section with a constant clamped sample.
+  if (curX < minX) {
+    int n = min(minX, endX) - curX;
+    auto src =
+        applyColor(unpack(bit_cast<packed_type>(V4<P>(row[minX]))), color);
+    commit_solid_span<BLEND>(buf, src, n);
+    buf += n;
+    curX += n;
+  }
+  // Here we only deal with valid samples within the sample bounds. No clamping
+  // should occur here within these inner loops.
+  int n = max(min(maxX + 1, endX) - curX, 0);
+  // Try to process as many chunks as possible with full loads and stores.
+  for (int end = curX + (n & ~3); curX < end; curX += 4, buf += 4) {
+    auto src = applyColor(unaligned_load<packed_type>(&row[curX]), color);
+    commit_blend_span<BLEND>(buf, src);
+  }
+  n &= 3;
+  // If we have any leftover samples after processing chunks, use partial loads
+  // and stores.
+  if (n > 0) {
+    auto src = applyColor(partial_load_span<packed_type>(&row[curX], n), color);
+    commit_blend_span<BLEND>(buf, src, n);
+    buf += n;
+    curX += n;
+  }
+  // If we still have samples left above the valid sample bounds, then we again
+  // need to fill this section with a constant clamped sample.
+  if (curX < endX) {
+    auto src =
+        applyColor(unpack(bit_cast<packed_type>(V4<P>(row[maxX]))), color);
+    commit_solid_span<BLEND>(buf, src, endX - curX);
+  }
+  return span;
+}
+
+// We need to verify that the pixel step reasonably approximates stepping by a
+// single texel for every pixel we need to reproduce. Try to ensure that the
+// margin of error is no more than approximately 2^-7. Also, we check here if
+// the scaling can be quantized for acceleration.
+template <typename T>
+static ALWAYS_INLINE int spanNeedsScale(int span, T P) {
+  span &= ~(128 - 1);
+  span += 128;
+  int scaled = round((P.x.y - P.x.x) * span);
+  return scaled != span ? (scaled == span * 2 ? 2 : 1) : 0;
+}
+
+// Helper function to decide whether we can safely apply 1:1 nearest filtering
+// without diverging too much from the linear filter.
+template <typename S, typename T>
+static inline LinearFilter needsTextureLinear(S sampler, T P, int span) {
+  // If each row is not wide enough for linear filtering, then just use nearest
+  // filtering.
+  if (sampler->width < 2) {
+    return LINEAR_FILTER_NEAREST;
+  }
+  // First verify if the row Y doesn't change across samples
+  if (P.y.x != P.y.y) {
+    return LINEAR_FILTER_FALLBACK;
+  }
+  P = samplerScale(sampler, P);
+  if (int scale = spanNeedsScale(span, P)) {
+    // If the source region is not flipped and smaller than the destination,
+    // then we can use the upscaling filter since row Y is constant.
+    return P.x.x < P.x.y && P.x.y - P.x.x <= 1
+               ? LINEAR_FILTER_UPSCALE
+               : (scale == 2 ? LINEAR_FILTER_DOWNSCALE
+                             : LINEAR_FILTER_FALLBACK);
+  }
+  // Also verify that we're reasonably close to the center of a texel
+  // so that it doesn't look that much different than if a linear filter
+  // was used.
+  if ((int(P.x.x * 4.0f + 0.5f) & 3) != 2 ||
+      (int(P.y.x * 4.0f + 0.5f) & 3) != 2) {
+    // The source and destination regions are the same, but there is a
+    // significant subpixel offset. We can use a faster linear filter to deal
+    // with the offset in this case.
+    return LINEAR_FILTER_FAST;
+  }
+  // Otherwise, we have a constant 1:1 step and we're stepping reasonably close
+  // to the center of each pixel, so it's safe to disable the linear filter and
+  // use nearest.
+  return LINEAR_FILTER_NEAREST;
+}
+
+// Commit an entire span with linear filtering
+#define swgl_commitTextureLinear(format, s, p, uv_rect, color, n)              \
+  do {                                                                         \
+    auto packed_color = packColor(swgl_Out##format, color);                    \
+    int len = (n);                                                             \
+    int drawn = 0;                                                             \
+    if (LinearFilter filter = needsTextureLinear(s, p, len)) {                 \
+      if (blend_key) {                                                         \
+        drawn = blendTextureLinear<true>(s, p, len, uv_rect, packed_color,     \
+                                         swgl_Out##format, filter);            \
+      } else {                                                                 \
+        drawn = blendTextureLinear<false>(s, p, len, uv_rect, packed_color,    \
+                                          swgl_Out##format, filter);           \
+      }                                                                        \
+    } else if (blend_key) {                                                    \
+      drawn = blendTextureNearestFast<true>(s, p, len, uv_rect, packed_color,  \
+                                            swgl_Out##format);                 \
+    } else {                                                                   \
+      drawn = blendTextureNearestFast<false>(s, p, len, uv_rect, packed_color, \
+                                             swgl_Out##format);                \
+    }                                                                          \
+    swgl_Out##format += drawn;                                                 \
+    swgl_SpanLength -= drawn;                                                  \
+  } while (0)
+#define swgl_commitTextureLinearRGBA8(s, p, uv_rect) \
+  swgl_commitTextureLinear(RGBA8, s, p, uv_rect, NoColor(), swgl_SpanLength)
+#define swgl_commitTextureLinearR8(s, p, uv_rect) \
+  swgl_commitTextureLinear(R8, s, p, uv_rect, NoColor(), swgl_SpanLength)
+
+// Commit a partial span with linear filtering, optionally inverting the color
+#define swgl_commitPartialTextureLinearR8(len, s, p, uv_rect) \
+  swgl_commitTextureLinear(R8, s, p, uv_rect, NoColor(),      \
+                           min(int(len), swgl_SpanLength))
+#define swgl_commitPartialTextureLinearInvertR8(len, s, p, uv_rect) \
+  swgl_commitTextureLinear(R8, s, p, uv_rect, InvertColor(),        \
+                           min(int(len), swgl_SpanLength))
+
+// Commit an entire span with linear filtering that is scaled by a color
+#define swgl_commitTextureLinearColorRGBA8(s, p, uv_rect, color) \
+  swgl_commitTextureLinear(RGBA8, s, p, uv_rect, color, swgl_SpanLength)
+#define swgl_commitTextureLinearColorR8(s, p, uv_rect, color) \
+  swgl_commitTextureLinear(R8, s, p, uv_rect, color, swgl_SpanLength)
+
+// Helper function that samples from an R8 texture while expanding it to support
+// a differing framebuffer format.
+template <bool BLEND, typename S, typename C, typename P>
+static inline int blendTextureLinearR8(S sampler, vec2 uv, int span,
+                                       const vec4_scalar& uv_rect, C color,
+                                       P* buf) {
+  if (!swgl_isTextureR8(sampler) || sampler->width < 2) {
+    return 0;
+  }
+  LINEAR_QUANTIZE_UV(sampler, uv, uv_step, uv_rect, min_uv, max_uv);
+  for (P* end = buf + span; buf < end; buf += swgl_StepSize, uv += uv_step) {
+    commit_blend_span<BLEND>(
+        buf, applyColor(expand_mask(buf, textureLinearUnpackedR8(
+                                             sampler,
+                                             ivec2(clamp(uv, min_uv, max_uv)))),
+                        color));
+  }
+  return span;
+}
+
+// Commit an entire span with linear filtering while expanding from R8 to RGBA8
+#define swgl_commitTextureLinearColorR8ToRGBA8(s, p, uv_rect, color)      \
+  do {                                                                    \
+    auto packed_color = packColor(swgl_OutRGBA8, color);                  \
+    int drawn = 0;                                                        \
+    if (blend_key) {                                                      \
+      drawn = blendTextureLinearR8<true>(s, p, swgl_SpanLength, uv_rect,  \
+                                         packed_color, swgl_OutRGBA8);    \
+    } else {                                                              \
+      drawn = blendTextureLinearR8<false>(s, p, swgl_SpanLength, uv_rect, \
+                                          packed_color, swgl_OutRGBA8);   \
+    }                                                                     \
+    swgl_OutRGBA8 += drawn;                                               \
+    swgl_SpanLength -= drawn;                                             \
+  } while (0)
+#define swgl_commitTextureLinearR8ToRGBA8(s, p, uv_rect) \
+  swgl_commitTextureLinearColorR8ToRGBA8(s, p, uv_rect, NoColor())
+
+// Compute repeating UVs, possibly constrained by tile repeat limits
+static inline vec2 tileRepeatUV(vec2 uv, const vec2_scalar& tile_repeat) {
+  if (tile_repeat.x > 0.0f) {
+    // Clamp to a number slightly less than the tile repeat limit so that
+    // it results in a number close to but not equal to 1 after fract().
+    // This avoids fract() yielding 0 if the limit was left as whole integer.
+    uv = clamp(uv, vec2_scalar(0.0f), tile_repeat - 1.0e-6f);
+  }
+  return fract(uv);
+}
+
+// Compute the number of non-repeating steps before we need to potentially
+// repeat the UVs.
+static inline int computeNoRepeatSteps(Float uv, float uv_step,
+                                       float tile_repeat, int steps) {
+  if (uv.w < uv.x) {
+    // Ensure the UV taps are ordered low to high.
+    uv = uv.wzyx;
+  }
+  // Check if the samples cross the boundary of the next whole integer or the
+  // tile repeat limit, whichever is lower.
+  float limit = floor(uv.x) + 1.0f;
+  if (tile_repeat > 0.0f) {
+    limit = min(limit, tile_repeat);
+  }
+  return uv.x >= 0.0f && uv.w < limit
+             ? (uv_step != 0.0f
+                    ? int(clamp((limit - uv.x) / uv_step, 0.0f, float(steps)))
+                    : steps)
+             : 0;
+}
+
+// Blends an entire span of texture with linear filtering and repeating UVs.
+template <bool BLEND, typename S, typename C, typename P>
+static int blendTextureLinearRepeat(S sampler, vec2 uv, int span,
+                                    const vec2_scalar& tile_repeat,
+                                    const vec4_scalar& uv_repeat,
+                                    const vec4_scalar& uv_rect, C color,
+                                    P* buf) {
+  if (!matchTextureFormat(sampler, buf)) {
+    return 0;
+  }
+  vec2_scalar uv_scale = {uv_repeat.z - uv_repeat.x, uv_repeat.w - uv_repeat.y};
+  vec2_scalar uv_offset = {uv_repeat.x, uv_repeat.y};
+  // Choose a linear filter to use for no-repeat sub-spans
+  LinearFilter filter =
+      needsTextureLinear(sampler, uv * uv_scale + uv_offset, span);
+  // We need to step UVs unscaled and unquantized so that we can modulo them
+  // with fract. We use uv_scale and uv_offset to map them into the correct
+  // range.
+  vec2_scalar uv_step =
+      float(swgl_StepSize) * vec2_scalar{uv.x.y - uv.x.x, uv.y.y - uv.y.x};
+  uv_scale = swgl_linearQuantizeStep(sampler, uv_scale);
+  uv_offset = swgl_linearQuantize(sampler, uv_offset);
+  vec2_scalar min_uv = max(
+      swgl_linearQuantize(sampler, vec2_scalar{uv_rect.x, uv_rect.y}), 0.0f);
+  vec2_scalar max_uv = max(
+      swgl_linearQuantize(sampler, vec2_scalar{uv_rect.z, uv_rect.w}), min_uv);
+  for (P* end = buf + span; buf < end; buf += swgl_StepSize, uv += uv_step) {
+    int steps = int(end - buf) / swgl_StepSize;
+    // Find the sub-span before UVs repeat to avoid expensive repeat math
+    steps = computeNoRepeatSteps(uv.x, uv_step.x, tile_repeat.x, steps);
+    if (steps > 0) {
+      steps = computeNoRepeatSteps(uv.y, uv_step.y, tile_repeat.y, steps);
+      if (steps > 0) {
+        buf = blendTextureLinearDispatch<BLEND>(
+            sampler, fract(uv) * uv_scale + uv_offset, steps * swgl_StepSize,
+            uv_step * uv_scale, min_uv, max_uv, color, buf, filter);
+        if (buf >= end) {
+          break;
+        }
+        uv += steps * uv_step;
+      }
+    }
+    // UVs might repeat within this step, so explicitly compute repeated UVs
+    vec2 repeated_uv = clamp(
+        tileRepeatUV(uv, tile_repeat) * uv_scale + uv_offset, min_uv, max_uv);
+    commit_blend_span<BLEND>(
+        buf, applyColor(textureLinearUnpacked(buf, sampler, ivec2(repeated_uv)),
+                        color));
+  }
+  return span;
+}
+
+// Commit an entire span with linear filtering and repeating UVs
+#define swgl_commitTextureLinearRepeat(format, s, p, tile_repeat, uv_repeat,   \
+                                       uv_rect, color)                         \
+  do {                                                                         \
+    auto packed_color = packColor(swgl_Out##format, color);                    \
+    int drawn = 0;                                                             \
+    if (blend_key) {                                                           \
+      drawn = blendTextureLinearRepeat<true>(s, p, swgl_SpanLength,            \
+                                             tile_repeat, uv_repeat, uv_rect,  \
+                                             packed_color, swgl_Out##format);  \
+    } else {                                                                   \
+      drawn = blendTextureLinearRepeat<false>(s, p, swgl_SpanLength,           \
+                                              tile_repeat, uv_repeat, uv_rect, \
+                                              packed_color, swgl_Out##format); \
+    }                                                                          \
+    swgl_Out##format += drawn;                                                 \
+    swgl_SpanLength -= drawn;                                                  \
+  } while (0)
+#define swgl_commitTextureLinearRepeatRGBA8(s, p, tile_repeat, uv_repeat,      \
+                                            uv_rect)                           \
+  swgl_commitTextureLinearRepeat(RGBA8, s, p, tile_repeat, uv_repeat, uv_rect, \
+                                 NoColor())
+#define swgl_commitTextureLinearRepeatColorRGBA8(s, p, tile_repeat, uv_repeat, \
+                                                 uv_rect, color)               \
+  swgl_commitTextureLinearRepeat(RGBA8, s, p, tile_repeat, uv_repeat, uv_rect, \
+                                 color)
+
+template <typename S>
+static ALWAYS_INLINE PackedRGBA8 textureNearestPacked(UNUSED uint32_t* buf,
+                                                      S sampler, ivec2 i) {
+  return textureNearestPackedRGBA8(sampler, i);
+}
+
+// Blends an entire span of texture with nearest filtering and either
+// repeated or clamped UVs.
+template <bool BLEND, bool REPEAT, typename S, typename C, typename P>
+static int blendTextureNearestRepeat(S sampler, vec2 uv, int span,
+                                     const vec2_scalar& tile_repeat,
+                                     const vec4_scalar& uv_rect, C color,
+                                     P* buf) {
+  if (!matchTextureFormat(sampler, buf)) {
+    return 0;
+  }
+  if (!REPEAT) {
+    // If clamping, then we step pre-scaled to the sampler. For repeat modes,
+    // this will be accomplished via uv_scale instead.
+    uv = samplerScale(sampler, uv);
+  }
+  vec2_scalar uv_step =
+      float(swgl_StepSize) * vec2_scalar{uv.x.y - uv.x.x, uv.y.y - uv.y.x};
+  vec2_scalar min_uv = samplerScale(sampler, vec2_scalar{uv_rect.x, uv_rect.y});
+  vec2_scalar max_uv = samplerScale(sampler, vec2_scalar{uv_rect.z, uv_rect.w});
+  vec2_scalar uv_scale = max_uv - min_uv;
+  // If the effective sampling area of this texture is only a single pixel, then
+  // treat it as a solid span. For repeat modes, the bounds are specified on
+  // pixel boundaries, whereas for clamp modes, bounds are on pixel centers, so
+  // the test varies depending on which. If the sample range on an axis is
+  // greater than one pixel, we can still check if we don't move far enough from
+  // the pixel center on that axis to hit the next pixel.
+  if ((int(min_uv.x) + (REPEAT ? 1 : 0) >= int(max_uv.x) ||
+       (abs(uv_step.x) * span * (REPEAT ? uv_scale.x : 1.0f) < 0.5f)) &&
+      (int(min_uv.y) + (REPEAT ? 1 : 0) >= int(max_uv.y) ||
+       (abs(uv_step.y) * span * (REPEAT ? uv_scale.y : 1.0f) < 0.5f))) {
+    vec2 repeated_uv = REPEAT
+                           ? tileRepeatUV(uv, tile_repeat) * uv_scale + min_uv
+                           : clamp(uv, min_uv, max_uv);
+    commit_solid_span<BLEND>(buf,
+                             applyColor(unpack(textureNearestPacked(
+                                            buf, sampler, ivec2(repeated_uv))),
+                                        color),
+                             span);
+  } else {
+    for (P* end = buf + span; buf < end; buf += swgl_StepSize, uv += uv_step) {
+      if (REPEAT) {
+        int steps = int(end - buf) / swgl_StepSize;
+        // Find the sub-span before UVs repeat to avoid expensive repeat math
+        steps = computeNoRepeatSteps(uv.x, uv_step.x, tile_repeat.x, steps);
+        if (steps > 0) {
+          steps = computeNoRepeatSteps(uv.y, uv_step.y, tile_repeat.y, steps);
+          if (steps > 0) {
+            vec2 inside_uv = fract(uv) * uv_scale + min_uv;
+            vec2 inside_step = uv_step * uv_scale;
+            for (P* outside = &buf[steps * swgl_StepSize]; buf < outside;
+                 buf += swgl_StepSize, inside_uv += inside_step) {
+              commit_blend_span<BLEND>(
+                  buf, applyColor(
+                           textureNearestPacked(buf, sampler, ivec2(inside_uv)),
+                           color));
+            }
+            if (buf >= end) {
+              break;
+            }
+            uv += steps * uv_step;
+          }
+        }
+      }
+
+      // UVs might repeat within this step, so explicitly compute repeated UVs
+      vec2 repeated_uv = REPEAT
+                             ? tileRepeatUV(uv, tile_repeat) * uv_scale + min_uv
+                             : clamp(uv, min_uv, max_uv);
+      commit_blend_span<BLEND>(
+          buf,
+          applyColor(textureNearestPacked(buf, sampler, ivec2(repeated_uv)),
+                     color));
+    }
+  }
+  return span;
+}
+
+// Determine if we can use the fast nearest filter for the given nearest mode.
+// If the Y coordinate varies more than half a pixel over
+// the span (which might cause the texel to alias to the next one), or the span
+// needs X scaling, then we have to use the fallback.
+template <typename S, typename T>
+static ALWAYS_INLINE bool needsNearestFallback(S sampler, T P, int span) {
+  P = samplerScale(sampler, P);
+  return (P.y.y - P.y.x) * span >= 0.5f || spanNeedsScale(span, P);
+}
+
+// Commit an entire span with nearest filtering and either clamped or repeating
+// UVs
+#define swgl_commitTextureNearest(format, s, p, uv_rect, color)               \
+  do {                                                                        \
+    auto packed_color = packColor(swgl_Out##format, color);                   \
+    int drawn = 0;                                                            \
+    if (needsNearestFallback(s, p, swgl_SpanLength)) {                        \
+      if (blend_key) {                                                        \
+        drawn = blendTextureNearestRepeat<true, false>(                       \
+            s, p, swgl_SpanLength, 0.0f, uv_rect, packed_color,               \
+            swgl_Out##format);                                                \
+      } else {                                                                \
+        drawn = blendTextureNearestRepeat<false, false>(                      \
+            s, p, swgl_SpanLength, 0.0f, uv_rect, packed_color,               \
+            swgl_Out##format);                                                \
+      }                                                                       \
+    } else if (blend_key) {                                                   \
+      drawn = blendTextureNearestFast<true>(s, p, swgl_SpanLength, uv_rect,   \
+                                            packed_color, swgl_Out##format);  \
+    } else {                                                                  \
+      drawn = blendTextureNearestFast<false>(s, p, swgl_SpanLength, uv_rect,  \
+                                             packed_color, swgl_Out##format); \
+    }                                                                         \
+    swgl_Out##format += drawn;                                                \
+    swgl_SpanLength -= drawn;                                                 \
+  } while (0)
+#define swgl_commitTextureNearestRGBA8(s, p, uv_rect) \
+  swgl_commitTextureNearest(RGBA8, s, p, uv_rect, NoColor())
+#define swgl_commitTextureNearestColorRGBA8(s, p, uv_rect, color) \
+  swgl_commitTextureNearest(RGBA8, s, p, uv_rect, color)
+
+#define swgl_commitTextureNearestRepeat(format, s, p, tile_repeat, uv_rect, \
+                                        color)                              \
+  do {                                                                      \
+    auto packed_color = packColor(swgl_Out##format, color);                 \
+    int drawn = 0;                                                          \
+    if (blend_key) {                                                        \
+      drawn = blendTextureNearestRepeat<true, true>(                        \
+          s, p, swgl_SpanLength, tile_repeat, uv_rect, packed_color,        \
+          swgl_Out##format);                                                \
+    } else {                                                                \
+      drawn = blendTextureNearestRepeat<false, true>(                       \
+          s, p, swgl_SpanLength, tile_repeat, uv_rect, packed_color,        \
+          swgl_Out##format);                                                \
+    }                                                                       \
+    swgl_Out##format += drawn;                                              \
+    swgl_SpanLength -= drawn;                                               \
+  } while (0)
+#define swgl_commitTextureNearestRepeatRGBA8(s, p, tile_repeat, uv_repeat, \
+                                             uv_rect)                      \
+  swgl_commitTextureNearestRepeat(RGBA8, s, p, tile_repeat, uv_repeat,     \
+                                  NoColor())
+#define swgl_commitTextureNearestRepeatColorRGBA8(s, p, tile_repeat,         \
+                                                  uv_repeat, uv_rect, color) \
+  swgl_commitTextureNearestRepeat(RGBA8, s, p, tile_repeat, uv_repeat, color)
+
+// Commit an entire span of texture with filtering determined by sampler state.
+#define swgl_commitTexture(format, s, ...)               \
+  do {                                                   \
+    if (s->filter == TextureFilter::LINEAR) {            \
+      swgl_commitTextureLinear##format(s, __VA_ARGS__);  \
+    } else {                                             \
+      swgl_commitTextureNearest##format(s, __VA_ARGS__); \
+    }                                                    \
+  } while (0)
+#define swgl_commitTextureRGBA8(...) swgl_commitTexture(RGBA8, __VA_ARGS__)
+#define swgl_commitTextureColorRGBA8(...) \
+  swgl_commitTexture(ColorRGBA8, __VA_ARGS__)
+#define swgl_commitTextureRepeatRGBA8(...) \
+  swgl_commitTexture(RepeatRGBA8, __VA_ARGS__)
+#define swgl_commitTextureRepeatColorRGBA8(...) \
+  swgl_commitTexture(RepeatColorRGBA8, __VA_ARGS__)
+
+// Commit an entire span of a separable pass of a Gaussian blur that falls
+// within the given radius scaled by supplied coefficients, clamped to uv_rect
+// bounds.
+template <bool BLEND, typename S, typename P>
+static int blendGaussianBlur(S sampler, vec2 uv, const vec4_scalar& uv_rect,
+                             P* buf, int span, bool hori, int radius,
+                             vec2_scalar coeffs) {
+  if (!matchTextureFormat(sampler, buf)) {
+    return 0;
+  }
+  vec2_scalar size = {float(sampler->width), float(sampler->height)};
+  ivec2_scalar curUV = make_ivec2(force_scalar(uv) * size);
+  ivec4_scalar bounds = make_ivec4(uv_rect * make_vec4(size, size));
+  int startX = curUV.x;
+  int endX = min(min(bounds.z, curUV.x + span), int(size.x));
+  if (hori) {
+    for (; curUV.x + swgl_StepSize <= endX;
+         buf += swgl_StepSize, curUV.x += swgl_StepSize) {
+      commit_blend_span<BLEND>(
+          buf, gaussianBlurHorizontal<P>(sampler, curUV, bounds.x, bounds.z,
+                                         radius, coeffs.x, coeffs.y));
+    }
+  } else {
+    for (; curUV.x + swgl_StepSize <= endX;
+         buf += swgl_StepSize, curUV.x += swgl_StepSize) {
+      commit_blend_span<BLEND>(
+          buf, gaussianBlurVertical<P>(sampler, curUV, bounds.y, bounds.w,
+                                       radius, coeffs.x, coeffs.y));
+    }
+  }
+  return curUV.x - startX;
+}
+
+#define swgl_commitGaussianBlur(format, s, p, uv_rect, hori, radius, coeffs)   \
+  do {                                                                         \
+    int drawn = 0;                                                             \
+    if (blend_key) {                                                           \
+      drawn = blendGaussianBlur<true>(s, p, uv_rect, swgl_Out##format,         \
+                                      swgl_SpanLength, hori, radius, coeffs);  \
+    } else {                                                                   \
+      drawn = blendGaussianBlur<false>(s, p, uv_rect, swgl_Out##format,        \
+                                       swgl_SpanLength, hori, radius, coeffs); \
+    }                                                                          \
+    swgl_Out##format += drawn;                                                 \
+    swgl_SpanLength -= drawn;                                                  \
+  } while (0)
+#define swgl_commitGaussianBlurRGBA8(s, p, uv_rect, hori, radius, coeffs) \
+  swgl_commitGaussianBlur(RGBA8, s, p, uv_rect, hori, radius, coeffs)
+#define swgl_commitGaussianBlurR8(s, p, uv_rect, hori, radius, coeffs) \
+  swgl_commitGaussianBlur(R8, s, p, uv_rect, hori, radius, coeffs)
+
+// Convert and pack planar YUV samples to RGB output using a color space
+static ALWAYS_INLINE PackedRGBA8 convertYUV(const YUVMatrix& rgb_from_ycbcr,
+                                            U16 y, U16 u, U16 v) {
+  auto yy = V8<int16_t>(zip(y, y));
+  auto uv = V8<int16_t>(zip(u, v));
+  return rgb_from_ycbcr.convert(yy, uv);
+}
+
+// Helper functions to sample from planar YUV textures before converting to RGB
+template <typename S0>
+static ALWAYS_INLINE PackedRGBA8 sampleYUV(S0 sampler0, ivec2 uv0,
+                                           const YUVMatrix& rgb_from_ycbcr,
+                                           UNUSED int rescaleFactor) {
+  switch (sampler0->format) {
+    case TextureFormat::RGBA8: {
+      auto planar = textureLinearPlanarRGBA8(sampler0, uv0);
+      return convertYUV(rgb_from_ycbcr, highHalf(planar.rg), lowHalf(planar.rg),
+                        lowHalf(planar.ba));
+    }
+    case TextureFormat::YUV422: {
+      auto planar = textureLinearPlanarYUV422(sampler0, uv0);
+      return convertYUV(rgb_from_ycbcr, planar.y, planar.u, planar.v);
+    }
+    default:
+      assert(false);
+      return PackedRGBA8(0);
+  }
+}
+
+template <bool BLEND, typename S0, typename P, typename C = NoColor>
+static int blendYUV(P* buf, int span, S0 sampler0, vec2 uv0,
+                    const vec4_scalar& uv_rect0, const vec3_scalar& ycbcr_bias,
+                    const mat3_scalar& rgb_from_debiased_ycbcr,
+                    int rescaleFactor, C color = C()) {
+  if (!swgl_isTextureLinear(sampler0)) {
+    return 0;
+  }
+  LINEAR_QUANTIZE_UV(sampler0, uv0, uv_step0, uv_rect0, min_uv0, max_uv0);
+  const auto rgb_from_ycbcr =
+      YUVMatrix::From(ycbcr_bias, rgb_from_debiased_ycbcr, rescaleFactor);
+  auto c = packColor(buf, color);
+  auto* end = buf + span;
+  for (; buf < end; buf += swgl_StepSize, uv0 += uv_step0) {
+    commit_blend_span<BLEND>(
+        buf, applyColor(sampleYUV(sampler0, ivec2(clamp(uv0, min_uv0, max_uv0)),
+                                  rgb_from_ycbcr, rescaleFactor),
+                        c));
+  }
+  return span;
+}
+
+template <typename S0, typename S1>
+static ALWAYS_INLINE PackedRGBA8 sampleYUV(S0 sampler0, ivec2 uv0, S1 sampler1,
+                                           ivec2 uv1,
+                                           const YUVMatrix& rgb_from_ycbcr,
+                                           int rescaleFactor) {
+  switch (sampler1->format) {
+    case TextureFormat::RG8: {
+      assert(sampler0->format == TextureFormat::R8);
+      auto y = textureLinearUnpackedR8(sampler0, uv0);
+      auto planar = textureLinearPlanarRG8(sampler1, uv1);
+      return convertYUV(rgb_from_ycbcr, y, lowHalf(planar.rg),
+                        highHalf(planar.rg));
+    }
+    case TextureFormat::RGBA8: {
+      assert(sampler0->format == TextureFormat::R8);
+      auto y = textureLinearUnpackedR8(sampler0, uv0);
+      auto planar = textureLinearPlanarRGBA8(sampler1, uv1);
+      return convertYUV(rgb_from_ycbcr, y, lowHalf(planar.ba),
+                        highHalf(planar.rg));
+    }
+    case TextureFormat::RG16: {
+      assert(sampler0->format == TextureFormat::R16);
+      // The rescaling factor represents how many bits to add to renormalize the
+      // texture to 16 bits, and so the color depth is actually 16 minus the
+      // rescaling factor.
+      // Need to right shift the sample by the amount of bits over 8 it
+      // occupies. On output from textureLinearUnpackedR16, we have lost 1 bit
+      // of precision at the low end already, hence 1 is subtracted from the
+      // color depth.
+      int colorDepth = 16 - rescaleFactor;
+      int rescaleBits = (colorDepth - 1) - 8;
+      auto y = textureLinearUnpackedR16(sampler0, uv0) >> rescaleBits;
+      auto uv = textureLinearUnpackedRG16(sampler1, uv1) >> rescaleBits;
+      return rgb_from_ycbcr.convert(zip(y, y), uv);
+    }
+    default:
+      assert(false);
+      return PackedRGBA8(0);
+  }
+}
+
+template <bool BLEND, typename S0, typename S1, typename P,
+          typename C = NoColor>
+static int blendYUV(P* buf, int span, S0 sampler0, vec2 uv0,
+                    const vec4_scalar& uv_rect0, S1 sampler1, vec2 uv1,
+                    const vec4_scalar& uv_rect1, const vec3_scalar& ycbcr_bias,
+                    const mat3_scalar& rgb_from_debiased_ycbcr,
+                    int rescaleFactor, C color = C()) {
+  if (!swgl_isTextureLinear(sampler0) || !swgl_isTextureLinear(sampler1)) {
+    return 0;
+  }
+  LINEAR_QUANTIZE_UV(sampler0, uv0, uv_step0, uv_rect0, min_uv0, max_uv0);
+  LINEAR_QUANTIZE_UV(sampler1, uv1, uv_step1, uv_rect1, min_uv1, max_uv1);
+  const auto rgb_from_ycbcr =
+      YUVMatrix::From(ycbcr_bias, rgb_from_debiased_ycbcr, rescaleFactor);
+  auto c = packColor(buf, color);
+  auto* end = buf + span;
+  for (; buf < end; buf += swgl_StepSize, uv0 += uv_step0, uv1 += uv_step1) {
+    commit_blend_span<BLEND>(
+        buf, applyColor(sampleYUV(sampler0, ivec2(clamp(uv0, min_uv0, max_uv0)),
+                                  sampler1, ivec2(clamp(uv1, min_uv1, max_uv1)),
+                                  rgb_from_ycbcr, rescaleFactor),
+                        c));
+  }
+  return span;
+}
+
+template <typename S0, typename S1, typename S2>
+static ALWAYS_INLINE PackedRGBA8 sampleYUV(S0 sampler0, ivec2 uv0, S1 sampler1,
+                                           ivec2 uv1, S2 sampler2, ivec2 uv2,
+                                           const YUVMatrix& rgb_from_ycbcr,
+                                           int rescaleFactor) {
+  assert(sampler0->format == sampler1->format &&
+         sampler0->format == sampler2->format);
+  switch (sampler0->format) {
+    case TextureFormat::R8: {
+      auto y = textureLinearUnpackedR8(sampler0, uv0);
+      auto u = textureLinearUnpackedR8(sampler1, uv1);
+      auto v = textureLinearUnpackedR8(sampler2, uv2);
+      return convertYUV(rgb_from_ycbcr, y, u, v);
+    }
+    case TextureFormat::R16: {
+      // The rescaling factor represents how many bits to add to renormalize the
+      // texture to 16 bits, and so the color depth is actually 16 minus the
+      // rescaling factor.
+      // Need to right shift the sample by the amount of bits over 8 it
+      // occupies. On output from textureLinearUnpackedR16, we have lost 1 bit
+      // of precision at the low end already, hence 1 is subtracted from the
+      // color depth.
+      int colorDepth = 16 - rescaleFactor;
+      int rescaleBits = (colorDepth - 1) - 8;
+      auto y = textureLinearUnpackedR16(sampler0, uv0) >> rescaleBits;
+      auto u = textureLinearUnpackedR16(sampler1, uv1) >> rescaleBits;
+      auto v = textureLinearUnpackedR16(sampler2, uv2) >> rescaleBits;
+      return convertYUV(rgb_from_ycbcr, U16(y), U16(u), U16(v));
+    }
+    default:
+      assert(false);
+      return PackedRGBA8(0);
+  }
+}
+
+// Fallback helper for when we can't specifically accelerate YUV with
+// composition.
+template <bool BLEND, typename S0, typename S1, typename S2, typename P,
+          typename C>
+static void blendYUVFallback(P* buf, int span, S0 sampler0, vec2 uv0,
+                             vec2_scalar uv_step0, vec2_scalar min_uv0,
+                             vec2_scalar max_uv0, S1 sampler1, vec2 uv1,
+                             vec2_scalar uv_step1, vec2_scalar min_uv1,
+                             vec2_scalar max_uv1, S2 sampler2, vec2 uv2,
+                             vec2_scalar uv_step2, vec2_scalar min_uv2,
+                             vec2_scalar max_uv2, const vec3_scalar& ycbcr_bias,
+                             const mat3_scalar& rgb_from_debiased_ycbcr,
+                             int rescaleFactor, C color) {
+  const auto rgb_from_ycbcr =
+      YUVMatrix::From(ycbcr_bias, rgb_from_debiased_ycbcr, rescaleFactor);
+  for (auto* end = buf + span; buf < end; buf += swgl_StepSize, uv0 += uv_step0,
+             uv1 += uv_step1, uv2 += uv_step2) {
+    commit_blend_span<BLEND>(
+        buf, applyColor(sampleYUV(sampler0, ivec2(clamp(uv0, min_uv0, max_uv0)),
+                                  sampler1, ivec2(clamp(uv1, min_uv1, max_uv1)),
+                                  sampler2, ivec2(clamp(uv2, min_uv2, max_uv2)),
+                                  rgb_from_ycbcr, rescaleFactor),
+                        color));
+  }
+}
+
+template <bool BLEND, typename S0, typename S1, typename S2, typename P,
+          typename C = NoColor>
+static int blendYUV(P* buf, int span, S0 sampler0, vec2 uv0,
+                    const vec4_scalar& uv_rect0, S1 sampler1, vec2 uv1,
+                    const vec4_scalar& uv_rect1, S2 sampler2, vec2 uv2,
+                    const vec4_scalar& uv_rect2, const vec3_scalar& ycbcr_bias,
+                    const mat3_scalar& rgb_from_debiased_ycbcr,
+                    int rescaleFactor, C color = C()) {
+  if (!swgl_isTextureLinear(sampler0) || !swgl_isTextureLinear(sampler1) ||
+      !swgl_isTextureLinear(sampler2)) {
+    return 0;
+  }
+  LINEAR_QUANTIZE_UV(sampler0, uv0, uv_step0, uv_rect0, min_uv0, max_uv0);
+  LINEAR_QUANTIZE_UV(sampler1, uv1, uv_step1, uv_rect1, min_uv1, max_uv1);
+  LINEAR_QUANTIZE_UV(sampler2, uv2, uv_step2, uv_rect2, min_uv2, max_uv2);
+  auto c = packColor(buf, color);
+  blendYUVFallback<BLEND>(buf, span, sampler0, uv0, uv_step0, min_uv0, max_uv0,
+                          sampler1, uv1, uv_step1, min_uv1, max_uv1, sampler2,
+                          uv2, uv_step2, min_uv2, max_uv2, ycbcr_bias,
+                          rgb_from_debiased_ycbcr, rescaleFactor, c);
+  return span;
+}
+
+// A variant of the blendYUV that attempts to reuse the inner loops from the
+// CompositeYUV infrastructure. CompositeYUV imposes stricter requirements on
+// the source data, which in turn allows it to be much faster than blendYUV.
+// At a minimum, we need to ensure that we are outputting to a BGRA8 framebuffer
+// and that no color scaling is applied, which we can accomplish via template
+// specialization. We need to further validate inside that texture formats
+// and dimensions are sane for video and that the video is axis-aligned before
+// acceleration can proceed.
+template <bool BLEND>
+static int blendYUV(uint32_t* buf, int span, sampler2DRect sampler0, vec2 uv0,
+                    const vec4_scalar& uv_rect0, sampler2DRect sampler1,
+                    vec2 uv1, const vec4_scalar& uv_rect1,
+                    sampler2DRect sampler2, vec2 uv2,
+                    const vec4_scalar& uv_rect2, const vec3_scalar& ycbcr_bias,
+                    const mat3_scalar& rgb_from_debiased_ycbcr,
+                    int rescaleFactor, NoColor noColor = NoColor()) {
+  if (!swgl_isTextureLinear(sampler0) || !swgl_isTextureLinear(sampler1) ||
+      !swgl_isTextureLinear(sampler2)) {
+    return 0;
+  }
+  LINEAR_QUANTIZE_UV(sampler0, uv0, uv_step0, uv_rect0, min_uv0, max_uv0);
+  LINEAR_QUANTIZE_UV(sampler1, uv1, uv_step1, uv_rect1, min_uv1, max_uv1);
+  LINEAR_QUANTIZE_UV(sampler2, uv2, uv_step2, uv_rect2, min_uv2, max_uv2);
+  auto* end = buf + span;
+  // CompositeYUV imposes further restrictions on the source textures, such that
+  // the the Y/U/V samplers must all have a matching format, the U/V samplers
+  // must have matching sizes and sample coordinates, and there must be no
+  // change in row across the entire span.
+  if (sampler0->format == sampler1->format &&
+      sampler1->format == sampler2->format &&
+      sampler1->width == sampler2->width &&
+      sampler1->height == sampler2->height && uv_step0.y == 0 &&
+      uv_step0.x > 0 && uv_step1.y == 0 && uv_step1.x > 0 &&
+      uv_step1 == uv_step2 && uv1.x.x == uv2.x.x && uv1.y.x == uv2.y.x) {
+    // CompositeYUV does not support a clamp rect, so we must take care to
+    // advance till we're inside the bounds of the clamp rect.
+    int outside = min(int(ceil(max((min_uv0.x - uv0.x.x) / uv_step0.x,
+                                   (min_uv1.x - uv1.x.x) / uv_step1.x))),
+                      (end - buf) / swgl_StepSize);
+    if (outside > 0) {
+      blendYUVFallback<BLEND>(buf, outside * swgl_StepSize, sampler0, uv0,
+                              uv_step0, min_uv0, max_uv0, sampler1, uv1,
+                              uv_step1, min_uv1, max_uv1, sampler2, uv2,
+                              uv_step2, min_uv2, max_uv2, ycbcr_bias,
+                              rgb_from_debiased_ycbcr, rescaleFactor, noColor);
+      buf += outside * swgl_StepSize;
+      uv0.x += outside * uv_step0.x;
+      uv1.x += outside * uv_step1.x;
+      uv2.x += outside * uv_step2.x;
+    }
+    // Find the amount of chunks inside the clamp rect before we hit the
+    // maximum. If there are any chunks inside, we can finally dispatch to
+    // CompositeYUV.
+    int inside = min(int(min((max_uv0.x - uv0.x.x) / uv_step0.x,
+                             (max_uv1.x - uv1.x.x) / uv_step1.x)),
+                     (end - buf) / swgl_StepSize);
+    if (inside > 0) {
+      // We need the color depth, which is relative to the texture format and
+      // rescale factor.
+      int colorDepth =
+          (sampler0->format == TextureFormat::R16 ? 16 : 8) - rescaleFactor;
+      // Finally, call the inner loop of CompositeYUV.
+      const auto rgb_from_ycbcr =
+          YUVMatrix::From(ycbcr_bias, rgb_from_debiased_ycbcr, rescaleFactor);
+      linear_row_yuv<BLEND>(
+          buf, inside * swgl_StepSize, sampler0, force_scalar(uv0),
+          uv_step0.x / swgl_StepSize, sampler1, sampler2, force_scalar(uv1),
+          uv_step1.x / swgl_StepSize, colorDepth, rgb_from_ycbcr);
+      // Now that we're done, advance past the processed inside portion.
+      buf += inside * swgl_StepSize;
+      uv0.x += inside * uv_step0.x;
+      uv1.x += inside * uv_step1.x;
+      uv2.x += inside * uv_step2.x;
+    }
+  }
+  // We either got here because we have some samples outside the clamp rect, or
+  // because some of the preconditions were not satisfied. Process whatever is
+  // left of the span.
+  blendYUVFallback<BLEND>(buf, end - buf, sampler0, uv0, uv_step0, min_uv0,
+                          max_uv0, sampler1, uv1, uv_step1, min_uv1, max_uv1,
+                          sampler2, uv2, uv_step2, min_uv2, max_uv2, ycbcr_bias,
+                          rgb_from_debiased_ycbcr, rescaleFactor, noColor);
+  return span;
+}
+
+// Commit a single chunk of a YUV surface represented by multiple planar
+// textures. This requires a color space specifier selecting how to convert
+// from YUV to RGB output. In the case of HDR formats, a rescaling factor
+// selects how many bits of precision must be utilized on conversion. See the
+// sampleYUV dispatcher functions for the various supported plane
+// configurations this intrinsic accepts.
+#define swgl_commitTextureLinearYUV(...)                                    \
+  do {                                                                      \
+    int drawn = 0;                                                          \
+    if (blend_key) {                                                        \
+      drawn = blendYUV<true>(swgl_OutRGBA8, swgl_SpanLength, __VA_ARGS__);  \
+    } else {                                                                \
+      drawn = blendYUV<false>(swgl_OutRGBA8, swgl_SpanLength, __VA_ARGS__); \
+    }                                                                       \
+    swgl_OutRGBA8 += drawn;                                                 \
+    swgl_SpanLength -= drawn;                                               \
+  } while (0)
+
+// Commit a single chunk of a YUV surface scaled by a color.
+#define swgl_commitTextureLinearColorYUV(...) \
+  swgl_commitTextureLinearYUV(__VA_ARGS__)
+
+// Each gradient stops entry is a pair of RGBA32F start color and end step.
+struct GradientStops {
+  Float startColor;
+  union {
+    Float stepColor;
+    vec4_scalar stepData;
+  };
+
+  // Whether this gradient entry can be merged with an adjacent entry. The
+  // step will be equal with the adjacent step if and only if they can be
+  // merged, or rather, that the stops are actually part of a single larger
+  // gradient.
+  bool can_merge(const GradientStops& next) const {
+    return stepData == next.stepData;
+  }
+
+  // Get the interpolated color within the entry based on the offset from its
+  // start.
+  Float interpolate(float offset) const {
+    return startColor + stepColor * offset;
+  }
+
+  // Get the end color of the entry where interpolation stops.
+  Float end_color() const { return startColor + stepColor; }
+};
+
+// Checks if a gradient table of the specified size exists at the UV coords of
+// the address within an RGBA32F texture. If so, a linear address within the
+// texture is returned that may be used to sample the gradient table later. If
+// the address doesn't describe a valid gradient, then a negative value is
+// returned.
+static inline int swgl_validateGradient(sampler2D sampler, ivec2_scalar address,
+                                        int entries) {
+  return sampler->format == TextureFormat::RGBA32F && address.y >= 0 &&
+                 address.y < int(sampler->height) && address.x >= 0 &&
+                 address.x < int(sampler->width) && entries > 0 &&
+                 address.x +
+                         int(sizeof(GradientStops) / sizeof(Float)) * entries <=
+                     int(sampler->width)
+             ? address.y * sampler->stride + address.x * 4
+             : -1;
+}
+
+static inline WideRGBA8 sampleGradient(sampler2D sampler, int address,
+                                       Float entry) {
+  assert(sampler->format == TextureFormat::RGBA32F);
+  assert(address >= 0 && address < int(sampler->height * sampler->stride));
+  // Get the integer portion of the entry index to find the entry colors.
+  I32 index = cast(entry);
+  // Use the fractional portion of the entry index to control blending between
+  // entry colors.
+  Float offset = entry - cast(index);
+  // Every entry is a pair of colors blended by the fractional offset.
+  assert(test_all(index >= 0 &&
+                  index * int(sizeof(GradientStops) / sizeof(Float)) <
+                      int(sampler->width)));
+  GradientStops* stops = (GradientStops*)&sampler->buf[address];
+  // Blend between the colors for each SIMD lane, then pack them to RGBA8
+  // result. Since the layout of the RGBA8 framebuffer is actually BGRA while
+  // the gradient table has RGBA colors, swizzling is required.
+  return combine(
+      packRGBA8(round_pixel(stops[index.x].interpolate(offset.x).zyxw),
+                round_pixel(stops[index.y].interpolate(offset.y).zyxw)),
+      packRGBA8(round_pixel(stops[index.z].interpolate(offset.z).zyxw),
+                round_pixel(stops[index.w].interpolate(offset.w).zyxw)));
+}
+
+// Samples a gradient entry from the gradient at the provided linearized
+// address. The integer portion of the entry index is used to find the entry
+// within the table whereas the fractional portion is used to blend between
+// adjacent table entries.
+#define swgl_commitGradientRGBA8(sampler, address, entry) \
+  swgl_commitChunk(RGBA8, sampleGradient(sampler, address, entry))
+
+// Variant that allows specifying a color multiplier of the gradient result.
+#define swgl_commitGradientColorRGBA8(sampler, address, entry, color)         \
+  swgl_commitChunk(RGBA8, applyColor(sampleGradient(sampler, address, entry), \
+                                     packColor(swgl_OutRGBA, color)))
+
+// Samples an entire span of a linear gradient by crawling the gradient table
+// and looking for consecutive stops that can be merged into a single larger
+// gradient, then interpolating between those larger gradients within the span.
+template <bool BLEND>
+static bool commitLinearGradient(sampler2D sampler, int address, float size,
+                                 bool tileRepeat, bool gradientRepeat, vec2 pos,
+                                 const vec2_scalar& scaleDir, float startOffset,
+                                 uint32_t* buf, int span) {
+  assert(sampler->format == TextureFormat::RGBA32F);
+  assert(address >= 0 && address < int(sampler->height * sampler->stride));
+  GradientStops* stops = (GradientStops*)&sampler->buf[address];
+  // Get the chunk delta from the difference in offset steps. This represents
+  // how far within the gradient table we advance for every step in output,
+  // normalized to gradient table size.
+  vec2_scalar posStep = dFdx(pos) * 4.0f;
+  float delta = dot(posStep, scaleDir);
+  if (!isfinite(delta)) {
+    return false;
+  }
+  // If we have a repeating brush, then the position will be modulo the [0,1)
+  // interval. Compute coefficients that can be used to quickly evaluate the
+  // distance to the interval boundary where the offset will wrap.
+  vec2_scalar distCoeffsX = {0.25f * span, 0.0f};
+  vec2_scalar distCoeffsY = distCoeffsX;
+  if (tileRepeat) {
+    if (posStep.x != 0.0f) {
+      distCoeffsX = vec2_scalar{step(0.0f, posStep.x), 1.0f} * recip(posStep.x);
+    }
+    if (posStep.y != 0.0f) {
+      distCoeffsY = vec2_scalar{step(0.0f, posStep.y), 1.0f} * recip(posStep.y);
+    }
+  }
+  for (; span > 0;) {
+    // Try to process as many chunks as are within the span if possible.
+    float chunks = 0.25f * span;
+    vec2 repeatPos = pos;
+    if (tileRepeat) {
+      // If this is a repeating brush, then limit the chunks to not cross the
+      // interval boundaries.
+      repeatPos = fract(pos);
+      chunks = min(chunks, distCoeffsX.x - repeatPos.x.x * distCoeffsX.y);
+      chunks = min(chunks, distCoeffsY.x - repeatPos.y.x * distCoeffsY.y);
+    }
+    // Compute the gradient offset from the position.
+    Float offset =
+        repeatPos.x * scaleDir.x + repeatPos.y * scaleDir.y - startOffset;
+    // If repeat is desired, we need to limit the offset to a fractional value.
+    if (gradientRepeat) {
+      offset = fract(offset);
+    }
+    // To properly handle both clamping and repeating of the table offset, we
+    // need to ensure we don't run past the 0 and 1 points. Here we compute the
+    // intercept points depending on whether advancing forwards or backwards in
+    // the gradient table to ensure the chunk count is limited by the amount
+    // before intersection. If there is no delta, then we compute no intercept.
+    float startEntry;
+    int minIndex, maxIndex;
+    if (offset.x < 0) {
+      // If we're below the gradient table, use the first color stop. We can
+      // only intercept the table if walking forward.
+      startEntry = 0;
+      minIndex = int(startEntry);
+      maxIndex = minIndex;
+      if (delta > 0) {
+        chunks = min(chunks, -offset.x / delta);
+      }
+    } else if (offset.x < 1) {
+      // Otherwise, we're inside the gradient table. Depending on the direction
+      // we're walking the the table, we may intersect either the 0 or 1 offset.
+      // Compute the start entry based on our initial offset, and compute the
+      // end entry based on the available chunks limited by intercepts. Clamp
+      // them into the valid range of the table.
+      startEntry = 1.0f + offset.x * size;
+      if (delta < 0) {
+        chunks = min(chunks, -offset.x / delta);
+      } else if (delta > 0) {
+        chunks = min(chunks, (1 - offset.x) / delta);
+      }
+      float endEntry = clamp(1.0f + (offset.x + delta * int(chunks)) * size,
+                             0.0f, 1.0f + size);
+      // Now that we know the range of entries we need to sample, we want to
+      // find the largest possible merged gradient within that range. Depending
+      // on which direction we are advancing in the table, we either walk up or
+      // down the table trying to merge the current entry with the adjacent
+      // entry. We finally limit the chunks to only sample from this merged
+      // gradient.
+      minIndex = int(startEntry);
+      maxIndex = minIndex;
+      if (delta > 0) {
+        while (maxIndex + 1 < endEntry &&
+               stops[maxIndex].can_merge(stops[maxIndex + 1])) {
+          maxIndex++;
+        }
+        chunks = min(chunks, (maxIndex + 1 - startEntry) / (delta * size));
+      } else if (delta < 0) {
+        while (minIndex - 1 > endEntry &&
+               stops[minIndex - 1].can_merge(stops[minIndex])) {
+          minIndex--;
+        }
+        chunks = min(chunks, (minIndex - startEntry) / (delta * size));
+      }
+    } else {
+      // If we're above the gradient table, use the last color stop. We can
+      // only intercept the table if walking backward.
+      startEntry = 1.0f + size;
+      minIndex = int(startEntry);
+      maxIndex = minIndex;
+      if (delta < 0) {
+        chunks = min(chunks, (1 - offset.x) / delta);
+      }
+    }
+    // If there are any amount of whole chunks of a merged gradient found,
+    // then we want to process that as a single gradient span with the start
+    // and end colors from the min and max entries.
+    if (chunks >= 1.0f) {
+      int inside = int(chunks);
+      // Sample the start color from the min entry and the end color from the
+      // max entry of the merged gradient. These are scaled to a range of
+      // 0..0xFF00, as that is the largest shifted value that can fit in a U16.
+      // Since we are only doing addition with the step value, we can still
+      // represent negative step values without having to use an explicit sign
+      // bit, as the result will still come out the same, allowing us to gain an
+      // extra bit of precision. We will later shift these into 8 bit output
+      // range while committing the span, but stepping with higher precision to
+      // avoid banding. We convert from RGBA to BGRA here to avoid doing this in
+      // the inner loop.
+      auto minColorF = stops[minIndex].startColor.zyxw * float(0xFF00);
+      auto maxColorF = stops[maxIndex].end_color().zyxw * float(0xFF00);
+      // Get the color range of the merged gradient, normalized to its size.
+      auto colorRangeF =
+          (maxColorF - minColorF) * (1.0f / (maxIndex + 1 - minIndex));
+      // Compute the actual starting color of the current start offset within
+      // the merged gradient. The value 0.5 is added to the low bits (0x80) so
+      // that the color will effective round to the nearest increment below.
+      auto colorF =
+          minColorF + colorRangeF * (startEntry - minIndex) + float(0x80);
+      // Compute the portion of the color range that we advance on each chunk.
+      Float deltaColorF = colorRangeF * (delta * size);
+      // Quantize the color delta and current color. These have already been
+      // scaled to the 0..0xFF00 range, so we just need to round them to U16.
+      auto deltaColor = repeat4(CONVERT(round_pixel(deltaColorF, 1), U16));
+      auto color =
+          combine(CONVERT(round_pixel(colorF, 1), U16),
+                  CONVERT(round_pixel(colorF + deltaColorF * 0.25f, 1), U16),
+                  CONVERT(round_pixel(colorF + deltaColorF * 0.5f, 1), U16),
+                  CONVERT(round_pixel(colorF + deltaColorF * 0.75f, 1), U16));
+      // Finally, step the current color through the output chunks, shifting
+      // it into 8 bit range and outputting as we go.
+      for (auto* end = buf + inside * 4; buf < end; buf += 4) {
+        commit_blend_span<BLEND>(buf, bit_cast<WideRGBA8>(color >> 8));
+        color += deltaColor;
+      }
+      // Deduct the number of chunks inside the gradient from the remaining
+      // overall span. If we exhausted the span, bail out.
+      span -= inside * 4;
+      if (span <= 0) {
+        break;
+      }
+      // Otherwise, assume we're in a transitional section of the gradient that
+      // will probably require per-sample table lookups, so fall through below.
+      // We need to re-evaluate the position and offset first, though.
+      pos += posStep * float(inside);
+      repeatPos = tileRepeat ? fract(pos) : pos;
+      offset =
+          repeatPos.x * scaleDir.x + repeatPos.y * scaleDir.y - startOffset;
+      if (gradientRepeat) {
+        offset = fract(offset);
+      }
+    }
+    // If we get here, there were no whole chunks of a merged gradient found
+    // that we could process, but we still have a non-zero amount of span left.
+    // That means we have segments of gradient that begin or end at the current
+    // entry we're on. For this case, we just fall back to sampleGradient which
+    // will calculate a table entry for each sample, assuming the samples may
+    // have different table entries.
+    Float entry = clamp(offset * size + 1.0f, 0.0f, 1.0f + size);
+    commit_blend_span<BLEND>(buf, sampleGradient(sampler, address, entry));
+    span -= 4;
+    buf += 4;
+    pos += posStep;
+  }
+  return true;
+}
+
+// Commits an entire span of a linear gradient, given the address of a table
+// previously resolved with swgl_validateGradient. The size of the inner portion
+// of the table is given, assuming the table start and ends with a single entry
+// each to deal with clamping. Repeating will be handled if necessary. The
+// initial offset within the table is used to designate where to start the span
+// and how to step through the gradient table.
+#define swgl_commitLinearGradientRGBA8(sampler, address, size, tileRepeat,   \
+                                       gradientRepeat, pos, scaleDir,        \
+                                       startOffset)                          \
+  do {                                                                       \
+    bool drawn = false;                                                      \
+    if (blend_key) {                                                         \
+      drawn = commitLinearGradient<true>(                                    \
+          sampler, address, size, tileRepeat, gradientRepeat, pos, scaleDir, \
+          startOffset, swgl_OutRGBA8, swgl_SpanLength);                      \
+    } else {                                                                 \
+      drawn = commitLinearGradient<false>(                                   \
+          sampler, address, size, tileRepeat, gradientRepeat, pos, scaleDir, \
+          startOffset, swgl_OutRGBA8, swgl_SpanLength);                      \
+    }                                                                        \
+    if (drawn) {                                                             \
+      swgl_OutRGBA8 += swgl_SpanLength;                                      \
+      swgl_SpanLength = 0;                                                   \
+    }                                                                        \
+  } while (0)
+
+template <bool CLAMP, typename V>
+static ALWAYS_INLINE V fastSqrt(V v) {
+#if USE_SSE2 || USE_NEON
+  // Clamp to avoid zero in inversesqrt.
+  return v * inversesqrt(CLAMP ? max(v, V(1.0e-10f)) : v);
+#else
+  return sqrt(v);
+#endif
+}
+
+template <bool CLAMP, typename V>
+static ALWAYS_INLINE auto fastLength(V v) {
+  return fastSqrt<CLAMP>(dot(v, v));
+}
+
+// Samples an entire span of a radial gradient by crawling the gradient table
+// and looking for consecutive stops that can be merged into a single larger
+// gradient, then interpolating between those larger gradients within the span
+// based on the computed position relative to a radius.
+template <bool BLEND>
+static bool commitRadialGradient(sampler2D sampler, int address, float size,
+                                 bool repeat, vec2 pos, float radius,
+                                 uint32_t* buf, int span) {
+  assert(sampler->format == TextureFormat::RGBA32F);
+  assert(address >= 0 && address < int(sampler->height * sampler->stride));
+  GradientStops* stops = (GradientStops*)&sampler->buf[address];
+  // clang-format off
+  // Given position p, delta d, and radius r, we need to repeatedly solve the
+  // following quadratic for the pixel offset t:
+  //    length(p + t*d) = r
+  //    (px + t*dx)^2 + (py + t*dy)^2 = r^2
+  // Rearranged into quadratic equation form (t^2*a + t*b + c = 0) this is:
+  //    t^2*(dx^2+dy^2) + t*2*(dx*px+dy*py) + (px^2+py^2-r^2) = 0
+  //    t^2*d.d + t*2*d.p + (p.p-r^2) = 0
+  // The solution of the quadratic formula t=(-b+-sqrt(b^2-4ac))/2a reduces to:
+  //    t = -d.p/d.d +- sqrt((d.p/d.d)^2 - (p.p-r^2)/d.d)
+  // Note that d.p, d.d, p.p, and r^2 are constant across the gradient, and so
+  // we cache them below for faster computation.
+  //
+  // The quadratic has two solutions, representing the span intersecting the
+  // given radius of gradient, which can occur at two offsets. If there is only
+  // one solution (where b^2-4ac = 0), this represents the point at which the
+  // span runs tangent to the radius. This middle point is significant in that
+  // before it, we walk down the gradient ramp, and after it, we walk up the
+  // ramp.
+  // clang-format on
+  vec2_scalar pos0 = {pos.x.x, pos.y.x};
+  vec2_scalar delta = {pos.x.y - pos.x.x, pos.y.y - pos.y.x};
+  float deltaDelta = dot(delta, delta);
+  if (!isfinite(deltaDelta) || !isfinite(radius)) {
+    return false;
+  }
+  float invDelta, middleT, middleB;
+  if (deltaDelta > 0) {
+    invDelta = 1.0f / deltaDelta;
+    middleT = -dot(delta, pos0) * invDelta;
+    middleB = middleT * middleT - dot(pos0, pos0) * invDelta;
+  } else {
+    // If position is invariant, just set the coefficients so the quadratic
+    // always reduces to the end of the span.
+    invDelta = 0.0f;
+    middleT = float(span);
+    middleB = 0.0f;
+  }
+  // We only want search for merged gradients up to the minimum of either the
+  // mid-point or the span length. Cache those offsets here as they don't vary
+  // in the inner loop.
+  Float middleEndRadius = fastLength<true>(
+      pos0 + delta * (Float){middleT, float(span), 0.0f, 0.0f});
+  float middleRadius = span < middleT ? middleEndRadius.y : middleEndRadius.x;
+  float endRadius = middleEndRadius.y;
+  // Convert delta to change in position per chunk.
+  delta *= 4;
+  deltaDelta *= 4 * 4;
+  // clang-format off
+  // Given current position p and delta d, we reduce:
+  //    length(p) = sqrt(dot(p,p)) = dot(p,p) * invsqrt(dot(p,p))
+  // where dot(p+d,p+d) can be accumulated as:
+  //    (x+dx)^2+(y+dy)^2 = (x^2+y^2) + 2(x*dx+y*dy) + (dx^2+dy^2)
+  //                      = p.p + 2p.d + d.d
+  // Since p increases by d every loop iteration, p.d increases by d.d, and thus
+  // we can accumulate d.d to calculate 2p.d, then allowing us to get the next
+  // dot-product by adding it to dot-product p.p of the prior iteration. This
+  // saves us some multiplications and an expensive sqrt inside the inner loop.
+  // clang-format on
+  Float dotPos = dot(pos, pos);
+  Float dotPosDelta = 2.0f * dot(pos, delta) + deltaDelta;
+  float deltaDelta2 = 2.0f * deltaDelta;
+  for (int t = 0; t < span;) {
+    // Compute the gradient table offset from the current position.
+    Float offset = fastSqrt<true>(dotPos) - radius;
+    float startRadius = radius;
+    // If repeat is desired, we need to limit the offset to a fractional value.
+    if (repeat) {
+      // The non-repeating radius at which the gradient table actually starts,
+      // radius + floor(offset) = radius + (offset - fract(offset)).
+      startRadius += offset.x;
+      offset = fract(offset);
+      startRadius -= offset.x;
+    }
+    // We need to find the min/max index in the table of the gradient we want to
+    // use as well as the intercept point where we leave this gradient.
+    float intercept = -1;
+    int minIndex = 0;
+    int maxIndex = int(1.0f + size);
+    if (offset.x < 0) {
+      // If inside the inner radius of the gradient table, then use the first
+      // stop. Set the intercept to advance forward to the start of the gradient
+      // table.
+      maxIndex = minIndex;
+      if (t >= middleT) {
+        intercept = radius;
+      }
+    } else if (offset.x < 1) {
+      // Otherwise, we're inside the valid part of the gradient table.
+      minIndex = int(1.0f + offset.x * size);
+      maxIndex = minIndex;
+      // Find the offset in the gradient that corresponds to the search limit.
+      // We only search up to the minimum of either the mid-point or the span
+      // length. Get the table index that corresponds to this offset, clamped so
+      // that we avoid hitting the beginning (0) or end (1 + size) of the table.
+      float searchOffset =
+          (t >= middleT ? endRadius : middleRadius) - startRadius;
+      int searchIndex = int(clamp(1.0f + size * searchOffset, 1.0f, size));
+      // If we are past the mid-point, walk up the gradient table trying to
+      // merge stops. If we're below the mid-point, we need to walk down the
+      // table. We note the table index at which we need to look for an
+      // intercept to determine a valid span.
+      if (t >= middleT) {
+        while (maxIndex + 1 <= searchIndex &&
+               stops[maxIndex].can_merge(stops[maxIndex + 1])) {
+          maxIndex++;
+        }
+        intercept = maxIndex + 1;
+      } else {
+        while (minIndex - 1 >= searchIndex &&
+               stops[minIndex - 1].can_merge(stops[minIndex])) {
+          minIndex--;
+        }
+        intercept = minIndex;
+      }
+      // Convert from a table index into units of radius from the center of the
+      // gradient.
+      intercept = clamp((intercept - 1.0f) / size, 0.0f, 1.0f) + startRadius;
+    } else {
+      // If outside the outer radius of the gradient table, then use the last
+      // stop. Set the intercept to advance toward the valid part of the
+      // gradient table if going in, or just run to the end of the span if going
+      // away from the gradient.
+      minIndex = maxIndex;
+      if (t < middleT) {
+        intercept = radius + 1;
+      }
+    }
+    // Solve the quadratic for t to find where the merged gradient ends. If no
+    // intercept is found, just go to the middle or end of the span.
+    float endT = t >= middleT ? span : min(span, int(middleT));
+    if (intercept >= 0) {
+      float b = middleB + intercept * intercept * invDelta;
+      if (b > 0) {
+        b = fastSqrt<false>(b);
+        endT = min(endT, t >= middleT ? middleT + b : middleT - b);
+      }
+    }
+    // Figure out how many chunks are actually inside the merged gradient.
+    if (t + 4.0f <= endT) {
+      int inside = int(endT - t) & ~3;
+      // Convert start and end colors to BGRA and scale to 0..255 range later.
+      auto minColorF = stops[minIndex].startColor.zyxw * 255.0f;
+      auto maxColorF = stops[maxIndex].end_color().zyxw * 255.0f;
+      // Compute the change in color per change in gradient offset.
+      auto deltaColorF =
+          (maxColorF - minColorF) * (size / (maxIndex + 1 - minIndex));
+      // Subtract off the color difference of the beginning of the current span
+      // from the beginning of the gradient.
+      Float colorF =
+          minColorF - deltaColorF * (startRadius + (minIndex - 1) / size);
+      // Finally, walk over the span accumulating the position dot product and
+      // getting its sqrt as an offset into the color ramp. Since we're already
+      // in BGRA format and scaled to 255, we just need to round to an integer
+      // and pack down to pixel format.
+      for (auto* end = buf + inside; buf < end; buf += 4) {
+        Float offsetG = fastSqrt<false>(dotPos);
+        commit_blend_span<BLEND>(
+            buf,
+            combine(
+                packRGBA8(round_pixel(colorF + deltaColorF * offsetG.x, 1),
+                          round_pixel(colorF + deltaColorF * offsetG.y, 1)),
+                packRGBA8(round_pixel(colorF + deltaColorF * offsetG.z, 1),
+                          round_pixel(colorF + deltaColorF * offsetG.w, 1))));
+        dotPos += dotPosDelta;
+        dotPosDelta += deltaDelta2;
+      }
+      // Advance past the portion of gradient we just processed.
+      t += inside;
+      // If we hit the end of the span, exit out now.
+      if (t >= span) {
+        break;
+      }
+      // Otherwise, we are most likely in a transitional section of the gradient
+      // between stops that will likely require doing per-sample table lookups.
+      // Rather than having to redo all the searching above to figure that out,
+      // just assume that to be the case and fall through below to doing the
+      // table lookups to hopefully avoid an iteration.
+      offset = fastSqrt<true>(dotPos) - radius;
+      if (repeat) {
+        offset = fract(offset);
+      }
+    }
+    // If we got here, that means we still have span left to process but did not
+    // have any whole chunks that fell within a merged gradient. Just fall back
+    // to doing a table lookup for each sample.
+    Float entry = clamp(offset * size + 1.0f, 0.0f, 1.0f + size);
+    commit_blend_span<BLEND>(buf, sampleGradient(sampler, address, entry));
+    buf += 4;
+    t += 4;
+    dotPos += dotPosDelta;
+    dotPosDelta += deltaDelta2;
+  }
+  return true;
+}
+
+// Commits an entire span of a radial gradient similar to
+// swglcommitLinearGradient, but given a varying 2D position scaled to
+// gradient-space and a radius at which the distance from the origin maps to the
+// start of the gradient table.
+#define swgl_commitRadialGradientRGBA8(sampler, address, size, repeat, pos,    \
+                                       radius)                                 \
+  do {                                                                         \
+    bool drawn = false;                                                        \
+    if (blend_key) {                                                           \
+      drawn =                                                                  \
+          commitRadialGradient<true>(sampler, address, size, repeat, pos,      \
+                                     radius, swgl_OutRGBA8, swgl_SpanLength);  \
+    } else {                                                                   \
+      drawn =                                                                  \
+          commitRadialGradient<false>(sampler, address, size, repeat, pos,     \
+                                      radius, swgl_OutRGBA8, swgl_SpanLength); \
+    }                                                                          \
+    if (drawn) {                                                               \
+      swgl_OutRGBA8 += swgl_SpanLength;                                        \
+      swgl_SpanLength = 0;                                                     \
+    }                                                                          \
+  } while (0)
+
+// Extension to set a clip mask image to be sampled during blending. The offset
+// specifies the positioning of the clip mask image relative to the viewport
+// origin. The bounding box specifies the rectangle relative to the clip mask's
+// origin that constrains sampling within the clip mask. Blending must be
+// enabled for this to work.
+static sampler2D swgl_ClipMask = nullptr;
+static IntPoint swgl_ClipMaskOffset = {0, 0};
+static IntRect swgl_ClipMaskBounds = {0, 0, 0, 0};
+#define swgl_clipMask(mask, offset, bb_origin, bb_size)        \
+  do {                                                         \
+    if (bb_size != vec2_scalar(0.0f, 0.0f)) {                  \
+      swgl_ClipFlags |= SWGL_CLIP_FLAG_MASK;                   \
+      swgl_ClipMask = mask;                                    \
+      swgl_ClipMaskOffset = make_ivec2(offset);                \
+      swgl_ClipMaskBounds =                                    \
+          IntRect(make_ivec2(bb_origin), make_ivec2(bb_size)); \
+    }                                                          \
+  } while (0)
+
+// Extension to enable anti-aliasing for the given edges of a quad.
+// Blending must be enable for this to work.
+static int swgl_AAEdgeMask = 0;
+
+static ALWAYS_INLINE int calcAAEdgeMask(bool on) { return on ? 0xF : 0; }
+static ALWAYS_INLINE int calcAAEdgeMask(int mask) { return mask; }
+static ALWAYS_INLINE int calcAAEdgeMask(bvec4_scalar mask) {
+  return (mask.x ? 1 : 0) | (mask.y ? 2 : 0) | (mask.z ? 4 : 0) |
+         (mask.w ? 8 : 0);
+}
+
+#define swgl_antiAlias(edges)                \
+  do {                                       \
+    swgl_AAEdgeMask = calcAAEdgeMask(edges); \
+    if (swgl_AAEdgeMask) {                   \
+      swgl_ClipFlags |= SWGL_CLIP_FLAG_AA;   \
+    }                                        \
+  } while (0)
+
+#define swgl_blendDropShadow(color)                         \
+  do {                                                      \
+    swgl_ClipFlags |= SWGL_CLIP_FLAG_BLEND_OVERRIDE;        \
+    swgl_BlendOverride = BLEND_KEY(SWGL_BLEND_DROP_SHADOW); \
+    swgl_BlendColorRGBA8 = packColor<uint32_t>(color);      \
+  } while (0)
+
+#define swgl_blendSubpixelText(color)                         \
+  do {                                                        \
+    swgl_ClipFlags |= SWGL_CLIP_FLAG_BLEND_OVERRIDE;          \
+    swgl_BlendOverride = BLEND_KEY(SWGL_BLEND_SUBPIXEL_TEXT); \
+    swgl_BlendColorRGBA8 = packColor<uint32_t>(color);        \
+    swgl_BlendAlphaRGBA8 = alphas(swgl_BlendColorRGBA8);      \
+  } while (0)
+
+// Dispatch helper used by the GLSL translator to swgl_drawSpan functions.
+// The number of pixels committed is tracked by checking for the difference in
+// swgl_SpanLength. Any varying interpolants used will be advanced past the
+// committed part of the span in case the fragment shader must be executed for
+// any remaining pixels that were not committed by the span shader.
+#define DISPATCH_DRAW_SPAN(self, format)        \
+  do {                                          \
+    int total = self->swgl_SpanLength;          \
+    self->swgl_drawSpan##format();              \
+    int drawn = total - self->swgl_SpanLength;  \
+    if (drawn) self->step_interp_inputs(drawn); \
+    return drawn;                               \
+  } while (0)
diff --git a/gfx/wr/swgl/src/swgl_fns.rs b/gfx/wr/swgl/src/swgl_fns.rs
new file mode 100644
index 0000000000..79669bd205
--- /dev/null
+++ b/gfx/wr/swgl/src/swgl_fns.rs
@@ -0,0 +1,2489 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+#![allow(unused_variables)]
+
+use gleam::gl::*;
+use std::ffi::{CStr, CString};
+use std::os::raw::{c_char, c_int, c_void};
+use std::ptr;
+use std::str;
+
+#[allow(unused)]
+macro_rules! debug {
+    ($($x:tt)*) => {};
+}
+
+#[repr(C)]
+struct LockedTexture {
+    _private: [u8; 0],
+}
+
+#[allow(dead_code)]
+extern "C" {
+    fn ActiveTexture(texture: GLenum);
+    fn BindTexture(target: GLenum, texture: GLuint);
+    fn BindBuffer(target: GLenum, buffer: GLuint);
+    fn BindVertexArray(vao: GLuint);
+    fn BindFramebuffer(target: GLenum, fb: GLuint);
+    fn BindRenderbuffer(target: GLenum, rb: GLuint);
+    fn BlendFunc(srgb: GLenum, drgb: GLenum, sa: GLenum, da: GLenum);
+    fn BlendColor(r: GLfloat, g: GLfloat, b: GLfloat, a: GLfloat);
+    fn BlendEquation(mode: GLenum);
+    fn Enable(cap: GLenum);
+    fn Disable(cap: GLenum);
+    fn GenQueries(n: GLsizei, result: *mut GLuint);
+    fn BeginQuery(target: GLenum, id: GLuint);
+    fn EndQuery(target: GLenum);
+    fn GetQueryObjectui64v(id: GLuint, pname: GLenum, params: *mut GLuint64);
+    fn GenBuffers(n: i32, result: *mut GLuint);
+    fn GenTextures(n: i32, result: *mut GLuint);
+    fn GenFramebuffers(n: i32, result: *mut GLuint);
+    fn GenRenderbuffers(n: i32, result: *mut GLuint);
+    fn BufferData(target: GLenum, size: GLsizeiptr, data: *const GLvoid, usage: GLenum);
+    fn BufferSubData(target: GLenum, offset: GLintptr, size: GLsizeiptr, data: *const GLvoid);
+    fn MapBuffer(target: GLenum, access: GLbitfield) -> *mut c_void;
+    fn MapBufferRange(
+        target: GLenum,
+        offset: GLintptr,
+        length: GLsizeiptr,
+        access: GLbitfield,
+    ) -> *mut c_void;
+    fn UnmapBuffer(target: GLenum) -> GLboolean;
+    fn TexStorage2D(
+        target: GLenum,
+        levels: GLint,
+        internal_format: GLenum,
+        width: GLsizei,
+        height: GLsizei,
+    );
+    fn FramebufferTexture2D(
+        target: GLenum,
+        attachment: GLenum,
+        textarget: GLenum,
+        texture: GLuint,
+        level: GLint,
+    );
+    fn CheckFramebufferStatus(target: GLenum) -> GLenum;
+    fn InvalidateFramebuffer(target: GLenum, num_attachments: GLsizei, attachments: *const GLenum);
+    fn TexImage2D(
+        target: GLenum,
+        level: GLint,
+        internal_format: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        border: GLint,
+        format: GLenum,
+        ty: GLenum,
+        data: *const c_void,
+    );
+    fn TexSubImage2D(
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        format: GLenum,
+        ty: GLenum,
+        data: *const c_void,
+    );
+    fn GenerateMipmap(target: GLenum);
+    fn GetUniformLocation(program: GLuint, name: *const GLchar) -> GLint;
+    fn BindAttribLocation(program: GLuint, index: GLuint, name: *const GLchar);
+    fn GetAttribLocation(program: GLuint, name: *const GLchar) -> GLint;
+    fn GenVertexArrays(n: i32, result: *mut GLuint);
+    fn VertexAttribPointer(
+        index: GLuint,
+        size: GLint,
+        type_: GLenum,
+        normalized: GLboolean,
+        stride: GLsizei,
+        offset: *const GLvoid,
+    );
+    fn VertexAttribIPointer(
+        index: GLuint,
+        size: GLint,
+        type_: GLenum,
+        stride: GLsizei,
+        offset: *const GLvoid,
+    );
+    fn CreateShader(shader_type: GLenum) -> GLuint;
+    fn AttachShader(program: GLuint, shader: GLuint);
+    fn CreateProgram() -> GLuint;
+    fn Uniform1i(location: GLint, v0: GLint);
+    fn Uniform4fv(location: GLint, count: GLsizei, value: *const GLfloat);
+    fn UniformMatrix4fv(
+        location: GLint,
+        count: GLsizei,
+        transpose: GLboolean,
+        value: *const GLfloat,
+    );
+    fn DrawElementsInstanced(
+        mode: GLenum,
+        count: GLsizei,
+        type_: GLenum,
+        indices: GLintptr,
+        instancecount: GLsizei,
+    );
+    fn EnableVertexAttribArray(index: GLuint);
+    fn VertexAttribDivisor(index: GLuint, divisor: GLuint);
+    fn LinkProgram(program: GLuint);
+    fn GetLinkStatus(program: GLuint) -> GLint;
+    fn UseProgram(program: GLuint);
+    fn SetViewport(x: GLint, y: GLint, width: GLsizei, height: GLsizei);
+    fn FramebufferRenderbuffer(
+        target: GLenum,
+        attachment: GLenum,
+        renderbuffertarget: GLenum,
+        renderbuffer: GLuint,
+    );
+    fn RenderbufferStorage(target: GLenum, internalformat: GLenum, width: GLsizei, height: GLsizei);
+    fn DepthMask(flag: GLboolean);
+    fn DepthFunc(func: GLenum);
+    fn SetScissor(x: GLint, y: GLint, width: GLsizei, height: GLsizei);
+    fn ClearColor(r: GLfloat, g: GLfloat, b: GLfloat, a: GLfloat);
+    fn ClearDepth(depth: GLdouble);
+    fn Clear(mask: GLbitfield);
+    fn ClearTexSubImage(
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        zoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        depth: GLsizei,
+        format: GLenum,
+        ty: GLenum,
+        data: *const c_void,
+    );
+    fn ClearTexImage(target: GLenum, level: GLint, format: GLenum, ty: GLenum, data: *const c_void);
+    fn ClearColorRect(
+        fbo: GLuint,
+        xoffset: GLint,
+        yoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        r: GLfloat,
+        g: GLfloat,
+        b: GLfloat,
+        a: GLfloat,
+    );
+    fn PixelStorei(name: GLenum, param: GLint);
+    fn ReadPixels(
+        x: GLint,
+        y: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        format: GLenum,
+        ty: GLenum,
+        data: *mut c_void,
+    );
+    fn Finish();
+    fn ShaderSourceByName(shader: GLuint, name: *const GLchar);
+    fn TexParameteri(target: GLenum, pname: GLenum, param: GLint);
+    fn CopyImageSubData(
+        src_name: GLuint,
+        src_target: GLenum,
+        src_level: GLint,
+        src_x: GLint,
+        src_y: GLint,
+        src_z: GLint,
+        dst_name: GLuint,
+        dst_target: GLenum,
+        dst_level: GLint,
+        dst_x: GLint,
+        dst_y: GLint,
+        dst_z: GLint,
+        src_width: GLsizei,
+        src_height: GLsizei,
+        src_depth: GLsizei,
+    );
+    fn CopyTexSubImage2D(
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        x: GLint,
+        y: GLint,
+        width: GLsizei,
+        height: GLsizei,
+    );
+    fn BlitFramebuffer(
+        src_x0: GLint,
+        src_y0: GLint,
+        src_x1: GLint,
+        src_y1: GLint,
+        dst_x0: GLint,
+        dst_y0: GLint,
+        dst_x1: GLint,
+        dst_y1: GLint,
+        mask: GLbitfield,
+        filter: GLenum,
+    );
+    fn GetIntegerv(pname: GLenum, params: *mut GLint);
+    fn GetBooleanv(pname: GLenum, params: *mut GLboolean);
+    fn GetString(name: GLenum) -> *const c_char;
+    fn GetStringi(name: GLenum, index: GLuint) -> *const c_char;
+    fn GetError() -> GLenum;
+    fn InitDefaultFramebuffer(
+        x: i32,
+        y: i32,
+        width: i32,
+        height: i32,
+        stride: i32,
+        buf: *mut c_void,
+    );
+    fn GetColorBuffer(
+        fbo: GLuint,
+        flush: GLboolean,
+        width: *mut i32,
+        height: *mut i32,
+        stride: *mut i32,
+    ) -> *mut c_void;
+    fn ResolveFramebuffer(fbo: GLuint);
+    fn SetTextureBuffer(
+        tex: GLuint,
+        internal_format: GLenum,
+        width: GLsizei,
+        height: GLsizei,
+        stride: GLsizei,
+        buf: *mut c_void,
+        min_width: GLsizei,
+        min_height: GLsizei,
+    );
+    fn SetTextureParameter(tex: GLuint, pname: GLenum, param: GLint);
+    fn DeleteTexture(n: GLuint);
+    fn DeleteRenderbuffer(n: GLuint);
+    fn DeleteFramebuffer(n: GLuint);
+    fn DeleteBuffer(n: GLuint);
+    fn DeleteVertexArray(n: GLuint);
+    fn DeleteQuery(n: GLuint);
+    fn DeleteShader(shader: GLuint);
+    fn DeleteProgram(program: GLuint);
+    fn LockFramebuffer(fbo: GLuint) -> *mut LockedTexture;
+    fn LockTexture(tex: GLuint) -> *mut LockedTexture;
+    fn LockResource(resource: *mut LockedTexture);
+    fn UnlockResource(resource: *mut LockedTexture);
+    fn GetResourceBuffer(
+        resource: *mut LockedTexture,
+        width: *mut i32,
+        height: *mut i32,
+        stride: *mut i32,
+    ) -> *mut c_void;
+    fn Composite(
+        locked_dst: *mut LockedTexture,
+        locked_src: *mut LockedTexture,
+        src_x: GLint,
+        src_y: GLint,
+        src_width: GLsizei,
+        src_height: GLsizei,
+        dst_x: GLint,
+        dst_y: GLint,
+        dst_width: GLsizei,
+        dst_height: GLsizei,
+        opaque: GLboolean,
+        flip_x: GLboolean,
+        flip_y: GLboolean,
+        filter: GLenum,
+        clip_x: GLint,
+        clip_y: GLint,
+        clip_width: GLsizei,
+        clip_height: GLsizei,
+    );
+    fn CompositeYUV(
+        locked_dst: *mut LockedTexture,
+        locked_y: *mut LockedTexture,
+        locked_u: *mut LockedTexture,
+        locked_v: *mut LockedTexture,
+        color_space: YuvRangedColorSpace,
+        color_depth: GLuint,
+        src_x: GLint,
+        src_y: GLint,
+        src_width: GLsizei,
+        src_height: GLsizei,
+        dst_x: GLint,
+        dst_y: GLint,
+        dst_width: GLsizei,
+        dst_height: GLsizei,
+        flip_x: GLboolean,
+        flip_y: GLboolean,
+        clip_x: GLint,
+        clip_y: GLint,
+        clip_width: GLsizei,
+        clip_height: GLsizei,
+    );
+    fn CreateContext() -> *mut c_void;
+    fn ReferenceContext(ctx: *mut c_void);
+    fn DestroyContext(ctx: *mut c_void);
+    fn MakeCurrent(ctx: *mut c_void);
+    fn ReportMemory(ctx: *mut c_void, size_of_op: unsafe extern "C" fn(ptr: *const c_void) -> usize) -> usize;
+}
+
+#[derive(Clone, Copy)]
+pub struct Context(*mut c_void);
+
+impl Context {
+    pub fn create() -> Self {
+        Context(unsafe { CreateContext() })
+    }
+
+    pub fn reference(&self) {
+        unsafe {
+            ReferenceContext(self.0);
+        }
+    }
+
+    pub fn destroy(&self) {
+        unsafe {
+            DestroyContext(self.0);
+        }
+    }
+
+    pub fn make_current(&self) {
+        unsafe {
+            MakeCurrent(self.0);
+        }
+    }
+
+    pub fn init_default_framebuffer(
+        &self,
+        x: i32,
+        y: i32,
+        width: i32,
+        height: i32,
+        stride: i32,
+        buf: *mut c_void,
+    ) {
+        unsafe {
+            InitDefaultFramebuffer(x, y, width, height, stride, buf);
+        }
+    }
+
+    pub fn get_color_buffer(&self, fbo: GLuint, flush: bool) -> (*mut c_void, i32, i32, i32) {
+        unsafe {
+            let mut width: i32 = 0;
+            let mut height: i32 = 0;
+            let mut stride: i32 = 0;
+            let data_ptr = GetColorBuffer(
+                fbo,
+                flush as GLboolean,
+                &mut width,
+                &mut height,
+                &mut stride,
+            );
+            (data_ptr, width, height, stride)
+        }
+    }
+
+    pub fn resolve_framebuffer(&self, fbo: GLuint) {
+        unsafe {
+            ResolveFramebuffer(fbo);
+        }
+    }
+
+    pub fn clear_color_rect(
+        &self,
+        fbo: GLuint,
+        xoffset: GLint,
+        yoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        r: f32,
+        g: f32,
+        b: f32,
+        a: f32,
+    ) {
+        unsafe {
+            ClearColorRect(fbo, xoffset, yoffset, width, height, r, g, b, a);
+        }
+    }
+
+    pub fn set_texture_buffer(
+        &self,
+        tex: GLuint,
+        internal_format: GLenum,
+        width: GLsizei,
+        height: GLsizei,
+        stride: GLsizei,
+        buf: *mut c_void,
+        min_width: GLsizei,
+        min_height: GLsizei,
+    ) {
+        unsafe {
+            SetTextureBuffer(
+                tex,
+                internal_format,
+                width,
+                height,
+                stride,
+                buf,
+                min_width,
+                min_height,
+            );
+        }
+    }
+
+    pub fn set_texture_parameter(&self, tex: GLuint, pname: GLenum, param: GLint) {
+        unsafe {
+            SetTextureParameter(tex, pname, param);
+        }
+    }
+
+    pub fn lock_framebuffer(&self, fbo: GLuint) -> Option<LockedResource> {
+        unsafe {
+            let resource = LockFramebuffer(fbo);
+            if resource != ptr::null_mut() {
+                Some(LockedResource(resource))
+            } else {
+                None
+            }
+        }
+    }
+
+    pub fn lock_texture(&self, tex: GLuint) -> Option<LockedResource> {
+        unsafe {
+            let resource = LockTexture(tex);
+            if resource != ptr::null_mut() {
+                Some(LockedResource(resource))
+            } else {
+                None
+            }
+        }
+    }
+
+    pub fn report_memory(&self, size_of_op: unsafe extern "C" fn(ptr: *const c_void) -> usize) -> usize {
+        unsafe { ReportMemory(self.0, size_of_op) }
+    }
+}
+
+impl From<*mut c_void> for Context {
+    fn from(ptr: *mut c_void) -> Self {
+        Context(ptr)
+    }
+}
+
+impl From<Context> for *mut c_void {
+    fn from(ctx: Context) -> Self {
+        ctx.0
+    }
+}
+
+fn calculate_length(width: GLsizei, height: GLsizei, format: GLenum, pixel_type: GLenum) -> usize {
+    let colors = match format {
+        RED => 1,
+        RGB => 3,
+        BGR => 3,
+
+        RGBA => 4,
+        BGRA => 4,
+
+        ALPHA => 1,
+        R16 => 1,
+        LUMINANCE => 1,
+        DEPTH_COMPONENT => 1,
+        _ => panic!("unsupported format for read_pixels: {:?}", format),
+    };
+    let depth = match pixel_type {
+        UNSIGNED_BYTE => 1,
+        UNSIGNED_SHORT => 2,
+        SHORT => 2,
+        FLOAT => 4,
+        UNSIGNED_INT_8_8_8_8_REV => 1,
+        _ => panic!("unsupported pixel_type for read_pixels: {:?}", pixel_type),
+    };
+
+    return (width * height * colors * depth) as usize;
+}
+
+impl Gl for Context {
+    fn get_type(&self) -> GlType {
+        GlType::Gl
+    }
+
+    fn buffer_data_untyped(
+        &self,
+        target: GLenum,
+        size: GLsizeiptr,
+        data: *const GLvoid,
+        usage: GLenum,
+    ) {
+        debug!(
+            "buffer_data_untyped {} {} {:?} {}",
+            target, size, data, usage
+        );
+        //panic!();
+        unsafe {
+            BufferData(target, size, data, usage);
+        }
+    }
+
+    fn buffer_sub_data_untyped(
+        &self,
+        target: GLenum,
+        offset: isize,
+        size: GLsizeiptr,
+        data: *const GLvoid,
+    ) {
+        debug!(
+            "buffer_sub_data_untyped {} {} {} {:?}",
+            target, offset, size, data
+        );
+        //panic!();
+        unsafe {
+            BufferSubData(target, offset, size, data);
+        }
+    }
+
+    fn map_buffer(&self, target: GLenum, access: GLbitfield) -> *mut c_void {
+        unsafe { MapBuffer(target, access) }
+    }
+
+    fn map_buffer_range(
+        &self,
+        target: GLenum,
+        offset: GLintptr,
+        length: GLsizeiptr,
+        access: GLbitfield,
+    ) -> *mut c_void {
+        unsafe { MapBufferRange(target, offset, length, access) }
+    }
+
+    fn unmap_buffer(&self, target: GLenum) -> GLboolean {
+        unsafe { UnmapBuffer(target) }
+    }
+
+    fn shader_source(&self, shader: GLuint, strings: &[&[u8]]) {
+        //panic!();
+        debug!("shader_source {}", shader);
+        //for s in strings {
+        //    debug!("{}", str::from_utf8(s).unwrap());
+        //}
+        //panic!();
+        for s in strings {
+            let u = str::from_utf8(s).unwrap();
+            const PREFIX: &'static str = "// shader: ";
+            if let Some(start) = u.find(PREFIX) {
+                if let Some(end) = u[start..].find('\n') {
+                    let name = u[start + PREFIX.len()..start + end].trim();
+                    debug!("shader name: {}", name);
+                    unsafe {
+                        let c_string = CString::new(name).unwrap();
+                        ShaderSourceByName(shader, c_string.as_ptr());
+                        return;
+                    }
+                }
+            }
+        }
+        panic!("unknown shader");
+    }
+
+    fn tex_buffer(&self, target: GLenum, internal_format: GLenum, buffer: GLuint) {
+        panic!();
+    }
+
+    fn read_buffer(&self, mode: GLenum) {
+        panic!();
+    }
+
+    fn read_pixels_into_buffer(
+        &self,
+        x: GLint,
+        y: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        format: GLenum,
+        pixel_type: GLenum,
+        dst_buffer: &mut [u8],
+    ) {
+        // Assumes that the user properly allocated the size for dst_buffer.
+        assert!(calculate_length(width, height, format, pixel_type) == dst_buffer.len());
+
+        unsafe {
+            ReadPixels(
+                x,
+                y,
+                width,
+                height,
+                format,
+                pixel_type,
+                dst_buffer.as_mut_ptr() as *mut c_void,
+            );
+        }
+    }
+
+    fn read_pixels(
+        &self,
+        x: GLint,
+        y: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        format: GLenum,
+        pixel_type: GLenum,
+    ) -> Vec<u8> {
+        let len = calculate_length(width, height, format, pixel_type);
+        let mut pixels: Vec<u8> = Vec::new();
+        pixels.reserve(len);
+        unsafe {
+            pixels.set_len(len);
+        }
+
+        self.read_pixels_into_buffer(
+            x,
+            y,
+            width,
+            height,
+            format,
+            pixel_type,
+            pixels.as_mut_slice(),
+        );
+
+        pixels
+    }
+
+    unsafe fn read_pixels_into_pbo(
+        &self,
+        x: GLint,
+        y: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        format: GLenum,
+        pixel_type: GLenum,
+    ) {
+        ReadPixels(x, y, width, height, format, pixel_type, ptr::null_mut());
+    }
+
+    fn sample_coverage(&self, value: GLclampf, invert: bool) {
+        panic!();
+    }
+
+    fn polygon_offset(&self, factor: GLfloat, units: GLfloat) {
+        panic!();
+    }
+
+    fn pixel_store_i(&self, name: GLenum, param: GLint) {
+        //panic!();
+        debug!("pixel_store_i {:x} {}", name, param);
+        unsafe {
+            PixelStorei(name, param);
+        }
+    }
+
+    fn gen_buffers(&self, n: GLsizei) -> Vec<GLuint> {
+        //panic!();
+        let mut result = vec![0 as GLuint; n as usize];
+        unsafe {
+            GenBuffers(n, result.as_mut_ptr());
+        }
+        result
+    }
+
+    fn gen_renderbuffers(&self, n: GLsizei) -> Vec<GLuint> {
+        debug!("gen_renderbuffers {}", n);
+        //panic!();
+        let mut result = vec![0 as GLuint; n as usize];
+        unsafe {
+            GenRenderbuffers(n, result.as_mut_ptr());
+        }
+        result
+    }
+
+    fn gen_framebuffers(&self, n: GLsizei) -> Vec<GLuint> {
+        //panic!();
+        debug!("gen_framebuffers {}", n);
+        let mut result = vec![0 as GLuint; n as usize];
+        unsafe {
+            GenFramebuffers(n, result.as_mut_ptr());
+        }
+        result
+    }
+
+    fn gen_textures(&self, n: GLsizei) -> Vec<GLuint> {
+        //panic!();
+        let mut result = vec![0 as GLuint; n as usize];
+        unsafe {
+            GenTextures(n, result.as_mut_ptr());
+        }
+        result
+    }
+
+    fn gen_vertex_arrays(&self, n: GLsizei) -> Vec<GLuint> {
+        //panic!();
+        let mut result = vec![0 as GLuint; n as usize];
+        unsafe {
+            GenVertexArrays(n, result.as_mut_ptr());
+        }
+        result
+    }
+
+    fn gen_vertex_arrays_apple(&self, n: GLsizei) -> Vec<GLuint> {
+        self.gen_vertex_arrays(n)
+    }
+
+    fn gen_queries(&self, n: GLsizei) -> Vec<GLuint> {
+        let mut result = vec![0 as GLuint; n as usize];
+        unsafe {
+            GenQueries(n, result.as_mut_ptr());
+        }
+        result
+    }
+
+    fn begin_query(&self, target: GLenum, id: GLuint) {
+        unsafe {
+            BeginQuery(target, id);
+        }
+    }
+
+    fn end_query(&self, target: GLenum) {
+        unsafe {
+            EndQuery(target);
+        }
+    }
+
+    fn query_counter(&self, id: GLuint, target: GLenum) {
+        panic!();
+    }
+
+    fn get_query_object_iv(&self, id: GLuint, pname: GLenum) -> i32 {
+        panic!();
+        //0
+    }
+
+    fn get_query_object_uiv(&self, id: GLuint, pname: GLenum) -> u32 {
+        panic!();
+        //0
+    }
+
+    fn get_query_object_i64v(&self, id: GLuint, pname: GLenum) -> i64 {
+        panic!();
+        //0
+    }
+
+    fn get_query_object_ui64v(&self, id: GLuint, pname: GLenum) -> u64 {
+        let mut result = 0;
+        unsafe {
+            GetQueryObjectui64v(id, pname, &mut result);
+        }
+        result
+    }
+
+    fn delete_queries(&self, queries: &[GLuint]) {
+        unsafe {
+            for q in queries {
+                DeleteQuery(*q);
+            }
+        }
+    }
+
+    fn delete_vertex_arrays(&self, vertex_arrays: &[GLuint]) {
+        unsafe {
+            for v in vertex_arrays {
+                DeleteVertexArray(*v);
+            }
+        }
+    }
+
+    fn delete_vertex_arrays_apple(&self, vertex_arrays: &[GLuint]) {
+        self.delete_vertex_arrays(vertex_arrays)
+    }
+
+    fn delete_buffers(&self, buffers: &[GLuint]) {
+        unsafe {
+            for b in buffers {
+                DeleteBuffer(*b);
+            }
+        }
+    }
+
+    fn delete_renderbuffers(&self, renderbuffers: &[GLuint]) {
+        unsafe {
+            for r in renderbuffers {
+                DeleteRenderbuffer(*r);
+            }
+        }
+    }
+
+    fn delete_framebuffers(&self, framebuffers: &[GLuint]) {
+        unsafe {
+            for f in framebuffers {
+                DeleteFramebuffer(*f);
+            }
+        }
+    }
+
+    fn delete_textures(&self, textures: &[GLuint]) {
+        unsafe {
+            for t in textures {
+                DeleteTexture(*t);
+            }
+        }
+    }
+
+    fn framebuffer_renderbuffer(
+        &self,
+        target: GLenum,
+        attachment: GLenum,
+        renderbuffertarget: GLenum,
+        renderbuffer: GLuint,
+    ) {
+        debug!(
+            "framebufer_renderbuffer {} {} {} {}",
+            target, attachment, renderbuffertarget, renderbuffer
+        );
+        //panic!();
+        unsafe {
+            FramebufferRenderbuffer(target, attachment, renderbuffertarget, renderbuffer);
+        }
+    }
+
+    fn renderbuffer_storage(
+        &self,
+        target: GLenum,
+        internalformat: GLenum,
+        width: GLsizei,
+        height: GLsizei,
+    ) {
+        debug!(
+            "renderbuffer_storage {} {} {} {}",
+            target, internalformat, width, height
+        );
+        //panic!();
+        unsafe {
+            RenderbufferStorage(target, internalformat, width, height);
+        }
+    }
+
+    fn depth_func(&self, func: GLenum) {
+        debug!("depth_func {}", func);
+        //panic!();
+        unsafe {
+            DepthFunc(func);
+        }
+    }
+
+    fn active_texture(&self, texture: GLenum) {
+        //panic!();
+        unsafe {
+            ActiveTexture(texture);
+        }
+    }
+
+    fn attach_shader(&self, program: GLuint, shader: GLuint) {
+        debug!("attach shader {} {}", program, shader);
+        //panic!();
+        unsafe {
+            AttachShader(program, shader);
+        }
+    }
+
+    fn bind_attrib_location(&self, program: GLuint, index: GLuint, name: &str) {
+        debug!("bind_attrib_location {} {} {}", program, index, name);
+        //panic!();
+        let c_string = CString::new(name).unwrap();
+        unsafe { BindAttribLocation(program, index, c_string.as_ptr()) }
+    }
+
+    // https://www.khronos.org/registry/OpenGL-Refpages/es2.0/xhtml/glGetUniform.xml
+    unsafe fn get_uniform_iv(&self, program: GLuint, location: GLint, result: &mut [GLint]) {
+        panic!();
+        //assert!(!result.is_empty());
+    }
+
+    // https://www.khronos.org/registry/OpenGL-Refpages/es2.0/xhtml/glGetUniform.xml
+    unsafe fn get_uniform_fv(&self, program: GLuint, location: GLint, result: &mut [GLfloat]) {
+        panic!();
+        //assert!(!result.is_empty());
+    }
+
+    fn get_uniform_block_index(&self, program: GLuint, name: &str) -> GLuint {
+        panic!();
+        //0
+    }
+
+    fn get_uniform_indices(&self, program: GLuint, names: &[&str]) -> Vec<GLuint> {
+        panic!();
+        //Vec::new()
+    }
+
+    fn bind_buffer_base(&self, target: GLenum, index: GLuint, buffer: GLuint) {
+        panic!();
+    }
+
+    fn bind_buffer_range(
+        &self,
+        target: GLenum,
+        index: GLuint,
+        buffer: GLuint,
+        offset: GLintptr,
+        size: GLsizeiptr,
+    ) {
+        panic!();
+    }
+
+    fn uniform_block_binding(
+        &self,
+        program: GLuint,
+        uniform_block_index: GLuint,
+        uniform_block_binding: GLuint,
+    ) {
+        panic!();
+    }
+
+    fn bind_buffer(&self, target: GLenum, buffer: GLuint) {
+        //panic!();
+        unsafe {
+            BindBuffer(target, buffer);
+        }
+    }
+
+    fn bind_vertex_array(&self, vao: GLuint) {
+        //panic!();
+        unsafe {
+            BindVertexArray(vao);
+        }
+    }
+
+    fn bind_vertex_array_apple(&self, vao: GLuint) {
+        self.bind_vertex_array(vao)
+    }
+
+    fn bind_renderbuffer(&self, target: GLenum, renderbuffer: GLuint) {
+        debug!("bind_renderbuffer {} {}", target, renderbuffer);
+        //panic!();
+        unsafe {
+            BindRenderbuffer(target, renderbuffer);
+        }
+    }
+
+    fn bind_framebuffer(&self, target: GLenum, framebuffer: GLuint) {
+        debug!("bind_framebuffer {} {}", target, framebuffer);
+        //panic!();
+        unsafe {
+            BindFramebuffer(target, framebuffer);
+        }
+    }
+
+    fn bind_vertex_buffer(
+        &self,
+        binding_index: GLuint,
+        buffer: GLuint,
+        offset: GLintptr,
+        stride: GLint,
+    ) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn bind_texture(&self, target: GLenum, texture: GLuint) {
+        //panic!();
+        unsafe {
+            BindTexture(target, texture);
+        }
+    }
+
+    fn draw_buffers(&self, bufs: &[GLenum]) {
+        panic!();
+        //unsafe {}
+    }
+
+    // FIXME: Does not verify buffer size -- unsafe!
+    fn tex_image_2d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        internal_format: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        border: GLint,
+        format: GLenum,
+        ty: GLenum,
+        opt_data: Option<&[u8]>,
+    ) {
+        unsafe {
+            let pdata = match opt_data {
+                Some(data) => data.as_ptr() as *const GLvoid,
+                None => ptr::null(),
+            };
+            TexImage2D(
+                target,
+                level,
+                internal_format,
+                width,
+                height,
+                border,
+                format,
+                ty,
+                pdata,
+            );
+        }
+    }
+
+    fn compressed_tex_image_2d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        internal_format: GLenum,
+        width: GLsizei,
+        height: GLsizei,
+        border: GLint,
+        data: &[u8],
+    ) {
+        panic!();
+    }
+
+    fn compressed_tex_sub_image_2d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        format: GLenum,
+        data: &[u8],
+    ) {
+        panic!();
+    }
+
+    fn tex_image_3d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        internal_format: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        depth: GLsizei,
+        border: GLint,
+        format: GLenum,
+        ty: GLenum,
+        opt_data: Option<&[u8]>,
+    ) {
+        panic!();
+    }
+
+    fn copy_tex_image_2d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        internal_format: GLenum,
+        x: GLint,
+        y: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        border: GLint,
+    ) {
+        panic!();
+    }
+
+    fn copy_tex_sub_image_2d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        x: GLint,
+        y: GLint,
+        width: GLsizei,
+        height: GLsizei,
+    ) {
+        unsafe {
+            CopyTexSubImage2D(target, level, xoffset, yoffset, x, y, width, height);
+        }
+    }
+
+    fn copy_tex_sub_image_3d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        zoffset: GLint,
+        x: GLint,
+        y: GLint,
+        width: GLsizei,
+        height: GLsizei,
+    ) {
+        panic!();
+    }
+
+    fn tex_sub_image_2d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        format: GLenum,
+        ty: GLenum,
+        data: &[u8],
+    ) {
+        debug!(
+            "tex_sub_image_2d {} {} {} {} {} {} {} {}",
+            target, level, xoffset, yoffset, width, height, format, ty
+        );
+        //panic!();
+        unsafe {
+            TexSubImage2D(
+                target,
+                level,
+                xoffset,
+                yoffset,
+                width,
+                height,
+                format,
+                ty,
+                data.as_ptr() as *const c_void,
+            );
+        }
+    }
+
+    fn tex_sub_image_2d_pbo(
+        &self,
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        format: GLenum,
+        ty: GLenum,
+        offset: usize,
+    ) {
+        debug!(
+            "tex_sub_image_2d_pbo {} {} {} {} {} {} {} {} {}",
+            target, level, xoffset, yoffset, width, height, format, ty, offset
+        );
+        //panic!();
+        unsafe {
+            TexSubImage2D(
+                target,
+                level,
+                xoffset,
+                yoffset,
+                width,
+                height,
+                format,
+                ty,
+                offset as *const c_void,
+            );
+        }
+    }
+
+    fn tex_sub_image_3d(
+        &self,
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        zoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        depth: GLsizei,
+        format: GLenum,
+        ty: GLenum,
+        data: &[u8],
+    ) {
+        debug!("tex_sub_image_3d");
+        panic!();
+    }
+
+    fn tex_sub_image_3d_pbo(
+        &self,
+        target: GLenum,
+        level: GLint,
+        xoffset: GLint,
+        yoffset: GLint,
+        zoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+        depth: GLsizei,
+        format: GLenum,
+        ty: GLenum,
+        offset: usize,
+    ) {
+        panic!();
+    }
+
+    fn tex_storage_2d(
+        &self,
+        target: GLenum,
+        levels: GLint,
+        internal_format: GLenum,
+        width: GLsizei,
+        height: GLsizei,
+    ) {
+        //panic!();
+        unsafe {
+            TexStorage2D(target, levels, internal_format, width, height);
+        }
+    }
+
+    fn tex_storage_3d(
+        &self,
+        target: GLenum,
+        levels: GLint,
+        internal_format: GLenum,
+        width: GLsizei,
+        height: GLsizei,
+        depth: GLsizei,
+    ) {
+        panic!();
+    }
+
+    fn get_tex_image_into_buffer(
+        &self,
+        target: GLenum,
+        level: GLint,
+        format: GLenum,
+        ty: GLenum,
+        output: &mut [u8],
+    ) {
+        panic!();
+    }
+
+    unsafe fn copy_image_sub_data(
+        &self,
+        src_name: GLuint,
+        src_target: GLenum,
+        src_level: GLint,
+        src_x: GLint,
+        src_y: GLint,
+        src_z: GLint,
+        dst_name: GLuint,
+        dst_target: GLenum,
+        dst_level: GLint,
+        dst_x: GLint,
+        dst_y: GLint,
+        dst_z: GLint,
+        src_width: GLsizei,
+        src_height: GLsizei,
+        src_depth: GLsizei,
+    ) {
+        CopyImageSubData(
+            src_name, src_target, src_level, src_x, src_y, src_z, dst_name, dst_target, dst_level,
+            dst_x, dst_y, dst_z, src_width, src_height, src_depth,
+        );
+    }
+
+    fn invalidate_framebuffer(&self, target: GLenum, attachments: &[GLenum]) {
+        unsafe {
+            InvalidateFramebuffer(target, attachments.len() as GLsizei, attachments.as_ptr());
+        }
+    }
+
+    fn invalidate_sub_framebuffer(
+        &self,
+        target: GLenum,
+        attachments: &[GLenum],
+        xoffset: GLint,
+        yoffset: GLint,
+        width: GLsizei,
+        height: GLsizei,
+    ) {
+    }
+
+    #[inline]
+    unsafe fn get_integer_v(&self, name: GLenum, result: &mut [GLint]) {
+        //panic!();
+        assert!(!result.is_empty());
+        GetIntegerv(name, result.as_mut_ptr());
+    }
+
+    #[inline]
+    unsafe fn get_integer_64v(&self, name: GLenum, result: &mut [GLint64]) {
+        panic!();
+        //assert!(!result.is_empty());
+    }
+
+    #[inline]
+    unsafe fn get_integer_iv(&self, name: GLenum, index: GLuint, result: &mut [GLint]) {
+        panic!();
+        //assert!(!result.is_empty());
+    }
+
+    #[inline]
+    unsafe fn get_integer_64iv(&self, name: GLenum, index: GLuint, result: &mut [GLint64]) {
+        panic!();
+        //assert!(!result.is_empty());
+    }
+
+    #[inline]
+    unsafe fn get_boolean_v(&self, name: GLenum, result: &mut [GLboolean]) {
+        debug!("get_boolean_v {}", name);
+        //panic!();
+        assert!(!result.is_empty());
+        GetBooleanv(name, result.as_mut_ptr());
+    }
+
+    #[inline]
+    unsafe fn get_float_v(&self, name: GLenum, result: &mut [GLfloat]) {
+        panic!();
+        //assert!(!result.is_empty());
+    }
+
+    fn get_framebuffer_attachment_parameter_iv(
+        &self,
+        target: GLenum,
+        attachment: GLenum,
+        pname: GLenum,
+    ) -> GLint {
+        panic!();
+        //0
+    }
+
+    fn get_renderbuffer_parameter_iv(&self, target: GLenum, pname: GLenum) -> GLint {
+        panic!();
+        //0
+    }
+
+    fn get_tex_parameter_iv(&self, target: GLenum, pname: GLenum) -> GLint {
+        panic!();
+        //0
+    }
+
+    fn get_tex_parameter_fv(&self, target: GLenum, pname: GLenum) -> GLfloat {
+        panic!();
+        //0.0
+    }
+
+    fn tex_parameter_i(&self, target: GLenum, pname: GLenum, param: GLint) {
+        //panic!();
+        unsafe {
+            TexParameteri(target, pname, param);
+        }
+    }
+
+    fn tex_parameter_f(&self, target: GLenum, pname: GLenum, param: GLfloat) {
+        panic!();
+    }
+
+    fn framebuffer_texture_2d(
+        &self,
+        target: GLenum,
+        attachment: GLenum,
+        textarget: GLenum,
+        texture: GLuint,
+        level: GLint,
+    ) {
+        debug!(
+            "framebuffer_texture_2d {} {} {} {} {}",
+            target, attachment, textarget, texture, level
+        );
+        //panic!();
+        unsafe {
+            FramebufferTexture2D(target, attachment, textarget, texture, level);
+        }
+    }
+
+    fn framebuffer_texture_layer(
+        &self,
+        target: GLenum,
+        attachment: GLenum,
+        texture: GLuint,
+        level: GLint,
+        layer: GLint,
+    ) {
+        debug!(
+            "framebuffer_texture_layer {} {} {} {} {}",
+            target, attachment, texture, level, layer
+        );
+        panic!();
+    }
+
+    fn blit_framebuffer(
+        &self,
+        src_x0: GLint,
+        src_y0: GLint,
+        src_x1: GLint,
+        src_y1: GLint,
+        dst_x0: GLint,
+        dst_y0: GLint,
+        dst_x1: GLint,
+        dst_y1: GLint,
+        mask: GLbitfield,
+        filter: GLenum,
+    ) {
+        unsafe {
+            BlitFramebuffer(
+                src_x0, src_y0, src_x1, src_y1, dst_x0, dst_y0, dst_x1, dst_y1, mask, filter,
+            );
+        }
+    }
+
+    fn vertex_attrib_4f(&self, index: GLuint, x: GLfloat, y: GLfloat, z: GLfloat, w: GLfloat) {
+        panic!();
+    }
+
+    fn vertex_attrib_binding(&self, attrib_index: GLuint, binding_index: GLuint) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn vertex_attrib_pointer_f32(
+        &self,
+        index: GLuint,
+        size: GLint,
+        normalized: bool,
+        stride: GLsizei,
+        offset: GLuint,
+    ) {
+        panic!();
+    }
+
+    fn vertex_attrib_pointer(
+        &self,
+        index: GLuint,
+        size: GLint,
+        type_: GLenum,
+        normalized: bool,
+        stride: GLsizei,
+        offset: GLuint,
+    ) {
+        debug!(
+            "vertex_attrib_pointer {} {} {} {} {} {}",
+            index, size, type_, normalized, stride, offset
+        );
+        //panic!();
+        unsafe {
+            VertexAttribPointer(
+                index,
+                size,
+                type_,
+                normalized as GLboolean,
+                stride,
+                offset as *const GLvoid,
+            );
+        }
+    }
+
+    fn vertex_attrib_i_pointer(
+        &self,
+        index: GLuint,
+        size: GLint,
+        type_: GLenum,
+        stride: GLsizei,
+        offset: GLuint,
+    ) {
+        debug!(
+            "vertex_attrib_i_pointer {} {} {} {} {}",
+            index, size, type_, stride, offset
+        );
+        //panic!();
+        unsafe {
+            VertexAttribIPointer(index, size, type_, stride, offset as *const GLvoid);
+        }
+    }
+
+    fn vertex_attrib_divisor(&self, index: GLuint, divisor: GLuint) {
+        debug!("vertex_attrib_divisor {} {}", index, divisor);
+        //assert!(index == 0 && divisor == 0);
+        //panic!();
+        unsafe {
+            VertexAttribDivisor(index, divisor);
+        }
+    }
+
+    fn vertex_attrib_format(
+        &self,
+        attrib_index: GLuint,
+        size: GLint,
+        type_: GLenum,
+        normalized: bool,
+        relative_offset: GLuint,
+    ) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn vertex_attrib_i_format(
+        &self,
+        attrib_index: GLuint,
+        size: GLint,
+        type_: GLenum,
+        relative_offset: GLuint,
+    ) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn vertex_binding_divisor(&self, binding_index: GLuint, divisor: GLuint) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn viewport(&self, x: GLint, y: GLint, width: GLsizei, height: GLsizei) {
+        debug!("viewport {} {} {} {}", x, y, width, height);
+        //panic!();
+        unsafe {
+            SetViewport(x, y, width, height);
+        }
+    }
+
+    fn scissor(&self, x: GLint, y: GLint, width: GLsizei, height: GLsizei) {
+        //panic!();
+        unsafe {
+            SetScissor(x, y, width, height);
+        }
+    }
+
+    fn line_width(&self, width: GLfloat) {
+        panic!();
+    }
+
+    fn use_program(&self, program: GLuint) {
+        //panic!();
+        unsafe {
+            UseProgram(program);
+        }
+    }
+
+    fn validate_program(&self, program: GLuint) {
+        panic!();
+    }
+
+    fn draw_arrays(&self, mode: GLenum, first: GLint, count: GLsizei) {
+        unsafe {
+            DrawElementsInstanced(mode, count, NONE, first as GLintptr, 1);
+        }
+    }
+
+    fn draw_arrays_instanced(
+        &self,
+        mode: GLenum,
+        first: GLint,
+        count: GLsizei,
+        primcount: GLsizei,
+    ) {
+        unsafe {
+            DrawElementsInstanced(mode, count, NONE, first as GLintptr, primcount);
+        }
+    }
+
+    fn draw_elements(
+        &self,
+        mode: GLenum,
+        count: GLsizei,
+        element_type: GLenum,
+        indices_offset: GLuint,
+    ) {
+        debug!(
+            "draw_elements {} {} {} {} {}",
+            mode, count, element_type, indices_offset
+        );
+        //panic!();
+        unsafe {
+            DrawElementsInstanced(mode, count, element_type, indices_offset as GLintptr, 1);
+        }
+    }
+
+    fn draw_elements_instanced(
+        &self,
+        mode: GLenum,
+        count: GLsizei,
+        element_type: GLenum,
+        indices_offset: GLuint,
+        primcount: GLsizei,
+    ) {
+        debug!(
+            "draw_elements_instanced {} {} {} {} {}",
+            mode, count, element_type, indices_offset, primcount
+        );
+        //panic!();
+        unsafe {
+            DrawElementsInstanced(
+                mode,
+                count,
+                element_type,
+                indices_offset as GLintptr,
+                primcount,
+            );
+        }
+    }
+
+    fn blend_color(&self, r: f32, g: f32, b: f32, a: f32) {
+        unsafe {
+            BlendColor(r, g, b, a);
+        }
+    }
+
+    fn blend_func(&self, sfactor: GLenum, dfactor: GLenum) {
+        unsafe {
+            BlendFunc(sfactor, dfactor, sfactor, dfactor);
+        }
+    }
+
+    fn blend_func_separate(
+        &self,
+        src_rgb: GLenum,
+        dest_rgb: GLenum,
+        src_alpha: GLenum,
+        dest_alpha: GLenum,
+    ) {
+        unsafe {
+            BlendFunc(src_rgb, dest_rgb, src_alpha, dest_alpha);
+        }
+    }
+
+    fn blend_equation(&self, mode: GLenum) {
+        unsafe {
+            BlendEquation(mode);
+        }
+    }
+
+    fn blend_equation_separate(&self, mode_rgb: GLenum, mode_alpha: GLenum) {
+        panic!();
+    }
+
+    fn color_mask(&self, r: bool, g: bool, b: bool, a: bool) {
+        panic!();
+    }
+
+    fn cull_face(&self, mode: GLenum) {
+        panic!();
+    }
+
+    fn front_face(&self, mode: GLenum) {
+        panic!();
+    }
+
+    fn enable(&self, cap: GLenum) {
+        debug!("enable {}", cap);
+        //panic!();
+        unsafe {
+            Enable(cap);
+        }
+    }
+
+    fn disable(&self, cap: GLenum) {
+        debug!("disable {}", cap);
+        //panic!();
+        unsafe {
+            Disable(cap);
+        }
+    }
+
+    fn hint(&self, param_name: GLenum, param_val: GLenum) {
+        panic!();
+    }
+
+    fn is_enabled(&self, cap: GLenum) -> GLboolean {
+        panic!();
+        //0
+    }
+
+    fn is_shader(&self, shader: GLuint) -> GLboolean {
+        panic!();
+        //0
+    }
+
+    fn is_texture(&self, texture: GLenum) -> GLboolean {
+        panic!();
+        //0
+    }
+
+    fn is_framebuffer(&self, framebuffer: GLenum) -> GLboolean {
+        panic!();
+        //0
+    }
+
+    fn is_renderbuffer(&self, renderbuffer: GLenum) -> GLboolean {
+        panic!();
+        //0
+    }
+
+    fn check_frame_buffer_status(&self, target: GLenum) -> GLenum {
+        debug!("check_frame_buffer_status {}", target);
+        //panic!();
+        unsafe { CheckFramebufferStatus(target) }
+    }
+
+    fn enable_vertex_attrib_array(&self, index: GLuint) {
+        //panic!();
+        debug!("enable_vertex_attrib_array {}", index);
+        unsafe {
+            EnableVertexAttribArray(index);
+            //assert_eq!(index, 0);
+        }
+    }
+
+    fn disable_vertex_attrib_array(&self, index: GLuint) {
+        panic!();
+    }
+
+    fn uniform_1f(&self, location: GLint, v0: GLfloat) {
+        panic!();
+    }
+
+    fn uniform_1fv(&self, location: GLint, values: &[f32]) {
+        panic!();
+    }
+
+    fn uniform_1i(&self, location: GLint, v0: GLint) {
+        debug!("uniform_1i {} {}", location, v0);
+        //panic!();
+        unsafe {
+            Uniform1i(location, v0);
+        }
+    }
+
+    fn uniform_1iv(&self, location: GLint, values: &[i32]) {
+        panic!();
+    }
+
+    fn uniform_1ui(&self, location: GLint, v0: GLuint) {
+        panic!();
+    }
+
+    fn uniform_2f(&self, location: GLint, v0: GLfloat, v1: GLfloat) {
+        panic!();
+    }
+
+    fn uniform_2fv(&self, location: GLint, values: &[f32]) {
+        panic!();
+    }
+
+    fn uniform_2i(&self, location: GLint, v0: GLint, v1: GLint) {
+        panic!();
+    }
+
+    fn uniform_2iv(&self, location: GLint, values: &[i32]) {
+        panic!();
+    }
+
+    fn uniform_2ui(&self, location: GLint, v0: GLuint, v1: GLuint) {
+        panic!();
+    }
+
+    fn uniform_3f(&self, location: GLint, v0: GLfloat, v1: GLfloat, v2: GLfloat) {
+        panic!();
+    }
+
+    fn uniform_3fv(&self, location: GLint, values: &[f32]) {
+        panic!();
+    }
+
+    fn uniform_3i(&self, location: GLint, v0: GLint, v1: GLint, v2: GLint) {
+        panic!();
+    }
+
+    fn uniform_3iv(&self, location: GLint, values: &[i32]) {
+        panic!();
+    }
+
+    fn uniform_3ui(&self, location: GLint, v0: GLuint, v1: GLuint, v2: GLuint) {
+        panic!();
+    }
+
+    fn uniform_4f(&self, location: GLint, x: GLfloat, y: GLfloat, z: GLfloat, w: GLfloat) {
+        panic!();
+    }
+
+    fn uniform_4i(&self, location: GLint, x: GLint, y: GLint, z: GLint, w: GLint) {
+        panic!();
+    }
+
+    fn uniform_4iv(&self, location: GLint, values: &[i32]) {
+        panic!();
+    }
+
+    fn uniform_4ui(&self, location: GLint, x: GLuint, y: GLuint, z: GLuint, w: GLuint) {
+        panic!();
+    }
+
+    fn uniform_4fv(&self, location: GLint, values: &[f32]) {
+        unsafe {
+            Uniform4fv(location, (values.len() / 4) as GLsizei, values.as_ptr());
+        }
+    }
+
+    fn uniform_matrix_2fv(&self, location: GLint, transpose: bool, value: &[f32]) {
+        panic!();
+    }
+
+    fn uniform_matrix_3fv(&self, location: GLint, transpose: bool, value: &[f32]) {
+        panic!();
+    }
+
+    fn uniform_matrix_4fv(&self, location: GLint, transpose: bool, value: &[f32]) {
+        debug!("uniform_matrix_4fv {} {} {:?}", location, transpose, value);
+        //panic!();
+        unsafe {
+            UniformMatrix4fv(
+                location,
+                (value.len() / 16) as GLsizei,
+                transpose as GLboolean,
+                value.as_ptr(),
+            );
+        }
+    }
+
+    fn depth_mask(&self, flag: bool) {
+        debug!("depth_mask {}", flag);
+        //panic!();
+        unsafe {
+            DepthMask(flag as GLboolean);
+        }
+    }
+
+    fn depth_range(&self, near: f64, far: f64) {
+        panic!();
+    }
+
+    fn get_active_attrib(&self, program: GLuint, index: GLuint) -> (i32, u32, String) {
+        panic!();
+        //(0, 0, String::new())
+    }
+
+    fn get_active_uniform(&self, program: GLuint, index: GLuint) -> (i32, u32, String) {
+        panic!();
+        //(0, 0, String::new())
+    }
+
+    fn get_active_uniforms_iv(
+        &self,
+        program: GLuint,
+        indices: Vec<GLuint>,
+        pname: GLenum,
+    ) -> Vec<GLint> {
+        panic!();
+        //Vec::new()
+    }
+
+    fn get_active_uniform_block_i(&self, program: GLuint, index: GLuint, pname: GLenum) -> GLint {
+        panic!();
+        //0
+    }
+
+    fn get_active_uniform_block_iv(
+        &self,
+        program: GLuint,
+        index: GLuint,
+        pname: GLenum,
+    ) -> Vec<GLint> {
+        panic!();
+        //Vec::new()
+    }
+
+    fn get_active_uniform_block_name(&self, program: GLuint, index: GLuint) -> String {
+        panic!();
+        //String::new()
+    }
+
+    fn get_attrib_location(&self, program: GLuint, name: &str) -> c_int {
+        let name = CString::new(name).unwrap();
+        unsafe { GetAttribLocation(program, name.as_ptr()) }
+    }
+
+    fn get_frag_data_location(&self, program: GLuint, name: &str) -> c_int {
+        panic!();
+        //0
+    }
+
+    fn get_uniform_location(&self, program: GLuint, name: &str) -> c_int {
+        debug!("get_uniform_location {} {}", program, name);
+        //panic!();
+        let name = CString::new(name).unwrap();
+        unsafe { GetUniformLocation(program, name.as_ptr()) }
+    }
+
+    fn get_program_info_log(&self, program: GLuint) -> String {
+        debug!("get_program_info_log {}", program);
+        String::new()
+    }
+
+    #[inline]
+    unsafe fn get_program_iv(&self, program: GLuint, pname: GLenum, result: &mut [GLint]) {
+        debug!("get_program_iv {}", pname);
+        //panic!();
+        assert!(!result.is_empty());
+        //#define GL_LINK_STATUS                    0x8B82
+        if pname == 0x8b82 {
+            result[0] = GetLinkStatus(program);
+        }
+    }
+
+    fn get_program_binary(&self, program: GLuint) -> (Vec<u8>, GLenum) {
+        panic!();
+        //(Vec::new(), NONE)
+    }
+
+    fn program_binary(&self, program: GLuint, format: GLenum, binary: &[u8]) {
+        panic!();
+    }
+
+    fn program_parameter_i(&self, program: GLuint, pname: GLenum, value: GLint) {
+        panic!();
+    }
+
+    #[inline]
+    unsafe fn get_vertex_attrib_iv(&self, index: GLuint, pname: GLenum, result: &mut [GLint]) {
+        panic!();
+        //assert!(!result.is_empty());
+    }
+
+    #[inline]
+    unsafe fn get_vertex_attrib_fv(&self, index: GLuint, pname: GLenum, result: &mut [GLfloat]) {
+        panic!();
+        //assert!(!result.is_empty());
+    }
+
+    fn get_vertex_attrib_pointer_v(&self, index: GLuint, pname: GLenum) -> GLsizeiptr {
+        panic!();
+        //0
+    }
+
+    fn get_buffer_parameter_iv(&self, target: GLuint, pname: GLenum) -> GLint {
+        panic!();
+        //0
+    }
+
+    fn get_shader_info_log(&self, shader: GLuint) -> String {
+        debug!("get_shader_info_log {}", shader);
+        //panic!();
+        String::new()
+    }
+
+    fn get_string(&self, which: GLenum) -> String {
+        // panic!();
+        unsafe {
+            let llstr = GetString(which);
+            if !llstr.is_null() {
+                return str::from_utf8_unchecked(CStr::from_ptr(llstr).to_bytes()).to_string();
+            } else {
+                return "".to_string();
+            }
+        }
+    }
+
+    fn get_string_i(&self, which: GLenum, index: GLuint) -> String {
+        //panic!();
+        unsafe {
+            let llstr = GetStringi(which, index);
+            if !llstr.is_null() {
+                str::from_utf8_unchecked(CStr::from_ptr(llstr).to_bytes()).to_string()
+            } else {
+                "".to_string()
+            }
+        }
+    }
+
+    unsafe fn get_shader_iv(&self, shader: GLuint, pname: GLenum, result: &mut [GLint]) {
+        debug!("get_shader_iv");
+        //panic!();
+        assert!(!result.is_empty());
+        if pname == 0x8B81
+        /*gl::COMPILE_STATUS*/
+        {
+            result[0] = 1;
+        }
+    }
+
+    fn get_shader_precision_format(
+        &self,
+        _shader_type: GLuint,
+        precision_type: GLuint,
+    ) -> (GLint, GLint, GLint) {
+        // gl.GetShaderPrecisionFormat is not available until OpenGL 4.1.
+        // Fallback to OpenGL standard precissions that most desktop hardware support.
+        match precision_type {
+            LOW_FLOAT | MEDIUM_FLOAT | HIGH_FLOAT => {
+                // Fallback to IEEE 754 single precision
+                // Range: from -2^127 to 2^127
+                // Significand precision: 23 bits
+                (127, 127, 23)
+            }
+            LOW_INT | MEDIUM_INT | HIGH_INT => {
+                // Fallback to single precision integer
+                // Range: from -2^24 to 2^24
+                // Precision: For integer formats this value is always 0
+                (24, 24, 0)
+            }
+            _ => (0, 0, 0),
+        }
+    }
+
+    fn compile_shader(&self, shader: GLuint) {
+        debug!("compile_shader {}", shader);
+        //panic!();
+    }
+
+    fn create_program(&self) -> GLuint {
+        debug!("create_program");
+        //panic!();
+        unsafe { CreateProgram() }
+    }
+
+    fn delete_program(&self, program: GLuint) {
+        unsafe {
+            DeleteProgram(program);
+        }
+    }
+
+    fn create_shader(&self, shader_type: GLenum) -> GLuint {
+        debug!("create_shader {}", shader_type);
+        //panic!();
+        unsafe { CreateShader(shader_type) }
+    }
+
+    fn delete_shader(&self, shader: GLuint) {
+        debug!("delete_shader {}", shader);
+        //panic!();
+        unsafe {
+            DeleteShader(shader);
+        }
+    }
+
+    fn detach_shader(&self, program: GLuint, shader: GLuint) {
+        debug!("detach_shader {} {}", program, shader);
+        //panic!();
+    }
+
+    fn link_program(&self, program: GLuint) {
+        debug!("link_program {}", program);
+        //panic!();
+        unsafe {
+            LinkProgram(program);
+        }
+    }
+
+    fn clear_color(&self, r: f32, g: f32, b: f32, a: f32) {
+        //panic!();
+        unsafe {
+            ClearColor(r, g, b, a);
+        }
+    }
+
+    fn clear(&self, buffer_mask: GLbitfield) {
+        debug!("clear {}", buffer_mask);
+        //panic!();
+        unsafe {
+            Clear(buffer_mask);
+        }
+    }
+
+    fn clear_depth(&self, depth: f64) {
+        debug!("clear_depth {}", depth);
+        //panic!();
+        unsafe {
+            ClearDepth(depth as GLclampd);
+        }
+    }
+
+    fn clear_stencil(&self, s: GLint) {
+        panic!();
+    }
+
+    fn flush(&self) {}
+
+    fn finish(&self) {
+        unsafe {
+            Finish();
+        }
+    }
+
+    fn get_error(&self) -> GLenum {
+        //panic!();
+        unsafe { GetError() }
+    }
+
+    fn stencil_mask(&self, mask: GLuint) {
+        panic!();
+    }
+
+    fn stencil_mask_separate(&self, face: GLenum, mask: GLuint) {
+        panic!();
+    }
+
+    fn stencil_func(&self, func: GLenum, ref_: GLint, mask: GLuint) {
+        panic!();
+    }
+
+    fn stencil_func_separate(&self, face: GLenum, func: GLenum, ref_: GLint, mask: GLuint) {
+        panic!();
+    }
+
+    fn stencil_op(&self, sfail: GLenum, dpfail: GLenum, dppass: GLenum) {
+        panic!();
+    }
+
+    fn stencil_op_separate(&self, face: GLenum, sfail: GLenum, dpfail: GLenum, dppass: GLenum) {
+        panic!();
+    }
+
+    fn egl_image_target_texture2d_oes(&self, target: GLenum, image: GLeglImageOES) {
+        panic!("not supported")
+    }
+
+    fn egl_image_target_renderbuffer_storage_oes(&self, target: GLenum, image: GLeglImageOES) {
+        panic!("not supported")
+    }
+
+    fn generate_mipmap(&self, target: GLenum) {
+        unsafe {
+            GenerateMipmap(target);
+        }
+    }
+
+    fn insert_event_marker_ext(&self, message: &str) {
+        panic!();
+    }
+
+    fn push_group_marker_ext(&self, message: &str) {
+        debug!("push group {}", message);
+        panic!();
+    }
+
+    fn pop_group_marker_ext(&self) {
+        debug!("pop group");
+        panic!();
+    }
+
+    fn debug_message_insert_khr(
+        &self,
+        source: GLenum,
+        type_: GLenum,
+        id: GLuint,
+        severity: GLenum,
+        message: &str,
+    ) {
+        panic!();
+    }
+
+    fn push_debug_group_khr(&self, source: GLenum, id: GLuint, message: &str) {
+        panic!();
+    }
+
+    fn pop_debug_group_khr(&self) {
+        panic!();
+    }
+
+    fn fence_sync(&self, condition: GLenum, flags: GLbitfield) -> GLsync {
+        panic!();
+        //ptr::null()
+    }
+
+    fn client_wait_sync(&self, sync: GLsync, flags: GLbitfield, timeout: GLuint64) -> GLenum {
+        panic!();
+    }
+
+    fn wait_sync(&self, sync: GLsync, flags: GLbitfield, timeout: GLuint64) {
+        panic!();
+    }
+
+    fn texture_range_apple(&self, target: GLenum, data: &[u8]) {
+        panic!();
+    }
+
+    fn delete_sync(&self, sync: GLsync) {
+        panic!();
+    }
+
+    fn gen_fences_apple(&self, n: GLsizei) -> Vec<GLuint> {
+        panic!();
+        //Vec::new()
+    }
+
+    fn delete_fences_apple(&self, fences: &[GLuint]) {
+        panic!();
+    }
+
+    fn set_fence_apple(&self, fence: GLuint) {
+        panic!();
+    }
+
+    fn finish_fence_apple(&self, fence: GLuint) {
+        panic!();
+    }
+
+    fn test_fence_apple(&self, fence: GLuint) {
+        panic!();
+    }
+
+    fn test_object_apple(&self, object: GLenum, name: GLuint) -> GLboolean {
+        panic!();
+        //0
+    }
+
+    fn finish_object_apple(&self, object: GLenum, name: GLuint) {
+        panic!();
+    }
+
+    // GL_ARB_blend_func_extended
+    fn bind_frag_data_location_indexed(
+        &self,
+        program: GLuint,
+        color_number: GLuint,
+        index: GLuint,
+        name: &str,
+    ) {
+        panic!();
+    }
+
+    fn get_frag_data_index(&self, program: GLuint, name: &str) -> GLint {
+        panic!();
+        //-1
+    }
+
+    // GL_KHR_debug
+    fn get_debug_messages(&self) -> Vec<DebugMessage> {
+        Vec::new()
+    }
+
+    fn provoking_vertex_angle(&self, _mode: GLenum) {
+        unimplemented!("This extension is GLES only");
+    }
+
+    // GL_KHR_blend_equation_advanced
+    fn blend_barrier_khr(&self) {
+        // No barrier required, so nothing to do
+    }
+
+    // GL_CHROMIUM_copy_texture
+    fn copy_texture_chromium(
+        &self,
+        _source_id: GLuint,
+        _source_level: GLint,
+        _dest_target: GLenum,
+        _dest_id: GLuint,
+        _dest_level: GLint,
+        _internal_format: GLint,
+        _dest_type: GLenum,
+        _unpack_flip_y: GLboolean,
+        _unpack_premultiply_alpha: GLboolean,
+        _unpack_unmultiply_alpha: GLboolean,
+    ) {
+        unimplemented!("This extension is GLES only");
+    }
+    fn copy_sub_texture_chromium(
+        &self,
+        _source_id: GLuint,
+        _source_level: GLint,
+        _dest_target: GLenum,
+        _dest_id: GLuint,
+        _dest_level: GLint,
+        _x_offset: GLint,
+        _y_offset: GLint,
+        _x: GLint,
+        _y: GLint,
+        _width: GLsizei,
+        _height: GLsizei,
+        _unpack_flip_y: GLboolean,
+        _unpack_premultiply_alpha: GLboolean,
+        _unpack_unmultiply_alpha: GLboolean,
+    ) {
+        unimplemented!("This extension is GLES only");
+    }
+
+    // GL_ANGLE_copy_texture_3d
+    fn copy_texture_3d_angle(
+        &self,
+        _source_id: GLuint,
+        _source_level: GLint,
+        _dest_target: GLenum,
+        _dest_id: GLuint,
+        _dest_level: GLint,
+        _internal_format: GLint,
+        _dest_type: GLenum,
+        _unpack_flip_y: GLboolean,
+        _unpack_premultiply_alpha: GLboolean,
+        _unpack_unmultiply_alpha: GLboolean,
+    ) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn copy_sub_texture_3d_angle(
+        &self,
+        _source_id: GLuint,
+        _source_level: GLint,
+        _dest_target: GLenum,
+        _dest_id: GLuint,
+        _dest_level: GLint,
+        _x_offset: GLint,
+        _y_offset: GLint,
+        _z_offset: GLint,
+        _x: GLint,
+        _y: GLint,
+        _z: GLint,
+        _width: GLsizei,
+        _height: GLsizei,
+        _depth: GLsizei,
+        _unpack_flip_y: GLboolean,
+        _unpack_premultiply_alpha: GLboolean,
+        _unpack_unmultiply_alpha: GLboolean,
+    ) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn buffer_storage(
+        &self,
+        target: GLenum,
+        size: GLsizeiptr,
+        data: *const GLvoid,
+        flags: GLbitfield,
+    ) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn flush_mapped_buffer_range(&self, target: GLenum, offset: GLintptr, length: GLsizeiptr) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn start_tiling_qcom(
+        &self,
+        x: GLuint,
+        y: GLuint,
+        width: GLuint,
+        height: GLuint,
+        preserve_mask: GLbitfield,
+    ) {
+        unimplemented!("Not supported by SWGL");
+    }
+
+    fn end_tiling_qcom(&self, preserve_mask: GLbitfield) {
+        unimplemented!("Not supported by SWGL");
+    }
+}
+
+/// A resource that is intended for sharing between threads.
+/// Locked resources such as textures or framebuffers will
+/// not allow any further modifications while it remains
+/// locked. The resource will be unlocked when LockedResource
+/// is dropped.
+pub struct LockedResource(*mut LockedTexture);
+
+unsafe impl Send for LockedResource {}
+unsafe impl Sync for LockedResource {}
+
+#[repr(u8)]
+pub enum YuvRangedColorSpace {
+    Rec601Narrow = 0,
+    Rec601Full,
+    Rec709Narrow,
+    Rec709Full,
+    Rec2020Narrow,
+    Rec2020Full,
+    GbrIdentity,
+}
+
+impl LockedResource {
+    /// Composites from a locked resource to another locked resource. The band
+    /// offset and height are relative to the destination rectangle and specify
+    /// how to clip the composition into appropriate range for this band.
+    pub fn composite(
+        &self,
+        locked_src: &LockedResource,
+        src_x: GLint,
+        src_y: GLint,
+        src_width: GLsizei,
+        src_height: GLsizei,
+        dst_x: GLint,
+        dst_y: GLint,
+        dst_width: GLsizei,
+        dst_height: GLsizei,
+        opaque: bool,
+        flip_x: bool,
+        flip_y: bool,
+        filter: GLenum,
+        clip_x: GLint,
+        clip_y: GLint,
+        clip_width: GLsizei,
+        clip_height: GLsizei,
+    ) {
+        unsafe {
+            Composite(
+                self.0,
+                locked_src.0,
+                src_x,
+                src_y,
+                src_width,
+                src_height,
+                dst_x,
+                dst_y,
+                dst_width,
+                dst_height,
+                opaque as GLboolean,
+                flip_x as GLboolean,
+                flip_y as GLboolean,
+                filter,
+                clip_x,
+                clip_y,
+                clip_width,
+                clip_height,
+            );
+        }
+    }
+
+    /// Composites from locked resources representing YUV planes
+    pub fn composite_yuv(
+        &self,
+        locked_y: &LockedResource,
+        locked_u: &LockedResource,
+        locked_v: &LockedResource,
+        color_space: YuvRangedColorSpace,
+        color_depth: GLuint,
+        src_x: GLint,
+        src_y: GLint,
+        src_width: GLsizei,
+        src_height: GLsizei,
+        dst_x: GLint,
+        dst_y: GLint,
+        dst_width: GLsizei,
+        dst_height: GLsizei,
+        flip_x: bool,
+        flip_y: bool,
+        clip_x: GLint,
+        clip_y: GLint,
+        clip_width: GLsizei,
+        clip_height: GLsizei,
+    ) {
+        unsafe {
+            CompositeYUV(
+                self.0,
+                locked_y.0,
+                locked_u.0,
+                locked_v.0,
+                color_space,
+                color_depth,
+                src_x,
+                src_y,
+                src_width,
+                src_height,
+                dst_x,
+                dst_y,
+                dst_width,
+                dst_height,
+                flip_x as GLboolean,
+                flip_y as GLboolean,
+                clip_x,
+                clip_y,
+                clip_width,
+                clip_height,
+            );
+        }
+    }
+
+    /// Get the underlying buffer for a locked resource
+    pub fn get_buffer(&self) -> (*mut c_void, i32, i32, i32) {
+        unsafe {
+            let mut width: i32 = 0;
+            let mut height: i32 = 0;
+            let mut stride: i32 = 0;
+            let data_ptr = GetResourceBuffer(self.0, &mut width, &mut height, &mut stride);
+            (data_ptr, width, height, stride)
+        }
+    }
+}
+
+impl Clone for LockedResource {
+    fn clone(&self) -> Self {
+        unsafe {
+            LockResource(self.0);
+        }
+        LockedResource(self.0)
+    }
+}
+
+impl Drop for LockedResource {
+    fn drop(&mut self) {
+        unsafe {
+            UnlockResource(self.0);
+        }
+    }
+}
diff --git a/gfx/wr/swgl/src/texture.h b/gfx/wr/swgl/src/texture.h
new file mode 100644
index 0000000000..57d27ed05c
--- /dev/null
+++ b/gfx/wr/swgl/src/texture.h
@@ -0,0 +1,1299 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+namespace glsl {
+
+using PackedRGBA8 = V16<uint8_t>;
+using WideRGBA8 = V16<uint16_t>;
+using HalfRGBA8 = V8<uint16_t>;
+
+SI WideRGBA8 unpack(PackedRGBA8 p) { return CONVERT(p, WideRGBA8); }
+
+template <int N>
+UNUSED SI VectorType<uint8_t, N> genericPackWide(VectorType<uint16_t, N> p) {
+  typedef VectorType<uint8_t, N> packed_type;
+  // Generic conversions only mask off the low byte without actually clamping
+  // like a real pack. First force the word to all 1s if it overflows, and then
+  // add on the sign bit to cause it to roll over to 0 if it was negative.
+  p = (p | (p > 255)) + (p >> 15);
+  return CONVERT(p, packed_type);
+}
+
+SI PackedRGBA8 pack(WideRGBA8 p) {
+#if USE_SSE2
+  return _mm_packus_epi16(lowHalf(p), highHalf(p));
+#elif USE_NEON
+  return vcombine_u8(vqmovun_s16(bit_cast<V8<int16_t>>(lowHalf(p))),
+                     vqmovun_s16(bit_cast<V8<int16_t>>(highHalf(p))));
+#else
+  return genericPackWide(p);
+#endif
+}
+
+using PackedR8 = V4<uint8_t>;
+using WideR8 = V4<uint16_t>;
+
+SI WideR8 unpack(PackedR8 p) { return CONVERT(p, WideR8); }
+
+SI PackedR8 pack(WideR8 p) {
+#if USE_SSE2
+  auto m = expand(p);
+  auto r = bit_cast<V16<uint8_t>>(_mm_packus_epi16(m, m));
+  return SHUFFLE(r, r, 0, 1, 2, 3);
+#elif USE_NEON
+  return lowHalf(
+      bit_cast<V8<uint8_t>>(vqmovun_s16(bit_cast<V8<int16_t>>(expand(p)))));
+#else
+  return genericPackWide(p);
+#endif
+}
+
+using PackedRG8 = V8<uint8_t>;
+using WideRG8 = V8<uint16_t>;
+
+SI PackedRG8 pack(WideRG8 p) {
+#if USE_SSE2
+  return lowHalf(bit_cast<V16<uint8_t>>(_mm_packus_epi16(p, p)));
+#elif USE_NEON
+  return bit_cast<V8<uint8_t>>(vqmovun_s16(bit_cast<V8<int16_t>>(p)));
+#else
+  return genericPackWide(p);
+#endif
+}
+
+SI I32 clampCoord(I32 coord, int limit, int base = 0) {
+#if USE_SSE2
+  return _mm_min_epi16(_mm_max_epi16(coord, _mm_set1_epi32(base)),
+                       _mm_set1_epi32(limit - 1));
+#else
+  return clamp(coord, base, limit - 1);
+#endif
+}
+
+SI int clampCoord(int coord, int limit, int base = 0) {
+  return min(max(coord, base), limit - 1);
+}
+
+template <typename T, typename S>
+SI T clamp2D(T P, S sampler) {
+  return T{clampCoord(P.x, sampler->width), clampCoord(P.y, sampler->height)};
+}
+
+SI float to_float(uint32_t x) { return x * (1.f / 255.f); }
+
+SI vec4 pixel_to_vec4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
+  U32 pixels = {a, b, c, d};
+  return vec4(cast((pixels >> 16) & 0xFF), cast((pixels >> 8) & 0xFF),
+              cast(pixels & 0xFF), cast(pixels >> 24)) *
+         (1.0f / 255.0f);
+}
+
+SI vec4 pixel_float_to_vec4(Float a, Float b, Float c, Float d) {
+  return vec4(Float{a.x, b.x, c.x, d.x}, Float{a.y, b.y, c.y, d.y},
+              Float{a.z, b.z, c.z, d.z}, Float{a.w, b.w, c.w, d.w});
+}
+
+SI ivec4 pixel_int_to_ivec4(I32 a, I32 b, I32 c, I32 d) {
+  return ivec4(I32{a.x, b.x, c.x, d.x}, I32{a.y, b.y, c.y, d.y},
+               I32{a.z, b.z, c.z, d.z}, I32{a.w, b.w, c.w, d.w});
+}
+
+SI vec4_scalar pixel_to_vec4(uint32_t p) {
+  U32 i = {(p >> 16) & 0xFF, (p >> 8) & 0xFF, p & 0xFF, p >> 24};
+  Float f = cast(i) * (1.0f / 255.0f);
+  return vec4_scalar(f.x, f.y, f.z, f.w);
+}
+
+template <typename S>
+SI vec4 fetchOffsetsRGBA8(S sampler, I32 offset) {
+  return pixel_to_vec4(sampler->buf[offset.x], sampler->buf[offset.y],
+                       sampler->buf[offset.z], sampler->buf[offset.w]);
+}
+
+template <typename S>
+vec4 texelFetchRGBA8(S sampler, ivec2 P) {
+  I32 offset = P.x + P.y * sampler->stride;
+  return fetchOffsetsRGBA8(sampler, offset);
+}
+
+template <typename S>
+SI Float fetchOffsetsR8(S sampler, I32 offset) {
+  U32 i = {
+      ((uint8_t*)sampler->buf)[offset.x], ((uint8_t*)sampler->buf)[offset.y],
+      ((uint8_t*)sampler->buf)[offset.z], ((uint8_t*)sampler->buf)[offset.w]};
+  return cast(i) * (1.0f / 255.0f);
+}
+
+template <typename S>
+vec4 texelFetchR8(S sampler, ivec2 P) {
+  I32 offset = P.x + P.y * sampler->stride;
+  return vec4(fetchOffsetsR8(sampler, offset), 0.0f, 0.0f, 1.0f);
+}
+
+template <typename S>
+SI vec4 fetchOffsetsRG8(S sampler, I32 offset) {
+  uint16_t* buf = (uint16_t*)sampler->buf;
+  U16 pixels = {buf[offset.x], buf[offset.y], buf[offset.z], buf[offset.w]};
+  Float r = CONVERT(pixels & 0xFF, Float) * (1.0f / 255.0f);
+  Float g = CONVERT(pixels >> 8, Float) * (1.0f / 255.0f);
+  return vec4(r, g, 0.0f, 1.0f);
+}
+
+template <typename S>
+vec4 texelFetchRG8(S sampler, ivec2 P) {
+  I32 offset = P.x + P.y * sampler->stride;
+  return fetchOffsetsRG8(sampler, offset);
+}
+
+template <typename S>
+SI Float fetchOffsetsR16(S sampler, I32 offset) {
+  U32 i = {
+      ((uint16_t*)sampler->buf)[offset.x], ((uint16_t*)sampler->buf)[offset.y],
+      ((uint16_t*)sampler->buf)[offset.z], ((uint16_t*)sampler->buf)[offset.w]};
+  return cast(i) * (1.0f / 65535.0f);
+}
+
+template <typename S>
+vec4 texelFetchR16(S sampler, ivec2 P) {
+  I32 offset = P.x + P.y * sampler->stride;
+  return vec4(fetchOffsetsR16(sampler, offset), 0.0f, 0.0f, 1.0f);
+}
+
+template <typename S>
+SI vec4 fetchOffsetsRG16(S sampler, I32 offset) {
+  U32 pixels = {sampler->buf[offset.x], sampler->buf[offset.y],
+                sampler->buf[offset.z], sampler->buf[offset.w]};
+  Float r = cast(pixels & 0xFFFF) * (1.0f / 65535.0f);
+  Float g = cast(pixels >> 16) * (1.0f / 65535.0f);
+  return vec4(r, g, 0.0f, 1.0f);
+}
+
+template <typename S>
+vec4 texelFetchRG16(S sampler, ivec2 P) {
+  I32 offset = P.x + P.y * sampler->stride;
+  return fetchOffsetsRG16(sampler, offset);
+}
+
+SI vec4 fetchOffsetsFloat(const uint32_t* buf, I32 offset) {
+  return pixel_float_to_vec4(*(Float*)&buf[offset.x], *(Float*)&buf[offset.y],
+                             *(Float*)&buf[offset.z], *(Float*)&buf[offset.w]);
+}
+
+SI vec4 fetchOffsetsFloat(samplerCommon* sampler, I32 offset) {
+  return fetchOffsetsFloat(sampler->buf, offset);
+}
+
+vec4 texelFetchFloat(sampler2D sampler, ivec2 P) {
+  I32 offset = P.x * 4 + P.y * sampler->stride;
+  return fetchOffsetsFloat(sampler, offset);
+}
+
+template <typename S>
+SI vec4 fetchOffsetsYUV422(S sampler, I32 offset) {
+  // Layout is 2 pixel chunks (occupying 4 bytes) organized as: G0, B, G1, R.
+  // Offset is aligned to a chunk rather than a pixel, and selector specifies
+  // pixel within the chunk.
+  I32 selector = offset & 1;
+  offset &= ~1;
+  uint16_t* buf = (uint16_t*)sampler->buf;
+  U32 pixels = {*(uint32_t*)&buf[offset.x], *(uint32_t*)&buf[offset.y],
+                *(uint32_t*)&buf[offset.z], *(uint32_t*)&buf[offset.w]};
+  Float b = CONVERT((pixels >> 8) & 0xFF, Float) * (1.0f / 255.0f);
+  Float r = CONVERT((pixels >> 24), Float) * (1.0f / 255.0f);
+  Float g =
+      CONVERT(if_then_else(-selector, pixels >> 16, pixels) & 0xFF, Float) *
+      (1.0f / 255.0f);
+  return vec4(r, g, b, 1.0f);
+}
+
+template <typename S>
+vec4 texelFetchYUV422(S sampler, ivec2 P) {
+  I32 offset = P.x + P.y * sampler->stride;
+  return fetchOffsetsYUV422(sampler, offset);
+}
+
+vec4 texelFetch(sampler2D sampler, ivec2 P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  switch (sampler->format) {
+    case TextureFormat::RGBA32F:
+      return texelFetchFloat(sampler, P);
+    case TextureFormat::RGBA8:
+      return texelFetchRGBA8(sampler, P);
+    case TextureFormat::R8:
+      return texelFetchR8(sampler, P);
+    case TextureFormat::RG8:
+      return texelFetchRG8(sampler, P);
+    case TextureFormat::R16:
+      return texelFetchR16(sampler, P);
+    case TextureFormat::RG16:
+      return texelFetchRG16(sampler, P);
+    case TextureFormat::YUV422:
+      return texelFetchYUV422(sampler, P);
+    default:
+      assert(false);
+      return vec4();
+  }
+}
+
+vec4 texelFetch(sampler2DRGBA32F sampler, ivec2 P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::RGBA32F);
+  return texelFetchFloat(sampler, P);
+}
+
+vec4 texelFetch(sampler2DRGBA8 sampler, ivec2 P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::RGBA8);
+  return texelFetchRGBA8(sampler, P);
+}
+
+vec4 texelFetch(sampler2DR8 sampler, ivec2 P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::R8);
+  return texelFetchR8(sampler, P);
+}
+
+vec4 texelFetch(sampler2DRG8 sampler, ivec2 P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::RG8);
+  return texelFetchRG8(sampler, P);
+}
+
+vec4_scalar texelFetch(sampler2D sampler, ivec2_scalar P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  if (sampler->format == TextureFormat::RGBA32F) {
+    return *(vec4_scalar*)&sampler->buf[P.x * 4 + P.y * sampler->stride];
+  } else {
+    assert(sampler->format == TextureFormat::RGBA8);
+    return pixel_to_vec4(sampler->buf[P.x + P.y * sampler->stride]);
+  }
+}
+
+vec4_scalar texelFetch(sampler2DRGBA32F sampler, ivec2_scalar P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::RGBA32F);
+  return *(vec4_scalar*)&sampler->buf[P.x * 4 + P.y * sampler->stride];
+}
+
+vec4_scalar texelFetch(sampler2DRGBA8 sampler, ivec2_scalar P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::RGBA8);
+  return pixel_to_vec4(sampler->buf[P.x + P.y * sampler->stride]);
+}
+
+vec4_scalar texelFetch(sampler2DR8 sampler, ivec2_scalar P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::R8);
+  return vec4_scalar{
+      to_float(((uint8_t*)sampler->buf)[P.x + P.y * sampler->stride]), 0.0f,
+      0.0f, 1.0f};
+}
+
+vec4_scalar texelFetch(sampler2DRG8 sampler, ivec2_scalar P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::RG8);
+  uint16_t pixel = ((uint16_t*)sampler->buf)[P.x + P.y * sampler->stride];
+  return vec4_scalar{to_float(pixel & 0xFF), to_float(pixel >> 8), 0.0f, 1.0f};
+}
+
+vec4 texelFetch(sampler2DRect sampler, ivec2 P) {
+  P = clamp2D(P, sampler);
+  switch (sampler->format) {
+    case TextureFormat::RGBA8:
+      return texelFetchRGBA8(sampler, P);
+    case TextureFormat::R8:
+      return texelFetchR8(sampler, P);
+    case TextureFormat::RG8:
+      return texelFetchRG8(sampler, P);
+    case TextureFormat::R16:
+      return texelFetchR16(sampler, P);
+    case TextureFormat::RG16:
+      return texelFetchRG16(sampler, P);
+    case TextureFormat::YUV422:
+      return texelFetchYUV422(sampler, P);
+    default:
+      assert(false);
+      return vec4();
+  }
+}
+
+SI ivec4 fetchOffsetsInt(const uint32_t* buf, I32 offset) {
+  return pixel_int_to_ivec4(*(I32*)&buf[offset.x], *(I32*)&buf[offset.y],
+                            *(I32*)&buf[offset.z], *(I32*)&buf[offset.w]);
+}
+
+SI ivec4 fetchOffsetsInt(samplerCommon* sampler, I32 offset) {
+  return fetchOffsetsInt(sampler->buf, offset);
+}
+
+ivec4 texelFetch(isampler2D sampler, ivec2 P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::RGBA32I);
+  I32 offset = P.x * 4 + P.y * sampler->stride;
+  return fetchOffsetsInt(sampler, offset);
+}
+
+ivec4_scalar texelFetch(isampler2D sampler, ivec2_scalar P, int lod) {
+  assert(lod == 0);
+  P = clamp2D(P, sampler);
+  assert(sampler->format == TextureFormat::RGBA32I);
+  return *(ivec4_scalar*)&sampler->buf[P.x * 4 + P.y * sampler->stride];
+}
+
+constexpr int MAX_TEXEL_OFFSET = 8;
+
+// Fill texelFetchOffset outside the valid texture bounds with zeroes. The
+// stride will be set to 0 so that only one row of zeroes is needed.
+static const uint32_t
+    zeroFetchBuf[MAX_TEXEL_OFFSET * sizeof(Float) / sizeof(uint32_t)] = {0};
+
+struct FetchScalar {
+  const uint32_t* buf;
+  uint32_t stride;
+};
+
+template <typename S>
+SI FetchScalar texelFetchPtr(S sampler, ivec2_scalar P, int min_x, int max_x,
+                             int min_y, int max_y) {
+  assert(max_x < MAX_TEXEL_OFFSET);
+  if (P.x < -min_x || P.x >= int(sampler->width) - max_x || P.y < -min_y ||
+      P.y >= int(sampler->height) - max_y) {
+    return FetchScalar{zeroFetchBuf, 0};
+  }
+  return FetchScalar{&sampler->buf[P.x * 4 + P.y * sampler->stride],
+                     sampler->stride};
+}
+
+SI vec4_scalar texelFetchUnchecked(sampler2D sampler, FetchScalar ptr, int x,
+                                   int y = 0) {
+  assert(sampler->format == TextureFormat::RGBA32F);
+  return *(vec4_scalar*)&ptr.buf[x * 4 + y * ptr.stride];
+}
+
+SI ivec4_scalar texelFetchUnchecked(isampler2D sampler, FetchScalar ptr, int x,
+                                    int y = 0) {
+  assert(sampler->format == TextureFormat::RGBA32I);
+  return *(ivec4_scalar*)&ptr.buf[x * 4 + y * ptr.stride];
+}
+
+struct FetchVector {
+  const uint32_t* buf;
+  I32 offset;
+  uint32_t stride;
+};
+
+template <typename S>
+SI FetchVector texelFetchPtr(S sampler, ivec2 P, int min_x, int max_x,
+                             int min_y, int max_y) {
+  assert(max_x < MAX_TEXEL_OFFSET);
+  if (test_any(P.x < -min_x || P.x >= int(sampler->width) - max_x ||
+               P.y < -min_y || P.y >= int(sampler->height) - max_y)) {
+    return FetchVector{zeroFetchBuf, I32(0), 0};
+  }
+  return FetchVector{sampler->buf, P.x * 4 + P.y * sampler->stride,
+                     sampler->stride};
+}
+
+SI vec4 texelFetchUnchecked(sampler2D sampler, FetchVector ptr, int x,
+                            int y = 0) {
+  assert(sampler->format == TextureFormat::RGBA32F);
+  return fetchOffsetsFloat(&ptr.buf[x * 4 + y * ptr.stride], ptr.offset);
+}
+
+SI ivec4 texelFetchUnchecked(isampler2D sampler, FetchVector ptr, int x,
+                             int y = 0) {
+  assert(sampler->format == TextureFormat::RGBA32I);
+  return fetchOffsetsInt(&ptr.buf[x * 4 + y * ptr.stride], ptr.offset);
+}
+
+#define texelFetchOffset(sampler, P, lod, offset) \
+  texelFetch(sampler, (P) + (offset), lod)
+
+// Scale texture coords for quantization, subtract offset for filtering
+// (assuming coords already offset to texel centers), and round to nearest
+// 1/scale increment
+template <typename T>
+SI T linearQuantize(T P, float scale) {
+  return P * scale + (0.5f - 0.5f * scale);
+}
+
+// Helper version that also scales normalized texture coords for sampler
+template <typename T, typename S>
+SI T samplerScale(S sampler, T P) {
+  P.x *= sampler->width;
+  P.y *= sampler->height;
+  return P;
+}
+
+template <typename T>
+SI T samplerScale(UNUSED sampler2DRect sampler, T P) {
+  return P;
+}
+
+template <typename T, typename S>
+SI T linearQuantize(T P, float scale, S sampler) {
+  return linearQuantize(samplerScale(sampler, P), scale);
+}
+
+// Compute clamped offset of first row for linear interpolation
+template <typename S, typename I>
+SI auto computeRow(S sampler, I i, size_t margin = 1) -> decltype(i.x) {
+  return clampCoord(i.x, sampler->width - margin) +
+         clampCoord(i.y, sampler->height) * sampler->stride;
+}
+
+// Compute clamped offset of second row for linear interpolation from first row
+template <typename S, typename I>
+SI auto computeNextRowOffset(S sampler, I i) -> decltype(i.x) {
+  return if_then_else(i.y >= 0 && i.y < int32_t(sampler->height) - 1,
+                      sampler->stride, 0);
+}
+
+// Convert X coordinate to a 2^7 scale fraction for interpolation
+template <typename S>
+SI I16 computeFracX(S sampler, ivec2 i, ivec2 frac) {
+  auto overread = i.x > int32_t(sampler->width) - 2;
+  return CONVERT((((frac.x & (i.x >= 0)) | overread) & 0x7F) - overread, I16);
+}
+
+// Convert Y coordinate to a 2^7 scale fraction for interpolation
+SI I16 computeFracNoClamp(I32 frac) { return CONVERT(frac & 0x7F, I16); }
+SI I16 computeFracY(ivec2 frac) { return computeFracNoClamp(frac.y); }
+
+struct WidePlanarRGBA8 {
+  V8<uint16_t> rg;
+  V8<uint16_t> ba;
+};
+
+template <typename S>
+SI WidePlanarRGBA8 textureLinearPlanarRGBA8(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::RGBA8);
+
+  ivec2 frac = i;
+  i >>= 7;
+
+  I32 row0 = computeRow(sampler, i);
+  I32 row1 = row0 + computeNextRowOffset(sampler, i);
+  I16 fracx = computeFracX(sampler, i, frac);
+  I16 fracy = computeFracY(frac);
+
+  auto a0 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row0.x]), V8<int16_t>);
+  auto a1 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row1.x]), V8<int16_t>);
+  a0 += ((a1 - a0) * fracy.x) >> 7;
+
+  auto b0 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row0.y]), V8<int16_t>);
+  auto b1 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row1.y]), V8<int16_t>);
+  b0 += ((b1 - b0) * fracy.y) >> 7;
+
+  auto abl = zipLow(a0, b0);
+  auto abh = zipHigh(a0, b0);
+  abl += ((abh - abl) * fracx.xyxyxyxy) >> 7;
+
+  auto c0 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row0.z]), V8<int16_t>);
+  auto c1 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row1.z]), V8<int16_t>);
+  c0 += ((c1 - c0) * fracy.z) >> 7;
+
+  auto d0 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row0.w]), V8<int16_t>);
+  auto d1 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row1.w]), V8<int16_t>);
+  d0 += ((d1 - d0) * fracy.w) >> 7;
+
+  auto cdl = zipLow(c0, d0);
+  auto cdh = zipHigh(c0, d0);
+  cdl += ((cdh - cdl) * fracx.zwzwzwzw) >> 7;
+
+  auto rg = V8<uint16_t>(zip2Low(abl, cdl));
+  auto ba = V8<uint16_t>(zip2High(abl, cdl));
+  return WidePlanarRGBA8{rg, ba};
+}
+
+template <typename S>
+vec4 textureLinearRGBA8(S sampler, vec2 P) {
+  ivec2 i(linearQuantize(P, 128, sampler));
+  auto planar = textureLinearPlanarRGBA8(sampler, i);
+  auto rg = CONVERT(planar.rg, V8<float>);
+  auto ba = CONVERT(planar.ba, V8<float>);
+  auto r = lowHalf(rg);
+  auto g = highHalf(rg);
+  auto b = lowHalf(ba);
+  auto a = highHalf(ba);
+  return vec4(b, g, r, a) * (1.0f / 255.0f);
+}
+
+template <typename S>
+static inline U16 textureLinearUnpackedR8(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::R8);
+  ivec2 frac = i;
+  i >>= 7;
+
+  I32 row0 = computeRow(sampler, i);
+  I32 row1 = row0 + computeNextRowOffset(sampler, i);
+  I16 fracx = computeFracX(sampler, i, frac);
+  I16 fracy = computeFracY(frac);
+
+  uint8_t* buf = (uint8_t*)sampler->buf;
+  auto a0 = unaligned_load<V2<uint8_t>>(&buf[row0.x]);
+  auto b0 = unaligned_load<V2<uint8_t>>(&buf[row0.y]);
+  auto c0 = unaligned_load<V2<uint8_t>>(&buf[row0.z]);
+  auto d0 = unaligned_load<V2<uint8_t>>(&buf[row0.w]);
+  auto abcd0 = CONVERT(combine(a0, b0, c0, d0), V8<int16_t>);
+
+  auto a1 = unaligned_load<V2<uint8_t>>(&buf[row1.x]);
+  auto b1 = unaligned_load<V2<uint8_t>>(&buf[row1.y]);
+  auto c1 = unaligned_load<V2<uint8_t>>(&buf[row1.z]);
+  auto d1 = unaligned_load<V2<uint8_t>>(&buf[row1.w]);
+  auto abcd1 = CONVERT(combine(a1, b1, c1, d1), V8<int16_t>);
+
+  abcd0 += ((abcd1 - abcd0) * fracy.xxyyzzww) >> 7;
+
+  abcd0 = SHUFFLE(abcd0, abcd0, 0, 2, 4, 6, 1, 3, 5, 7);
+  auto abcdl = lowHalf(abcd0);
+  auto abcdh = highHalf(abcd0);
+  abcdl += ((abcdh - abcdl) * fracx) >> 7;
+
+  return U16(abcdl);
+}
+
+template <typename S>
+vec4 textureLinearR8(S sampler, vec2 P) {
+  assert(sampler->format == TextureFormat::R8);
+
+  ivec2 i(linearQuantize(P, 128, sampler));
+  Float r = CONVERT(textureLinearUnpackedR8(sampler, i), Float);
+  return vec4(r * (1.0f / 255.0f), 0.0f, 0.0f, 1.0f);
+}
+
+struct WidePlanarRG8 {
+  V8<uint16_t> rg;
+};
+
+template <typename S>
+SI WidePlanarRG8 textureLinearPlanarRG8(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::RG8);
+
+  ivec2 frac = i;
+  i >>= 7;
+
+  I32 row0 = computeRow(sampler, i);
+  I32 row1 = row0 + computeNextRowOffset(sampler, i);
+  I16 fracx = computeFracX(sampler, i, frac);
+  I16 fracy = computeFracY(frac);
+
+  uint16_t* buf = (uint16_t*)sampler->buf;
+
+  // Load RG bytes for two adjacent pixels - rgRG
+  auto a0 = unaligned_load<V4<uint8_t>>(&buf[row0.x]);
+  auto b0 = unaligned_load<V4<uint8_t>>(&buf[row0.y]);
+  auto ab0 = CONVERT(combine(a0, b0), V8<int16_t>);
+  // Load two pixels for next row
+  auto a1 = unaligned_load<V4<uint8_t>>(&buf[row1.x]);
+  auto b1 = unaligned_load<V4<uint8_t>>(&buf[row1.y]);
+  auto ab1 = CONVERT(combine(a1, b1), V8<int16_t>);
+  // Blend rows
+  ab0 += ((ab1 - ab0) * fracy.xxxxyyyy) >> 7;
+
+  auto c0 = unaligned_load<V4<uint8_t>>(&buf[row0.z]);
+  auto d0 = unaligned_load<V4<uint8_t>>(&buf[row0.w]);
+  auto cd0 = CONVERT(combine(c0, d0), V8<int16_t>);
+  auto c1 = unaligned_load<V4<uint8_t>>(&buf[row1.z]);
+  auto d1 = unaligned_load<V4<uint8_t>>(&buf[row1.w]);
+  auto cd1 = CONVERT(combine(c1, d1), V8<int16_t>);
+  // Blend rows
+  cd0 += ((cd1 - cd0) * fracy.zzzzwwww) >> 7;
+
+  // ab = a.rgRG,b.rgRG
+  // cd = c.rgRG,d.rgRG
+  // ... ac = ar,cr,ag,cg,aR,cR,aG,cG
+  // ... bd = br,dr,bg,dg,bR,dR,bG,dG
+  auto ac = zipLow(ab0, cd0);
+  auto bd = zipHigh(ab0, cd0);
+  // ar,br,cr,dr,ag,bg,cg,dg
+  // aR,bR,cR,dR,aG,bG,cG,dG
+  auto abcdl = zipLow(ac, bd);
+  auto abcdh = zipHigh(ac, bd);
+  // Blend columns
+  abcdl += ((abcdh - abcdl) * fracx.xyzwxyzw) >> 7;
+
+  auto rg = V8<uint16_t>(abcdl);
+  return WidePlanarRG8{rg};
+}
+
+template <typename S>
+vec4 textureLinearRG8(S sampler, vec2 P) {
+  ivec2 i(linearQuantize(P, 128, sampler));
+  auto planar = textureLinearPlanarRG8(sampler, i);
+  auto rg = CONVERT(planar.rg, V8<float>) * (1.0f / 255.0f);
+  auto r = lowHalf(rg);
+  auto g = highHalf(rg);
+  return vec4(r, g, 0.0f, 1.0f);
+}
+
+// Samples R16 texture with linear filtering and returns results packed as
+// signed I16. One bit of precision is shifted away from the bottom end to
+// accommodate the sign bit, so only 15 bits of precision is left.
+template <typename S>
+static inline I16 textureLinearUnpackedR16(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::R16);
+
+  ivec2 frac = i;
+  i >>= 7;
+
+  I32 row0 = computeRow(sampler, i);
+  I32 row1 = row0 + computeNextRowOffset(sampler, i);
+
+  I16 fracx =
+      CONVERT(
+          ((frac.x & (i.x >= 0)) | (i.x > int32_t(sampler->width) - 2)) & 0x7F,
+          I16)
+      << 8;
+  I16 fracy = computeFracY(frac) << 8;
+
+  // Sample the 16 bit data for both rows
+  uint16_t* buf = (uint16_t*)sampler->buf;
+  auto a0 = unaligned_load<V2<uint16_t>>(&buf[row0.x]);
+  auto b0 = unaligned_load<V2<uint16_t>>(&buf[row0.y]);
+  auto c0 = unaligned_load<V2<uint16_t>>(&buf[row0.z]);
+  auto d0 = unaligned_load<V2<uint16_t>>(&buf[row0.w]);
+  auto abcd0 = CONVERT(combine(a0, b0, c0, d0) >> 1, V8<int16_t>);
+
+  auto a1 = unaligned_load<V2<uint16_t>>(&buf[row1.x]);
+  auto b1 = unaligned_load<V2<uint16_t>>(&buf[row1.y]);
+  auto c1 = unaligned_load<V2<uint16_t>>(&buf[row1.z]);
+  auto d1 = unaligned_load<V2<uint16_t>>(&buf[row1.w]);
+  auto abcd1 = CONVERT(combine(a1, b1, c1, d1) >> 1, V8<int16_t>);
+
+  // The samples occupy 15 bits and the fraction occupies 15 bits, so that when
+  // they are multiplied together, the new scaled sample will fit in the high
+  // 14 bits of the result. It is left shifted once to make it 15 bits again
+  // for the final multiply.
+#if USE_SSE2
+  abcd0 += bit_cast<V8<int16_t>>(_mm_mulhi_epi16(abcd1 - abcd0, fracy.xxyyzzww))
+           << 1;
+#elif USE_NEON
+  // NEON has a convenient instruction that does both the multiply and the
+  // doubling, so doesn't need an extra shift.
+  abcd0 += bit_cast<V8<int16_t>>(vqrdmulhq_s16(abcd1 - abcd0, fracy.xxyyzzww));
+#else
+  abcd0 += CONVERT((CONVERT(abcd1 - abcd0, V8<int32_t>) *
+                    CONVERT(fracy.xxyyzzww, V8<int32_t>)) >>
+                       16,
+                   V8<int16_t>)
+           << 1;
+#endif
+
+  abcd0 = SHUFFLE(abcd0, abcd0, 0, 2, 4, 6, 1, 3, 5, 7);
+  auto abcdl = lowHalf(abcd0);
+  auto abcdh = highHalf(abcd0);
+#if USE_SSE2
+  abcdl += lowHalf(bit_cast<V8<int16_t>>(
+               _mm_mulhi_epi16(expand(abcdh - abcdl), expand(fracx))))
+           << 1;
+#elif USE_NEON
+  abcdl += bit_cast<V4<int16_t>>(vqrdmulh_s16(abcdh - abcdl, fracx));
+#else
+  abcdl += CONVERT((CONVERT(abcdh - abcdl, V4<int32_t>) *
+                    CONVERT(fracx, V4<int32_t>)) >>
+                       16,
+                   V4<int16_t>)
+           << 1;
+#endif
+
+  return abcdl;
+}
+
+template <typename S>
+vec4 textureLinearR16(S sampler, vec2 P) {
+  assert(sampler->format == TextureFormat::R16);
+
+  ivec2 i(linearQuantize(P, 128, sampler));
+  Float r = CONVERT(textureLinearUnpackedR16(sampler, i), Float);
+  return vec4(r * (1.0f / 32767.0f), 0.0f, 0.0f, 1.0f);
+}
+
+// Samples RG16 texture with linear filtering and returns results packed as
+// signed I16. One bit of precision is shifted away from the bottom end to
+// accommodate the sign bit, so only 15 bits of precision is left.
+template <typename S>
+static inline V8<int16_t> textureLinearUnpackedRG16(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::RG16);
+
+  ivec2 frac = i;
+  i >>= 7;
+
+  I32 row0 = computeRow(sampler, i);
+  I32 row1 = row0 + computeNextRowOffset(sampler, i);
+
+  I16 fracx =
+      CONVERT(
+          ((frac.x & (i.x >= 0)) | (i.x > int32_t(sampler->width) - 2)) & 0x7F,
+          I16)
+      << 8;
+  I16 fracy = computeFracY(frac) << 8;
+
+  // Sample the 2x16 bit data for both rows
+  auto a0 = unaligned_load<V4<uint16_t>>(&sampler->buf[row0.x]);
+  auto b0 = unaligned_load<V4<uint16_t>>(&sampler->buf[row0.y]);
+  auto ab0 = CONVERT(combine(a0, b0) >> 1, V8<int16_t>);
+  auto c0 = unaligned_load<V4<uint16_t>>(&sampler->buf[row0.z]);
+  auto d0 = unaligned_load<V4<uint16_t>>(&sampler->buf[row0.w]);
+  auto cd0 = CONVERT(combine(c0, d0) >> 1, V8<int16_t>);
+
+  auto a1 = unaligned_load<V4<uint16_t>>(&sampler->buf[row1.x]);
+  auto b1 = unaligned_load<V4<uint16_t>>(&sampler->buf[row1.y]);
+  auto ab1 = CONVERT(combine(a1, b1) >> 1, V8<int16_t>);
+  auto c1 = unaligned_load<V4<uint16_t>>(&sampler->buf[row1.z]);
+  auto d1 = unaligned_load<V4<uint16_t>>(&sampler->buf[row1.w]);
+  auto cd1 = CONVERT(combine(c1, d1) >> 1, V8<int16_t>);
+
+  // The samples occupy 15 bits and the fraction occupies 15 bits, so that when
+  // they are multiplied together, the new scaled sample will fit in the high
+  // 14 bits of the result. It is left shifted once to make it 15 bits again
+  // for the final multiply.
+#if USE_SSE2
+  ab0 += bit_cast<V8<int16_t>>(_mm_mulhi_epi16(ab1 - ab0, fracy.xxxxyyyy)) << 1;
+  cd0 += bit_cast<V8<int16_t>>(_mm_mulhi_epi16(cd1 - cd0, fracy.zzzzwwww)) << 1;
+#elif USE_NEON
+  // NEON has a convenient instruction that does both the multiply and the
+  // doubling, so doesn't need an extra shift.
+  ab0 += bit_cast<V8<int16_t>>(vqrdmulhq_s16(ab1 - ab0, fracy.xxxxyyyy));
+  cd0 += bit_cast<V8<int16_t>>(vqrdmulhq_s16(cd1 - cd0, fracy.zzzzwwww));
+#else
+  ab0 += CONVERT((CONVERT(ab1 - ab0, V8<int32_t>) *
+                  CONVERT(fracy.xxxxyyyy, V8<int32_t>)) >>
+                     16,
+                 V8<int16_t>)
+         << 1;
+  cd0 += CONVERT((CONVERT(cd1 - cd0, V8<int32_t>) *
+                  CONVERT(fracy.zzzzwwww, V8<int32_t>)) >>
+                     16,
+                 V8<int16_t>)
+         << 1;
+#endif
+
+  // ab = a.rgRG,b.rgRG
+  // cd = c.rgRG,d.rgRG
+  // ... ac = a.rg,c.rg,a.RG,c.RG
+  // ... bd = b.rg,d.rg,b.RG,d.RG
+  auto ac = zip2Low(ab0, cd0);
+  auto bd = zip2High(ab0, cd0);
+  // a.rg,b.rg,c.rg,d.rg
+  // a.RG,b.RG,c.RG,d.RG
+  auto abcdl = zip2Low(ac, bd);
+  auto abcdh = zip2High(ac, bd);
+  // Blend columns
+#if USE_SSE2
+  abcdl += bit_cast<V8<int16_t>>(_mm_mulhi_epi16(abcdh - abcdl, fracx.xxyyzzww))
+           << 1;
+#elif USE_NEON
+  abcdl += bit_cast<V8<int16_t>>(vqrdmulhq_s16(abcdh - abcdl, fracx.xxyyzzww));
+#else
+  abcdl += CONVERT((CONVERT(abcdh - abcdl, V8<int32_t>) *
+                    CONVERT(fracx.xxyyzzww, V8<int32_t>)) >>
+                       16,
+                   V8<int16_t>)
+           << 1;
+#endif
+
+  return abcdl;
+}
+
+template <typename S>
+vec4 textureLinearRG16(S sampler, vec2 P) {
+  assert(sampler->format == TextureFormat::RG16);
+
+  ivec2 i(linearQuantize(P, 128, sampler));
+  auto rg = bit_cast<V4<int32_t>>(textureLinearUnpackedRG16(sampler, i));
+  auto r = cast(rg & 0xFFFF) * (1.0f / 32767.0f);
+  auto g = cast(rg >> 16) * (1.0f / 32767.0f);
+  return vec4(r, g, 0.0f, 1.0f);
+}
+
+using PackedRGBA32F = V16<float>;
+using WideRGBA32F = V16<float>;
+
+template <typename S>
+vec4 textureLinearRGBA32F(S sampler, vec2 P) {
+  assert(sampler->format == TextureFormat::RGBA32F);
+  P = samplerScale(sampler, P);
+  P -= 0.5f;
+  vec2 f = floor(P);
+  vec2 r = P - f;
+  ivec2 i(f);
+  ivec2 c(clampCoord(i.x, sampler->width - 1),
+          clampCoord(i.y, sampler->height));
+  r.x = if_then_else(i.x >= 0, if_then_else(i.x < sampler->width - 1, r.x, 1.0),
+                     0.0f);
+  I32 offset0 = c.x * 4 + c.y * sampler->stride;
+  I32 offset1 = offset0 + computeNextRowOffset(sampler, i);
+
+  Float c0 = mix(mix(*(Float*)&sampler->buf[offset0.x],
+                     *(Float*)&sampler->buf[offset0.x + 4], r.x),
+                 mix(*(Float*)&sampler->buf[offset1.x],
+                     *(Float*)&sampler->buf[offset1.x + 4], r.x),
+                 r.y);
+  Float c1 = mix(mix(*(Float*)&sampler->buf[offset0.y],
+                     *(Float*)&sampler->buf[offset0.y + 4], r.x),
+                 mix(*(Float*)&sampler->buf[offset1.y],
+                     *(Float*)&sampler->buf[offset1.y + 4], r.x),
+                 r.y);
+  Float c2 = mix(mix(*(Float*)&sampler->buf[offset0.z],
+                     *(Float*)&sampler->buf[offset0.z + 4], r.x),
+                 mix(*(Float*)&sampler->buf[offset1.z],
+                     *(Float*)&sampler->buf[offset1.z + 4], r.x),
+                 r.y);
+  Float c3 = mix(mix(*(Float*)&sampler->buf[offset0.w],
+                     *(Float*)&sampler->buf[offset0.w + 4], r.x),
+                 mix(*(Float*)&sampler->buf[offset1.w],
+                     *(Float*)&sampler->buf[offset1.w + 4], r.x),
+                 r.y);
+  return pixel_float_to_vec4(c0, c1, c2, c3);
+}
+
+struct WidePlanarYUV8 {
+  U16 y;
+  U16 u;
+  U16 v;
+};
+
+template <typename S>
+SI WidePlanarYUV8 textureLinearPlanarYUV422(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::YUV422);
+
+  ivec2 frac = i;
+  i >>= 7;
+
+  I32 row0 = computeRow(sampler, i, 2);
+  // Layout is 2 pixel chunks (occupying 4 bytes) organized as: G0, B, G1, R.
+  // Get the selector for the pixel within the chunk.
+  I32 selector = row0 & 1;
+  // Align the row index to the chunk.
+  row0 &= ~1;
+  I32 row1 = row0 + computeNextRowOffset(sampler, i);
+  // G only needs to be clamped to a pixel boundary for safe interpolation,
+  // whereas the BR fraction needs to be clamped 1 extra pixel inside to a chunk
+  // boundary.
+  frac.x &= (i.x >= 0);
+  auto fracx =
+      CONVERT(combine(frac.x | (i.x > int32_t(sampler->width) - 3),
+                      (frac.x >> 1) | (i.x > int32_t(sampler->width) - 3)) &
+                  0x7F,
+              V8<int16_t>);
+  I16 fracy = computeFracY(frac);
+
+  uint16_t* buf = (uint16_t*)sampler->buf;
+
+  // Load bytes for two adjacent chunks - g0,b,g1,r,G0,B,G1,R
+  // We always need to interpolate between (b,r) and (B,R).
+  // Depending on selector we need to either interpolate between g0 and g1
+  // or between g1 and G0. So for now we just interpolate both cases for g
+  // and will select the appropriate one on output.
+  auto a0 = CONVERT(unaligned_load<V8<uint8_t>>(&buf[row0.x]), V8<int16_t>);
+  auto a1 = CONVERT(unaligned_load<V8<uint8_t>>(&buf[row1.x]), V8<int16_t>);
+  // Combine with next row.
+  a0 += ((a1 - a0) * fracy.x) >> 7;
+
+  auto b0 = CONVERT(unaligned_load<V8<uint8_t>>(&buf[row0.y]), V8<int16_t>);
+  auto b1 = CONVERT(unaligned_load<V8<uint8_t>>(&buf[row1.y]), V8<int16_t>);
+  b0 += ((b1 - b0) * fracy.y) >> 7;
+
+  auto c0 = CONVERT(unaligned_load<V8<uint8_t>>(&buf[row0.z]), V8<int16_t>);
+  auto c1 = CONVERT(unaligned_load<V8<uint8_t>>(&buf[row1.z]), V8<int16_t>);
+  c0 += ((c1 - c0) * fracy.z) >> 7;
+
+  auto d0 = CONVERT(unaligned_load<V8<uint8_t>>(&buf[row0.w]), V8<int16_t>);
+  auto d1 = CONVERT(unaligned_load<V8<uint8_t>>(&buf[row1.w]), V8<int16_t>);
+  d0 += ((d1 - d0) * fracy.w) >> 7;
+
+  // Shuffle things around so we end up with g0,g0,g0,g0,b,b,b,b and
+  // g1,g1,g1,g1,r,r,r,r.
+  auto abl = zipLow(a0, b0);
+  auto cdl = zipLow(c0, d0);
+  auto g0b = zip2Low(abl, cdl);
+  auto g1r = zip2High(abl, cdl);
+
+  // Need to zip g1,B,G0,R. Instead of using a bunch of complicated masks and
+  // and shifts, just shuffle here instead... We finally end up with
+  // g1,g1,g1,g1,B,B,B,B and G0,G0,G0,G0,R,R,R,R.
+  auto abh = SHUFFLE(a0, b0, 2, 10, 5, 13, 4, 12, 7, 15);
+  auto cdh = SHUFFLE(c0, d0, 2, 10, 5, 13, 4, 12, 7, 15);
+  auto g1B = zip2Low(abh, cdh);
+  auto G0R = zip2High(abh, cdh);
+
+  // Finally interpolate between adjacent columns.
+  g0b += ((g1B - g0b) * fracx) >> 7;
+  g1r += ((G0R - g1r) * fracx) >> 7;
+
+  // Choose either g0 or g1 based on selector.
+  return WidePlanarYUV8{
+      U16(if_then_else(CONVERT(-selector, I16), lowHalf(g1r), lowHalf(g0b))),
+      U16(highHalf(g0b)), U16(highHalf(g1r))};
+}
+
+template <typename S>
+vec4 textureLinearYUV422(S sampler, vec2 P) {
+  ivec2 i(linearQuantize(P, 128, sampler));
+  auto planar = textureLinearPlanarYUV422(sampler, i);
+  auto y = CONVERT(planar.y, Float) * (1.0f / 255.0f);
+  auto u = CONVERT(planar.u, Float) * (1.0f / 255.0f);
+  auto v = CONVERT(planar.v, Float) * (1.0f / 255.0f);
+  return vec4(v, y, u, 1.0f);
+}
+
+SI vec4 texture(sampler2D sampler, vec2 P) {
+  if (sampler->filter == TextureFilter::LINEAR) {
+    switch (sampler->format) {
+      case TextureFormat::RGBA32F:
+        return textureLinearRGBA32F(sampler, P);
+      case TextureFormat::RGBA8:
+        return textureLinearRGBA8(sampler, P);
+      case TextureFormat::R8:
+        return textureLinearR8(sampler, P);
+      case TextureFormat::RG8:
+        return textureLinearRG8(sampler, P);
+      case TextureFormat::R16:
+        return textureLinearR16(sampler, P);
+      case TextureFormat::RG16:
+        return textureLinearRG16(sampler, P);
+      case TextureFormat::YUV422:
+        return textureLinearYUV422(sampler, P);
+      default:
+        assert(false);
+        return vec4();
+    }
+  } else {
+    ivec2 coord(roundzero(P.x, sampler->width),
+                roundzero(P.y, sampler->height));
+    return texelFetch(sampler, coord, 0);
+  }
+}
+
+vec4 texture(sampler2DRect sampler, vec2 P) {
+  if (sampler->filter == TextureFilter::LINEAR) {
+    switch (sampler->format) {
+      case TextureFormat::RGBA8:
+        return textureLinearRGBA8(sampler, P);
+      case TextureFormat::R8:
+        return textureLinearR8(sampler, P);
+      case TextureFormat::RG8:
+        return textureLinearRG8(sampler, P);
+      case TextureFormat::R16:
+        return textureLinearR16(sampler, P);
+      case TextureFormat::RG16:
+        return textureLinearRG16(sampler, P);
+      case TextureFormat::YUV422:
+        return textureLinearYUV422(sampler, P);
+      default:
+        assert(false);
+        return vec4();
+    }
+  } else {
+    ivec2 coord(roundzero(P.x, 1.0f), roundzero(P.y, 1.0f));
+    return texelFetch(sampler, coord);
+  }
+}
+
+template <typename S>
+vec4_scalar texture(S sampler, vec2_scalar P) {
+  return force_scalar(texture(sampler, vec2(P)));
+}
+
+ivec2_scalar textureSize(sampler2D sampler, int) {
+  return ivec2_scalar{int32_t(sampler->width), int32_t(sampler->height)};
+}
+
+ivec2_scalar textureSize(sampler2DRect sampler) {
+  return ivec2_scalar{int32_t(sampler->width), int32_t(sampler->height)};
+}
+
+template <typename S>
+static WideRGBA8 textureLinearUnpackedRGBA8(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::RGBA8);
+  ivec2 frac = i;
+  i >>= 7;
+
+  I32 row0 = computeRow(sampler, i);
+  I32 row1 = row0 + computeNextRowOffset(sampler, i);
+  I16 fracx = computeFracX(sampler, i, frac);
+  I16 fracy = computeFracY(frac);
+
+  auto a0 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row0.x]), V8<int16_t>);
+  auto a1 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row1.x]), V8<int16_t>);
+  a0 += ((a1 - a0) * fracy.x) >> 7;
+
+  auto b0 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row0.y]), V8<int16_t>);
+  auto b1 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row1.y]), V8<int16_t>);
+  b0 += ((b1 - b0) * fracy.y) >> 7;
+
+  auto abl = combine(lowHalf(a0), lowHalf(b0));
+  auto abh = combine(highHalf(a0), highHalf(b0));
+  abl += ((abh - abl) * fracx.xxxxyyyy) >> 7;
+
+  auto c0 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row0.z]), V8<int16_t>);
+  auto c1 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row1.z]), V8<int16_t>);
+  c0 += ((c1 - c0) * fracy.z) >> 7;
+
+  auto d0 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row0.w]), V8<int16_t>);
+  auto d1 =
+      CONVERT(unaligned_load<V8<uint8_t>>(&sampler->buf[row1.w]), V8<int16_t>);
+  d0 += ((d1 - d0) * fracy.w) >> 7;
+
+  auto cdl = combine(lowHalf(c0), lowHalf(d0));
+  auto cdh = combine(highHalf(c0), highHalf(d0));
+  cdl += ((cdh - cdl) * fracx.zzzzwwww) >> 7;
+
+  return combine(HalfRGBA8(abl), HalfRGBA8(cdl));
+}
+
+template <typename S>
+static PackedRGBA8 textureLinearPackedRGBA8(S sampler, ivec2 i) {
+  return pack(textureLinearUnpackedRGBA8(sampler, i));
+}
+
+template <typename S>
+static PackedRGBA8 textureNearestPackedRGBA8(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::RGBA8);
+  I32 row = computeRow(sampler, i, 0);
+  return combine(unaligned_load<V4<uint8_t>>(&sampler->buf[row.x]),
+                 unaligned_load<V4<uint8_t>>(&sampler->buf[row.y]),
+                 unaligned_load<V4<uint8_t>>(&sampler->buf[row.z]),
+                 unaligned_load<V4<uint8_t>>(&sampler->buf[row.w]));
+}
+
+template <typename S>
+static PackedR8 textureLinearPackedR8(S sampler, ivec2 i) {
+  return pack(textureLinearUnpackedR8(sampler, i));
+}
+
+template <typename S>
+static WideRG8 textureLinearUnpackedRG8(S sampler, ivec2 i) {
+  assert(sampler->format == TextureFormat::RG8);
+  ivec2 frac = i & 0x7F;
+  i >>= 7;
+
+  I32 row0 = computeRow(sampler, i);
+  I32 row1 = row0 + computeNextRowOffset(sampler, i);
+  I16 fracx = computeFracX(sampler, i, frac);
+  I16 fracy = computeFracY(frac);
+
+  uint16_t* buf = (uint16_t*)sampler->buf;
+
+  // Load RG bytes for two adjacent pixels - rgRG
+  auto a0 = unaligned_load<V4<uint8_t>>(&buf[row0.x]);
+  auto b0 = unaligned_load<V4<uint8_t>>(&buf[row0.y]);
+  auto ab0 = CONVERT(combine(a0, b0), V8<int16_t>);
+  // Load two pixels for next row
+  auto a1 = unaligned_load<V4<uint8_t>>(&buf[row1.x]);
+  auto b1 = unaligned_load<V4<uint8_t>>(&buf[row1.y]);
+  auto ab1 = CONVERT(combine(a1, b1), V8<int16_t>);
+  // Blend rows
+  ab0 += ((ab1 - ab0) * fracy.xxxxyyyy) >> 7;
+
+  auto c0 = unaligned_load<V4<uint8_t>>(&buf[row0.z]);
+  auto d0 = unaligned_load<V4<uint8_t>>(&buf[row0.w]);
+  auto cd0 = CONVERT(combine(c0, d0), V8<int16_t>);
+  auto c1 = unaligned_load<V4<uint8_t>>(&buf[row1.z]);
+  auto d1 = unaligned_load<V4<uint8_t>>(&buf[row1.w]);
+  auto cd1 = CONVERT(combine(c1, d1), V8<int16_t>);
+  // Blend rows
+  cd0 += ((cd1 - cd0) * fracy.zzzzwwww) >> 7;
+
+  // ab = a.rgRG,b.rgRG
+  // cd = c.rgRG,d.rgRG
+  // ... ac = a.rg,c.rg,a.RG,c.RG
+  // ... bd = b.rg,d.rg,b.RG,d.RG
+  auto ac = zip2Low(ab0, cd0);
+  auto bd = zip2High(ab0, cd0);
+  // a.rg,b.rg,c.rg,d.rg
+  // a.RG,b.RG,c.RG,d.RG
+  auto abcdl = zip2Low(ac, bd);
+  auto abcdh = zip2High(ac, bd);
+  // Blend columns
+  abcdl += ((abcdh - abcdl) * fracx.xxyyzzww) >> 7;
+
+  return WideRG8(abcdl);
+}
+
+template <typename S>
+static PackedRG8 textureLinearPackedRG8(S sampler, ivec2 i) {
+  return pack(textureLinearUnpackedRG8(sampler, i));
+}
+
+template <int N>
+static ALWAYS_INLINE VectorType<uint16_t, N> addsat(VectorType<uint16_t, N> x,
+                                                    VectorType<uint16_t, N> y) {
+  auto r = x + y;
+  return r | (r < x);
+}
+
+template <typename P, typename S>
+static VectorType<uint16_t, 4 * sizeof(P)> gaussianBlurHorizontal(
+    S sampler, const ivec2_scalar& i, int minX, int maxX, int radius,
+    float coeff, float coeffStep) {
+  // Packed and unpacked vectors for a chunk of the given pixel type.
+  typedef VectorType<uint8_t, 4 * sizeof(P)> packed_type;
+  typedef VectorType<uint16_t, 4 * sizeof(P)> unpacked_type;
+
+  // Pre-scale the coefficient by 8 bits of fractional precision, so that when
+  // the sample is multiplied by it, it will yield a 16 bit unsigned integer
+  // that will use all 16 bits of precision to accumulate the sum.
+  coeff *= 1 << 8;
+  float coeffStep2 = coeffStep * coeffStep;
+
+  int row = computeRow(sampler, i);
+  P* buf = (P*)sampler->buf;
+  auto pixelsRight = unaligned_load<V4<P>>(&buf[row]);
+  auto pixelsLeft = pixelsRight;
+  auto sum = CONVERT(bit_cast<packed_type>(pixelsRight), unpacked_type) *
+             uint16_t(coeff + 0.5f);
+
+  // Here we use some trickery to reuse the pixels within a chunk, shifted over
+  // by one pixel, to get the next sample for the entire chunk. This allows us
+  // to sample only one pixel for each offset across the entire chunk in both
+  // the left and right directions. To avoid clamping within the loop to the
+  // texture bounds, we compute the valid radius that doesn't require clamping
+  // and fall back to a slower clamping loop outside of that valid radius.
+  int offset = 1;
+  // The left bound is how much we can offset the sample before the start of
+  // the row bounds.
+  int leftBound = i.x - max(minX, 0);
+  // The right bound is how much we can offset the sample before the end of the
+  // row bounds.
+  int rightBound = min(maxX, sampler->width - 1) - i.x;
+  int validRadius = min(radius, min(leftBound, rightBound - (4 - 1)));
+  for (; offset <= validRadius; offset++) {
+    // Overwrite the pixel that needs to be shifted out with the new pixel, and
+    // shift it into the correct location.
+    pixelsRight.x = unaligned_load<P>(&buf[row + offset + 4 - 1]);
+    pixelsRight = pixelsRight.yzwx;
+    pixelsLeft = pixelsLeft.wxyz;
+    pixelsLeft.x = unaligned_load<P>(&buf[row - offset]);
+
+    // Accumulate the Gaussian coefficients step-wise.
+    coeff *= coeffStep;
+    coeffStep *= coeffStep2;
+
+    // Both left and right samples at this offset use the same coefficient.
+    sum = addsat(sum,
+                 (CONVERT(bit_cast<packed_type>(pixelsRight), unpacked_type) +
+                  CONVERT(bit_cast<packed_type>(pixelsLeft), unpacked_type)) *
+                     uint16_t(coeff + 0.5f));
+  }
+
+  for (; offset <= radius; offset++) {
+    pixelsRight.x =
+        unaligned_load<P>(&buf[row + min(offset + 4 - 1, rightBound)]);
+    pixelsRight = pixelsRight.yzwx;
+    pixelsLeft = pixelsLeft.wxyz;
+    pixelsLeft.x = unaligned_load<P>(&buf[row - min(offset, leftBound)]);
+
+    coeff *= coeffStep;
+    coeffStep *= coeffStep2;
+
+    sum = addsat(sum,
+                 (CONVERT(bit_cast<packed_type>(pixelsRight), unpacked_type) +
+                  CONVERT(bit_cast<packed_type>(pixelsLeft), unpacked_type)) *
+                     uint16_t(coeff + 0.5f));
+  }
+
+  // Shift away the intermediate precision.
+  return sum >> 8;
+}
+
+template <typename P, typename S>
+static VectorType<uint16_t, 4 * sizeof(P)> gaussianBlurVertical(
+    S sampler, const ivec2_scalar& i, int minY, int maxY, int radius,
+    float coeff, float coeffStep) {
+  // Packed and unpacked vectors for a chunk of the given pixel type.
+  typedef VectorType<uint8_t, 4 * sizeof(P)> packed_type;
+  typedef VectorType<uint16_t, 4 * sizeof(P)> unpacked_type;
+
+  // Pre-scale the coefficient by 8 bits of fractional precision, so that when
+  // the sample is multiplied by it, it will yield a 16 bit unsigned integer
+  // that will use all 16 bits of precision to accumulate the sum.
+  coeff *= 1 << 8;
+  float coeffStep2 = coeffStep * coeffStep;
+
+  int rowAbove = computeRow(sampler, i);
+  int rowBelow = rowAbove;
+  P* buf = (P*)sampler->buf;
+  auto pixels = unaligned_load<V4<P>>(&buf[rowAbove]);
+  auto sum = CONVERT(bit_cast<packed_type>(pixels), unpacked_type) *
+             uint16_t(coeff + 0.5f);
+
+  // For the vertical loop we can't be quite as creative with reusing old values
+  // as we were in the horizontal loop. We just do the obvious implementation of
+  // loading a chunk from each row in turn and accumulating it into the sum. We
+  // compute a valid radius within which we don't need to clamp the sampled row
+  // and use that to avoid any clamping in the main inner loop. We fall back to
+  // a slower clamping loop outside of that valid radius.
+  int offset = 1;
+  int belowBound = i.y - max(minY, 0);
+  int aboveBound = min(maxY, sampler->height - 1) - i.y;
+  int validRadius = min(radius, min(belowBound, aboveBound));
+  for (; offset <= validRadius; offset++) {
+    rowAbove += sampler->stride;
+    rowBelow -= sampler->stride;
+    auto pixelsAbove = unaligned_load<V4<P>>(&buf[rowAbove]);
+    auto pixelsBelow = unaligned_load<V4<P>>(&buf[rowBelow]);
+
+    // Accumulate the Gaussian coefficients step-wise.
+    coeff *= coeffStep;
+    coeffStep *= coeffStep2;
+
+    // Both above and below samples at this offset use the same coefficient.
+    sum = addsat(sum,
+                 (CONVERT(bit_cast<packed_type>(pixelsAbove), unpacked_type) +
+                  CONVERT(bit_cast<packed_type>(pixelsBelow), unpacked_type)) *
+                     uint16_t(coeff + 0.5f));
+  }
+
+  for (; offset <= radius; offset++) {
+    if (offset <= aboveBound) {
+      rowAbove += sampler->stride;
+    }
+    if (offset <= belowBound) {
+      rowBelow -= sampler->stride;
+    }
+    auto pixelsAbove = unaligned_load<V4<P>>(&buf[rowAbove]);
+    auto pixelsBelow = unaligned_load<V4<P>>(&buf[rowBelow]);
+
+    coeff *= coeffStep;
+    coeffStep *= coeffStep2;
+
+    sum = addsat(sum,
+                 (CONVERT(bit_cast<packed_type>(pixelsAbove), unpacked_type) +
+                  CONVERT(bit_cast<packed_type>(pixelsBelow), unpacked_type)) *
+                     uint16_t(coeff + 0.5f));
+  }
+
+  // Shift away the intermediate precision.
+  return sum >> 8;
+}
+
+}  // namespace glsl
diff --git a/gfx/wr/swgl/src/vector_type.h b/gfx/wr/swgl/src/vector_type.h
new file mode 100644
index 0000000000..43364ffcce
--- /dev/null
+++ b/gfx/wr/swgl/src/vector_type.h
@@ -0,0 +1,563 @@
+/* This Source Code Form is subject to the terms of the Mozilla Public
+ * License, v. 2.0. If a copy of the MPL was not distributed with this
+ * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
+
+#ifdef __clang__
+#  ifdef __SSE2__
+#    include <xmmintrin.h>
+#    define USE_SSE2 1
+#  endif
+#  ifdef __ARM_NEON
+#    include <arm_neon.h>
+#    define USE_NEON 1
+#  endif
+#endif
+
+namespace glsl {
+
+#ifdef __clang__
+template <typename T, int N>
+using VectorType = T __attribute__((ext_vector_type(N)));
+
+#  define CONVERT(vector, type) __builtin_convertvector(vector, type)
+#  define SHUFFLE(a, b, ...) __builtin_shufflevector(a, b, __VA_ARGS__)
+
+template <typename T>
+SI VectorType<T, 4> combine(VectorType<T, 2> a, VectorType<T, 2> b) {
+  return __builtin_shufflevector(a, b, 0, 1, 2, 3);
+}
+
+template <typename T>
+SI VectorType<T, 8> combine(VectorType<T, 4> a, VectorType<T, 4> b) {
+  return __builtin_shufflevector(a, b, 0, 1, 2, 3, 4, 5, 6, 7);
+}
+
+template <typename T>
+SI VectorType<T, 16> combine(VectorType<T, 8> a, VectorType<T, 8> b) {
+  return __builtin_shufflevector(a, b, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
+                                 13, 14, 15);
+}
+
+template <typename T>
+SI VectorType<T, 2> lowHalf(VectorType<T, 4> a) {
+  return __builtin_shufflevector(a, a, 0, 1);
+}
+
+template <typename T>
+SI VectorType<T, 2> highHalf(VectorType<T, 4> a) {
+  return __builtin_shufflevector(a, a, 2, 3);
+}
+
+template <typename T>
+SI VectorType<T, 4> lowHalf(VectorType<T, 8> a) {
+  return __builtin_shufflevector(a, a, 0, 1, 2, 3);
+}
+
+template <typename T>
+SI VectorType<T, 4> highHalf(VectorType<T, 8> a) {
+  return __builtin_shufflevector(a, a, 4, 5, 6, 7);
+}
+
+template <typename T>
+SI VectorType<T, 8> lowHalf(VectorType<T, 16> a) {
+  return __builtin_shufflevector(a, a, 0, 1, 2, 3, 4, 5, 6, 7);
+}
+
+template <typename T>
+SI VectorType<T, 8> highHalf(VectorType<T, 16> a) {
+  return __builtin_shufflevector(a, a, 8, 9, 10, 11, 12, 13, 14, 15);
+}
+
+template <typename T>
+SI VectorType<T, 8> expand(VectorType<T, 4> a) {
+  return __builtin_shufflevector(a, a, 0, 1, 2, 3, -1, -1, -1, -1);
+}
+#else
+template <typename T>
+struct VectorMask {
+  typedef T type;
+};
+template <>
+struct VectorMask<uint32_t> {
+  typedef int32_t type;
+};
+template <>
+struct VectorMask<uint16_t> {
+  typedef int16_t type;
+};
+template <>
+struct VectorMask<uint8_t> {
+  typedef int8_t type;
+};
+template <>
+struct VectorMask<float> {
+  typedef int type;
+};
+
+template <typename T, int N>
+struct VectorType {
+  enum { SIZE = N };
+
+  typedef T data_type __attribute__((vector_size(sizeof(T) * N)));
+  typedef typename VectorMask<T>::type mask_index;
+  typedef mask_index mask_type
+      __attribute__((vector_size(sizeof(mask_index) * N)));
+  typedef T half_type __attribute__((vector_size(sizeof(T) * (N / 2))));
+  union {
+    data_type data;
+    struct {
+      T x, y, z, w;
+    };
+    T elements[N];
+    struct {
+      half_type low_half, high_half;
+    };
+  };
+
+  VectorType() : data{0} {}
+
+  constexpr VectorType(const VectorType& rhs) : data(rhs.data) {}
+  // GCC vector extensions only support broadcasting scalars on arithmetic ops,
+  // but not on initializers, hence the following...
+  constexpr VectorType(T n) : data((data_type){0} + n) {}
+  constexpr VectorType(T a, T b, T c, T d) : data{a, b, c, d} {}
+  constexpr VectorType(T a, T b, T c, T d, T e, T f, T g, T h)
+      : data{a, b, c, d, e, f, g, h} {}
+  constexpr VectorType(T a, T b, T c, T d, T e, T f, T g, T h, T i, T j, T k,
+                       T l, T m, T n, T o, T p)
+      : data{a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p} {}
+
+  SI VectorType wrap(const data_type& data) {
+    VectorType v;
+    v.data = data;
+    return v;
+  }
+
+  T& operator[](size_t i) { return elements[i]; }
+  T operator[](size_t i) const { return elements[i]; }
+
+  template <typename U>
+  operator VectorType<U, 2>() const {
+    return VectorType<U, 2>::wrap(
+        (typename VectorType<U, N>::data_type){U(x), U(y)});
+  }
+  template <typename U>
+  operator VectorType<U, 4>() const {
+    return VectorType<U, 4>::wrap(
+        (typename VectorType<U, N>::data_type){U(x), U(y), U(z), U(w)});
+  }
+  template <typename U>
+  operator VectorType<U, 8>() const {
+    return VectorType<U, 8>::wrap((typename VectorType<U, N>::data_type){
+        U(elements[0]), U(elements[1]), U(elements[2]), U(elements[3]),
+        U(elements[4]), U(elements[5]), U(elements[6]), U(elements[7])});
+  }
+  template <typename U>
+  operator VectorType<U, 16>() const {
+    return VectorType<U, 16>::wrap((typename VectorType<U, N>::data_type){
+        U(elements[0]),
+        U(elements[1]),
+        U(elements[2]),
+        U(elements[3]),
+        U(elements[4]),
+        U(elements[5]),
+        U(elements[6]),
+        U(elements[7]),
+        U(elements[8]),
+        U(elements[9]),
+        U(elements[10]),
+        U(elements[11]),
+        U(elements[12]),
+        U(elements[13]),
+        U(elements[14]),
+        U(elements[15]),
+    });
+  }
+
+  VectorType operator-() const { return wrap(-data); }
+  VectorType operator~() const { return wrap(~data); }
+
+  VectorType operator&(VectorType x) const { return wrap(data & x.data); }
+  VectorType operator&(T x) const { return wrap(data & x); }
+  VectorType operator|(VectorType x) const { return wrap(data | x.data); }
+  VectorType operator|(T x) const { return wrap(data | x); }
+  VectorType operator^(VectorType x) const { return wrap(data ^ x.data); }
+  VectorType operator^(T x) const { return wrap(data ^ x); }
+  VectorType operator<<(int x) const { return wrap(data << x); }
+  VectorType operator>>(int x) const { return wrap(data >> x); }
+  VectorType operator+(VectorType x) const { return wrap(data + x.data); }
+  VectorType operator+(T x) const { return wrap(data + x); }
+  friend VectorType operator+(T x, VectorType y) { return wrap(x + y.data); }
+  VectorType operator-(VectorType x) const { return wrap(data - x.data); }
+  VectorType operator-(T x) const { return wrap(data - x); }
+  friend VectorType operator-(T x, VectorType y) { return wrap(x - y.data); }
+  VectorType operator*(VectorType x) const { return wrap(data * x.data); }
+  VectorType operator*(T x) const { return wrap(data * x); }
+  friend VectorType operator*(T x, VectorType y) { return wrap(x * y.data); }
+  VectorType operator/(VectorType x) const { return wrap(data / x.data); }
+  VectorType operator/(T x) const { return wrap(data / x); }
+  friend VectorType operator/(T x, VectorType y) { return wrap(x / y.data); }
+  VectorType operator%(int x) const { return wrap(data % x); }
+
+  VectorType& operator&=(VectorType x) {
+    data &= x.data;
+    return *this;
+  }
+  VectorType& operator|=(VectorType x) {
+    data |= x.data;
+    return *this;
+  }
+  VectorType& operator^=(VectorType x) {
+    data ^= x.data;
+    return *this;
+  }
+  VectorType& operator<<=(int x) {
+    data <<= x;
+    return *this;
+  }
+  VectorType& operator>>=(int x) {
+    data >>= x;
+    return *this;
+  }
+  VectorType& operator+=(VectorType x) {
+    data += x.data;
+    return *this;
+  }
+  VectorType& operator-=(VectorType x) {
+    data -= x.data;
+    return *this;
+  }
+  VectorType& operator*=(VectorType x) {
+    data *= x.data;
+    return *this;
+  }
+  VectorType& operator/=(VectorType x) {
+    data /= x.data;
+    return *this;
+  }
+  VectorType& operator%=(int x) {
+    data %= x;
+    return *this;
+  }
+
+  VectorType<mask_type, N> operator==(VectorType x) const {
+    return VectorType<mask_type, N>::wrap(data == x.data);
+  }
+  VectorType<mask_type, N> operator!=(VectorType x) const {
+    return VectorType<mask_type, N>::wrap(data != x.data);
+  }
+  VectorType<mask_type, N> operator<(VectorType x) const {
+    return VectorType<mask_type, N>::wrap(data < x.data);
+  }
+  VectorType<mask_type, N> operator>(VectorType x) const {
+    return VectorType<mask_type, N>::wrap(data > x.data);
+  }
+  VectorType<mask_type, N> operator<=(VectorType x) const {
+    return VectorType<mask_type, N>::wrap(data <= x.data);
+  }
+  VectorType<mask_type, N> operator>=(VectorType x) const {
+    return VectorType<mask_type, N>::wrap(data >= x.data);
+  }
+
+  VectorType operator!() const { return wrap(!data); }
+  VectorType operator&&(VectorType x) const { return wrap(data & x.data); }
+  VectorType operator||(VectorType x) const { return wrap(data | x.data); }
+
+  VectorType& operator=(VectorType x) {
+    data = x.data;
+    return *this;
+  }
+
+  VectorType<T, 4> shuffle(VectorType b, mask_index x, mask_index y,
+                           mask_index z, mask_index w) const {
+    return VectorType<T, 4>::wrap(__builtin_shuffle(
+        data, b.data, (typename VectorType<T, 4>::mask_type){x, y, z, w}));
+  }
+  VectorType<T, 8> shuffle(VectorType b, mask_index x, mask_index y,
+                           mask_index z, mask_index w, mask_index s,
+                           mask_index t, mask_index u, mask_index v) const {
+    return VectorType<T, 8>::wrap(__builtin_shuffle(
+        data, b.data,
+        (typename VectorType<T, 8>::mask_type){x, y, z, w, s, t, u, v}));
+  }
+  VectorType<T, 16> shuffle(VectorType b, mask_index x, mask_index y,
+                            mask_index z, mask_index w, mask_index s,
+                            mask_index t, mask_index u, mask_index v,
+                            mask_index i, mask_index j, mask_index k,
+                            mask_index l, mask_index m, mask_index n,
+                            mask_index o, mask_index p) const {
+    return VectorType<T, 16>::wrap(
+        __builtin_shuffle(data, b.data,
+                          (typename VectorType<T, 16>::mask_type){
+                              x, y, z, w, s, t, u, v, i, j, k, l, m, n, o, p}));
+  }
+
+  VectorType<T, 4> swizzle(mask_index x, mask_index y, mask_index z,
+                           mask_index w) const {
+    return VectorType<T, 4>::wrap(__builtin_shuffle(
+        data, (typename VectorType<T, 4>::mask_type){x, y, z, w}));
+  }
+  VectorType<T, 8> swizzle(mask_index x, mask_index y, mask_index z,
+                           mask_index w, mask_index s, mask_index t,
+                           mask_index u, mask_index v) const {
+    return VectorType<T, 8>::wrap(__builtin_shuffle(
+        data, (typename VectorType<T, 8>::mask_type){x, y, z, w, s, t, u, v}));
+  }
+
+  SI VectorType wrap(half_type low, half_type high) {
+    VectorType v;
+    v.low_half = low;
+    v.high_half = high;
+    return v;
+  }
+
+  VectorType<T, N * 2> combine(VectorType high) const {
+    return VectorType<T, N * 2>::wrap(data, high.data);
+  }
+
+#  define xxxx swizzle(0, 0, 0, 0)
+#  define yyyy swizzle(1, 1, 1, 1)
+#  define zzzz swizzle(2, 2, 2, 2)
+#  define wwww swizzle(3, 3, 3, 3)
+#  define xxyy swizzle(0, 0, 1, 1)
+#  define xxzz swizzle(0, 0, 2, 2)
+#  define yyww swizzle(1, 1, 3, 3)
+#  define zzww swizzle(2, 2, 3, 3)
+#  define xyxy swizzle(0, 1, 0, 1)
+#  define xzxz swizzle(0, 2, 0, 2)
+#  define ywyw swizzle(1, 3, 1, 3)
+#  define zwzw swizzle(2, 3, 2, 3)
+#  define zwxy swizzle(2, 3, 0, 1)
+#  define zyxw swizzle(2, 1, 0, 3)
+#  define xxyz swizzle(0, 0, 1, 2)
+#  define xyyz swizzle(0, 1, 1, 2)
+#  define xyzz swizzle(0, 1, 2, 2)
+#  define xzyw swizzle(0, 2, 1, 3)
+#  define yzwx swizzle(1, 2, 3, 0)
+#  define wxyz swizzle(3, 0, 1, 2)
+#  define wzyx swizzle(3, 2, 1, 0)
+#  define xxxxyyyy XXXXYYYY()
+  VectorType<T, 8> XXXXYYYY() const {
+    return swizzle(0, 0, 0, 0).combine(swizzle(1, 1, 1, 1));
+  }
+#  define zzzzwwww ZZZZWWWW()
+  VectorType<T, 8> ZZZZWWWW() const {
+    return swizzle(2, 2, 2, 2).combine(swizzle(3, 3, 3, 3));
+  }
+#  define xyzwxyzw XYZWXYZW()
+  VectorType<T, 8> XYZWXYZW() const { return combine(*this); }
+#  define xyxyxyxy XYXYXYXY()
+  VectorType<T, 8> XYXYXYXY() const {
+    return swizzle(0, 1, 0, 1).combine(swizzle(0, 1, 0, 1));
+  }
+#  define zwzwzwzw ZWZWZWZW()
+  VectorType<T, 8> ZWZWZWZW() const {
+    return swizzle(2, 3, 2, 3).combine(swizzle(2, 3, 2, 3));
+  }
+#  define xxyyzzww XXYYZZWW()
+  VectorType<T, 8> XXYYZZWW() const {
+    return swizzle(0, 0, 1, 1).combine(swizzle(2, 2, 3, 3));
+  }
+#  define xxxxyyyyzzzzwwww XXXXYYYYZZZZWWWW()
+  VectorType<T, 16> XXXXYYYYZZZZWWWW() {
+    return XXXXYYYY().combine(ZZZZWWWW());
+  }
+};
+
+template <typename T>
+struct VectorType<T, 2> {
+  typedef T data_type __attribute__((vector_size(sizeof(T) * 2)));
+  union {
+    data_type data;
+    struct {
+      T x, y;
+    };
+    T elements[2];
+  };
+
+  SI VectorType wrap(const data_type& data) {
+    VectorType v;
+    v.data = data;
+    return v;
+  }
+
+  VectorType operator&(VectorType x) const { return wrap(data & x.data); }
+  VectorType operator&(T x) const { return wrap(data & x); }
+  VectorType operator|(VectorType x) const { return wrap(data | x.data); }
+  VectorType operator|(T x) const { return wrap(data | x); }
+};
+
+#  define CONVERT(vector, type) ((type)(vector))
+#  define SHUFFLE(a, b, ...) a.shuffle(b, __VA_ARGS__)
+
+template <typename T, int N>
+SI VectorType<T, N * 2> combine(VectorType<T, N> a, VectorType<T, N> b) {
+  return VectorType<T, N * 2>::wrap(a.data, b.data);
+}
+
+template <typename T, int N>
+SI VectorType<T, N / 2> lowHalf(VectorType<T, N> a) {
+  return VectorType<T, N / 2>::wrap(a.low_half);
+}
+
+template <typename T, int N>
+SI VectorType<T, N / 2> highHalf(VectorType<T, N> a) {
+  return VectorType<T, N / 2>::wrap(a.high_half);
+}
+
+template <typename T, int N>
+SI VectorType<T, N * 2> expand(VectorType<T, N> a) {
+  return combine(a, a);
+}
+#endif
+
+template <typename T, int N>
+SI VectorType<T, N * 4> combine(VectorType<T, N> a, VectorType<T, N> b,
+                                VectorType<T, N> c, VectorType<T, N> d) {
+  return combine(combine(a, b), combine(c, d));
+}
+
+template <typename T, int N>
+SI VectorType<T, N> combineLow(VectorType<T, N> a, VectorType<T, N> b) {
+  return combine(lowHalf(a), lowHalf(b));
+}
+
+template <typename T, int N>
+SI VectorType<T, N> combineHigh(VectorType<T, N> a, VectorType<T, N> b) {
+  return combine(highHalf(a), highHalf(b));
+}
+
+template <typename T, int N>
+SI VectorType<T, N * 2> repeat2(VectorType<T, N> a) {
+  return combine(a, a);
+}
+
+template <typename T, int N>
+SI VectorType<T, N * 4> repeat4(VectorType<T, N> a) {
+  return combine(a, a, a, a);
+}
+
+template <typename T>
+SI VectorType<T, 4> zipLow(VectorType<T, 4> a, VectorType<T, 4> b) {
+  return SHUFFLE(a, b, 0, 4, 1, 5);
+}
+
+template <typename T>
+SI VectorType<T, 4> zipHigh(VectorType<T, 4> a, VectorType<T, 4> b) {
+  return SHUFFLE(a, b, 2, 6, 3, 7);
+}
+
+template <typename T>
+SI VectorType<T, 8> zipLow(VectorType<T, 8> a, VectorType<T, 8> b) {
+  return SHUFFLE(a, b, 0, 8, 1, 9, 2, 10, 3, 11);
+}
+
+template <typename T>
+SI VectorType<T, 8> zipHigh(VectorType<T, 8> a, VectorType<T, 8> b) {
+  return SHUFFLE(a, b, 4, 12, 5, 13, 6, 14, 7, 15);
+}
+
+template <typename T>
+SI VectorType<T, 16> zipLow(VectorType<T, 16> a, VectorType<T, 16> b) {
+  return SHUFFLE(a, b, 0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23);
+}
+
+template <typename T>
+SI VectorType<T, 16> zipHigh(VectorType<T, 16> a, VectorType<T, 16> b) {
+  return SHUFFLE(a, b, 8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30,
+                 31);
+}
+
+template <typename T>
+SI VectorType<T, 8> zip2Low(VectorType<T, 8> a, VectorType<T, 8> b) {
+  return SHUFFLE(a, b, 0, 1, 8, 9, 2, 3, 10, 11);
+}
+
+template <typename T>
+SI VectorType<T, 8> zip2High(VectorType<T, 8> a, VectorType<T, 8> b) {
+  return SHUFFLE(a, b, 4, 5, 12, 13, 6, 7, 14, 15);
+}
+
+#ifdef __clang__
+template <typename T>
+SI VectorType<T, 8> zip(VectorType<T, 4> a, VectorType<T, 4> b) {
+  return SHUFFLE(a, b, 0, 4, 1, 5, 2, 6, 3, 7);
+}
+
+template <typename T>
+SI VectorType<T, 16> zip(VectorType<T, 8> a, VectorType<T, 8> b) {
+  return SHUFFLE(a, b, 0, 8, 1, 9, 2, 10, 3, 11, 4, 12, 5, 13, 6, 14, 7, 15);
+}
+#else
+template <typename T, int N>
+SI VectorType<T, N * 2> zip(VectorType<T, N> a, VectorType<T, N> b) {
+  return combine(zipLow(a, b), zipHigh(a, b));
+}
+#endif
+
+template <typename T>
+struct Unaligned {
+  template <typename P>
+  SI T load(const P* p) {
+    T v;
+    memcpy(&v, p, sizeof(v));
+    return v;
+  }
+
+  template <typename P>
+  SI void store(P* p, T v) {
+    memcpy(p, &v, sizeof(v));
+  }
+};
+
+#ifndef __clang__
+template <typename T, int N>
+struct Unaligned<VectorType<T, N>> {
+  template <typename P>
+  SI VectorType<T, N> load(const P* p) {
+    VectorType<T, N> v;
+    memcpy(v.elements, p, sizeof(v));
+    return v;
+  }
+
+  template <typename P>
+  SI void store(P* p, VectorType<T, N> v) {
+    memcpy(p, v.elements, sizeof(v));
+  }
+};
+#endif
+
+template <typename T, typename P>
+SI T unaligned_load(const P* p) {
+  return Unaligned<T>::load(p);
+}
+
+template <typename T, typename P>
+SI void unaligned_store(P* p, T v) {
+  Unaligned<T>::store(p, v);
+}
+
+template <typename D, typename S>
+SI D bit_cast(const S& src) {
+  static_assert(sizeof(D) == sizeof(S), "");
+  return unaligned_load<D>(&src);
+}
+
+template <typename T>
+using V2 = VectorType<T, 2>;
+template <typename T>
+using V4 = VectorType<T, 4>;
+using Float = V4<float>;
+using I32 = V4<int32_t>;
+using I16 = V4<int16_t>;
+using U64 = V4<uint64_t>;
+using U32 = V4<uint32_t>;
+using U16 = V4<uint16_t>;
+using U8 = V4<uint8_t>;
+using Bool = V4<int>;
+template <typename T>
+using V8 = VectorType<T, 8>;
+template <typename T>
+using V16 = VectorType<T, 16>;
+
+}  // namespace glsl