1 files changed, 261 insertions, 0 deletions
diff --git a/third_party/jpeg-xl/lib/jxl/convolve_separable5.cc b/third_party/jpeg-xl/lib/jxl/convolve_separable5.cc
new file mode 100644
index 0000000000..b26ff54bbc
--- /dev/null
+++ b/third_party/jpeg-xl/lib/jxl/convolve_separable5.cc
@@ -0,0 +1,261 @@
+// Copyright (c) the JPEG XL Project Authors. All rights reserved.
+//
+// Use of this source code is governed by a BSD-style
+// license that can be found in the LICENSE file.
+
+#include "lib/jxl/convolve.h"
+
+#undef HWY_TARGET_INCLUDE
+#define HWY_TARGET_INCLUDE "lib/jxl/convolve_separable5.cc"
+#include <hwy/foreach_target.h>
+#include <hwy/highway.h>
+
+#include "lib/jxl/convolve-inl.h"
+
+HWY_BEFORE_NAMESPACE();
+namespace jxl {
+namespace HWY_NAMESPACE {
+
+// These templates are not found via ADL.
+using hwy::HWY_NAMESPACE::Add;
+using hwy::HWY_NAMESPACE::Mul;
+using hwy::HWY_NAMESPACE::MulAdd;
+using hwy::HWY_NAMESPACE::Vec;
+
+// 5x5 convolution by separable kernel with a single scan through the input.
+// This is more cache-efficient than separate horizontal/vertical passes, and
+// possibly faster (given enough registers) than tiling and/or transposing.
+//
+// Overview: imagine a 5x5 window around a central pixel. First convolve the
+// rows by multiplying the pixels with the corresponding weights from
+// WeightsSeparable5.horz[abs(x_offset) * 4]. Then multiply each of these
+// intermediate results by the corresponding vertical weight, i.e.
+// vert[abs(y_offset) * 4]. Finally, store the sum of these values as the
+// convolution result at the position of the central pixel in the output.
+//
+// Each of these operations uses SIMD vectors. The central pixel and most
+// importantly the output are aligned, so neighnoring pixels (e.g. x_offset=1)
+// require unaligned loads. Because weights are supplied in identical groups of
+// 4, we can use LoadDup128 to load them (slightly faster).
+//
+// Uses mirrored boundary handling. Until x >= kRadius, the horizontal
+// convolution uses Neighbors class to shuffle vectors as if each of its lanes
+// had been loaded from the mirrored offset. Similarly, the last full vector to
+// write uses mirroring. In the case of scalar vectors, Neighbors is not usable
+// and the value is loaded directly. Otherwise, the number of valid pixels
+// modulo the vector size enables a small optimization: for smaller offsets,
+// a non-mirrored load is sufficient.
+class Separable5Strategy {
+  using D = HWY_CAPPED(float, 16);
+  using V = Vec<D>;
+
+ public:
+  static constexpr int64_t kRadius = 2;
+
+  template <size_t kSizeModN, class WrapRow>
+  static JXL_MAYBE_INLINE void ConvolveRow(
+      const float* const JXL_RESTRICT row_m, const size_t xsize,
+      const int64_t stride, const WrapRow& wrap_row,
+      const WeightsSeparable5& weights, float* const JXL_RESTRICT row_out) {
+    const D d;
+    const int64_t neg_stride = -stride;  // allows LEA addressing.
+    const float* const JXL_RESTRICT row_t2 =
+        wrap_row(row_m + 2 * neg_stride, stride);
+    const float* const JXL_RESTRICT row_t1 =
+        wrap_row(row_m + 1 * neg_stride, stride);
+    const float* const JXL_RESTRICT row_b1 =
+        wrap_row(row_m + 1 * stride, stride);
+    const float* const JXL_RESTRICT row_b2 =
+        wrap_row(row_m + 2 * stride, stride);
+
+    const V wh0 = LoadDup128(d, weights.horz + 0 * 4);
+    const V wh1 = LoadDup128(d, weights.horz + 1 * 4);
+    const V wh2 = LoadDup128(d, weights.horz + 2 * 4);
+    const V wv0 = LoadDup128(d, weights.vert + 0 * 4);
+    const V wv1 = LoadDup128(d, weights.vert + 1 * 4);
+    const V wv2 = LoadDup128(d, weights.vert + 2 * 4);
+
+    size_t x = 0;
+
+    // More than one iteration for scalars.
+    for (; x < kRadius; x += Lanes(d)) {
+      const V conv0 =
+          Mul(HorzConvolveFirst(row_m, x, xsize, wh0, wh1, wh2), wv0);
+
+      const V conv1t = HorzConvolveFirst(row_t1, x, xsize, wh0, wh1, wh2);
+      const V conv1b = HorzConvolveFirst(row_b1, x, xsize, wh0, wh1, wh2);
+      const V conv1 = MulAdd(Add(conv1t, conv1b), wv1, conv0);
+
+      const V conv2t = HorzConvolveFirst(row_t2, x, xsize, wh0, wh1, wh2);
+      const V conv2b = HorzConvolveFirst(row_b2, x, xsize, wh0, wh1, wh2);
+      const V conv2 = MulAdd(Add(conv2t, conv2b), wv2, conv1);
+      Store(conv2, d, row_out + x);
+    }
+
+    // Main loop: load inputs without padding
+    for (; x + Lanes(d) + kRadius <= xsize; x += Lanes(d)) {
+      const V conv0 = Mul(HorzConvolve(row_m + x, wh0, wh1, wh2), wv0);
+
+      const V conv1t = HorzConvolve(row_t1 + x, wh0, wh1, wh2);
+      const V conv1b = HorzConvolve(row_b1 + x, wh0, wh1, wh2);
+      const V conv1 = MulAdd(Add(conv1t, conv1b), wv1, conv0);
+
+      const V conv2t = HorzConvolve(row_t2 + x, wh0, wh1, wh2);
+      const V conv2b = HorzConvolve(row_b2 + x, wh0, wh1, wh2);
+      const V conv2 = MulAdd(Add(conv2t, conv2b), wv2, conv1);
+      Store(conv2, d, row_out + x);
+    }
+
+    // Last full vector to write (the above loop handled mod >= kRadius)
+#if HWY_TARGET == HWY_SCALAR
+    while (x < xsize) {
+#else
+    if (kSizeModN < kRadius) {
+#endif
+      const V conv0 =
+          Mul(HorzConvolveLast<kSizeModN>(row_m, x, xsize, wh0, wh1, wh2), wv0);
+
+      const V conv1t =
+          HorzConvolveLast<kSizeModN>(row_t1, x, xsize, wh0, wh1, wh2);
+      const V conv1b =
+          HorzConvolveLast<kSizeModN>(row_b1, x, xsize, wh0, wh1, wh2);
+      const V conv1 = MulAdd(Add(conv1t, conv1b), wv1, conv0);
+
+      const V conv2t =
+          HorzConvolveLast<kSizeModN>(row_t2, x, xsize, wh0, wh1, wh2);
+      const V conv2b =
+          HorzConvolveLast<kSizeModN>(row_b2, x, xsize, wh0, wh1, wh2);
+      const V conv2 = MulAdd(Add(conv2t, conv2b), wv2, conv1);
+      Store(conv2, d, row_out + x);
+      x += Lanes(d);
+    }
+
+    // If mod = 0, the above vector was the last.
+    if (kSizeModN != 0) {
+      for (; x < xsize; ++x) {
+        float mul = 0.0f;
+        for (int64_t dy = -kRadius; dy <= kRadius; ++dy) {
+          const float wy = weights.vert[std::abs(dy) * 4];
+          const float* clamped_row = wrap_row(row_m + dy * stride, stride);
+          for (int64_t dx = -kRadius; dx <= kRadius; ++dx) {
+            const float wx = weights.horz[std::abs(dx) * 4];
+            const int64_t clamped_x = Mirror(x + dx, xsize);
+            mul += clamped_row[clamped_x] * wx * wy;
+          }
+        }
+        row_out[x] = mul;
+      }
+    }
+  }
+
+ private:
+  // Same as HorzConvolve for the first/last vector in a row.
+  static JXL_MAYBE_INLINE V HorzConvolveFirst(
+      const float* const JXL_RESTRICT row, const int64_t x, const int64_t xsize,
+      const V wh0, const V wh1, const V wh2) {
+    const D d;
+    const V c = LoadU(d, row + x);
+    const V mul0 = Mul(c, wh0);
+
+#if HWY_TARGET == HWY_SCALAR
+    const V l1 = LoadU(d, row + Mirror(x - 1, xsize));
+    const V l2 = LoadU(d, row + Mirror(x - 2, xsize));
+#else
+    (void)xsize;
+    const V l1 = Neighbors::FirstL1(c);
+    const V l2 = Neighbors::FirstL2(c);
+#endif
+
+    const V r1 = LoadU(d, row + x + 1);
+    const V r2 = LoadU(d, row + x + 2);
+
+    const V mul1 = MulAdd(Add(l1, r1), wh1, mul0);
+    const V mul2 = MulAdd(Add(l2, r2), wh2, mul1);
+    return mul2;
+  }
+
+  template <size_t kSizeModN>
+  static JXL_MAYBE_INLINE V
+  HorzConvolveLast(const float* const JXL_RESTRICT row, const int64_t x,
+                   const int64_t xsize, const V wh0, const V wh1, const V wh2) {
+    const D d;
+    const V c = LoadU(d, row + x);
+    const V mul0 = Mul(c, wh0);
+
+    const V l1 = LoadU(d, row + x - 1);
+    const V l2 = LoadU(d, row + x - 2);
+
+    V r1, r2;
+#if HWY_TARGET == HWY_SCALAR
+    r1 = LoadU(d, row + Mirror(x + 1, xsize));
+    r2 = LoadU(d, row + Mirror(x + 2, xsize));
+#else
+    const size_t N = Lanes(d);
+    if (kSizeModN == 0) {
+      r2 = TableLookupLanes(c, SetTableIndices(d, MirrorLanes(N - 2)));
+      r1 = TableLookupLanes(c, SetTableIndices(d, MirrorLanes(N - 1)));
+    } else {  // == 1
+      const auto last = LoadU(d, row + xsize - N);
+      r2 = TableLookupLanes(last, SetTableIndices(d, MirrorLanes(N - 1)));
+      r1 = last;
+    }
+#endif
+
+    // Sum of pixels with Manhattan distance i, multiplied by weights[i].
+    const V sum1 = Add(l1, r1);
+    const V mul1 = MulAdd(sum1, wh1, mul0);
+    const V sum2 = Add(l2, r2);
+    const V mul2 = MulAdd(sum2, wh2, mul1);
+    return mul2;
+  }
+
+  // Requires kRadius valid pixels before/after pos.
+  static JXL_MAYBE_INLINE V HorzConvolve(const float* const JXL_RESTRICT pos,
+                                         const V wh0, const V wh1,
+                                         const V wh2) {
+    const D d;
+    const V c = LoadU(d, pos);
+    const V mul0 = Mul(c, wh0);
+
+    // Loading anew is faster than combining vectors.
+    const V l1 = LoadU(d, pos - 1);
+    const V r1 = LoadU(d, pos + 1);
+    const V l2 = LoadU(d, pos - 2);
+    const V r2 = LoadU(d, pos + 2);
+    // Sum of pixels with Manhattan distance i, multiplied by weights[i].
+    const V sum1 = Add(l1, r1);
+    const V mul1 = MulAdd(sum1, wh1, mul0);
+    const V sum2 = Add(l2, r2);
+    const V mul2 = MulAdd(sum2, wh2, mul1);
+    return mul2;
+  }
+};
+
+void Separable5(const ImageF& in, const Rect& rect,
+                const WeightsSeparable5& weights, ThreadPool* pool,
+                ImageF* out) {
+  using Conv = ConvolveT<Separable5Strategy>;
+  if (rect.xsize() >= Conv::MinWidth()) {
+    return Conv::Run(in, rect, weights, pool, out);
+  }
+
+  return SlowSeparable5(in, rect, weights, pool, out);
+}
+
+// NOLINTNEXTLINE(google-readability-namespace-comments)
+}  // namespace HWY_NAMESPACE
+}  // namespace jxl
+HWY_AFTER_NAMESPACE();
+
+#if HWY_ONCE
+namespace jxl {
+
+HWY_EXPORT(Separable5);
+void Separable5(const ImageF& in, const Rect& rect,
+                const WeightsSeparable5& weights, ThreadPool* pool,
+                ImageF* out) {
+  return HWY_DYNAMIC_DISPATCH(Separable5)(in, rect, weights, pool, out);
+}
+
+}  // namespace jxl
+#endif  // HWY_ONCE