summaryrefslogtreecommitdiffstats
path: root/media/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c
diff options
context:
space:
mode:
Diffstat (limited to 'media/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c')
-rw-r--r--media/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c965
1 files changed, 965 insertions, 0 deletions
diff --git a/media/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c b/media/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c
new file mode 100644
index 0000000000..65fb67c984
--- /dev/null
+++ b/media/libvpx/libvpx/vpx_dsp/arm/vpx_convolve8_neon.c
@@ -0,0 +1,965 @@
+/*
+ * Copyright (c) 2014 The WebM 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 in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include <assert.h>
+
+#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx/vpx_integer.h"
+#include "vpx_dsp/arm/mem_neon.h"
+#include "vpx_dsp/arm/transpose_neon.h"
+#include "vpx_dsp/arm/vpx_convolve8_neon.h"
+#include "vpx_dsp/vpx_filter.h"
+#include "vpx_ports/mem.h"
+
+// Note:
+// 1. src is not always 32-bit aligned, so don't call vld1_lane_u32(src).
+// 2. After refactoring the shared code in kernel loops with inline functions,
+// the decoder speed dropped a lot when using gcc compiler. Therefore there is
+// no refactoring for those parts by now.
+// 3. For horizontal convolve, there is an alternative optimization that
+// convolves a single row in each loop. For each row, 8 sample banks with 4 or 8
+// samples in each are read from memory: src, (src+1), (src+2), (src+3),
+// (src+4), (src+5), (src+6), (src+7), or prepared by vector extract
+// instructions. This optimization is much faster in speed unit test, but slowed
+// down the whole decoder by 5%.
+
+static INLINE void vpx_convolve_4tap_horiz_neon(const uint8_t *src,
+ ptrdiff_t src_stride,
+ uint8_t *dst,
+ ptrdiff_t dst_stride, int w,
+ int h, const int16x4_t filter) {
+ if (w == 4) {
+ do {
+ int16x4_t s0[4], s1[4];
+
+ int16x8_t t0 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(src)));
+ s0[0] = vget_low_s16(vextq_s16(t0, t0, 0));
+ s0[1] = vget_low_s16(vextq_s16(t0, t0, 1));
+ s0[2] = vget_low_s16(vextq_s16(t0, t0, 2));
+ s0[3] = vget_low_s16(vextq_s16(t0, t0, 3));
+
+ int16x8_t t1 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(src + src_stride)));
+ s1[0] = vget_low_s16(vextq_s16(t1, t1, 0));
+ s1[1] = vget_low_s16(vextq_s16(t1, t1, 1));
+ s1[2] = vget_low_s16(vextq_s16(t1, t1, 2));
+ s1[3] = vget_low_s16(vextq_s16(t1, t1, 3));
+
+ int16x4_t d0 = convolve4_4(s0[0], s0[1], s0[2], s0[3], filter);
+ int16x4_t d1 = convolve4_4(s1[0], s1[1], s1[2], s1[3], filter);
+ uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
+
+ store_u8(dst, dst_stride, d01);
+
+ src += 2 * src_stride;
+ dst += 2 * dst_stride;
+ h -= 2;
+ } while (h > 0);
+ } else {
+ do {
+ const uint8_t *s = src;
+ uint8_t *d = dst;
+ int width = w;
+
+ do {
+ int16x8_t t0[2], t1[2];
+ int16x8_t s0[4], s1[4];
+
+ t0[0] = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s)));
+ t0[1] = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + 8)));
+ s0[0] = vextq_s16(t0[0], t0[1], 0);
+ s0[1] = vextq_s16(t0[0], t0[1], 1);
+ s0[2] = vextq_s16(t0[0], t0[1], 2);
+ s0[3] = vextq_s16(t0[0], t0[1], 3);
+
+ t1[0] = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + src_stride)));
+ t1[1] = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s + src_stride + 8)));
+ s1[0] = vextq_s16(t1[0], t1[1], 0);
+ s1[1] = vextq_s16(t1[0], t1[1], 1);
+ s1[2] = vextq_s16(t1[0], t1[1], 2);
+ s1[3] = vextq_s16(t1[0], t1[1], 3);
+
+ uint8x8_t d0 = convolve4_8(s0[0], s0[1], s0[2], s0[3], filter);
+ uint8x8_t d1 = convolve4_8(s1[0], s1[1], s1[2], s1[3], filter);
+
+ vst1_u8(d, d0);
+ vst1_u8(d + dst_stride, d1);
+ s += 8;
+ d += 8;
+ width -= 8;
+ } while (width != 0);
+ src += 2 * src_stride;
+ dst += 2 * dst_stride;
+ h -= 2;
+ } while (h > 0);
+ }
+}
+
+static INLINE void vpx_convolve_8tap_horiz_neon(const uint8_t *src,
+ ptrdiff_t src_stride,
+ uint8_t *dst,
+ ptrdiff_t dst_stride, int w,
+ int h, const int16x8_t filter) {
+ uint8x8_t t0, t1, t2, t3;
+
+ if (h == 4) {
+ uint8x8_t d01, d23;
+ int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
+
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ transpose_u8_8x4(&t0, &t1, &t2, &t3);
+ s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+ s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ src += 7;
+
+ do {
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ transpose_u8_8x4(&t0, &t1, &t2, &t3);
+ s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+
+ d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+ d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
+ d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
+
+ transpose_u8_4x4(&d01, &d23);
+
+ store_u8(dst + 0 * dst_stride, 2 * dst_stride, d01);
+ store_u8(dst + 1 * dst_stride, 2 * dst_stride, d23);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ src += 4;
+ dst += 4;
+ w -= 4;
+ } while (w != 0);
+ } else {
+ int width;
+ const uint8_t *s;
+ uint8x8_t t4, t5, t6, t7, d04, d15, d26, d37;
+ int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
+
+ if (w == 4) {
+ do {
+ load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ load_u8_8x8(src + 7, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6,
+ &t7);
+ src += 8 * src_stride;
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ __builtin_prefetch(dst + 4 * dst_stride);
+ __builtin_prefetch(dst + 5 * dst_stride);
+ __builtin_prefetch(dst + 6 * dst_stride);
+ __builtin_prefetch(dst + 7 * dst_stride);
+ transpose_u8_4x8(&t0, &t1, &t2, &t3, t4, t5, t6, t7);
+ s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+ __builtin_prefetch(src + 4 * src_stride);
+ __builtin_prefetch(src + 5 * src_stride);
+ __builtin_prefetch(src + 6 * src_stride);
+ __builtin_prefetch(src + 7 * src_stride);
+ d04 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ d15 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ d26 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ d37 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+
+ transpose_u8_8x4(&d04, &d15, &d26, &d37);
+
+ store_u8(dst + 0 * dst_stride, 4 * dst_stride, d04);
+ store_u8(dst + 1 * dst_stride, 4 * dst_stride, d15);
+ store_u8(dst + 2 * dst_stride, 4 * dst_stride, d26);
+ store_u8(dst + 3 * dst_stride, 4 * dst_stride, d37);
+
+ dst += 8 * dst_stride;
+ h -= 8;
+ } while (h > 0);
+ } else {
+ uint8_t *d;
+ uint8x8_t d0, d1, d2, d3, d4, d5, d6, d7;
+ int16x8_t s11, s12, s13, s14;
+
+ do {
+ load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ width = w;
+ s = src + 7;
+ d = dst;
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ __builtin_prefetch(dst + 4 * dst_stride);
+ __builtin_prefetch(dst + 5 * dst_stride);
+ __builtin_prefetch(dst + 6 * dst_stride);
+ __builtin_prefetch(dst + 7 * dst_stride);
+
+ do {
+ load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s11 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s12 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s13 = vreinterpretq_s16_u16(vmovl_u8(t6));
+ s14 = vreinterpretq_s16_u16(vmovl_u8(t7));
+
+ d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+ d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filter);
+ d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filter);
+ d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filter);
+ d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filter);
+
+ transpose_u8_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
+
+ store_u8_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
+
+ s0 = s8;
+ s1 = s9;
+ s2 = s10;
+ s3 = s11;
+ s4 = s12;
+ s5 = s13;
+ s6 = s14;
+ s += 8;
+ d += 8;
+ width -= 8;
+ } while (width != 0);
+ src += 8 * src_stride;
+ dst += 8 * dst_stride;
+ h -= 8;
+ } while (h > 0);
+ }
+ }
+}
+
+void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *filter, int x0_q4,
+ int x_step_q4, int y0_q4, int y_step_q4, int w,
+ int h) {
+ assert((intptr_t)dst % 4 == 0);
+ assert(dst_stride % 4 == 0);
+ assert(x_step_q4 == 16);
+
+ (void)x_step_q4;
+ (void)y0_q4;
+ (void)y_step_q4;
+
+ if (vpx_get_filter_taps(filter[x0_q4]) <= 4) {
+ /* All 4-tap and bilinear filter values are even, so halve them to reduce
+ * intermediate precision requirements.
+ */
+ const int16x4_t x_filter_4tap = vshr_n_s16(vld1_s16(filter[x0_q4] + 2), 1);
+ vpx_convolve_4tap_horiz_neon(src - 1, src_stride, dst, dst_stride, w, h,
+ x_filter_4tap);
+ } else {
+ const int16x8_t x_filter_8tap = vld1q_s16(filter[x0_q4]);
+ vpx_convolve_8tap_horiz_neon(src - 3, src_stride, dst, dst_stride, w, h,
+ x_filter_8tap);
+ }
+}
+
+void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *filter, int x0_q4,
+ int x_step_q4, int y0_q4, int y_step_q4,
+ int w, int h) {
+ const int16x8_t filters = vld1q_s16(filter[x0_q4]);
+ uint8x8_t t0, t1, t2, t3;
+
+ assert((intptr_t)dst % 4 == 0);
+ assert(dst_stride % 4 == 0);
+ assert(x_step_q4 == 16);
+
+ (void)x_step_q4;
+ (void)y0_q4;
+ (void)y_step_q4;
+
+ src -= 3;
+
+ if (h == 4) {
+ uint8x8_t d01, d23, dd01, dd23;
+ int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
+
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ transpose_u8_8x4(&t0, &t1, &t2, &t3);
+ s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+ s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ src += 7;
+
+ do {
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ transpose_u8_8x4(&t0, &t1, &t2, &t3);
+ s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+
+ d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+ d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+ d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+ d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+ d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
+ d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
+
+ transpose_u8_4x4(&d01, &d23);
+
+ dd01 = load_u8(dst + 0 * dst_stride, 2 * dst_stride);
+ dd23 = load_u8(dst + 1 * dst_stride, 2 * dst_stride);
+
+ d01 = vrhadd_u8(d01, dd01);
+ d23 = vrhadd_u8(d23, dd23);
+
+ store_u8(dst + 0 * dst_stride, 2 * dst_stride, d01);
+ store_u8(dst + 1 * dst_stride, 2 * dst_stride, d23);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ src += 4;
+ dst += 4;
+ w -= 4;
+ } while (w != 0);
+ } else {
+ int width;
+ const uint8_t *s;
+ uint8x8_t t4, t5, t6, t7;
+ int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
+
+ if (w == 4) {
+ uint8x8_t d04, d15, d26, d37, dd04, dd15, dd26, dd37;
+
+ do {
+ load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ load_u8_8x8(src + 7, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6,
+ &t7);
+ src += 8 * src_stride;
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ __builtin_prefetch(dst + 4 * dst_stride);
+ __builtin_prefetch(dst + 5 * dst_stride);
+ __builtin_prefetch(dst + 6 * dst_stride);
+ __builtin_prefetch(dst + 7 * dst_stride);
+ transpose_u8_4x8(&t0, &t1, &t2, &t3, t4, t5, t6, t7);
+ s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+ __builtin_prefetch(src + 4 * src_stride);
+ __builtin_prefetch(src + 5 * src_stride);
+ __builtin_prefetch(src + 6 * src_stride);
+ __builtin_prefetch(src + 7 * src_stride);
+ d04 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+ d15 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+ d26 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+ d37 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+
+ transpose_u8_8x4(&d04, &d15, &d26, &d37);
+
+ dd04 = load_u8(dst + 0 * dst_stride, 4 * dst_stride);
+ dd15 = load_u8(dst + 1 * dst_stride, 4 * dst_stride);
+ dd26 = load_u8(dst + 2 * dst_stride, 4 * dst_stride);
+ dd37 = load_u8(dst + 3 * dst_stride, 4 * dst_stride);
+
+ d04 = vrhadd_u8(d04, dd04);
+ d15 = vrhadd_u8(d15, dd15);
+ d26 = vrhadd_u8(d26, dd26);
+ d37 = vrhadd_u8(d37, dd37);
+
+ store_u8(dst + 0 * dst_stride, 4 * dst_stride, d04);
+ store_u8(dst + 1 * dst_stride, 4 * dst_stride, d15);
+ store_u8(dst + 2 * dst_stride, 4 * dst_stride, d26);
+ store_u8(dst + 3 * dst_stride, 4 * dst_stride, d37);
+
+ dst += 8 * dst_stride;
+ h -= 8;
+ } while (h != 0);
+ } else {
+ uint8_t *d;
+ uint8x8_t d0, d1, d2, d3, d4, d5, d6, d7;
+ int16x8_t s11, s12, s13, s14;
+
+ do {
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+ __builtin_prefetch(src + 4 * src_stride);
+ __builtin_prefetch(src + 5 * src_stride);
+ __builtin_prefetch(src + 6 * src_stride);
+ __builtin_prefetch(src + 7 * src_stride);
+ load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ width = w;
+ s = src + 7;
+ d = dst;
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ __builtin_prefetch(dst + 4 * dst_stride);
+ __builtin_prefetch(dst + 5 * dst_stride);
+ __builtin_prefetch(dst + 6 * dst_stride);
+ __builtin_prefetch(dst + 7 * dst_stride);
+
+ do {
+ load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s11 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s12 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s13 = vreinterpretq_s16_u16(vmovl_u8(t6));
+ s14 = vreinterpretq_s16_u16(vmovl_u8(t7));
+
+ d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+ d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+ d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+ d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+ d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters);
+ d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters);
+ d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters);
+ d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters);
+
+ transpose_u8_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
+
+ d0 = vrhadd_u8(d0, vld1_u8(d + 0 * dst_stride));
+ d1 = vrhadd_u8(d1, vld1_u8(d + 1 * dst_stride));
+ d2 = vrhadd_u8(d2, vld1_u8(d + 2 * dst_stride));
+ d3 = vrhadd_u8(d3, vld1_u8(d + 3 * dst_stride));
+ d4 = vrhadd_u8(d4, vld1_u8(d + 4 * dst_stride));
+ d5 = vrhadd_u8(d5, vld1_u8(d + 5 * dst_stride));
+ d6 = vrhadd_u8(d6, vld1_u8(d + 6 * dst_stride));
+ d7 = vrhadd_u8(d7, vld1_u8(d + 7 * dst_stride));
+
+ store_u8_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
+
+ s0 = s8;
+ s1 = s9;
+ s2 = s10;
+ s3 = s11;
+ s4 = s12;
+ s5 = s13;
+ s6 = s14;
+ s += 8;
+ d += 8;
+ width -= 8;
+ } while (width != 0);
+ src += 8 * src_stride;
+ dst += 8 * dst_stride;
+ h -= 8;
+ } while (h != 0);
+ }
+ }
+}
+
+static INLINE void vpx_convolve_4tap_vert_neon(const uint8_t *src,
+ ptrdiff_t src_stride,
+ uint8_t *dst,
+ ptrdiff_t dst_stride, int w,
+ int h, const int16x4_t filter) {
+ if (w == 4) {
+ uint8x8_t t0, t1, t2, t3, d01, d23;
+ int16x4_t s0, s1, s2, s3, s4, s5, s6, d0, d1, d2, d3;
+
+ load_u8_8x3(src, src_stride, &t0, &t1, &t2);
+ s0 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t0)));
+ s1 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t1)));
+ s2 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t2)));
+
+ src += 3 * src_stride;
+
+ do {
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ s3 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t0)));
+ s4 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t1)));
+ s5 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t2)));
+ s6 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t3)));
+
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+
+ d0 = convolve4_4(s0, s1, s2, s3, filter);
+ d1 = convolve4_4(s1, s2, s3, s4, filter);
+ d2 = convolve4_4(s2, s3, s4, s5, filter);
+ d3 = convolve4_4(s3, s4, s5, s6, filter);
+ /* We halved the filter values so -1 from right shift. */
+ d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
+ d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS - 1);
+
+ store_u8(dst + 0 * dst_stride, dst_stride, d01);
+ store_u8(dst + 2 * dst_stride, dst_stride, d23);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ src += 4 * src_stride;
+ dst += 4 * dst_stride;
+ h -= 4;
+ } while (h != 0);
+ } else {
+ int height;
+ const uint8_t *s;
+ uint8_t *d;
+ uint8x8_t t0, t1, t2, t3, d0, d1, d2, d3;
+ int16x8_t s0, s1, s2, s3, s4, s5, s6;
+
+ do {
+ load_u8_8x3(src, src_stride, &t0, &t1, &t2);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+
+ s = src + 3 * src_stride;
+ d = dst;
+ height = h;
+
+ do {
+ load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t3));
+
+ __builtin_prefetch(d + 0 * dst_stride);
+ __builtin_prefetch(d + 1 * dst_stride);
+ __builtin_prefetch(d + 2 * dst_stride);
+ __builtin_prefetch(d + 3 * dst_stride);
+ __builtin_prefetch(s + 0 * src_stride);
+ __builtin_prefetch(s + 1 * src_stride);
+ __builtin_prefetch(s + 2 * src_stride);
+ __builtin_prefetch(s + 3 * src_stride);
+
+ d0 = convolve4_8(s0, s1, s2, s3, filter);
+ d1 = convolve4_8(s1, s2, s3, s4, filter);
+ d2 = convolve4_8(s2, s3, s4, s5, filter);
+ d3 = convolve4_8(s3, s4, s5, s6, filter);
+
+ store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s += 4 * src_stride;
+ d += 4 * dst_stride;
+ height -= 4;
+ } while (height != 0);
+ src += 8;
+ dst += 8;
+ w -= 8;
+ } while (w != 0);
+ }
+}
+
+static INLINE void vpx_convolve_8tap_vert_neon(const uint8_t *src,
+ ptrdiff_t src_stride,
+ uint8_t *dst,
+ ptrdiff_t dst_stride, int w,
+ int h, const int16x8_t filter) {
+ if (w == 4) {
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6, d01, d23;
+ int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
+
+ load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6);
+ s0 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t0)));
+ s1 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t1)));
+ s2 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t2)));
+ s3 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t3)));
+ s4 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t4)));
+ s5 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t5)));
+ s6 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t6)));
+
+ src += 7 * src_stride;
+
+ do {
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ s7 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t0)));
+ s8 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t1)));
+ s9 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t2)));
+ s10 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t3)));
+
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+
+ d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+ d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
+ d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
+
+ store_u8(dst + 0 * dst_stride, dst_stride, d01);
+ store_u8(dst + 2 * dst_stride, dst_stride, d23);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ src += 4 * src_stride;
+ dst += 4 * dst_stride;
+ h -= 4;
+ } while (h != 0);
+ } else {
+ int height;
+ const uint8_t *s;
+ uint8_t *d;
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6, d0, d1, d2, d3;
+ int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
+
+ do {
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+ __builtin_prefetch(src + 4 * src_stride);
+ __builtin_prefetch(src + 5 * src_stride);
+ __builtin_prefetch(src + 6 * src_stride);
+
+ load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ s = src + 7 * src_stride;
+ d = dst;
+ height = h;
+
+ do {
+ load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
+ s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+
+ __builtin_prefetch(d + 0 * dst_stride);
+ __builtin_prefetch(d + 1 * dst_stride);
+ __builtin_prefetch(d + 2 * dst_stride);
+ __builtin_prefetch(d + 3 * dst_stride);
+ __builtin_prefetch(s + 0 * src_stride);
+ __builtin_prefetch(s + 1 * src_stride);
+ __builtin_prefetch(s + 2 * src_stride);
+ __builtin_prefetch(s + 3 * src_stride);
+
+ d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filter);
+ d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filter);
+ d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filter);
+ d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filter);
+
+ store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ s += 4 * src_stride;
+ d += 4 * dst_stride;
+ height -= 4;
+ } while (height != 0);
+ src += 8;
+ dst += 8;
+ w -= 8;
+ } while (w != 0);
+ }
+}
+
+void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *filter, int x0_q4,
+ int x_step_q4, int y0_q4, int y_step_q4, int w,
+ int h) {
+ assert((intptr_t)dst % 4 == 0);
+ assert(dst_stride % 4 == 0);
+ assert(y_step_q4 == 16);
+
+ (void)x0_q4;
+ (void)x_step_q4;
+ (void)y_step_q4;
+
+ if (vpx_get_filter_taps(filter[y0_q4]) <= 4) {
+ /* All 4-tap and bilinear filter values are even, so halve them to reduce
+ * intermediate precision requirements.
+ */
+ const int16x4_t y_filter_4tap = vshr_n_s16(vld1_s16(filter[y0_q4] + 2), 1);
+ vpx_convolve_4tap_vert_neon(src - src_stride, src_stride, dst, dst_stride,
+ w, h, y_filter_4tap);
+ } else {
+ const int16x8_t y_filter_8tap = vld1q_s16(filter[y0_q4]);
+ vpx_convolve_8tap_vert_neon(src - 3 * src_stride, src_stride, dst,
+ dst_stride, w, h, y_filter_8tap);
+ }
+}
+
+void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *filter, int x0_q4,
+ int x_step_q4, int y0_q4, int y_step_q4, int w,
+ int h) {
+ const int16x8_t filters = vld1q_s16(filter[y0_q4]);
+
+ assert((intptr_t)dst % 4 == 0);
+ assert(dst_stride % 4 == 0);
+ assert(y_step_q4 == 16);
+
+ (void)x0_q4;
+ (void)x_step_q4;
+ (void)y_step_q4;
+
+ src -= 3 * src_stride;
+
+ if (w == 4) {
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6, d01, d23, dd01, dd23;
+ int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
+
+ load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6);
+ s0 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t0)));
+ s1 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t1)));
+ s2 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t2)));
+ s3 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t3)));
+ s4 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t4)));
+ s5 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t5)));
+ s6 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t6)));
+
+ src += 7 * src_stride;
+
+ do {
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ s7 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t0)));
+ s8 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t1)));
+ s9 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t2)));
+ s10 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t3)));
+
+ __builtin_prefetch(dst + 0 * dst_stride);
+ __builtin_prefetch(dst + 1 * dst_stride);
+ __builtin_prefetch(dst + 2 * dst_stride);
+ __builtin_prefetch(dst + 3 * dst_stride);
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+
+ d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+ d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+ d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+ d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+ d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
+ d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
+
+ dd01 = load_u8(dst + 0 * dst_stride, dst_stride);
+ dd23 = load_u8(dst + 2 * dst_stride, dst_stride);
+
+ d01 = vrhadd_u8(d01, dd01);
+ d23 = vrhadd_u8(d23, dd23);
+
+ store_u8(dst + 0 * dst_stride, dst_stride, d01);
+ store_u8(dst + 2 * dst_stride, dst_stride, d23);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ src += 4 * src_stride;
+ dst += 4 * dst_stride;
+ h -= 4;
+ } while (h != 0);
+ } else {
+ int height;
+ const uint8_t *s;
+ uint8_t *d;
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6, d0, d1, d2, d3;
+ int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
+
+ do {
+ __builtin_prefetch(src + 0 * src_stride);
+ __builtin_prefetch(src + 1 * src_stride);
+ __builtin_prefetch(src + 2 * src_stride);
+ __builtin_prefetch(src + 3 * src_stride);
+ __builtin_prefetch(src + 4 * src_stride);
+ __builtin_prefetch(src + 5 * src_stride);
+ __builtin_prefetch(src + 6 * src_stride);
+
+ load_u8_8x7(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ s = src + 7 * src_stride;
+ d = dst;
+ height = h;
+
+ do {
+ load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
+ s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+
+ __builtin_prefetch(d + 0 * dst_stride);
+ __builtin_prefetch(d + 1 * dst_stride);
+ __builtin_prefetch(d + 2 * dst_stride);
+ __builtin_prefetch(d + 3 * dst_stride);
+ __builtin_prefetch(s + 0 * src_stride);
+ __builtin_prefetch(s + 1 * src_stride);
+ __builtin_prefetch(s + 2 * src_stride);
+ __builtin_prefetch(s + 3 * src_stride);
+
+ d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+ d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+ d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+ d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+
+ d0 = vrhadd_u8(d0, vld1_u8(d + 0 * dst_stride));
+ d1 = vrhadd_u8(d1, vld1_u8(d + 1 * dst_stride));
+ d2 = vrhadd_u8(d2, vld1_u8(d + 2 * dst_stride));
+ d3 = vrhadd_u8(d3, vld1_u8(d + 3 * dst_stride));
+
+ store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ height -= 4;
+ s += 4 * src_stride;
+ d += 4 * dst_stride;
+ } while (height != 0);
+ src += 8;
+ dst += 8;
+ w -= 8;
+ } while (w != 0);
+ }
+}
generated by cgit v1.2.3 (git 2.39.1) at 2024-06-06 13:33:24 +0000