vp9_convolve8_neon placeholder Call the individually optimized horizontal and vertical functions. This implementation abuses the temp buffer. This will be replaced with a custom optimized function. Over 2x speedup. Change-Id: I5b908d2a73d264e9810d6022bbff73207a3055dd
diff --git a/test/convolve_test.cc b/test/convolve_test.cc index 3320a16..3b72129 100644 --- a/test/convolve_test.cc +++ b/test/convolve_test.cc
@@ -551,7 +551,7 @@ const ConvolveFunctions convolve8_neon( vp9_convolve8_horiz_neon, vp9_convolve8_avg_horiz_neon, vp9_convolve8_vert_neon, vp9_convolve8_avg_vert_neon, - vp9_convolve8_c, vp9_convolve8_avg_c); + vp9_convolve8_neon, vp9_convolve8_avg_neon); INSTANTIATE_TEST_CASE_P(NEON, ConvolveTest, ::testing::Values( make_tuple(4, 4, &convolve8_neon),
diff --git a/vp9/common/arm/neon/vp9_convolve_neon.c b/vp9/common/arm/neon/vp9_convolve_neon.c new file mode 100644 index 0000000..6e37ff6 --- /dev/null +++ b/vp9/common/arm/neon/vp9_convolve_neon.c
@@ -0,0 +1,77 @@ +/* + * Copyright (c) 2013 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 "./vp9_rtcd.h" +#include "vp9/common/vp9_common.h" + +void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, + int w, int h) { + /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the + * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4). + */ + uint8_t temp[64 * 72]; + + // Account for the vertical phase needing 3 lines prior and 4 lines post + int intermediate_height = h + 7; + + if (x_step_q4 != 16 || y_step_q4 != 16) + return vp9_convolve8_c(src, src_stride, + dst, dst_stride, + filter_x, x_step_q4, + filter_y, y_step_q4, + w, h); + + /* Filter starting 3 lines back. The neon implementation will ignore the + * given height and filter a multiple of 4 lines. Since this goes in to + * the temp buffer which has lots of extra room and is subsequently discarded + * this is safe if somewhat less than ideal. + */ + vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride, + temp, 64, + filter_x, x_step_q4, filter_y, y_step_q4, + w, intermediate_height); + + /* Step into the temp buffer 3 lines to get the actual frame data */ + vp9_convolve8_vert_neon(temp + 64 * 3, 64, + dst, dst_stride, + filter_x, x_step_q4, filter_y, y_step_q4, + w, h); +} + +void vp9_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride, + uint8_t *dst, ptrdiff_t dst_stride, + const int16_t *filter_x, int x_step_q4, + const int16_t *filter_y, int y_step_q4, + int w, int h) { + uint8_t temp[64 * 72]; + int intermediate_height = h + 7; + + if (x_step_q4 != 16 || y_step_q4 != 16) + return vp9_convolve8_avg_c(src, src_stride, + dst, dst_stride, + filter_x, x_step_q4, + filter_y, y_step_q4, + w, h); + + /* This implementation has the same issues as above. In addition, we only want + * to average the values after both passes. + */ + vp9_convolve8_horiz_neon(src - src_stride * 3, src_stride, + temp, 64, + filter_x, x_step_q4, filter_y, y_step_q4, + w, intermediate_height); + vp9_convolve8_avg_vert_neon(temp + 64 * 3, + 64, dst, dst_stride, + filter_x, x_step_q4, filter_y, y_step_q4, + w, h); +}
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh index 8bab9f0..25986db 100644 --- a/vp9/common/vp9_rtcd_defs.sh +++ b/vp9/common/vp9_rtcd_defs.sh
@@ -271,7 +271,7 @@ specialize vp9_convolve_avg sse2 prototype void vp9_convolve8 "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h" -specialize vp9_convolve8 ssse3 +specialize vp9_convolve8 ssse3 neon prototype void vp9_convolve8_horiz "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h" specialize vp9_convolve8_horiz ssse3 neon @@ -280,7 +280,7 @@ specialize vp9_convolve8_vert ssse3 neon prototype void vp9_convolve8_avg "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h" -specialize vp9_convolve8_avg ssse3 +specialize vp9_convolve8_avg ssse3 neon prototype void vp9_convolve8_avg_horiz "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h" specialize vp9_convolve8_avg_horiz ssse3 neon
diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk index d23709b..bd956ce 100644 --- a/vp9/vp9_common.mk +++ b/vp9/vp9_common.mk
@@ -88,6 +88,7 @@ VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.c +VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve_neon.c VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve8_neon$(ASM) VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_convolve8_avg_neon$(ASM) VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_loopfilter_neon$(ASM)