diff options
author | chiyotsai <chiyotsai@google.com> | 2019-01-31 12:38:44 -0800 |
---|---|---|
committer | chiyotsai <chiyotsai@google.com> | 2019-02-01 09:20:19 -0800 |
commit | c447a1d51f9b2912e63c3a1143957db09a7924a0 (patch) | |
tree | b42f05a01c850811eb037d59e54cddb42d33fd6c | |
parent | 16b02219433156915d8e7c265e44e7a54dcd21ef (diff) | |
download | libvpx-c447a1d51f9b2912e63c3a1143957db09a7924a0.tar libvpx-c447a1d51f9b2912e63c3a1143957db09a7924a0.tar.gz libvpx-c447a1d51f9b2912e63c3a1143957db09a7924a0.tar.bz2 libvpx-c447a1d51f9b2912e63c3a1143957db09a7924a0.zip |
Remove old version of temporal_filter_apply
BUG=webm:1591
Change-Id: I926566ac1bf4bac8cb1ce1c6ded9ba940109283e
-rw-r--r-- | test/temporal_filter_test.cc | 280 | ||||
-rw-r--r-- | test/test.mk | 1 | ||||
-rw-r--r-- | vp9/common/vp9_rtcd_defs.pl | 3 | ||||
-rw-r--r-- | vp9/encoder/vp9_temporal_filter.c | 130 | ||||
-rw-r--r-- | vp9/encoder/x86/temporal_filter_sse4.c | 207 |
5 files changed, 0 insertions, 621 deletions
diff --git a/test/temporal_filter_test.cc b/test/temporal_filter_test.cc deleted file mode 100644 index d14a4826e..000000000 --- a/test/temporal_filter_test.cc +++ /dev/null @@ -1,280 +0,0 @@ -/* - * Copyright (c) 2016 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 <limits> - -#include "third_party/googletest/src/include/gtest/gtest.h" - -#include "./vp9_rtcd.h" -#include "test/acm_random.h" -#include "test/buffer.h" -#include "test/register_state_check.h" -#include "vpx_ports/vpx_timer.h" - -namespace { - -using ::libvpx_test::ACMRandom; -using ::libvpx_test::Buffer; - -typedef void (*TemporalFilterFunc)(const uint8_t *a, unsigned int stride, - const uint8_t *b, unsigned int w, - unsigned int h, int filter_strength, - int filter_weight, unsigned int *accumulator, - uint16_t *count); - -// Calculate the difference between 'a' and 'b', sum in blocks of 9, and apply -// filter based on strength and weight. Store the resulting filter amount in -// 'count' and apply it to 'b' and store it in 'accumulator'. -void reference_filter(const Buffer<uint8_t> &a, const Buffer<uint8_t> &b, int w, - int h, int filter_strength, int filter_weight, - Buffer<unsigned int> *accumulator, - Buffer<uint16_t> *count) { - Buffer<int> diff_sq = Buffer<int>(w, h, 0); - ASSERT_TRUE(diff_sq.Init()); - diff_sq.Set(0); - - int rounding = 0; - if (filter_strength > 0) { - rounding = 1 << (filter_strength - 1); - } - - ASSERT_TRUE(a.TopLeftPixel() != NULL); - ASSERT_TRUE(b.TopLeftPixel() != NULL); - ASSERT_TRUE(diff_sq.TopLeftPixel() != NULL); - // Calculate all the differences. Avoids re-calculating a bunch of extra - // values. - for (int height = 0; height < h; ++height) { - for (int width = 0; width < w; ++width) { - int diff = a.TopLeftPixel()[height * a.stride() + width] - - b.TopLeftPixel()[height * b.stride() + width]; - diff_sq.TopLeftPixel()[height * diff_sq.stride() + width] = diff * diff; - } - } - - // For any given point, sum the neighboring values and calculate the - // modifier. - for (int height = 0; height < h; ++height) { - for (int width = 0; width < w; ++width) { - // Determine how many values are being summed. - int summed_values = 9; - - if (height == 0 || height == (h - 1)) { - summed_values -= 3; - } - - if (width == 0 || width == (w - 1)) { - if (summed_values == 6) { // corner - summed_values -= 2; - } else { - summed_values -= 3; - } - } - - // Sum the diff_sq of the surrounding values. - int sum = 0; - for (int idy = -1; idy <= 1; ++idy) { - for (int idx = -1; idx <= 1; ++idx) { - const int y = height + idy; - const int x = width + idx; - - // If inside the border. - if (y >= 0 && y < h && x >= 0 && x < w) { - sum += diff_sq.TopLeftPixel()[y * diff_sq.stride() + x]; - } - } - } - - sum *= 3; - sum /= summed_values; - sum += rounding; - sum >>= filter_strength; - - // Clamp the value and invert it. - if (sum > 16) sum = 16; - sum = 16 - sum; - - sum *= filter_weight; - - count->TopLeftPixel()[height * count->stride() + width] += sum; - accumulator->TopLeftPixel()[height * accumulator->stride() + width] += - sum * b.TopLeftPixel()[height * b.stride() + width]; - } - } -} - -class TemporalFilterTest : public ::testing::TestWithParam<TemporalFilterFunc> { - public: - virtual void SetUp() { - filter_func_ = GetParam(); - rnd_.Reset(ACMRandom::DeterministicSeed()); - } - - protected: - TemporalFilterFunc filter_func_; - ACMRandom rnd_; -}; - -TEST_P(TemporalFilterTest, SizeCombinations) { - // Depending on subsampling this function may be called with values of 8 or 16 - // for width and height, in any combination. - Buffer<uint8_t> a = Buffer<uint8_t>(16, 16, 8); - ASSERT_TRUE(a.Init()); - - const int filter_weight = 2; - const int filter_strength = 6; - - for (int width = 8; width <= 16; width += 8) { - for (int height = 8; height <= 16; height += 8) { - // The second buffer must not have any border. - Buffer<uint8_t> b = Buffer<uint8_t>(width, height, 0); - ASSERT_TRUE(b.Init()); - Buffer<unsigned int> accum_ref = Buffer<unsigned int>(width, height, 0); - ASSERT_TRUE(accum_ref.Init()); - Buffer<unsigned int> accum_chk = Buffer<unsigned int>(width, height, 0); - ASSERT_TRUE(accum_chk.Init()); - Buffer<uint16_t> count_ref = Buffer<uint16_t>(width, height, 0); - ASSERT_TRUE(count_ref.Init()); - Buffer<uint16_t> count_chk = Buffer<uint16_t>(width, height, 0); - ASSERT_TRUE(count_chk.Init()); - - // The difference between the buffers must be small to pass the threshold - // to apply the filter. - a.Set(&rnd_, 0, 7); - b.Set(&rnd_, 0, 7); - - accum_ref.Set(rnd_.Rand8()); - accum_chk.CopyFrom(accum_ref); - count_ref.Set(rnd_.Rand8()); - count_chk.CopyFrom(count_ref); - reference_filter(a, b, width, height, filter_strength, filter_weight, - &accum_ref, &count_ref); - ASM_REGISTER_STATE_CHECK( - filter_func_(a.TopLeftPixel(), a.stride(), b.TopLeftPixel(), width, - height, filter_strength, filter_weight, - accum_chk.TopLeftPixel(), count_chk.TopLeftPixel())); - EXPECT_TRUE(accum_chk.CheckValues(accum_ref)); - EXPECT_TRUE(count_chk.CheckValues(count_ref)); - if (HasFailure()) { - printf("Width: %d Height: %d\n", width, height); - count_chk.PrintDifference(count_ref); - accum_chk.PrintDifference(accum_ref); - return; - } - } - } -} - -TEST_P(TemporalFilterTest, CompareReferenceRandom) { - for (int width = 8; width <= 16; width += 8) { - for (int height = 8; height <= 16; height += 8) { - Buffer<uint8_t> a = Buffer<uint8_t>(width, height, 8); - ASSERT_TRUE(a.Init()); - // The second buffer must not have any border. - Buffer<uint8_t> b = Buffer<uint8_t>(width, height, 0); - ASSERT_TRUE(b.Init()); - Buffer<unsigned int> accum_ref = Buffer<unsigned int>(width, height, 0); - ASSERT_TRUE(accum_ref.Init()); - Buffer<unsigned int> accum_chk = Buffer<unsigned int>(width, height, 0); - ASSERT_TRUE(accum_chk.Init()); - Buffer<uint16_t> count_ref = Buffer<uint16_t>(width, height, 0); - ASSERT_TRUE(count_ref.Init()); - Buffer<uint16_t> count_chk = Buffer<uint16_t>(width, height, 0); - ASSERT_TRUE(count_chk.Init()); - - for (int filter_strength = 0; filter_strength <= 6; ++filter_strength) { - for (int filter_weight = 0; filter_weight <= 2; ++filter_weight) { - for (int repeat = 0; repeat < 100; ++repeat) { - if (repeat < 50) { - a.Set(&rnd_, 0, 7); - b.Set(&rnd_, 0, 7); - } else { - // Check large (but close) values as well. - a.Set(&rnd_, std::numeric_limits<uint8_t>::max() - 7, - std::numeric_limits<uint8_t>::max()); - b.Set(&rnd_, std::numeric_limits<uint8_t>::max() - 7, - std::numeric_limits<uint8_t>::max()); - } - - accum_ref.Set(rnd_.Rand8()); - accum_chk.CopyFrom(accum_ref); - count_ref.Set(rnd_.Rand8()); - count_chk.CopyFrom(count_ref); - reference_filter(a, b, width, height, filter_strength, - filter_weight, &accum_ref, &count_ref); - ASM_REGISTER_STATE_CHECK(filter_func_( - a.TopLeftPixel(), a.stride(), b.TopLeftPixel(), width, height, - filter_strength, filter_weight, accum_chk.TopLeftPixel(), - count_chk.TopLeftPixel())); - EXPECT_TRUE(accum_chk.CheckValues(accum_ref)); - EXPECT_TRUE(count_chk.CheckValues(count_ref)); - if (HasFailure()) { - printf("Weight: %d Strength: %d\n", filter_weight, - filter_strength); - count_chk.PrintDifference(count_ref); - accum_chk.PrintDifference(accum_ref); - return; - } - } - } - } - } - } -} - -TEST_P(TemporalFilterTest, DISABLED_Speed) { - Buffer<uint8_t> a = Buffer<uint8_t>(16, 16, 8); - ASSERT_TRUE(a.Init()); - - const int filter_weight = 2; - const int filter_strength = 6; - - for (int width = 8; width <= 16; width += 8) { - for (int height = 8; height <= 16; height += 8) { - // The second buffer must not have any border. - Buffer<uint8_t> b = Buffer<uint8_t>(width, height, 0); - ASSERT_TRUE(b.Init()); - Buffer<unsigned int> accum_ref = Buffer<unsigned int>(width, height, 0); - ASSERT_TRUE(accum_ref.Init()); - Buffer<unsigned int> accum_chk = Buffer<unsigned int>(width, height, 0); - ASSERT_TRUE(accum_chk.Init()); - Buffer<uint16_t> count_ref = Buffer<uint16_t>(width, height, 0); - ASSERT_TRUE(count_ref.Init()); - Buffer<uint16_t> count_chk = Buffer<uint16_t>(width, height, 0); - ASSERT_TRUE(count_chk.Init()); - - a.Set(&rnd_, 0, 7); - b.Set(&rnd_, 0, 7); - - accum_chk.Set(0); - count_chk.Set(0); - - vpx_usec_timer timer; - vpx_usec_timer_start(&timer); - for (int i = 0; i < 10000; ++i) { - filter_func_(a.TopLeftPixel(), a.stride(), b.TopLeftPixel(), width, - height, filter_strength, filter_weight, - accum_chk.TopLeftPixel(), count_chk.TopLeftPixel()); - } - vpx_usec_timer_mark(&timer); - const int elapsed_time = static_cast<int>(vpx_usec_timer_elapsed(&timer)); - printf("Temporal filter %dx%d time: %5d us\n", width, height, - elapsed_time); - } - } -} - -INSTANTIATE_TEST_CASE_P(C, TemporalFilterTest, - ::testing::Values(&vp9_temporal_filter_apply_c)); - -#if HAVE_SSE4_1 -INSTANTIATE_TEST_CASE_P(SSE4_1, TemporalFilterTest, - ::testing::Values(&vp9_temporal_filter_apply_sse4_1)); -#endif // HAVE_SSE4_1 -} // namespace diff --git a/test/test.mk b/test/test.mk index 2b7636185..61eb6060f 100644 --- a/test/test.mk +++ b/test/test.mk @@ -170,7 +170,6 @@ LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += hadamard_test.cc LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += minmax_test.cc LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_scale_test.cc ifneq ($(CONFIG_REALTIME_ONLY),yes) -LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += temporal_filter_test.cc LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += yuv_temporal_filter_test.cc endif LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += variance_test.cc diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl index 42705dd26..00c4414ad 100644 --- a/vp9/common/vp9_rtcd_defs.pl +++ b/vp9/common/vp9_rtcd_defs.pl @@ -187,9 +187,6 @@ specialize qw/vp9_diamond_search_sad avx/; # Apply temporal filter # if (vpx_config("CONFIG_REALTIME_ONLY") ne "yes") { -add_proto qw/void vp9_temporal_filter_apply/, "const uint8_t *frame1, unsigned int stride, const uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, uint32_t *accumulator, uint16_t *count"; -specialize qw/vp9_temporal_filter_apply sse4_1/; - add_proto qw/void vp9_apply_temporal_filter/, "const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre, int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src, int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre, int uv_pre_stride, unsigned int block_width, unsigned int block_height, int ss_x, int ss_y, int strength, const int *const blk_fw, int use_32x32, uint32_t *y_accumulator, uint16_t *y_count, uint32_t *u_accumulator, uint16_t *u_count, uint32_t *v_accumulator, uint16_t *v_count"; specialize qw/vp9_apply_temporal_filter sse4_1/; diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c index c622fd88b..d02603615 100644 --- a/vp9/encoder/vp9_temporal_filter.c +++ b/vp9/encoder/vp9_temporal_filter.c @@ -386,137 +386,7 @@ void vp9_apply_temporal_filter_c( } } -// TODO(any): This function is not used anymore. Should be removed. -void vp9_temporal_filter_apply_c(const uint8_t *frame1, unsigned int stride, - const uint8_t *frame2, - unsigned int block_width, - unsigned int block_height, int strength, - int filter_weight, uint32_t *accumulator, - uint16_t *count) { - unsigned int i, j, k; - int modifier; - int byte = 0; - const int rounding = (1 << strength) >> 1; - - assert(strength >= 0); - assert(strength <= 6); - - assert(filter_weight >= 0); - assert(filter_weight <= 2); - - for (i = 0, k = 0; i < block_height; i++) { - for (j = 0; j < block_width; j++, k++) { - int pixel_value = *frame2; - - // non-local mean approach - int diff_sse[9] = { 0 }; - int idx, idy, index = 0; - - for (idy = -1; idy <= 1; ++idy) { - for (idx = -1; idx <= 1; ++idx) { - int row = (int)i + idy; - int col = (int)j + idx; - - if (row >= 0 && row < (int)block_height && col >= 0 && - col < (int)block_width) { - int diff = frame1[byte + idy * (int)stride + idx] - - frame2[idy * (int)block_width + idx]; - diff_sse[index] = diff * diff; - ++index; - } - } - } - - assert(index > 0); - - modifier = 0; - for (idx = 0; idx < 9; ++idx) modifier += diff_sse[idx]; - - modifier *= 3; - modifier /= index; - - ++frame2; - - modifier += rounding; - modifier >>= strength; - - if (modifier > 16) modifier = 16; - - modifier = 16 - modifier; - modifier *= filter_weight; - - count[k] += modifier; - accumulator[k] += modifier * pixel_value; - - byte++; - } - - byte += stride - block_width; - } -} - #if CONFIG_VP9_HIGHBITDEPTH -void vp9_highbd_temporal_filter_apply_c( - const uint8_t *frame1_8, unsigned int stride, const uint8_t *frame2_8, - unsigned int block_width, unsigned int block_height, int strength, - int *blk_fw, int use_32x32, uint32_t *accumulator, uint16_t *count) { - const uint16_t *frame1 = CONVERT_TO_SHORTPTR(frame1_8); - const uint16_t *frame2 = CONVERT_TO_SHORTPTR(frame2_8); - unsigned int i, j, k; - int modifier; - const int rounding = strength > 0 ? 1 << (strength - 1) : 0; - - int diff_sse[BLK_PELS] = { 0 }; - int this_idx = 0; - - for (i = 0; i < block_height; i++) { - for (j = 0; j < block_width; j++) { - const int diff = - frame1[i * (int)stride + j] - frame2[i * (int)block_width + j]; - diff_sse[this_idx++] = diff * diff; - } - } - - modifier = 0; - for (i = 0, k = 0; i < block_height; i++) { - for (j = 0; j < block_width; j++, k++) { - int pixel_value = frame2[i * (int)block_width + j]; - int filter_weight = - get_filter_weight(i, j, block_height, block_width, blk_fw, use_32x32); - - int idx, idy, index = 0; - - for (idy = -1; idy <= 1; ++idy) { - for (idx = -1; idx <= 1; ++idx) { - int row = (int)i + idy; - int col = (int)j + idx; - - if (row >= 0 && row < (int)block_height && col >= 0 && - col < (int)block_width) { - modifier += diff_sse[row * (int)block_width + col]; - ++index; - } - } - } - assert(index > 0); - - modifier *= 3; - modifier /= index; - - modifier += rounding; - modifier >>= strength; - - if (modifier > 16) modifier = 16; - - modifier = 16 - modifier; - modifier *= filter_weight; - - count[k] += modifier; - accumulator[k] += modifier * pixel_value; - } - } -} - void vp9_highbd_apply_temporal_filter_c( const uint16_t *y_src, int y_src_stride, const uint16_t *y_pre, int y_pre_stride, const uint16_t *u_src, const uint16_t *v_src, diff --git a/vp9/encoder/x86/temporal_filter_sse4.c b/vp9/encoder/x86/temporal_filter_sse4.c index b560e2218..9f9483a9b 100644 --- a/vp9/encoder/x86/temporal_filter_sse4.c +++ b/vp9/encoder/x86/temporal_filter_sse4.c @@ -18,71 +18,6 @@ #include "vp9/encoder/vp9_temporal_filter.h" #include "vp9/encoder/x86/temporal_filter_constants.h" -// Load values from 'a' and 'b'. Compute the difference squared and sum -// neighboring values such that: -// sum[1] = (a[0]-b[0])^2 + (a[1]-b[1])^2 + (a[2]-b[2])^2 -// Values to the left and right of the row are set to 0. -// The values are returned in sum_0 and sum_1 as *unsigned* 16 bit values. -static void sum_8(const uint8_t *a, const uint8_t *b, __m128i *sum) { - const __m128i a_u8 = _mm_loadl_epi64((const __m128i *)a); - const __m128i b_u8 = _mm_loadl_epi64((const __m128i *)b); - - const __m128i a_u16 = _mm_cvtepu8_epi16(a_u8); - const __m128i b_u16 = _mm_cvtepu8_epi16(b_u8); - - const __m128i diff_s16 = _mm_sub_epi16(a_u16, b_u16); - const __m128i diff_sq_u16 = _mm_mullo_epi16(diff_s16, diff_s16); - - // Shift all the values one place to the left/right so we can efficiently sum - // diff_sq_u16[i - 1] + diff_sq_u16[i] + diff_sq_u16[i + 1]. - const __m128i shift_left = _mm_slli_si128(diff_sq_u16, 2); - const __m128i shift_right = _mm_srli_si128(diff_sq_u16, 2); - - // It becomes necessary to treat the values as unsigned at this point. The - // 255^2 fits in uint16_t but not int16_t. Use saturating adds from this point - // forward since the filter is only applied to smooth small pixel changes. - // Once the value has saturated to uint16_t it is well outside the useful - // range. - __m128i sum_u16 = _mm_adds_epu16(diff_sq_u16, shift_left); - sum_u16 = _mm_adds_epu16(sum_u16, shift_right); - - *sum = sum_u16; -} - -static void sum_16(const uint8_t *a, const uint8_t *b, __m128i *sum_0, - __m128i *sum_1) { - const __m128i zero = _mm_setzero_si128(); - const __m128i a_u8 = _mm_loadu_si128((const __m128i *)a); - const __m128i b_u8 = _mm_loadu_si128((const __m128i *)b); - - const __m128i a_0_u16 = _mm_cvtepu8_epi16(a_u8); - const __m128i a_1_u16 = _mm_unpackhi_epi8(a_u8, zero); - const __m128i b_0_u16 = _mm_cvtepu8_epi16(b_u8); - const __m128i b_1_u16 = _mm_unpackhi_epi8(b_u8, zero); - - const __m128i diff_0_s16 = _mm_sub_epi16(a_0_u16, b_0_u16); - const __m128i diff_1_s16 = _mm_sub_epi16(a_1_u16, b_1_u16); - const __m128i diff_sq_0_u16 = _mm_mullo_epi16(diff_0_s16, diff_0_s16); - const __m128i diff_sq_1_u16 = _mm_mullo_epi16(diff_1_s16, diff_1_s16); - - __m128i shift_left = _mm_slli_si128(diff_sq_0_u16, 2); - // Use _mm_alignr_epi8() to "shift in" diff_sq_u16[8]. - __m128i shift_right = _mm_alignr_epi8(diff_sq_1_u16, diff_sq_0_u16, 2); - - __m128i sum_u16 = _mm_adds_epu16(diff_sq_0_u16, shift_left); - sum_u16 = _mm_adds_epu16(sum_u16, shift_right); - - *sum_0 = sum_u16; - - shift_left = _mm_alignr_epi8(diff_sq_1_u16, diff_sq_0_u16, 14); - shift_right = _mm_srli_si128(diff_sq_1_u16, 2); - - sum_u16 = _mm_adds_epu16(diff_sq_1_u16, shift_left); - sum_u16 = _mm_adds_epu16(sum_u16, shift_right); - - *sum_1 = sum_u16; -} - // Read in 8 pixels from a and b as 8-bit unsigned integers, compute the // difference squared, and store as unsigned 16-bit integer to dst. static INLINE void store_dist_8(const uint8_t *a, const uint8_t *b, @@ -312,148 +247,6 @@ static INLINE void get_sum_16(const uint16_t *y_dist, __m128i *sum_first, get_sum_8(y_dist + 8, sum_second); } -void vp9_temporal_filter_apply_sse4_1(const uint8_t *a, unsigned int stride, - const uint8_t *b, unsigned int width, - unsigned int height, int strength, - int weight, uint32_t *accumulator, - uint16_t *count) { - unsigned int h; - const int rounding = (1 << strength) >> 1; - - assert(strength >= 0); - assert(strength <= 6); - - assert(weight >= 0); - assert(weight <= 2); - - assert(width == 8 || width == 16); - - if (width == 8) { - __m128i sum_row_a, sum_row_b, sum_row_c; - __m128i mul_constants = _mm_setr_epi16( - NEIGHBOR_CONSTANT_4, NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_4); - - sum_8(a, b, &sum_row_a); - sum_8(a + stride, b + width, &sum_row_b); - sum_row_c = _mm_adds_epu16(sum_row_a, sum_row_b); - sum_row_c = - average_8(sum_row_c, &mul_constants, strength, rounding, weight); - accumulate_and_store_8(sum_row_c, b, count, accumulator); - - a += stride + stride; - b += width; - count += width; - accumulator += width; - - mul_constants = _mm_setr_epi16(NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_6); - - for (h = 0; h < height - 2; ++h) { - sum_8(a, b + width, &sum_row_c); - sum_row_a = _mm_adds_epu16(sum_row_a, sum_row_b); - sum_row_a = _mm_adds_epu16(sum_row_a, sum_row_c); - sum_row_a = - average_8(sum_row_a, &mul_constants, strength, rounding, weight); - accumulate_and_store_8(sum_row_a, b, count, accumulator); - - a += stride; - b += width; - count += width; - accumulator += width; - - sum_row_a = sum_row_b; - sum_row_b = sum_row_c; - } - - mul_constants = _mm_setr_epi16(NEIGHBOR_CONSTANT_4, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_4); - sum_row_a = _mm_adds_epu16(sum_row_a, sum_row_b); - sum_row_a = - average_8(sum_row_a, &mul_constants, strength, rounding, weight); - accumulate_and_store_8(sum_row_a, b, count, accumulator); - - } else { // width == 16 - __m128i sum_row_a_0, sum_row_a_1; - __m128i sum_row_b_0, sum_row_b_1; - __m128i sum_row_c_0, sum_row_c_1; - __m128i mul_constants_0 = _mm_setr_epi16( - NEIGHBOR_CONSTANT_4, NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6), - mul_constants_1 = _mm_setr_epi16( - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_4); - - sum_16(a, b, &sum_row_a_0, &sum_row_a_1); - sum_16(a + stride, b + width, &sum_row_b_0, &sum_row_b_1); - - sum_row_c_0 = _mm_adds_epu16(sum_row_a_0, sum_row_b_0); - sum_row_c_1 = _mm_adds_epu16(sum_row_a_1, sum_row_b_1); - - average_16(&sum_row_c_0, &sum_row_c_1, &mul_constants_0, &mul_constants_1, - strength, rounding, weight); - accumulate_and_store_16(sum_row_c_0, sum_row_c_1, b, count, accumulator); - - a += stride + stride; - b += width; - count += width; - accumulator += width; - - mul_constants_0 = _mm_setr_epi16(NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_9); - mul_constants_1 = _mm_setr_epi16(NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_9, - NEIGHBOR_CONSTANT_9, NEIGHBOR_CONSTANT_6); - for (h = 0; h < height - 2; ++h) { - sum_16(a, b + width, &sum_row_c_0, &sum_row_c_1); - - sum_row_a_0 = _mm_adds_epu16(sum_row_a_0, sum_row_b_0); - sum_row_a_0 = _mm_adds_epu16(sum_row_a_0, sum_row_c_0); - sum_row_a_1 = _mm_adds_epu16(sum_row_a_1, sum_row_b_1); - sum_row_a_1 = _mm_adds_epu16(sum_row_a_1, sum_row_c_1); - - average_16(&sum_row_a_0, &sum_row_a_1, &mul_constants_0, &mul_constants_1, - strength, rounding, weight); - accumulate_and_store_16(sum_row_a_0, sum_row_a_1, b, count, accumulator); - - a += stride; - b += width; - count += width; - accumulator += width; - - sum_row_a_0 = sum_row_b_0; - sum_row_a_1 = sum_row_b_1; - sum_row_b_0 = sum_row_c_0; - sum_row_b_1 = sum_row_c_1; - } - - mul_constants_0 = _mm_setr_epi16(NEIGHBOR_CONSTANT_4, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6); - mul_constants_1 = _mm_setr_epi16(NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6, - NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_4); - sum_row_c_0 = _mm_adds_epu16(sum_row_a_0, sum_row_b_0); - sum_row_c_1 = _mm_adds_epu16(sum_row_a_1, sum_row_b_1); - - average_16(&sum_row_c_0, &sum_row_c_1, &mul_constants_0, &mul_constants_1, - strength, rounding, weight); - accumulate_and_store_16(sum_row_c_0, sum_row_c_1, b, count, accumulator); - } -} - // Read in a row of chroma values corresponds to a row of 16 luma values. static INLINE void read_chroma_dist_row_16(int ss_x, const uint16_t *u_dist, const uint16_t *v_dist, |