From a9ebbcc33836d8666f98de7a6d416a6dcd0cae9a Mon Sep 17 00:00:00 2001 From: John Koleszar Date: Thu, 18 Apr 2013 13:05:38 -0700 Subject: [PATCH] convolve: support larger blocks, fix asm saturation bug Updates the common convoloution code to support blocks larger than 16x16, and rectangular blocks. This uncovered a bug in the SSSE3 filtering routines due to the order of application of saturation. This commit fixes that bug, adjusts the unit test to bias its random values towards the extremes, and adds a test to ensure that all filters conform to the expected pairwise addition structure. Change-Id: I81f69668b1de0de5a8ed43f0643845641525c8f0 --- test/acm_random.h | 7 ++ test/convolve_test.cc | 93 +++++++++++++++++------- vp9/common/vp9_convolve.c | 38 +++++----- vp9/common/x86/vp9_asm_stubs.c | 12 +-- vp9/common/x86/vp9_subpixel_8t_ssse3.asm | 16 ++-- 5 files changed, 105 insertions(+), 61 deletions(-) diff --git a/test/acm_random.h b/test/acm_random.h index 514894eda..84c6c7529 100644 --- a/test/acm_random.h +++ b/test/acm_random.h @@ -35,6 +35,13 @@ class ACMRandom { return (rand() >> 8) & 0xff; } + uint8_t Rand8Extremes(void) { + // Returns a random value near 0 or near 255, to better exercise + // saturation behavior. + const uint8_t r = Rand8(); + return r < 128 ? r << 4 : r >> 4; + } + int PseudoUniform(int range) { return (rand() >> 8) % range; } diff --git a/test/convolve_test.cc b/test/convolve_test.cc index 35065a41f..a8139cbd3 100644 --- a/test/convolve_test.cc +++ b/test/convolve_test.cc @@ -66,7 +66,7 @@ static void filter_block2d_8_c(const uint8_t *src_ptr, // support. const int kInterp_Extend = 4; const unsigned int intermediate_height = - (kInterp_Extend - 1) + output_height + kInterp_Extend; + (kInterp_Extend - 1) + output_height + kInterp_Extend; /* Size of intermediate_buffer is max_intermediate_height * filter_max_width, * where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height @@ -75,7 +75,7 @@ static void filter_block2d_8_c(const uint8_t *src_ptr, * = 23 * and filter_max_width = 16 */ - uint8_t intermediate_buffer[23 * 16]; + uint8_t intermediate_buffer[71 * 64]; const int intermediate_next_stride = 1 - intermediate_height * output_width; // Horizontal pass (src -> transposed intermediate). @@ -158,13 +158,13 @@ static void filter_average_block2d_8_c(const uint8_t *src_ptr, unsigned int dst_stride, unsigned int output_width, unsigned int output_height) { - uint8_t tmp[16*16]; + uint8_t tmp[64*64]; - assert(output_width <= 16); - assert(output_height <= 16); - filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 16, + assert(output_width <= 64); + assert(output_height <= 64); + filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 64, output_width, output_height); - block2d_average_c(tmp, 16, dst_ptr, dst_stride, + block2d_average_c(tmp, 64, dst_ptr, dst_stride, output_width, output_height); } @@ -188,10 +188,10 @@ class ConvolveTest : public PARAMS(int, int, const ConvolveFunctions*) { protected: static const int kDataAlignment = 16; - static const int kOuterBlockSize = 32; + static const int kOuterBlockSize = 128; static const int kInputStride = kOuterBlockSize; static const int kOutputStride = kOuterBlockSize; - static const int kMaxDimension = 16; + static const int kMaxDimension = 64; int Width() const { return GET_PARAM(0); } int Height() const { return GET_PARAM(1); } @@ -221,7 +221,7 @@ class ConvolveTest : public PARAMS(int, int, const ConvolveFunctions*) { ::libvpx_test::ACMRandom prng; for (int i = 0; i < kOuterBlockSize * kOuterBlockSize; ++i) - input_[i] = prng.Rand8(); + input_[i] = prng.Rand8Extremes(); } void CheckGuardBlocks() { @@ -308,6 +308,29 @@ const int16_t (*kTestFilterList[])[8] = { vp9_sub_pel_filters_8s, vp9_sub_pel_filters_8lp }; +const int kNumFilterBanks = sizeof(kTestFilterList) / + sizeof(kTestFilterList[0]); +const int kNumFilters = 16; + +TEST(ConvolveTest, FiltersWontSaturateWhenAddedPairwise) { + for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) { + const int16_t (*filters)[8] = kTestFilterList[filter_bank]; + for (int i = 0; i < kNumFilters; i++) { + const int p0 = filters[i][0] + filters[i][1]; + const int p1 = filters[i][2] + filters[i][3]; + const int p2 = filters[i][4] + filters[i][5]; + const int p3 = filters[i][6] + filters[i][7]; + EXPECT_LE(p0, 128); + EXPECT_LE(p1, 128); + EXPECT_LE(p2, 128); + EXPECT_LE(p3, 128); + EXPECT_LE(p0 + p3, 128); + EXPECT_LE(p0 + p3 + p1, 128); + EXPECT_LE(p0 + p3 + p1 + p2, 128); + EXPECT_EQ(p0 + p1 + p2 + p3, 128); + } + } +} const int16_t kInvalidFilter[8] = { 0 }; @@ -316,12 +339,9 @@ TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) { uint8_t* const out = output(); uint8_t ref[kOutputStride * kMaxDimension]; - const int kNumFilterBanks = sizeof(kTestFilterList) / - sizeof(kTestFilterList[0]); for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) { const int16_t (*filters)[8] = kTestFilterList[filter_bank]; - const int kNumFilters = 16; for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) { for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) { @@ -368,7 +388,7 @@ TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) { ::libvpx_test::ACMRandom prng; for (int y = 0; y < Height(); ++y) { for (int x = 0; x < Width(); ++x) { - const uint8_t r = prng.Rand8(); + const uint8_t r = prng.Rand8Extremes(); out[y * kOutputStride + x] = r; ref[y * kOutputStride + x] = r; @@ -440,16 +460,17 @@ DECLARE_ALIGNED(256, const int16_t, kChangeFilters[16][8]) = { TEST_P(ConvolveTest, ChangeFilterWorks) { uint8_t* const in = input(); uint8_t* const out = output(); + const int kPixelSelected = 4; REGISTER_STATE_CHECK(UUT_->h8_(in, kInputStride, out, kOutputStride, kChangeFilters[8], 17, kChangeFilters[4], 16, Width(), Height())); for (int x = 0; x < Width(); ++x) { - if (x < 8) - ASSERT_EQ(in[4], out[x]) << "x == " << x; - else - ASSERT_EQ(in[12], out[x]) << "x == " << x; + const int kQ4StepAdjust = x >> 4; + const int kFilterPeriodAdjust = (x >> 3) << 3; + const int ref_x = kQ4StepAdjust + kFilterPeriodAdjust + kPixelSelected; + ASSERT_EQ(in[ref_x], out[x]) << "x == " << x; } REGISTER_STATE_CHECK(UUT_->v8_(in, kInputStride, out, kOutputStride, @@ -457,10 +478,10 @@ TEST_P(ConvolveTest, ChangeFilterWorks) { Width(), Height())); for (int y = 0; y < Height(); ++y) { - if (y < 8) - ASSERT_EQ(in[4 * kInputStride], out[y * kOutputStride]) << "y == " << y; - else - ASSERT_EQ(in[12 * kInputStride], out[y * kOutputStride]) << "y == " << y; + const int kQ4StepAdjust = y >> 4; + const int kFilterPeriodAdjust = (y >> 3) << 3; + const int ref_y = kQ4StepAdjust + kFilterPeriodAdjust + kPixelSelected; + ASSERT_EQ(in[ref_y * kInputStride], out[y * kInputStride]) << "y == " << y; } REGISTER_STATE_CHECK(UUT_->hv8_(in, kInputStride, out, kOutputStride, @@ -468,9 +489,13 @@ TEST_P(ConvolveTest, ChangeFilterWorks) { Width(), Height())); for (int y = 0; y < Height(); ++y) { + const int kQ4StepAdjustY = y >> 4; + const int kFilterPeriodAdjustY = (y >> 3) << 3; + const int ref_y = kQ4StepAdjustY + kFilterPeriodAdjustY + kPixelSelected; for (int x = 0; x < Width(); ++x) { - const int ref_x = x < 8 ? 4 : 12; - const int ref_y = y < 8 ? 4 : 12; + const int kQ4StepAdjustX = x >> 4; + const int kFilterPeriodAdjustX = (x >> 3) << 3; + const int ref_x = kQ4StepAdjustX + kFilterPeriodAdjustX + kPixelSelected; ASSERT_EQ(in[ref_y * kInputStride + ref_x], out[y * kOutputStride + x]) << "x == " << x << ", y == " << y; @@ -489,9 +514,17 @@ const ConvolveFunctions convolve8_c( INSTANTIATE_TEST_CASE_P(C, ConvolveTest, ::testing::Values( make_tuple(4, 4, &convolve8_c), make_tuple(8, 4, &convolve8_c), + make_tuple(4, 8, &convolve8_c), make_tuple(8, 8, &convolve8_c), make_tuple(16, 8, &convolve8_c), - make_tuple(16, 16, &convolve8_c))); + make_tuple(8, 16, &convolve8_c), + make_tuple(16, 16, &convolve8_c), + make_tuple(32, 16, &convolve8_c), + make_tuple(16, 32, &convolve8_c), + make_tuple(32, 32, &convolve8_c), + make_tuple(64, 32, &convolve8_c), + make_tuple(32, 64, &convolve8_c), + make_tuple(64, 64, &convolve8_c))); } #if HAVE_SSSE3 @@ -503,7 +536,15 @@ const ConvolveFunctions convolve8_ssse3( INSTANTIATE_TEST_CASE_P(SSSE3, ConvolveTest, ::testing::Values( make_tuple(4, 4, &convolve8_ssse3), make_tuple(8, 4, &convolve8_ssse3), + make_tuple(4, 8, &convolve8_ssse3), make_tuple(8, 8, &convolve8_ssse3), make_tuple(16, 8, &convolve8_ssse3), - make_tuple(16, 16, &convolve8_ssse3))); + make_tuple(8, 16, &convolve8_ssse3), + make_tuple(16, 16, &convolve8_ssse3), + make_tuple(32, 16, &convolve8_ssse3), + make_tuple(16, 32, &convolve8_ssse3), + make_tuple(32, 32, &convolve8_ssse3), + make_tuple(64, 32, &convolve8_ssse3), + make_tuple(32, 64, &convolve8_ssse3), + make_tuple(64, 64, &convolve8_ssse3))); #endif diff --git a/vp9/common/vp9_convolve.c b/vp9/common/vp9_convolve.c index 3ab8bec7a..a27ca6f5d 100644 --- a/vp9/common/vp9_convolve.c +++ b/vp9/common/vp9_convolve.c @@ -331,14 +331,14 @@ static void convolve_c(const uint8_t *src, int src_stride, const int16_t *filter_y, int y_step_q4, int w, int h, int taps) { /* Fixed size intermediate buffer places limits on parameters. - * Maximum intermediate_height is 39, for y_step_q4 == 32, - * h == 16, taps == 8. + * Maximum intermediate_height is 135, for y_step_q4 == 32, + * h == 64, taps == 8. */ - uint8_t temp[16 * 39]; + uint8_t temp[64 * 135]; int intermediate_height = ((h * y_step_q4) >> 4) + taps - 1; - assert(w <= 16); - assert(h <= 16); + assert(w <= 64); + assert(h <= 64); assert(taps <= 8); assert(y_step_q4 <= 32); @@ -346,10 +346,10 @@ static void convolve_c(const uint8_t *src, int src_stride, intermediate_height = h; convolve_horiz_c(src - src_stride * (taps / 2 - 1), src_stride, - temp, 16, + temp, 64, filter_x, x_step_q4, filter_y, y_step_q4, w, intermediate_height, taps); - convolve_vert_c(temp + 16 * (taps / 2 - 1), 16, dst, dst_stride, + convolve_vert_c(temp + 64 * (taps / 2 - 1), 64, dst, dst_stride, filter_x, x_step_q4, filter_y, y_step_q4, w, h, taps); } @@ -360,14 +360,14 @@ static void convolve_avg_c(const uint8_t *src, int src_stride, const int16_t *filter_y, int y_step_q4, int w, int h, int taps) { /* Fixed size intermediate buffer places limits on parameters. - * Maximum intermediate_height is 39, for y_step_q4 == 32, - * h == 16, taps == 8. + * Maximum intermediate_height is 135, for y_step_q4 == 32, + * h == 64, taps == 8. */ - uint8_t temp[16 * 39]; + uint8_t temp[64 * 135]; int intermediate_height = ((h * y_step_q4) >> 4) + taps - 1; - assert(w <= 16); - assert(h <= 16); + assert(w <= 64); + assert(h <= 64); assert(taps <= 8); assert(y_step_q4 <= 32); @@ -375,10 +375,10 @@ static void convolve_avg_c(const uint8_t *src, int src_stride, intermediate_height = h; convolve_horiz_c(src - src_stride * (taps / 2 - 1), src_stride, - temp, 16, + temp, 64, filter_x, x_step_q4, filter_y, y_step_q4, w, intermediate_height, taps); - convolve_avg_vert_c(temp + 16 * (taps / 2 - 1), 16, dst, dst_stride, + convolve_avg_vert_c(temp + 64 * (taps / 2 - 1), 64, dst, dst_stride, filter_x, x_step_q4, filter_y, y_step_q4, w, h, taps); } @@ -563,16 +563,16 @@ void vp9_convolve8_avg_c(const uint8_t *src, int src_stride, const int16_t *filter_y, int y_step_q4, int w, int h) { /* Fixed size intermediate buffer places limits on parameters. */ - DECLARE_ALIGNED_ARRAY(16, uint8_t, temp, 16 * 16); - assert(w <= 16); - assert(h <= 16); + DECLARE_ALIGNED_ARRAY(16, uint8_t, temp, 64 * 64); + assert(w <= 64); + assert(h <= 64); vp9_convolve8(src, src_stride, - temp, 16, + temp, 64, filter_x, x_step_q4, filter_y, y_step_q4, w, h); - vp9_convolve_avg(temp, 16, + vp9_convolve_avg(temp, 64, dst, dst_stride, NULL, 0, /* These unused parameter should be removed! */ NULL, 0, /* These unused parameter should be removed! */ diff --git a/vp9/common/x86/vp9_asm_stubs.c b/vp9/common/x86/vp9_asm_stubs.c index 6d3bb021a..310f8ed24 100644 --- a/vp9/common/x86/vp9_asm_stubs.c +++ b/vp9/common/x86/vp9_asm_stubs.c @@ -278,11 +278,9 @@ void vp9_convolve8_ssse3(const uint8_t *src, int src_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h) { - DECLARE_ALIGNED_ARRAY(16, unsigned char, fdata2, 16*23); + DECLARE_ALIGNED_ARRAY(16, unsigned char, fdata2, 16*71); - // check w/h due to fixed size fdata2 array - assert(w <= 16); - assert(h <= 16); + assert(h <= 64); if (x_step_q4 == 16 && y_step_q4 == 16 && filter_x[3] != 128 && filter_y[3] != 128) { @@ -324,11 +322,9 @@ void vp9_convolve8_avg_ssse3(const uint8_t *src, int src_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h) { - DECLARE_ALIGNED_ARRAY(16, unsigned char, fdata2, 16*23); + DECLARE_ALIGNED_ARRAY(16, unsigned char, fdata2, 16*71); - // check w/h due to fixed size fdata2 array - assert(w <= 16); - assert(h <= 16); + assert(h <= 64); if (x_step_q4 == 16 && y_step_q4 == 16 && filter_x[3] != 128 && filter_y[3] != 128) { diff --git a/vp9/common/x86/vp9_subpixel_8t_ssse3.asm b/vp9/common/x86/vp9_subpixel_8t_ssse3.asm index 32f00e289..bbf9888ca 100644 --- a/vp9/common/x86/vp9_subpixel_8t_ssse3.asm +++ b/vp9/common/x86/vp9_subpixel_8t_ssse3.asm @@ -81,10 +81,10 @@ pmaddubsw xmm4, k4k5 pmaddubsw xmm6, k6k7 + paddsw xmm0, xmm6 paddsw xmm0, xmm2 - paddsw xmm0, krd - paddsw xmm4, xmm6 paddsw xmm0, xmm4 + paddsw xmm0, krd psraw xmm0, 7 packuswb xmm0, xmm0 @@ -165,10 +165,10 @@ pmaddubsw xmm4, k4k5 pmaddubsw xmm6, k6k7 + paddsw xmm0, xmm6 paddsw xmm0, xmm2 - paddsw xmm0, krd - paddsw xmm4, xmm6 paddsw xmm0, xmm4 + paddsw xmm0, krd psraw xmm0, 7 packuswb xmm0, xmm0 @@ -250,10 +250,10 @@ pmaddubsw xmm4, k4k5 pmaddubsw xmm6, k6k7 + paddsw xmm0, xmm6 paddsw xmm0, xmm2 - paddsw xmm0, krd - paddsw xmm4, xmm6 paddsw xmm0, xmm4 + paddsw xmm0, krd psraw xmm0, 7 packuswb xmm0, xmm0 @@ -285,10 +285,10 @@ pmaddubsw xmm4, k4k5 pmaddubsw xmm6, k6k7 + paddsw xmm0, xmm6 paddsw xmm0, xmm2 - paddsw xmm4, xmm6 - paddsw xmm0, krd paddsw xmm0, xmm4 + paddsw xmm0, krd psraw xmm0, 7 packuswb xmm0, xmm0