From fa540837aa23ae8cdd926961ddbef71998a2a525 Mon Sep 17 00:00:00 2001 From: chiyotsai Date: Fri, 25 Jan 2019 19:30:49 -0800 Subject: [PATCH] Fix mismatch between SIMD/C version of vp9_apply_temporal_filter Change-Id: I6503ebc79beaac2947992437ac133f3ac4379019 --- test/yuv_temporal_filter_test.cc | 117 +++++++++++++++++++++++++ vp9/encoder/vp9_temporal_filter.c | 3 +- vp9/encoder/x86/temporal_filter_sse4.c | 4 +- 3 files changed, 121 insertions(+), 3 deletions(-) diff --git a/test/yuv_temporal_filter_test.cc b/test/yuv_temporal_filter_test.cc index 2febe9e90..dcb5990c5 100644 --- a/test/yuv_temporal_filter_test.cc +++ b/test/yuv_temporal_filter_test.cc @@ -480,6 +480,123 @@ TEST_P(YUVTemporalFilterTest, Use16x16) { } } +TEST_P(YUVTemporalFilterTest, SaturationTest) { + const int width = 32, height = 32; + const int use_32x32 = 1; + + Buffer y_src = Buffer(width, height, 8); + Buffer y_pre = Buffer(width, height, 0); + Buffer y_count_ref = Buffer(width, height, 0); + Buffer y_accum_ref = Buffer(width, height, 0); + Buffer y_count_tst = Buffer(width, height, 0); + Buffer y_accum_tst = Buffer(width, height, 0); + ASSERT_TRUE(y_src.Init()); + ASSERT_TRUE(y_pre.Init()); + ASSERT_TRUE(y_count_ref.Init()); + ASSERT_TRUE(y_accum_ref.Init()); + ASSERT_TRUE(y_count_tst.Init()); + ASSERT_TRUE(y_accum_tst.Init()); + + for (int ss_x = 0; ss_x <= 1; ss_x++) { + for (int ss_y = 0; ss_y <= 1; ss_y++) { + for (int filter_strength = 0; filter_strength <= 6; + filter_strength += 2) { + for (int filter_weight = 0; filter_weight <= 2; filter_weight++) { + const int uv_width = width >> ss_x, uv_height = height >> ss_y; + Buffer u_src = Buffer(uv_width, uv_height, 8); + Buffer u_pre = Buffer(uv_width, uv_height, 0); + Buffer u_count_ref = + Buffer(uv_width, uv_height, 0); + Buffer u_accum_ref = + Buffer(uv_width, uv_height, 0); + Buffer u_count_tst = + Buffer(uv_width, uv_height, 0); + Buffer u_accum_tst = + Buffer(uv_width, uv_height, 0); + ASSERT_TRUE(u_src.Init()); + ASSERT_TRUE(u_pre.Init()); + ASSERT_TRUE(u_count_ref.Init()); + ASSERT_TRUE(u_accum_ref.Init()); + ASSERT_TRUE(u_count_tst.Init()); + ASSERT_TRUE(u_accum_tst.Init()); + Buffer v_src = Buffer(uv_width, uv_height, 8); + Buffer v_pre = Buffer(uv_width, uv_height, 0); + Buffer v_count_ref = + Buffer(uv_width, uv_height, 0); + Buffer v_accum_ref = + Buffer(uv_width, uv_height, 0); + Buffer v_count_tst = + Buffer(uv_width, uv_height, 0); + Buffer v_accum_tst = + Buffer(uv_width, uv_height, 0); + ASSERT_TRUE(v_src.Init()); + ASSERT_TRUE(v_pre.Init()); + ASSERT_TRUE(v_count_ref.Init()); + ASSERT_TRUE(v_accum_ref.Init()); + ASSERT_TRUE(v_count_tst.Init()); + ASSERT_TRUE(v_accum_tst.Init()); + + // The difference between the buffers must be small to pass the + // threshold to apply the filter. + y_src.Set(255); + y_pre.Set(0); + u_src.Set(255); + u_pre.Set(0); + v_src.Set(255); + v_pre.Set(0); + + y_accum_ref.Set(rnd_.Rand8()); + y_accum_tst.CopyFrom(y_accum_ref); + y_count_ref.Set(rnd_.Rand8()); + y_count_tst.CopyFrom(y_count_ref); + u_accum_ref.Set(rnd_.Rand8()); + u_accum_tst.CopyFrom(u_accum_ref); + u_count_ref.Set(rnd_.Rand8()); + u_count_tst.CopyFrom(u_count_ref); + v_accum_ref.Set(rnd_.Rand8()); + v_accum_tst.CopyFrom(v_accum_ref); + v_count_ref.Set(rnd_.Rand8()); + v_count_tst.CopyFrom(v_count_ref); + + ApplyReferenceFilter(y_src, y_pre, u_src, v_src, u_pre, v_pre, width, + height, ss_x, ss_y, filter_strength, + &filter_weight, use_32x32, &y_accum_ref, + &y_count_ref, &u_accum_ref, &u_count_ref, + &v_accum_ref, &v_count_ref); + ASM_REGISTER_STATE_CHECK(filter_func_( + y_src.TopLeftPixel(), y_src.stride(), y_pre.TopLeftPixel(), + y_pre.stride(), u_src.TopLeftPixel(), v_src.TopLeftPixel(), + u_src.stride(), u_pre.TopLeftPixel(), v_pre.TopLeftPixel(), + u_pre.stride(), width, height, ss_x, ss_y, filter_strength, + &filter_weight, use_32x32, y_accum_tst.TopLeftPixel(), + y_count_tst.TopLeftPixel(), u_accum_tst.TopLeftPixel(), + u_count_tst.TopLeftPixel(), v_accum_tst.TopLeftPixel(), + v_count_tst.TopLeftPixel())); + + EXPECT_TRUE(y_accum_tst.CheckValues(y_accum_ref)); + EXPECT_TRUE(y_count_tst.CheckValues(y_count_ref)); + EXPECT_TRUE(u_accum_tst.CheckValues(u_accum_ref)); + EXPECT_TRUE(u_count_tst.CheckValues(u_count_ref)); + EXPECT_TRUE(v_accum_tst.CheckValues(v_accum_ref)); + EXPECT_TRUE(v_count_tst.CheckValues(v_count_ref)); + + if (HasFailure()) { + printf("SS_X: %d, SS_Y: %d, Weight: %d, Strength: %d\n", ss_x, ss_y, + filter_weight, filter_strength); + y_accum_tst.PrintDifference(y_accum_ref); + y_count_tst.PrintDifference(y_count_ref); + u_accum_tst.PrintDifference(u_accum_ref); + u_count_tst.PrintDifference(u_count_ref); + v_accum_tst.PrintDifference(v_accum_ref); + v_count_tst.PrintDifference(v_count_ref); + return; + } + } + } + } + } +} + TEST_P(YUVTemporalFilterTest, DISABLED_Speed) { const int width = 32, height = 32; Buffer y_src = Buffer(width, height, 8); diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c index 04b589ac3..e10946aca 100644 --- a/vp9/encoder/vp9_temporal_filter.c +++ b/vp9/encoder/vp9_temporal_filter.c @@ -192,7 +192,8 @@ static INLINE int mod_index(int sum_dist, int index, int rounding, int strength, assert(index >= 0 && index <= 13); assert(index_mult[index] != 0); - mod = (clamp(sum_dist, 0, UINT16_MAX) * index_mult[index]) >> 16; + mod = + ((unsigned int)clamp(sum_dist, 0, UINT16_MAX) * index_mult[index]) >> 16; mod += rounding; mod >>= strength; diff --git a/vp9/encoder/x86/temporal_filter_sse4.c b/vp9/encoder/x86/temporal_filter_sse4.c index 18c4f02d5..b560e2218 100644 --- a/vp9/encoder/x86/temporal_filter_sse4.c +++ b/vp9/encoder/x86/temporal_filter_sse4.c @@ -967,8 +967,8 @@ static void vp9_apply_temporal_filter_chroma_8( v_sum_row_2 = v_sum_row_3; // Add chroma values - u_sum_row = _mm_adds_epu8(u_sum_row_1, u_sum_row_2); - v_sum_row = _mm_adds_epu8(v_sum_row_1, v_sum_row_2); + u_sum_row = _mm_adds_epu16(u_sum_row_1, u_sum_row_2); + v_sum_row = _mm_adds_epu16(v_sum_row_1, v_sum_row_2); // Add luma values add_luma_dist_to_8_chroma_mod(y_dist, ss_x, ss_y, &u_sum_row, &v_sum_row); -- 2.40.0