X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=resample_effect_test.cpp;h=71354f209c4a2858553c3f00b939bce3fec1bb70;hp=9082b26b42bc3ebe490046747fd8deef5a004328;hb=037ee905bc05fcbb278ac3a0b1f73249efada9b2;hpb=42f0fd5ccbb3560a76d55f3e725416a5e0f93523 diff --git a/resample_effect_test.cpp b/resample_effect_test.cpp index 9082b26..71354f2 100644 --- a/resample_effect_test.cpp +++ b/resample_effect_test.cpp @@ -4,12 +4,18 @@ #include #include +#include + #include "effect_chain.h" #include "flat_input.h" +#include "fp16.h" #include "image_format.h" +#include "init.h" #include "resample_effect.h" #include "test_util.h" +using namespace std; + namespace movit { namespace { @@ -70,7 +76,7 @@ TEST(ResampleEffectTest, UpscaleByTwoGetsCorrectPixelCenters) { } } - EffectChainTester tester(NULL, size * 2, size * 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR); + EffectChainTester tester(nullptr, size * 2, size * 2, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR); ImageFormat format; format.color_space = COLORSPACE_sRGB; @@ -96,11 +102,11 @@ TEST(ResampleEffectTest, DownscaleByTwoGetsCorrectPixelCenters) { // the texel center right (everything is nicely symmetric). // The approximate magnitudes have been checked against ImageMagick. float expected_data[size * size] = { - 0.0046, -0.0068, -0.0611, -0.0068, 0.0047, - -0.0068, 0.0100, 0.0895, 0.0100, -0.0068, - -0.0603, 0.0892, 0.7993, 0.0895, -0.0611, - -0.0067, 0.0100, 0.0892, 0.0100, -0.0068, - 0.0045, -0.0067, -0.0603, -0.0068, 0.0046, + 0.0045, -0.0067, -0.0599, -0.0067, 0.0045, + -0.0067, 0.0100, 0.0892, 0.0100, -0.0067, + -0.0599, 0.0890, 0.7925, 0.0892, -0.0599, + -0.0067, 0.0100, 0.0890, 0.0100, -0.0067, + 0.0045, -0.0067, -0.0599, -0.0067, 0.0045, }; float data[size * size * 4], out_data[size * size]; @@ -112,7 +118,7 @@ TEST(ResampleEffectTest, DownscaleByTwoGetsCorrectPixelCenters) { } } - EffectChainTester tester(NULL, size, size, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR); + EffectChainTester tester(nullptr, size, size, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR); ImageFormat format; format.color_space = COLORSPACE_sRGB; @@ -142,7 +148,7 @@ TEST(ResampleEffectTest, UpscaleByThreeGetsCorrectPixelCenters) { }; float out_data[size * size * 9]; - EffectChainTester tester(NULL, size * 3, size * 3, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR); + EffectChainTester tester(nullptr, size * 3, size * 3, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR); ImageFormat format; format.color_space = COLORSPACE_sRGB; @@ -158,12 +164,17 @@ TEST(ResampleEffectTest, UpscaleByThreeGetsCorrectPixelCenters) { tester.run(out_data, GL_RED, COLORSPACE_sRGB, GAMMA_LINEAR); // We only bother checking that the middle pixel is still correct, - // and that symmetry holds. - EXPECT_FLOAT_EQ(1.0, out_data[7 * (size * 3) + 7]); + // and that symmetry holds. Note that the middle weight in practice + // becomes something like 0.99999 due to the normalization + // (some supposedly zero weights become 1e-6 or so), and then after + // squaring, the error compounds. Ironically, less texture precision + // here will give a more accurate result, since the weight can get + // rounded towards 1.0. + EXPECT_NEAR(1.0, out_data[7 * (size * 3) + 7], 1e-3); for (unsigned y = 0; y < size * 3; ++y) { for (unsigned x = 0; x < size * 3; ++x) { - EXPECT_FLOAT_EQ(out_data[y * (size * 3) + x], out_data[(size * 3 - y - 1) * (size * 3) + x]); - EXPECT_FLOAT_EQ(out_data[y * (size * 3) + x], out_data[y * (size * 3) + (size * 3 - x - 1)]); + EXPECT_NEAR(out_data[y * (size * 3) + x], out_data[(size * 3 - y - 1) * (size * 3) + x], 1e-6); + EXPECT_NEAR(out_data[y * (size * 3) + x], out_data[y * (size * 3) + (size * 3 - x - 1)], 1e-6); } } } @@ -187,7 +198,7 @@ TEST(ResampleEffectTest, HeavyResampleGetsSumRight) { } } - EffectChainTester tester(NULL, dwidth, dheight, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR, GL_RGBA32F); + EffectChainTester tester(nullptr, dwidth, dheight, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR, GL_RGBA32F); ImageFormat format; format.color_space = COLORSPACE_sRGB; @@ -205,7 +216,7 @@ TEST(ResampleEffectTest, HeavyResampleGetsSumRight) { // Require that we are within 10-bit accuracy. Note that this limit is for // one pass only, but the limit is tight enough that it should be good enough // for 10-bit accuracy even after two passes. - expect_equal(expected_data, out_data, dwidth, dheight, 0.1 / 1023.0); + expect_equal(expected_data, out_data, dwidth, dheight, 0.12 / 1023.0); } TEST(ResampleEffectTest, ReadWholePixelFromLeft) { @@ -317,7 +328,7 @@ TEST(ResampleEffectTest, ReadHalfPixelFromLeftAndScale) { }; float out_data[dst_width * 1]; - EffectChainTester tester(NULL, dst_width, 1, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR); + EffectChainTester tester(nullptr, dst_width, 1, FORMAT_GRAYSCALE, COLORSPACE_sRGB, GAMMA_LINEAR); ImageFormat format; format.color_space = COLORSPACE_sRGB; @@ -402,12 +413,12 @@ TEST(ResampleEffectTest, VerticalZoomFromTop) { } TEST(ResampleEffectTest, Precision) { - const int size = 2048; + const int size = 1920; // Difficult non-power-of-two size. const int offset = 5; // Deliberately put the data of interest very close to the right, // where texture coordinates are farther from 0 and thus less precise. - float data[size] = {0}; + float data[size * 2] = {0}; data[size - offset] = 1.0f; float expected_data[size * 2] = {0}; for (int x = 0; x < size * 2; ++x) { @@ -425,4 +436,68 @@ TEST(ResampleEffectTest, Precision) { expect_equal(expected_data, out_data, size, 1); } +#ifdef HAVE_BENCHMARK +template<> inline uint8_t from_fp32(float x) { return x; } + +template +void BM_ResampleEffect(benchmark::State &state, GammaCurve gamma_curve, GLenum output_format, const std::string &shader_type) +{ + DisableComputeShadersTemporarily disabler(shader_type == "fragment"); + if (disabler.should_skip(&state)) return; + + unsigned in_width = state.range(0), in_height = state.range(1); + unsigned out_width = state.range(2), out_height = state.range(3); + + unique_ptr data(new T[in_width * in_height * 4]); + unique_ptr out_data(new T[out_width * out_height * 4]); + + for (unsigned i = 0; i < in_width * in_height * 4; ++i) { + data[i] = from_fp32(rand() / (RAND_MAX + 1.0)); + } + + EffectChainTester tester(nullptr, out_width, out_height, FORMAT_BGRA_POSTMULTIPLIED_ALPHA, COLORSPACE_sRGB, gamma_curve, output_format); + tester.add_input(data.get(), FORMAT_BGRA_POSTMULTIPLIED_ALPHA, COLORSPACE_sRGB, gamma_curve, in_width, in_height); + Effect *resample_effect = tester.get_chain()->add_effect(new ResampleEffect()); + + ASSERT_TRUE(resample_effect->set_int("width", out_width)); + ASSERT_TRUE(resample_effect->set_int("height", out_height)); + + tester.benchmark(state, out_data.get(), GL_BGRA, COLORSPACE_sRGB, gamma_curve, OUTPUT_ALPHA_FORMAT_PREMULTIPLIED); +} + +void BM_ResampleEffectHalf(benchmark::State &state, GammaCurve gamma_curve, const std::string &shader_type) +{ + BM_ResampleEffect(state, gamma_curve, GL_RGBA16F, shader_type); +} + +void BM_ResampleEffectInt8(benchmark::State &state, GammaCurve gamma_curve, const std::string &shader_type) +{ + BM_ResampleEffect(state, gamma_curve, GL_RGBA8, shader_type); +} + +BENCHMARK_CAPTURE(BM_ResampleEffectInt8, Int8Upscale, GAMMA_REC_709, "fragment")->Args({640, 360, 1280, 720})->Args({320, 180, 1280, 720})->Args({321, 181, 1280, 720})->UseRealTime()->Unit(benchmark::kMicrosecond); +BENCHMARK_CAPTURE(BM_ResampleEffectHalf, Float16Upscale, GAMMA_LINEAR, "fragment")->Args({640, 360, 1280, 720})->Args({320, 180, 1280, 720})->Args({321, 181, 1280, 720})->UseRealTime()->Unit(benchmark::kMicrosecond); +BENCHMARK_CAPTURE(BM_ResampleEffectInt8, Int8Downscale, GAMMA_REC_709, "fragment")->Args({1280, 720, 640, 360})->Args({1280, 720, 320, 180})->Args({1280, 720, 321, 181})->UseRealTime()->Unit(benchmark::kMicrosecond); +BENCHMARK_CAPTURE(BM_ResampleEffectHalf, Float16Downscale, GAMMA_LINEAR, "fragment")->Args({1280, 720, 640, 360})->Args({1280, 720, 320, 180})->Args({1280, 720, 321, 181})->UseRealTime()->Unit(benchmark::kMicrosecond); + +void BM_ComputeScalingWeights(benchmark::State &state) +{ + constexpr unsigned src_size = 1280; + constexpr unsigned dst_size = 35; + int old_precision = movit_texel_subpixel_precision; + movit_texel_subpixel_precision = 64; // To get consistent results across GPUs; this is a CPU test. + + // One iteration warmup to make sure the Lanczos table is computed. + calculate_scaling_weights(src_size, dst_size, 0.999f, 0.0f); + + for (auto _ : state) { + ScalingWeights weights = calculate_scaling_weights(src_size, dst_size, 0.999f, 0.0f); + } + + movit_texel_subpixel_precision = old_precision; +} +BENCHMARK(BM_ComputeScalingWeights)->Unit(benchmark::kMicrosecond); + +#endif + } // namespace movit