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_NEAR(out_data[y * (size * 3) + x], out_data[(size * 3 - y - 1) * (size * 3) + x], 1e-6);
}
#ifdef HAVE_BENCHMARK
-template<> inline uint8_t from_fp32<uint8_t>(float x) { return x; }
+template<> inline uint8_t from_fp32<uint8_t>(float x) { return lrintf(x * 255.0f); }
template<class T>
void BM_ResampleEffect(benchmark::State &state, GammaCurve gamma_curve, GLenum output_format, const std::string &shader_type)
unique_ptr<T[]> out_data(new T[out_width * out_height * 4]);
for (unsigned i = 0; i < in_width * in_height * 4; ++i) {
- data[i] = from_fp32<T>(float(rand()));
+ data[i] = from_fp32<T>(rand() / (RAND_MAX + 1.0));
}
EffectChainTester tester(nullptr, out_width, out_height, FORMAT_BGRA_POSTMULTIPLIED_ALPHA, COLORSPACE_sRGB, gamma_curve, output_format);
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)
+void BM_ComputeBilinearScalingWeights(benchmark::State &state)
{
constexpr unsigned src_size = 1280;
constexpr unsigned dst_size = 35;
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);
+ calculate_bilinear_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);
+ ScalingWeights weights = calculate_bilinear_scaling_weights(src_size, dst_size, 0.999f, 0.0f);
}
movit_texel_subpixel_precision = old_precision;
}
-BENCHMARK(BM_ComputeScalingWeights)->Unit(benchmark::kMicrosecond);
+BENCHMARK(BM_ComputeBilinearScalingWeights)->Unit(benchmark::kMicrosecond);
#endif