Refactor calculate_scaling_weights() so that it is possible to get weights without...

diff --git a/resample_effect.cpp b/resample_effect.cpp
index b6da2e2aeea3be8f5605d6ec10525a37c082b152..802a8968a893482460988b910f2db707d68b692b 100644 (file)
--- a/resample_effect.cpp
+++ b/resample_effect.cpp
@@ -502,7 +502,7 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
                assert(false);
        }
 
-       ScalingWeights weights = calculate_scaling_weights(src_size, dst_size, zoom, offset);
+       ScalingWeights weights = calculate_bilinear_scaling_weights(src_size, dst_size, zoom, offset);
        src_bilinear_samples = weights.src_bilinear_samples;
        num_loops = weights.num_loops;
        slice_height = 1.0f / weights.num_loops;
@@ -529,6 +529,8 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str
        tex.update(weights.src_bilinear_samples, weights.dst_samples, internal_format, GL_RG, type, pixels);
 }
 
+namespace {
+
 ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset)
 {
        // Only needed if run from outside ResampleEffect.
@@ -627,16 +629,33 @@ ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, f
                }
        }
 
+       ScalingWeights ret;
+       ret.src_bilinear_samples = src_samples;
+       ret.dst_samples = dst_samples;
+       ret.num_loops = num_loops;
+       ret.bilinear_weights_fp16 = nullptr;
+       ret.bilinear_weights_fp32 = move(weights);
+       return ret;
+}
+
+}  // namespace
+
+ScalingWeights calculate_bilinear_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset)
+{
+       ScalingWeights ret = calculate_scaling_weights(src_size, dst_size, zoom, offset);
+       unique_ptr<Tap<float>[]> weights = move(ret.bilinear_weights_fp32);
+       const int src_samples = ret.src_bilinear_samples;
+
        // Now make use of the bilinear filtering in the GPU to reduce the number of samples
        // we need to make. Try fp16 first; if it's not accurate enough, we go to fp32.
        // Our tolerance level for total error is a bit higher than the one for invididual
        // samples, since one would assume overall errors in the shape don't matter as much.
        const float max_error = 2.0f / (255.0f * 255.0f);
        unique_ptr<Tap<fp16_int_t>[]> bilinear_weights_fp16;
-       int src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, dst_samples, &bilinear_weights_fp16);
+       int src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, ret.dst_samples, &bilinear_weights_fp16);
        unique_ptr<Tap<float>[]> bilinear_weights_fp32 = nullptr;
        double max_sum_sq_error_fp16 = 0.0;
-       for (unsigned y = 0; y < dst_samples; ++y) {
+       for (unsigned y = 0; y < ret.dst_samples; ++y) {
                double sum_sq_error_fp16 = compute_sum_sq_error(
                        weights.get() + y * src_samples, src_samples,
                        bilinear_weights_fp16.get() + y * src_bilinear_samples, src_bilinear_samples,
@@ -649,13 +668,10 @@ ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, f
 
        if (max_sum_sq_error_fp16 > max_error) {
                bilinear_weights_fp16.reset();
-               src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, dst_samples, &bilinear_weights_fp32);
+               src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, ret.dst_samples, &bilinear_weights_fp32);
        }
 
-       ScalingWeights ret;
        ret.src_bilinear_samples = src_bilinear_samples;
-       ret.dst_samples = dst_samples;
-       ret.num_loops = num_loops;
        ret.bilinear_weights_fp16 = move(bilinear_weights_fp16);
        ret.bilinear_weights_fp32 = move(bilinear_weights_fp32);
        return ret;

diff --git a/resample_effect.h b/resample_effect.h
index 027a497cc5fb5d6616ad43190117e4f91ed025d9..6a96ded56e41334637c667ca480194b7f26433f4 100644 (file)
--- a/resample_effect.h
+++ b/resample_effect.h
@@ -44,7 +44,7 @@ struct ScalingWeights {
        std::unique_ptr<Tap<fp16_int_t>[]> bilinear_weights_fp16;
        std::unique_ptr<Tap<float>[]> bilinear_weights_fp32;
 };
-ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset);
+ScalingWeights calculate_bilinear_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset);
 
 // A simple manager for support data stored in a 2D texture.
 // Consider moving it to a shared location of more classes

diff --git a/resample_effect_test.cpp b/resample_effect_test.cpp
index a51c51d664d54801f2ea8389a69fbc94b632c0a0..16bd70c552c5627a63caea619fb913822c3628f1 100644 (file)
--- a/resample_effect_test.cpp
+++ b/resample_effect_test.cpp
@@ -480,7 +480,7 @@ BENCHMARK_CAPTURE(BM_ResampleEffectHalf, Float16Upscale, GAMMA_LINEAR, "fragment
 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;
@@ -488,15 +488,15 @@ void BM_ComputeScalingWeights(benchmark::State &state)
        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