X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=resample_effect.cpp;h=9838cd429f0129b3e8b6bb947bfa3c34d0169e78;hp=792d86134ac54f875e5724bd63f715b7434d916c;hb=185ced44b129739c9b5438da691e71d664d6443a;hpb=27d066774adf38a774934c09746d638196d7d392 diff --git a/resample_effect.cpp b/resample_effect.cpp index 792d861..9838cd4 100644 --- a/resample_effect.cpp +++ b/resample_effect.cpp @@ -20,6 +20,12 @@ namespace movit { namespace { +template +struct Tap { + T weight; + T pos; +}; + float sinc(float x) { if (fabs(x) < 1e-6) { @@ -49,14 +55,15 @@ unsigned gcd(unsigned a, unsigned b) return a; } -unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsigned max_samples_saved) +template +unsigned combine_samples(Tap *src, Tap *dst, unsigned src_size, unsigned num_src_samples, unsigned max_samples_saved) { unsigned num_samples_saved = 0; for (unsigned i = 0, j = 0; i < num_src_samples; ++i, ++j) { // Copy the sample directly; it will be overwritten later if we can combine. if (dst != NULL) { - dst[j * 2 + 0] = src[i * 2 + 0]; - dst[j * 2 + 1] = src[i * 2 + 1]; + dst[j].weight = convert_float(src[i].weight); + dst[j].pos = convert_float(src[i].pos); } if (i == num_src_samples - 1) { @@ -69,19 +76,20 @@ unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsig continue; } - float w1 = src[i * 2 + 0]; - float w2 = src[(i + 1) * 2 + 0]; + float w1 = src[i].weight; + float w2 = src[i + 1].weight; if (w1 * w2 < 0.0f) { // Differing signs; cannot combine. continue; } - float pos1 = src[i * 2 + 1]; - float pos2 = src[(i + 1) * 2 + 1]; + float pos1 = src[i].pos; + float pos2 = src[i + 1].pos; assert(pos2 > pos1); - float offset, total_weight, sum_sq_error; - combine_two_samples(w1, w2, &offset, &total_weight, &sum_sq_error); + fp16_int_t pos, total_weight; + float sum_sq_error; + combine_two_samples(w1, w2, pos1, pos2, src_size, &pos, &total_weight, &sum_sq_error); // If the interpolation error is larger than that of about sqrt(2) of // a level at 8-bit precision, don't combine. (You'd think 1.0 was enough, @@ -94,8 +102,8 @@ unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsig // OK, we can combine this and the next sample. if (dst != NULL) { - dst[j * 2 + 0] = total_weight; - dst[j * 2 + 1] = pos1 + offset * (pos2 - pos1); + dst[j].weight = total_weight; + dst[j].pos = pos; } ++i; // Skip the next sample. @@ -370,7 +378,7 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str float radius_scaling_factor = min(scaling_factor, 1.0f); int int_radius = lrintf(LANCZOS_RADIUS / radius_scaling_factor); int src_samples = int_radius * 2 + 1; - float *weights = new float[dst_samples * src_samples * 2]; + Tap *weights = new Tap[dst_samples * src_samples]; float subpixel_offset = offset - lrintf(offset); // The part not covered by whole_pixel_offset. assert(subpixel_offset >= -0.5f && subpixel_offset <= 0.5f); for (unsigned y = 0; y < dst_samples; ++y) { @@ -383,8 +391,8 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str for (int i = 0; i < src_samples; ++i) { int src_y = base_src_y + i - int_radius; float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y - subpixel_offset), LANCZOS_RADIUS); - weights[(y * src_samples + i) * 2 + 0] = weight * radius_scaling_factor; - weights[(y * src_samples + i) * 2 + 1] = (src_y + 0.5) / float(src_size); + weights[y * src_samples + i].weight = weight * radius_scaling_factor; + weights[y * src_samples + i].pos = (src_y + 0.5) / float(src_size); } } @@ -398,39 +406,32 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str // The greedy strategy for combining samples is optimal. src_bilinear_samples = 0; for (unsigned y = 0; y < dst_samples; ++y) { - unsigned num_samples_saved = combine_samples(weights + (y * src_samples) * 2, NULL, src_samples, UINT_MAX); + unsigned num_samples_saved = combine_samples(weights + y * src_samples, NULL, src_size, src_samples, UINT_MAX); src_bilinear_samples = max(src_bilinear_samples, src_samples - num_samples_saved); } // Now that we know the right width, actually combine the samples. - float *bilinear_weights = new float[dst_samples * src_bilinear_samples * 2]; - fp16_int_t *bilinear_weights_fp16 = new fp16_int_t[dst_samples * src_bilinear_samples * 2]; + Tap *bilinear_weights = new Tap[dst_samples * src_bilinear_samples]; for (unsigned y = 0; y < dst_samples; ++y) { - float *bilinear_weights_ptr = bilinear_weights + (y * src_bilinear_samples) * 2; - fp16_int_t *bilinear_weights_fp16_ptr = bilinear_weights_fp16 + (y * src_bilinear_samples) * 2; + Tap *bilinear_weights_ptr = bilinear_weights + y * src_bilinear_samples; unsigned num_samples_saved = combine_samples( - weights + (y * src_samples) * 2, + weights + y * src_samples, bilinear_weights_ptr, + src_size, src_samples, src_samples - src_bilinear_samples); assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples); - // Convert to fp16. - for (int i = 0; i < src_bilinear_samples; ++i) { - bilinear_weights_fp16_ptr[i * 2 + 0] = fp64_to_fp16(bilinear_weights_ptr[i * 2 + 0]); - bilinear_weights_fp16_ptr[i * 2 + 1] = fp64_to_fp16(bilinear_weights_ptr[i * 2 + 1]); - } - // Normalize so that the sum becomes one. Note that we do it twice; // this sometimes helps a tiny little bit when we have many samples. for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) { double sum = 0.0; for (int i = 0; i < src_bilinear_samples; ++i) { - sum += fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]); + sum += fp16_to_fp64(bilinear_weights_ptr[i].weight); } for (int i = 0; i < src_bilinear_samples; ++i) { - bilinear_weights_fp16_ptr[i * 2 + 0] = fp64_to_fp16( - fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]) / sum); + bilinear_weights_ptr[i].weight = fp64_to_fp16( + fp16_to_fp64(bilinear_weights_ptr[i].weight) / sum); } } } @@ -446,12 +447,11 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str check_error(); glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT); check_error(); - glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_HALF_FLOAT, bilinear_weights_fp16); + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_HALF_FLOAT, bilinear_weights); check_error(); delete[] weights; delete[] bilinear_weights; - delete[] bilinear_weights_fp16; } void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num) @@ -484,7 +484,7 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const strin check_error(); set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num); - ++sampler_num; + ++*sampler_num; set_uniform_int(glsl_program_num, prefix, "num_samples", src_bilinear_samples); set_uniform_float(glsl_program_num, prefix, "num_loops", num_loops); set_uniform_float(glsl_program_num, prefix, "slice_height", slice_height);