X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=resample_effect.cpp;h=453a83832f7cb5d905a11b49951ac63a6f14cc12;hp=958ed25b9c035a9ca1ad59109294b9cc96ed97ff;hb=11f2afecc201d7c4c845f5782059860ce3209a28;hpb=6bf496a34a84fb14b1b3b6debfd42ba36568b879 diff --git a/resample_effect.cpp b/resample_effect.cpp index 958ed25..453a838 100644 --- a/resample_effect.cpp +++ b/resample_effect.cpp @@ -1,7 +1,7 @@ // Three-lobed Lanczos, the most common choice. // Note that if you change this, the accuracy for LANCZOS_TABLE_SIZE // needs to be recomputed. -#define LANCZOS_RADIUS 3.0 +#define LANCZOS_RADIUS 3.0f #include #include @@ -27,12 +27,6 @@ namespace movit { namespace { -template -struct Tap { - T weight; - T pos; -}; - float sinc(float x) { if (fabs(x) < 1e-6) { @@ -74,6 +68,9 @@ float lanczos_weight(float x) // // Solve for e = 1e-6 yields a step size of 0.0027, which to cover the range // 0..3 needs 1109 steps. We round up to the next power of two, just to be sure. +// +// You need to call lanczos_table_init_done before the first call to +// lanczos_weight_cached. #define LANCZOS_TABLE_SIZE 2048 bool lanczos_table_init_done = false; float lanczos_table[LANCZOS_TABLE_SIZE + 2]; @@ -88,17 +85,12 @@ void init_lanczos_table() float lanczos_weight_cached(float x) { - if (!lanczos_table_init_done) { - // Could in theory race between two threads if we are unlucky, - // but that is harmless, since they'll write the same data. - init_lanczos_table(); - } x = fabs(x); if (x > LANCZOS_RADIUS) { return 0.0f; } float table_pos = x * (LANCZOS_TABLE_SIZE / LANCZOS_RADIUS); - int table_pos_int = int(table_pos); // Truncate towards zero. + unsigned table_pos_int = int(table_pos); // Truncate towards zero. float table_pos_frac = table_pos - table_pos_int; assert(table_pos < LANCZOS_TABLE_SIZE + 2); return lanczos_table[table_pos_int] + @@ -117,7 +109,7 @@ unsigned gcd(unsigned a, unsigned b) } template -unsigned combine_samples(const Tap *src, Tap *dst, float num_subtexels, float inv_num_subtexels, unsigned num_src_samples, unsigned max_samples_saved) +unsigned combine_samples(const Tap *src, Tap *dst, float num_subtexels, float inv_num_subtexels, unsigned num_src_samples, unsigned max_samples_saved, float pos1_pos2_diff, float inv_pos1_pos2_diff) { // Cut off near-zero values at both sides. unsigned num_samples_saved = 0; @@ -165,7 +157,7 @@ unsigned combine_samples(const Tap *src, Tap *dst, float num_s DestFloat pos, total_weight; float sum_sq_error; - combine_two_samples(w1, w2, pos1, pos2, num_subtexels, inv_num_subtexels, &pos, &total_weight, &sum_sq_error); + combine_two_samples(w1, w2, pos1, pos1_pos2_diff, inv_pos1_pos2_diff, num_subtexels, inv_num_subtexels, &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, @@ -194,13 +186,13 @@ template void normalize_sum(Tap* vals, unsigned num) { for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) { - double sum = 0.0; + float sum = 0.0; for (unsigned i = 0; i < num; ++i) { - sum += to_fp64(vals[i].weight); + sum += to_fp32(vals[i].weight); } - double inv_sum = 1.0 / sum; + float inv_sum = 1.0 / sum; for (unsigned i = 0; i < num; ++i) { - vals[i].weight = from_fp64(to_fp64(vals[i].weight) * inv_sum); + vals[i].weight = from_fp32(to_fp32(vals[i].weight) * inv_sum); } } } @@ -218,15 +210,18 @@ unsigned combine_many_samples(const Tap *weights, unsigned src_size, unsi { float num_subtexels = src_size / movit_texel_subpixel_precision; float inv_num_subtexels = movit_texel_subpixel_precision / src_size; + float pos1_pos2_diff = 1.0f / src_size; + float inv_pos1_pos2_diff = src_size; unsigned max_samples_saved = UINT_MAX; for (unsigned y = 0; y < dst_samples && max_samples_saved > 0; ++y) { - unsigned num_samples_saved = combine_samples(weights + y * src_samples, NULL, num_subtexels, inv_num_subtexels, src_samples, max_samples_saved); + unsigned num_samples_saved = combine_samples(weights + y * src_samples, NULL, num_subtexels, inv_num_subtexels, src_samples, max_samples_saved, pos1_pos2_diff, inv_pos1_pos2_diff); max_samples_saved = min(max_samples_saved, num_samples_saved); } // Now that we know the right width, actually combine the samples. unsigned src_bilinear_samples = src_samples - max_samples_saved; + if (*bilinear_weights != NULL) delete[] *bilinear_weights; *bilinear_weights = new Tap[dst_samples * src_bilinear_samples]; for (unsigned y = 0; y < dst_samples; ++y) { Tap *bilinear_weights_ptr = *bilinear_weights + y * src_bilinear_samples; @@ -236,7 +231,9 @@ unsigned combine_many_samples(const Tap *weights, unsigned src_size, unsi num_subtexels, inv_num_subtexels, src_samples, - max_samples_saved); + max_samples_saved, + pos1_pos2_diff, + inv_pos1_pos2_diff); assert(num_samples_saved == max_samples_saved); normalize_sum(bilinear_weights_ptr, src_bilinear_samples); } @@ -257,10 +254,10 @@ double compute_sum_sq_error(const Tap* weights, unsigned num_weights, // Find the effective range of the bilinear-optimized kernel. // Due to rounding of the positions, this is not necessarily the same // as the intended range (ie., the range of the original weights). - int lower_pos = int(floor(to_fp64(bilinear_weights[0].pos) * size - 0.5)); - int upper_pos = int(ceil(to_fp64(bilinear_weights[num_bilinear_weights - 1].pos) * size - 0.5)) + 2; - lower_pos = min(lower_pos, lrintf(weights[0].pos * size - 0.5)); - upper_pos = max(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5) + 1); + int lower_pos = int(floor(to_fp32(bilinear_weights[0].pos) * size - 0.5f)); + int upper_pos = int(ceil(to_fp32(bilinear_weights[num_bilinear_weights - 1].pos) * size - 0.5f)) + 2; + lower_pos = min(lower_pos, lrintf(weights[0].pos * size - 0.5f)); + upper_pos = max(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5f) + 1); float* effective_weights = new float[upper_pos - lower_pos]; for (int i = 0; i < upper_pos - lower_pos; ++i) { @@ -269,7 +266,7 @@ double compute_sum_sq_error(const Tap* weights, unsigned num_weights, // Now find the effective weights that result from this sampling. for (unsigned i = 0; i < num_bilinear_weights; ++i) { - const float pixel_pos = to_fp64(bilinear_weights[i].pos) * size - 0.5f; + const float pixel_pos = to_fp32(bilinear_weights[i].pos) * size - 0.5f; const int x0 = int(floor(pixel_pos)) - lower_pos; const int x1 = x0 + 1; const float f = lrintf((pixel_pos - (x0 + lower_pos)) / movit_texel_subpixel_precision) * movit_texel_subpixel_precision; @@ -279,8 +276,8 @@ double compute_sum_sq_error(const Tap* weights, unsigned num_weights, assert(x0 < upper_pos - lower_pos); assert(x1 < upper_pos - lower_pos); - effective_weights[x0] += to_fp64(bilinear_weights[i].weight) * (1.0 - f); - effective_weights[x1] += to_fp64(bilinear_weights[i].weight) * f; + effective_weights[x0] += to_fp32(bilinear_weights[i].weight) * (1.0f - f); + effective_weights[x1] += to_fp32(bilinear_weights[i].weight) * f; } // Subtract the desired weights to get the error. @@ -304,7 +301,8 @@ double compute_sum_sq_error(const Tap* weights, unsigned num_weights, } // namespace ResampleEffect::ResampleEffect() - : input_width(1280), + : owns_effects(true), + input_width(1280), input_height(720), offset_x(0.0f), offset_y(0.0f), zoom_x(1.0f), zoom_y(1.0f), @@ -322,6 +320,14 @@ ResampleEffect::ResampleEffect() update_size(); } +ResampleEffect::~ResampleEffect() +{ + if (owns_effects) { + delete hpass; + delete vpass; + } +} + void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self) { Node *hpass_node = graph->add_node(hpass); @@ -330,6 +336,7 @@ void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self) graph->replace_receiver(self, hpass_node); graph->replace_sender(self, vpass_node); self->disabled = true; + owns_effects = false; } // We get this information forwarded from the first blur pass, @@ -461,6 +468,12 @@ SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent) register_uniform_float("whole_pixel_offset", &uniform_whole_pixel_offset); glGenTextures(1, &texnum); + + if (!lanczos_table_init_done) { + // Could in theory race between two threads if we are unlucky, + // but that is harmless, since they'll write the same data. + init_lanczos_table(); + } } SingleResamplePassEffect::~SingleResamplePassEffect() @@ -502,12 +515,71 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str assert(false); } + ScalingWeights weights = calculate_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; + + // Encode as a two-component texture. Note the GL_REPEAT. + glActiveTexture(GL_TEXTURE0 + *sampler_num); + check_error(); + glBindTexture(GL_TEXTURE_2D, texnum); + check_error(); + if (last_texture_width == -1) { + // Need to set this state the first time. + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + check_error(); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT); + check_error(); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT); + check_error(); + } + + GLenum type, internal_format; + void *pixels; + assert((weights.bilinear_weights_fp16 == NULL) != (weights.bilinear_weights_fp32 == NULL)); + if (weights.bilinear_weights_fp32 != NULL) { + type = GL_FLOAT; + internal_format = GL_RG32F; + pixels = weights.bilinear_weights_fp32; + } else { + type = GL_HALF_FLOAT; + internal_format = GL_RG16F; + pixels = weights.bilinear_weights_fp16; + } + + if (int(weights.src_bilinear_samples) == last_texture_width && + int(weights.dst_samples) == last_texture_height && + internal_format == last_texture_internal_format) { + // Texture dimensions and type are unchanged; it is more efficient + // to just update it rather than making an entirely new texture. + glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, weights.src_bilinear_samples, weights.dst_samples, GL_RG, type, pixels); + } else { + glTexImage2D(GL_TEXTURE_2D, 0, internal_format, weights.src_bilinear_samples, weights.dst_samples, 0, GL_RG, type, pixels); + last_texture_width = weights.src_bilinear_samples; + last_texture_height = weights.dst_samples; + last_texture_internal_format = internal_format; + } + check_error(); + + delete[] weights.bilinear_weights_fp16; + delete[] weights.bilinear_weights_fp32; +} + +ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset) +{ + if (!lanczos_table_init_done) { + // Only needed if run from outside ResampleEffect. + init_lanczos_table(); + } + // For many resamplings (e.g. 640 -> 1280), we will end up with the same // set of samples over and over again in a loop. Thus, we can compute only // the first such loop, and then ask the card to repeat the texture for us. // This is both easier on the texture cache and lowers our CPU cost for // generating the kernel somewhat. float scaling_factor; + int num_loops; if (fabs(zoom - 1.0f) < 1e-6) { num_loops = gcd(src_size, dst_size); scaling_factor = float(dst_size) / float(src_size); @@ -520,7 +592,6 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str num_loops = 1; scaling_factor = zoom * float(dst_size) / float(src_size); } - slice_height = 1.0f / num_loops; unsigned dst_samples = dst_size / num_loops; // Sample the kernel in the right place. A diagram with a triangular kernel @@ -585,11 +656,12 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str int base_src_y = lrintf(center_src_y); // Now sample pixels on each side around that point. + float inv_src_size = 1.0 / float(src_size); for (int i = 0; i < src_samples; ++i) { int src_y = base_src_y + i - int_radius; float weight = lanczos_weight_cached(radius_scaling_factor * (src_y - center_src_y - subpixel_offset)); weights[y * src_samples + i].weight = weight * radius_scaling_factor; - weights[y * src_samples + i].pos = (src_y + 0.5) / float(src_size); + weights[y * src_samples + i].pos = (src_y + 0.5f) * inv_src_size; } } @@ -598,10 +670,9 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str // 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); - Tap *bilinear_weights_fp16; - src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp16); + Tap *bilinear_weights_fp16 = NULL; + int src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp16); Tap *bilinear_weights_fp32 = NULL; - bool fallback_to_fp32 = false; double max_sum_sq_error_fp16 = 0.0; for (unsigned y = 0; y < dst_samples; ++y) { double sum_sq_error_fp16 = compute_sum_sq_error( @@ -615,54 +686,20 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str } if (max_sum_sq_error_fp16 > max_error) { - fallback_to_fp32 = true; + delete[] bilinear_weights_fp16; + bilinear_weights_fp16 = NULL; src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp32); } - // Encode as a two-component texture. Note the GL_REPEAT. - glActiveTexture(GL_TEXTURE0 + *sampler_num); - check_error(); - glBindTexture(GL_TEXTURE_2D, texnum); - check_error(); - if (last_texture_width == -1) { - // Need to set this state the first time. - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - check_error(); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT); - check_error(); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT); - check_error(); - } - - GLenum type, internal_format; - void *pixels; - if (fallback_to_fp32) { - type = GL_FLOAT; - internal_format = GL_RG32F; - pixels = bilinear_weights_fp32; - } else { - type = GL_HALF_FLOAT; - internal_format = GL_RG16F; - pixels = bilinear_weights_fp16; - } - - if (int(src_bilinear_samples) == last_texture_width && - int(dst_samples) == last_texture_height && - internal_format == last_texture_internal_format) { - // Texture dimensions and type are unchanged; it is more efficient - // to just update it rather than making an entirely new texture. - glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, src_bilinear_samples, dst_samples, GL_RG, type, pixels); - } else { - glTexImage2D(GL_TEXTURE_2D, 0, internal_format, src_bilinear_samples, dst_samples, 0, GL_RG, type, pixels); - last_texture_width = src_bilinear_samples; - last_texture_height = dst_samples; - last_texture_internal_format = internal_format; - } - check_error(); - delete[] weights; - delete[] bilinear_weights_fp16; - delete[] 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 = bilinear_weights_fp16; + ret.bilinear_weights_fp32 = bilinear_weights_fp32; + return ret; } void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num) @@ -713,10 +750,12 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const strin // We specifically do not want mipmaps on the input texture; // they break minification. Node *self = chain->find_node_for_effect(this); - glActiveTexture(chain->get_input_sampler(self, 0)); - check_error(); - glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); - check_error(); + if (chain->has_input_sampler(self, 0)) { + glActiveTexture(chain->get_input_sampler(self, 0)); + check_error(); + glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR); + check_error(); + } } } // namespace movit