X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=resample_effect.cpp;h=690f15a90e6622ed9a1c53194251491622eed855;hp=9c8caf3bb77ed58bf2d143dcc2b3b837b81a001c;hb=29888c096ef7aeea7d4d15ae40a5fece05b016ac;hpb=35ab97543afb74f72dd1d4c0d7d3206efe867a5e diff --git a/resample_effect.cpp b/resample_effect.cpp index 9c8caf3..690f15a 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 @@ -9,6 +9,7 @@ #include #include #include +#include #include #include #include @@ -27,12 +28,6 @@ namespace movit { namespace { -template -struct Tap { - T weight; - T pos; -}; - float sinc(float x) { if (fabs(x) < 1e-6) { @@ -78,7 +73,7 @@ float lanczos_weight(float x) // 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; +static once_flag lanczos_table_init_done; float lanczos_table[LANCZOS_TABLE_SIZE + 2]; void init_lanczos_table() @@ -86,7 +81,6 @@ void init_lanczos_table() for (unsigned i = 0; i < LANCZOS_TABLE_SIZE + 2; ++i) { lanczos_table[i] = lanczos_weight(float(i) * (LANCZOS_RADIUS / LANCZOS_TABLE_SIZE)); } - lanczos_table_init_done = true; } float lanczos_weight_cached(float x) @@ -96,7 +90,7 @@ float lanczos_weight_cached(float x) 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] + @@ -115,7 +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; @@ -135,7 +129,7 @@ unsigned combine_samples(const Tap *src, Tap *dst, float num_s 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) { + if (dst != nullptr) { dst[j].weight = convert_float(src[i].weight); dst[j].pos = convert_float(src[i].pos); } @@ -163,7 +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, @@ -175,7 +169,7 @@ unsigned combine_samples(const Tap *src, Tap *dst, float num_s } // OK, we can combine this and the next sample. - if (dst != NULL) { + if (dst != nullptr) { dst[j].weight = total_weight; dst[j].pos = pos; } @@ -212,29 +206,33 @@ void normalize_sum(Tap* vals, unsigned num) // // The greedy strategy for combining samples is optimal. template -unsigned combine_many_samples(const Tap *weights, unsigned src_size, unsigned src_samples, unsigned dst_samples, Tap **bilinear_weights) +unsigned combine_many_samples(const Tap *weights, unsigned src_size, unsigned src_samples, unsigned dst_samples, unique_ptr[]> *bilinear_weights) { 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, nullptr, 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; - *bilinear_weights = new Tap[dst_samples * src_bilinear_samples]; + bilinear_weights->reset(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; + Tap *bilinear_weights_ptr = bilinear_weights->get() + y * src_bilinear_samples; unsigned num_samples_saved = combine_samples( weights + y * src_samples, bilinear_weights_ptr, 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); } @@ -255,10 +253,10 @@ 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_fp32(bilinear_weights[0].pos) * size - 0.5)); - int upper_pos = int(ceil(to_fp32(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) { @@ -277,7 +275,7 @@ 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_fp32(bilinear_weights[i].weight) * (1.0 - f); + effective_weights[x0] += to_fp32(bilinear_weights[i].weight) * (1.0f - f); effective_weights[x1] += to_fp32(bilinear_weights[i].weight) * f; } @@ -312,18 +310,24 @@ ResampleEffect::ResampleEffect() register_int("height", &output_height); // The first blur pass will forward resolution information to us. - hpass = new SingleResamplePassEffect(this); + hpass_owner.reset(new SingleResamplePassEffect(this)); + hpass = hpass_owner.get(); CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL)); - vpass = new SingleResamplePassEffect(NULL); + vpass_owner.reset(new SingleResamplePassEffect(this)); + vpass = vpass_owner.get(); CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL)); update_size(); } +ResampleEffect::~ResampleEffect() +{ +} + void ResampleEffect::rewrite_graph(EffectChain *graph, Node *self) { - Node *hpass_node = graph->add_node(hpass); - Node *vpass_node = graph->add_node(vpass); + Node *hpass_node = graph->add_node(hpass_owner.release()); + Node *vpass_node = graph->add_node(vpass_owner.release()); graph->connect_nodes(hpass_node, vpass_node); graph->replace_receiver(self, hpass_node); graph->replace_sender(self, vpass_node); @@ -431,8 +435,8 @@ bool ResampleEffect::set_float(const string &key, float value) { SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent) : parent(parent), direction(HORIZONTAL), - input_width(1280), - input_height(720), + input_width(1280), + input_height(720), offset(0.0), zoom(1.0), last_input_width(-1), @@ -440,8 +444,7 @@ SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent) last_output_width(-1), last_output_height(-1), last_offset(0.0 / 0.0), // NaN. - last_zoom(0.0 / 0.0), // NaN. - last_texture_width(-1), last_texture_height(-1) + last_zoom(0.0 / 0.0) // NaN. { register_int("direction", (int *)&direction); register_int("input_width", &input_width); @@ -458,18 +461,11 @@ SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent) register_uniform_float("sample_x_offset", &uniform_sample_x_offset); 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(); - } + call_once(lanczos_table_init_done, init_lanczos_table); } SingleResamplePassEffect::~SingleResamplePassEffect() { - glDeleteTextures(1, &texnum); } string SingleResamplePassEffect::output_fragment_shader() @@ -506,12 +502,45 @@ 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, tex.get_texnum()); + check_error(); + + GLenum type, internal_format; + void *pixels; + assert((weights.bilinear_weights_fp16 == nullptr) != (weights.bilinear_weights_fp32 == nullptr)); + if (weights.bilinear_weights_fp32 != nullptr) { + type = GL_FLOAT; + internal_format = GL_RG32F; + pixels = weights.bilinear_weights_fp32.get(); + } else { + type = GL_HALF_FLOAT; + internal_format = GL_RG16F; + pixels = weights.bilinear_weights_fp16.get(); + } + + tex.update(weights.src_bilinear_samples, weights.dst_samples, internal_format, GL_RG, type, pixels); +} + +ScalingWeights calculate_scaling_weights(unsigned src_size, unsigned dst_size, float zoom, float offset) +{ + // Only needed if run from outside ResampleEffect. + call_once(lanczos_table_init_done, 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); @@ -524,7 +553,6 @@ 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 @@ -579,7 +607,7 @@ 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; - Tap *weights = new Tap[dst_samples * src_samples]; + unique_ptr[]> 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) { @@ -589,11 +617,12 @@ 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; } } @@ -602,15 +631,14 @@ 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_fp32 = NULL; - bool fallback_to_fp32 = false; + unique_ptr[]> bilinear_weights_fp16; + int src_bilinear_samples = combine_many_samples(weights.get(), src_size, src_samples, dst_samples, &bilinear_weights_fp16); + unique_ptr[]> bilinear_weights_fp32 = nullptr; 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( - weights + y * src_samples, src_samples, - bilinear_weights_fp16 + y * src_bilinear_samples, src_bilinear_samples, + weights.get() + y * src_samples, src_samples, + bilinear_weights_fp16.get() + y * src_bilinear_samples, src_bilinear_samples, src_size); max_sum_sq_error_fp16 = std::max(max_sum_sq_error_fp16, sum_sq_error_fp16); if (max_sum_sq_error_fp16 > max_error) { @@ -619,54 +647,17 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str } if (max_sum_sq_error_fp16 > max_error) { - fallback_to_fp32 = true; - src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp32); + bilinear_weights_fp16.reset(); + src_bilinear_samples = combine_many_samples(weights.get(), 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 = move(bilinear_weights_fp16); + ret.bilinear_weights_fp32 = move(bilinear_weights_fp32); + return ret; } void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num) @@ -695,7 +686,7 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const strin glActiveTexture(GL_TEXTURE0 + *sampler_num); check_error(); - glBindTexture(GL_TEXTURE_2D, texnum); + glBindTexture(GL_TEXTURE_2D, tex.get_texnum()); check_error(); uniform_sample_tex = *sampler_num; @@ -725,4 +716,44 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const strin } } +Support2DTexture::Support2DTexture() +{ + glGenTextures(1, &texnum); + check_error(); + glBindTexture(GL_TEXTURE_2D, texnum); + check_error(); + 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(); +} + +Support2DTexture::~Support2DTexture() +{ + glDeleteTextures(1, &texnum); + check_error(); +} + +void Support2DTexture::update(GLint width, GLint height, GLenum internal_format, GLenum format, GLenum type, const GLvoid * data) +{ + glBindTexture(GL_TEXTURE_2D, texnum); + check_error(); + if (width == last_texture_width && + height == 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, width, height, format, type, data); + check_error(); + } else { + glTexImage2D(GL_TEXTURE_2D, 0, internal_format, width, height, 0, format, type, data); + check_error(); + last_texture_width = width; + last_texture_height = height; + last_texture_internal_format = internal_format; + } +} + } // namespace movit