X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=resample_effect.cpp;h=aa3a4033ee8e5dec5b2f9826de529dfb98c22567;hp=c0afe0b1964ec06c2ab757883327c6a13570276a;hb=641053e9fc86b2166e361a983075febc3bb69acd;hpb=17195eb5da34f03778cde297eb77c486fa0c51bc diff --git a/resample_effect.cpp b/resample_effect.cpp index c0afe0b..aa3a403 100644 --- a/resample_effect.cpp +++ b/resample_effect.cpp @@ -1,16 +1,36 @@ // Three-lobed Lanczos, the most common choice. #define LANCZOS_RADIUS 3.0 -#include +#include #include +#include +#include +#include +#include +#include +#include +#include -#include "resample_effect.h" #include "effect_chain.h" +#include "effect_util.h" +#include "fp16.h" +#include "init.h" +#include "resample_effect.h" #include "util.h" -#include "opengl.h" + +using namespace Eigen; +using namespace std; + +namespace movit { namespace { +template +struct Tap { + T weight; + T pos; +}; + float sinc(float x) { if (fabs(x) < 1e-6) { @@ -40,14 +60,15 @@ 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(const 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) { @@ -60,30 +81,34 @@ 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 one level - // at 8-bit precision, don't combine. - if (sum_sq_error > 1.0f / (256.0f * 256.0f)) { + // 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, + // but since the artifacts are not really random, they can get quite + // visible. On the other hand, going to 0.25f, I can see no change at + // all with 8-bit output, so it would not seem to be worth it.) + if (sum_sq_error > 0.5f / (255.0f * 255.0f)) { continue; } // 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. @@ -92,20 +117,130 @@ unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsig return num_samples_saved; } +// 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. +template +void normalize_sum(Tap* vals, unsigned num) +{ + for (int normalize_pass = 0; normalize_pass < 2; ++normalize_pass) { + double sum = 0.0; + for (unsigned i = 0; i < num; ++i) { + sum += to_fp64(vals[i].weight); + } + for (unsigned i = 0; i < num; ++i) { + vals[i].weight = from_fp64(to_fp64(vals[i].weight) / sum); + } + } +} + +// Make use of the bilinear filtering in the GPU to reduce the number of samples +// we need to make. This is a bit more complex than BlurEffect since we cannot combine +// two neighboring samples if their weights have differing signs, so we first need to +// figure out the maximum number of samples. Then, we downconvert all the weights to +// that number -- we could have gone for a variable-length system, but this is simpler, +// and the gains would probably be offset by the extra cost of checking when to stop. +// +// 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) +{ + int src_bilinear_samples = 0; + for (unsigned y = 0; y < dst_samples; ++y) { + 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. + *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; + unsigned num_samples_saved = combine_samples( + 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); + normalize_sum(bilinear_weights_ptr, src_bilinear_samples); + } + return src_bilinear_samples; +} + +// Compute the sum of squared errors between the ideal weights (which are +// assumed to fall exactly on pixel centers) and the weights that result +// from sampling at . The primary reason for the difference +// is inaccuracy in the sampling positions, both due to limited precision +// in storing them (already inherent in sending them in as fp16_int_t) +// and in subtexel sampling precision (which we calculate in this function). +template +double compute_sum_sq_error(const Tap* weights, unsigned num_weights, + const Tap* bilinear_weights, unsigned num_bilinear_weights, + unsigned size) +{ + // 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)); + + float* effective_weights = new float[upper_pos - lower_pos]; + for (int i = 0; i < upper_pos - lower_pos; ++i) { + effective_weights[i] = 0.0f; + } + + // 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 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; + + assert(x0 >= 0); + assert(x1 >= 0); + 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; + } + + // Subtract the desired weights to get the error. + for (unsigned i = 0; i < num_weights; ++i) { + const int x = lrintf(weights[i].pos * size - 0.5f) - lower_pos; + assert(x >= 0); + assert(x < upper_pos - lower_pos); + + effective_weights[x] -= weights[i].weight; + } + + double sum_sq_error = 0.0; + for (unsigned i = 0; i < num_weights; ++i) { + sum_sq_error += effective_weights[i] * effective_weights[i]; + } + + delete[] effective_weights; + return sum_sq_error; +} + } // namespace ResampleEffect::ResampleEffect() : input_width(1280), - input_height(720) + input_height(720), + offset_x(0.0f), offset_y(0.0f), + zoom_x(1.0f), zoom_y(1.0f), + zoom_center_x(0.5f), zoom_center_y(0.5f) { register_int("width", &output_width); register_int("height", &output_height); // The first blur pass will forward resolution information to us. hpass = new SingleResamplePassEffect(this); - hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL); + CHECK(hpass->set_int("direction", SingleResamplePassEffect::HORIZONTAL)); vpass = new SingleResamplePassEffect(NULL); - vpass->set_int("direction", SingleResamplePassEffect::VERTICAL); + CHECK(vpass->set_int("direction", SingleResamplePassEffect::VERTICAL)); update_size(); } @@ -131,7 +266,7 @@ void ResampleEffect::inform_input_size(unsigned input_num, unsigned width, unsig input_height = height; update_size(); } - + void ResampleEffect::update_size() { bool ok = true; @@ -146,9 +281,29 @@ void ResampleEffect::update_size() ok |= vpass->set_int("output_height", output_height); assert(ok); + + // The offset added due to zoom may have changed with the size. + update_offset_and_zoom(); } -bool ResampleEffect::set_float(const std::string &key, float value) { +void ResampleEffect::update_offset_and_zoom() +{ + bool ok = true; + + // Zoom from the right origin. (zoom_center is given in normalized coordinates, + // i.e. 0..1.) + float extra_offset_x = zoom_center_x * (1.0f - 1.0f / zoom_x) * input_width; + float extra_offset_y = (1.0f - zoom_center_y) * (1.0f - 1.0f / zoom_y) * input_height; + + ok |= hpass->set_float("offset", extra_offset_x + offset_x); + ok |= vpass->set_float("offset", extra_offset_y - offset_y); // Compensate for the bottom-left origin. + ok |= hpass->set_float("zoom", zoom_x); + ok |= vpass->set_float("zoom", zoom_y); + + assert(ok); +} + +bool ResampleEffect::set_float(const string &key, float value) { if (key == "width") { output_width = value; update_size(); @@ -159,6 +314,42 @@ bool ResampleEffect::set_float(const std::string &key, float value) { update_size(); return true; } + if (key == "top") { + offset_y = value; + update_offset_and_zoom(); + return true; + } + if (key == "left") { + offset_x = value; + update_offset_and_zoom(); + return true; + } + if (key == "zoom_x") { + if (value <= 0.0f) { + return false; + } + zoom_x = value; + update_offset_and_zoom(); + return true; + } + if (key == "zoom_y") { + if (value <= 0.0f) { + return false; + } + zoom_y = value; + update_offset_and_zoom(); + return true; + } + if (key == "zoom_center_x") { + zoom_center_x = value; + update_offset_and_zoom(); + return true; + } + if (key == "zoom_center_y") { + zoom_center_y = value; + update_offset_and_zoom(); + return true; + } return false; } @@ -167,16 +358,22 @@ SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent) direction(HORIZONTAL), input_width(1280), input_height(720), + offset(0.0), + zoom(1.0), last_input_width(-1), last_input_height(-1), last_output_width(-1), - last_output_height(-1) + last_output_height(-1), + last_offset(0.0 / 0.0), // NaN. + last_zoom(0.0 / 0.0) // NaN. { register_int("direction", (int *)&direction); register_int("input_width", &input_width); register_int("input_height", &input_height); register_int("output_width", &output_width); register_int("output_height", &output_height); + register_float("offset", &offset); + register_float("zoom", &zoom); glGenTextures(1, &texnum); } @@ -186,7 +383,7 @@ SingleResamplePassEffect::~SingleResamplePassEffect() glDeleteTextures(1, &texnum); } -std::string SingleResamplePassEffect::output_fragment_shader() +string SingleResamplePassEffect::output_fragment_shader() { char buf[256]; sprintf(buf, "#define DIRECTION_VERTICAL %d\n", (direction == VERTICAL)); @@ -204,8 +401,8 @@ std::string SingleResamplePassEffect::output_fragment_shader() // so out[0] will read from parameters = <0,0>, <1,0>, <2,0> and so on. // // For horizontal scaling, we fill in the exact same texture; -// the shader just interprets is differently. -void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num) +// the shader just interprets it differently. +void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num) { unsigned src_size, dst_size; if (direction == SingleResamplePassEffect::HORIZONTAL) { @@ -220,13 +417,24 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std assert(false); } - // 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. - num_loops = gcd(src_size, dst_size); + float scaling_factor; + if (fabs(zoom - 1.0f) < 1e-6) { + num_loops = gcd(src_size, dst_size); + scaling_factor = float(dst_size) / float(src_size); + } else { + // If zooming is enabled (ie., zoom != 1), we turn off the looping. + // We _could_ perhaps do it for rational zoom levels (especially + // things like 2:1), but it doesn't seem to be worth it, given that + // the most common use case would seem to be varying the zoom + // from frame to frame. + 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; @@ -279,49 +487,52 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std // Anyhow, in this case we clearly need to look at more source pixels // to compute the destination pixel, and how many depend on the scaling factor. // Thus, the kernel width will vary with how much we scale. - float radius_scaling_factor = std::min(float(dst_size) / float(src_size), 1.0f); + 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) { // Find the point around which we want to sample the source image, // compensating for differing pixel centers as the scale changes. - float center_src_y = (y + 0.5f) * float(src_size) / float(dst_size) - 0.5f; + float center_src_y = (y + 0.5f) / scaling_factor - 0.5f; int base_src_y = lrintf(center_src_y); // Now sample pixels on each side around that point. 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), 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); + float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y - subpixel_offset), LANCZOS_RADIUS); + weights[y * src_samples + i].weight = weight * radius_scaling_factor; + weights[y * src_samples + i].pos = (src_y + 0.5) / float(src_size); } } // Now make use of the bilinear filtering in the GPU to reduce the number of samples - // we need to make. This is a bit more complex than BlurEffect since we cannot combine - // two neighboring samples if their weights have differing signs, so we first need to - // figure out the maximum number of samples. Then, we downconvert all the weights to - // that number -- we could have gone for a variable-length system, but this is simpler, - // and the gains would probably be offset by the extra cost of checking when to stop. - // - // The greedy strategy for combining samples is optimal. - src_bilinear_samples = 0; + // 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); + 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; + double max_sum_sq_error_fp16 = 0.0; for (unsigned y = 0; y < dst_samples; ++y) { - unsigned num_samples_saved = combine_samples(weights + (y * src_samples) * 2, NULL, src_samples, UINT_MAX); - src_bilinear_samples = std::max(src_bilinear_samples, src_samples - num_samples_saved); + 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, + 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) { + break; + } } - // Now that we know the right width, actually combine the samples. - float *bilinear_weights = new float[dst_samples * src_bilinear_samples * 2]; - for (unsigned y = 0; y < dst_samples; ++y) { - unsigned num_samples_saved = combine_samples( - weights + (y * src_samples) * 2, - bilinear_weights + (y * src_bilinear_samples) * 2, - src_samples, - src_samples - src_bilinear_samples); - assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples); - } + 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); + } // Encode as a two-component texture. Note the GL_REPEAT. glActiveTexture(GL_TEXTURE0 + *sampler_num); @@ -334,26 +545,40 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std 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_FLOAT, bilinear_weights); + if (fallback_to_fp32) { + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG32F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_FLOAT, bilinear_weights_fp32); + } else { + glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_HALF_FLOAT, bilinear_weights_fp16); + } check_error(); delete[] weights; - delete[] bilinear_weights; + delete[] bilinear_weights_fp16; + delete[] bilinear_weights_fp32; } -void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num) +void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num) { Effect::set_gl_state(glsl_program_num, prefix, sampler_num); + assert(input_width > 0); + assert(input_height > 0); + assert(output_width > 0); + assert(output_height > 0); + if (input_width != last_input_width || input_height != last_input_height || output_width != last_output_width || - output_height != last_output_height) { + output_height != last_output_height || + offset != last_offset || + zoom != last_zoom) { update_texture(glsl_program_num, prefix, sampler_num); last_input_width = input_width; last_input_height = input_height; last_output_width = output_width; last_output_height = output_height; + last_offset = offset; + last_zoom = zoom; } glActiveTexture(GL_TEXTURE0 + *sampler_num); @@ -362,7 +587,7 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const std:: 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); @@ -371,10 +596,21 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const std:: set_uniform_float(glsl_program_num, prefix, "sample_x_scale", 1.0f / src_bilinear_samples); set_uniform_float(glsl_program_num, prefix, "sample_x_offset", 0.5f / src_bilinear_samples); + float whole_pixel_offset; + if (direction == SingleResamplePassEffect::VERTICAL) { + whole_pixel_offset = lrintf(offset) / float(input_height); + } else { + whole_pixel_offset = lrintf(offset) / float(input_width); + } + set_uniform_float(glsl_program_num, prefix, "whole_pixel_offset", whole_pixel_offset); + // We specifically do not want mipmaps on the input texture; // they break minification. - glActiveTexture(GL_TEXTURE0); + 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(); } + +} // namespace movit