X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=resample_effect.cpp;h=4dd3ea87e261086939a91e2d26d51ef52c813168;hp=f438a873b095114554bb1b53a41cebe263d16f4b;hb=5e34f7a8969f4afc169f034d34fb908019b3a389;hpb=0af958592e20bde3721a6deb16a6e32edfeb6cdc diff --git a/resample_effect.cpp b/resample_effect.cpp index f438a87..4dd3ea8 100644 --- a/resample_effect.cpp +++ b/resample_effect.cpp @@ -63,7 +63,22 @@ unsigned gcd(unsigned a, unsigned b) template unsigned combine_samples(const Tap *src, Tap *dst, unsigned src_size, unsigned num_src_samples, unsigned max_samples_saved) { + // Cut off near-zero values at both sides. unsigned num_samples_saved = 0; + while (num_samples_saved < max_samples_saved && + num_src_samples > 0 && + fabs(src[0].weight) < 1e-6) { + ++src; + --num_src_samples; + ++num_samples_saved; + } + while (num_samples_saved < max_samples_saved && + num_src_samples > 0 && + fabs(src[num_src_samples - 1].weight) < 1e-6) { + --num_src_samples; + ++num_samples_saved; + } + 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) { @@ -183,7 +198,7 @@ double compute_sum_sq_error(const Tap* weights, unsigned num_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)); + upper_pos = max(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5) + 1); float* effective_weights = new float[upper_pos - lower_pos]; for (int i = 0; i < upper_pos - lower_pos; ++i) { @@ -224,112 +239,6 @@ double compute_sum_sq_error(const Tap* weights, unsigned num_weights, return sum_sq_error; } -// Given a predefined, fixed set of bilinear weight positions, try to optimize -// their weights through some linear algebra. This can do a better job than -// the weight calculation in combine_samples() because it can look at the entire -// picture (an effective weight can sometimes be affected by multiple samples). -// It will also optimize weights for non-combined samples, which is useful when -// a sample happens in-between texels for numerical reasons. -// -// The math goes as follows: The desired result is a weighted sum, where the -// weights are the coefficients in : -// -// y = sum(c_j x_j, j) -// -// We try to approximate this by a different set of coefficients, which have -// weights d_i and are placed at some fraction to the right of a source texel x_j. -// This means it will influence two texels (x_j and x_{j+1}); generalizing this, -// let us define that w_ij means the amount texel influences bilinear weight -// (keeping in mind that w_ij = 0 for all but at most two different j). -// This means the actually computed result is: -// -// y' = sum(d_i w_ij x_j, j) -// -// We assume w_ij fixed and wish to find {d_i} so that y' gets as close to y -// as possible. Specifically, let us consider the sum of squred errors of the -// coefficients: -// -// ε² = sum((sum( d_i w_ij, i ) - c_j)², j) -// -// The standard trick, which also applies just fine here, is to differentiate -// the error with respect to each variable we wish to optimize, and set each -// such expression to zero. Solving this equation set (which we can do efficiently -// by letting Eigen invert a sparse matrix for us) yields the minimum possible -// error. To see the form each such equation takes, pick any value k and -// differentiate the expression by d_k: -// -// ∂(ε²)/∂(d_k) = sum(2(sum( d_i w_ij, i ) - c_j) w_kj, j) -// -// Setting this expression equal to zero, dropping the irrelevant factor 2 and -// rearranging yields: -// -// sum(w_kj sum( d_i w_ij, i ), j) = sum(w_kj c_j, j) -// -// where again, we remember where the sums over j are over at most two elements, -// since w_kj is nonzero for at most two values of j. -template -void optimize_sum_sq_error(const Tap* weights, unsigned num_weights, - Tap* bilinear_weights, unsigned num_bilinear_weights, - unsigned size) -{ - // Find the range of the desired weights. - int c_lower_pos = lrintf(weights[0].pos * size - 0.5); - int c_upper_pos = lrintf(weights[num_weights - 1].pos * size - 0.5) + 1; - - SparseMatrix A(num_bilinear_weights, num_bilinear_weights); - SparseVector b(num_bilinear_weights); - - // Convert each bilinear weight to the (x, frac) form for less junk in the code below. - int* pos = new int[num_bilinear_weights]; - float* fracs = new float[num_bilinear_weights]; - for (unsigned i = 0; i < num_bilinear_weights; ++i) { - const float pixel_pos = to_fp64(bilinear_weights[i].pos) * size - 0.5f; - const float f = pixel_pos - floor(pixel_pos); - pos[i] = int(floor(pixel_pos)); - fracs[i] = lrintf(f / movit_texel_subpixel_precision) * movit_texel_subpixel_precision; - } - - // The index ordering is a bit unusual to fit better with the - // notation in the derivation above. - for (unsigned k = 0; k < num_bilinear_weights; ++k) { - for (int j = pos[k]; j <= pos[k] + 1; ++j) { - const float w_kj = (j == pos[k]) ? (1.0f - fracs[k]) : fracs[k]; - for (unsigned i = 0; i < num_bilinear_weights; ++i) { - float w_ij; - if (j == pos[i]) { - w_ij = 1.0f - fracs[i]; - } else if (j == pos[i] + 1) { - w_ij = fracs[i]; - } else { - // w_ij = 0 - continue; - } - A.coeffRef(i, k) += w_kj * w_ij; - } - float c_j; - if (j >= c_lower_pos && j < c_upper_pos) { - c_j = weights[j - c_lower_pos].weight; - } else { - c_j = 0.0f; - } - b.coeffRef(k) += w_kj * c_j; - } - } - delete[] pos; - delete[] fracs; - - A.makeCompressed(); - SparseQR, COLAMDOrdering > qr(A); - assert(qr.info() == Success); - SparseMatrix new_weights = qr.solve(b); - assert(qr.info() == Success); - - for (unsigned i = 0; i < num_bilinear_weights; ++i) { - bilinear_weights[i].weight = from_fp64(new_weights.coeff(i, 0)); - } - normalize_sum(bilinear_weights, num_bilinear_weights); -} - } // namespace ResampleEffect::ResampleEffect() @@ -471,7 +380,8 @@ 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_zoom(0.0 / 0.0), // NaN. + last_texture_width(-1), last_texture_height(-1) { register_int("direction", (int *)&direction); register_int("input_width", &input_width); @@ -480,6 +390,13 @@ SingleResamplePassEffect::SingleResamplePassEffect(ResampleEffect *parent) register_int("output_height", &output_height); register_float("offset", &offset); register_float("zoom", &zoom); + register_uniform_sampler2d("sample_tex", &uniform_sample_tex); + register_uniform_int("num_samples", &uniform_num_samples); // FIXME: What about GLSL pre-1.30? + register_uniform_float("num_loops", &uniform_num_loops); + register_uniform_float("slice_height", &uniform_slice_height); + register_uniform_float("sample_x_scale", &uniform_sample_x_scale); + register_uniform_float("sample_x_offset", &uniform_sample_x_offset); + register_uniform_float("whole_pixel_offset", &uniform_whole_pixel_offset); glGenTextures(1, &texnum); } @@ -616,34 +533,28 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str // 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); 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) { - optimize_sum_sq_error( - weights + y * src_samples, src_samples, - bilinear_weights_fp16 + y * src_bilinear_samples, src_bilinear_samples, - src_size); 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; + } } - // 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. - if (max_sum_sq_error_fp16 > 2.0f / (255.0f * 255.0f)) { + 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); - for (unsigned y = 0; y < dst_samples; ++y) { - optimize_sum_sq_error( - weights + y * src_samples, src_samples, - bilinear_weights_fp32 + y * src_bilinear_samples, src_bilinear_samples, - src_size); - } } // Encode as a two-component texture. Note the GL_REPEAT. @@ -651,16 +562,39 @@ void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const str 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(); + 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) { - glTexImage2D(GL_TEXTURE_2D, 0, GL_RG32F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_FLOAT, bilinear_weights_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, GL_RG16F, src_bilinear_samples, dst_samples, 0, GL_RG, GL_HALF_FLOAT, bilinear_weights_fp16); + 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(); @@ -698,23 +632,21 @@ void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const strin glBindTexture(GL_TEXTURE_2D, texnum); check_error(); - set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num); + uniform_sample_tex = *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); + uniform_num_samples = src_bilinear_samples; + uniform_num_loops = num_loops; + uniform_slice_height = slice_height; // Instructions for how to convert integer sample numbers to positions in the weight texture. - 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); + uniform_sample_x_scale = 1.0f / src_bilinear_samples; + uniform_sample_x_offset = 0.5f / src_bilinear_samples; - float whole_pixel_offset; if (direction == SingleResamplePassEffect::VERTICAL) { - whole_pixel_offset = lrintf(offset) / float(input_height); + uniform_whole_pixel_offset = lrintf(offset) / float(input_height); } else { - whole_pixel_offset = lrintf(offset) / float(input_width); + uniform_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.