#include <math.h>
#include <stdio.h>
#include <algorithm>
+#include <Eigen/Sparse>
+#include <Eigen/SparseQR>
+#include <Eigen/OrderingMethods>
#include "effect_chain.h"
#include "effect_util.h"
#include "resample_effect.h"
#include "util.h"
+using namespace Eigen;
using namespace std;
namespace movit {
}
template<class DestFloat>
-unsigned combine_samples(const Tap<float> *src, Tap<DestFloat> *dst, unsigned src_size, unsigned num_src_samples, unsigned max_samples_saved)
+unsigned combine_samples(const Tap<float> *src, Tap<DestFloat> *dst, float num_subtexels, float inv_num_subtexels, 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) {
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);
+ combine_two_samples(w1, w2, pos1, pos2, 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,
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<class T>
+void normalize_sum(Tap<T>* 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<T>(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
template<class DestFloat>
unsigned combine_many_samples(const Tap<float> *weights, unsigned src_size, unsigned src_samples, unsigned dst_samples, Tap<DestFloat> **bilinear_weights)
{
- int src_bilinear_samples = 0;
- for (unsigned y = 0; y < dst_samples; ++y) {
- unsigned num_samples_saved = combine_samples<DestFloat>(weights + y * src_samples, NULL, src_size, src_samples, UINT_MAX);
- src_bilinear_samples = max<int>(src_bilinear_samples, src_samples - num_samples_saved);
+ float num_subtexels = src_size / movit_texel_subpixel_precision;
+ float inv_num_subtexels = movit_texel_subpixel_precision / 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<DestFloat>(weights + y * src_samples, NULL, num_subtexels, inv_num_subtexels, src_samples, max_samples_saved);
+ 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<DestFloat>[dst_samples * src_bilinear_samples];
for (unsigned y = 0; y < dst_samples; ++y) {
Tap<DestFloat> *bilinear_weights_ptr = *bilinear_weights + y * src_bilinear_samples;
unsigned num_samples_saved = combine_samples(
weights + y * src_samples,
bilinear_weights_ptr,
- src_size,
+ num_subtexels,
+ inv_num_subtexels,
src_samples,
- src_samples - src_bilinear_samples);
- assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
-
- // 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 += to_fp64(bilinear_weights_ptr[i].weight);
- }
- for (int i = 0; i < src_bilinear_samples; ++i) {
- bilinear_weights_ptr[i].weight = from_fp64<DestFloat>(
- to_fp64(bilinear_weights_ptr[i].weight) / sum);
- }
- }
+ max_samples_saved);
+ assert(num_samples_saved == max_samples_saved);
+ normalize_sum(bilinear_weights_ptr, src_bilinear_samples);
}
return src_bilinear_samples;
}
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<int>(lower_pos, lrintf(weights[0].pos * size - 0.5));
- upper_pos = max<int>(upper_pos, lrintf(weights[num_weights - 1].pos * size - 0.5));
+ upper_pos = max<int>(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) {
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);
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);
}
// 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<fp16_int_t> *bilinear_weights_fp16;
src_bilinear_samples = combine_many_samples(weights, src_size, src_samples, dst_samples, &bilinear_weights_fp16);
Tap<float> *bilinear_weights_fp32 = NULL;
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);
}
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();
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.