// Three-lobed Lanczos, the most common choice.
#define LANCZOS_RADIUS 3.0
-#include <GL/glew.h>
+#include <epoxy/gl.h>
#include <assert.h>
#include <limits.h>
#include <math.h>
#include "effect_chain.h"
#include "effect_util.h"
+#include "fp16.h"
#include "resample_effect.h"
#include "util.h"
input_height = height;
update_size();
}
-
+
void ResampleEffect::update_size()
{
bool ok = true;
update_size();
return true;
}
+ if (key == "top") {
+ // Compensate for the bottom-left origin.
+ return vpass->set_float("offset", -value);
+ }
+ if (key == "left") {
+ return hpass->set_float("offset", value);
+ }
return false;
}
direction(HORIZONTAL),
input_width(1280),
input_height(720),
+ offset(0.0),
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.
{
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);
glGenTextures(1, &texnum);
}
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];
+ 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.
// Now sample <int_radius> 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);
+ float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y - subpixel_offset), 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);
}
-
}
// Now make use of the bilinear filtering in the GPU to reduce the number of samples
// Now that we know the right width, actually combine the samples.
float *bilinear_weights = new float[dst_samples * src_bilinear_samples * 2];
+ fp16_int_t *bilinear_weights_fp16 = new fp16_int_t[dst_samples * src_bilinear_samples * 2];
for (unsigned y = 0; y < dst_samples; ++y) {
+ float *bilinear_weights_ptr = bilinear_weights + (y * src_bilinear_samples) * 2;
+ fp16_int_t *bilinear_weights_fp16_ptr = bilinear_weights_fp16 + (y * src_bilinear_samples) * 2;
unsigned num_samples_saved = combine_samples(
weights + (y * src_samples) * 2,
- bilinear_weights + (y * src_bilinear_samples) * 2,
+ bilinear_weights_ptr,
src_samples,
src_samples - src_bilinear_samples);
assert(int(src_samples) - int(num_samples_saved) == src_bilinear_samples);
+ // Convert to fp16.
+ for (int i = 0; i < src_bilinear_samples; ++i) {
+ bilinear_weights_fp16_ptr[i * 2 + 0] = fp64_to_fp16(bilinear_weights_ptr[i * 2 + 0]);
+ bilinear_weights_fp16_ptr[i * 2 + 1] = fp64_to_fp16(bilinear_weights_ptr[i * 2 + 1]);
+ }
+
// 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) {
- float sum = 0.0;
+ double sum = 0.0;
for (int i = 0; i < src_bilinear_samples; ++i) {
- sum += bilinear_weights[(y * src_bilinear_samples + i) * 2 + 0];
+ sum += fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]);
}
for (int i = 0; i < src_bilinear_samples; ++i) {
- bilinear_weights[(y * src_bilinear_samples + i) * 2 + 0] /= sum;
+ bilinear_weights_fp16_ptr[i * 2 + 0] = fp64_to_fp16(
+ fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]) / sum);
}
}
- }
+ }
// Encode as a two-component texture. Note the GL_REPEAT.
glActiveTexture(GL_TEXTURE0 + *sampler_num);
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);
+ 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;
}
void SingleResamplePassEffect::set_gl_state(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
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) {
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;
}
glActiveTexture(GL_TEXTURE0 + *sampler_num);
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();
projects / movit / blobdiff