// 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"
+using namespace std;
+
+namespace movit {
+
namespace {
float sinc(float x)
input_height = height;
update_size();
}
-
+
void ResampleEffect::update_size()
{
bool ok = true;
assert(ok);
}
-bool ResampleEffect::set_float(const std::string &key, float value) {
+bool ResampleEffect::set_float(const string &key, float value) {
if (key == "width") {
output_width = value;
update_size();
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));
//
// For horizontal scaling, we fill in the exact same texture;
// the shader just interprets it differently.
-void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const std::string &prefix, unsigned *sampler_num)
+void SingleResamplePassEffect::update_texture(GLuint glsl_program_num, const string &prefix, unsigned *sampler_num)
{
unsigned src_size, dst_size;
if (direction == SingleResamplePassEffect::HORIZONTAL) {
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.
// 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(float(dst_size) / float(src_size), 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];
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
src_bilinear_samples = 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<int>(src_bilinear_samples, src_samples - num_samples_saved);
+ src_bilinear_samples = max<int>(src_bilinear_samples, src_samples - num_samples_saved);
}
// 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) {
+ double sum = 0.0;
+ for (int i = 0; i < src_bilinear_samples; ++i) {
+ sum += fp16_to_fp64(bilinear_weights_fp16_ptr[i * 2 + 0]);
+ }
+ for (int i = 0; i < src_bilinear_samples; ++i) {
+ 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 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);
// 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
projects / movit / blobdiff