// Three-lobed Lanczos, the most common choice.
#define LANCZOS_RADIUS 3.0
-#include <math.h>
+#include <epoxy/gl.h>
#include <assert.h>
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <algorithm>
-#include "resample_effect.h"
#include "effect_chain.h"
+#include "effect_util.h"
+#include "fp16.h"
+#include "resample_effect.h"
#include "util.h"
-#include "opengl.h"
+
+using namespace std;
+
+namespace movit {
namespace {
}
}
+// Euclid's algorithm, from Wikipedia.
+unsigned gcd(unsigned a, unsigned b)
+{
+ while (b != 0) {
+ unsigned t = b;
+ b = a % b;
+ a = t;
+ }
+ return a;
+}
+
+unsigned combine_samples(float *src, float *dst, 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];
+ }
+
+ if (i == num_src_samples - 1) {
+ // Last sample; cannot combine.
+ continue;
+ }
+ assert(num_samples_saved <= max_samples_saved);
+ if (num_samples_saved == max_samples_saved) {
+ // We could maybe save more here, but other rows can't, so don't bother.
+ continue;
+ }
+
+ float w1 = src[i * 2 + 0];
+ float w2 = src[(i + 1) * 2 + 0];
+ if (w1 * w2 < 0.0f) {
+ // Differing signs; cannot combine.
+ continue;
+ }
+
+ float pos1 = src[i * 2 + 1];
+ float pos2 = src[(i + 1) * 2 + 1];
+ assert(pos2 > pos1);
+
+ float offset, total_weight, sum_sq_error;
+ combine_two_samples(w1, w2, &offset, &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,
+ // 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 / (256.0f * 256.0f)) {
+ 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);
+ }
+
+ ++i; // Skip the next sample.
+ ++num_samples_saved;
+ }
+ return num_samples_saved;
+}
+
} // namespace
ResampleEffect::ResampleEffect()
// 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();
}
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));
// so out[0] will read from parameters <x,y> = <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.
-//
-// TODO: Support optimization of wrapping the sample texture.
-// TODO: Support optimization using free linear sampling, like in BlurEffect.
-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) {
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);
+ slice_height = 1.0f / num_loops;
+ unsigned dst_samples = dst_size / num_loops;
+
// Sample the kernel in the right place. A diagram with a triangular kernel
// (corresponding to linear filtering, and obviously with radius 1)
// for easier ASCII art drawing:
// 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);
- src_samples = int_radius * 2 + 1;
- float *weights = new float[dst_size * src_samples * 2];
- for (unsigned y = 0; y < dst_size; ++y) {
+ int src_samples = int_radius * 2 + 1;
+ float *weights = new float[dst_samples * src_samples * 2];
+ 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;
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);
}
+
}
- // Encode as a two-component texture. Note the GL_REPEAT, which is not relevant
- // right now, but will be later.
+ // 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;
+ 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 = 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_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();
glBindTexture(GL_TEXTURE_2D, texnum);
check_error();
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
check_error();
- glTexImage2D(GL_TEXTURE_2D, 0, GL_RG16F, src_samples, dst_size, 0, GL_RG, GL_FLOAT, 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);
+ 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 ||
set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num);
++sampler_num;
- set_uniform_int(glsl_program_num, prefix, "num_samples", src_samples);
+ 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);
// 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_samples);
- set_uniform_float(glsl_program_num, prefix, "sample_x_offset", 0.5f / src_samples);
+ 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);
// 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