// 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 {
+template<class T>
+struct Tap {
+ T weight;
+ T pos;
+};
+
float sinc(float x)
{
if (fabs(x) < 1e-6) {
return a;
}
-unsigned combine_samples(float *src, float *dst, unsigned num_src_samples, unsigned max_samples_saved)
+unsigned combine_samples(Tap<float> *src, Tap<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];
+ dst[j] = src[i];
}
if (i == num_src_samples - 1) {
continue;
}
- float w1 = src[i * 2 + 0];
- float w2 = src[(i + 1) * 2 + 0];
+ float w1 = src[i].weight;
+ float w2 = src[i + 1].weight;
if (w1 * w2 < 0.0f) {
// Differing signs; cannot combine.
continue;
}
- float pos1 = src[i * 2 + 1];
- float pos2 = src[(i + 1) * 2 + 1];
+ float pos1 = src[i].pos;
+ float pos2 = src[i + 1].pos;
assert(pos2 > pos1);
float offset, total_weight, sum_sq_error;
// 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);
+ dst[j].weight = total_weight;
+ dst[j].pos = pos1 + offset * (pos2 - pos1);
}
++i; // Skip the next sample.
ResampleEffect::ResampleEffect()
: input_width(1280),
- input_height(720)
+ input_height(720),
+ offset_x(0.0f), offset_y(0.0f),
+ zoom_x(1.0f), zoom_y(1.0f),
+ zoom_center_x(0.5f), zoom_center_y(0.5f)
{
register_int("width", &output_width);
register_int("height", &output_height);
input_height = height;
update_size();
}
-
+
void ResampleEffect::update_size()
{
bool ok = true;
ok |= vpass->set_int("output_height", output_height);
assert(ok);
+
+ // The offset added due to zoom may have changed with the size.
+ update_offset_and_zoom();
+}
+
+void ResampleEffect::update_offset_and_zoom()
+{
+ bool ok = true;
+
+ // Zoom from the right origin. (zoom_center is given in normalized coordinates,
+ // i.e. 0..1.)
+ float extra_offset_x = zoom_center_x * (1.0f - 1.0f / zoom_x) * input_width;
+ float extra_offset_y = (1.0f - zoom_center_y) * (1.0f - 1.0f / zoom_y) * input_height;
+
+ ok |= hpass->set_float("offset", extra_offset_x + offset_x);
+ ok |= vpass->set_float("offset", extra_offset_y - offset_y); // Compensate for the bottom-left origin.
+ ok |= hpass->set_float("zoom", zoom_x);
+ ok |= vpass->set_float("zoom", zoom_y);
+
+ assert(ok);
}
bool ResampleEffect::set_float(const string &key, float value) {
update_size();
return true;
}
+ if (key == "top") {
+ offset_y = value;
+ update_offset_and_zoom();
+ return true;
+ }
+ if (key == "left") {
+ offset_x = value;
+ update_offset_and_zoom();
+ return true;
+ }
+ if (key == "zoom_x") {
+ if (value <= 0.0f) {
+ return false;
+ }
+ zoom_x = value;
+ update_offset_and_zoom();
+ return true;
+ }
+ if (key == "zoom_y") {
+ if (value <= 0.0f) {
+ return false;
+ }
+ zoom_y = value;
+ update_offset_and_zoom();
+ return true;
+ }
+ if (key == "zoom_center_x") {
+ zoom_center_x = value;
+ update_offset_and_zoom();
+ return true;
+ }
+ if (key == "zoom_center_y") {
+ zoom_center_y = value;
+ update_offset_and_zoom();
+ return true;
+ }
return false;
}
direction(HORIZONTAL),
input_width(1280),
input_height(720),
+ offset(0.0),
+ zoom(1.0),
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.
+ last_zoom(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);
+ register_float("zoom", &zoom);
glGenTextures(1, &texnum);
}
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);
+ float scaling_factor;
+ if (fabs(zoom - 1.0f) < 1e-6) {
+ num_loops = gcd(src_size, dst_size);
+ scaling_factor = float(dst_size) / float(src_size);
+ } else {
+ // If zooming is enabled (ie., zoom != 1), we turn off the looping.
+ // We _could_ perhaps do it for rational zoom levels (especially
+ // things like 2:1), but it doesn't seem to be worth it, given that
+ // the most common use case would seem to be varying the zoom
+ // from frame to frame.
+ num_loops = 1;
+ scaling_factor = zoom * float(dst_size) / float(src_size);
+ }
slice_height = 1.0f / num_loops;
unsigned dst_samples = dst_size / num_loops;
// 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 = min(float(dst_size) / float(src_size), 1.0f);
+ float radius_scaling_factor = min(scaling_factor, 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];
+ Tap<float> *weights = new Tap<float>[dst_samples * src_samples];
+ 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.
- float center_src_y = (y + 0.5f) * float(src_size) / float(dst_size) - 0.5f;
+ float center_src_y = (y + 0.5f) / scaling_factor - 0.5f;
int base_src_y = lrintf(center_src_y);
// 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);
- 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);
+ float weight = lanczos_weight(radius_scaling_factor * (src_y - center_src_y - subpixel_offset), LANCZOS_RADIUS);
+ weights[y * src_samples + i].weight = weight * radius_scaling_factor;
+ weights[y * src_samples + i].pos = (src_y + 0.5) / float(src_size);
}
}
// 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);
+ unsigned num_samples_saved = combine_samples(weights + y * src_samples, 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];
+ Tap<float> *bilinear_weights = new Tap<float>[dst_samples * src_bilinear_samples];
+ Tap<fp16_int_t> *bilinear_weights_fp16 = new Tap<fp16_int_t>[dst_samples * src_bilinear_samples];
for (unsigned y = 0; y < dst_samples; ++y) {
+ Tap<float> *bilinear_weights_ptr = bilinear_weights + y * src_bilinear_samples;
+ Tap<fp16_int_t> *bilinear_weights_fp16_ptr = bilinear_weights_fp16 + y * src_bilinear_samples;
unsigned num_samples_saved = combine_samples(
- weights + (y * src_samples) * 2,
- bilinear_weights + (y * src_bilinear_samples) * 2,
+ weights + y * src_samples,
+ 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].weight = fp64_to_fp16(bilinear_weights_ptr[i].weight);
+ bilinear_weights_fp16_ptr[i].pos = fp64_to_fp16(bilinear_weights_ptr[i].pos);
+ }
+
+ // 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].weight);
+ }
+ for (int i = 0; i < src_bilinear_samples; ++i) {
+ bilinear_weights_fp16_ptr[i].weight = fp64_to_fp16(
+ fp16_to_fp64(bilinear_weights_fp16_ptr[i].weight) / 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 ||
+ zoom != last_zoom) {
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;
+ last_zoom = zoom;
}
glActiveTexture(GL_TEXTURE0 + *sampler_num);
check_error();
set_uniform_int(glsl_program_num, prefix, "sample_tex", *sampler_num);
- ++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);
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();
}
+
+} // namespace movit