#define NO_SDL_GLEXT 1
-#include <epoxy/gl.h>
-
-#include <SDL2/SDL.h>
-#include <SDL2/SDL_error.h>
-#include <SDL2/SDL_events.h>
-#include <SDL2/SDL_image.h>
-#include <SDL2/SDL_keyboard.h>
-#include <SDL2/SDL_mouse.h>
-#include <SDL2/SDL_video.h>
-
-#include <assert.h>
-#include <getopt.h>
-#include <stdio.h>
-#include <unistd.h>
-
#include "flow.h"
+
#include "gpu_timers.h"
#include "util.h"
#include <algorithm>
+#include <assert.h>
#include <deque>
-#include <memory>
+#include <epoxy/gl.h>
#include <map>
+#include <memory>
#include <stack>
+#include <stdio.h>
+#include <string.h>
+#include <unistd.h>
#include <vector>
#define BUFFER_OFFSET(i) ((char *)nullptr + (i))
using namespace std;
-SDL_Window *window;
-
-// Operating point 3 (10 Hz on CPU, excluding preprocessing).
-constexpr float patch_overlap_ratio = 0.75f;
-constexpr unsigned coarsest_level = 5;
-constexpr unsigned finest_level = 1;
-constexpr unsigned patch_size_pixels = 12;
-
// Weighting constants for the different parts of the variational refinement.
// These don't correspond 1:1 to the values given in the DIS paper,
// since we have different normalizations and ranges in some cases.
// although the error (EPE) seems to be fairly insensitive to the precise values.
// Only the relative values matter, so we fix alpha (the smoothness constant)
// at unity and tweak the others.
+//
+// TODO: Maybe this should not be global.
float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
-bool enable_timing = true;
-bool detailed_timing = false;
-bool enable_warmup = false;
-bool in_warmup = false;
-bool enable_variational_refinement = true; // Just for debugging.
-bool enable_interpolation = false;
-
// Some global OpenGL objects.
// TODO: These should really be part of DISComputeFlow.
GLuint nearest_sampler, linear_sampler, zero_border_sampler;
GLuint vertex_vbo;
-// Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F).
-struct ReadInProgress {
- GLuint pbo;
- string filename0, filename1;
- string flow_filename, ppm_filename; // Either may be empty for no write.
-};
-stack<GLuint> spare_pbos;
-deque<ReadInProgress> reads_in_progress;
-
int find_num_levels(int width, int height)
{
int levels = 1;
}
if (ret == 0) {
fprintf(stderr, "Short read when trying to read %d bytes from %s\n",
- size, filename.c_str());
+ size, filename.c_str());
exit(1);
}
fclose(fp);
return str;
}
-
GLuint compile_shader(const string &shader_src, GLenum type)
{
GLuint obj = glCreateShader(type);
- const GLchar* source[] = { shader_src.data() };
+ const GLchar *source[] = { shader_src.data() };
const GLint length[] = { (GLint)shader_src.size() };
glShaderSource(obj, 1, source, length);
glCompileShader(obj);
return obj;
}
-enum MipmapPolicy {
- WITHOUT_MIPMAPS,
- WITH_MIPMAPS
-};
-
-GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret, MipmapPolicy mipmaps)
-{
- SDL_Surface *surf = IMG_Load(filename);
- if (surf == nullptr) {
- fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
- exit(1);
- }
-
- // For whatever reason, SDL doesn't support converting to YUV surfaces
- // nor grayscale, so we'll do it ourselves.
- SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA32, /*flags=*/0);
- if (rgb_surf == nullptr) {
- fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
- exit(1);
- }
-
- SDL_FreeSurface(surf);
-
- unsigned width = rgb_surf->w, height = rgb_surf->h;
- const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
- unique_ptr<uint8_t[]> pix(new uint8_t[width * height * 4]);
-
- // Extract the Y component, and convert to bottom-left origin.
- for (unsigned y = 0; y < height; ++y) {
- unsigned y2 = height - 1 - y;
- memcpy(pix.get() + y * width * 4, sptr + y2 * rgb_surf->pitch, width * 4);
- }
- SDL_FreeSurface(rgb_surf);
-
- int num_levels = (mipmaps == WITH_MIPMAPS) ? find_num_levels(width, height) : 1;
-
- GLuint tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &tex);
- glTextureStorage2D(tex, num_levels, GL_RGBA8, width, height);
- glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RGBA, GL_UNSIGNED_BYTE, pix.get());
-
- if (mipmaps == WITH_MIPMAPS) {
- glGenerateTextureMipmap(tex);
- }
-
- *width_ret = width;
- *height_ret = height;
-
- return tex;
-}
-
GLuint link_program(GLuint vs_obj, GLuint fs_obj)
{
GLuint program = glCreateProgram();
glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-MotionSearch::MotionSearch()
+MotionSearch::MotionSearch(const OperatingPoint &op)
+ : op(op)
{
motion_vs_obj = compile_shader(read_file("motion_search.vert"), GL_VERTEX_SHADER);
motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
uniform_image_tex = glGetUniformLocation(motion_search_program, "image_tex");
uniform_grad_tex = glGetUniformLocation(motion_search_program, "grad_tex");
uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
+ uniform_patch_size = glGetUniformLocation(motion_search_program, "patch_size");
+ uniform_num_iterations = glGetUniformLocation(motion_search_program, "num_iterations");
}
void MotionSearch::exec(GLuint tex_view, GLuint grad_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int prev_level_width, int prev_level_height, int width_patches, int height_patches, int num_layers)
glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
glProgramUniform2f(motion_search_program, uniform_out_flow_size, width_patches, height_patches);
+ glProgramUniform1ui(motion_search_program, uniform_patch_size, op.patch_size_pixels);
+ glProgramUniform1ui(motion_search_program, uniform_num_iterations, op.search_iterations);
glViewport(0, 0, width_patches, height_patches);
fbos.render_to(flow_out_tex);
glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-Densify::Densify()
+Densify::Densify(const OperatingPoint &op)
+ : op(op)
{
densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
bind_sampler(densify_program, uniform_flow_tex, 1, flow_tex, nearest_sampler);
glProgramUniform2f(densify_program, uniform_patch_size,
- float(patch_size_pixels) / level_width,
- float(patch_size_pixels) / level_height);
+ float(op.patch_size_pixels) / level_width,
+ float(op.patch_size_pixels) / level_height);
glViewport(0, 0, level_width, level_height);
glEnable(GL_BLEND);
glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, num_layers);
}
-DISComputeFlow::DISComputeFlow(int width, int height)
- : width(width), height(height)
+DISComputeFlow::DISComputeFlow(int width, int height, const OperatingPoint &op)
+ : width(width), height(height), op(op), motion_search(op), densify(op)
{
// Make some samplers.
glCreateSamplers(1, &nearest_sampler);
GPUTimers timers;
glBindVertexArray(vao);
+ glDisable(GL_DITHER);
ScopedTimer total_timer("Compute flow", &timers);
- for (int level = coarsest_level; level >= int(finest_level); --level) {
+ for (int level = op.coarsest_level; level >= int(op.finest_level); --level) {
char timer_name[256];
snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
ScopedTimer level_timer(timer_name, &total_timer);
int level_width = width >> level;
int level_height = height >> level;
- float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
+ float patch_spacing_pixels = op.patch_size_pixels * (1.0f - op.patch_overlap_ratio);
// Make sure we have patches at least every Nth pixel, e.g. for width=9
// and patch_spacing=3 (the default), we put out patch centers in
pool.release_texture(dense_flow_tex);
glDeleteTextures(1, &tex_view);
- // Calculate I_x and I_y. We're only calculating first derivatives;
- // the others will be taken on-the-fly in order to sample from fewer
- // textures overall, since sampling from the L1 cache is cheap.
- // (TODO: Verify that this is indeed faster than making separate
- // double-derivative textures.)
- GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
- GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
- {
- ScopedTimer timer("First derivatives", &varref_timer);
- derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers);
- }
- pool.release_texture(I_tex);
-
- // We need somewhere to store du and dv (the flow increment, relative
- // to the non-refined base flow u0 and v0). It's initially garbage,
- // but not read until we've written something sane to it.
- GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
-
- // And for diffusivity.
- GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
-
- // And finally for the equation set. See SetupEquations for
- // the storage format.
- GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
- GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
-
- for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
- // Calculate the diffusivity term for each pixel.
+ // TODO: If we don't have variational refinement, we don't need I and I_t,
+ // so computing them is a waste.
+ if (op.variational_refinement) {
+ // Calculate I_x and I_y. We're only calculating first derivatives;
+ // the others will be taken on-the-fly in order to sample from fewer
+ // textures overall, since sampling from the L1 cache is cheap.
+ // (TODO: Verify that this is indeed faster than making separate
+ // double-derivative textures.)
+ GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
+ GLuint beta_0_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
{
- ScopedTimer timer("Compute diffusivity", &varref_timer);
- compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0, num_layers);
+ ScopedTimer timer("First derivatives", &varref_timer);
+ derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height, num_layers);
}
-
- // Set up the 2x2 equation system for each pixel.
- {
- ScopedTimer timer("Set up equations", &varref_timer);
- setup_equations.exec(I_x_y_tex, I_t_tex, diff_flow_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0, num_layers);
+ pool.release_texture(I_tex);
+
+ // We need somewhere to store du and dv (the flow increment, relative
+ // to the non-refined base flow u0 and v0). It's initially garbage,
+ // but not read until we've written something sane to it.
+ GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height, num_layers);
+
+ // And for diffusivity.
+ GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height, num_layers);
+
+ // And finally for the equation set. See SetupEquations for
+ // the storage format.
+ GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
+ GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height, num_layers);
+
+ for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
+ // Calculate the diffusivity term for each pixel.
+ {
+ ScopedTimer timer("Compute diffusivity", &varref_timer);
+ compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0, num_layers);
+ }
+
+ // Set up the 2x2 equation system for each pixel.
+ {
+ ScopedTimer timer("Set up equations", &varref_timer);
+ setup_equations.exec(I_x_y_tex, I_t_tex, diff_flow_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0, num_layers);
+ }
+
+ // Run a few SOR iterations. Note that these are to/from the same texture.
+ {
+ ScopedTimer timer("SOR", &varref_timer);
+ sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer);
+ }
}
- // Run a few SOR iterations. Note that these are to/from the same texture.
+ pool.release_texture(I_t_tex);
+ pool.release_texture(I_x_y_tex);
+ pool.release_texture(beta_0_tex);
+ pool.release_texture(diffusivity_tex);
+ pool.release_texture(equation_red_tex);
+ pool.release_texture(equation_black_tex);
+
+ // Add the differential flow found by the variational refinement to the base flow,
+ // giving the final flow estimate for this level.
+ // The output is in base_flow_tex; we don't need to make a new texture.
{
- ScopedTimer timer("SOR", &varref_timer);
- sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, num_layers, &timer);
+ ScopedTimer timer("Add differential flow", &varref_timer);
+ add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers);
}
+ pool.release_texture(diff_flow_tex);
}
- pool.release_texture(I_t_tex);
- pool.release_texture(I_x_y_tex);
- pool.release_texture(beta_0_tex);
- pool.release_texture(diffusivity_tex);
- pool.release_texture(equation_red_tex);
- pool.release_texture(equation_black_tex);
-
- // Add the differential flow found by the variational refinement to the base flow,
- // giving the final flow estimate for this level.
- // The output is in diff_flow_tex; we don't need to make a new texture.
- //
- // Disabling this doesn't save any time (although we could easily make it so that
- // it is more efficient), but it helps debug the motion search.
- if (enable_variational_refinement) {
- ScopedTimer timer("Add differential flow", &varref_timer);
- add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height, num_layers);
- }
- pool.release_texture(diff_flow_tex);
-
if (prev_level_flow_tex != initial_flow_tex) {
pool.release_texture(prev_level_flow_tex);
}
}
// Scale up the flow to the final size (if needed).
- if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
+ if (op.finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
return prev_level_flow_tex;
} else {
GLuint final_tex = pool.get_texture(GL_RG16F, width, height, num_layers);
}
}
-Splat::Splat()
+Splat::Splat(const OperatingPoint &op)
+ : op(op)
{
splat_vs_obj = compile_shader(read_file("splat.vert"), GL_VERTEX_SHADER);
splat_fs_obj = compile_shader(read_file("splat.frag"), GL_FRAGMENT_SHADER);
uniform_splat_size = glGetUniformLocation(splat_program, "splat_size");
uniform_alpha = glGetUniformLocation(splat_program, "alpha");
- uniform_image_tex = glGetUniformLocation(splat_program, "image_tex");
+ uniform_gray_tex = glGetUniformLocation(splat_program, "gray_tex");
uniform_flow_tex = glGetUniformLocation(splat_program, "flow_tex");
uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
}
-void Splat::exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha)
+void Splat::exec(GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha)
{
glUseProgram(splat_program);
- bind_sampler(splat_program, uniform_image_tex, 0, image_tex, linear_sampler);
+ bind_sampler(splat_program, uniform_gray_tex, 0, gray_tex, linear_sampler);
bind_sampler(splat_program, uniform_flow_tex, 1, bidirectional_flow_tex, nearest_sampler);
- // FIXME: This is set to 1.0 right now so not to trigger Haswell's “PMA stall”.
- // Move to 2.0 later, or even 4.0.
- // (Since we have hole filling, it's not critical, but larger values seem to do
- // better than hole filling for large motion, blurs etc.)
- float splat_size = 1.0f; // 4x4 splat means 16x overdraw, 2x2 splat means 4x overdraw.
- glProgramUniform2f(splat_program, uniform_splat_size, splat_size / width, splat_size / height);
+ glProgramUniform2f(splat_program, uniform_splat_size, op.splat_size / width, op.splat_size / height);
glProgramUniform1f(splat_program, uniform_alpha, alpha);
glProgramUniform2f(splat_program, uniform_inv_flow_size, 1.0f / width, 1.0f / height);
glViewport(0, 0, width, height);
glDisable(GL_BLEND);
glEnable(GL_DEPTH_TEST);
+ glDepthMask(GL_TRUE);
glDepthFunc(GL_LESS); // We store the difference between I_0 and I_1, where less difference is good. (Default 1.0 is effectively +inf, which always loses.)
fbos.render_to(depth_rb, flow_tex);
glDisable(GL_DEPTH_TEST);
}
-Blend::Blend()
+Blend::Blend(bool split_ycbcr_output)
+ : split_ycbcr_output(split_ycbcr_output)
{
+ string frag_shader = read_file("blend.frag");
+ if (split_ycbcr_output) {
+ // Insert after the first #version line.
+ size_t offset = frag_shader.find('\n');
+ assert(offset != string::npos);
+ frag_shader = frag_shader.substr(0, offset + 1) + "#define SPLIT_YCBCR_OUTPUT 1\n" + frag_shader.substr(offset + 1);
+ }
+
blend_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
- blend_fs_obj = compile_shader(read_file("blend.frag"), GL_FRAGMENT_SHADER);
+ blend_fs_obj = compile_shader(frag_shader, GL_FRAGMENT_SHADER);
blend_program = link_program(blend_vs_obj, blend_fs_obj);
uniform_image_tex = glGetUniformLocation(blend_program, "image_tex");
uniform_flow_consistency_tolerance = glGetUniformLocation(blend_program, "flow_consistency_tolerance");
}
-void Blend::exec(GLuint image_tex, GLuint flow_tex, GLuint output_tex, int level_width, int level_height, float alpha)
+void Blend::exec(GLuint image_tex, GLuint flow_tex, GLuint output_tex, GLuint output2_tex, int level_width, int level_height, float alpha)
{
glUseProgram(blend_program);
bind_sampler(blend_program, uniform_image_tex, 0, image_tex, linear_sampler);
glProgramUniform1f(blend_program, uniform_alpha, alpha);
glViewport(0, 0, level_width, level_height);
- fbos.render_to(output_tex);
+ if (split_ycbcr_output) {
+ fbos_split.render_to(output_tex, output2_tex);
+ } else {
+ fbos.render_to(output_tex);
+ }
glDisable(GL_BLEND); // A bit ironic, perhaps.
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
}
-Interpolate::Interpolate(int width, int height, int flow_level)
- : width(width), height(height), flow_level(flow_level) {
+Interpolate::Interpolate(const OperatingPoint &op, bool split_ycbcr_output)
+ : flow_level(op.finest_level),
+ split_ycbcr_output(split_ycbcr_output),
+ splat(op),
+ blend(split_ycbcr_output) {
// Set up the vertex data that will be shared between all passes.
float vertices[] = {
0.0f, 1.0f,
glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
}
-GLuint Interpolate::exec(GLuint image_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha)
+pair<GLuint, GLuint> Interpolate::exec(GLuint image_tex, GLuint gray_tex, GLuint bidirectional_flow_tex, GLuint width, GLuint height, float alpha)
{
GPUTimers timers;
ScopedTimer total_timer("Interpolate", &timers);
glBindVertexArray(vao);
+ glDisable(GL_DITHER);
// Pick out the right level to test splatting results on.
GLuint tex_view;
glGenTextures(1, &tex_view);
- glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, image_tex, GL_RGBA8, flow_level, 1, 0, 2);
+ glTextureView(tex_view, GL_TEXTURE_2D_ARRAY, gray_tex, GL_R8, flow_level, 1, 0, 2);
int flow_width = width >> flow_level;
int flow_height = height >> flow_level;
pool.release_texture(temp_tex[2]);
pool.release_renderbuffer(depth_rb);
- GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
- {
- ScopedTimer timer("Blend", &total_timer);
- blend.exec(image_tex, flow_tex, output_tex, width, height, alpha);
+ GLuint output_tex, output2_tex = 0;
+ if (split_ycbcr_output) {
+ output_tex = pool.get_texture(GL_R8, width, height);
+ output2_tex = pool.get_texture(GL_RG8, width, height);
+ {
+ ScopedTimer timer("Blend", &total_timer);
+ blend.exec(image_tex, flow_tex, output_tex, output2_tex, width, height, alpha);
+ }
+ } else {
+ output_tex = pool.get_texture(GL_RGBA8, width, height);
+ {
+ ScopedTimer timer("Blend", &total_timer);
+ blend.exec(image_tex, flow_tex, output_tex, 0, width, height, alpha);
+ }
}
pool.release_texture(flow_tex);
total_timer.end();
timers.print();
}
- return output_tex;
+ return make_pair(output_tex, output2_tex);
}
GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height, GLuint num_layers)
{
- for (Texture &tex : textures) {
- if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
- tex.width == width && tex.height == height && tex.num_layers == num_layers) {
- tex.in_use = true;
- return tex.tex_num;
+ {
+ lock_guard<mutex> lock(mu);
+ for (Texture &tex : textures) {
+ if (!tex.in_use && !tex.is_renderbuffer && tex.format == format &&
+ tex.width == width && tex.height == height && tex.num_layers == num_layers) {
+ tex.in_use = true;
+ return tex.tex_num;
+ }
}
}
tex.num_layers = num_layers;
tex.in_use = true;
tex.is_renderbuffer = false;
- textures.push_back(tex);
+ {
+ lock_guard<mutex> lock(mu);
+ textures.push_back(tex);
+ }
return tex.tex_num;
}
GLuint TexturePool::get_renderbuffer(GLenum format, GLuint width, GLuint height)
{
- for (Texture &tex : textures) {
- if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
- tex.width == width && tex.height == height) {
- tex.in_use = true;
- return tex.tex_num;
+ {
+ lock_guard<mutex> lock(mu);
+ for (Texture &tex : textures) {
+ if (!tex.in_use && tex.is_renderbuffer && tex.format == format &&
+ tex.width == width && tex.height == height) {
+ tex.in_use = true;
+ return tex.tex_num;
+ }
}
}
tex.height = height;
tex.in_use = true;
tex.is_renderbuffer = true;
- textures.push_back(tex);
+ {
+ lock_guard<mutex> lock(mu);
+ textures.push_back(tex);
+ }
return tex.tex_num;
}
void TexturePool::release_texture(GLuint tex_num)
{
+ lock_guard<mutex> lock(mu);
for (Texture &tex : textures) {
if (!tex.is_renderbuffer && tex.tex_num == tex_num) {
assert(tex.in_use);
void TexturePool::release_renderbuffer(GLuint tex_num)
{
+ lock_guard<mutex> lock(mu);
for (Texture &tex : textures) {
if (tex.is_renderbuffer && tex.tex_num == tex_num) {
assert(tex.in_use);
}
//assert(false);
}
-
-// OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
-void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
-{
- for (unsigned i = 0; i < width * height; ++i) {
- dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
- }
-}
-
-// Not relevant for RGB.
-void flip_coordinate_system(uint8_t *dense_flow, unsigned width, unsigned height)
-{
-}
-
-void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
-{
- FILE *flowfp = fopen(filename, "wb");
- fprintf(flowfp, "FEIH");
- fwrite(&width, 4, 1, flowfp);
- fwrite(&height, 4, 1, flowfp);
- for (unsigned y = 0; y < height; ++y) {
- int yy = height - y - 1;
- fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
- }
- fclose(flowfp);
-}
-
-// Not relevant for RGB.
-void write_flow(const char *filename, const uint8_t *dense_flow, unsigned width, unsigned height)
-{
- assert(false);
-}
-
-void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
-{
- FILE *fp = fopen(filename, "wb");
- fprintf(fp, "P6\n%d %d\n255\n", width, height);
- for (unsigned y = 0; y < unsigned(height); ++y) {
- int yy = height - y - 1;
- for (unsigned x = 0; x < unsigned(width); ++x) {
- float du = dense_flow[(yy * width + x) * 2 + 0];
- float dv = dense_flow[(yy * width + x) * 2 + 1];
-
- uint8_t r, g, b;
- flow2rgb(du, dv, &r, &g, &b);
- putc(r, fp);
- putc(g, fp);
- putc(b, fp);
- }
- }
- fclose(fp);
-}
-
-void write_ppm(const char *filename, const uint8_t *rgba, unsigned width, unsigned height)
-{
- unique_ptr<uint8_t[]> rgb_line(new uint8_t[width * 3 + 1]);
-
- FILE *fp = fopen(filename, "wb");
- fprintf(fp, "P6\n%d %d\n255\n", width, height);
- for (unsigned y = 0; y < height; ++y) {
- unsigned y2 = height - 1 - y;
- for (size_t x = 0; x < width; ++x) {
- memcpy(&rgb_line[x * 3], &rgba[(y2 * width + x) * 4], 4);
- }
- fwrite(rgb_line.get(), width * 3, 1, fp);
- }
- fclose(fp);
-}
-
-struct FlowType {
- using type = float;
- static constexpr GLenum gl_format = GL_RG;
- static constexpr GLenum gl_type = GL_FLOAT;
- static constexpr int num_channels = 2;
-};
-
-struct RGBAType {
- using type = uint8_t;
- static constexpr GLenum gl_format = GL_RGBA;
- static constexpr GLenum gl_type = GL_UNSIGNED_BYTE;
- static constexpr int num_channels = 4;
-};
-
-template <class Type>
-void finish_one_read(GLuint width, GLuint height)
-{
- using T = typename Type::type;
- constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
-
- assert(!reads_in_progress.empty());
- ReadInProgress read = reads_in_progress.front();
- reads_in_progress.pop_front();
-
- unique_ptr<T[]> flow(new typename Type::type[width * height * Type::num_channels]);
- void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * bytes_per_pixel, GL_MAP_READ_BIT); // Blocks if the read isn't done yet.
- memcpy(flow.get(), buf, width * height * bytes_per_pixel); // TODO: Unneeded for RGBType, since flip_coordinate_system() does nothing.:
- glUnmapNamedBuffer(read.pbo);
- spare_pbos.push(read.pbo);
-
- flip_coordinate_system(flow.get(), width, height);
- if (!read.flow_filename.empty()) {
- write_flow(read.flow_filename.c_str(), flow.get(), width, height);
- fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str());
- }
- if (!read.ppm_filename.empty()) {
- write_ppm(read.ppm_filename.c_str(), flow.get(), width, height);
- }
-}
-
-template <class Type>
-void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename)
-{
- using T = typename Type::type;
- constexpr int bytes_per_pixel = Type::num_channels * sizeof(T);
-
- if (spare_pbos.empty()) {
- finish_one_read<Type>(width, height);
- }
- assert(!spare_pbos.empty());
- reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename });
- glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top());
- spare_pbos.pop();
- glGetTextureImage(tex, 0, Type::gl_format, Type::gl_type, width * height * bytes_per_pixel, nullptr);
- glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
-}
-
-void compute_flow_only(int argc, char **argv, int optind)
-{
- const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
- const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
- const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo";
-
- // Load pictures.
- unsigned width1, height1, width2, height2;
- GLuint tex0 = load_texture(filename0, &width1, &height1, WITHOUT_MIPMAPS);
- GLuint tex1 = load_texture(filename1, &width2, &height2, WITHOUT_MIPMAPS);
-
- if (width1 != width2 || height1 != height2) {
- fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
- width1, height1, width2, height2);
- exit(1);
- }
-
- // Move them into an array texture, since that's how the rest of the code
- // would like them.
- GLuint image_tex;
- glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &image_tex);
- glTextureStorage3D(image_tex, 1, GL_RGBA8, width1, height1, 2);
- glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
- glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
- glDeleteTextures(1, &tex0);
- glDeleteTextures(1, &tex1);
-
- // Set up some PBOs to do asynchronous readback.
- GLuint pbos[5];
- glCreateBuffers(5, pbos);
- for (int i = 0; i < 5; ++i) {
- glNamedBufferData(pbos[i], width1 * height1 * 2 * 2 * sizeof(float), nullptr, GL_STREAM_READ);
- spare_pbos.push(pbos[i]);
- }
-
- int levels = find_num_levels(width1, height1);
-
- GLuint tex_gray;
- glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex_gray);
- glTextureStorage3D(tex_gray, levels, GL_R8, width1, height1, 2);
-
- GrayscaleConversion gray;
- gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
- glGenerateTextureMipmap(tex_gray);
-
- DISComputeFlow compute_flow(width1, height1);
-
- if (enable_warmup) {
- in_warmup = true;
- for (int i = 0; i < 10; ++i) {
- GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
- compute_flow.release_texture(final_tex);
- }
- in_warmup = false;
- }
-
- GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
- //GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
-
- schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
- compute_flow.release_texture(final_tex);
-
- // See if there are more flows on the command line (ie., more than three arguments),
- // and if so, process them.
- int num_flows = (argc - optind) / 3;
- for (int i = 1; i < num_flows; ++i) {
- const char *filename0 = argv[optind + i * 3 + 0];
- const char *filename1 = argv[optind + i * 3 + 1];
- const char *flow_filename = argv[optind + i * 3 + 2];
- GLuint width, height;
- GLuint tex0 = load_texture(filename0, &width, &height, WITHOUT_MIPMAPS);
- if (width != width1 || height != height1) {
- fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
- filename0, width, height, width1, height1);
- exit(1);
- }
- glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
- glDeleteTextures(1, &tex0);
-
- GLuint tex1 = load_texture(filename1, &width, &height, WITHOUT_MIPMAPS);
- if (width != width1 || height != height1) {
- fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
- filename1, width, height, width1, height1);
- exit(1);
- }
- glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
- glDeleteTextures(1, &tex1);
-
- gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
- glGenerateTextureMipmap(tex_gray);
-
- GLuint final_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
-
- schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "");
- compute_flow.release_texture(final_tex);
- }
- glDeleteTextures(1, &tex_gray);
-
- while (!reads_in_progress.empty()) {
- finish_one_read<FlowType>(width1, height1);
- }
-}
-
-// Interpolate images based on
-//
-// Herbst, Seitz, Baker: “Occlusion Reasoning for Temporal Interpolation
-// Using Optical Flow”
-//
-// or at least a reasonable subset thereof. Unfinished.
-void interpolate_image(int argc, char **argv, int optind)
-{
- const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png";
- const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png";
- //const char *out_filename = argc >= (optind + 3) ? argv[optind + 2] : "interpolated.png";
-
- // Load pictures.
- unsigned width1, height1, width2, height2;
- GLuint tex0 = load_texture(filename0, &width1, &height1, WITH_MIPMAPS);
- GLuint tex1 = load_texture(filename1, &width2, &height2, WITH_MIPMAPS);
-
- if (width1 != width2 || height1 != height2) {
- fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
- width1, height1, width2, height2);
- exit(1);
- }
-
- // Move them into an array texture, since that's how the rest of the code
- // would like them.
- int levels = find_num_levels(width1, height1);
- GLuint image_tex;
- glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &image_tex);
- glTextureStorage3D(image_tex, levels, GL_RGBA8, width1, height1, 2);
- glCopyImageSubData(tex0, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 0, width1, height1, 1);
- glCopyImageSubData(tex1, GL_TEXTURE_2D, 0, 0, 0, 0, image_tex, GL_TEXTURE_2D_ARRAY, 0, 0, 0, 1, width1, height1, 1);
- glDeleteTextures(1, &tex0);
- glDeleteTextures(1, &tex1);
- glGenerateTextureMipmap(image_tex);
-
- // Set up some PBOs to do asynchronous readback.
- GLuint pbos[5];
- glCreateBuffers(5, pbos);
- for (int i = 0; i < 5; ++i) {
- glNamedBufferData(pbos[i], width1 * height1 * 4 * sizeof(uint8_t), nullptr, GL_STREAM_READ);
- spare_pbos.push(pbos[i]);
- }
-
- DISComputeFlow compute_flow(width1, height1);
- GrayscaleConversion gray;
- Interpolate interpolate(width1, height1, finest_level);
-
- GLuint tex_gray;
- glCreateTextures(GL_TEXTURE_2D_ARRAY, 1, &tex_gray);
- glTextureStorage3D(tex_gray, levels, GL_R8, width1, height1, 2);
- gray.exec(image_tex, tex_gray, width1, height1, /*num_layers=*/2);
- glGenerateTextureMipmap(tex_gray);
-
- if (enable_warmup) {
- in_warmup = true;
- for (int i = 0; i < 10; ++i) {
- GLuint bidirectional_flow_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
- GLuint interpolated_tex = interpolate.exec(image_tex, bidirectional_flow_tex, width1, height1, 0.5f);
- compute_flow.release_texture(bidirectional_flow_tex);
- interpolate.release_texture(interpolated_tex);
- }
- in_warmup = false;
- }
-
- GLuint bidirectional_flow_tex = compute_flow.exec(tex_gray, DISComputeFlow::FORWARD_AND_BACKWARD, DISComputeFlow::DO_NOT_RESIZE_FLOW);
-
- for (int frameno = 1; frameno < 60; ++frameno) {
- char ppm_filename[256];
- snprintf(ppm_filename, sizeof(ppm_filename), "interp%04d.ppm", frameno);
-
- float alpha = frameno / 60.0f;
- GLuint interpolated_tex = interpolate.exec(image_tex, bidirectional_flow_tex, width1, height1, alpha);
-
- schedule_read<RGBAType>(interpolated_tex, width1, height1, filename0, filename1, "", ppm_filename);
- interpolate.release_texture(interpolated_tex);
- }
-
- while (!reads_in_progress.empty()) {
- finish_one_read<RGBAType>(width1, height1);
- }
-}
-
-int main(int argc, char **argv)
-{
- static const option long_options[] = {
- { "smoothness-relative-weight", required_argument, 0, 's' }, // alpha.
- { "intensity-relative-weight", required_argument, 0, 'i' }, // delta.
- { "gradient-relative-weight", required_argument, 0, 'g' }, // gamma.
- { "disable-timing", no_argument, 0, 1000 },
- { "detailed-timing", no_argument, 0, 1003 },
- { "ignore-variational-refinement", no_argument, 0, 1001 }, // Still calculates it, just doesn't apply it.
- { "interpolate", no_argument, 0, 1002 },
- { "warmup", no_argument, 0, 1004 }
- };
-
- for ( ;; ) {
- int option_index = 0;
- int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
-
- if (c == -1) {
- break;
- }
- switch (c) {
- case 's':
- vr_alpha = atof(optarg);
- break;
- case 'i':
- vr_delta = atof(optarg);
- break;
- case 'g':
- vr_gamma = atof(optarg);
- break;
- case 1000:
- enable_timing = false;
- break;
- case 1001:
- enable_variational_refinement = false;
- break;
- case 1002:
- enable_interpolation = true;
- break;
- case 1003:
- detailed_timing = true;
- break;
- case 1004:
- enable_warmup = true;
- break;
- default:
- fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
- exit(1);
- };
- }
-
- if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
- fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
- exit(1);
- }
- SDL_GL_SetAttribute(SDL_GL_ALPHA_SIZE, 8);
- SDL_GL_SetAttribute(SDL_GL_DEPTH_SIZE, 0);
- SDL_GL_SetAttribute(SDL_GL_STENCIL_SIZE, 0);
- SDL_GL_SetAttribute(SDL_GL_DOUBLEBUFFER, 1);
-
- SDL_GL_SetAttribute(SDL_GL_CONTEXT_PROFILE_MASK, SDL_GL_CONTEXT_PROFILE_CORE);
- SDL_GL_SetAttribute(SDL_GL_CONTEXT_MAJOR_VERSION, 4);
- SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
- // SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
- window = SDL_CreateWindow("OpenGL window",
- SDL_WINDOWPOS_UNDEFINED,
- SDL_WINDOWPOS_UNDEFINED,
- 64, 64,
- SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
- SDL_GLContext context = SDL_GL_CreateContext(window);
- assert(context != nullptr);
-
- glDisable(GL_DITHER);
-
- // FIXME: Should be part of DISComputeFlow (but needs to be initialized
- // before all the render passes).
- float vertices[] = {
- 0.0f, 1.0f,
- 0.0f, 0.0f,
- 1.0f, 1.0f,
- 1.0f, 0.0f,
- };
- glCreateBuffers(1, &vertex_vbo);
- glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
- glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
-
- if (enable_interpolation) {
- interpolate_image(argc, argv, optind);
- } else {
- compute_flow_only(argc, argv, optind);
- }
-}