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;
template<size_t num_elements>
class PersistentFBOSetWithDepth {
public:
- void render_to(GLuint depth_tex, const array<GLuint, num_elements> &textures);
+ void render_to(GLuint depth_rb, const array<GLuint, num_elements> &textures);
// Convenience wrappers.
- void render_to(GLuint depth_tex, GLuint texture0) {
- render_to(depth_tex, {{texture0}});
+ void render_to(GLuint depth_rb, GLuint texture0) {
+ render_to(depth_rb, {{texture0}});
}
- void render_to(GLuint depth_tex, GLuint texture0, GLuint texture1) {
- render_to(depth_tex, {{texture0, texture1}});
+ void render_to(GLuint depth_rb, GLuint texture0, GLuint texture1) {
+ render_to(depth_rb, {{texture0, texture1}});
}
- void render_to(GLuint depth_tex, GLuint texture0, GLuint texture1, GLuint texture2) {
- render_to(depth_tex, {{texture0, texture1, texture2}});
+ void render_to(GLuint depth_rb, GLuint texture0, GLuint texture1, GLuint texture2) {
+ render_to(depth_rb, {{texture0, texture1, texture2}});
}
- void render_to(GLuint depth_tex, GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3) {
- render_to(depth_tex, {{texture0, texture1, texture2, texture3}});
+ void render_to(GLuint depth_rb, GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3) {
+ render_to(depth_rb, {{texture0, texture1, texture2, texture3}});
}
private:
};
template<size_t num_elements>
-void PersistentFBOSetWithDepth<num_elements>::render_to(GLuint depth_tex, const array<GLuint, num_elements> &textures)
+void PersistentFBOSetWithDepth<num_elements>::render_to(GLuint depth_rb, const array<GLuint, num_elements> &textures)
{
- auto key = make_pair(depth_tex, textures);
+ auto key = make_pair(depth_rb, textures);
auto it = fbos.find(key);
if (it != fbos.end()) {
GLuint fbo;
glCreateFramebuffers(1, &fbo);
GLenum bufs[num_elements];
- glNamedFramebufferTexture(fbo, GL_DEPTH_ATTACHMENT, depth_tex, 0);
+ glNamedFramebufferRenderbuffer(fbo, GL_DEPTH_ATTACHMENT, GL_RENDERBUFFER, depth_rb);
for (size_t i = 0; i < num_elements; ++i) {
glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
bufs[i] = GL_COLOR_ATTACHMENT0 + i;
GLuint motion_fs_obj;
GLuint motion_search_program;
- GLuint uniform_inv_image_size, uniform_inv_prev_level_size;
+ GLuint uniform_inv_image_size, uniform_inv_prev_level_size, uniform_out_flow_size;
GLuint uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
};
uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_size");
uniform_inv_prev_level_size = glGetUniformLocation(motion_search_program, "inv_prev_level_size");
+ uniform_out_flow_size = glGetUniformLocation(motion_search_program, "out_flow_size");
uniform_image1_tex = glGetUniformLocation(motion_search_program, "image1_tex");
uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
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);
glViewport(0, 0, width_patches, height_patches);
fbos.render_to(flow_out_tex);
GLuint uniform_diff_flow_tex;
GLuint uniform_equation_red_tex, uniform_equation_black_tex;
GLuint uniform_diffusivity_tex;
- GLuint uniform_phase, uniform_zero_diff_flow;
+ GLuint uniform_phase, uniform_num_nonzero_phases;
};
SOR::SOR()
uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex");
uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
uniform_phase = glGetUniformLocation(sor_program, "phase");
- uniform_zero_diff_flow = glGetUniformLocation(sor_program, "zero_diff_flow");
+ uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases");
}
void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer)
bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler);
bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler);
- glProgramUniform1i(sor_program, uniform_zero_diff_flow, zero_diff_flow);
+ if (!zero_diff_flow) {
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
+ }
// NOTE: We bind to the texture we are rendering from, but we never write any value
// that we read in the same shader pass (we call discard for red values when we compute
for (int i = 0; i < num_iterations; ++i) {
{
ScopedTimer timer("Red pass", sor_timer);
+ if (zero_diff_flow && i == 0) {
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0);
+ }
glProgramUniform1i(sor_program, uniform_phase, 0);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
glTextureBarrier();
{
ScopedTimer timer("Black pass", sor_timer);
if (zero_diff_flow && i == 0) {
- // Not zero anymore.
- glProgramUniform1i(sor_program, uniform_zero_diff_flow, 0);
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1);
}
glProgramUniform1i(sor_program, uniform_phase, 1);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ if (zero_diff_flow && i == 0) {
+ glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
+ }
if (i != num_iterations - 1) {
glTextureBarrier();
}
public:
GLuint get_texture(GLenum format, GLuint width, GLuint height);
void release_texture(GLuint tex_num);
+ GLuint get_renderbuffer(GLenum format, GLuint width, GLuint height);
+ void release_renderbuffer(GLuint tex_num);
private:
struct Texture {
GLenum format;
GLuint width, height;
bool in_use = false;
+ bool is_renderbuffer = false;
};
vector<Texture> textures;
};
glBindVertexArray(vao);
- ScopedTimer total_timer("Total", &timers);
+ ScopedTimer total_timer("Compute flow", &timers);
for (int level = coarsest_level; level >= int(finest_level); --level) {
char timer_name[256];
snprintf(timer_name, sizeof(timer_name), "Level %d (%d x %d)", level, width >> level, height >> level);
}
total_timer.end();
- timers.print();
+ if (!in_warmup) {
+ timers.print();
+ }
// Scale up the flow to the final size (if needed).
if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
Splat();
// alpha is the time of the interpolated frame (0..1).
- void exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint flow_tex, GLuint depth_tex, int width, int height, float alpha);
+ void exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha);
private:
PersistentFBOSetWithDepth<1> fbos;
uniform_inv_flow_size = glGetUniformLocation(splat_program, "inv_flow_size");
}
-void Splat::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint flow_tex, GLuint depth_tex, int width, int height, float alpha)
+void Splat::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLuint backward_flow_tex, GLuint flow_tex, GLuint depth_rb, int width, int height, float alpha)
{
glUseProgram(splat_program);
glEnable(GL_DEPTH_TEST);
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_tex, flow_tex);
+ fbos.render_to(depth_rb, flow_tex);
// Evidently NVIDIA doesn't use fast clears for glClearTexImage, so clear now that
// we've got it bound.
// Output will be in flow_tex, temp_tex[0, 1, 2], representing the filling
// from the down, left, right and up, respectively. Use HoleBlend to merge
// them into one.
- void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
+ void exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height);
private:
PersistentFBOSetWithDepth<1> fbos;
uniform_sample_offset = glGetUniformLocation(fill_program, "sample_offset");
}
-void HoleFill::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
+void HoleFill::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
{
glUseProgram(fill_program);
glEnable(GL_DEPTH_TEST);
glDepthFunc(GL_LESS); // Only update the values > 0.999f (ie., only invalid pixels).
- fbos.render_to(depth_tex, flow_tex); // NOTE: Reading and writing to the same texture.
+ fbos.render_to(depth_rb, flow_tex); // NOTE: Reading and writing to the same texture.
// Fill holes from the left, by shifting 1, 2, 4, 8, etc. pixels to the right.
for (int offs = 1; offs < width; offs *= 2) {
public:
HoleBlend();
- void exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height);
+ void exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height);
private:
PersistentFBOSetWithDepth<1> fbos;
uniform_sample_offset = glGetUniformLocation(blend_program, "sample_offset");
}
-void HoleBlend::exec(GLuint flow_tex, GLuint depth_tex, GLuint temp_tex[3], int width, int height)
+void HoleBlend::exec(GLuint flow_tex, GLuint depth_rb, GLuint temp_tex[3], int width, int height)
{
glUseProgram(blend_program);
glEnable(GL_DEPTH_TEST);
glDepthFunc(GL_LEQUAL); // Skip over all of the pixels that were never holes to begin with.
- fbos.render_to(depth_tex, flow_tex); // NOTE: Reading and writing to the same texture.
+ fbos.render_to(depth_rb, flow_tex); // NOTE: Reading and writing to the same texture.
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
{
GPUTimers timers;
- ScopedTimer total_timer("Total", &timers);
+ ScopedTimer total_timer("Interpolate", &timers);
glBindVertexArray(vao);
int flow_height = height >> flow_level;
GLuint flow_tex = pool.get_texture(GL_RG16F, flow_width, flow_height);
- GLuint depth_tex = pool.get_texture(GL_DEPTH_COMPONENT32F, flow_width, flow_height); // Used for ranking flows.
+ GLuint depth_rb = pool.get_renderbuffer(GL_DEPTH_COMPONENT32F, flow_width, flow_height); // Used for ranking flows.
{
ScopedTimer timer("Splat", &total_timer);
- splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_tex, flow_width, flow_height, alpha);
+ splat.exec(tex0_view, tex1_view, forward_flow_tex, backward_flow_tex, flow_tex, depth_rb, flow_width, flow_height, alpha);
}
glDeleteTextures(1, &tex0_view);
glDeleteTextures(1, &tex1_view);
{
ScopedTimer timer("Fill holes", &total_timer);
- hole_fill.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
- hole_blend.exec(flow_tex, depth_tex, temp_tex, flow_width, flow_height);
+ hole_fill.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
+ hole_blend.exec(flow_tex, depth_rb, temp_tex, flow_width, flow_height);
}
pool.release_texture(temp_tex[0]);
pool.release_texture(temp_tex[1]);
pool.release_texture(temp_tex[2]);
- pool.release_texture(depth_tex);
+ pool.release_renderbuffer(depth_rb);
GLuint output_tex = pool.get_texture(GL_RGBA8, width, height);
{
}
pool.release_texture(flow_tex);
total_timer.end();
- timers.print();
+ if (!in_warmup) {
+ timers.print();
+ }
return output_tex;
}
GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height)
{
for (Texture &tex : textures) {
- if (!tex.in_use && tex.format == format &&
+ 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.width = width;
tex.height = height;
tex.in_use = true;
+ tex.is_renderbuffer = false;
+ 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;
+ }
+ }
+
+ Texture tex;
+ glCreateRenderbuffers(1, &tex.tex_num);
+ glNamedRenderbufferStorage(tex.tex_num, format, width, height);
+
+ tex.format = format;
+ tex.width = width;
+ tex.height = height;
+ tex.in_use = true;
+ tex.is_renderbuffer = true;
textures.push_back(tex);
return tex.tex_num;
}
void TexturePool::release_texture(GLuint tex_num)
{
for (Texture &tex : textures) {
- if (tex.tex_num == tex_num) {
+ if (!tex.is_renderbuffer && tex.tex_num == tex_num) {
+ assert(tex.in_use);
+ tex.in_use = false;
+ return;
+ }
+ }
+ assert(false);
+}
+
+void TexturePool::release_renderbuffer(GLuint tex_num)
+{
+ for (Texture &tex : textures) {
+ if (tex.is_renderbuffer && tex.tex_num == tex_num) {
assert(tex.in_use);
tex.in_use = false;
return;
glGenerateTextureMipmap(tex1_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(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
+ compute_flow.release_texture(final_tex);
+ }
+ in_warmup = false;
+ }
+
GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
gray.exec(tex1, tex1_gray, width1, height1);
glGenerateTextureMipmap(tex1_gray);
+ if (enable_warmup) {
+ in_warmup = true;
+ for (int i = 0; i < 10; ++i) {
+ GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
+ GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
+ GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, 0.5f);
+ compute_flow.release_texture(forward_flow_tex);
+ compute_flow.release_texture(backward_flow_tex);
+ interpolate.release_texture(interpolated_tex);
+ }
+ in_warmup = false;
+ }
+
GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
{ "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 }
+ { "interpolate", no_argument, 0, 1002 },
+ { "warmup", no_argument, 0, 1004 }
};
for ( ;; ) {
case 1003:
detailed_timing = true;
break;
+ case 1004:
+ enable_warmup = true;
+ break;
default:
fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
exit(1);