#include <algorithm>
#include <memory>
+#include <vector>
#define BUFFER_OFFSET(i) ((char *)nullptr + (i))
class MotionSearch {
public:
MotionSearch();
- void exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int width_patches, int height_patches);
+ void exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_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);
private:
GLuint motion_vs_obj;
GLuint motion_search_program;
GLuint motion_search_vao;
- GLuint uniform_image_size, uniform_inv_image_size;
+ GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size;
GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
};
uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
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_image0_tex = glGetUniformLocation(motion_search_program, "image0_tex");
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");
}
-void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GLuint flow_tex, GLuint flow_out_tex, int level_width, int level_height, int width_patches, int height_patches)
+void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_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)
{
glUseProgram(motion_search_program);
glProgramUniform2f(motion_search_program, uniform_image_size, level_width, level_height);
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);
GLuint flow_fbo; // TODO: cleanup
glCreateFramebuffers(1, &flow_fbo);
float patch_spacing_x = float(level_width - patch_size_pixels) / (width_patches - 1);
float patch_spacing_y = float(level_height - patch_size_pixels) / (height_patches - 1);
+ if (width_patches == 1) patch_spacing_x = 0.0f; // Avoid infinities.
+ if (height_patches == 1) patch_spacing_y = 0.0f;
glProgramUniform2f(densify_program, uniform_patch_spacing,
patch_spacing_x / level_width,
patch_spacing_y / level_height);
// I_0 and I_w. The prewarping is what enables us to solve the variational
// flow for du,dv instead of u,v.
//
+// Also calculates the normalized flow, ie. divides by z (this is needed because
+// Densify works by additive blending) and multiplies by the image size.
+//
// See variational_refinement.txt for more information.
class Prewarp {
public:
Prewarp();
- void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height);
+ void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint normalized_flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height);
private:
GLuint prewarp_vs_obj;
GLuint prewarp_vao;
GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
+ GLuint uniform_image_size;
};
Prewarp::Prewarp()
uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
+
+ uniform_image_size = glGetUniformLocation(prewarp_program, "image_size");
}
-void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height)
+void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I_tex, GLuint I_t_tex, GLuint normalized_flow_tex, int level_width, int level_height)
{
glUseProgram(prewarp_program);
bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
+ glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
+
GLuint prewarp_fbo; // TODO: cleanup
glCreateFramebuffers(1, &prewarp_fbo);
- GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
- glNamedFramebufferDrawBuffers(prewarp_fbo, 2, bufs);
+ GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1, GL_COLOR_ATTACHMENT2 };
+ glNamedFramebufferDrawBuffers(prewarp_fbo, 3, bufs);
glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
+ glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT2, normalized_flow_tex, 0);
glViewport(0, 0, level_width, level_height);
glDisable(GL_BLEND);
glViewport(0, 0, level_width, level_height);
- // Make sure the smoothness on the right and upper borders is zero.
- // We could have done this by making a (W-1)x(H-1) texture instead
- // (we're sampling smoothness with all-zero border color), but we'd
- // have to adjust the sampling coordinates, which is annoying.
- glScissor(0, 0, level_width - 1, level_height - 1);
- glEnable(GL_SCISSOR_TEST);
-
glDisable(GL_BLEND);
glBindVertexArray(smoothness_vao);
glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
- glDisable(GL_SCISSOR_TEST);
+ // Make sure the smoothness on the right and upper borders is zero.
+ // We could have done this by making (W-1)xH and Wx(H-1) textures instead
+ // (we're sampling smoothness with all-zero border color), but we'd
+ // have to adjust the sampling coordinates, which is annoying.
+ glClearTexSubImage(smoothness_x_tex, 0, level_width - 1, 0, 0, 1, level_height, 1, GL_RED, GL_FLOAT, nullptr);
+ glClearTexSubImage(smoothness_y_tex, 0, 0, level_height - 1, 0, level_width, 1, 1, GL_RED, GL_FLOAT, nullptr);
}
// Set up the equations set (two equations in two unknowns, per pixel).
GLuint uniform_diff_flow_tex, uniform_flow_tex;
GLuint uniform_beta_0_tex;
GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
-
};
SetupEquations::SetupEquations()
}
// Simply add the differential flow found by the variational refinement to the base flow.
-// The output is in diff_flow_tex; we don't need to make a new texture.
+// The output is in base_flow_tex; we don't need to make a new texture.
class AddBaseFlow {
public:
AddBaseFlow();
GLuint add_flow_program;
GLuint add_flow_vao;
- GLuint uniform_base_flow_tex;
+ GLuint uniform_diff_flow_tex;
};
AddBaseFlow::AddBaseFlow()
glEnableVertexArrayAttrib(add_flow_vao, position_attrib);
glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
- uniform_base_flow_tex = glGetUniformLocation(add_flow_program, "base_flow_tex");
+ uniform_diff_flow_tex = glGetUniformLocation(add_flow_program, "diff_flow_tex");
}
void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height)
{
glUseProgram(add_flow_program);
- bind_sampler(add_flow_program, uniform_base_flow_tex, 0, base_flow_tex, nearest_sampler);
+ bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
GLuint add_flow_fbo; // TODO: cleanup
glCreateFramebuffers(1, &add_flow_fbo);
- glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0);
+ glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, base_flow_tex, 0);
glViewport(0, 0, level_width, level_height);
glEnable(GL_BLEND);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
}
+class GPUTimers {
+public:
+ void print();
+ pair<GLuint, GLuint> begin_timer(const string &name, int level);
+
+private:
+ struct Timer {
+ string name;
+ int level;
+ pair<GLuint, GLuint> query;
+ };
+ vector<Timer> timers;
+};
+
+pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
+{
+ GLuint queries[2];
+ glGenQueries(2, queries);
+ glQueryCounter(queries[0], GL_TIMESTAMP);
+
+ Timer timer;
+ timer.name = name;
+ timer.level = level;
+ timer.query.first = queries[0];
+ timer.query.second = queries[1];
+ timers.push_back(timer);
+ return timer.query;
+}
+
+void GPUTimers::print()
+{
+ for (const Timer &timer : timers) {
+ // NOTE: This makes the CPU wait for the GPU.
+ GLuint64 time_start, time_end;
+ glGetQueryObjectui64v(timer.query.first, GL_QUERY_RESULT, &time_start);
+ glGetQueryObjectui64v(timer.query.second, GL_QUERY_RESULT, &time_end);
+ //fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
+ for (int i = 0; i < timer.level * 2; ++i) {
+ fprintf(stderr, " ");
+ }
+ fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), (time_end - time_start) / 1e6);
+ }
+}
+
+// A simple RAII class for timing until the end of the scope.
+class ScopedTimer {
+public:
+ ScopedTimer(const string &name, GPUTimers *timers)
+ : timers(timers), level(0)
+ {
+ query = timers->begin_timer(name, level);
+ }
+
+ ScopedTimer(const string &name, ScopedTimer *parent_timer)
+ : timers(parent_timer->timers),
+ level(parent_timer->level + 1)
+ {
+ query = timers->begin_timer(name, level);
+ }
+
+ ~ScopedTimer()
+ {
+ end();
+ }
+
+ void end()
+ {
+ if (!ended) {
+ glQueryCounter(query.second, GL_TIMESTAMP);
+ ended = true;
+ }
+ }
+
+private:
+ GPUTimers *timers;
+ int level;
+ pair<GLuint, GLuint> query;
+ bool ended = false;
+};
+
int main(void)
{
if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
GLuint initial_flow_tex;
glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
+ int prev_level_width = 1, prev_level_height = 1;
GLuint prev_level_flow_tex = initial_flow_tex;
glGenQueries(1, &query);
glBeginQuery(GL_TIME_ELAPSED, query);
+ GPUTimers timers;
+
+ ScopedTimer total_timer("Total", &timers);
for (int level = coarsest_level; level >= int(finest_level); --level) {
+ char timer_name[256];
+ snprintf(timer_name, sizeof(timer_name), "Level %d", 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);
glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
// Find the derivative.
- sobel.exec(tex0_view, grad0_tex, level_width, level_height);
+ {
+ ScopedTimer timer("Sobel", &level_timer);
+ sobel.exec(tex0_view, grad0_tex, level_width, level_height);
+ }
// Motion search to find the initial flow. We use the flow from the previous
// level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
glTextureStorage2D(flow_out_tex, 1, GL_RGB16F, width_patches, height_patches);
// And draw.
- motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, width_patches, height_patches);
+ {
+ ScopedTimer timer("Motion search", &level_timer);
+ motion_search.exec(tex0_view, tex1_view, grad0_tex, prev_level_flow_tex, flow_out_tex, level_width, level_height, prev_level_width, prev_level_height, width_patches, height_patches);
+ }
// Densification.
GLuint dense_flow_tex;
glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
+ glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
// And draw.
- densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
+ {
+ ScopedTimer timer("Densification", &level_timer);
+ densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
+ }
// Everything below here in the loop belongs to variational refinement.
- // It is not done yet.
-
- // Prewarping; create I and I_t.
- GLuint I_tex, I_t_tex;
+ ScopedTimer varref_timer("Variational refinement", &level_timer);
+
+ // Prewarping; create I and I_t, and a normalized base flow (so we don't
+ // have to normalize it over and over again, and also save some bandwidth).
+ //
+ // During the entire rest of the variational refinement, flow will be measured
+ // in pixels, not 0..1 normalized OpenGL texture coordinates.
+ // This is because variational refinement depends so heavily on derivatives,
+ // which are measured in intensity levels per pixel.
+ GLuint I_tex, I_t_tex, base_flow_tex;
glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
glCreateTextures(GL_TEXTURE_2D, 1, &I_t_tex);
+ glCreateTextures(GL_TEXTURE_2D, 1, &base_flow_tex);
glTextureStorage2D(I_tex, 1, GL_R16F, level_width, level_height);
glTextureStorage2D(I_t_tex, 1, GL_R16F, level_width, level_height);
- prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, level_width, level_height);
+ glTextureStorage2D(base_flow_tex, 1, GL_RG16F, level_width, level_height);
+ {
+ ScopedTimer timer("Prewarping", &varref_timer);
+ prewarp.exec(tex0_view, tex1_view, dense_flow_tex, I_tex, I_t_tex, base_flow_tex, level_width, level_height);
+ }
// 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
glCreateTextures(GL_TEXTURE_2D, 1, &beta_0_tex);
glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
glTextureStorage2D(beta_0_tex, 1, GL_R16F, level_width, level_height);
- derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
+ {
+ ScopedTimer timer("First derivatives", &varref_timer);
+ derivatives.exec(I_tex, I_x_y_tex, beta_0_tex, level_width, level_height);
+ }
// We need somewhere to store du and dv (the flow increment, relative
// to the non-refined base flow u0 and v0). It starts at zero.
GLuint du_dv_tex;
glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
glTextureStorage2D(du_dv_tex, 1, GL_RG16F, level_width, level_height);
+ glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
// And for smoothness.
GLuint smoothness_x_tex, smoothness_y_tex;
for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
// Calculate the smoothness terms between the neighboring pixels,
// both in x and y direction.
- compute_smoothness.exec(dense_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
+ {
+ ScopedTimer timer("Compute smoothness", &varref_timer);
+ compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height);
+ }
// Set up the 2x2 equation system for each pixel.
- setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, dense_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height);
+ {
+ ScopedTimer timer("Set up equations", &varref_timer);
+ setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height);
+ }
// Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
// Note that these are to/from the same texture.
- sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
+ {
+ ScopedTimer timer("SOR", &varref_timer);
+ sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5);
+ }
}
// 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.
- add_base_flow.exec(dense_flow_tex, du_dv_tex, level_width, level_height);
+ // You can comment out this prat if you wish to test disabling of the variational refinement.
+ {
+ ScopedTimer timer("Add differential flow", &varref_timer);
+ add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
+ }
- prev_level_flow_tex = du_dv_tex;
+ prev_level_flow_tex = base_flow_tex;
+ prev_level_width = level_width;
+ prev_level_height = level_height;
}
- glEndQuery(GL_TIME_ELAPSED);
+ total_timer.end();
- GLint available;
- do {
- glGetQueryObjectiv(query, GL_QUERY_RESULT_AVAILABLE, &available);
- usleep(1000);
- } while (!available);
- GLuint64 time_elapsed;
- glGetQueryObjectui64v(query, GL_QUERY_RESULT, &time_elapsed);
- fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
+ timers.print();
int level_width = WIDTH >> finest_level;
int level_height = HEIGHT >> finest_level;
- unique_ptr<float[]> dense_flow(new float[level_width * level_height * 3]);
- glGetTextureImage(prev_level_flow_tex, 0, GL_RGB, GL_FLOAT, level_width * level_height * 3 * sizeof(float), dense_flow.get());
+ unique_ptr<float[]> dense_flow(new float[level_width * level_height * 2]);
+ glGetTextureImage(prev_level_flow_tex, 0, GL_RG, GL_FLOAT, level_width * level_height * 2 * sizeof(float), dense_flow.get());
FILE *fp = fopen("flow.ppm", "wb");
FILE *flowfp = fopen("flow.flo", "wb");
for (unsigned y = 0; y < unsigned(level_height); ++y) {
int yy = level_height - y - 1;
for (unsigned x = 0; x < unsigned(level_width); ++x) {
- float du = dense_flow[(yy * level_width + x) * 3 + 0];
- float dv = dense_flow[(yy * level_width + x) * 3 + 1];
- float w = dense_flow[(yy * level_width + x) * 3 + 2];
+ float du = dense_flow[(yy * level_width + x) * 2 + 0];
+ float dv = dense_flow[(yy * level_width + x) * 2 + 1];
- du = (du / w) * level_width;
- dv = (-dv / w) * level_height;
+ dv = -dv;
fwrite(&du, 4, 1, flowfp);
fwrite(&dv, 4, 1, flowfp);