#define NO_SDL_GLEXT 1
-
+
#define WIDTH 1280
#define HEIGHT 720
#include <assert.h>
#include <stdio.h>
+#include "util.h"
+
#include <algorithm>
#include <memory>
// Operating point 3 (10 Hz on CPU, excluding preprocessing).
constexpr float patch_overlap_ratio = 0.75f;
-constexpr unsigned coarsest_level = 0;
-constexpr unsigned finest_level = 0;
+constexpr unsigned coarsest_level = 5;
+constexpr unsigned finest_level = 1;
constexpr unsigned patch_size_pixels = 12;
+// Some global OpenGL objects.
+GLuint nearest_sampler, linear_sampler, mipmap_sampler;
+GLuint vertex_vbo;
+
string read_file(const string &filename)
{
FILE *fp = fopen(filename.c_str(), "r");
glBindBuffer(GL_ARRAY_BUFFER, 0);
return vbo;
-}
+}
+
+void bind_sampler(GLuint program, const char *uniform_name, GLuint texture_unit, GLuint tex, GLuint sampler)
+{
+ GLint location = glGetUniformLocation(program, uniform_name);
+ if (location == -1) {
+ return;
+ }
+
+ glBindTextureUnit(texture_unit, tex);
+ glBindSampler(texture_unit, sampler);
+ glProgramUniform1i(program, location, texture_unit);
+}
+
+// Compute gradients in every point, used for the motion search.
+// The DIS paper doesn't actually mention how these are computed,
+// but seemingly, a 3x3 Sobel operator is used here (at least in
+// later versions of the code), while a [1 -8 0 8 -1] kernel is
+// used for all the derivatives in the variational refinement part
+// (which borrows code from DeepFlow). This is inconsistent,
+// but I guess we're better off with staying with the original
+// decisions until we actually know having different ones would be better.
+class Sobel {
+public:
+ Sobel();
+ void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
+
+private:
+ GLuint sobel_vs_obj;
+ GLuint sobel_fs_obj;
+ GLuint sobel_program;
+ GLuint sobel_vao;
+};
+
+Sobel::Sobel()
+{
+ sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+ sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
+ sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
+
+ // Set up the VAO containing all the required position/texcoord data.
+ glCreateVertexArrays(1, &sobel_vao);
+ glBindVertexArray(sobel_vao);
+
+ GLint position_attrib = glGetAttribLocation(sobel_program, "position");
+ glEnableVertexArrayAttrib(sobel_vao, position_attrib);
+ glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+
+ GLint texcoord_attrib = glGetAttribLocation(sobel_program, "texcoord");
+ glEnableVertexArrayAttrib(sobel_vao, texcoord_attrib);
+ glVertexAttribPointer(texcoord_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+}
+
+void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height)
+{
+ glUseProgram(sobel_program);
+ glBindTextureUnit(0, tex0_view);
+ glBindSampler(0, nearest_sampler);
+ glProgramUniform1i(sobel_program, glGetUniformLocation(sobel_program, "tex"), 0);
+ glProgramUniform2f(sobel_program, glGetUniformLocation(sobel_program, "image_size"), level_width, level_height);
+ glProgramUniform2f(sobel_program, glGetUniformLocation(sobel_program, "inv_image_size"), 1.0f / level_width, 1.0f / level_height);
+
+ GLuint grad0_fbo; // TODO: cleanup
+ glCreateFramebuffers(1, &grad0_fbo);
+ glNamedFramebufferTexture(grad0_fbo, GL_COLOR_ATTACHMENT0, grad0_tex, 0);
+
+ glViewport(0, 0, level_width, level_height);
+ glBindFramebuffer(GL_FRAMEBUFFER, grad0_fbo);
+ glBindVertexArray(sobel_vao);
+ glUseProgram(sobel_program);
+ glDisable(GL_BLEND);
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+}
+
+// Motion search to find the initial flow. See motion_search.frag for documentation.
+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);
+
+private:
+ GLuint motion_vs_obj;
+ GLuint motion_fs_obj;
+ GLuint motion_search_program;
+ GLuint motion_search_vao;
+};
+
+MotionSearch::MotionSearch()
+{
+ 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);
+ motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
+
+ // Set up the VAO containing all the required position/texcoord data.
+ glCreateVertexArrays(1, &motion_search_vao);
+ glBindVertexArray(motion_search_vao);
+ glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
+
+ GLint position_attrib = glGetAttribLocation(motion_search_program, "position");
+ glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
+ glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+
+ GLint texcoord_attrib = glGetAttribLocation(motion_search_program, "texcoord");
+ glEnableVertexArrayAttrib(motion_search_vao, texcoord_attrib);
+ glVertexAttribPointer(texcoord_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+}
+
+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)
+{
+ glUseProgram(motion_search_program);
+
+ bind_sampler(motion_search_program, "image0_tex", 0, tex0_view, nearest_sampler);
+ bind_sampler(motion_search_program, "image1_tex", 1, tex1_view, linear_sampler);
+ bind_sampler(motion_search_program, "grad0_tex", 2, grad0_tex, nearest_sampler);
+ bind_sampler(motion_search_program, "flow_tex", 3, flow_tex, linear_sampler);
+
+ glProgramUniform2f(motion_search_program, glGetUniformLocation(motion_search_program, "image_size"), level_width, level_height);
+ glProgramUniform2f(motion_search_program, glGetUniformLocation(motion_search_program, "inv_image_size"), 1.0f / level_width, 1.0f / level_height);
+
+ GLuint flow_fbo; // TODO: cleanup
+ glCreateFramebuffers(1, &flow_fbo);
+ glNamedFramebufferTexture(flow_fbo, GL_COLOR_ATTACHMENT0, flow_out_tex, 0);
+
+ glViewport(0, 0, width_patches, height_patches);
+ glBindFramebuffer(GL_FRAMEBUFFER, flow_fbo);
+ glBindVertexArray(motion_search_vao);
+ glUseProgram(motion_search_program);
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+}
+
+// Do “densification”, ie., upsampling of the flow patches to the flow field
+// (the same size as the image at this level). We draw one quad per patch
+// over its entire covered area (using instancing in the vertex shader),
+// and then weight the contributions in the pixel shader by post-warp difference.
+// This is equation (3) in the paper.
+//
+// We accumulate the flow vectors in the R/G channels (for u/v) and the total
+// weight in the B channel. Dividing R and G by B gives the normalized values.
+class Densify {
+public:
+ Densify();
+ void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches);
+
+private:
+ GLuint densify_vs_obj;
+ GLuint densify_fs_obj;
+ GLuint densify_program;
+ GLuint densify_vao;
+};
+
+Densify::Densify()
+{
+ densify_vs_obj = compile_shader(read_file("densify.vert"), GL_VERTEX_SHADER);
+ densify_fs_obj = compile_shader(read_file("densify.frag"), GL_FRAGMENT_SHADER);
+ densify_program = link_program(densify_vs_obj, densify_fs_obj);
+
+ // Set up the VAO containing all the required position/texcoord data.
+ glCreateVertexArrays(1, &densify_vao);
+ glBindVertexArray(densify_vao);
+ glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
+
+ GLint position_attrib = glGetAttribLocation(densify_program, "position");
+ glEnableVertexArrayAttrib(densify_vao, position_attrib);
+ glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+}
+
+void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint dense_flow_tex, int level_width, int level_height, int width_patches, int height_patches)
+{
+ glUseProgram(densify_program);
+
+ bind_sampler(densify_program, "image0_tex", 0, tex0_view, nearest_sampler);
+ bind_sampler(densify_program, "image1_tex", 1, tex1_view, linear_sampler);
+ bind_sampler(densify_program, "flow_tex", 2, flow_tex, nearest_sampler);
+
+ glProgramUniform1i(densify_program, glGetUniformLocation(densify_program, "width_patches"), width_patches);
+ glProgramUniform2f(densify_program, glGetUniformLocation(densify_program, "patch_size"),
+ float(patch_size_pixels) / level_width,
+ float(patch_size_pixels) / level_height);
+
+ 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);
+ glProgramUniform2f(densify_program, glGetUniformLocation(densify_program, "patch_spacing"),
+ patch_spacing_x / level_width,
+ patch_spacing_y / level_height);
+
+ GLuint dense_flow_fbo; // TODO: cleanup
+ glCreateFramebuffers(1, &dense_flow_fbo);
+ glNamedFramebufferTexture(dense_flow_fbo, GL_COLOR_ATTACHMENT0, dense_flow_tex, 0);
+
+ glViewport(0, 0, level_width, level_height);
+ glEnable(GL_BLEND);
+ glBlendFunc(GL_ONE, GL_ONE);
+ glBindVertexArray(densify_vao);
+ glBindFramebuffer(GL_FRAMEBUFFER, dense_flow_fbo);
+ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
+}
int main(void)
{
GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
- // Load shaders.
- GLuint motion_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
- GLuint motion_fs_obj = compile_shader(read_file("motion_search.frag"), GL_FRAGMENT_SHADER);
- GLuint motion_search_program = link_program(motion_vs_obj, motion_fs_obj);
-
- GLuint sobel_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
- GLuint sobel_fs_obj = compile_shader(read_file("sobel.frag"), GL_FRAGMENT_SHADER);
- GLuint sobel_program = link_program(sobel_vs_obj, sobel_fs_obj);
-
- // Coarsest level.
- int level_width = WIDTH >> coarsest_level;
- int level_height = HEIGHT >> coarsest_level;
- float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
- int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
- int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
-
- // Compute gradients in every point, used for the motion search.
- // The DIS paper doesn't actually mention how these are computed,
- // but seemingly, a 3x3 Sobel operator is used here (at least in
- // later versions of the code), while a [1 -8 0 8 -1] kernel is
- // used for all the derivatives in the variational refinement part
- // (which borrows code from DeepFlow). This is inconsistent,
- // but I guess we're better off with staying with the original
- // decisions until we actually know having different ones would be better.
-
- // Create a new texture; we could be fancy and render use a multi-level
- // texture, but meh.
- GLuint grad_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &grad_tex);
- glTextureStorage2D(grad_tex, 1, GL_RG16F, level_width, level_height);
-
- GLuint grad_fbo;
- glCreateFramebuffers(1, &grad_fbo);
- glNamedFramebufferTexture(grad_fbo, GL_COLOR_ATTACHMENT0, grad_tex, 0);
-
- glUseProgram(sobel_program);
- glBindTextureUnit(0, tex0);
- glUniform1i(glGetUniformLocation(sobel_program, "tex"), 0);
- glUniform1f(glGetUniformLocation(sobel_program, "inv_width"), 1.0f / level_width);
- glUniform1f(glGetUniformLocation(sobel_program, "inv_height"), 1.0f / level_height);
+ // Make some samplers.
+ glCreateSamplers(1, &nearest_sampler);
+ glSamplerParameteri(nearest_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+ glSamplerParameteri(nearest_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+ glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+ glSamplerParameteri(nearest_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
+ glCreateSamplers(1, &linear_sampler);
+ glSamplerParameteri(linear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
+ glSamplerParameteri(linear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+ glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+ glSamplerParameteri(linear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+
+ glCreateSamplers(1, &mipmap_sampler);
+ glSamplerParameteri(mipmap_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
+ glSamplerParameteri(mipmap_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
+ glSamplerParameteri(mipmap_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+ glSamplerParameteri(mipmap_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
- // Set up the VAO containing all the required position/texcoord data.
- GLuint sobel_vao;
- glCreateVertexArrays(1, &sobel_vao);
- glBindVertexArray(sobel_vao);
float vertices[] = {
0.0f, 1.0f,
0.0f, 0.0f,
1.0f, 1.0f,
1.0f, 0.0f,
};
- GLuint vertex_vbo;
glCreateBuffers(1, &vertex_vbo);
glNamedBufferData(vertex_vbo, sizeof(vertices), vertices, GL_STATIC_DRAW);
glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
- int position_attrib = glGetAttribLocation(sobel_program, "position");
- glEnableVertexArrayAttrib(sobel_vao, position_attrib);
- glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+ // Initial flow is zero, 1x1.
+ GLuint initial_flow_tex;
+ glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
+ glTextureStorage2D(initial_flow_tex, 1, GL_RGB32F, 1, 1);
- int texcoord_attrib = glGetAttribLocation(sobel_program, "texcoord");
- glEnableVertexArrayAttrib(sobel_vao, texcoord_attrib);
- glVertexAttribPointer(texcoord_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+ GLuint prev_level_flow_tex = initial_flow_tex;
- glBindBuffer(GL_ARRAY_BUFFER, 0);
+ Sobel sobel;
+ MotionSearch motion_search;
+ Densify densify;
- // Now finally draw.
- glViewport(0, 0, level_width, level_height);
- glBindFramebuffer(GL_FRAMEBUFFER, grad_fbo);
- glUseProgram(sobel_program);
- glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
- glUseProgram(0);
- glBindFramebuffer(GL_FRAMEBUFFER, 0);
- glBindVertexArray(0);
+ GLuint query;
+ glGenQueries(1, &query);
+ glBeginQuery(GL_TIME_ELAPSED, query);
+
+ for (int level = coarsest_level; level >= int(finest_level); --level) {
+ int level_width = WIDTH >> level;
+ int level_height = HEIGHT >> level;
+ float patch_spacing_pixels = patch_size_pixels * (1.0f - patch_overlap_ratio);
+ int width_patches = 1 + lrintf((level_width - patch_size_pixels) / patch_spacing_pixels);
+ int height_patches = 1 + lrintf((level_height - patch_size_pixels) / patch_spacing_pixels);
+
+ // Make sure we always read from the correct level; the chosen
+ // mipmapping could otherwise be rather unpredictable, especially
+ // during motion search.
+ // TODO: create these beforehand, and stop leaking them.
+ GLuint tex0_view, tex1_view;
+ glGenTextures(1, &tex0_view);
+ glTextureView(tex0_view, GL_TEXTURE_2D, tex0, GL_R8, level, 1, 0, 1);
+ glGenTextures(1, &tex1_view);
+ glTextureView(tex1_view, GL_TEXTURE_2D, tex1, GL_R8, level, 1, 0, 1);
+
+ // Create a new texture; we could be fancy and render use a multi-level
+ // texture, but meh.
+ GLuint grad0_tex;
+ glCreateTextures(GL_TEXTURE_2D, 1, &grad0_tex);
+ glTextureStorage2D(grad0_tex, 1, GL_RG16F, level_width, level_height);
- GLuint flow_tex;
- glCreateTextures(GL_TEXTURE_2D, 1, &flow_tex);
- glTextureStorage2D(flow_tex, 1, GL_RG16F, width_patches, height_patches);
+ // Find the derivative.
+ 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.
+
+ // Create an output flow texture.
+ GLuint flow_out_tex;
+ glCreateTextures(GL_TEXTURE_2D, 1, &flow_out_tex);
+ glTextureStorage2D(flow_out_tex, 1, GL_RG16F, 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);
+
+ // Densification.
+
+ // Set up an output texture (initially zero).
+ GLuint dense_flow_tex;
+ glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
+ //glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
+ glTextureStorage2D(dense_flow_tex, 1, GL_RGBA32F, level_width, level_height);
+
+ // And draw.
+ densify.exec(tex0_view, tex1_view, flow_out_tex, dense_flow_tex, level_width, level_height, width_patches, height_patches);
+
+ // TODO: Variational refinement.
+
+ prev_level_flow_tex = dense_flow_tex;
+ }
+ glEndQuery(GL_TIME_ELAPSED);
+
+ GLint available;
+ do {
+ glGetQueryObjectiv(query, GL_QUERY_RESULT_AVAILABLE, &available);
+ } while (!available);
+ GLuint64 time_elapsed;
+ glGetQueryObjectui64v(query, GL_QUERY_RESULT, &time_elapsed);
+ fprintf(stderr, "GPU time used = %.1f ms\n", time_elapsed / 1e6);
+
+ 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());
+
+ FILE *fp = fopen("flow.ppm", "wb");
+ FILE *flowfp = fopen("flow.flo", "wb");
+ fprintf(fp, "P6\n%d %d\n255\n", level_width, level_height);
+ fprintf(flowfp, "FEIH");
+ fwrite(&level_width, 4, 1, flowfp);
+ fwrite(&level_height, 4, 1, flowfp);
+ 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];
+
+ du = (du / w) * level_width;
+ dv = (-dv / w) * level_height;
+
+ fwrite(&du, 4, 1, flowfp);
+ fwrite(&dv, 4, 1, flowfp);
+
+ uint8_t r, g, b;
+ flow2rgb(du, dv, &r, &g, &b);
+ putc(r, fp);
+ putc(g, fp);
+ putc(b, fp);
+ }
+ }
+ fclose(fp);
+ fclose(flowfp);
-// printf("%d x %d patches on this level\n", width_patches, height_patches);
+ fprintf(stderr, "err = %d\n", glGetError());
}