#include <assert.h>
#include <stdio.h>
+#include <unistd.h>
#include "util.h"
constexpr unsigned patch_size_pixels = 12;
// Some global OpenGL objects.
-GLuint nearest_sampler, linear_sampler, mipmap_sampler;
+GLuint nearest_sampler, linear_sampler;
GLuint vertex_vbo;
string read_file(const string &filename)
return vbo;
}
-void bind_sampler(GLuint program, const char *uniform_name, GLuint texture_unit, GLuint tex, GLuint sampler)
+void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint tex, GLuint sampler)
{
- GLint location = glGetUniformLocation(program, uniform_name);
if (location == -1) {
return;
}
GLuint sobel_fs_obj;
GLuint sobel_program;
GLuint sobel_vao;
+
+ GLuint uniform_tex, uniform_image_size;
};
Sobel::Sobel()
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));
+ uniform_tex = glGetUniformLocation(sobel_program, "tex");
}
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);
+ glProgramUniform1i(sobel_program, uniform_tex, 0);
GLuint grad0_fbo; // TODO: cleanup
glCreateFramebuffers(1, &grad0_fbo);
GLuint motion_fs_obj;
GLuint motion_search_program;
GLuint motion_search_vao;
+
+ GLuint uniform_image_size, uniform_inv_image_size;
+ GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
};
MotionSearch::MotionSearch()
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));
+ uniform_image_size = glGetUniformLocation(motion_search_program, "image_size");
+ uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_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)
{
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);
+ bind_sampler(motion_search_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
+ bind_sampler(motion_search_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
+ bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, nearest_sampler);
+ bind_sampler(motion_search_program, uniform_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);
+ 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);
GLuint flow_fbo; // TODO: cleanup
glCreateFramebuffers(1, &flow_fbo);
GLuint densify_fs_obj;
GLuint densify_program;
GLuint densify_vao;
+
+ GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
+ GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
};
Densify::Densify()
GLint position_attrib = glGetAttribLocation(densify_program, "position");
glEnableVertexArrayAttrib(densify_vao, position_attrib);
glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+
+ uniform_width_patches = glGetUniformLocation(densify_program, "width_patches");
+ uniform_patch_size = glGetUniformLocation(densify_program, "patch_size");
+ uniform_patch_spacing = glGetUniformLocation(densify_program, "patch_spacing");
+ uniform_image0_tex = glGetUniformLocation(densify_program, "image0_tex");
+ uniform_image1_tex = glGetUniformLocation(densify_program, "image1_tex");
+ uniform_flow_tex = glGetUniformLocation(densify_program, "flow_tex");
}
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);
+ bind_sampler(densify_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
+ bind_sampler(densify_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
+ bind_sampler(densify_program, uniform_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"),
+ glProgramUniform1i(densify_program, uniform_width_patches, width_patches);
+ glProgramUniform2f(densify_program, uniform_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"),
+ glProgramUniform2f(densify_program, uniform_patch_spacing,
patch_spacing_x / level_width,
patch_spacing_y / level_height);
glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
}
+// Warp I_1 to I_w, and then compute the mean (I) and difference (I_t) of
+// I_0 and I_w. The prewarping is what enables us to solve the variational
+// flow for du,dv instead of u,v.
+//
+// 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);
+
+private:
+ GLuint prewarp_vs_obj;
+ GLuint prewarp_fs_obj;
+ GLuint prewarp_program;
+ GLuint prewarp_vao;
+
+ GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
+};
+
+Prewarp::Prewarp()
+{
+ prewarp_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+ prewarp_fs_obj = compile_shader(read_file("prewarp.frag"), GL_FRAGMENT_SHADER);
+ prewarp_program = link_program(prewarp_vs_obj, prewarp_fs_obj);
+
+ // Set up the VAO containing all the required position/texcoord data.
+ glCreateVertexArrays(1, &prewarp_vao);
+ glBindVertexArray(prewarp_vao);
+ glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
+
+ GLint position_attrib = glGetAttribLocation(prewarp_program, "position");
+ glEnableVertexArrayAttrib(prewarp_vao, position_attrib);
+ glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+
+ uniform_image0_tex = glGetUniformLocation(prewarp_program, "image0_tex");
+ uniform_image1_tex = glGetUniformLocation(prewarp_program, "image1_tex");
+ uniform_flow_tex = glGetUniformLocation(prewarp_program, "flow_tex");
+}
+
+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)
+{
+ glUseProgram(prewarp_program);
+
+ bind_sampler(prewarp_program, uniform_image0_tex, 0, tex0_view, nearest_sampler);
+ bind_sampler(prewarp_program, uniform_image1_tex, 1, tex1_view, linear_sampler);
+ bind_sampler(prewarp_program, uniform_flow_tex, 2, flow_tex, nearest_sampler);
+
+ GLuint prewarp_fbo; // TODO: cleanup
+ glCreateFramebuffers(1, &prewarp_fbo);
+ GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
+ glNamedFramebufferDrawBuffers(prewarp_fbo, 2, bufs);
+ glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
+ glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
+
+ glViewport(0, 0, level_width, level_height);
+ glDisable(GL_BLEND);
+ glBindVertexArray(prewarp_vao);
+ glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+}
+
+// From I, calculate the partial derivatives I_x and I_y. We use a four-tap
+// central difference filter, since apparently, that's tradition (I haven't
+// measured quality versus a more normal 0.5 (I[x+1] - I[x-1]).)
+// The coefficients come from
+//
+// https://en.wikipedia.org/wiki/Finite_difference_coefficient
+class Derivatives {
+public:
+ Derivatives();
+ void exec(GLuint input_tex, GLuint output_tex, int level_width, int level_height);
+
+private:
+ GLuint derivatives_vs_obj;
+ GLuint derivatives_fs_obj;
+ GLuint derivatives_program;
+ GLuint derivatives_vao;
+
+ GLuint uniform_tex;
+};
+
+Derivatives::Derivatives()
+{
+ derivatives_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+ derivatives_fs_obj = compile_shader(read_file("derivatives.frag"), GL_FRAGMENT_SHADER);
+ derivatives_program = link_program(derivatives_vs_obj, derivatives_fs_obj);
+
+ // Set up the VAO containing all the required position/texcoord data.
+ glCreateVertexArrays(1, &derivatives_vao);
+ glBindVertexArray(derivatives_vao);
+ glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
+
+ GLint position_attrib = glGetAttribLocation(derivatives_program, "position");
+ glEnableVertexArrayAttrib(derivatives_vao, position_attrib);
+ glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+
+ uniform_tex = glGetUniformLocation(derivatives_program, "tex");
+}
+
+void Derivatives::exec(GLuint input_tex, GLuint output_tex, int level_width, int level_height)
+{
+ glUseProgram(derivatives_program);
+
+ bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
+
+ GLuint derivatives_fbo; // TODO: cleanup
+ glCreateFramebuffers(1, &derivatives_fbo);
+ glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, output_tex, 0);
+
+ glViewport(0, 0, level_width, level_height);
+ glDisable(GL_BLEND);
+ glBindVertexArray(derivatives_vao);
+ glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+}
+
int main(void)
{
if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
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);
-
float vertices[] = {
0.0f, 1.0f,
0.0f, 0.0f,
Sobel sobel;
MotionSearch motion_search;
Densify densify;
+ Prewarp prewarp;
+ Derivatives derivatives;
GLuint query;
glGenQueries(1, &query);
// 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);
+ 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);
// 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);
+ glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, 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.
+ // 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;
+ glCreateTextures(GL_TEXTURE_2D, 1, &I_tex);
+ glCreateTextures(GL_TEXTURE_2D, 1, &I_t_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);
+
+ // Derivatives of the images. 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.)
+
+ // Calculate I_x and I_y.
+ GLuint I_x_y_tex;
+ glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
+ glTextureStorage2D(I_x_y_tex, 1, GL_RG16F, level_width, level_height);
+ derivatives.exec(I_tex, I_x_y_tex, level_width, level_height);
prev_level_flow_tex = dense_flow_tex;
}
GLint available;
do {
glGetQueryObjectiv(query, GL_QUERY_RESULT_AVAILABLE, &available);
+ usleep(1000);
} while (!available);
GLuint64 time_elapsed;
glGetQueryObjectui64v(query, GL_QUERY_RESULT, &time_elapsed);
projects / nageru / blobdiff