GLuint sobel_program;
GLuint sobel_vao;
- GLuint uniform_tex, uniform_image_size;
+ GLuint uniform_tex;
};
Sobel::Sobel()
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, uniform_tex, 0);
+ bind_sampler(sobel_program, uniform_tex, 0, tex0_view, nearest_sampler);
glViewport(0, 0, level_width, level_height);
fbos.render_to(grad0_tex);
GLuint motion_search_program;
GLuint motion_search_vao;
- GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_flow_size, uniform_inv_prev_level_size;
+ GLuint uniform_inv_image_size, uniform_inv_prev_level_size;
GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
};
glEnableVertexArrayAttrib(motion_search_vao, position_attrib);
glVertexAttribPointer(position_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_inv_flow_size = glGetUniformLocation(motion_search_program, "inv_flow_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");
bind_sampler(motion_search_program, uniform_grad0_tex, 2, grad0_tex, zero_border_sampler);
bind_sampler(motion_search_program, uniform_flow_tex, 3, flow_tex, linear_sampler);
- 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_flow_size, 1.0f / width_patches, 1.0f / height_patches);
glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_level_height);
glViewport(0, 0, width_patches, height_patches);
GLuint densify_program;
GLuint densify_vao;
- GLuint uniform_width_patches, uniform_patch_size, uniform_patch_spacing;
+ GLuint uniform_patch_size, uniform_patch_spacing;
GLuint uniform_image0_tex, uniform_image1_tex, uniform_flow_tex;
};
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");
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, uniform_width_patches, width_patches);
glProgramUniform2f(densify_program, uniform_patch_size,
float(patch_size_pixels) / level_width,
float(patch_size_pixels) / level_height);
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, GLuint normalized_flow_tex, int level_width, int level_height)
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);
-
glViewport(0, 0, level_width, level_height);
glDisable(GL_BLEND);
glBindVertexArray(prewarp_vao);
GLuint uniform_diff_flow_tex;
GLuint uniform_equation_tex;
GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
+ GLuint uniform_phase;
};
SOR::SOR()
{
- sor_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+ sor_vs_obj = compile_shader(read_file("sor.vert"), GL_VERTEX_SHADER);
sor_fs_obj = compile_shader(read_file("sor.frag"), GL_FRAGMENT_SHADER);
sor_program = link_program(sor_vs_obj, sor_fs_obj);
uniform_equation_tex = glGetUniformLocation(sor_program, "equation_tex");
uniform_smoothness_x_tex = glGetUniformLocation(sor_program, "smoothness_x_tex");
uniform_smoothness_y_tex = glGetUniformLocation(sor_program, "smoothness_y_tex");
+ uniform_phase = glGetUniformLocation(sor_program, "phase");
}
void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, int num_iterations)
bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, zero_border_sampler);
bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
+ // 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
+ // black, and vice versa), and we have barriers between the passes, so we're fine
+ // as per the spec.
glViewport(0, 0, level_width, level_height);
glDisable(GL_BLEND);
glBindVertexArray(sor_vao);
- fbos.render_to(diff_flow_tex); // NOTE: Bind to same as we render from!
+ fbos.render_to(diff_flow_tex);
for (int i = 0; i < num_iterations; ++i) {
+ glProgramUniform1i(sor_program, uniform_phase, 0);
+ glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+ glTextureBarrier();
+ glProgramUniform1i(sor_program, uniform_phase, 1);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
if (i != num_iterations - 1) {
glTextureBarrier();
bool ended = false;
};
+class TexturePool {
+public:
+ GLuint get_texture(GLenum format, GLuint width, GLuint height);
+ void release_texture(GLuint tex_num);
+
+private:
+ struct Texture {
+ GLuint tex_num;
+ GLenum format;
+ GLuint width, height;
+ bool in_use = false;
+ };
+ vector<Texture> textures;
+};
+
class DISComputeFlow {
public:
DISComputeFlow(int width, int height);
// Returns a texture that must be released with release_texture()
// after use.
GLuint exec(GLuint tex0, GLuint tex1);
- void release_texture(GLuint tex);
+
+ void release_texture(GLuint tex) {
+ pool.release_texture(tex);
+ }
private:
int width, height;
GLuint initial_flow_tex;
+ TexturePool pool;
// The various passes.
Sobel sobel;
SOR sor;
AddBaseFlow add_base_flow;
ResizeFlow resize_flow;
-
- struct Texture {
- GLuint tex_num;
- GLenum format;
- GLuint width, height;
- bool in_use = false;
- };
- vector<Texture> textures;
-
- GLuint get_texture(GLenum format, GLuint width, GLuint height);
};
DISComputeFlow::DISComputeFlow(int width, int height)
GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
{
- for (const Texture &tex : textures) {
- assert(!tex.in_use);
- }
-
int prev_level_width = 1, prev_level_height = 1;
GLuint prev_level_flow_tex = initial_flow_tex;
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 have patches at least every Nth pixel, e.g. for width=9
+ // and patch_spacing=3 (the default), we put out patch centers in
+ // x=0, x=3, x=6, x=9, which is four patches. The fragment shader will
+ // lock all the centers to integer coordinates if needed.
+ int width_patches = 1 + ceil(level_width / patch_spacing_pixels);
+ int height_patches = 1 + ceil(level_height / 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);
// Create a new texture; we could be fancy and render use a multi-level
// texture, but meh.
- GLuint grad0_tex = get_texture(GL_RG16F, level_width, level_height);
+ GLuint grad0_tex = pool.get_texture(GL_RG16F, level_width, level_height);
// Find the derivative.
{
// level (sampled bilinearly; no fancy tricks) as a guide, then search from there.
// Create an output flow texture.
- GLuint flow_out_tex = get_texture(GL_RGB16F, width_patches, height_patches);
+ GLuint flow_out_tex = pool.get_texture(GL_RGB16F, width_patches, height_patches);
// And draw.
{
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);
}
- release_texture(grad0_tex);
+ pool.release_texture(grad0_tex);
// Densification.
// Set up an output texture (initially zero).
- GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height);
+ GLuint dense_flow_tex = pool.get_texture(GL_RGB16F, level_width, level_height);
glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
// And draw.
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);
}
- release_texture(flow_out_tex);
+ pool.release_texture(flow_out_tex);
// Everything below here in the loop belongs to variational refinement.
ScopedTimer varref_timer("Variational refinement", &level_timer);
// 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 = get_texture(GL_R16F, level_width, level_height);
- GLuint I_t_tex = get_texture(GL_R16F, level_width, level_height);
- GLuint base_flow_tex = get_texture(GL_RG16F, level_width, level_height);
+ GLuint I_tex = pool.get_texture(GL_R16F, level_width, level_height);
+ GLuint I_t_tex = pool.get_texture(GL_R16F, level_width, level_height);
+ GLuint base_flow_tex = pool.get_texture(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);
}
- release_texture(dense_flow_tex);
+ pool.release_texture(dense_flow_tex);
+ glDeleteTextures(1, &tex0_view);
+ glDeleteTextures(1, &tex1_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 = get_texture(GL_RG16F, level_width, level_height);
- GLuint beta_0_tex = get_texture(GL_R16F, level_width, level_height);
+ GLuint I_x_y_tex = pool.get_texture(GL_RG16F, level_width, level_height);
+ GLuint beta_0_tex = pool.get_texture(GL_R16F, 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);
}
- release_texture(I_tex);
+ 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 starts at zero.
- GLuint du_dv_tex = get_texture(GL_RG16F, level_width, level_height);
+ GLuint du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
glClearTexImage(du_dv_tex, 0, GL_RG, GL_FLOAT, nullptr);
// And for smoothness.
- GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height);
- GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height);
+ GLuint smoothness_x_tex = pool.get_texture(GL_R16F, level_width, level_height);
+ GLuint smoothness_y_tex = pool.get_texture(GL_R16F, level_width, level_height);
// And finally for the equation set. See SetupEquations for
// the storage format.
- GLuint equation_tex = get_texture(GL_RGBA32UI, level_width, level_height);
+ GLuint equation_tex = pool.get_texture(GL_RGBA32UI, level_width, level_height);
for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
// Calculate the smoothness terms between the neighboring pixels,
}
}
- release_texture(I_t_tex);
- release_texture(I_x_y_tex);
- release_texture(beta_0_tex);
- release_texture(smoothness_x_tex);
- release_texture(smoothness_y_tex);
- release_texture(equation_tex);
+ pool.release_texture(I_t_tex);
+ pool.release_texture(I_x_y_tex);
+ pool.release_texture(beta_0_tex);
+ pool.release_texture(smoothness_x_tex);
+ pool.release_texture(smoothness_y_tex);
+ pool.release_texture(equation_tex);
// Add the differential flow found by the variational refinement to the base flow,
// giving the final flow estimate for this level.
ScopedTimer timer("Add differential flow", &varref_timer);
add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
}
- release_texture(du_dv_tex);
+ pool.release_texture(du_dv_tex);
if (prev_level_flow_tex != initial_flow_tex) {
- release_texture(prev_level_flow_tex);
+ pool.release_texture(prev_level_flow_tex);
}
prev_level_flow_tex = base_flow_tex;
prev_level_width = level_width;
if (finest_level == 0) {
return prev_level_flow_tex;
} else {
- GLuint final_tex = get_texture(GL_RG16F, width, height);
+ GLuint final_tex = pool.get_texture(GL_RG16F, width, height);
resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width, height);
- release_texture(prev_level_flow_tex);
+ pool.release_texture(prev_level_flow_tex);
return final_tex;
}
}
-GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height)
+GLuint TexturePool::get_texture(GLenum format, GLuint width, GLuint height)
{
for (Texture &tex : textures) {
if (!tex.in_use && tex.format == format &&
return tex.tex_num;
}
-void DISComputeFlow::release_texture(GLuint tex_num)
+void TexturePool::release_texture(GLuint tex_num)
{
for (Texture &tex : textures) {
if (tex.tex_num == tex_num) {
SDL_GL_SetAttribute(SDL_GL_CONTEXT_MINOR_VERSION, 5);
// SDL_GL_SetAttribute(SDL_GL_CONTEXT_FLAGS, SDL_GL_CONTEXT_DEBUG_FLAG);
SDL_Window *window = SDL_CreateWindow("OpenGL window",
- SDL_WINDOWPOS_UNDEFINED,
- SDL_WINDOWPOS_UNDEFINED,
- 64, 64,
- SDL_WINDOW_OPENGL);
+ SDL_WINDOWPOS_UNDEFINED,
+ SDL_WINDOWPOS_UNDEFINED,
+ 64, 64,
+ SDL_WINDOW_OPENGL | SDL_WINDOW_HIDDEN);
SDL_GLContext context = SDL_GL_CreateContext(window);
assert(context != nullptr);
projects / nageru / blobdiff