X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=flow.cpp;h=aedac9d5e63410f8c60a047d4903afb2722c787e;hb=c1e1b590ad5720a7459decce76d944c86e2743c3;hp=df6ac6df625d96586049234fc9c0e9094eca9017;hpb=d51c933ab4f9cbd90b4c5577d5d9ed5d89afefb4;p=nageru diff --git a/flow.cpp b/flow.cpp index df6ac6d..aedac9d 100644 --- a/flow.cpp +++ b/flow.cpp @@ -18,8 +18,10 @@ #include "util.h" #include +#include #include #include +#include #include #define BUFFER_OFFSET(i) ((char *)nullptr + (i)) @@ -42,9 +44,18 @@ bool enable_variational_refinement = true; // Just for debugging. // Some global OpenGL objects. // TODO: These should really be part of DISComputeFlow. -GLuint nearest_sampler, linear_sampler, smoothness_sampler; +GLuint nearest_sampler, linear_sampler, zero_border_sampler; GLuint vertex_vbo; +// Structures for asynchronous readback. We assume everything is the same size (and GL_RG16F). +struct ReadInProgress { + GLuint pbo; + string filename0, filename1; + string flow_filename, ppm_filename; // Either may be empty for no write. +}; +stack spare_pbos; +deque reads_in_progress; + string read_file(const string &filename) { FILE *fp = fopen(filename.c_str(), "r"); @@ -388,7 +399,7 @@ void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GL 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_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); @@ -731,8 +742,8 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler); bind_sampler(equations_program, uniform_base_flow_tex, 3, base_flow_tex, nearest_sampler); bind_sampler(equations_program, uniform_beta_0_tex, 4, beta_0_tex, nearest_sampler); - bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, smoothness_sampler); - bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, smoothness_sampler); + bind_sampler(equations_program, uniform_smoothness_x_tex, 5, smoothness_x_tex, zero_border_sampler); + bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler); glProgramUniform1f(equations_program, uniform_delta, vr_delta); glProgramUniform1f(equations_program, uniform_gamma, vr_gamma); @@ -791,8 +802,8 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_te glUseProgram(sor_program); bind_sampler(sor_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler); - bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, smoothness_sampler); - bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler); + bind_sampler(sor_program, uniform_smoothness_x_tex, 1, smoothness_x_tex, zero_border_sampler); + 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); glViewport(0, 0, level_width, level_height); @@ -1050,13 +1061,15 @@ DISComputeFlow::DISComputeFlow(int width, int height) // The smoothness is sampled so that once we get to a smoothness involving // a value outside the border, the diffusivity between the two becomes zero. - glCreateSamplers(1, &smoothness_sampler); - glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST); - glSamplerParameteri(smoothness_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST); - glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER); - glSamplerParameteri(smoothness_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER); + // Similarly, gradients are zero outside the border, since the edge is taken + // to be constant. + glCreateSamplers(1, &zero_border_sampler); + glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MIN_FILTER, GL_NEAREST); + glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_NEAREST); + glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_BORDER); + glSamplerParameteri(zero_border_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_BORDER); float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f }; - glSamplerParameterfv(smoothness_sampler, GL_TEXTURE_BORDER_COLOR, zero); + glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero); // Initial flow is zero, 1x1. glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex); @@ -1315,6 +1328,41 @@ void write_ppm(const char *filename, const float *dense_flow, unsigned width, un fclose(fp); } +void finish_one_read(GLuint width, GLuint height) +{ + assert(!reads_in_progress.empty()); + ReadInProgress read = reads_in_progress.front(); + reads_in_progress.pop_front(); + + unique_ptr flow(new float[width * height * 2]); + void *buf = glMapNamedBufferRange(read.pbo, 0, width * height * 2 * sizeof(float), GL_MAP_READ_BIT); // Blocks if the read isn't done yet. + memcpy(flow.get(), buf, width * height * 2 * sizeof(float)); + glUnmapNamedBuffer(read.pbo); + spare_pbos.push(read.pbo); + + flip_coordinate_system(flow.get(), width, height); + if (!read.flow_filename.empty()) { + write_flow(read.flow_filename.c_str(), flow.get(), width, height); + fprintf(stderr, "%s %s -> %s\n", read.filename0.c_str(), read.filename1.c_str(), read.flow_filename.c_str()); + } + if (!read.ppm_filename.empty()) { + write_ppm(read.ppm_filename.c_str(), flow.get(), width, height); + } +} + +void schedule_read(GLuint tex, GLuint width, GLuint height, const char *filename0, const char *filename1, const char *flow_filename, const char *ppm_filename) +{ + if (spare_pbos.empty()) { + finish_one_read(width, height); + } + assert(!spare_pbos.empty()); + reads_in_progress.emplace_back(ReadInProgress{ spare_pbos.top(), filename0, filename1, flow_filename, ppm_filename }); + glBindBuffer(GL_PIXEL_PACK_BUFFER, spare_pbos.top()); + spare_pbos.pop(); + glGetTextureImage(tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), nullptr); + glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); +} + int main(int argc, char **argv) { static const option long_options[] = { @@ -1378,7 +1426,6 @@ int main(int argc, char **argv) const char *filename0 = argc >= (optind + 1) ? argv[optind] : "test1499.png"; const char *filename1 = argc >= (optind + 2) ? argv[optind + 1] : "test1500.png"; const char *flow_filename = argc >= (optind + 3) ? argv[optind + 2] : "flow.flo"; - fprintf(stderr, "%s %s -> %s\n", filename0, filename1, flow_filename); // Load pictures. unsigned width1, height1, width2, height2; @@ -1391,6 +1438,14 @@ int main(int argc, char **argv) exit(1); } + // Set up some PBOs to do asynchronous readback. + GLuint pbos[5]; + glCreateBuffers(5, pbos); + for (int i = 0; i < 5; ++i) { + glNamedBufferData(pbos[i], width1 * height1 * 2 * sizeof(float), nullptr, GL_STREAM_READ); + spare_pbos.push(pbos[i]); + } + // FIXME: Should be part of DISComputeFlow (but needs to be initialized // before all the render passes). float vertices[] = { @@ -1406,17 +1461,9 @@ int main(int argc, char **argv) DISComputeFlow compute_flow(width1, height1); GLuint final_tex = compute_flow.exec(tex0, tex1); - unique_ptr dense_flow(new float[width1 * height1 * 2]); - glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width1 * height1 * 2 * sizeof(float), dense_flow.get()); - + schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm"); compute_flow.release_texture(final_tex); - flip_coordinate_system(dense_flow.get(), width1, height1); - write_flow(flow_filename, dense_flow.get(), width1, height1); - write_ppm("flow.ppm", dense_flow.get(), width1, height1); - - dense_flow.reset(); - // See if there are more flows on the command line (ie., more than three arguments), // and if so, process them. int num_flows = (argc - optind) / 3; @@ -1424,8 +1471,6 @@ int main(int argc, char **argv) const char *filename0 = argv[optind + i * 3 + 0]; const char *filename1 = argv[optind + i * 3 + 1]; const char *flow_filename = argv[optind + i * 3 + 2]; - fprintf(stderr, "%s %s -> %s\n", filename0, filename1, flow_filename); - GLuint width, height; GLuint tex0 = load_texture(filename0, &width, &height); if (width != width1 || height != height1) { @@ -1442,15 +1487,11 @@ int main(int argc, char **argv) } GLuint final_tex = compute_flow.exec(tex0, tex1); - - unique_ptr dense_flow(new float[width * height * 2]); - glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width * height * 2 * sizeof(float), dense_flow.get()); - + schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, ""); compute_flow.release_texture(final_tex); - - flip_coordinate_system(dense_flow.get(), width, height); - write_flow(flow_filename, dense_flow.get(), width, height); } - fprintf(stderr, "err = %d\n", glGetError()); + while (!reads_in_progress.empty()) { + finish_one_read(width1, height1); + } }