]> git.sesse.net Git - nageru/blobdiff - flow.cpp
Give the variational refinement terms slightly less mysterious names.
[nageru] / flow.cpp
index ff4c747f6d0f678d8e3c4284b50212f919d71f3f..b6f0517ea746a3698cecee814b10ab9ea3b37dcf 100644 (file)
--- a/flow.cpp
+++ b/flow.cpp
 #include "util.h"
 
 #include <algorithm>
+#include <deque>
 #include <memory>
 #include <map>
+#include <stack>
 #include <vector>
 
 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
@@ -37,11 +39,23 @@ constexpr unsigned patch_size_pixels = 12;
 // since we have different normalizations and ranges in some cases.
 float vr_gamma = 10.0f, vr_delta = 5.0f, vr_alpha = 10.0f;
 
+bool enable_timing = true;
+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<GLuint> spare_pbos;
+deque<ReadInProgress> reads_in_progress;
+
 string read_file(const string &filename)
 {
        FILE *fp = fopen(filename.c_str(), "r");
@@ -350,7 +364,7 @@ private:
        GLuint motion_search_program;
        GLuint motion_search_vao;
 
-       GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_prev_level_size;
+       GLuint uniform_image_size, uniform_inv_image_size, uniform_inv_flow_size, uniform_inv_prev_level_size;
        GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
 };
 
@@ -371,6 +385,7 @@ MotionSearch::MotionSearch()
 
        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");
@@ -384,11 +399,12 @@ 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);
        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);
@@ -726,8 +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);
 
@@ -786,8 +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);
@@ -923,6 +939,10 @@ private:
 
 pair<GLuint, GLuint> GPUTimers::begin_timer(const string &name, int level)
 {
+       if (!enable_timing) {
+               return make_pair(0, 0);
+       }
+
        GLuint queries[2];
        glGenQueries(2, queries);
        glQueryCounter(queries[0], GL_TIMESTAMP);
@@ -974,7 +994,7 @@ public:
 
        void end()
        {
-               if (!ended) {
+               if (enable_timing && !ended) {
                        glQueryCounter(query.second, GL_TIMESTAMP);
                        ended = true;
                }
@@ -990,7 +1010,11 @@ private:
 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);
 
 private:
        int width, height;
@@ -1007,6 +1031,16 @@ private:
        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)
@@ -1027,13 +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);
@@ -1043,6 +1079,10 @@ 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;
 
@@ -1072,9 +1112,7 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
 
                // 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 grad0_tex = get_texture(GL_RG16F, level_width, level_height);
 
                // Find the derivative.
                {
@@ -1086,22 +1124,19 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
                // 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_RGB16F, width_patches, height_patches);
+               GLuint flow_out_tex = 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);
 
                // 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);
+               GLuint dense_flow_tex = get_texture(GL_RGB16F, level_width, level_height);
                glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
 
                // And draw.
@@ -1109,6 +1144,7 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
                        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);
 
                // Everything below here in the loop belongs to variational refinement.
                ScopedTimer varref_timer("Variational refinement", &level_timer);
@@ -1120,52 +1156,40 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
                // 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);
-               glTextureStorage2D(base_flow_tex, 1, GL_RG16F, level_width, level_height);
+               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);
                {
                        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);
 
                // 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, beta_0_tex;
-               glCreateTextures(GL_TEXTURE_2D, 1, &I_x_y_tex);
-               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);
+               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);
                {
                        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);
 
                // 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);
+               GLuint du_dv_tex = 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, smoothness_y_tex;
-               glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_x_tex);
-               glCreateTextures(GL_TEXTURE_2D, 1, &smoothness_y_tex);
-               glTextureStorage2D(smoothness_x_tex, 1, GL_R16F, level_width, level_height);
-               glTextureStorage2D(smoothness_y_tex, 1, GL_R16F, level_width, level_height);
+               GLuint smoothness_x_tex = get_texture(GL_R16F, level_width, level_height);
+               GLuint smoothness_y_tex = get_texture(GL_R16F, level_width, level_height);
 
                // And finally for the equation set. See SetupEquations for
                // the storage format.
-               GLuint equation_tex;
-               glCreateTextures(GL_TEXTURE_2D, 1, &equation_tex);
-               glTextureStorage2D(equation_tex, 1, GL_RGBA32UI, level_width, level_height);
+               GLuint equation_tex = 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,
@@ -1189,15 +1213,28 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
                        }
                }
 
+               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);
+
                // 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.
-               // You can comment out this part if you wish to test disabling of the variational refinement.
-               {
+               //
+               // Disabling this doesn't save any time (although we could easily make it so that
+               // it is more efficient), but it helps debug the motion search.
+               if (enable_variational_refinement) {
                        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);
 
+               if (prev_level_flow_tex != initial_flow_tex) {
+                       release_texture(prev_level_flow_tex);
+               }
                prev_level_flow_tex = base_flow_tex;
                prev_level_width = level_width;
                prev_level_height = level_height;
@@ -1210,39 +1247,155 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1)
        if (finest_level == 0) {
                return prev_level_flow_tex;
        } else {
-               GLuint final_tex;
-               glCreateTextures(GL_TEXTURE_2D, 1, &final_tex);
-               glTextureStorage2D(final_tex, 1, GL_RG16F, width, height);
+               GLuint final_tex = 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);
                return final_tex;
        }
 }
 
+GLuint DISComputeFlow::get_texture(GLenum format, GLuint width, GLuint height)
+{
+       for (Texture &tex : textures) {
+               if (!tex.in_use && tex.format == format &&
+                   tex.width == width && tex.height == height) {
+                       tex.in_use = true;
+                       return tex.tex_num;
+               }
+       }
+
+       Texture tex;
+       glCreateTextures(GL_TEXTURE_2D, 1, &tex.tex_num);
+       glTextureStorage2D(tex.tex_num, 1, format, width, height);
+       tex.format = format;
+       tex.width = width;
+       tex.height = height;
+       tex.in_use = true;
+       textures.push_back(tex);
+       return tex.tex_num;
+}
+
+void DISComputeFlow::release_texture(GLuint tex_num)
+{
+       for (Texture &tex : textures) {
+               if (tex.tex_num == tex_num) {
+                       assert(tex.in_use);
+                       tex.in_use = false;
+                       return;
+               }
+       }
+       assert(false);
+}
+
+// OpenGL uses a bottom-left coordinate system, .flo files use a top-left coordinate system.
+void flip_coordinate_system(float *dense_flow, unsigned width, unsigned height)
+{
+       for (unsigned i = 0; i < width * height; ++i) {
+               dense_flow[i * 2 + 1] = -dense_flow[i * 2 + 1];
+       }
+}
+
+void write_flow(const char *filename, const float *dense_flow, unsigned width, unsigned height)
+{
+       FILE *flowfp = fopen(filename, "wb");
+       fprintf(flowfp, "FEIH");
+       fwrite(&width, 4, 1, flowfp);
+       fwrite(&height, 4, 1, flowfp);
+       for (unsigned y = 0; y < height; ++y) {
+               int yy = height - y - 1;
+               fwrite(&dense_flow[yy * width * 2], width * 2 * sizeof(float), 1, flowfp);
+       }
+       fclose(flowfp);
+}
+
+void write_ppm(const char *filename, const float *dense_flow, unsigned width, unsigned height)
+{
+       FILE *fp = fopen(filename, "wb");
+       fprintf(fp, "P6\n%d %d\n255\n", width, height);
+       for (unsigned y = 0; y < unsigned(height); ++y) {
+               int yy = height - y - 1;
+               for (unsigned x = 0; x < unsigned(width); ++x) {
+                       float du = dense_flow[(yy * width + x) * 2 + 0];
+                       float dv = dense_flow[(yy * width + x) * 2 + 1];
+
+                       uint8_t r, g, b;
+                       flow2rgb(du, dv, &r, &g, &b);
+                       putc(r, fp);
+                       putc(g, fp);
+                       putc(b, fp);
+               }
+       }
+       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<float[]> 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[] = {
-                { "alpha", required_argument, 0, 'a' },
-                { "delta", required_argument, 0, 'd' },
-                { "gamma", required_argument, 0, 'g' }
+               { "smoothness-relative-weight", required_argument, 0, 's' },  // alpha.
+               { "intensity-relative-weight", required_argument, 0, 'i' },  // delta.
+               { "gradient-relative-weight", required_argument, 0, 'g' },  // gamma.
+               { "disable-timing", no_argument, 0, 1000 },
+               { "ignore-variational-refinement", no_argument, 0, 1001 }  // Still calculates it, just doesn't apply it.
        };
 
        for ( ;; ) {
                int option_index = 0;
-               int c = getopt_long(argc, argv, "a:d:g:", long_options, &option_index);
+               int c = getopt_long(argc, argv, "s:i:g:", long_options, &option_index);
 
                if (c == -1) {
                        break;
                }
                switch (c) {
-               case 'a':
+               case 's':
                        vr_alpha = atof(optarg);
                        break;
-               case 'd':
+               case 'i':
                        vr_delta = atof(optarg);
                        break;
                case 'g':
                        vr_gamma = atof(optarg);
                        break;
+               case 1000:
+                       enable_timing = false;
+                       break;
+               case 1001:
+                       enable_variational_refinement = false;
+                       break;
                default:
                        fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
                        exit(1);
@@ -1270,10 +1423,14 @@ int main(int argc, char **argv)
        SDL_GLContext context = SDL_GL_CreateContext(window);
        assert(context != nullptr);
 
+       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";
+
        // Load pictures.
        unsigned width1, height1, width2, height2;
-       GLuint tex0 = load_texture(argc >= (optind + 1) ? argv[optind] : "test1499.png", &width1, &height1);
-       GLuint tex1 = load_texture(argc >= (optind + 2) ? argv[optind + 1] : "test1500.png", &width2, &height2);
+       GLuint tex0 = load_texture(filename0, &width1, &height1);
+       GLuint tex1 = load_texture(filename1, &width2, &height2);
 
        if (width1 != width2 || height1 != height2) {
                fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
@@ -1281,6 +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[] = {
@@ -1296,35 +1461,37 @@ int main(int argc, char **argv)
        DISComputeFlow compute_flow(width1, height1);
        GLuint final_tex = compute_flow.exec(tex0, tex1);
 
-       unique_ptr<float[]> dense_flow(new float[width1 * height1 * 2]);
-       glGetTextureImage(final_tex, 0, GL_RG, GL_FLOAT, width1 * height1 * 2 * 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", width1, height1);
-       fprintf(flowfp, "FEIH");
-       fwrite(&width1, 4, 1, flowfp);
-       fwrite(&height1, 4, 1, flowfp);
-       for (unsigned y = 0; y < unsigned(height1); ++y) {
-               int yy = height1 - y - 1;
-               for (unsigned x = 0; x < unsigned(width1); ++x) {
-                       float du = dense_flow[(yy * width1 + x) * 2 + 0];
-                       float dv = dense_flow[(yy * width1 + x) * 2 + 1];
-
-                       dv = -dv;
-
-                       fwrite(&du, 4, 1, flowfp);
-                       fwrite(&dv, 4, 1, flowfp);
+       schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
+       compute_flow.release_texture(final_tex);
+
+       // 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;
+       for (int i = 1; i < num_flows; ++i) {
+               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];
+               GLuint width, height;
+               GLuint tex0 = load_texture(filename0, &width, &height);
+               if (width != width1 || height != height1) {
+                       fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
+                               filename0, width, height, width1, height1);
+                       exit(1);
+               }
 
-                       uint8_t r, g, b;
-                       flow2rgb(du, dv, &r, &g, &b);
-                       putc(r, fp);
-                       putc(g, fp);
-                       putc(b, fp);
+               GLuint tex1 = load_texture(filename1, &width, &height);
+               if (width != width1 || height != height1) {
+                       fprintf(stderr, "%s: Image dimensions don't match (%dx%d versus %dx%d)\n",
+                               filename1, width, height, width1, height1);
+                       exit(1);
                }
+
+               GLuint final_tex = compute_flow.exec(tex0, tex1);
+               schedule_read(final_tex, width1, height1, filename0, filename1, flow_filename, "");
+               compute_flow.release_texture(final_tex);
        }
-       fclose(fp);
-       fclose(flowfp);
 
-       fprintf(stderr, "err = %d\n", glGetError());
+       while (!reads_in_progress.empty()) {
+               finish_one_read(width1, height1);
+       }
 }