]> git.sesse.net Git - nageru/blobdiff - flow.cpp
Refactor FBO creation. A step on the way to persistent FBOs and temporary textures.
[nageru] / flow.cpp
index 3c69089e0805d192d60ebdef6702e5c736936a80..cf208a64d0efcb8f4dfe3a1278e28f450bdb0434 100644 (file)
--- a/flow.cpp
+++ b/flow.cpp
@@ -1,8 +1,5 @@
 #define NO_SDL_GLEXT 1
 
-#define WIDTH 1280
-#define HEIGHT 720
-
 #include <epoxy/gl.h>
 
 #include <SDL2/SDL.h>
@@ -14,6 +11,7 @@
 #include <SDL2/SDL_video.h>
 
 #include <assert.h>
+#include <getopt.h>
 #include <stdio.h>
 #include <unistd.h>
 
@@ -21,6 +19,7 @@
 
 #include <algorithm>
 #include <memory>
+#include <map>
 #include <vector>
 
 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
@@ -33,6 +32,11 @@ constexpr unsigned coarsest_level = 5;
 constexpr unsigned finest_level = 1;
 constexpr unsigned patch_size_pixels = 12;
 
+// Weighting constants for the different parts of the variational refinement.
+// These don't correspond 1:1 to the values given in the DIS paper,
+// since we have different normalizations and ranges in some cases.
+float vr_gamma = 10.0f, vr_delta = 5.0f, vr_alpha = 10.0f;
+
 // Some global OpenGL objects.
 GLuint nearest_sampler, linear_sampler, smoothness_sampler;
 GLuint vertex_vbo;
@@ -115,26 +119,41 @@ GLuint compile_shader(const string &shader_src, GLenum type)
        return obj;
 }
 
-
-GLuint load_texture(const char *filename, unsigned width, unsigned height)
+GLuint load_texture(const char *filename, unsigned *width_ret, unsigned *height_ret)
 {
-       FILE *fp = fopen(filename, "rb");
-       if (fp == nullptr) {
-               perror(filename);
+       SDL_Surface *surf = IMG_Load(filename);
+       if (surf == nullptr) {
+               fprintf(stderr, "IMG_Load(%s): %s\n", filename, IMG_GetError());
                exit(1);
        }
-       unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
-       if (fread(pix.get(), width * height, 1, fp) != 1) {
-               fprintf(stderr, "Short read from %s\n", filename);
+
+       // For whatever reason, SDL doesn't support converting to YUV surfaces
+       // nor grayscale, so we'll do it (slowly) ourselves.
+       SDL_Surface *rgb_surf = SDL_ConvertSurfaceFormat(surf, SDL_PIXELFORMAT_RGBA8888, /*flags=*/0);
+       if (rgb_surf == nullptr) {
+               fprintf(stderr, "SDL_ConvertSurfaceFormat(%s): %s\n", filename, SDL_GetError());
                exit(1);
        }
-       fclose(fp);
 
-       // Convert to bottom-left origin.
-       for (unsigned y = 0; y < height / 2; ++y) {
+       SDL_FreeSurface(surf);
+
+       unsigned width = rgb_surf->w, height = rgb_surf->h;
+       const uint8_t *sptr = (uint8_t *)rgb_surf->pixels;
+       unique_ptr<uint8_t[]> pix(new uint8_t[width * height]);
+
+       // Extract the Y component, and convert to bottom-left origin.
+       for (unsigned y = 0; y < height; ++y) {
                unsigned y2 = height - 1 - y;
-               swap_ranges(&pix[y * width], &pix[y * width + width], &pix[y2 * width]);
+               for (unsigned x = 0; x < width; ++x) {
+                       uint8_t r = sptr[(y2 * width + x) * 4 + 3];
+                       uint8_t g = sptr[(y2 * width + x) * 4 + 2];
+                       uint8_t b = sptr[(y2 * width + x) * 4 + 1];
+
+                       // Rec. 709.
+                       pix[y * width + x] = lrintf(r * 0.2126f + g * 0.7152f + b * 0.0722f);
+               }
        }
+       SDL_FreeSurface(rgb_surf);
 
        int levels = 1;
        for (int w = width, h = height; w > 1 || h > 1; ) {
@@ -149,6 +168,9 @@ GLuint load_texture(const char *filename, unsigned width, unsigned height)
        glTextureSubImage2D(tex, 0, 0, 0, width, height, GL_RED, GL_UNSIGNED_BYTE, pix.get());
        glGenerateTextureMipmap(tex);
 
+       *width_ret = width;
+       *height_ret = height;
+
        return tex;
 }
 
@@ -206,6 +228,58 @@ void bind_sampler(GLuint program, GLint location, GLuint texture_unit, GLuint te
        glProgramUniform1i(program, location, texture_unit);
 }
 
+// A class that caches FBOs that render to a given set of textures.
+// It never frees anything, so it is only suitable for rendering to
+// the same (small) set of textures over and over again.
+template<size_t num_elements>
+class PersistentFBOSet {
+public:
+       void render_to(const array<GLuint, num_elements> &textures);
+
+       // Convenience wrappers.
+       void render_to(GLuint texture0, enable_if<num_elements == 1> * = nullptr) {
+               render_to({{texture0}});
+       }
+
+       void render_to(GLuint texture0, GLuint texture1, enable_if<num_elements == 2> * = nullptr) {
+               render_to({{texture0, texture1}});
+       }
+
+       void render_to(GLuint texture0, GLuint texture1, GLuint texture2, enable_if<num_elements == 3> * = nullptr) {
+               render_to({{texture0, texture1, texture2}});
+       }
+
+       void render_to(GLuint texture0, GLuint texture1, GLuint texture2, GLuint texture3, enable_if<num_elements == 4> * = nullptr) {
+               render_to({{texture0, texture1, texture2, texture3}});
+       }
+
+private:
+       // TODO: Delete these on destruction.
+       map<array<GLuint, num_elements>, GLuint> fbos;
+};
+
+template<size_t num_elements>
+void PersistentFBOSet<num_elements>::render_to(const array<GLuint, num_elements> &textures)
+{
+       auto it = fbos.find(textures);
+       if (it != fbos.end()) {
+               glBindFramebuffer(GL_FRAMEBUFFER, it->second);
+               return;
+       }
+
+       GLuint fbo;
+       glCreateFramebuffers(1, &fbo);
+       GLenum bufs[num_elements];
+       for (size_t i = 0; i < num_elements; ++i) {
+               glNamedFramebufferTexture(fbo, GL_COLOR_ATTACHMENT0 + i, textures[i], 0);
+               bufs[i] = GL_COLOR_ATTACHMENT0 + i;
+       }
+       glNamedFramebufferDrawBuffers(fbo, num_elements, bufs);
+
+       fbos[textures] = fbo;
+       glBindFramebuffer(GL_FRAMEBUFFER, fbo);
+}
+
 // 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
@@ -220,6 +294,7 @@ public:
        void exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_height);
 
 private:
+       PersistentFBOSet<1> fbos;
        GLuint sobel_vs_obj;
        GLuint sobel_fs_obj;
        GLuint sobel_program;
@@ -252,12 +327,8 @@ void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_he
        glBindSampler(0, nearest_sampler);
        glProgramUniform1i(sobel_program, uniform_tex, 0);
 
-       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);
+       fbos.render_to(grad0_tex);
        glBindVertexArray(sobel_vao);
        glUseProgram(sobel_program);
        glDisable(GL_BLEND);
@@ -271,6 +342,8 @@ public:
        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 prev_level_width, int prev_level_height, int width_patches, int height_patches);
 
 private:
+       PersistentFBOSet<1> fbos;
+
        GLuint motion_vs_obj;
        GLuint motion_fs_obj;
        GLuint motion_search_program;
@@ -317,12 +390,8 @@ void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GL
        glProgramUniform2f(motion_search_program, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
        glProgramUniform2f(motion_search_program, uniform_inv_prev_level_size, 1.0f / prev_level_width, 1.0f / prev_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);
+       fbos.render_to(flow_out_tex);
        glBindVertexArray(motion_search_vao);
        glUseProgram(motion_search_program);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
@@ -342,6 +411,8 @@ public:
        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:
+       PersistentFBOSet<1> fbos;
+
        GLuint densify_vs_obj;
        GLuint densify_fs_obj;
        GLuint densify_program;
@@ -395,15 +466,11 @@ void Densify::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint d
                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);
+       fbos.render_to(dense_flow_tex);
        glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, width_patches * height_patches);
 }
 
@@ -421,6 +488,8 @@ public:
        void exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint normalized_flow_tex, GLuint I_tex, GLuint I_t_tex, int level_width, int level_height);
 
 private:
+       PersistentFBOSet<3> fbos;
+
        GLuint prewarp_vs_obj;
        GLuint prewarp_fs_obj;
        GLuint prewarp_program;
@@ -462,18 +531,10 @@ void Prewarp::exec(GLuint tex0_view, GLuint tex1_view, GLuint flow_tex, GLuint I
 
        glProgramUniform2f(prewarp_program, uniform_image_size, level_width, level_height);
 
-       GLuint prewarp_fbo;  // TODO: cleanup
-       glCreateFramebuffers(1, &prewarp_fbo);
-       GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1, GL_COLOR_ATTACHMENT2 };
-       glNamedFramebufferDrawBuffers(prewarp_fbo, 3, bufs);
-       glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT0, I_tex, 0);
-       glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT1, I_t_tex, 0);
-       glNamedFramebufferTexture(prewarp_fbo, GL_COLOR_ATTACHMENT2, normalized_flow_tex, 0);
-
        glViewport(0, 0, level_width, level_height);
        glDisable(GL_BLEND);
        glBindVertexArray(prewarp_vao);
-       glBindFramebuffer(GL_FRAMEBUFFER, prewarp_fbo);
+       fbos.render_to(I_tex, I_t_tex, normalized_flow_tex);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
 
@@ -491,6 +552,8 @@ public:
        void exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, int level_width, int level_height);
 
 private:
+       PersistentFBOSet<2> fbos;
+
        GLuint derivatives_vs_obj;
        GLuint derivatives_fs_obj;
        GLuint derivatives_program;
@@ -523,17 +586,10 @@ void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, in
 
        bind_sampler(derivatives_program, uniform_tex, 0, input_tex, nearest_sampler);
 
-       GLuint derivatives_fbo;  // TODO: cleanup
-       glCreateFramebuffers(1, &derivatives_fbo);
-       GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
-       glNamedFramebufferDrawBuffers(derivatives_fbo, 2, bufs);
-       glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT0, I_x_y_tex, 0);
-       glNamedFramebufferTexture(derivatives_fbo, GL_COLOR_ATTACHMENT1, beta_0_tex, 0);
-
        glViewport(0, 0, level_width, level_height);
        glDisable(GL_BLEND);
        glBindVertexArray(derivatives_vao);
-       glBindFramebuffer(GL_FRAMEBUFFER, derivatives_fbo);
+       fbos.render_to(I_x_y_tex, beta_0_tex);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
 
@@ -550,12 +606,15 @@ public:
        void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height);
 
 private:
+       PersistentFBOSet<2> fbos;
+
        GLuint smoothness_vs_obj;
        GLuint smoothness_fs_obj;
        GLuint smoothness_program;
        GLuint smoothness_vao;
 
        GLuint uniform_flow_tex, uniform_diff_flow_tex;
+       GLuint uniform_alpha;
 };
 
 ComputeSmoothness::ComputeSmoothness()
@@ -575,6 +634,7 @@ ComputeSmoothness::ComputeSmoothness()
 
        uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
        uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
+       uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
 }
 
 void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height)
@@ -583,32 +643,21 @@ void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoot
 
        bind_sampler(smoothness_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
        bind_sampler(smoothness_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
-
-       GLuint smoothness_fbo;  // TODO: cleanup
-       glCreateFramebuffers(1, &smoothness_fbo);
-       GLenum bufs[] = { GL_COLOR_ATTACHMENT0, GL_COLOR_ATTACHMENT1 };
-       glNamedFramebufferDrawBuffers(smoothness_fbo, 2, bufs);
-       glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT0, smoothness_x_tex, 0);
-       glNamedFramebufferTexture(smoothness_fbo, GL_COLOR_ATTACHMENT1, smoothness_y_tex, 0);
+       glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
 
        glViewport(0, 0, level_width, level_height);
 
-       // Make sure the smoothness on the right and upper borders is zero.
-       // We could have done this by making (W-1)xH and Wx(H-1) textures instead
-       // (we're sampling smoothness with all-zero border color), but we'd
-       // have to adjust the sampling coordinates, which is annoying.
-       //
-       // FIXME: We shouldn't scissor width for horizontal,
-       // and we shouldn't scissor height for vertical
-       glScissor(0, 0, level_width - 1, level_height - 1);
-       glEnable(GL_SCISSOR_TEST);
-
        glDisable(GL_BLEND);
        glBindVertexArray(smoothness_vao);
-       glBindFramebuffer(GL_FRAMEBUFFER, smoothness_fbo);
+       fbos.render_to(smoothness_x_tex, smoothness_y_tex);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 
-       glDisable(GL_SCISSOR_TEST);
+       // Make sure the smoothness on the right and upper borders is zero.
+       // We could have done this by making (W-1)xH and Wx(H-1) textures instead
+       // (we're sampling smoothness with all-zero border color), but we'd
+       // have to adjust the sampling coordinates, which is annoying.
+       glClearTexSubImage(smoothness_x_tex, 0,  level_width - 1, 0, 0,   1, level_height, 1,  GL_RED, GL_FLOAT, nullptr);
+       glClearTexSubImage(smoothness_y_tex, 0,  0, level_height - 1, 0,  level_width, 1, 1,   GL_RED, GL_FLOAT, nullptr);
 }
 
 // Set up the equations set (two equations in two unknowns, per pixel).
@@ -627,15 +676,18 @@ public:
        void exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height);
 
 private:
+       PersistentFBOSet<1> fbos;
+
        GLuint equations_vs_obj;
        GLuint equations_fs_obj;
        GLuint equations_program;
        GLuint equations_vao;
 
        GLuint uniform_I_x_y_tex, uniform_I_t_tex;
-       GLuint uniform_diff_flow_tex, uniform_flow_tex;
+       GLuint uniform_diff_flow_tex, uniform_base_flow_tex;
        GLuint uniform_beta_0_tex;
        GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
+       GLuint uniform_gamma, uniform_delta;
 };
 
 SetupEquations::SetupEquations()
@@ -656,40 +708,37 @@ SetupEquations::SetupEquations()
        uniform_I_x_y_tex = glGetUniformLocation(equations_program, "I_x_y_tex");
        uniform_I_t_tex = glGetUniformLocation(equations_program, "I_t_tex");
        uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_flow_tex");
-       uniform_flow_tex = glGetUniformLocation(equations_program, "flow_tex");
+       uniform_base_flow_tex = glGetUniformLocation(equations_program, "base_flow_tex");
        uniform_beta_0_tex = glGetUniformLocation(equations_program, "beta_0_tex");
        uniform_smoothness_x_tex = glGetUniformLocation(equations_program, "smoothness_x_tex");
        uniform_smoothness_y_tex = glGetUniformLocation(equations_program, "smoothness_y_tex");
+       uniform_gamma = glGetUniformLocation(equations_program, "gamma");
+       uniform_delta = glGetUniformLocation(equations_program, "delta");
 }
 
-void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height)
+void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint base_flow_tex, GLuint beta_0_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, GLuint equation_tex, int level_width, int level_height)
 {
        glUseProgram(equations_program);
 
        bind_sampler(equations_program, uniform_I_x_y_tex, 0, I_x_y_tex, nearest_sampler);
        bind_sampler(equations_program, uniform_I_t_tex, 1, I_t_tex, nearest_sampler);
        bind_sampler(equations_program, uniform_diff_flow_tex, 2, diff_flow_tex, nearest_sampler);
-       bind_sampler(equations_program, uniform_flow_tex, 3, 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);
-
-       GLuint equations_fbo;  // TODO: cleanup
-       glCreateFramebuffers(1, &equations_fbo);
-       glNamedFramebufferTexture(equations_fbo, GL_COLOR_ATTACHMENT0, equation_tex, 0);
+       glProgramUniform1f(equations_program, uniform_delta, vr_delta);
+       glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
 
        glViewport(0, 0, level_width, level_height);
        glDisable(GL_BLEND);
        glBindVertexArray(equations_vao);
-       glBindFramebuffer(GL_FRAMEBUFFER, equations_fbo);
+       fbos.render_to(equation_tex);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
 
-// Calculate the smoothness constraints between neighboring pixels;
-// s_x(x,y) stores smoothness between pixel (x,y) and (x+1,y),
-// and s_y(x,y) stores between (x,y) and (x,y+1). We'll sample with
-// border color (0,0) later, so that there's zero diffusion out of
-// the border.
+// Actually solve the equation sets made by SetupEquations, by means of
+// successive over-relaxation (SOR).
 //
 // See variational_refinement.txt for more information.
 class SOR {
@@ -698,6 +747,8 @@ public:
        void 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);
 
 private:
+       PersistentFBOSet<1> fbos;
+
        GLuint sor_vs_obj;
        GLuint sor_fs_obj;
        GLuint sor_program;
@@ -738,14 +789,10 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_tex, GLuint smoothness_x_te
        bind_sampler(sor_program, uniform_smoothness_y_tex, 2, smoothness_y_tex, smoothness_sampler);
        bind_sampler(sor_program, uniform_equation_tex, 3, equation_tex, nearest_sampler);
 
-       GLuint sor_fbo;  // TODO: cleanup
-       glCreateFramebuffers(1, &sor_fbo);
-       glNamedFramebufferTexture(sor_fbo, GL_COLOR_ATTACHMENT0, diff_flow_tex, 0);  // NOTE: Bind to same as we render from!
-
        glViewport(0, 0, level_width, level_height);
        glDisable(GL_BLEND);
        glBindVertexArray(sor_vao);
-       glBindFramebuffer(GL_FRAMEBUFFER, sor_fbo);
+       fbos.render_to(diff_flow_tex);  // NOTE: Bind to same as we render from!
 
        for (int i = 0; i < num_iterations; ++i) {
                glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
@@ -763,6 +810,8 @@ public:
        void exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_width, int level_height);
 
 private:
+       PersistentFBOSet<1> fbos;
+
        GLuint add_flow_vs_obj;
        GLuint add_flow_fs_obj;
        GLuint add_flow_program;
@@ -795,15 +844,64 @@ void AddBaseFlow::exec(GLuint base_flow_tex, GLuint diff_flow_tex, int level_wid
 
        bind_sampler(add_flow_program, uniform_diff_flow_tex, 0, diff_flow_tex, nearest_sampler);
 
-       GLuint add_flow_fbo;  // TODO: cleanup
-       glCreateFramebuffers(1, &add_flow_fbo);
-       glNamedFramebufferTexture(add_flow_fbo, GL_COLOR_ATTACHMENT0, base_flow_tex, 0);
-
        glViewport(0, 0, level_width, level_height);
        glEnable(GL_BLEND);
        glBlendFunc(GL_ONE, GL_ONE);
        glBindVertexArray(add_flow_vao);
-       glBindFramebuffer(GL_FRAMEBUFFER, add_flow_fbo);
+       fbos.render_to(base_flow_tex);
+
+       glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+}
+
+// Take a copy of the flow, bilinearly interpolated and scaled up.
+class ResizeFlow {
+public:
+       ResizeFlow();
+       void exec(GLuint in_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height);
+
+private:
+       PersistentFBOSet<1> fbos;
+
+       GLuint resize_flow_vs_obj;
+       GLuint resize_flow_fs_obj;
+       GLuint resize_flow_program;
+       GLuint resize_flow_vao;
+
+       GLuint uniform_flow_tex;
+       GLuint uniform_scale_factor;
+};
+
+ResizeFlow::ResizeFlow()
+{
+       resize_flow_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+       resize_flow_fs_obj = compile_shader(read_file("resize_flow.frag"), GL_FRAGMENT_SHADER);
+       resize_flow_program = link_program(resize_flow_vs_obj, resize_flow_fs_obj);
+
+       // Set up the VAO containing all the required position/texcoord data.
+       glCreateVertexArrays(1, &resize_flow_vao);
+       glBindVertexArray(resize_flow_vao);
+       glBindBuffer(GL_ARRAY_BUFFER, vertex_vbo);
+
+       GLint position_attrib = glGetAttribLocation(resize_flow_program, "position");
+       glEnableVertexArrayAttrib(resize_flow_vao, position_attrib);
+       glVertexAttribPointer(position_attrib, 2, GL_FLOAT, GL_FALSE, 0, BUFFER_OFFSET(0));
+
+       uniform_flow_tex = glGetUniformLocation(resize_flow_program, "flow_tex");
+       uniform_scale_factor = glGetUniformLocation(resize_flow_program, "scale_factor");
+}
+
+void ResizeFlow::exec(GLuint flow_tex, GLuint out_tex, int input_width, int input_height, int output_width, int output_height)
+{
+       glUseProgram(resize_flow_program);
+
+       bind_sampler(resize_flow_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
+
+       glProgramUniform2f(resize_flow_program, uniform_scale_factor, float(output_width) / input_width, float(output_height) / input_height);
+
+       glViewport(0, 0, output_width, output_height);
+       glDisable(GL_BLEND);
+       glBindVertexArray(resize_flow_vao);
+       fbos.render_to(out_tex);
 
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
@@ -848,7 +946,7 @@ void GPUTimers::print()
                for (int i = 0; i < timer.level * 2; ++i) {
                        fprintf(stderr, " ");
                }
-               fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), (time_end - time_start) / 1e6);
+               fprintf(stderr, "%-30s %4.1f ms\n", timer.name.c_str(), GLint64(time_end - time_start) / 1e6);
        }
 }
 
@@ -888,8 +986,37 @@ private:
        bool ended = false;
 };
 
-int main(void)
+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' }
+       };
+
+       for ( ;; ) {
+               int option_index = 0;
+               int c = getopt_long(argc, argv, "a:d:g:", long_options, &option_index);
+
+               if (c == -1) {
+                       break;
+               }
+               switch (c) {
+               case 'a':
+                       vr_alpha = atof(optarg);
+                       break;
+               case 'd':
+                       vr_delta = atof(optarg);
+                       break;
+               case 'g':
+                       vr_gamma = atof(optarg);
+                       break;
+               default:
+                       fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
+                       exit(1);
+               };
+       }
+
        if (SDL_Init(SDL_INIT_EVERYTHING) == -1) {
                fprintf(stderr, "SDL_Init failed: %s\n", SDL_GetError());
                exit(1);
@@ -912,8 +1039,15 @@ int main(void)
        assert(context != nullptr);
 
        // Load pictures.
-       GLuint tex0 = load_texture("test1499.pgm", WIDTH, HEIGHT);
-       GLuint tex1 = load_texture("test1500.pgm", WIDTH, HEIGHT);
+       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);
+
+       if (width1 != width2 || height1 != height2) {
+               fprintf(stderr, "Image dimensions don't match (%dx%d versus %dx%d)\n",
+                       width1, height1, width2, height2);
+               exit(1);
+       }
 
        // Make some samplers.
        glCreateSamplers(1, &nearest_sampler);
@@ -952,6 +1086,7 @@ int main(void)
        GLuint initial_flow_tex;
        glCreateTextures(GL_TEXTURE_2D, 1, &initial_flow_tex);
        glTextureStorage2D(initial_flow_tex, 1, GL_RG16F, 1, 1);
+       glClearTexImage(initial_flow_tex, 0, GL_RG, GL_FLOAT, nullptr);
        int prev_level_width = 1, prev_level_height = 1;
 
        GLuint prev_level_flow_tex = initial_flow_tex;
@@ -965,6 +1100,7 @@ int main(void)
        SetupEquations setup_equations;
        SOR sor;
        AddBaseFlow add_base_flow;
+       ResizeFlow resize_flow;
 
        GLuint query;
        glGenQueries(1, &query);
@@ -978,8 +1114,8 @@ int main(void)
                snprintf(timer_name, sizeof(timer_name), "Level %d", level);
                ScopedTimer level_timer(timer_name, &total_timer);
 
-               int level_width = WIDTH >> level;
-               int level_height = HEIGHT >> level;
+               int level_width = width1 >> level;
+               int level_height = height1 >> 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);
@@ -1026,6 +1162,7 @@ int main(void)
                GLuint dense_flow_tex;
                glCreateTextures(GL_TEXTURE_2D, 1, &dense_flow_tex);
                glTextureStorage2D(dense_flow_tex, 1, GL_RGB16F, level_width, level_height);
+               glClearTexImage(dense_flow_tex, 0, GL_RGB, GL_FLOAT, nullptr);
 
                // And draw.
                {
@@ -1075,6 +1212,7 @@ int main(void)
                GLuint du_dv_tex;
                glCreateTextures(GL_TEXTURE_2D, 1, &du_dv_tex);
                glTextureStorage2D(du_dv_tex, 1, 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;
@@ -1114,7 +1252,7 @@ int main(void)
                // 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 prat if you wish to test disabling of the variational refinement.
+               // You can comment out this part if you wish to test disabling of the variational refinement.
                {
                        ScopedTimer timer("Add differential flow", &varref_timer);
                        add_base_flow.exec(base_flow_tex, du_dv_tex, level_width, level_height);
@@ -1128,22 +1266,30 @@ int main(void)
 
        timers.print();
 
-       int level_width = WIDTH >> finest_level;
-       int level_height = HEIGHT >> finest_level;
-       unique_ptr<float[]> dense_flow(new float[level_width * level_height * 2]);
-       glGetTextureImage(prev_level_flow_tex, 0, GL_RG, GL_FLOAT, level_width * level_height * 2 * sizeof(float), dense_flow.get());
+       // Scale up the flow to the final size (if needed).
+       GLuint final_tex;
+       if (finest_level == 0) {
+               final_tex = prev_level_flow_tex;
+       } else {
+               glCreateTextures(GL_TEXTURE_2D, 1, &final_tex);
+               glTextureStorage2D(final_tex, 1, GL_RG16F, width1, height1);
+               resize_flow.exec(prev_level_flow_tex, final_tex, prev_level_width, prev_level_height, width1, height1);
+       }
+
+       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", level_width, level_height);
+       fprintf(fp, "P6\n%d %d\n255\n", width1, height1);
        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) * 2 + 0];
-                       float dv = dense_flow[(yy * level_width + x) * 2 + 1];
+       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;