]> git.sesse.net Git - nageru/blobdiff - flow.cpp
Rename du_dv_tex to diff_flow_tex, for consistency.
[nageru] / flow.cpp
index 54d28d98aa70418b856894fc7e6e093e26d81cf1..76589429aed187faefeec955ac20d2943a5bcf83 100644 (file)
--- a/flow.cpp
+++ b/flow.cpp
@@ -379,7 +379,6 @@ void GrayscaleConversion::exec(GLint tex, GLint gray_tex, int width, int height)
        glViewport(0, 0, width, height);
        fbos.render_to(gray_tex);
        glBindVertexArray(gray_vao);
-       glUseProgram(gray_program);
        glDisable(GL_BLEND);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
@@ -422,7 +421,6 @@ void Sobel::exec(GLint tex0_view, GLint grad0_tex, int level_width, int level_he
 
        glViewport(0, 0, level_width, level_height);
        fbos.render_to(grad0_tex);
-       glUseProgram(sobel_program);
        glDisable(GL_BLEND);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
@@ -441,7 +439,7 @@ private:
        GLuint motion_search_program;
 
        GLuint uniform_inv_image_size, uniform_inv_prev_level_size;
-       GLuint uniform_image0_tex, uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
+       GLuint uniform_image1_tex, uniform_grad0_tex, uniform_flow_tex;
 };
 
 MotionSearch::MotionSearch()
@@ -452,7 +450,6 @@ MotionSearch::MotionSearch()
 
        uniform_inv_image_size = glGetUniformLocation(motion_search_program, "inv_image_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");
        uniform_grad0_tex = glGetUniformLocation(motion_search_program, "grad0_tex");
        uniform_flow_tex = glGetUniformLocation(motion_search_program, "flow_tex");
@@ -462,9 +459,8 @@ void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GL
 {
        glUseProgram(motion_search_program);
 
-       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, zero_border_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, uniform_inv_image_size, 1.0f / level_width, 1.0f / level_height);
@@ -472,7 +468,6 @@ void MotionSearch::exec(GLuint tex0_view, GLuint tex1_view, GLuint grad0_tex, GL
 
        glViewport(0, 0, width_patches, height_patches);
        fbos.render_to(flow_out_tex);
-       glUseProgram(motion_search_program);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
 
@@ -625,62 +620,54 @@ void Derivatives::exec(GLuint input_tex, GLuint I_x_y_tex, GLuint beta_0_tex, in
        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.
+// Calculate the diffusivity for each pixels, g(x,y). Smoothness (s) will
+// be calculated in the shaders on-the-fly by sampling in-between two
+// neighboring g(x,y) pixels, plus a border tweak to make sure we get
+// zero smoothness at the border.
 //
 // See variational_refinement.txt for more information.
-class ComputeSmoothness {
+class ComputeDiffusivity {
 public:
-       ComputeSmoothness();
-       void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, bool zero_diff_flow);
+       ComputeDiffusivity();
+       void exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow);
 
 private:
-       PersistentFBOSet<2> fbos;
+       PersistentFBOSet<1> fbos;
 
-       GLuint smoothness_vs_obj;
-       GLuint smoothness_fs_obj;
-       GLuint smoothness_program;
+       GLuint diffusivity_vs_obj;
+       GLuint diffusivity_fs_obj;
+       GLuint diffusivity_program;
 
        GLuint uniform_flow_tex, uniform_diff_flow_tex;
        GLuint uniform_alpha, uniform_zero_diff_flow;
 };
 
-ComputeSmoothness::ComputeSmoothness()
+ComputeDiffusivity::ComputeDiffusivity()
 {
-       smoothness_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
-       smoothness_fs_obj = compile_shader(read_file("smoothness.frag"), GL_FRAGMENT_SHADER);
-       smoothness_program = link_program(smoothness_vs_obj, smoothness_fs_obj);
-
-       uniform_flow_tex = glGetUniformLocation(smoothness_program, "flow_tex");
-       uniform_diff_flow_tex = glGetUniformLocation(smoothness_program, "diff_flow_tex");
-       uniform_alpha = glGetUniformLocation(smoothness_program, "alpha");
-       uniform_zero_diff_flow = glGetUniformLocation(smoothness_program, "zero_diff_flow");
+       diffusivity_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+       diffusivity_fs_obj = compile_shader(read_file("diffusivity.frag"), GL_FRAGMENT_SHADER);
+       diffusivity_program = link_program(diffusivity_vs_obj, diffusivity_fs_obj);
+
+       uniform_flow_tex = glGetUniformLocation(diffusivity_program, "flow_tex");
+       uniform_diff_flow_tex = glGetUniformLocation(diffusivity_program, "diff_flow_tex");
+       uniform_alpha = glGetUniformLocation(diffusivity_program, "alpha");
+       uniform_zero_diff_flow = glGetUniformLocation(diffusivity_program, "zero_diff_flow");
 }
 
-void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoothness_x_tex, GLuint smoothness_y_tex, int level_width, int level_height, bool zero_diff_flow)
+void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diffusivity_tex, int level_width, int level_height, bool zero_diff_flow)
 {
-       glUseProgram(smoothness_program);
+       glUseProgram(diffusivity_program);
 
-       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);
-       glProgramUniform1f(smoothness_program, uniform_alpha, vr_alpha);
-       glProgramUniform1i(smoothness_program, uniform_zero_diff_flow, zero_diff_flow);
+       bind_sampler(diffusivity_program, uniform_flow_tex, 0, flow_tex, nearest_sampler);
+       bind_sampler(diffusivity_program, uniform_diff_flow_tex, 1, diff_flow_tex, nearest_sampler);
+       glProgramUniform1f(diffusivity_program, uniform_alpha, vr_alpha);
+       glProgramUniform1i(diffusivity_program, uniform_zero_diff_flow, zero_diff_flow);
 
        glViewport(0, 0, level_width, level_height);
 
        glDisable(GL_BLEND);
-       fbos.render_to(smoothness_x_tex, smoothness_y_tex);
+       fbos.render_to(diffusivity_tex);
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
-
-       // 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).
@@ -692,14 +679,23 @@ void ComputeSmoothness::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint smoot
 // All the values of the energy term (E_I, E_G, E_S), except the smoothness
 // terms that depend on other pixels, are calculated in one pass.
 //
-// See variational_refinement.txt for more information.
+// The equation set is split in two; one contains only the pixels needed for
+// the red pass, and one only for the black pass (see sor.frag). This reduces
+// the amount of data the SOR shader has to pull in, at the cost of some
+// complexity when the equation texture ends up with half the size and we need
+// to adjust texture coordinates.  The contraction is done along the horizontal
+// axis, so that on even rows (0, 2, 4, ...), the “red” texture will contain
+// pixels 0, 2, 4, 6, etc., and on odd rows 1, 3, 5, etc..
+//
+// See variational_refinement.txt for more information about the actual
+// equations in use.
 class SetupEquations {
 public:
        SetupEquations();
-       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, bool zero_diff_flow);
+       void exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex, GLuint flow_tex, GLuint beta_0_tex, GLuint diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow);
 
 private:
-       PersistentFBOSet<1> fbos;
+       PersistentFBOSet<2> fbos;
 
        GLuint equations_vs_obj;
        GLuint equations_fs_obj;
@@ -708,13 +704,13 @@ private:
        GLuint uniform_I_x_y_tex, uniform_I_t_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_diffusivity_tex;
        GLuint uniform_gamma, uniform_delta, uniform_zero_diff_flow;
 };
 
 SetupEquations::SetupEquations()
 {
-       equations_vs_obj = compile_shader(read_file("vs.vert"), GL_VERTEX_SHADER);
+       equations_vs_obj = compile_shader(read_file("equations.vert"), GL_VERTEX_SHADER);
        equations_fs_obj = compile_shader(read_file("equations.frag"), GL_FRAGMENT_SHADER);
        equations_program = link_program(equations_vs_obj, equations_fs_obj);
 
@@ -723,14 +719,13 @@ SetupEquations::SetupEquations()
        uniform_diff_flow_tex = glGetUniformLocation(equations_program, "diff_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_diffusivity_tex = glGetUniformLocation(equations_program, "diffusivity_tex");
        uniform_gamma = glGetUniformLocation(equations_program, "gamma");
        uniform_delta = glGetUniformLocation(equations_program, "delta");
        uniform_zero_diff_flow = glGetUniformLocation(equations_program, "zero_diff_flow");
 }
 
-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, bool zero_diff_flow)
+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 diffusivity_tex, GLuint equation_red_tex, GLuint equation_black_tex, int level_width, int level_height, bool zero_diff_flow)
 {
        glUseProgram(equations_program);
 
@@ -739,15 +734,14 @@ 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, zero_border_sampler);
-       bind_sampler(equations_program, uniform_smoothness_y_tex, 6, smoothness_y_tex, zero_border_sampler);
+       bind_sampler(equations_program, uniform_diffusivity_tex, 5, diffusivity_tex, zero_border_sampler);
        glProgramUniform1f(equations_program, uniform_delta, vr_delta);
        glProgramUniform1f(equations_program, uniform_gamma, vr_gamma);
        glProgramUniform1i(equations_program, uniform_zero_diff_flow, zero_diff_flow);
 
-       glViewport(0, 0, level_width, level_height);
+       glViewport(0, 0, (level_width + 1) / 2, level_height);
        glDisable(GL_BLEND);
-       fbos.render_to(equation_tex);
+       fbos.render_to({equation_red_tex, equation_black_tex});
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
 
@@ -758,7 +752,7 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex
 class SOR {
 public:
        SOR();
-       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, bool zero_diff_flow, ScopedTimer *sor_timer);
+       void exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer);
 
 private:
        PersistentFBOSet<1> fbos;
@@ -768,8 +762,8 @@ private:
        GLuint sor_program;
 
        GLuint uniform_diff_flow_tex;
-       GLuint uniform_equation_tex;
-       GLuint uniform_smoothness_x_tex, uniform_smoothness_y_tex;
+       GLuint uniform_equation_red_tex, uniform_equation_black_tex;
+       GLuint uniform_diffusivity_tex;
        GLuint uniform_phase, uniform_zero_diff_flow;
 };
 
@@ -780,21 +774,21 @@ SOR::SOR()
        sor_program = link_program(sor_vs_obj, sor_fs_obj);
 
        uniform_diff_flow_tex = glGetUniformLocation(sor_program, "diff_flow_tex");
-       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_equation_red_tex = glGetUniformLocation(sor_program, "equation_red_tex");
+       uniform_equation_black_tex = glGetUniformLocation(sor_program, "equation_black_tex");
+       uniform_diffusivity_tex = glGetUniformLocation(sor_program, "diffusivity_tex");
        uniform_phase = glGetUniformLocation(sor_program, "phase");
        uniform_zero_diff_flow = glGetUniformLocation(sor_program, "zero_diff_flow");
 }
 
-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, bool zero_diff_flow, ScopedTimer *sor_timer)
+void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_black_tex, GLuint diffusivity_tex, int level_width, int level_height, int num_iterations, bool zero_diff_flow, ScopedTimer *sor_timer)
 {
        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, 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);
+       bind_sampler(sor_program, uniform_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
+       bind_sampler(sor_program, uniform_equation_red_tex, 2, equation_red_tex, nearest_sampler);
+       bind_sampler(sor_program, uniform_equation_black_tex, 3, equation_black_tex, nearest_sampler);
 
        glProgramUniform1i(sor_program, uniform_zero_diff_flow, zero_diff_flow);
 
@@ -954,7 +948,7 @@ private:
        Densify densify;
        Prewarp prewarp;
        Derivatives derivatives;
-       ComputeSmoothness compute_smoothness;
+       ComputeDiffusivity compute_diffusivity;
        SetupEquations setup_equations;
        SOR sor;
        AddBaseFlow add_base_flow;
@@ -982,11 +976,11 @@ DISComputeFlow::DISComputeFlow(int width, int height)
        // 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_MIN_FILTER, GL_LINEAR);
+       glSamplerParameteri(zero_border_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
        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 };
+       float zero[] = { 0.0f, 0.0f, 0.0f, 0.0f };  // Note that zero alpha means we can also see whether we sampled outside the border or not.
        glSamplerParameterfv(zero_border_sampler, GL_TEXTURE_BORDER_COLOR, zero);
 
        // Initial flow is zero, 1x1.
@@ -1050,7 +1044,7 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
 
                // Create a new texture; we could be fancy and render use a multi-level
                // texture, but meh.
-               GLuint grad0_tex = pool.get_texture(GL_RG16F, level_width, level_height);
+               GLuint grad0_tex = pool.get_texture(GL_R32UI, level_width, level_height);
 
                // Find the derivative.
                {
@@ -1120,44 +1114,42 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                // We need somewhere to store du and dv (the flow increment, relative
                // to the non-refined base flow u0 and v0). It's initially garbage,
                // but not read until we've written something sane to it.
-               GLuint du_dv_tex = pool.get_texture(GL_RG16F, level_width, level_height);
+               GLuint diff_flow_tex = pool.get_texture(GL_RG16F, level_width, level_height);
 
-               // And for smoothness.
-               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 for diffusivity.
+               GLuint diffusivity_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 = pool.get_texture(GL_RGBA32UI, level_width, level_height);
+               GLuint equation_red_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height);
+               GLuint equation_black_tex = pool.get_texture(GL_RGBA32UI, (level_width + 1) / 2, level_height);
 
                for (int outer_idx = 0; outer_idx < level + 1; ++outer_idx) {
-                       // Calculate the smoothness terms between the neighboring pixels,
-                       // both in x and y direction.
+                       // Calculate the diffusivity term for each pixel.
                        {
-                               ScopedTimer timer("Compute smoothness", &varref_timer);
-                               compute_smoothness.exec(base_flow_tex, du_dv_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, outer_idx == 0);
+                               ScopedTimer timer("Compute diffusivity", &varref_timer);
+                               compute_diffusivity.exec(base_flow_tex, diff_flow_tex, diffusivity_tex, level_width, level_height, outer_idx == 0);
                        }
 
                        // Set up the 2x2 equation system for each pixel.
                        {
                                ScopedTimer timer("Set up equations", &varref_timer);
-                               setup_equations.exec(I_x_y_tex, I_t_tex, du_dv_tex, base_flow_tex, beta_0_tex, smoothness_x_tex, smoothness_y_tex, equation_tex, level_width, level_height, outer_idx == 0);
+                               setup_equations.exec(I_x_y_tex, I_t_tex, diff_flow_tex, base_flow_tex, beta_0_tex, diffusivity_tex, equation_red_tex, equation_black_tex, level_width, level_height, outer_idx == 0);
                        }
 
-                       // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
-                       // Note that these are to/from the same texture.
+                       // Run a few SOR iterations. Note that these are to/from the same texture.
                        {
                                ScopedTimer timer("SOR", &varref_timer);
-                               sor.exec(du_dv_tex, equation_tex, smoothness_x_tex, smoothness_y_tex, level_width, level_height, 5, outer_idx == 0, &timer);
+                               sor.exec(diff_flow_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer);
                        }
                }
 
                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);
+               pool.release_texture(diffusivity_tex);
+               pool.release_texture(equation_red_tex);
+               pool.release_texture(equation_black_tex);
 
                // Add the differential flow found by the variational refinement to the base flow,
                // giving the final flow estimate for this level.
@@ -1167,9 +1159,9 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                // 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);
+                       add_base_flow.exec(base_flow_tex, diff_flow_tex, level_width, level_height);
                }
-               pool.release_texture(du_dv_tex);
+               pool.release_texture(diff_flow_tex);
 
                if (prev_level_flow_tex != initial_flow_tex) {
                        pool.release_texture(prev_level_flow_tex);
@@ -1460,11 +1452,9 @@ void Blend::exec(GLuint tex0, GLuint tex1, GLuint flow_tex, GLuint output_tex, i
        bind_sampler(blend_program, uniform_image1_tex, 1, tex1, linear_sampler);
        bind_sampler(blend_program, uniform_flow_tex, 2, flow_tex, linear_sampler);  // May be upsampled.
        glProgramUniform1f(blend_program, uniform_alpha, alpha);
-       //glProgramUniform1f(blend_program, uniform_flow_consistency_tolerance, 1.0f / 
 
        glViewport(0, 0, level_width, level_height);
        fbos.render_to(output_tex);
-       glUseProgram(blend_program);
        glDisable(GL_BLEND);  // A bit ironic, perhaps.
        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
 }
@@ -1563,6 +1553,7 @@ GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLui
                ScopedTimer timer("Blend", &total_timer);
                blend.exec(tex0, tex1, flow_tex, output_tex, width, height, alpha);
        }
+       pool.release_texture(flow_tex);
        total_timer.end();
        timers.print();