]> git.sesse.net Git - nageru/blobdiff - flow.cpp
Add a warmup option to get somewhat more consistent timings.
[nageru] / flow.cpp
index b5656cb10766cdf4ee6cb6ce6b06c20a195fb799..4fa779473a400a3baa5e9c6c1cfd0a26f9c174cc 100644 (file)
--- a/flow.cpp
+++ b/flow.cpp
@@ -48,6 +48,8 @@ float vr_alpha = 1.0f, vr_delta = 0.25f, vr_gamma = 0.25f;
 
 bool enable_timing = true;
 bool detailed_timing = false;
+bool enable_warmup = false;
+bool in_warmup = false;
 bool enable_variational_refinement = true;  // Just for debugging.
 bool enable_interpolation = false;
 
@@ -764,7 +766,7 @@ private:
        GLuint uniform_diff_flow_tex;
        GLuint uniform_equation_red_tex, uniform_equation_black_tex;
        GLuint uniform_diffusivity_tex;
-       GLuint uniform_phase, uniform_zero_diff_flow;
+       GLuint uniform_phase, uniform_num_nonzero_phases;
 };
 
 SOR::SOR()
@@ -778,7 +780,7 @@ SOR::SOR()
        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");
+       uniform_num_nonzero_phases = glGetUniformLocation(sor_program, "num_nonzero_phases");
 }
 
 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)
@@ -790,7 +792,9 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl
        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);
+       if (!zero_diff_flow) {
+               glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
+       }
 
        // 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
@@ -803,6 +807,9 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl
        for (int i = 0; i < num_iterations; ++i) {
                {
                        ScopedTimer timer("Red pass", sor_timer);
+                       if (zero_diff_flow && i == 0) {
+                               glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 0);
+                       }
                        glProgramUniform1i(sor_program, uniform_phase, 0);
                        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
                        glTextureBarrier();
@@ -810,11 +817,13 @@ void SOR::exec(GLuint diff_flow_tex, GLuint equation_red_tex, GLuint equation_bl
                {
                        ScopedTimer timer("Black pass", sor_timer);
                        if (zero_diff_flow && i == 0) {
-                               // Not zero anymore.
-                               glProgramUniform1i(sor_program, uniform_zero_diff_flow, 0);
+                               glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 1);
                        }
                        glProgramUniform1i(sor_program, uniform_phase, 1);
                        glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
+                       if (zero_diff_flow && i == 0) {
+                               glProgramUniform1i(sor_program, uniform_num_nonzero_phases, 2);
+                       }
                        if (i != num_iterations - 1) {
                                glTextureBarrier();
                        }
@@ -1114,7 +1123,7 @@ 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 diffusivity.
                GLuint diffusivity_tex = pool.get_texture(GL_R16F, level_width, level_height);
@@ -1128,19 +1137,19 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                        // Calculate the diffusivity term for each pixel.
                        {
                                ScopedTimer timer("Compute diffusivity", &varref_timer);
-                               compute_diffusivity.exec(base_flow_tex, du_dv_tex, diffusivity_tex, level_width, level_height, outer_idx == 0);
+                               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, diffusivity_tex, equation_red_tex, equation_black_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 iterations. Note that these are to/from the same texture.
                        {
                                ScopedTimer timer("SOR", &varref_timer);
-                               sor.exec(du_dv_tex, equation_red_tex, equation_black_tex, diffusivity_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);
                        }
                }
 
@@ -1159,9 +1168,9 @@ 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);
@@ -1172,7 +1181,9 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
        }
        total_timer.end();
 
-       timers.print();
+       if (!in_warmup) {
+               timers.print();
+       }
 
        // Scale up the flow to the final size (if needed).
        if (finest_level == 0 || resize_strategy == DO_NOT_RESIZE_FLOW) {
@@ -1555,7 +1566,9 @@ GLuint Interpolate::exec(GLuint tex0, GLuint tex1, GLuint forward_flow_tex, GLui
        }
        pool.release_texture(flow_tex);
        total_timer.end();
-       timers.print();
+       if (!in_warmup) {
+               timers.print();
+       }
 
        return output_tex;
 }
@@ -1760,6 +1773,16 @@ void compute_flow_only(int argc, char **argv, int optind)
        glGenerateTextureMipmap(tex1_gray);
 
        DISComputeFlow compute_flow(width1, height1);
+
+       if (enable_warmup) {
+               in_warmup = true;
+               for (int i = 0; i < 10; ++i) {
+                       GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
+                       compute_flow.release_texture(final_tex);
+               }
+               in_warmup = false;
+       }
+
        GLuint final_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::RESIZE_FLOW_TO_FULL_SIZE);
 
        schedule_read<FlowType>(final_tex, width1, height1, filename0, filename1, flow_filename, "flow.ppm");
@@ -1854,6 +1877,19 @@ void interpolate_image(int argc, char **argv, int optind)
        gray.exec(tex1, tex1_gray, width1, height1);
        glGenerateTextureMipmap(tex1_gray);
 
+       if (enable_warmup) {
+               in_warmup = true;
+               for (int i = 0; i < 10; ++i) {
+                       GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
+                       GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
+                       GLuint interpolated_tex = interpolate.exec(tex0, tex1, forward_flow_tex, backward_flow_tex, width1, height1, 0.5f);
+                       compute_flow.release_texture(forward_flow_tex);
+                       compute_flow.release_texture(backward_flow_tex);
+                       interpolate.release_texture(interpolated_tex);
+               }
+               in_warmup = false;
+       }
+
        GLuint forward_flow_tex = compute_flow.exec(tex0_gray, tex1_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
        GLuint backward_flow_tex = compute_flow.exec(tex1_gray, tex0_gray, DISComputeFlow::DO_NOT_RESIZE_FLOW);
 
@@ -1882,7 +1918,8 @@ int main(int argc, char **argv)
                { "disable-timing", no_argument, 0, 1000 },
                { "detailed-timing", no_argument, 0, 1003 },
                { "ignore-variational-refinement", no_argument, 0, 1001 },  // Still calculates it, just doesn't apply it.
-               { "interpolate", no_argument, 0, 1002 }
+               { "interpolate", no_argument, 0, 1002 },
+               { "warmup", no_argument, 0, 1004 }
        };
 
        for ( ;; ) {
@@ -1914,6 +1951,9 @@ int main(int argc, char **argv)
                case 1003:
                        detailed_timing = true;
                        break;
+               case 1004:
+                       enable_warmup = true;
+                       break;
                default:
                        fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
                        exit(1);