]> git.sesse.net Git - nageru/commitdiff
Split the equation texture in two, which speeds up SOR by ~30%.
authorSteinar H. Gunderson <sgunderson@bigfoot.com>
Sat, 4 Aug 2018 12:24:11 +0000 (14:24 +0200)
committerSteinar H. Gunderson <sgunderson@bigfoot.com>
Sat, 4 Aug 2018 12:24:11 +0000 (14:24 +0200)
equations.frag
equations.vert
flow.cpp
sor.frag
sor.vert

index ae32ed7fd72e48b166211a828a3d331335827a45..8bf0cbb4f7a68bd49fc79be400365fc973ee7ad4 100644 (file)
@@ -1,7 +1,9 @@
 #version 450 core
 
-in vec2 tc, tc_left, tc_down;
-out uvec4 equation;
+in vec2 tc0, tc_left0, tc_down0;
+in vec2 tc1, tc_left1, tc_down1;
+in float line_offset;
+out uvec4 equation_red, equation_black;
 
 uniform sampler2D I_x_y_tex, I_t_tex;
 uniform sampler2D diff_flow_tex, base_flow_tex;
@@ -65,7 +67,7 @@ float zero_if_outside_border(vec4 val)
        }
 }
 
-void main()
+uvec4 compute_equation(vec2 tc, vec2 tc_left, vec2 tc_down)
 {
        // Read the flow (on top of the u0/v0 flow).
        float du, dv;
@@ -161,8 +163,25 @@ void main()
        b2 += laplacian.y;
 
        // Encode the equation down into four uint32s.
-       equation.x = floatBitsToUint(1.0 / A11);
-       equation.y = floatBitsToUint(A12);
-       equation.z = floatBitsToUint(1.0 / A22);
-       equation.w = pack_floats_shared(b1, b2);
+       uvec4 ret;
+       ret.x = floatBitsToUint(1.0 / A11);
+       ret.y = floatBitsToUint(A12);
+       ret.z = floatBitsToUint(1.0 / A22);
+       ret.w = pack_floats_shared(b1, b2);
+       return ret;
+}
+
+void main()
+{
+       uvec4 eq0 = compute_equation(tc0, tc_left0, tc_down0);
+       uvec4 eq1 = compute_equation(tc1, tc_left1, tc_down1);
+
+       if ((int(round(line_offset)) & 1) == 1) {
+               // Odd line, so the right value is red.
+               equation_red = eq1;
+               equation_black = eq0;
+       } else {
+               equation_red = eq0;
+               equation_black = eq1;
+       }
 }
index 6ace5775e0017f69261dd20f35a9ca68602c651d..eec3c170aa6c7cb2361d1d2a6daa744ff0ad85fd 100644 (file)
@@ -1,7 +1,9 @@
 #version 450 core
 
 layout(location=0) in vec2 position;
-out vec2 tc, tc_left, tc_down;
+out vec2 tc0, tc_left0, tc_down0;
+out vec2 tc1, tc_left1, tc_down1;
+out float line_offset;
 
 uniform sampler2D diffusivity_tex;
 
@@ -14,7 +16,21 @@ void main()
        //   0.000  0.000 -2.000 -1.000
        //   0.000  0.000  0.000  1.000
        gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0);
-       tc = position;
-       tc_left = vec2(tc.x - 0.5f / textureSize(diffusivity_tex, 0).x, tc.y);
-       tc_down = vec2(tc.x, tc.y - 0.5f / textureSize(diffusivity_tex, 0).y);
+
+       const vec2 half_texel = 0.5f / textureSize(diffusivity_tex, 0);
+
+       vec2 tc = position;
+       vec2 tc_left = vec2(tc.x - half_texel.x, tc.y);
+       vec2 tc_down = vec2(tc.x, tc.y - half_texel.y);
+
+       // Adjust for different texel centers.
+       tc0 = vec2(tc.x - half_texel.x, tc.y);
+       tc_left0 = vec2(tc_left.x - half_texel.x, tc_left.y);
+       tc_down0 = vec2(tc_down.x - half_texel.x, tc_down.y);
+
+       tc1 = vec2(tc.x + half_texel.x, tc.y);
+       tc_left1 = vec2(tc_left.x + half_texel.x, tc_left.y);
+       tc_down1 = vec2(tc_down.x + half_texel.x, tc_down.y);
+
+       line_offset = position.y * textureSize(diffusivity_tex, 0).y - 0.5f;
 }
index 620b511a19175a185d44cbedb2628c1ac7fc509a..1e4873362c1586e0f3a1095e74b38ec1b4a6bb2a 100644 (file)
--- a/flow.cpp
+++ b/flow.cpp
@@ -679,14 +679,23 @@ void ComputeDiffusivity::exec(GLuint flow_tex, GLuint diff_flow_tex, GLuint diff
 // 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 diffusivity_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;
@@ -716,7 +725,7 @@ SetupEquations::SetupEquations()
        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 diffusivity_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);
 
@@ -730,9 +739,9 @@ void SetupEquations::exec(GLuint I_x_y_tex, GLuint I_t_tex, GLuint diff_flow_tex
        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);
 }
 
@@ -743,7 +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 diffusivity_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;
@@ -753,7 +762,7 @@ private:
        GLuint sor_program;
 
        GLuint uniform_diff_flow_tex;
-       GLuint uniform_equation_tex;
+       GLuint uniform_equation_red_tex, uniform_equation_black_tex;
        GLuint uniform_diffusivity_tex;
        GLuint uniform_phase, uniform_zero_diff_flow;
 };
@@ -765,19 +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_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 diffusivity_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_diffusivity_tex, 1, diffusivity_tex, zero_border_sampler);
-       bind_sampler(sor_program, uniform_equation_tex, 2, equation_tex, nearest_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);
 
@@ -1110,7 +1121,8 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
 
                // 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 diffusivity term for each pixel.
@@ -1122,14 +1134,14 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                        // 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_tex, level_width, level_height, outer_idx == 0);
+                               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);
                        }
 
                        // Run a few SOR (or quasi-SOR, since we're not really Jacobi) iterations.
                        // Note that these are to/from the same texture.
                        {
                                ScopedTimer timer("SOR", &varref_timer);
-                               sor.exec(du_dv_tex, equation_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer);
+                               sor.exec(du_dv_tex, equation_red_tex, equation_black_tex, diffusivity_tex, level_width, level_height, 5, outer_idx == 0, &timer);
                        }
                }
 
@@ -1137,7 +1149,8 @@ GLuint DISComputeFlow::exec(GLuint tex0, GLuint tex1, ResizeStrategy resize_stra
                pool.release_texture(I_x_y_tex);
                pool.release_texture(beta_0_tex);
                pool.release_texture(diffusivity_tex);
-               pool.release_texture(equation_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.
index 91bf831e212a52b48204af6246d28c6419c60120..047526cb0899a32db07d9e8c143578b03a3e09e9 100644 (file)
--- a/sor.frag
+++ b/sor.frag
@@ -1,11 +1,12 @@
 #version 450 core
 
 in vec2 tc, tc_left, tc_down;
-in float element_sum_idx;
+in vec2 equation_tc_assuming_left, equation_tc_assuming_right;
+in float element_x_idx, element_sum_idx;
 out vec2 diff_flow;
 
 uniform sampler2D diff_flow_tex, diffusivity_tex;
-uniform usampler2D equation_tex;
+uniform usampler2D equation_red_tex, equation_black_tex;
 uniform int phase;
 
 uniform bool zero_diff_flow;
@@ -49,7 +50,18 @@ void main()
        int color = int(round(element_sum_idx)) & 1;
        if (color != phase) discard;
 
-       uvec4 equation = texture(equation_tex, tc);
+       uvec4 equation;
+       vec2 equation_tc;
+       if ((int(round(element_x_idx)) & 1) == 0) {
+               equation_tc = equation_tc_assuming_left;
+       } else {
+               equation_tc = equation_tc_assuming_right;
+       }
+       if (phase == 0) {
+               equation = texture(equation_red_tex, equation_tc);
+       } else {
+               equation = texture(equation_black_tex, equation_tc);
+       }
        float inv_A11 = uintBitsToFloat(equation.x);
        float A12 = uintBitsToFloat(equation.y);
        float inv_A22 = uintBitsToFloat(equation.z);
index 55d1a90f40e013da429098147aa2d68c2d2509d1..32498decf948f4afe6972e313e822918416405f4 100644 (file)
--- a/sor.vert
+++ b/sor.vert
@@ -2,9 +2,12 @@
 
 layout(location=0) in vec2 position;
 out vec2 tc, tc_left, tc_down;
+out vec2 equation_tc_assuming_left, equation_tc_assuming_right;
+out float element_x_idx;
 out float element_sum_idx;
 
 uniform sampler2D diff_flow_tex, diffusivity_tex;
+uniform usampler2D equation_red_tex;
 
 void main()
 {
@@ -19,6 +22,17 @@ void main()
        tc_left = vec2(tc.x - 0.5f / textureSize(diffusivity_tex, 0).x, tc.y);
        tc_down = vec2(tc.x, tc.y - 0.5f / textureSize(diffusivity_tex, 0).y);
 
-       vec2 element_idx = position * textureSize(diff_flow_tex, 0) - 0.5;
+       // The equation textures have half the horizontal width, so we need to adjust the texel centers.
+       // It becomes extra tricky since the SOR texture might be of odd size, and then
+       // the equation texture is not exactly half the size.
+       vec2 element_idx = position * textureSize(diff_flow_tex, 0) - 0.5f;
+       float equation_texel_number_assuming_left = element_idx.x / 2.0f;
+       float equation_texel_number_assuming_right = (element_idx.x - 1.0f) / 2.0f;
+       equation_tc_assuming_left.x = (equation_texel_number_assuming_left + 0.5f) / textureSize(equation_red_tex, 0).x;
+       equation_tc_assuming_right.x = (equation_texel_number_assuming_right + 0.5f) / textureSize(equation_red_tex, 0).x;
+       equation_tc_assuming_left.y = tc.y;
+       equation_tc_assuming_right.y = tc.y;
+
+       element_x_idx = element_idx.x;
        element_sum_idx = element_idx.x + element_idx.y;
 }