Pack the gradients and image together into a single 32-bit texture; seems to help...

[nageru] / motion_search.frag

diff --git a/motion_search.frag b/motion_search.frag
index f67cbba51cc3cb06869bbebc2cb037189404d04d..9ef607c3fae79146b3b5ab97ab5298d8456b9faa 100644 (file)
--- a/motion_search.frag
+++ b/motion_search.frag
@@ -36,17 +36,44 @@
  */
 
 const uint patch_size = 12;
-const uint num_iterations = 16;
+const uint num_iterations = 8;
 
 in vec2 flow_tc;
 in vec2 patch_center;
 out vec3 out_flow;
 
-uniform sampler2D flow_tex, grad0_tex, image0_tex, image1_tex;
-uniform vec2 image_size, inv_image_size, inv_prev_level_size;
+uniform sampler2D flow_tex, image1_tex;
+uniform usampler2D grad0_tex;  // Also contains image0.
+uniform vec2 inv_image_size, inv_prev_level_size;
+
+vec3 unpack_gradients(uint v)
+{
+       uint vi = v & 0xff;
+       uint xi = (v >> 8) & 0xfff;
+       uint yi = v >> 20;
+       vec3 r = vec3(xi * (1.0f / 4095.0f) - 0.5f, yi * (1.0f / 4095.0f) - 0.5f, vi * (1.0f / 255.0f));
+       return r;
+}
+
+// Note: The third variable is the actual pixel value.
+vec3 get_gradients(vec2 tc)
+{
+       vec3 grad = unpack_gradients(texture(grad0_tex, tc).x);
+
+       // Zero gradients outside the image. (We'd do this with a sampler,
+       // but we want the repeat behavior for the actual texels, in the
+       // z channel.)
+       if (any(lessThan(tc, vec2(0.0f))) || any(greaterThan(tc, vec2(1.0f)))) {
+               grad.xy = vec2(0.0f);
+       }
+
+       return grad;
+}
 
 void main()
 {
+       vec2 image_size = textureSize(grad0_tex, 0);
+
        // Lock the patch center to an integer, so that we never get
        // any bilinear artifacts for the gradient. (NOTE: This assumes an
        // even patch size.) Then calculate the bottom-left texel of the patch.
@@ -69,13 +96,13 @@ void main()
        for (uint y = 0; y < patch_size; ++y) {
                for (uint x = 0; x < patch_size; ++x) {
                        vec2 tc = base + uvec2(x, y) * inv_image_size;
-                       vec2 grad = texture(grad0_tex, tc).xy;
+                       vec3 grad = get_gradients(tc);
                        H[0][0] += grad.x * grad.x;
                        H[1][1] += grad.y * grad.y;
                        H[0][1] += grad.x * grad.y;
 
-                       template_sum += texture(image0_tex, tc).x;
-                       grad_sum += grad;
+                       template_sum += grad.z;  // The actual template pixel value.
+                       grad_sum += grad.xy;
                }
        }
        H[1][0] = H[0][1];
@@ -103,10 +130,10 @@ void main()
                for (uint y = 0; y < patch_size; ++y) {
                        for (uint x = 0; x < patch_size; ++x) {
                                vec2 tc = base + uvec2(x, y) * inv_image_size;
-                               vec2 grad = texture(grad0_tex, tc).xy;
-                               float t = texture(image0_tex, tc).x;
+                               vec3 grad = get_gradients(tc);
+                               float t = grad.z;
                                float warped = texture(image1_tex, tc + u_norm).x;
-                               du += grad * (warped - t);
+                               du += grad.xy * (warped - t);
                                warped_sum += warped;
                        }
                }
@@ -122,7 +149,7 @@ void main()
                //   sum(S^T * (x - y)) = [what we calculated] - (µ1 - µ2) sum(S^T)
                //
                // so we can just subtract away the mean difference here.
-               mean_diff = (warped_sum - template_sum) * (1.0 / (patch_size * patch_size));
+               mean_diff = (warped_sum - template_sum) * (1.0 / float(patch_size * patch_size));
                du -= grad_sum * mean_diff;
 
                if (i == 0) {