Fix a GLSL syntax error that tripped up NVIDIA.

[nageru] / motion_search.frag

diff --git a/motion_search.frag b/motion_search.frag
index 163db2cbaf4be29dc74d749e07ce4037dda9b9f6..136d316ef4cfe43ac4e4604e12379cec81be9385 100644 (file)
--- a/motion_search.frag
+++ b/motion_search.frag
@@ -36,20 +36,23 @@
  */
 
 const uint patch_size = 12;
-const uint num_iterations = 16;
+const uint num_iterations = 8;
 
 in vec2 flow_tc;
-in vec2 patch_bottom_left_texel;  // Center of bottom-left texel of patch.
+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 vec2 inv_image_size, inv_prev_level_size;
 
 void main()
 {
-       // Lock patch_bottom_left_texel to an integer, so that we never get
-       // any bilinear artifacts for the gradient.
-       vec2 base = (round(patch_bottom_left_texel * image_size - vec2(0.5, 0.5)) + vec2(0.5, 0.5))
+       vec2 image_size = textureSize(image0_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.
+       vec2 base = (round(patch_center * image_size) - (0.5f * patch_size - 0.5f))
                * inv_image_size;
 
        // First, precompute the pseudo-Hessian for the template patch.
@@ -90,24 +93,21 @@ void main()
 
        mat2 H_inv = inverse(H);
 
-       // Fetch the initial guess for the flow.
-       vec2 initial_u = texture(flow_tex, flow_tc).xy * inv_prev_level_size;
-
-       // Note: The flow is in OpenGL coordinates [0..1], but the calculations
-       // generally come out in pixels since the gradient is in pixels,
-       // so we need to convert at the end.
+       // Fetch the initial guess for the flow, and convert from the previous size to this one.
+       vec2 initial_u = texture(flow_tex, flow_tc).xy * (image_size * inv_prev_level_size);
        vec2 u = initial_u;
        float mean_diff, first_mean_diff;
 
        for (uint i = 0; i < num_iterations; ++i) {
                vec2 du = vec2(0.0, 0.0);
                float warped_sum = 0.0f;
+               vec2 u_norm = u * inv_image_size;  // In [0..1] coordinates instead of pixels.
                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;
-                               float warped = texture(image1_tex, tc + u).x;
+                               float warped = texture(image1_tex, tc + u_norm).x;
                                du += grad * (warped - t);
                                warped_sum += warped;
                        }
@@ -124,7 +124,7 @@ 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) {
@@ -132,22 +132,28 @@ void main()
                }
 
                // Do the actual update.
-               u -= (H_inv * du) * inv_image_size;
+               u -= H_inv * du;
        }
 
-       // Reject if we moved too far. Also reject if the patch goes out-of-bounds
-       // (the paper does not mention this, but the code does, and it seems to be
-       // critical to avoid really bad behavior at the edges).
-       if ((length((u - initial_u) * image_size) > patch_size) ||
-            u.x * image_size.x < -(patch_size * 0.5f) ||
-            (1.0 - u.x) * image_size.x < -(patch_size * 0.5f) ||
-            u.y * image_size.y < -(patch_size * 0.5f) ||
-            (1.0 - u.y) * image_size.y < -(patch_size * 0.5f)) {
+       // Reject if we moved too far. Note that the paper says “too far” is the
+       // patch size, but the DIS code uses half of a patch size. The latter seems
+       // to give much better overall results.
+       //
+       // Also reject if the patch goes out-of-bounds (the paper does not mention this,
+       // but the code does, and it seems to be critical to avoid really bad behavior
+       // at the edges).
+       vec2 patch_center = (base * image_size - 0.5f) + patch_size * 0.5f + u;
+       if (length(u - initial_u) > (patch_size * 0.5f) ||
+           patch_center.x < -(patch_size * 0.5f) ||
+           image_size.x - patch_center.x < -(patch_size * 0.5f) ||
+           patch_center.y < -(patch_size * 0.5f) ||
+           image_size.y - patch_center.y < -(patch_size * 0.5f)) {
                u = initial_u;
                mean_diff = first_mean_diff;
        }
 
        // NOTE: The mean patch diff will be for the second-to-last patch,
        // not the true position of du. But hopefully, it will be very close.
+       u *= inv_image_size;
        out_flow = vec3(u.x, u.y, mean_diff);
 }