X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=motion_search.frag;h=eb4f7c77fe2c6f8cb056a46caeaa89d58f0ee486;hb=3795723be95f2fe82f3c8b8b45b1a905b2c811fd;hp=a417834ff3fc72c5121fb144c3a373df6d71da3b;hpb=d567402be6554018899290ede7608c07540ff18c;p=nageru

diff --git a/motion_search.frag b/motion_search.frag
index a417834..eb4f7c7 100644
--- a/motion_search.frag
+++ b/motion_search.frag
@@ -35,21 +35,49 @@
   be ideal.
  */
 
-const uint patch_size = 12;
-const uint num_iterations = 16;
+in vec3 flow_tc;
+in vec2 patch_center;
+flat in int ref_layer, search_layer;
+out vec3 out_flow;
+
+uniform sampler2DArray flow_tex, image_tex;
+uniform usampler2DArray grad_tex;  // Also contains the corresponding reference image.
+uniform vec2 inv_image_size, inv_prev_level_size;
+uniform uint patch_size;
+uniform uint num_iterations;
+
+vec3 unpack_gradients(uint v)
+{
+	uint vi = v & 0xffu;
+	uint xi = (v >> 8) & 0xfffu;
+	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(vec3 tc)
+{
+	vec3 grad = unpack_gradients(texture(grad_tex, tc).x);
 
-in vec2 flow_tc;
-in vec2 patch_bottom_left_texel;  // Center of bottom-left texel of patch.
-out vec2 out_flow;
+	// 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.xy, vec2(0.0f))) || any(greaterThan(tc.xy, vec2(1.0f)))) {
+		grad.xy = vec2(0.0f);
+	}
 
-uniform sampler2D flow_tex, grad0_tex, image0_tex, image1_tex;
-uniform vec2 image_size, inv_image_size;
+	return grad;
+}
 
 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 image_size = textureSize(grad_tex, 0).xy;
+
+	// 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.
@@ -68,13 +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(vec3(tc, ref_layer));
 			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];
@@ -90,32 +118,22 @@ void main()
 
 	mat2 H_inv = inverse(H);
 
-	// Fetch the initial guess for the flow. (We need the normalization step
-	// because densification works by accumulating; see the comments on the
-	// Densify class.)
-	vec3 prev_flow = texture(flow_tex, flow_tc).xyz;
-	vec2 initial_u;
-	if (prev_flow.z < 1e-3) {
-		initial_u = vec2(0.0, 0.0);
-	} else {
-		initial_u = prev_flow.xy / prev_flow.z;
-	}
-
-	// 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;
-				du += grad * (warped - t);
+				vec3 grad = get_gradients(vec3(tc, ref_layer));
+				float t = grad.z;
+				float warped = texture(image_tex, vec3(tc + u_norm, search_layer)).x;
+				du += grad.xy * (warped - t);
 				warped_sum += warped;
 			}
 		}
@@ -131,16 +149,36 @@ 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.
-		du -= grad_sum * (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) {
+			first_mean_diff = mean_diff;
+		}
 
 		// Do the actual update.
-		u -= (H_inv * du) * inv_image_size;
+		u -= H_inv * du;
 	}
 
-	// Reject if we moved too far.
-	if (length((u - initial_u) * image_size) > patch_size) {
+	// 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;
 	}
 
-	out_flow = u;
+	// 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);
 }