X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=equations.frag;h=04e5370fd9d53baf5afda39ee140a526d6937945;hb=0c3bfa8f2fb956efce124dea05e006231603c333;hp=d4328decfb997cbdd2e12600c909ed44bdcb2df8;hpb=a10be5d6a8a0652e434b725528265729aff19d27;p=nageru

diff --git a/equations.frag b/equations.frag
index d4328de..04e5370 100644
--- a/equations.frag
+++ b/equations.frag
@@ -1,12 +1,14 @@
 #version 450 core
 
-in vec2 tc;
-out uvec4 equation;
+in vec3 tc0, tc_left0, tc_down0;
+in vec3 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;
-uniform sampler2D beta_0_tex;
-uniform sampler2D smoothness_x_tex, smoothness_y_tex;
+uniform sampler2DArray I_x_y_tex, I_t_tex;
+uniform sampler2DArray diff_flow_tex, base_flow_tex;
+uniform sampler2DArray beta_0_tex;
+uniform sampler2DArray diffusivity_tex;
 
 // Relative weighting of intensity term.
 uniform float delta;
@@ -14,6 +16,8 @@ uniform float delta;
 // Relative weighting of gradient term.
 uniform float gamma;
 
+uniform bool zero_diff_flow;
+
 // Similar to packHalf2x16, but the two values share exponent, and are stored
 // as 12-bit fixed point numbers multiplied by that exponent (the leading one
 // can't be implicit in this kind of format). This allows us to store a much
@@ -53,12 +57,27 @@ uint pack_floats_shared(float a, float b)
 	return (qa & 0xfffu) | ((qb & 0xfffu) << 12) | e;
 }
 
-void main()
+float zero_if_outside_border(vec4 val)
+{
+	if (val.w < 1.0f) {
+		// We hit the border (or more like half-way to it), so zero smoothness.
+		return 0.0f;
+	} else {
+		return val.x;
+	}
+}
+
+uvec4 compute_equation(vec3 tc, vec3 tc_left, vec3 tc_down)
 {
 	// Read the flow (on top of the u0/v0 flow).
-	vec2 diff_flow = texture(diff_flow_tex, tc).xy;
-	float du = diff_flow.x;
-	float dv = diff_flow.y;
+	float du, dv;
+	if (zero_diff_flow) {
+		du = dv = 0.0f;
+	} else {
+		vec2 diff_flow = texture(diff_flow_tex, tc).xy;
+		du = diff_flow.x;
+		dv = diff_flow.y;
+	}
 
 	// Read the first derivatives.
 	vec2 I_x_y = texture(I_x_y_tex, tc).xy;
@@ -69,8 +88,6 @@ void main()
 	// E_I term. Note that we don't square β_0, in line with DeepFlow;
 	// it's probably an error (see variational_refinement.txt),
 	// but squaring it seems to give worse results.
-	//
-  	// FIXME: Should the penalizer be adjusted for 0..1 intensity range instead of 0..255?
 	float beta_0 = texture(beta_0_tex, tc).x;
 	float k1 = delta * beta_0 * inversesqrt(beta_0 * (I_x * du + I_y * dv + I_t) * (I_x * du + I_y * dv + I_t) + 1e-6);
 	float A11 = k1 * I_x * I_x;
@@ -106,8 +123,7 @@ void main()
 	float I_xt = I_tx_ty.x;
 	float I_yt = I_tx_ty.y;
 
-	// E_G term. Same TODOs as E_I. Same normalization as beta_0
-	// (see derivatives.frag).
+	// E_G term. Same normalization as beta_0 (see derivatives.frag).
 	float beta_x = 1.0 / (I_xx * I_xx + I_xy * I_xy + 1e-7);
 	float beta_y = 1.0 / (I_xy * I_xy + I_yy * I_yy + 1e-7);
 	float k2 = gamma * inversesqrt(
@@ -124,10 +140,15 @@ void main()
 
 	// E_S term, sans the part on the right-hand side that deals with
 	// the neighboring pixels. The gamma is multiplied in in smoothness.frag.
-	float smooth_l = textureOffset(smoothness_x_tex, tc, ivec2(-1,  0)).x;
-	float smooth_r = texture(smoothness_x_tex, tc).x;
-	float smooth_d = textureOffset(smoothness_y_tex, tc, ivec2( 0, -1)).x;
-	float smooth_u = texture(smoothness_y_tex, tc).x;
+	//
+	// Note that we sample in-between two texels, which gives us the 0.5 *
+	// (x[-1] + x[0]) part for free. If one of the texels is a border
+	// texel, it will have zero alpha, and zero_if_outside_border() will
+	// set smoothness to zero.
+	float smooth_l = zero_if_outside_border(texture(diffusivity_tex, tc_left));
+	float smooth_r = zero_if_outside_border(textureOffset(diffusivity_tex, tc_left, ivec2(1, 0)));
+	float smooth_d = zero_if_outside_border(texture(diffusivity_tex, tc_down));
+	float smooth_u = zero_if_outside_border(textureOffset(diffusivity_tex, tc_down, ivec2(0, 1)));
 	A11 += smooth_l + smooth_r + smooth_d + smooth_u;
 	A22 += smooth_l + smooth_r + smooth_d + smooth_u;
 
@@ -142,8 +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;
+	}
 }