X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=sor.frag;h=9a8e1e40aa0a32490091c5e2eb7b1b671ad3ff2e;hb=3795723be95f2fe82f3c8b8b45b1a905b2c811fd;hp=047526cb0899a32db07d9e8c143578b03a3e09e9;hpb=20989aaee7ca26bb7aede452c55597fbcf26ebb8;p=nageru

diff --git a/sor.frag b/sor.frag
index 047526c..9a8e1e4 100644
--- a/sor.frag
+++ b/sor.frag
@@ -1,15 +1,15 @@
 #version 450 core
 
-in vec2 tc, tc_left, tc_down;
-in vec2 equation_tc_assuming_left, equation_tc_assuming_right;
+in vec3 tc, tc_left, tc_down;
+in vec3 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_red_tex, equation_black_tex;
+uniform sampler2DArray diff_flow_tex, diffusivity_tex;
+uniform usampler2DArray equation_red_tex, equation_black_tex;
 uniform int phase;
 
-uniform bool zero_diff_flow;
+uniform int num_nonzero_phases;
 
 // See pack_floats_shared() in equations.frag.
 vec2 unpack_floats_shared(uint c)
@@ -45,13 +45,14 @@ void main()
 	// just immediately throws away half of the warp, but it helps convergence
 	// a _lot_ (rough testing indicates that five iterations of SOR is as good
 	// as ~50 iterations of Jacobi). We could probably do better by reorganizing
-	// the data into two-values-per-pixel, so-called “twinning buffering”,
-	// but it makes for rather annoying code in the rest of the pipeline.
+	// the data into two-values-per-pixel, so-called “twinned buffering”;
+	// seemingly, it helps Haswell by ~15% on the SOR code, but GTX 950 not at all
+	// (at least not on 720p). Presumably the latter is already bandwidth bound.
 	int color = int(round(element_sum_idx)) & 1;
 	if (color != phase) discard;
 
 	uvec4 equation;
-	vec2 equation_tc;
+	vec3 equation_tc;
 	if ((int(round(element_x_idx)) & 1) == 0) {
 		equation_tc = equation_tc_assuming_left;
 	} else {
@@ -67,8 +68,12 @@ void main()
 	float inv_A22 = uintBitsToFloat(equation.z);
 	vec2 b = unpack_floats_shared(equation.w);
 
-	if (zero_diff_flow) {
-		diff_flow = vec2(0.0f);
+	const float omega = 1.8;  // Marginally better than 1.6, it seems.
+
+	if (num_nonzero_phases == 0) {
+		// Simplified version of the code below, assuming diff_flow == 0.0f everywhere.
+		diff_flow.x = omega * b.x * inv_A11;
+		diff_flow.y = omega * b.y * inv_A22;
 	} else {
 		// Subtract the missing terms from the right-hand side
 		// (it couldn't be done earlier, because we didn't know
@@ -82,14 +87,17 @@ void main()
 		b += smooth_r * textureOffset(diff_flow_tex, tc, ivec2( 1,  0)).xy;
 		b += smooth_d * textureOffset(diff_flow_tex, tc, ivec2( 0, -1)).xy;
 		b += smooth_u * textureOffset(diff_flow_tex, tc, ivec2( 0,  1)).xy;
-		diff_flow = texture(diff_flow_tex, tc).xy;
-	}
 
-	const float omega = 1.8;  // Marginally better than 1.6, it seems.
+		if (num_nonzero_phases == 1) {
+			diff_flow = vec2(0.0f);
+		} else {
+			diff_flow = texture(diff_flow_tex, tc).xy;
+		}
 
-	// From https://en.wikipedia.org/wiki/Successive_over-relaxation.
-	float sigma_u = A12 * diff_flow.y;
-	diff_flow.x += omega * ((b.x - sigma_u) * inv_A11 - diff_flow.x);
-	float sigma_v = A12 * diff_flow.x;
-	diff_flow.y += omega * ((b.y - sigma_v) * inv_A22 - diff_flow.y);
+		// From https://en.wikipedia.org/wiki/Successive_over-relaxation.
+		float sigma_u = A12 * diff_flow.y;
+		diff_flow.x += omega * ((b.x - sigma_u) * inv_A11 - diff_flow.x);
+		float sigma_v = A12 * diff_flow.x;
+		diff_flow.y += omega * ((b.y - sigma_v) * inv_A22 - diff_flow.y);
+	}
 }