Object Mode Engine: Optimize outline passes.

Group texture fetches to hide latency. 3.2ms -> 2.2ms (constant time improvement, not depending on scene complexity)

Could optimize further with textureGather (require OpenGL 4.0).

3 files changed, 50 insertions, 39 deletions

diff --git a/source/blender/draw/modes/object_mode.c b/source/blender/draw/modes/object_mode.c
index 8176fbcda5c..1e70e942d07 100644
--- a/source/blender/draw/modes/object_mode.c
+++ b/source/blender/draw/modes/object_mode.c
@@ -696,7 +696,7 @@ static void OBJECT_cache_init(void *vedata)
 
 	{
 		DRWState state = DRW_STATE_WRITE_COLOR | DRW_STATE_WRITE_DEPTH | DRW_STATE_DEPTH_LESS | DRW_STATE_WIRE;
-		psl->outlines = DRW_pass_create("Outlines Pass", state);
+		psl->outlines = DRW_pass_create("Outlines Depth Pass", state);
 
 		GPUShader *sh = GPU_shader_get_builtin_shader(GPU_SHADER_3D_UNIFORM_COLOR);
 
@@ -725,7 +725,7 @@ static void OBJECT_cache_init(void *vedata)
 		static bool bTrue = true;
 		static bool bFalse = false;
 
-		psl->outlines_search = DRW_pass_create("Outlines Expand Pass", state);
+		psl->outlines_search = DRW_pass_create("Outlines Detect Pass", state);
 
 		DRWShadingGroup *grp = DRW_shgroup_create(e_data.outline_detect_sh, psl->outlines_search);
 		DRW_shgroup_uniform_buffer(grp, "outlineColor", &e_data.outlines_color_tx);
diff --git a/source/blender/draw/modes/shaders/object_outline_detect_frag.glsl b/source/blender/draw/modes/shaders/object_outline_detect_frag.glsl
index 5565a0f1e09..dc0ea938436 100644
--- a/source/blender/draw/modes/shaders/object_outline_detect_frag.glsl
+++ b/source/blender/draw/modes/shaders/object_outline_detect_frag.glsl
@@ -32,30 +32,48 @@ void search_outline(ivec2 uv, vec4 ref_col, inout bool ref_occlu, inout bool out
 void main()
 {
 	ivec2 uv = ivec2(gl_FragCoord.xy);
+
+	vec4 color[4];
+	/* Idea : Use a 16bit ID to identify the color
+	 * and store the colors in a UBO. And fetch all ids
+	 * for discontinuity check with one textureGather \o/ */
 	vec4 ref_col = texelFetch(outlineColor, uv, 0).rgba;
+	color[0] = texelFetchOffset(outlineColor, uv, 0, ivec2( 1,  0)).rgba;
+	color[1] = texelFetchOffset(outlineColor, uv, 0, ivec2( 0,  1)).rgba;
+	color[2] = texelFetchOffset(outlineColor, uv, 0, ivec2(-1,  0)).rgba;
+	color[3] = texelFetchOffset(outlineColor, uv, 0, ivec2( 0, -1)).rgba;
 
+	/* TODO GATHER */
+	vec4 depths;
 	float depth = texelFetch(outlineDepth, uv, 0).r;
-	/* Modulate color if occluded */
+	depths.x = texelFetchOffset(outlineDepth, uv, 0, ivec2( 1,  0)).r;
+	depths.y = texelFetchOffset(outlineDepth, uv, 0, ivec2( 0,  1)).r;
+	depths.z = texelFetchOffset(outlineDepth, uv, 0, ivec2(-1,  0)).r;
+	depths.w = texelFetchOffset(outlineDepth, uv, 0, ivec2( 0, -1)).r;
+
+	vec4 scene_depths;
 	float scene_depth = texelFetch(sceneDepth, uv, 0).r;
+	scene_depths.x = texelFetchOffset(sceneDepth, uv, 0, ivec2( 1,  0)).r;
+	scene_depths.y = texelFetchOffset(sceneDepth, uv, 0, ivec2( 0,  1)).r;
+	scene_depths.z = texelFetchOffset(sceneDepth, uv, 0, ivec2(-1,  0)).r;
+	scene_depths.w = texelFetchOffset(sceneDepth, uv, 0, ivec2( 0, -1)).r;
 
 	bool ref_occlu = (depth > scene_depth);
-
 	bool outline = false;
 
+#if 1
+	bvec4 occlu = (!ref_occlu) ? notEqual(greaterThan(depths, scene_depths), bvec4(ref_occlu)) : bvec4(false);
+	outline = (!outline) ? (color[0] != ref_col) || occlu.x : true;
+	outline = (!outline) ? (color[1] != ref_col) || occlu.y : true;
+	outline = (!outline) ? (color[2] != ref_col) || occlu.z : true;
+	outline = (!outline) ? (color[3] != ref_col) || occlu.w : true;
+#else
 	search_outline(uv + ivec2( 1,  0), ref_col, ref_occlu, outline);
 	search_outline(uv + ivec2( 0,  1), ref_col, ref_occlu, outline);
 	search_outline(uv + ivec2(-1,  0), ref_col, ref_occlu, outline);
 	search_outline(uv + ivec2( 0, -1), ref_col, ref_occlu, outline);
+#endif
 
 	FragColor = ref_col;
-
-	/* We Hit something ! */
-	if (outline) {
-		if (ref_occlu) {
-			FragColor.a *= alphaOcclu;
-		}
-	}
-	else {
-		FragColor.a = 0.0;
-	}
+	FragColor.a *= (outline) ? (ref_occlu) ? alphaOcclu : 1.0 : 0.0;
 }
diff --git a/source/blender/draw/modes/shaders/object_outline_expand_frag.glsl b/source/blender/draw/modes/shaders/object_outline_expand_frag.glsl
index e0568d1157a..c753e3ab39c 100644
--- a/source/blender/draw/modes/shaders/object_outline_expand_frag.glsl
+++ b/source/blender/draw/modes/shaders/object_outline_expand_frag.glsl
@@ -9,37 +9,30 @@ uniform sampler2D outlineDepth;
 uniform float alpha;
 uniform bool doExpand;
 
-void search_outline(ivec2 uv, inout bool found_edge)
-{
-	if (!found_edge) {
-		vec4 color = texelFetch(outlineColor, uv, 0).rgba;
-		if (color.a != 0.0) {
-			if (doExpand || color.a != 1.0) {
-				FragColor = color;
-				found_edge = true;
-			}
-		}
-	}
-}
-
 void main()
 {
 	ivec2 uv = ivec2(gl_FragCoord.xy);
 	FragColor = texelFetch(outlineColor, uv, 0).rgba;
 	float depth = texelFetch(outlineDepth, uv, 0).r;
 
-	if (FragColor.a != 0.0 || (depth == 1.0 && !doExpand))
-		return;
+	vec4 color[4];
+	color[0] = texelFetchOffset(outlineColor, uv, 0, ivec2( 1,  0)).rgba;
+	color[1] = texelFetchOffset(outlineColor, uv, 0, ivec2( 0,  1)).rgba;
+	color[2] = texelFetchOffset(outlineColor, uv, 0, ivec2(-1,  0)).rgba;
+	color[3] = texelFetchOffset(outlineColor, uv, 0, ivec2( 0, -1)).rgba;
+
+	vec4 values = vec4(color[0].a, color[1].a, color[2].a, color[3].a);
+
+	bool is_blank_pixel = !(FragColor.a != 0.0 || (depth == 1.0 && !doExpand));
+	vec4 tests = vec4(is_blank_pixel);
+	tests *= step(vec4(1e-6), values); /* (color.a != 0.0) */
+	tests *= (doExpand) ? vec4(1.0) : step(values, vec4(0.999)); /* (doExpand || color.a != 1.0) */
+	bvec4 btests = equal(tests, vec4(1.0));
 
-	bool found_edge = false;
-	search_outline(uv + ivec2( 1,  0), found_edge);
-	search_outline(uv + ivec2( 0,  1), found_edge);
-	search_outline(uv + ivec2(-1,  0), found_edge);
-	search_outline(uv + ivec2( 0, -1), found_edge);
+	FragColor = (btests.x) ? color[0] : FragColor;
+	FragColor = (btests.y) ? color[1] : FragColor;
+	FragColor = (btests.z) ? color[2] : FragColor;
+	FragColor = (btests.w) ? color[3] : FragColor;
 
-	/* We Hit something ! */
-	if (found_edge) {
-		/* only change alpha */
-		FragColor.a *= alpha;
-	}
+	FragColor *= (is_blank_pixel) ? alpha : 1.0;
 }