summaryrefslogtreecommitdiff
path: root/servers/rendering/renderer_rd/shaders/copy.glsl
diff options
context:
space:
mode:
Diffstat (limited to 'servers/rendering/renderer_rd/shaders/copy.glsl')
-rw-r--r--servers/rendering/renderer_rd/shaders/copy.glsl91
1 files changed, 40 insertions, 51 deletions
diff --git a/servers/rendering/renderer_rd/shaders/copy.glsl b/servers/rendering/renderer_rd/shaders/copy.glsl
index d4d0ed0f56..ecf7bb9817 100644
--- a/servers/rendering/renderer_rd/shaders/copy.glsl
+++ b/servers/rendering/renderer_rd/shaders/copy.glsl
@@ -61,7 +61,7 @@ layout(rgba8, set = 3, binding = 0) uniform restrict writeonly image2D dest_buff
layout(rgba32f, set = 3, binding = 0) uniform restrict writeonly image2D dest_buffer;
#endif
-#ifdef MODE_GAUSSIAN_GLOW
+#ifdef MODE_GAUSSIAN_BLUR
shared vec4 local_cache[256];
shared vec4 temp_cache[128];
#endif
@@ -70,7 +70,7 @@ void main() {
// Pixel being shaded
ivec2 pos = ivec2(gl_GlobalInvocationID.xy);
-#ifndef MODE_GAUSSIAN_GLOW // Glow needs the extra threads
+#ifndef MODE_GAUSSIAN_BLUR // Gaussian blur needs the extra threads
if (any(greaterThanEqual(pos, params.section.zw))) { //too large, do nothing
return;
}
@@ -92,35 +92,11 @@ void main() {
#ifdef MODE_GAUSSIAN_BLUR
- //Simpler blur uses SIGMA2 for the gaussian kernel for a stronger effect
-
- if (bool(params.flags & FLAG_HORIZONTAL)) {
- ivec2 base_pos = (pos + params.section.xy) << 1;
- vec4 color = texelFetch(source_color, base_pos + ivec2(0, 0), 0) * 0.214607;
- color += texelFetch(source_color, base_pos + ivec2(1, 0), 0) * 0.189879;
- color += texelFetch(source_color, base_pos + ivec2(2, 0), 0) * 0.131514;
- color += texelFetch(source_color, base_pos + ivec2(3, 0), 0) * 0.071303;
- color += texelFetch(source_color, base_pos + ivec2(-1, 0), 0) * 0.189879;
- color += texelFetch(source_color, base_pos + ivec2(-2, 0), 0) * 0.131514;
- color += texelFetch(source_color, base_pos + ivec2(-3, 0), 0) * 0.071303;
- imageStore(dest_buffer, pos + params.target, color);
- } else {
- ivec2 base_pos = (pos + params.section.xy);
- vec4 color = texelFetch(source_color, base_pos + ivec2(0, 0), 0) * 0.38774;
- color += texelFetch(source_color, base_pos + ivec2(0, 1), 0) * 0.24477;
- color += texelFetch(source_color, base_pos + ivec2(0, 2), 0) * 0.06136;
- color += texelFetch(source_color, base_pos + ivec2(0, -1), 0) * 0.24477;
- color += texelFetch(source_color, base_pos + ivec2(0, -2), 0) * 0.06136;
- imageStore(dest_buffer, pos + params.target, color);
- }
-#endif
-
-#ifdef MODE_GAUSSIAN_GLOW
-
// First pass copy texture into 16x16 local memory for every 8x8 thread block
vec2 quad_center_uv = clamp(vec2(gl_GlobalInvocationID.xy + gl_LocalInvocationID.xy - 3.5) / params.section.zw, vec2(0.5 / params.section.zw), vec2(1.0 - 1.5 / params.section.zw));
uint dest_index = gl_LocalInvocationID.x * 2 + gl_LocalInvocationID.y * 2 * 16;
+#ifdef MODE_GLOW
if (bool(params.flags & FLAG_HIGH_QUALITY_GLOW)) {
vec2 quad_offset_uv = clamp((vec2(gl_GlobalInvocationID.xy + gl_LocalInvocationID.xy - 3.0)) / params.section.zw, vec2(0.5 / params.section.zw), vec2(1.0 - 1.5 / params.section.zw));
@@ -128,12 +104,15 @@ void main() {
local_cache[dest_index + 1] = (textureLod(source_color, quad_center_uv + vec2(1.0 / params.section.z, 0.0), 0) + textureLod(source_color, quad_offset_uv + vec2(1.0 / params.section.z, 0.0), 0)) * 0.5;
local_cache[dest_index + 16] = (textureLod(source_color, quad_center_uv + vec2(0.0, 1.0 / params.section.w), 0) + textureLod(source_color, quad_offset_uv + vec2(0.0, 1.0 / params.section.w), 0)) * 0.5;
local_cache[dest_index + 16 + 1] = (textureLod(source_color, quad_center_uv + vec2(1.0 / params.section.zw), 0) + textureLod(source_color, quad_offset_uv + vec2(1.0 / params.section.zw), 0)) * 0.5;
- } else {
+ } else
+#endif
+ {
local_cache[dest_index] = textureLod(source_color, quad_center_uv, 0);
local_cache[dest_index + 1] = textureLod(source_color, quad_center_uv + vec2(1.0 / params.section.z, 0.0), 0);
local_cache[dest_index + 16] = textureLod(source_color, quad_center_uv + vec2(0.0, 1.0 / params.section.w), 0);
local_cache[dest_index + 16 + 1] = textureLod(source_color, quad_center_uv + vec2(1.0 / params.section.zw), 0);
}
+#ifdef MODE_GLOW
if (bool(params.flags & FLAG_GLOW_FIRST_PASS)) {
// Tonemap initial samples to reduce weight of fireflies: https://graphicrants.blogspot.com/2013/12/tone-mapping.html
local_cache[dest_index] /= 1.0 + dot(local_cache[dest_index].rgb, vec3(0.299, 0.587, 0.114));
@@ -141,29 +120,33 @@ void main() {
local_cache[dest_index + 16] /= 1.0 + dot(local_cache[dest_index + 16].rgb, vec3(0.299, 0.587, 0.114));
local_cache[dest_index + 16 + 1] /= 1.0 + dot(local_cache[dest_index + 16 + 1].rgb, vec3(0.299, 0.587, 0.114));
}
-
+ const float kernel[4] = { 0.174938, 0.165569, 0.140367, 0.106595 };
+#else
+ // Simpler blur uses SIGMA2 for the gaussian kernel for a stronger effect.
+ const float kernel[4] = { 0.214607, 0.189879, 0.131514, 0.071303 };
+#endif
memoryBarrierShared();
barrier();
// Horizontal pass. Needs to copy into 8x16 chunk of local memory so vertical pass has full resolution
uint read_index = gl_LocalInvocationID.x + gl_LocalInvocationID.y * 32 + 4;
vec4 color_top = vec4(0.0);
- color_top += local_cache[read_index] * 0.174938;
- color_top += local_cache[read_index + 1] * 0.165569;
- color_top += local_cache[read_index + 2] * 0.140367;
- color_top += local_cache[read_index + 3] * 0.106595;
- color_top += local_cache[read_index - 1] * 0.165569;
- color_top += local_cache[read_index - 2] * 0.140367;
- color_top += local_cache[read_index - 3] * 0.106595;
+ color_top += local_cache[read_index] * kernel[0];
+ color_top += local_cache[read_index + 1] * kernel[1];
+ color_top += local_cache[read_index + 2] * kernel[2];
+ color_top += local_cache[read_index + 3] * kernel[3];
+ color_top += local_cache[read_index - 1] * kernel[1];
+ color_top += local_cache[read_index - 2] * kernel[2];
+ color_top += local_cache[read_index - 3] * kernel[3];
vec4 color_bottom = vec4(0.0);
- color_bottom += local_cache[read_index + 16] * 0.174938;
- color_bottom += local_cache[read_index + 1 + 16] * 0.165569;
- color_bottom += local_cache[read_index + 2 + 16] * 0.140367;
- color_bottom += local_cache[read_index + 3 + 16] * 0.106595;
- color_bottom += local_cache[read_index - 1 + 16] * 0.165569;
- color_bottom += local_cache[read_index - 2 + 16] * 0.140367;
- color_bottom += local_cache[read_index - 3 + 16] * 0.106595;
+ color_bottom += local_cache[read_index + 16] * kernel[0];
+ color_bottom += local_cache[read_index + 1 + 16] * kernel[1];
+ color_bottom += local_cache[read_index + 2 + 16] * kernel[2];
+ color_bottom += local_cache[read_index + 3 + 16] * kernel[3];
+ color_bottom += local_cache[read_index - 1 + 16] * kernel[1];
+ color_bottom += local_cache[read_index - 2 + 16] * kernel[2];
+ color_bottom += local_cache[read_index - 3 + 16] * kernel[3];
// rotate samples to take advantage of cache coherency
uint write_index = gl_LocalInvocationID.y * 2 + gl_LocalInvocationID.x * 16;
@@ -174,18 +157,24 @@ void main() {
memoryBarrierShared();
barrier();
+ // If destination outside of texture, can stop doing work now
+ if (any(greaterThanEqual(pos, params.section.zw))) {
+ return;
+ }
+
// Vertical pass
uint index = gl_LocalInvocationID.y + gl_LocalInvocationID.x * 16 + 4;
vec4 color = vec4(0.0);
- color += temp_cache[index] * 0.174938;
- color += temp_cache[index + 1] * 0.165569;
- color += temp_cache[index + 2] * 0.140367;
- color += temp_cache[index + 3] * 0.106595;
- color += temp_cache[index - 1] * 0.165569;
- color += temp_cache[index - 2] * 0.140367;
- color += temp_cache[index - 3] * 0.106595;
+ color += temp_cache[index] * kernel[0];
+ color += temp_cache[index + 1] * kernel[1];
+ color += temp_cache[index + 2] * kernel[2];
+ color += temp_cache[index + 3] * kernel[3];
+ color += temp_cache[index - 1] * kernel[1];
+ color += temp_cache[index - 2] * kernel[2];
+ color += temp_cache[index - 3] * kernel[3];
+#ifdef MODE_GLOW
if (bool(params.flags & FLAG_GLOW_FIRST_PASS)) {
// Undo tonemap to restore range: https://graphicrants.blogspot.com/2013/12/tone-mapping.html
color /= 1.0 - dot(color.rgb, vec3(0.299, 0.587, 0.114));
@@ -205,7 +194,7 @@ void main() {
color = min(color * feedback, vec4(params.glow_luminance_cap));
}
-
+#endif
imageStore(dest_buffer, pos + params.target, color);
#endif
generated by cgit v1.2.3 (git 2.25.1) at 2024-11-08 06:40:37 +0000