rendervulkan: Implement CTM

src/rendervulkan.cpp

@@ -3280,11 +3280,13 @@ float g_flInternalDisplayBrightnessNits = 500.0f;
 float g_flHDRItmSdrNits = 100.f;
 float g_flHDRItmTargetNits = 1000.f;
 
+#pragma pack(push, 1)
 struct BlitPushData_t
 {
 	vec2_t scale[k_nMaxLayers];
 	vec2_t offset[k_nMaxLayers];
 	float opacity[k_nMaxLayers];
+	glm::mat3x4 ctm[k_nMaxLayers];
 	uint32_t borderMask;
 	uint32_t frameId;
 	uint32_t blurRadius;
@@ -3309,7 +3311,21 @@ struct BlitPushData_t
             else
                 u_shaderFilter[i] = (uint32_t)layer->filter;
 
+			if (layer->ctm)
+			{
+				ctm[i] = layer->ctm->matrix;
+			}
+			else
+			{
+				ctm[i] = glm::mat3x4
+				{
+					1, 0, 0, 0,
+					0, 1, 0, 0,
+					0, 0, 1, 0
+				};
+			}
 		}
+
 		borderMask = frameInfo->borderMask();
 		frameId = s_frameId++;
 		blurRadius = frameInfo->blurRadius ? ( frameInfo->blurRadius * 2 ) - 1 : 0;
@@ -3340,16 +3356,16 @@ struct CaptureConvertBlitData_t
 	vec2_t scale[1];
 	vec2_t offset[1];
 	float opacity[1];
+	mat3x4 ctm[1];
 	uint32_t borderMask;
 	uint32_t halfExtent[2];
-	mat3x4 colorMatrix;
 
 	explicit CaptureConvertBlitData_t(float blit_scale, const mat3x4 &color_matrix) {
 		scale[0] = { blit_scale, blit_scale };
 		offset[0] = { 0.0f, 0.0f };
 		opacity[0] = 1.0f;
 		borderMask = 0;
-		colorMatrix = color_matrix;
+		ctm[0] = color_matrix;
 	}
 };
 
@@ -3385,6 +3401,7 @@ struct RcasPushData_t
 	vec2_t u_scale[k_nMaxLayers - 1];
 	vec2_t u_offset[k_nMaxLayers - 1];
 	float u_opacity[k_nMaxLayers];
+	glm::mat3x4 ctm[k_nMaxLayers];
 	uint32_t u_borderMask;
 	uint32_t u_frameId;
 	uint32_t u_c1;
@@ -3439,6 +3456,7 @@ struct NisPushData_t
 			tempX, tempY);
 	}
 };
+#pragma pack(pop)
 
 void bind_all_layers(CVulkanCmdBuffer* cmdBuffer, const struct FrameInfo_t *frameInfo)
 {

src/shaders/blit_push_data.h

@@ -1,8 +1,9 @@
-layout(binding = 0, std430)
+layout(binding = 0, scalar)
 uniform layers_t {
     vec2 u_scale[VKR_MAX_LAYERS];
     vec2 u_offset[VKR_MAX_LAYERS];
     float u_opacity[VKR_MAX_LAYERS];
+    mat3x4 u_ctm[VKR_MAX_LAYERS];
     uint u_borderMask;
     uint u_frameId;
     uint u_blur_radius;

src/shaders/composite.h

@@ -171,6 +171,7 @@ vec4 sampleLayerEx(sampler2D layerSampler, uint offsetLayerIdx, uint colorspaceL
     else {
         color = sampleRegular(layerSampler, coord, colorspace);
     }
+    color.rgb = color.rgba * u_ctm[offsetLayerIdx];
     color.rgb = apply_layer_color_mgmt(color.rgb, colorspace);
 
     return color;

src/shaders/cs_composite_rcas.comp

@@ -11,12 +11,13 @@ layout(
   local_size_y = 1,
   local_size_z = 1) in;
 
-layout(binding = 0, std430)
+layout(binding = 0, scalar)
 uniform layers_t {
     uvec2 u_layer0Offset;
     vec2 u_scale[VKR_MAX_LAYERS - 1];
     vec2 u_offset[VKR_MAX_LAYERS - 1];
     float u_opacity[VKR_MAX_LAYERS];
+    mat3x4 u_ctm[VKR_MAX_LAYERS];
     uint u_borderMask;
     uint u_frameId;
     uint u_c1;
@@ -72,6 +73,7 @@ void rcasComposite(uvec2 pos)
             }
 
             outputValue.rgb = colorspace_plane_degamma_tf(outputValue.rgb, colorspace);
+            outputValue.rgb = (vec4(outputValue.rgb, 1.0f) * u_ctm[0]).rgb;
             outputValue.rgb = apply_layer_color_mgmt(outputValue.rgb, colorspace);
             outputValue *= u_opacity[0];
         }

src/shaders/cs_easu.comp

@@ -11,7 +11,7 @@ layout(
   local_size_y = 1,
   local_size_z = 1) in;
 
-layout(binding = 0, std430)
+layout(binding = 0, scalar)
 uniform layers_t {
     uvec4 c1, c2, c3, c4;
 };

src/shaders/cs_easu_fp16.comp

@@ -12,7 +12,7 @@ layout(
   local_size_y = 1,
   local_size_z = 1) in;
 
-layout(binding = 0, std430)
+layout(binding = 0, scalar)
 uniform layers_t {
     uvec4 c1, c2, c3, c4;
 };

src/shaders/cs_gaussian_blur_horizontal.comp

@@ -11,11 +11,12 @@ layout(
   local_size_y = 8,
   local_size_z = 1) in;
 
-layout(binding = 0, std430)
+layout(binding = 0, scalar)
 uniform layers_t {
     vec2 u_scale[VKR_MAX_LAYERS];
     vec2 u_offset[VKR_MAX_LAYERS];
     float u_opacity[VKR_MAX_LAYERS];
+    mat3x4 u_ctm[VKR_MAX_LAYERS];
     uint u_borderMask;
     uint u_frameId;
     uint u_blur_radius;

src/shaders/cs_nis.comp

@@ -9,7 +9,7 @@
 
 #include "descriptor_set.h"
 
-layout(binding = 0, std430)
+layout(binding = 0, scalar)
 uniform const_buffer
 {
     float kDetectRatio;

src/shaders/cs_nis_fp16.comp

@@ -11,7 +11,7 @@
 
 #include "descriptor_set.h"
 
-layout(binding = 0, std430)
+layout(binding = 0, scalar)
 uniform const_buffer
 {
     float kDetectRatio;

src/shaders/cs_rgb_to_nv12.comp

@@ -23,14 +23,14 @@ const float u_nitsToLinear = 1.0f / 100.0f;
 const float u_itmSdrNits = 100.f;
 const float u_itmTargetNits = 1000.f;
 
-layout(binding = 0, std430)
+layout(binding = 0, scalar)
 uniform layers_t {
     vec2 u_scale[1];
     vec2 u_offset[1];
     float u_opacity[1];
+    mat3x4 u_ctm[1];
     uint u_borderMask;
     uvec2 u_halfExtent;
-    mat3x4 u_colorMatrix;
 };
 
 #include "composite.h"
@@ -63,11 +63,11 @@ void main() {
     };
 
     vec3 avg_color = (color[0] + color[1] + color[2] + color[3]) / 4.0f;
-    vec2 uv = applyColorMatrix(avg_color, u_colorMatrix).yz;
+    vec2 uv = applyColorMatrix(avg_color, u_ctm[0]).yz;
     imageStore(dst_chroma, chroma_uv, vec4(uv, 0.0f, 1.0f));
 
     for (int i = 0; i < 4; i++) {
-      float y = applyColorMatrix(color[i], u_colorMatrix).x;
+      float y = applyColorMatrix(color[i], u_ctm[0]).x;
       imageStore(dst_luma, luma_uv + offset_table[i], vec4(y, 0.0f, 0.0f, 1.0f));
     }
   }