GS/DX12/VK: ROV support.

Author: TJnotJT

bin/resources/shaders/dx11/tfx.fx

@@ -83,6 +83,11 @@ layout(std140, binding = 0) uniform cb21
 
 	float ScaledScaleFactor;
 	float RcpScaleFactor;
+
+	float pad0;
+	float pad1;
+	
+	uvec4 ColorMask;
 };
 
 in SHADER
@@ -1106,7 +1111,7 @@ void ps_main()
 
 	bool atst_pass = atst(C);
 
-#if PS_AFAIL == AFAIL_KEEP
+#if PS_ATST != PS_ATST_NONE && PS_AFAIL == AFAIL_KEEP
 	if (!atst_pass)
 		discard;
 #endif

bin/resources/shaders/opengl/tfx_fs.glsl

@@ -168,20 +168,6 @@ void ps_colclip_resolve()
 }
 #endif
 
-#ifdef ps_convert_float32_depth_to_color
-void ps_convert_float32_depth_to_color()
-{
-	o_col0 = sample_c(v_tex).r;
-}
-#endif
-
-#ifdef ps_convert_float32_color_to_depth
-void ps_convert_float32_color_to_depth()
-{
-	gl_FragDepth = sample_c(v_tex).r;
-}
-#endif
-
 #ifdef ps_convert_float32_32bits
 void ps_convert_float32_32bits()
 {
@@ -232,6 +218,20 @@ float rgb5a1_to_depth16(vec4 unorm)
 	return float(((c.r & 0xF8u) >> 3) | ((c.g & 0xF8u) << 2) | ((c.b & 0xF8u) << 7) | ((c.a & 0x80u) << 8)) * exp2(-32.0f);
 }
 
+#ifdef ps_convert_float32_depth_to_color
+void ps_convert_float32_depth_to_color()
+{
+	o_col0 = sample_c(v_tex).r;
+}
+#endif
+
+#ifdef ps_convert_float32_color_to_depth
+void ps_convert_float32_color_to_depth()
+{
+	gl_FragDepth = sample_c(v_tex).r;
+}
+#endif
+
 #ifdef ps_convert_float32_float24
 void ps_convert_float32_float24()
 {

bin/resources/shaders/vulkan/convert.glsl

@@ -318,9 +318,10 @@ void main()
 #define PS_NO_COLOR 0
 #define PS_NO_COLOR1 0
 #define PS_DATE 0
-#define PS_TEX_IS_FB 0
 #define PS_COLOR_FEEDBACK 0
 #define PS_DEPTH_FEEDBACK 0
+#define PS_ROV_COLOR 0
+#define PS_ROV_DEPTH 0
 #endif
 
 #define SW_BLEND (PS_BLEND_A || PS_BLEND_B || PS_BLEND_D)
@@ -331,10 +332,21 @@ void main()
 #define ZTST_NEEDS_DEPTH (PS_ZTST == ZTST_GEQUAL || PS_ZTST == ZTST_GREATER)
 
 #define PS_FEEDBACK_LOOP_IS_NEEDED_RT (PS_TEX_IS_FB == 1 || AFAIL_NEEDS_RT || PS_FBMASK || SW_BLEND_NEEDS_RT || SW_AD_TO_HW || (PS_DATE >= 5) || PS_COLOR_FEEDBACK)
-#define PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH (PS_DEPTH_FEEDBACK && (AFAIL_NEEDS_DEPTH || ZTST_NEEDS_DEPTH))
+#define PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH (PS_DEPTH_FEEDBACK && (AFAIL_NEEDS_DEPTH || ZTST_NEEDS_DEPTH || PS_ROV_DEPTH))
+
+#define PS_RETURN_COLOR_ROV (!PS_NO_COLOR && PS_ROV_COLOR)
+#define PS_RETURN_COLOR (!PS_NO_COLOR && !PS_ROV_COLOR)
+#define PS_RETURN_DEPTH_ROV (PS_ZWRITE && PS_ROV_DEPTH)
+#define PS_RETURN_DEPTH (PS_ZWRITE && !PS_ROV_DEPTH)
+#define PS_ROV_EARLYDEPTHSTENCIL (PS_ROV_COLOR && !PS_ROV_DEPTH && !PS_ZWRITE)
 
 #define NEEDS_TEX (PS_TFX != 4)
 
+#if PS_ROV_COLOR || PS_ROV_DEPTH
+#extension GL_ARB_fragment_shader_interlock : enable
+#extension GL_ARB_shader_image_load_store : enable
+#endif
+
 layout(std140, set = 0, binding = 1) uniform cb1
 {
 	vec3 FogColor;
@@ -355,6 +367,9 @@ layout(std140, set = 0, binding = 1) uniform cb1
 	mat4 DitherMatrix;
 	float ScaledScaleFactor;
 	float RcpScaleFactor;
+	float pad0;
+	float pad1;
+	uvec4 ColorMask;
 };
 
 layout(location = 0) in VSOutput
@@ -368,11 +383,31 @@ layout(location = 0) in VSOutput
 	#endif
 } vsIn;
 
-#if !PS_NO_COLOR && !PS_NO_COLOR1
-layout(location = 0, index = 0) out vec4 o_col0;
-layout(location = 0, index = 1) out vec4 o_col1;
-#elif !PS_NO_COLOR
-layout(location = 0) out vec4 o_col0;
+#if PS_RETURN_COLOR
+	#if !PS_NO_COLOR1
+		layout(location = 0, index = 0) out vec4 o_col0;
+		layout(location = 0, index = 1) out vec4 o_col1;
+	#elif !PS_NO_COLOR
+		layout(location = 0) out vec4 o_col0;
+	#endif
+#elif PS_RETURN_COLOR_ROV
+	vec4 o_col0;
+#endif
+
+#if PS_ROV_COLOR
+	layout(set = 1, binding = 5, rgba8) uniform restrict coherent image2D RtImageRov;
+	#if PS_FEEDBACK_LOOP_IS_NEEDED_RT
+		vec4 cachedRtValue;
+		vec4 sample_from_rt() { return cachedRtValue; }
+	#endif
+#endif
+
+#if PS_ROV_DEPTH
+	layout(set = 1, binding = 6, r32f) uniform restrict coherent image2D DepthImageRov;
+	#if PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH
+		float cachedDepthValue;
+		vec4 sample_from_depth() { return vec4(cachedDepthValue, 0.0f, 0.0f, 0.0f); }
+	#endif
 #endif
 
 #if NEEDS_TEX
@@ -382,25 +417,25 @@ layout(set = 1, binding = 1) uniform texture2D Palette;
 
 #if PS_FEEDBACK_LOOP_IS_NEEDED_RT || PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH
 	#if defined(DISABLE_TEXTURE_BARRIER) || defined(HAS_FEEDBACK_LOOP_LAYOUT)
-		#if PS_FEEDBACK_LOOP_IS_NEEDED_RT
+		#if (PS_FEEDBACK_LOOP_IS_NEEDED_RT && !PS_ROV_COLOR)
 			layout(set = 1, binding = 2) uniform texture2D RtSampler;
 			vec4 sample_from_rt() { return texelFetch(RtSampler, ivec2(gl_FragCoord.xy), 0); }
 		#endif
-		#if PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH
+		#if (PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH && !PS_ROV_DEPTH)
 			layout(set = 1, binding = 4) uniform texture2D DepthSampler;
 			vec4 sample_from_depth() { return texelFetch(DepthSampler, ivec2(gl_FragCoord.xy), 0); }
 		#endif
 	#else
 		// Must consider each case separately since the input attachment indices must be consecutive.
-		#if PS_FEEDBACK_LOOP_IS_NEEDED_RT && PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH
+		#if (PS_FEEDBACK_LOOP_IS_NEEDED_RT && !PS_ROV_COLOR) && (PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH && !PS_ROV_DEPTH)
 			layout(input_attachment_index = 0, set = 1, binding = 2) uniform subpassInput RtSampler;
 			layout(input_attachment_index = 1, set = 1, binding = 4) uniform subpassInput DepthSampler;
 			vec4 sample_from_rt() { return subpassLoad(RtSampler); }
 			vec4 sample_from_depth() { return subpassLoad(DepthSampler); }
-		#elif PS_FEEDBACK_LOOP_IS_NEEDED_RT
+		#elif (PS_FEEDBACK_LOOP_IS_NEEDED_RT && !PS_ROV_COLOR)
 			layout(input_attachment_index = 0, set = 1, binding = 2) uniform subpassInput RtSampler;
 			vec4 sample_from_rt() { return subpassLoad(RtSampler); }
-		#elif PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH
+		#elif (PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH && !PS_ROV_DEPTH)
 			layout(input_attachment_index = 0, set = 1, binding = 4) uniform subpassInput DepthSampler;
 			vec4 sample_from_depth() { return subpassLoad(DepthSampler); }
 		#endif
@@ -1282,6 +1317,20 @@ void ps_blend(inout vec4 Color, inout vec4 As_rgba)
 	#endif
 }
 
+#if PS_ROV_COLOR || PS_ROV_DEPTH
+layout(pixel_interlock_ordered) in;
+#endif
+
+#if PS_ROV_EARLYDEPTHSTENCIL
+layout(early_fragment_tests) in;
+#endif
+
+#if PS_ROV_COLOR || PS_ROV_DEPTH
+#define DISCARD rov_discard = true
+#else
+#define DISCARD discard
+#endif
+
 void main()
 {
 	float input_z = gl_FragCoord.z;
@@ -1291,20 +1340,36 @@ void main()
 	input_z = floor(input_z * exp2(32.0f)) * exp2(-32.0f);
 #endif
 
-#if PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH && (PS_ZTST == ZTST_GEQUAL || PS_ZTST == ZTST_GREATER)
+#if PS_ROV_COLOR || PS_ROV_DEPTH
+	beginInvocationInterlockARB();
+#endif
+
+#if PS_ROV_COLOR && PS_FEEDBACK_LOOP_IS_NEEDED_RT
+	cachedRtValue = imageLoad(RtImageRov, ivec2(gl_FragCoord.xy));
+#endif
+
+#if PS_ROV_DEPTH && PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH
+	cachedDepthValue = imageLoad(DepthImageRov, ivec2(gl_FragCoord.xy)).r;
+#endif
+
+#if (PS_ROV_COLOR && PS_FEEDBACK_LOOP_IS_NEEDED_RT) || (PS_ROV_DEPTH && PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH)
+	bool rov_discard = gl_HelperInvocation;
+#endif
+
+#if PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH && ZTST_NEEDS_DEPTH
 	#if PS_ZTST == ZTST_GEQUAL
 		if (input_z < sample_from_depth().r)
-			discard;
+			DISCARD;
 	#elif PS_ZTST == ZTST_GREATER
 		if (input_z <= sample_from_depth().r)
-			discard;
+			DISCARD;
 	#endif
 #endif // PS_ZTST
 
 #if PS_SCANMSK & 2
 	// fail depth test on prohibited lines
 	if ((int(gl_FragCoord.y) & 1) == (PS_SCANMSK & 1))
-		discard;
+		DISCARD;
 #endif
 #if PS_DATE >= 5
 
@@ -1332,7 +1397,7 @@ void main()
 #endif
 
 	if (bad) {
-		discard;
+		DISCARD;
 	}
 
 #endif		// PS_DATE >= 5
@@ -1343,7 +1408,7 @@ void main()
 	// the bad alpha value so we must keep it.
 
 	if (gl_PrimitiveID > stencil_ceil) {
-		discard;
+		DISCARD;
 	}
 #endif
 
@@ -1356,9 +1421,10 @@ void main()
 
 	bool atst_pass = atst(C);
 
-#if PS_AFAIL == AFAIL_KEEP
-	if (!atst_pass)
-		discard;
+#if PS_ATST != PS_ATST_NONE && PS_AFAIL == AFAIL_KEEP
+	if (!atst_pass) {
+		DISCARD;
+	}
 #endif
 
 #if SW_AD_TO_HW
@@ -1450,6 +1516,7 @@ void main()
 		alpha_blend.a = float(atst_pass);
 	#endif
 
+	// Output color scaling
 	#if !PS_NO_COLOR
 		#if PS_RTA_CORRECTION
 			o_col0.a = C.a / 128.0f;
@@ -1489,8 +1556,33 @@ void main()
 		input_z = min(input_z, MaxDepthPS);
 	#endif
 
-	#if PS_ZWRITE
+	// Writing back color (nothing to do for non-ROV)
+	#if PS_RETURN_COLOR_ROV
+		#if PS_FEEDBACK_LOOP_IS_NEEDED_RT
+			vec4 rt_col = sample_from_rt();
+
+			o_col0.r = bool(ColorMask.r) ? o_col0.r : rt_col.r;
+			o_col0.g = bool(ColorMask.g) ? o_col0.g : rt_col.g;
+			o_col0.b = bool(ColorMask.b) ? o_col0.b : rt_col.b;
+			o_col0.a = bool(ColorMask.a) ? o_col0.a : rt_col.a;
+
+			o_col0 = rov_discard ? rt_col : o_col0;
+		#endif
+
+		imageStore(RtImageRov, ivec2(gl_FragCoord.xy), o_col0);
+	#endif
+	
+	#if PS_RETURN_DEPTH
 		gl_FragDepth = input_z;
+	#elif PS_RETURN_DEPTH_ROV
+		#if PS_FEEDBACK_LOOP_IS_NEEDED_DEPTH
+			input_z = rov_discard ? sample_from_depth().r : input_z;
+		#endif
+		imageStore(DepthImageRov, ivec2(gl_FragCoord.xy), vec4(input_z, 0, 0, 1.0f));
+	#endif
+
+	#if PS_ROV_COLOR || PS_ROV_DEPTH
+		endInvocationInterlockARB();
 	#endif
 #endif // PS_DATE
 }

bin/resources/shaders/vulkan/tfx.glsl

@@ -91,6 +91,7 @@ static double s_last_barriers = 0;
 static double s_last_copies = 0;
 static double s_last_uploads = 0;
 static double s_last_readbacks = 0;
+static double s_last_target_transitions = 0;
 static u64 s_total_internal_draws = 0;
 static u64 s_total_draws = 0;
 static u64 s_total_render_passes = 0;
@@ -927,10 +928,16 @@ void GSRunner::DumpStats()
 #define main real_main
 #endif
 
+// FIXME: DEBUGGING - REMOVE
+std::string dump_name;
+
 static void CPUThreadMain(VMBootParameters* params, std::atomic<int>* ret)
 {
 	ret->store(EXIT_FAILURE);
 
+	// FIXME: DEBUGGING - REMOVE
+	dump_name = Path::GetFileName(params->filename);
+
 	if (VMManager::Internal::CPUThreadInitialize())
 	{
 		// apply new settings (e.g. pick up renderer change)

pcsx2-gsrunner/Main.cpp

@@ -789,6 +789,8 @@ struct Pcsx2Config
 					Mipmap : 1,
 					HWMipmap : 1,
 					HWAccurateAlphaTest: 1,
+					HWROV : 1,
+					HWROVLogging : 1,
 					ManualUserHacks : 1,
 					UserHacks_AlignSpriteX : 1,
 					UserHacks_CPUFBConversion : 1,
@@ -883,6 +885,18 @@ struct Pcsx2Config
 		s8 OverrideTextureBarriers = -1;
 		GSDepthFeedbackMode DepthFeedbackMode = GSDepthFeedbackMode::Auto;
 
+		u32 HWROVHistoryTextures = 16;
+		u32 HWROVHistoryDraws = 32;
+		u32 HWROVMaxBarriers = 16;
+		float HWROVHistoryWeightColor = 0.75f;
+		float HWROVHistoryWeightDepth = 0.75f;
+		float HWROVBarriersEnableColor = 2.0f;
+		float HWROVBarriersEnableDepth = 4.0f;
+		float HWROVBarriersDisableColor = 1.125f;
+		float HWROVBarriersDisableDepth = 1.25f;
+		u32 HWROVPreset = 1;
+		u32 HWROVUseBarriersVK = false;
+
 		u8 CAS_Sharpness = 50;
 		u8 ShadeBoost_Brightness = DEFAULT_SHADEBOOST_BRIGHTNESS;
 		u8 ShadeBoost_Contrast = DEFAULT_SHADEBOOST_CONTRAST;