GS:HW: Avoid using blend + fbfetch for AFAIL RGB_ONLY

Author: TellowKrinkle

bin/resources/shaders/opengl/tfx_fs.glsl

@@ -25,7 +25,8 @@
 #define SW_AD_TO_HW (PS_BLEND_C == 1 && PS_A_MASKED)
 #define PS_PRIMID_INIT (PS_DATE == 1 || PS_DATE == 2)
 #define NEEDS_RT_EARLY (PS_TEX_IS_FB == 1 || PS_DATE >= 5)
-#define NEEDS_RT (NEEDS_RT_EARLY || (!PS_PRIMID_INIT && (PS_FBMASK || SW_BLEND_NEEDS_RT || SW_AD_TO_HW)))
+#define NEEDS_RT_FOR_AFAIL (PS_AFAIL == 3 && PS_NO_COLOR1)
+#define NEEDS_RT (NEEDS_RT_EARLY || NEEDS_RT_FOR_AFAIL || (!PS_PRIMID_INIT && (PS_FBMASK || SW_BLEND_NEEDS_RT || SW_AD_TO_HW)))
 #define NEEDS_TEX (PS_TFX != 4)
 
 layout(std140, binding = 0) uniform cb21
@@ -1114,7 +1115,7 @@ void ps_main()
 
 	ps_fbmask(C);
 
-#if PS_AFAIL == 3 // RGB_ONLY
+#if PS_AFAIL == 3 && !PS_NO_COLOR1 // RGB_ONLY
 	// Use alpha blend factor to determine whether to update A.
 	alpha_blend.a = float(atst_pass);
 #endif
@@ -1130,6 +1131,10 @@ void ps_main()
 	#else
 		SV_Target0.rgb = C.rgb / 255.0f;
 	#endif
+	#if PS_AFAIL == 3 && !PS_NO_COLOR1 // RGB_ONLY, no dual src blend
+		if (!atst_pass)
+			SV_Target0.a = sample_from_rt().a;
+	#endif
 	#if !PS_NO_COLOR1
 		SV_Target1 = alpha_blend;
 	#endif

bin/resources/shaders/vulkan/tfx.glsl

@@ -299,8 +299,9 @@ void main()
 #define SW_BLEND (PS_BLEND_A || PS_BLEND_B || PS_BLEND_D)
 #define SW_BLEND_NEEDS_RT (SW_BLEND && (PS_BLEND_A == 1 || PS_BLEND_B == 1 || PS_BLEND_C == 1 || PS_BLEND_D == 1))
 #define SW_AD_TO_HW (PS_BLEND_C == 1 && PS_A_MASKED)
+#define AFAIL_NEEDS_RT (PS_AFAIL == 3 && PS_NO_COLOR1)
 
-#define PS_FEEDBACK_LOOP_IS_NEEDED (PS_TEX_IS_FB == 1 || PS_FBMASK || SW_BLEND_NEEDS_RT || SW_AD_TO_HW || (PS_DATE >= 5))
+#define PS_FEEDBACK_LOOP_IS_NEEDED (PS_TEX_IS_FB == 1 || AFAIL_NEEDS_RT || PS_FBMASK || SW_BLEND_NEEDS_RT || SW_AD_TO_HW || (PS_DATE >= 5))
 
 #define NEEDS_TEX (PS_TFX != 4)
 
@@ -1381,7 +1382,7 @@ void main()
 
 	ps_fbmask(C);
 
-	#if PS_AFAIL == 3 // RGB_ONLY
+	#if PS_AFAIL == 3 && !PS_NO_COLOR1 // RGB_ONLY
 		// Use alpha blend factor to determine whether to update A.
 		alpha_blend.a = float(atst_pass);
 	#endif
@@ -1400,6 +1401,10 @@ void main()
 		#if !PS_NO_COLOR1
 			o_col1 = alpha_blend;
 		#endif
+		#if PS_AFAIL == 3 && PS_NO_COLOR1 // RGB_ONLY, no dual src blend
+			if (!atst_pass)
+				o_col0.a = sample_from_rt().a;
+		#endif
 	#endif
 
 	#if PS_ZCLAMP

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -6240,23 +6240,32 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 			// Blending might be off, ensure it's enabled.
 			// We write the alpha pass/fail to SRC1_ALPHA, which is used to update A.
 			m_conf.ps.afail = AFAIL_RGB_ONLY;
-			m_conf.ps.no_color1 = false;
-			if (!m_conf.blend.enable)
+			if ((features.framebuffer_fetch && m_conf.require_one_barrier) || m_conf.require_full_barrier)
 			{
-				m_conf.blend = GSHWDrawConfig::BlendState(true, GSDevice::CONST_ONE, GSDevice::CONST_ZERO,
-					GSDevice::OP_ADD, GSDevice::SRC1_ALPHA, GSDevice::INV_SRC1_ALPHA, false, 0);
+				// We're reading the rt anyways, use it for AFAIL
+				// This ensures we don't attempt to use fbfetch + blend, which breaks Intel GPUs on Metal
+				// Setting afail to RGB_ONLY without enabling color1 will enable this mode in the shader, so nothing more to do here.
 			}
 			else
 			{
-				if (m_conf.blend_multi_pass.enable)
+				m_conf.ps.no_color1 = false;
+				if (!m_conf.blend.enable)
 				{
-					m_conf.blend_multi_pass.blend.src_factor_alpha = GSDevice::SRC1_ALPHA;
-					m_conf.blend_multi_pass.blend.dst_factor_alpha = GSDevice::INV_SRC1_ALPHA;
+					m_conf.blend = GSHWDrawConfig::BlendState(true, GSDevice::CONST_ONE, GSDevice::CONST_ZERO,
+						GSDevice::OP_ADD, GSDevice::SRC1_ALPHA, GSDevice::INV_SRC1_ALPHA, false, 0);
 				}
 				else
 				{
-					m_conf.blend.src_factor_alpha = GSDevice::SRC1_ALPHA;
-					m_conf.blend.dst_factor_alpha = GSDevice::INV_SRC1_ALPHA;
+					if (m_conf.blend_multi_pass.enable)
+					{
+						m_conf.blend_multi_pass.blend.src_factor_alpha = GSDevice::SRC1_ALPHA;
+						m_conf.blend_multi_pass.blend.dst_factor_alpha = GSDevice::INV_SRC1_ALPHA;
+					}
+					else
+					{
+						m_conf.blend.src_factor_alpha = GSDevice::SRC1_ALPHA;
+						m_conf.blend.dst_factor_alpha = GSDevice::INV_SRC1_ALPHA;
+					}
 				}
 			}
 

pcsx2/GS/Renderers/Metal/tfx.metal

@@ -97,7 +97,8 @@ constant bool SW_BLEND = (PS_BLEND_A != PS_BLEND_B) || PS_BLEND_D;
 constant bool SW_AD_TO_HW = (PS_BLEND_C == 1 && PS_A_MASKED);
 constant bool NEEDS_RT_FOR_BLEND = (((PS_BLEND_A != PS_BLEND_B) && (PS_BLEND_A == 1 || PS_BLEND_B == 1 || PS_BLEND_C == 1)) || PS_BLEND_D == 1 || SW_AD_TO_HW);
 constant bool NEEDS_RT_EARLY = PS_TEX_IS_FB || PS_DATE >= 5;
-constant bool NEEDS_RT = NEEDS_RT_EARLY || (!PS_PRIM_CHECKING_INIT && (PS_FBMASK || NEEDS_RT_FOR_BLEND));
+constant bool NEEDS_RT_FOR_AFAIL = PS_AFAIL == 3 && PS_NO_COLOR1;
+constant bool NEEDS_RT = NEEDS_RT_FOR_AFAIL || NEEDS_RT_EARLY || (!PS_PRIM_CHECKING_INIT && (PS_FBMASK || NEEDS_RT_FOR_BLEND));
 
 constant bool PS_COLOR0 = !PS_NO_COLOR;
 constant bool PS_COLOR1 = !PS_NO_COLOR1;
@@ -1202,8 +1203,12 @@ struct PSMain
 			alpha_blend.a = float(atst_pass);
 
 		if (PS_COLOR0)
+		{
 			out.c0.a = PS_RTA_CORRECTION ? C.a / 128.f : C.a / 255.f;
 			out.c0.rgb = PS_HDR ? float3(C.rgb / 65535.f) : C.rgb / 255.f;
+			if (PS_AFAIL == 3 && !PS_COLOR1 && !atst_pass) // Doing RGB_ONLY without COLOR1
+				out.c0.a = current_color.a;
+		}
 		if (PS_COLOR1)
 			out.c1 = alpha_blend;
 		if (PS_ZCLAMP)

pcsx2/ShaderCacheVersion.h

@@ -3,4 +3,4 @@
 
 /// Version number for GS and other shaders. Increment whenever any of the contents of the
 /// shaders change, to invalidate the cache.
-static constexpr u32 SHADER_CACHE_VERSION = 59;
+static constexpr u32 SHADER_CACHE_VERSION = 60;