GS/HW: Prefer blend second pass over barriers for AMD.

Allow to use blend second pass on VK/GL on AMD, barriers are slower so let's use blend second pass if possible.

pcsx2/GS/Renderers/Common/GSDevice.h

@@ -734,20 +734,21 @@ class GSDevice : public GSAlignedClass<32>
 	// clang-format off
 	struct FeatureSupport
 	{
-		bool broken_point_sampler : 1; ///< Issue with AMD cards, see tfx shader for details
-		bool vs_expand            : 1; ///< Supports expanding points/lines/sprites in the vertex shader
-		bool primitive_id         : 1; ///< Supports primitive ID for use with prim tracking destination alpha algorithm
-		bool texture_barrier      : 1; ///< Supports sampling rt and hopefully texture barrier
-		bool provoking_vertex_last: 1; ///< Supports using the last vertex in a primitive as the value for flat shading.
-		bool point_expand         : 1; ///< Supports point expansion in hardware.
-		bool line_expand          : 1; ///< Supports line expansion in hardware.
-		bool prefer_new_textures  : 1; ///< Allocate textures up to the pool size before reusing them, to avoid render pass restarts.
-		bool dxt_textures         : 1; ///< Supports DXTn texture compression, i.e. S3TC and BC1-3.
-		bool bptc_textures        : 1; ///< Supports BC6/7 texture compression.
-		bool framebuffer_fetch    : 1; ///< Can sample from the framebuffer without texture barriers.
-		bool stencil_buffer       : 1; ///< Supports stencil buffer, and can use for DATE.
-		bool cas_sharpening       : 1; ///< Supports sufficient functionality for contrast adaptive sharpening.
-		bool test_and_sample_depth: 1; ///< Supports concurrently binding the depth-stencil buffer for sampling and depth testing.
+		bool broken_point_sampler    : 1; ///< Issue with AMD cards, see tfx shader for details
+		bool vs_expand               : 1; ///< Supports expanding points/lines/sprites in the vertex shader
+		bool primitive_id            : 1; ///< Supports primitive ID for use with prim tracking destination alpha algorithm
+		bool texture_barrier         : 1; ///< Supports sampling rt and hopefully texture barrier
+		bool provoking_vertex_last   : 1; ///< Supports using the last vertex in a primitive as the value for flat shading.
+		bool point_expand            : 1; ///< Supports point expansion in hardware.
+		bool line_expand             : 1; ///< Supports line expansion in hardware.
+		bool prefer_new_textures     : 1; ///< Allocate textures up to the pool size before reusing them, to avoid render pass restarts.
+		bool dxt_textures            : 1; ///< Supports DXTn texture compression, i.e. S3TC and BC1-3.
+		bool bptc_textures           : 1; ///< Supports BC6/7 texture compression.
+		bool framebuffer_fetch       : 1; ///< Can sample from the framebuffer without texture barriers.
+		bool stencil_buffer          : 1; ///< Supports stencil buffer, and can use for DATE.
+		bool cas_sharpening          : 1; ///< Supports sufficient functionality for contrast adaptive sharpening.
+		bool test_and_sample_depth   : 1; ///< Supports concurrently binding the depth-stencil buffer for sampling and depth testing.
+		bool prefer_blend_second_pass: 1; ///< Barriers are slow on AMD, we can prefer blend 2 pass instead.
 		FeatureSupport()
 		{
 			memset(this, 0, sizeof(*this));

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -56,6 +56,7 @@ GSDevice11::GSDevice11()
 	m_features.stencil_buffer = true;
 	m_features.cas_sharpening = true;
 	m_features.test_and_sample_depth = false;
+	m_features.prefer_blend_second_pass = false;
 }
 
 GSDevice11::~GSDevice11() = default;

pcsx2/GS/Renderers/DX12/GSDevice12.cpp

@@ -1209,6 +1209,7 @@ bool GSDevice12::CheckFeatures()
 	m_features.stencil_buffer = true;
 	m_features.cas_sharpening = true;
 	m_features.test_and_sample_depth = false;
+	m_features.prefer_blend_second_pass = false;
 	m_features.vs_expand = !GSConfig.DisableVertexShaderExpand;
 
 	m_features.dxt_textures = SupportsTextureFormat(DXGI_FORMAT_BC1_UNORM) &&

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -4048,7 +4048,7 @@ void GSRendererHW::EmulateBlending(int rt_alpha_min, int rt_alpha_max, bool& DAT
 	// HW blend can be done in multiple passes when there's no overlap.
 	// Blend second pass is only useful when texture barriers aren't supported.
 	// Speed wise Texture barriers > blend second pass > texture copies.
-	const bool blend_second_pass_support = !features.texture_barrier && no_prim_overlap && is_basic_blend;
+	const bool blend_second_pass_support = (!features.texture_barrier || features.prefer_blend_second_pass) && no_prim_overlap && is_basic_blend;
 	const bool bmix1_second_pass = blend_second_pass_support && blend_mix1 && (alpha_c0_high_max_one || alpha_c2_high_one) && m_conf.ps.blend_d == 2;
 	// We don't want to enable blend mix if we are doing a second pass, it's useless.
 	blend_mix &= !bmix1_second_pass;
@@ -4066,6 +4066,23 @@ void GSRendererHW::EmulateBlending(int rt_alpha_min, int rt_alpha_max, bool& DAT
 	bool sw_blending = false;
 	if (features.texture_barrier)
 	{
+		// Prefer blend second pass over sw blend when primitives don't overlap on AMD.
+		// Barriers are slower on AMD so we can use blend second pass if possible.
+		const bool ad_second_pass_no_rta_scale = alpha_c1_high_no_rta_correct && (blend_flag & (BLEND_HW3 | BLEND_HW5 | BLEND_HW6 | BLEND_HW7 | BLEND_HW9));
+		const bool blend_hw4_second_pass = (alpha_c0_high_max_one || alpha_c2_high_one) && (blend_flag & BLEND_HW4);
+		const bool blend_hw8_second_pass = (blend_flag & BLEND_HW8);
+		bool prefer_blend_second_pass = features.prefer_blend_second_pass && blend_second_pass_support // Check if blend second pass is supported.
+										// Don't enable if we want to do sw blend on draw.
+										&& !prefer_sw_blend
+										// Limited to clamp 1, it will sw blend anyway on wrap.
+										&& COLCLAMP.CLAMP
+										// No dithering headaches, just sw blend it.
+										&& m_conf.ps.dither == 0
+										// PABE will require sw blend.
+										&& !(m_draw_env->PABE.PABE && GetAlphaMinMax().min < 128)
+										// Type of blends/formulas.
+										&& (bmix1_second_pass || ad_second_pass_no_rta_scale || blend_hw4_second_pass || blend_hw8_second_pass);
+
 		const bool blend_requires_barrier = (blend_flag & BLEND_A_MAX) // Impossible blending
 			// Sw blend, either full barrier or one barrier with no overlap.
 			|| prefer_sw_blend
@@ -4095,11 +4112,15 @@ void GSRendererHW::EmulateBlending(int rt_alpha_min, int rt_alpha_max, bool& DAT
 				sw_blending |= m_vt.m_primclass == GS_SPRITE_CLASS && m_drawlist.size() < 100;
 				[[fallthrough]];
 			case AccBlendLevel::Basic:
+				// Disable prefer_blend_second_pass if higher blending requires sw blend.
+				prefer_blend_second_pass &= sw_blending;
 				// Prefer sw blend if possible.
 				color_dest_blend &= !prefer_sw_blend;
 				accumulation_blend &= !prefer_sw_blend;
 				// Enable sw blending for barriers.
 				sw_blending |= blend_requires_barrier;
+				// Disable any sw blend if we prefer blend second pass.
+				sw_blending &= !prefer_blend_second_pass;
 				// Enable sw blending for free blending.
 				sw_blending |= free_blend;
 				// Do not run BLEND MIX if sw blending is already present, it's less accurate.

pcsx2/GS/Renderers/Metal/GSDeviceMTL.mm

@@ -900,6 +900,7 @@ static void OnMainThread(Fn&& fn)
 	m_features.stencil_buffer = true;
 	m_features.cas_sharpening = true;
 	m_features.test_and_sample_depth = true;
+	m_features.prefer_blend_second_pass = false;
 
 	// Init metal stuff
 	m_fn_constants = MRCTransfer([MTLFunctionConstantValues new]);

pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp

@@ -742,6 +742,7 @@ bool GSDeviceOGL::CheckFeatures(bool& buggy_pbo)
 	m_features.prefer_new_textures = false;
 	m_features.stencil_buffer = true;
 	m_features.test_and_sample_depth = m_features.texture_barrier;
+	m_features.prefer_blend_second_pass = vendor_id_amd;
 
 	// NVIDIA GPUs prior to Kepler appear to have broken vertex shader buffer loading.
 	// Use bindless textures (introduced in Kepler) to differentiate.

pcsx2/GS/Renderers/Vulkan/GSDeviceVK.cpp

@@ -2667,6 +2667,9 @@ bool GSDeviceVK::CheckFeatures()
 	// Buggy drivers with broken barriers probably have no chance using GENERAL layout for depth either...
 	m_features.test_and_sample_depth = m_features.texture_barrier;
 
+	// 2 pass is faster than barriers on amd.
+	m_features.prefer_blend_second_pass = isAMD;
+
 	// Use D32F depth instead of D32S8 when we have framebuffer fetch.
 	m_features.stencil_buffer &= !m_features.framebuffer_fetch;