GS/HW: Change config settings for depth feedback.

Author: TJnotJT

bin/resources/shaders/dx11/tfx.fx

@@ -1268,7 +1268,7 @@ PS_OUTPUT ps_main(PS_INPUT input)
 #endif // PS_DATE != 1/2
 
 #if PS_ZCLAMP
-	output.depth = min(input.p.z, MaxDepthPS);	
+	output.depth = min(input.p.z, MaxDepthPS);
 #endif
 
 	return output;

pcsx2/Config.h

@@ -732,6 +732,7 @@ struct Pcsx2Config
 					DisableShaderCache : 1,
 					DisableFramebufferFetch : 1,
 					DisableVertexShaderExpand : 1,
+					DisableDepthFeedback : 1,
 					SkipDuplicateFrames : 1,
 					OsdShowSpeed : 1,
 					OsdShowFPS : 1,
@@ -757,7 +758,6 @@ struct Pcsx2Config
 					PreloadFrameWithGSData : 1,
 					Mipmap : 1,
 					HWMipmap : 1,
-					HWAFAILFeedback : 1,
 					ManualUserHacks : 1,
 					UserHacks_AlignSpriteX : 1,
 					UserHacks_CPUFBConversion : 1,

pcsx2/GS/Renderers/Common/GSDevice.h

@@ -900,7 +900,8 @@ class GSDevice : public GSAlignedClass<32>
 		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 depth_as_rt_feedback : 1; ///< Depth feedback loops/barriers by converting depth to a temporary color target.
+		bool depth_feedback       : 1; ///< Depth feedback loops by directly binding DepthStencil target for read/write.
+		bool depth_as_rt_feedback : 1; ///< Depth feedback loops by converting depth to a temporary color target.
 		FeatureSupport()
 		{
 			memset(this, 0, sizeof(*this));

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -59,6 +59,8 @@ GSDevice11::GSDevice11()
 	m_features.stencil_buffer = true;
 	m_features.cas_sharpening = true;
 	m_features.test_and_sample_depth = true;
+	m_features.depth_feedback = m_features.multidraw_fb_copy && !GSConfig.DisableDepthFeedback;
+	m_features.depth_as_rt_feedback = false;
 }
 
 GSDevice11::~GSDevice11() = default;

pcsx2/GS/Renderers/DX12/GSDevice12.cpp

@@ -16,7 +16,6 @@
 #include "common/BitUtils.h"
 #include "common/Error.h"
 #include "common/HostSys.h"
-#include "common/ScopedGuard.h"
 #include "common/SmallString.h"
 #include "common/StringUtil.h"
 
@@ -1240,7 +1239,8 @@ bool GSDevice12::CheckFeatures(const u32& vendor_id)
 	m_features.cas_sharpening = true;
 	m_features.test_and_sample_depth = true;
 	m_features.vs_expand = !GSConfig.DisableVertexShaderExpand;
-	m_features.depth_as_rt_feedback = m_features.texture_barrier;
+	m_features.depth_feedback = false;
+	m_features.depth_as_rt_feedback = m_features.test_and_sample_depth && !GSConfig.DisableDepthFeedback;
 
 	m_features.dxt_textures = SupportsTextureFormat(DXGI_FORMAT_BC1_UNORM) &&
 	                          SupportsTextureFormat(DXGI_FORMAT_BC2_UNORM) &&
@@ -4225,7 +4225,7 @@ void GSDevice12::SendHWDraw(const PipelineSelector& pipe, const GSHWDrawConfig&
 			PSSetShaderResource(0, draw_rt, false, true);
 		if ((one_barrier || full_barrier) && feedback_depth)
 			PSSetShaderResource(4, draw_ds, false, true);
-
+		
 		std::array<D3D12_RESOURCE_BARRIER, 2> barriers;
 		u32 n_barriers = 0;
 		// Specify null for the after resource as both resources are used after the barrier.
@@ -4257,7 +4257,7 @@ void GSDevice12::SendHWDraw(const PipelineSelector& pipe, const GSHWDrawConfig&
 				const u32 count = (*config.drawlist)[n] * indices_per_prim;
 
 				EndRenderPass();
-				
+
 				GetCommandList()->ResourceBarrier(n_barriers, barriers.data());
 
 				if (BindDrawPipeline(pipe))

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -7346,22 +7346,43 @@ void GSRendererHW::EmulateAlphaTest(const bool& DATE, bool& DATE_BARRIER, bool&
 	// Determine whether the feedback methods require a single pass.
 	const bool feedback_one_pass = simple_fb_only || simple_rgb_only || simple_zb_only;
 
+	// Determine if the method for doing depth feedback uses multiple render targets.
+	// This should not be used in conjunction with dual source blend.
+	// This also requires 2 full screen copies of depth <-> color so avoid if possible.
+	const bool depth_as_rt_feedback = afail_needs_depth && features.depth_as_rt_feedback;
+
 	// If we are already have the required barriers for the accurate feedback path.
-	const bool already_have_barriers =
+	const bool free_to_use_barriers =
 		((m_conf.require_one_barrier && feedback_one_pass) || m_conf.require_full_barrier) &&
-		(features.texture_barrier || features.multidraw_fb_copy);
+		(features.texture_barrier || features.multidraw_fb_copy) &&
+		!depth_as_rt_feedback;
 
 	// Dual source blend comes with some restrictions, so detect and avoid them here.
 	const bool using_dual_source_blend = !m_conf.ps.no_color1;
 
-	// Determines if the method for doing depth feedback uses multiple render targets.
-	// This should not be used in conjunction with dual source blend.
-	const bool depth_as_rt_feedback = afail_needs_depth && features.depth_as_rt_feedback;
+	// Determine if we have the correct features for depth feedback.
+	const bool depth_feedback_supported = features.depth_feedback || features.depth_as_rt_feedback;
+
+	const bool avoid_feedback =
+		// Do not use dual source blend with FB-fetch, which breaks Intel GPUs on Metal.
+		// Also do not use dual source blend with multiple render targets, which is against spec.
+		((features.framebuffer_fetch || depth_as_rt_feedback) && using_dual_source_blend) ||
+		// We need depth feedback but do not have the correct features.
+		(afail_needs_depth && !depth_feedback_supported);
+	
+	// Determine if we can use FB-fetch for color only feedback.
+	const bool free_fbfetch_feedback = features.framebuffer_fetch && !depth_as_rt_feedback;
 
-	// Also do not use dual source blend with FB-fetch, which breaks Intel GPUs on Metal.
-	const bool avoid_feedback = (features.framebuffer_fetch || depth_as_rt_feedback) && using_dual_source_blend;
+	// Prefer feedback method only if blend level requests it or it's free.
+	const bool prefer_feedback = (GSConfig.AccurateBlendingUnit >= AccBlendLevel::Maximum) ||
+	                             free_to_use_barriers || free_fbfetch_feedback;
 
-	if ((GSConfig.HWAFAILFeedback || already_have_barriers || features.framebuffer_fetch) && !avoid_feedback)
+	// The two simple cases can be handle accurately in two passes so no point
+	// in requiring barriers if they are not already being used.
+	const bool prefer_two_pass = !(m_conf.require_full_barrier || m_conf.require_full_barrier) &&
+	                             (simple_fb_only || simple_rgb_only);
+	
+	if (prefer_feedback && !prefer_two_pass && !avoid_feedback)
 	{
 		// Use RT and/or depth sampling for accurate AFAIL in the shader.
 		GL_INS("Alpha test with RT/depth feedback (accurate)");
@@ -7377,7 +7398,7 @@ void GSRendererHW::EmulateAlphaTest(const bool& DATE, bool& DATE_BARRIER, bool&
 		// In that case, we don't need barriers.
 		const bool use_full_fbfetch = features.framebuffer_fetch &&
 		                              (!afail_needs_depth || features.depth_as_rt_feedback);
-		if (!use_full_fbfetch)
+		if (!use_full_fbfetch && (features.texture_barrier || features.multidraw_fb_copy))
 		{
 			m_conf.require_one_barrier |= feedback_one_pass;
 			m_conf.require_full_barrier |= !feedback_one_pass;
@@ -8320,14 +8341,15 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 		if (m_conf.alpha_second_pass.enable)
 		{
 			m_conf.alpha_second_pass.depth.zwe = 0;
-			m_conf.alpha_second_pass.depth.ztst = 0;
+			m_conf.alpha_second_pass.depth.ztst = ZTST_ALWAYS;
 		}
 
 		// Create the temporary depth copy
 		m_conf.ds_as_rt = g_gs_device->CreateRenderTarget(m_conf.ds->GetWidth(), m_conf.ds->GetHeight(),
 			GSTexture::Format::Float32, false, true);
 		const GSVector4 dRect(0.0f, 0.0f, static_cast<float>(m_conf.ds->GetWidth()), static_cast<float>(m_conf.ds->GetHeight()));
 		g_gs_device->StretchRect(m_conf.ds, m_conf.ds_as_rt, dRect, ShaderConvert::FLOAT32_DEPTH_TO_COLOR, false);
+		g_perfmon.Put(GSPerfMon::TextureCopies, 1.0);
 	}
 
 	// rs
@@ -8362,6 +8384,7 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 		g_gs_device->StretchRect(m_conf.ds_as_rt, m_conf.ds, dRect, ShaderConvert::FLOAT32_COLOR_TO_DEPTH, false);
 		g_gs_device->Recycle(m_conf.ds_as_rt);
 		m_conf.ds_as_rt = nullptr;
+		g_perfmon.Put(GSPerfMon::TextureCopies, 1.0);
 	}
 }
 

pcsx2/GS/Renderers/Metal/GSDeviceMTL.mm

@@ -933,6 +933,8 @@ static void OnMainThread(Fn&& fn)
 	m_features.stencil_buffer = true;
 	m_features.cas_sharpening = true;
 	m_features.test_and_sample_depth = true;
+	m_features.depth_feedback = false;
+	m_features.depth_as_rt_feedback = false;
 	m_max_texture_size = m_dev.features.max_texsize;
 
 	// Init metal stuff

pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp

@@ -741,7 +741,8 @@ bool GSDeviceOGL::CheckFeatures()
 	m_features.prefer_new_textures = false;
 	m_features.stencil_buffer = true;
 	m_features.test_and_sample_depth = m_features.texture_barrier;
-	m_features.depth_as_rt_feedback = m_features.texture_barrier;
+	m_features.depth_feedback = false;
+	m_features.depth_as_rt_feedback = m_features.test_and_sample_depth && !GSConfig.DisableDepthFeedback;
 
 	if (GLAD_GL_ARB_shader_storage_buffer_object)
 	{

pcsx2/GS/Renderers/Vulkan/GSDeviceVK.cpp

@@ -2705,6 +2705,9 @@ bool GSDeviceVK::CheckFeatures()
 	m_features.line_expand =
 		(m_device_features.wideLines && limits.lineWidthRange[0] <= f_upscale && limits.lineWidthRange[1] >= f_upscale);
 
+	m_features.depth_feedback = m_features.test_and_sample_depth && !GSConfig.DisableDepthFeedback;
+	m_features.depth_as_rt_feedback = false;
+
 	DevCon.WriteLn("Optional features:%s%s%s%s%s", m_features.primitive_id ? " primitive_id" : "",
 		m_features.texture_barrier ? " texture_barrier" : "", m_features.framebuffer_fetch ? " framebuffer_fetch" : "",
 		m_features.provoking_vertex_last ? " provoking_vertex_last" : "", m_features.vs_expand ? " vs_expand" : "");

pcsx2/Pcsx2Config.cpp

@@ -722,6 +722,7 @@ Pcsx2Config::GSOptions::GSOptions()
 	DisableShaderCache = false;
 	DisableFramebufferFetch = false;
 	DisableVertexShaderExpand = false;
+	DisableDepthFeedback = false;
 	SkipDuplicateFrames = false;
 	OsdMessagesPos = OsdOverlayPos::TopLeft;
 	OsdPerformancePos = OsdOverlayPos::TopRight;
@@ -751,7 +752,6 @@ Pcsx2Config::GSOptions::GSOptions()
 	PreloadFrameWithGSData = false;
 	Mipmap = true;
 	HWMipmap = true;
-	HWAFAILFeedback = false;
 
 	ManualUserHacks = false;
 	UserHacks_AlignSpriteX = false;
@@ -900,6 +900,7 @@ bool Pcsx2Config::GSOptions::RestartOptionsAreEqual(const GSOptions& right) cons
 		   OpEqu(DisableShaderCache) &&
 		   OpEqu(DisableFramebufferFetch) &&
 		   OpEqu(DisableVertexShaderExpand) &&
+		   OpEqu(DisableDepthFeedback) &&
 		   OpEqu(OverrideTextureBarriers) &&
 		   OpEqu(ExclusiveFullscreenControl);
 }
@@ -945,6 +946,7 @@ void Pcsx2Config::GSOptions::LoadSave(SettingsWrapper& wrap)
 	SettingsWrapBitBool(DisableShaderCache);
 	SettingsWrapBitBool(DisableFramebufferFetch);
 	SettingsWrapBitBool(DisableVertexShaderExpand);
+	SettingsWrapBitBool(DisableDepthFeedback);
 	SettingsWrapBitBool(SkipDuplicateFrames);
 	SettingsWrapBitBool(OsdShowSpeed);
 	SettingsWrapBitBool(OsdShowFPS);
@@ -1022,7 +1024,6 @@ void Pcsx2Config::GSOptions::LoadSave(SettingsWrapper& wrap)
 	SettingsWrapEntryEx(UpscaleMultiplier, "upscale_multiplier");
 
 	SettingsWrapBitBoolEx(HWMipmap, "hw_mipmap");
-	SettingsWrapBitBoolEx(HWAFAILFeedback, "HWAFAILFeedback");
 	SettingsWrapIntEnumEx(AccurateBlendingUnit, "accurate_blending_unit");
 	SettingsWrapIntEnumEx(TextureFiltering, "filter");
 	SettingsWrapIntEnumEx(TexturePreloading, "texture_preloading");