GS/HW: Reduce number of copies for HDR

bin/resources/shaders/dx11/tfx.fx

@@ -799,7 +799,8 @@ void ps_fbmask(inout float4 C, float2 pos_xy)
 {
 	if (PS_FBMASK)
 	{
-		float4 RT = trunc(RtTexture.Load(int3(pos_xy, 0)) * 255.0f + 0.1f);
+		float multi = PS_HDR ? 65535.0f : 255.0f;
+		float4 RT = trunc(RtTexture.Load(int3(pos_xy, 0)) * multi + 0.1f);
 		C = (float4)(((uint4)C & ~FbMask) | ((uint4)RT & FbMask));
 	}
 }
@@ -895,7 +896,8 @@ void ps_blend(inout float4 Color, inout float4 As_rgba, float2 pos_xy)
 		}
 
 		float Ad = PS_RTA_CORRECTION ? trunc(RT.a * 128.0f + 0.1f) / 128.0f : trunc(RT.a * 255.0f + 0.1f) / 128.0f;
-		float3 Cd = trunc(RT.rgb * 255.0f + 0.1f);
+		float color_multi = PS_HDR ? 65535.0f : 255.0f;
+		float3 Cd = trunc(RT.rgb * color_multi + 0.1f);
 		float3 Cs = Color.rgb;
 
 		float3 A = (PS_BLEND_A == 0) ? Cs : ((PS_BLEND_A == 1) ? Cd : (float3)0.0f);

bin/resources/shaders/opengl/tfx_fs.glsl

@@ -708,7 +708,11 @@ void ps_fbmask(inout vec4 C)
 {
 	// FIXME do I need special case for 16 bits
 #if PS_FBMASK
-	vec4 RT = trunc(fetch_rt() * 255.0f + 0.1f);
+	#if PS_HDR == 1
+		vec4 RT = trunc(fetch_rt() * 65535.0f);
+	#else
+		vec4 RT = trunc(fetch_rt() * 255.0f + 0.1f);
+	#endif
 	C = vec4((uvec4(C) & ~FbMask) | (uvec4(RT) & FbMask));
 #endif
 }
@@ -823,7 +827,11 @@ float As = As_rgba.a;
 	#endif
 
 	// Let the compiler do its jobs !
-	vec3 Cd = trunc(RT.rgb * 255.0f + 0.1f);
+	#if PS_HDR == 1
+		vec3 Cd = trunc(RT.rgb * 65535.0f);
+	#else
+		vec3 Cd = trunc(RT.rgb * 255.0f + 0.1f);
+	#endif
 	vec3 Cs = Color.rgb;
 
 #if PS_BLEND_A == 0

bin/resources/shaders/vulkan/tfx.glsl

@@ -972,7 +972,12 @@ vec4 ps_color()
 void ps_fbmask(inout vec4 C)
 {
 	#if PS_FBMASK
-		vec4 RT = trunc(sample_from_rt() * 255.0f + 0.1f);
+
+		#if PS_HDR == 1
+			vec4 RT = trunc(sample_from_rt() * 65535.0f);
+		#else
+			vec4 RT = trunc(sample_from_rt() * 255.0f + 0.1f);
+		#endif
 		C = vec4((uvec4(C) & ~FbMask) | (uvec4(RT) & FbMask));
 	#endif
 }
@@ -1090,7 +1095,11 @@ void ps_blend(inout vec4 Color, inout vec4 As_rgba)
 		#endif
 
 			// Let the compiler do its jobs !
+			#if PS_HDR == 1
+			vec3 Cd = trunc(RT.rgb * 65535.0f);
+			#else
 			vec3 Cd = trunc(RT.rgb * 255.0f + 0.1f);
+			#endif
 			vec3 Cs = Color.rgb;
 
 		#if PS_BLEND_A == 0

pcsx2/GS/Renderers/Common/GSDevice.h

@@ -8,6 +8,7 @@
 #include "GS/GS.h"
 #include "GS/Renderers/Common/GSFastList.h"
 #include "GS/Renderers/Common/GSTexture.h"
+#include "GS/Renderers/Vulkan/GSTextureVK.h"
 #include "GS/Renderers/Common/GSVertex.h"
 #include "GS/GSAlignedClass.h"
 #include "GS/GSExtra.h"
@@ -674,6 +675,15 @@ struct alignas(16) GSHWDrawConfig
 		Full,           ///< Full emulation (using barriers / ROV)
 	};
 
+	enum class HDRMode : u8
+	{
+		NoModify = 0,
+		ConvertOnly = 1,
+		ResolveOnly = 2,
+		ConvertAndResolve = 3,
+		EarlyResolve = 4
+	};
+
 	GSTexture* rt;        ///< Render target
 	GSTexture* ds;        ///< Depth stencil
 	GSTexture* tex;       ///< Source texture
@@ -730,6 +740,11 @@ struct alignas(16) GSHWDrawConfig
 
 	VSConstantBuffer cb_vs;
 	PSConstantBuffer cb_ps;
+
+	// These are here as they need to be preserved between draws, and the state clear only does up to the constant buffers.
+	HDRMode hdr_mode;
+	GIFRegFRAME hdr_frame;
+	GSVector4i hdr_update_area; ///< Area in the framebuffer which HDR will modify;
 };
 
 class GSDevice : public GSAlignedClass<32>
@@ -850,6 +865,7 @@ class GSDevice : public GSAlignedClass<32>
 	GSTexture* m_target_tmp = nullptr;
 	GSTexture* m_current = nullptr;
 	GSTexture* m_cas = nullptr;
+	GSTexture* m_hdr_rt = nullptr; ///< Temp HDR texture
 
 	bool AcquireWindow(bool recreate_window);
 
@@ -874,6 +890,10 @@ class GSDevice : public GSAlignedClass<32>
 	/// Returns a string containing current adapter in use.
 	const std::string& GetName() const { return m_name; }
 
+	GSTexture* GetHDRTexture() const { return m_hdr_rt; }
+
+	void SetHDRTexture(GSTexture* tex) { m_hdr_rt = tex; }
+
 	/// Returns a string representing the specified API.
 	static const char* RenderAPIToString(RenderAPI api);
 

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -2521,14 +2521,55 @@ void GSDevice11::RenderHW(GSHWDrawConfig& config)
 
 	GSVector2i rtsize = (config.rt ? config.rt : config.ds)->GetSize();
 
+	GSTexture* hdr_rt = g_gs_device->GetHDRTexture();
+
+	if (hdr_rt)
+	{
+		if (config.hdr_mode == GSHWDrawConfig::HDRMode::EarlyResolve)
+		{
+			const GSVector2i size = config.rt->GetSize();
+			const GSVector4 dRect(config.hdr_update_area);
+			const GSVector4 sRect = dRect / GSVector4(size.x, size.y).xyxy();
+			StretchRect(hdr_rt, sRect, config.rt, dRect, ShaderConvert::HDR_RESOLVE, false);
+			g_perfmon.Put(GSPerfMon::TextureCopies, 1);
+			Recycle(hdr_rt);
+
+			g_gs_device->SetHDRTexture(nullptr);
+
+			hdr_rt = nullptr;
+		}
+		else
+			config.ps.hdr = 1;
+	}
+
+	if (config.ps.hdr)
+	{
+		if (!hdr_rt)
+		{
+			config.hdr_update_area = config.drawarea;
+
+			const GSVector4 dRect = GSVector4((config.hdr_mode == GSHWDrawConfig::HDRMode::ConvertOnly) ? GSVector4i::loadh(rtsize) : config.drawarea);
+			const GSVector4 sRect = dRect / GSVector4(rtsize.x, rtsize.y).xyxy();
+			hdr_rt = CreateRenderTarget(rtsize.x, rtsize.y, GSTexture::Format::HDRColor);
+			if (!hdr_rt)
+				return;
+
+			g_gs_device->SetHDRTexture(hdr_rt);
+			// Warning: StretchRect must be called before BeginScene otherwise
+			// vertices will be overwritten. Trust me you don't want to do that.
+			StretchRect(config.rt, sRect, hdr_rt, dRect, ShaderConvert::HDR_INIT, false);
+			g_perfmon.Put(GSPerfMon::TextureCopies, 1);
+		}
+	}
+
 	GSTexture* primid_tex = nullptr;
 	if (config.destination_alpha == GSHWDrawConfig::DestinationAlphaMode::PrimIDTracking)
 	{
 		primid_tex = CreateRenderTarget(rtsize.x, rtsize.y, GSTexture::Format::PrimID, false);
 		if (!primid_tex)
 			return;
 
-		StretchRect(config.rt, GSVector4(config.drawarea) / GSVector4(rtsize).xyxy(),
+		StretchRect(hdr_rt ? hdr_rt : config.rt, GSVector4(config.drawarea) / GSVector4(rtsize).xyxy(),
 			primid_tex, GSVector4(config.drawarea), m_date.primid_init_ps[static_cast<u8>(config.datm)].get(), nullptr, false);
 	}
 	else if (config.destination_alpha != GSHWDrawConfig::DestinationAlphaMode::Off)
@@ -2544,22 +2585,7 @@ void GSDevice11::RenderHW(GSHWDrawConfig& config)
 			{GSVector4(dst.z, -dst.w, 0.5f, 1.0f), GSVector2(src.z, src.w)},
 		};
 
-		SetupDATE(config.rt, config.ds, vertices, config.datm);
-	}
-
-	GSTexture* hdr_rt = nullptr;
-	if (config.ps.hdr)
-	{
-		const GSVector4 dRect(config.drawarea);
-		const GSVector4 sRect = dRect / GSVector4(rtsize.x, rtsize.y).xyxy();
-		hdr_rt = CreateRenderTarget(rtsize.x, rtsize.y, GSTexture::Format::HDRColor);
-		if (!hdr_rt)
-			return;
-
-		// Warning: StretchRect must be called before BeginScene otherwise
-		// vertices will be overwritten. Trust me you don't want to do that.
-		StretchRect(config.rt, sRect, hdr_rt, dRect, ShaderConvert::HDR_INIT, false);
-		g_perfmon.Put(GSPerfMon::TextureCopies, 1);
+		SetupDATE(hdr_rt ? hdr_rt : config.rt, config.ds, vertices, config.datm);
 	}
 
 	if (config.vs.expand != GSHWDrawConfig::VSExpand::None)
@@ -2623,7 +2649,7 @@ void GSDevice11::RenderHW(GSHWDrawConfig& config)
 		// Do not always bind the rt when it's not needed,
 		// only bind it when effects use it such as fbmask emulation currently
 		// because we copy the frame buffer and it is quite slow.
-		CloneTexture(config.rt, &rt_copy, config.drawarea);
+		CloneTexture(hdr_rt ? hdr_rt : config.rt, &rt_copy, config.drawarea);
 		if (rt_copy)
 		{
 			if (config.require_one_barrier)
@@ -2690,11 +2716,18 @@ void GSDevice11::RenderHW(GSHWDrawConfig& config)
 
 	if (hdr_rt)
 	{
-		const GSVector2i size = config.rt->GetSize();
-		const GSVector4 dRect(config.drawarea);
-		const GSVector4 sRect = dRect / GSVector4(size.x, size.y).xyxy();
-		StretchRect(hdr_rt, sRect, config.rt, dRect, ShaderConvert::HDR_RESOLVE, false);
-		g_perfmon.Put(GSPerfMon::TextureCopies, 1);
-		Recycle(hdr_rt);
+		config.hdr_update_area = config.hdr_update_area.runion(config.drawarea);
+
+		if (config.hdr_mode == GSHWDrawConfig::HDRMode::ResolveOnly || config.hdr_mode == GSHWDrawConfig::HDRMode::ConvertAndResolve)
+		{
+			const GSVector2i size = config.rt->GetSize();
+			const GSVector4 dRect(config.hdr_update_area);
+			const GSVector4 sRect = dRect / GSVector4(size.x, size.y).xyxy();
+			StretchRect(hdr_rt, sRect, config.rt, dRect, ShaderConvert::HDR_RESOLVE, false);
+			g_perfmon.Put(GSPerfMon::TextureCopies, 1);
+			Recycle(hdr_rt);
+
+			g_gs_device->SetHDRTexture(nullptr);
+		}
 	}
 }