GS/HW: Allow source is rt to use corrected alpha in most cases

bin/resources/shaders/dx11/convert.fx

@@ -139,15 +139,15 @@ PS_OUTPUT ps_rta_correction(PS_INPUT input)
 {
 	PS_OUTPUT output;
 	float4 value = sample_c(input.t);
-	output.c = float4(value.rgb, (value.a * 255.0f + 0.1f) / 127.5f);
+	output.c = float4(value.rgb, value.a / (128.5f / 255.0f));
 	return output;
 }
 
 PS_OUTPUT ps_rta_decorrection(PS_INPUT input)
 {
 	PS_OUTPUT output;
 	float4 value = sample_c(input.t);
-	output.c = float4(value.rgb, (value.a * 127.5f + 0.1f) / 255.0f);
+	output.c = float4(value.rgb, value.a * (128.5f / 255.0f));
 	return output;
 }
 

bin/resources/shaders/dx11/tfx.fx

@@ -51,6 +51,7 @@
 #define PS_URBAN_CHAOS_HLE 0
 #define PS_HDR 0
 #define PS_RTA_CORRECTION 0
+#define PS_RTA_SRC_CORRECTION 0
 #define PS_COLCLIP 0
 #define PS_BLEND_A 0
 #define PS_BLEND_B 0
@@ -330,7 +331,16 @@ uint4 sample_4_index(float4 uv, float uv_w)
 	c.w = sample_c(uv.zw, uv_w).a;
 
 	// Denormalize value
-	uint4 i = uint4(c * 255.5f);
+	uint4 i;
+
+	if (PS_RTA_SRC_CORRECTION)
+	{
+		i = uint4(c * 128.25f); // Denormalize value
+	}
+	else
+	{
+		i = uint4(c * 255.5f); // Denormalize value
+	}
 
 	if (PS_PAL_FMT == 1)
 	{
@@ -650,6 +660,9 @@ float4 sample_color(float2 st, float uv_w)
 		t = c[0];
 	}
 
+	if (PS_AEM_FMT == FMT_32 && PS_PAL_FMT == 0 && PS_RTA_SRC_CORRECTION)
+		t.a = t.a * (128.5f / 255.0f);
+
 	return trunc(t * 255.0f + 0.05f);
 }
 
@@ -850,7 +863,7 @@ void ps_blend(inout float4 Color, inout float4 As_rgba, float2 pos_xy)
 
 		float4 RT = SW_BLEND_NEEDS_RT ? RtTexture.Load(int3(pos_xy, 0)) : (float4)0.0f;
 
-		float Ad = PS_RTA_CORRECTION ? trunc(RT.a * 127.5f + 0.05f) / 128.0f : trunc(RT.a * 255.0f + 0.1f) / 128.0f;
+		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);
 		float3 Cs = Color.rgb;
 
@@ -969,7 +982,7 @@ PS_OUTPUT ps_main(PS_INPUT input)
 	float4 alpha_blend = (float4)0.0f;
 	if (SW_AD_TO_HW)
 	{
-		float4 RT = PS_RTA_CORRECTION ? trunc(RtTexture.Load(int3(input.p.xy, 0)) * 127.5f + 0.05f) : trunc(RtTexture.Load(int3(input.p.xy, 0)) * 255.0f + 0.1f);
+		float4 RT = PS_RTA_CORRECTION ? trunc(RtTexture.Load(int3(input.p.xy, 0)) * 128.0f + 0.1f) : trunc(RtTexture.Load(int3(input.p.xy, 0)) * 255.0f + 0.1f);
 		alpha_blend = (float4)(RT.a / 128.0f);
 	}
 	else

bin/resources/shaders/opengl/convert.glsl

@@ -337,15 +337,15 @@ void ps_datm0_rta_correction()
 void ps_rta_correction()
 {
 	vec4 value = sample_c();
-	SV_Target0 = vec4(value.rgb, (value.a * 255.0f + 0.1f) / 127.5f);
+	SV_Target0 = vec4(value.rgb, value.a / (128.5f / 255.0f));
 }
 #endif
 
 #ifdef ps_rta_decorrection
 void ps_rta_decorrection()
 {
 	vec4 value = sample_c();
-	SV_Target0 = vec4(value.rgb, (value.a * 127.5f + 0.1f) / 255.0f);
+	SV_Target0 = vec4(value.rgb, value.a * (128.5f / 255.0f));
 }
 #endif
 

bin/resources/shaders/opengl/tfx_fs.glsl

@@ -281,8 +281,12 @@ uvec4 sample_4_index(vec4 uv)
 	c.y = sample_c(uv.zy).a;
 	c.z = sample_c(uv.xw).a;
 	c.w = sample_c(uv.zw).a;
-
+
+#if PS_RTA_SRC_CORRECTION 
+	uvec4 i = uvec4(c * 128.25f); // Denormalize value
+#else
 	uvec4 i = uvec4(c * 255.5f); // Denormalize value
+#endif
 
 #if PS_PAL_FMT == 1
 	// 4HL
@@ -591,6 +595,10 @@ vec4 sample_color(vec2 st)
 	t = c[0];
 #endif
 
+#if PS_AEM_FMT == FMT_32 && PS_PAL_FMT == 0 && PS_RTA_SRC_CORRECTION
+	t.a = t.a * (128.5f / 255.0f);
+#endif
+
 	// The 0.05f helps to fix the overbloom of sotc
 	// I think the issue is related to the rounding of texture coodinate. The linear (from fixed unit)
 	// interpolation could be slightly below the correct one.
@@ -803,7 +811,7 @@ float As = As_rgba.a;
 #endif
 
 	#if PS_RTA_CORRECTION
-		float Ad = trunc(RT.a * 127.5f + 0.05f) / 128.0f;
+		float Ad = trunc(RT.a * 128.0f + 0.1f) / 128.0f;
 	#else
 		float Ad = trunc(RT.a * 255.0f + 0.1f) / 128.0f;
 	#endif
@@ -985,7 +993,7 @@ void ps_main()
 
 #if SW_AD_TO_HW
 	#if PS_RTA_CORRECTION
-		vec4 RT = trunc(fetch_rt() * 127.5f + 0.05f);
+		vec4 RT = trunc(fetch_rt() * 128.0f + 0.1f);
 	#else
 		vec4 RT = trunc(fetch_rt() * 255.0f + 0.1f);
 	#endif

bin/resources/shaders/vulkan/convert.glsl

@@ -114,15 +114,15 @@ void ps_datm0_rta_correction()
 void ps_rta_correction()
 {
 	vec4 value = sample_c(v_tex);
-	o_col0 = vec4(value.rgb, (value.a * 255.0f + 0.1f) / 127.5f);
+	o_col0 = vec4(value.rgb, value.a / (128.5f / 255.0f));
 }
 #endif
 
 #ifdef ps_rta_decorrection
 void ps_rta_decorrection()
 {
 	vec4 value = sample_c(v_tex);
-	o_col0 = vec4(value.rgb, (value.a * 127.5f + 0.1f) / 255.0f);
+	o_col0 = vec4(value.rgb, value.a * (128.5f / 255.0f));
 }
 #endif
 

bin/resources/shaders/vulkan/tfx.glsl

@@ -523,7 +523,12 @@ uvec4 sample_4_index(vec4 uv)
 	c.w = sample_c(uv.zw).a;
 
 	// Denormalize value
+
+#if PS_RTA_SRC_CORRECTION
+	uvec4 i = uvec4(c * 128.25f);
+#else
 	uvec4 i = uvec4(c * 255.5f);
+#endif
 
 	#if PS_PAL_FMT == 1
 		// 4HL
@@ -835,7 +840,9 @@ vec4 sample_color(vec2 st)
 		t = c[0];
 	}
 	#endif
-
+#if PS_AEM_FMT == FMT_32 && PS_PAL_FMT == 0 && PS_RTA_SRC_CORRECTION
+	t.a = t.a * (128.5f / 255.0f);
+#endif
 	return trunc(t * 255.0f + 0.05f);
 }
 
@@ -1056,7 +1063,7 @@ void ps_blend(inout vec4 Color, inout vec4 As_rgba)
 		#endif
 
 		#if PS_RTA_CORRECTION
-			float Ad = trunc(RT.a * 127.5f + 0.05f) / 128.0f;
+			float Ad = trunc(RT.a * 128.0f + 0.1f) / 128.0f;
 		#else
 			float Ad = trunc(RT.a * 255.0f + 0.1f) / 128.0f;
 		#endif
@@ -1236,7 +1243,7 @@ void main()
 
 #if SW_AD_TO_HW
 	#if PS_RTA_CORRECTION
-		vec4 RT = trunc(sample_from_rt() * 127.5f + 0.05f);
+		vec4 RT = trunc(sample_from_rt() * 128.0f + 0.1f);
 	#else
 		vec4 RT = trunc(sample_from_rt() * 255.0f + 0.1f);
 	#endif

pcsx2/GS/Renderers/Common/GSDevice.h

@@ -330,6 +330,7 @@ struct alignas(16) GSHWDrawConfig
 				u32 a_masked       : 1;
 				u32 hdr            : 1;
 				u32 rta_correction : 1;
+				u32 rta_source_correction : 1;
 				u32 colclip        : 1;
 				u32 blend_mix      : 2;
 				u32 round_inv      : 1; // Blending will invert the value, so rounding needs to go the other way

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -1689,6 +1689,7 @@ void GSDevice11::SetupPS(const PSSelector& sel, const GSHWDrawConfig::PSConstant
 		sm.AddMacro("PS_PAL_FMT", sel.pal_fmt);
 		sm.AddMacro("PS_HDR", sel.hdr);
 		sm.AddMacro("PS_RTA_CORRECTION", sel.rta_correction);
+		sm.AddMacro("PS_RTA_SRC_CORRECTION", sel.rta_source_correction);
 		sm.AddMacro("PS_COLCLIP", sel.colclip);
 		sm.AddMacro("PS_BLEND_A", sel.blend_a);
 		sm.AddMacro("PS_BLEND_B", sel.blend_b);

pcsx2/GS/Renderers/DX12/GSDevice12.cpp

@@ -2817,6 +2817,7 @@ const ID3DBlob* GSDevice12::GetTFXPixelShader(const GSHWDrawConfig::PSSelector&
 	sm.AddMacro("PS_PAL_FMT", sel.pal_fmt);
 	sm.AddMacro("PS_HDR", sel.hdr);
 	sm.AddMacro("PS_RTA_CORRECTION", sel.rta_correction);
+	sm.AddMacro("PS_RTA_SRC_CORRECTION", sel.rta_source_correction);
 	sm.AddMacro("PS_COLCLIP", sel.colclip);
 	sm.AddMacro("PS_BLEND_A", sel.blend_a);
 	sm.AddMacro("PS_BLEND_B", sel.blend_b);

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -4358,6 +4358,9 @@ __ri void GSRendererHW::EmulateTextureSampler(const GSTextureCache::Target* rt,
 	float scale = tex->GetScale();
 	HandleTextureHazards(rt, ds, tex, tmm, source_region, target_region, unscaled_size, scale, src_copy);
 
+	if (tex->m_target && tex->m_from_target && tex->m_target_direct && tex->m_from_target->m_rt_alpha_scale)
+		m_conf.ps.rta_source_correction = 1;
+
 	// Warning fetch the texture PSM format rather than the context format. The latter could have been corrected in the texture cache for depth.
 	//const GSLocalMemory::psm_t &psm = GSLocalMemory::m_psm[m_cached_ctx.TEX0.PSM];
 	const GSLocalMemory::psm_t& psm = GSLocalMemory::m_psm[tex->m_TEX0.PSM];
@@ -5265,7 +5268,10 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 			ds->m_alpha_min &= 128;
 		}
 	}
-
+
+	// If we Correct/Decorrect and tex is rt, we will need to update the texture reference
+	const bool req_src_update = tex && rt && tex->m_target && tex->m_target_direct && tex->m_texture == rt->m_texture;
+
 	if (rt)
 	{
 		const bool rta_decorrection = m_channel_shuffle || m_texture_shuffle || std::max(blend_alpha_max, rt->m_alpha_max) > 128 || m_conf.ps.fbmask || m_conf.ps.tex_is_fb;
@@ -5278,6 +5284,9 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 				{
 					rt->RTADecorrect(rt);
 					m_conf.rt = rt->m_texture;
+
+					if (req_src_update)
+						tex->m_texture = rt->m_texture;
 				}
 				else if (m_conf.colormask.wa)
 				{
@@ -5289,6 +5298,9 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 					{
 						rt->RTADecorrect(rt);
 						m_conf.rt = rt->m_texture;
+
+						if (req_src_update)
+							tex->m_texture = rt->m_texture;
 					}
 				}
 			}
@@ -5298,6 +5310,9 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 				{
 					rt->RTADecorrect(rt);
 					m_conf.rt = rt->m_texture;
+
+					if (req_src_update)
+						tex->m_texture = rt->m_texture;
 				}
 			}
 			else if (rt->m_last_draw == s_n)
@@ -5308,6 +5323,9 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 			{
 				rt->RTADecorrect(rt);
 				m_conf.rt = rt->m_texture;
+
+				if (req_src_update)
+					tex->m_texture = rt->m_texture;
 			}
 		}
 
@@ -5318,6 +5336,9 @@ __ri void GSRendererHW::DrawPrims(GSTextureCache::Target* rt, GSTextureCache::Ta
 	if ((!IsOpaque() || m_context->ALPHA.IsBlack()) && rt && ((m_conf.colormask.wrgba & 0x7) || (m_texture_shuffle && !m_copy_16bit_to_target_shuffle && !m_same_group_texture_shuffle)))
 	{
 		EmulateBlending(blend_alpha_min, blend_alpha_max, DATE_PRIMID, DATE_BARRIER, blending_alpha_pass, rt);
+
+		if (req_src_update && tex->m_texture != rt->m_texture)
+			tex->m_texture = rt->m_texture;
 	}
 	else
 	{

pcsx2/GS/Renderers/HW/GSTextureCache.cpp

@@ -1690,9 +1690,6 @@ GSTextureCache::Source* GSTextureCache::LookupSource(const bool is_color, const
 		}
 #endif
 
-		 if (dst && (GSUtil::GetChannelMask(TEX0.PSM) & 0x8))
-			dst->RTADecorrect(dst);
-
 		src = CreateSource(TEX0, TEXA, dst, half_right, x_offset, y_offset, lod, &r, gpu_clut, region);
 		if (!src) [[unlikely]]
 			return nullptr;
@@ -4299,7 +4296,16 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 			// copy the rt in
 			const GSVector4i area(GSVector4i(x, y, x + w, y + h).rintersect(GSVector4i(sTex->GetSize()).zwxy()));
 			if (!area.rempty())
-				g_gs_device->CopyRect(sTex, dTex, area, 0, 0);
+			{
+				if (dst->m_rt_alpha_scale)
+				{
+					const GSVector4 sRectF = GSVector4(area) / GSVector4(1, 1, sTex->GetWidth(), sTex->GetHeight());
+					g_gs_device->StretchRect(
+						sTex, sRectF, dTex, GSVector4(area), ShaderConvert::RTA_DECORRECTION, false);
+				}
+				else
+					g_gs_device->CopyRect(sTex, dTex, area, 0, 0);
+			}
 
 			src->m_texture = dTex;
 			src->m_unscaled_size = GSVector2i(tw, th);
@@ -4317,6 +4323,8 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 				src->m_region.SetY(y_offset, region.GetMaxY() + y_offset);
 			else
 				src->m_region.SetY(y_offset, y_offset + th);
+
+			src->m_target_direct = true;
 			src->m_texture = dst->m_texture;
 			src->m_unscaled_size = dst->m_unscaled_size;
 			src->m_shared_texture = true;
@@ -4554,6 +4562,7 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 		{
 			// sample the target directly
 			src->m_texture = dst->m_texture;
+			src->m_target_direct = true;
 			src->m_scale = dst->m_scale;
 			src->m_unscaled_size = dst->m_unscaled_size;
 			src->m_shared_texture = true;
@@ -4600,7 +4609,14 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 
 			if (use_texture)
 			{
-				g_gs_device->CopyRect(sTex, dTex, sRect, destX, destY);
+				if (dst->m_rt_alpha_scale)
+				{
+					const GSVector4 sRectF = GSVector4(sRect) / GSVector4(1, 1, sTex->GetWidth(), sTex->GetHeight());
+					g_gs_device->StretchRect(
+						sTex, sRectF, dTex, GSVector4(destX, destY, sRect.width(), sRect.height()), ShaderConvert::RTA_DECORRECTION, false);
+				}
+				else
+					g_gs_device->CopyRect(sTex, dTex, sRect, destX, destY);
 				g_perfmon.Put(GSPerfMon::TextureCopies, 1);
 
 #ifdef PCSX2_DEVBUILD
@@ -4615,12 +4631,21 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 			{
 				if (is_8bits)
 				{
+					if (dst->m_rt_alpha_scale)
+					{
+						dst->RTADecorrect(dst);
+						sTex = dst->m_texture;
+					}
+
 					g_gs_device->ConvertToIndexedTexture(sTex, dst->m_scale, x_offset, y_offset,
 						std::max<u32>(dst->m_TEX0.TBW, 1u) * 64, dst->m_TEX0.PSM, dTex,
 						std::max<u32>(TEX0.TBW, 1u) * 64, TEX0.PSM);
 				}
 				else
 				{
+					if (dst->m_rt_alpha_scale && shader == ShaderConvert::COPY)
+						shader = ShaderConvert::RTA_DECORRECTION;
+
 					const GSVector4 sRectF = GSVector4(sRect) / GSVector4(1, 1, sTex->GetWidth(), sTex->GetHeight());
 					g_gs_device->StretchRect(
 						sTex, sRectF, dTex, GSVector4(destX, destY, new_size.x, new_size.y), shader, false);

pcsx2/GS/Renderers/HW/GSTextureCache.h

@@ -283,6 +283,7 @@ class GSTextureCache
 		u8 m_valid_hashes = 0;
 		u8 m_complete_layers = 0;
 		bool m_target = false;
+		bool m_target_direct = false;
 		bool m_repeating = false;
 		std::pair<u8, u8> m_alpha_minmax = {0u, 255u};
 		std::vector<GSVector2i>* m_p2t = nullptr;

pcsx2/GS/Renderers/Metal/GSDeviceMTL.mm

@@ -1821,6 +1821,7 @@ static GSMTLExpandType ConvertVSExpand(GSHWDrawConfig::VSExpand generic)
 		setFnConstantB(m_fn_constants, pssel.a_masked,           GSMTLConstantIndex_PS_A_MASKED);
 		setFnConstantB(m_fn_constants, pssel.hdr,                GSMTLConstantIndex_PS_HDR);
 		setFnConstantB(m_fn_constants, pssel.rta_correction,     GSMTLConstantIndex_PS_RTA_CORRECTION);
+		setFnConstantB(m_fn_constants, pssel.rta_source_correction, GSMTLConstantIndex_PS_RTA_SRC_CORRECTION);
 		setFnConstantB(m_fn_constants, pssel.colclip,            GSMTLConstantIndex_PS_COLCLIP);
 		setFnConstantI(m_fn_constants, pssel.blend_mix,          GSMTLConstantIndex_PS_BLEND_MIX);
 		setFnConstantB(m_fn_constants, pssel.round_inv,          GSMTLConstantIndex_PS_ROUND_INV);

pcsx2/GS/Renderers/Metal/GSMTLSharedHeader.h

@@ -183,6 +183,7 @@ enum GSMTLFnConstants
 	GSMTLConstantIndex_PS_A_MASKED,
 	GSMTLConstantIndex_PS_HDR,
 	GSMTLConstantIndex_PS_RTA_CORRECTION,
+	GSMTLConstantIndex_PS_RTA_SRC_CORRECTION,
 	GSMTLConstantIndex_PS_COLCLIP,
 	GSMTLConstantIndex_PS_BLEND_MIX,
 	GSMTLConstantIndex_PS_ROUND_INV,