GS/DX11: Allow to pick whenever to update sr or ss, and some reordering.

No need to call sampler update when updating conflicting srvs.

pcsx2/GS/Renderers/DX11/GSDevice11.cpp

@@ -1105,14 +1105,14 @@ float GSDevice11::GetAndResetAccumulatedGPUTime()
 void GSDevice11::DrawPrimitive()
 {
 	g_perfmon.Put(GSPerfMon::DrawCalls, 1);
-	PSUpdateShaderState();
+	PSUpdateShaderState(true, true);
 	m_ctx->Draw(m_vertex.count, m_vertex.start);
 }
 
 void GSDevice11::DrawIndexedPrimitive()
 {
 	g_perfmon.Put(GSPerfMon::DrawCalls, 1);
-	PSUpdateShaderState();
+	PSUpdateShaderState(true, true);
 	m_ctx->DrawIndexed(m_index.count, m_index.start, m_vertex.start);
 }
 
@@ -1120,7 +1120,7 @@ void GSDevice11::DrawIndexedPrimitive(int offset, int count)
 {
 	pxAssert(offset + count <= (int)m_index.count);
 	g_perfmon.Put(GSPerfMon::DrawCalls, 1);
-	PSUpdateShaderState();
+	PSUpdateShaderState(true, true);
 	m_ctx->DrawIndexed(count, m_index.start + offset, m_vertex.start);
 }
 
@@ -1280,8 +1280,10 @@ void GSDevice11::StretchRect(GSTexture* sTex, const GSVector4& sRect, GSTexture*
 	GSVector2i ds;
 	if (dTex)
 	{
-		ds = dTex->GetSize();
+		// ps unbind conflicting srvs
 		PSUnbindConflictingSRVs(dTex);
+
+		ds = dTex->GetSize();
 		if (draw_in_depth)
 			OMSetRenderTargets(nullptr, dTex);
 		else
@@ -1293,6 +1295,7 @@ void GSDevice11::StretchRect(GSTexture* sTex, const GSVector4& sRect, GSTexture*
 	}
 
 	// om
+
 	if (draw_in_depth)
 		OMSetDepthStencilState(m_convert.dss_write.get(), 0);
 	else
@@ -1342,6 +1345,9 @@ void GSDevice11::PresentRect(GSTexture* sTex, const GSVector4& sRect, GSTexture*
 	GSVector2i ds;
 	if (dTex)
 	{
+		// ps unbind conflicting srvs
+		PSUnbindConflictingSRVs(dTex);
+
 		ds = dTex->GetSize();
 		OMSetRenderTargets(dTex, nullptr);
 	}
@@ -1357,6 +1363,7 @@ void GSDevice11::PresentRect(GSTexture* sTex, const GSVector4& sRect, GSTexture*
 	m_ctx->UpdateSubresource(m_present.ps_cb.get(), 0, nullptr, &cb, 0, 0);
 
 	// om
+
 	OMSetDepthStencilState(m_convert.dss.get(), 0);
 	OMSetBlendState(m_convert.bs[D3D11_COLOR_WRITE_ENABLE_ALL].get(), 0);
 
@@ -1455,9 +1462,9 @@ void GSDevice11::DrawMultiStretchRects(const MultiStretchRect* rects, u32 num_re
 
 	VSSetShader(m_convert.vs.get(), nullptr);
 	PSSetShader(m_convert.ps[static_cast<int>(shader)].get(), nullptr);
+	PSUnbindConflictingSRVs(dTex);
 
 	OMSetDepthStencilState(dTex->IsRenderTarget() ? m_convert.dss.get() : m_convert.dss_write.get(), 0);
-	PSUnbindConflictingSRVs(dTex);
 	OMSetRenderTargets(dTex->IsRenderTarget() ? dTex : nullptr, dTex->IsDepthStencil() ? dTex : nullptr);
 
 	const GSVector2 ds(static_cast<float>(dTex->GetWidth()), static_cast<float>(dTex->GetHeight()));
@@ -2112,7 +2119,7 @@ void GSDevice11::RenderImGui()
 
 			// Since we don't have the GSTexture...
 			m_state.ps_sr_views[0] = reinterpret_cast<ID3D11ShaderResourceView*>(pcmd->GetTexID());
-			PSUpdateShaderState();
+			PSUpdateShaderState(true, true);
 
 			m_ctx->DrawIndexed(pcmd->ElemCount, m_index.start + pcmd->IdxOffset, vertex_offset + pcmd->VtxOffset);
 		}
@@ -2132,6 +2139,10 @@ void GSDevice11::SetupDATE(GSTexture* rt, GSTexture* ds, const GSVertexPT1* vert
 
 	m_ctx->ClearDepthStencilView(*static_cast<GSTexture11*>(ds), D3D11_CLEAR_STENCIL, 0.0f, 0);
 
+	// ps unbind conflicting srvs
+
+	PSUnbindConflictingSRVs(ds);
+
 	// om
 
 	OMSetDepthStencilState(m_date.dss.get(), 1);
@@ -2360,10 +2371,46 @@ void GSDevice11::PSSetShader(ID3D11PixelShader* ps, ID3D11Buffer* ps_cb)
 	}
 }
 
-void GSDevice11::PSUpdateShaderState()
+void GSDevice11::PSUpdateShaderState(const bool sr_update, const bool ss_update)
 {
-	m_ctx->PSSetShaderResources(0, m_state.ps_sr_views.size(), m_state.ps_sr_views.data());
-	m_ctx->PSSetSamplers(0, m_state.ps_ss.size(), m_state.ps_ss.data());
+	// Shader resource caching requires srv/rtv hazards to be resolved, ensure PSUnbindConflictingSRVs handle.
+	if (sr_update)
+	{
+		bool sr_changed = false;
+		for (size_t i = 0; i < m_state.ps_sr_views.size(); ++i)
+		{
+			if (m_state.ps_cached_sr_views[i] != m_state.ps_sr_views[i])
+			{
+				sr_changed = true;
+				break;
+			}
+		}
+
+		if (sr_changed)
+		{
+			m_state.ps_cached_sr_views = m_state.ps_sr_views;
+			m_ctx->PSSetShaderResources(0, m_state.ps_sr_views.size(), m_state.ps_sr_views.data());
+		}
+	}
+
+	if (ss_update)
+	{
+		bool ss_changed = false;
+		for (size_t i = 0; i < m_state.ps_ss.size(); ++i)
+		{
+			if (m_state.ps_cached_ss[i] != m_state.ps_ss[i])
+			{
+				ss_changed = true;
+				break;
+			}
+		}
+
+		if (ss_changed)
+		{
+			m_state.ps_cached_ss = m_state.ps_ss;
+			m_ctx->PSSetSamplers(0, m_state.ps_ss.size(), m_state.ps_ss.data());
+		}
+	}
 }
 
 void GSDevice11::PSUnbindConflictingSRVs(GSTexture* tex1, GSTexture* tex2)
@@ -2380,7 +2427,7 @@ void GSDevice11::PSUnbindConflictingSRVs(GSTexture* tex1, GSTexture* tex2)
 	}
 
 	if (changed)
-		PSUpdateShaderState();
+		PSUpdateShaderState(true, false);
 }
 
 void GSDevice11::OMSetDepthStencilState(ID3D11DepthStencilState* dss, u8 sref)

pcsx2/GS/Renderers/DX11/GSDevice11.h

@@ -145,9 +145,11 @@ private:
 		ID3D11VertexShader* vs;
 		ID3D11Buffer* vs_cb;
 		std::array<ID3D11ShaderResourceView*, MAX_TEXTURES> ps_sr_views;
+		std::array<ID3D11ShaderResourceView*, MAX_TEXTURES> ps_cached_sr_views;
 		ID3D11PixelShader* ps;
 		ID3D11Buffer* ps_cb;
 		std::array<ID3D11SamplerState*, MAX_SAMPLERS> ps_ss;
+		std::array<ID3D11SamplerState*, MAX_SAMPLERS> ps_cached_ss;
 		GSVector2i viewport;
 		GSVector4i scissor;
 		u32 vb_stride;
@@ -324,7 +326,7 @@ public:
 
 	void PSSetShaderResource(int i, GSTexture* sr);
 	void PSSetShader(ID3D11PixelShader* ps, ID3D11Buffer* ps_cb);
-	void PSUpdateShaderState();
+	void PSUpdateShaderState(const bool sr_update, const bool ss_update);
 	void PSUnbindConflictingSRVs(GSTexture* tex1 = nullptr, GSTexture* tex2 = nullptr);
 	void PSSetSamplerState(ID3D11SamplerState* ss0);
 

pcsx2/GS/Renderers/OpenGL/GSDeviceOGL.cpp

@@ -2585,6 +2585,7 @@ void GSDeviceOGL::RenderHW(GSHWDrawConfig& config)
 	{
 		// Ensure all depth writes are finished before sampling
 		GL_INS("GL: Texture barrier to flush depth or rt before reading");
+		g_perfmon.Put(GSPerfMon::Barriers, 1);
 		glTextureBarrier();
 	}
 	// additional non-pipeline config stuff