GS: Use tex-is-fb for sprites/shuffles

pcsx2/GS/Renderers/HW/GSRendererHW.cpp

@@ -28,6 +28,7 @@ GSRendererHW::GSRendererHW()
 	, m_tc(new GSTextureCache())
 	, m_src(nullptr)
 	, m_reset(false)
+	, m_tex_is_fb(false)
 	, m_channel_shuffle(false)
 	, m_userhacks_tcoffset(false)
 	, m_userhacks_tcoffset_x(0)
@@ -120,10 +121,41 @@ void GSRendererHW::PurgeTextureCache()
 	m_tc->RemoveAll();
 }
 
-bool GSRendererHW::IsPossibleTextureShuffle(GSTextureCache::Source* src) const
+bool GSRendererHW::UpdateTexIsFB(GSTextureCache::Target* dst, const GIFRegTEX0& TEX0)
+{
+	if (GSConfig.AccurateBlendingUnit == AccBlendLevel::Minimum || !g_gs_device->Features().texture_barrier)
+		return false;
+
+	// Texture is actually the frame buffer. Stencil emulation to compute shadow (Jak series/tri-ace game)
+	// Will hit the "m_ps_sel.tex_is_fb = 1" path in the draw
+	if (m_vt.m_primclass == GS_TRIANGLE_CLASS)
+	{
+		if (m_context->FRAME.FBMSK == 0x00FFFFFF && TEX0.TBP0 == m_context->FRAME.Block())
+			m_tex_is_fb = true;
+	}
+	else if (m_vt.m_primclass == GS_SPRITE_CLASS)
+	{
+		if (TEX0.TBP0 == m_context->FRAME.Block())
+		{
+			m_tex_is_fb = IsPossibleTextureShuffle(dst, TEX0);
+
+			if (!m_tex_is_fb && !m_vt.IsLinear())
+			{
+				// Make sure that we're not sampling away from the area we're rendering.
+				const GSVector4 diff(m_vt.m_min.p.xyxy(m_vt.m_max.p) - m_vt.m_min.t.xyxy(m_vt.m_max.t));
+				if ((diff < GSVector4(1.0f)).alltrue())
+					m_tex_is_fb = true;
+			}
+		}
+	}
+
+	return m_tex_is_fb;
+}
+
+bool GSRendererHW::IsPossibleTextureShuffle(GSTextureCache::Target* dst, const GIFRegTEX0& TEX0) const
 {
 	return (PRIM->TME && m_vt.m_primclass == GS_SPRITE_CLASS &&
-		src->m_32_bits_fmt && GSLocalMemory::m_psm[src->m_TEX0.PSM].bpp == 16 &&
+		dst->m_32_bits_fmt && GSLocalMemory::m_psm[TEX0.PSM].bpp == 16 &&
 		GSLocalMemory::m_psm[m_context->FRAME.PSM].bpp == 16);
 }
 
@@ -1320,6 +1352,7 @@ void GSRendererHW::Draw()
 
 	m_src = nullptr;
 	m_texture_shuffle = false;
+	m_tex_is_fb = false;
 
 	if (PRIM->TME)
 	{
@@ -2256,21 +2289,8 @@ void GSRendererHW::EmulateChannelShuffle(const GSTextureCache::Source* tex)
 		m_conf.tex = *tex->m_from_target;
 		if (m_conf.tex)
 		{
-			if (m_conf.tex == m_conf.rt)
+			// Identify when we're sampling the current buffer, defer fixup for later.
-			{
+			m_tex_is_fb |= (m_conf.tex == m_conf.rt || m_conf.tex == m_conf.ds);
-				// sample from fb instead
-				m_conf.tex = nullptr;
-				m_conf.ps.tex_is_fb = true;
-				m_conf.require_one_barrier = true;
-			}
-			else if (m_conf.tex == m_conf.ds)
-			{
-				// if depth testing is disabled, we don't need to copy, and can just unbind the depth buffer
-				// no need for a barrier for GL either, since it's not bound to depth and texture concurrently
-				// otherwise, the backend should recognise the hazard, and copy the buffer (D3D/Vulkan).
-				if (m_conf.depth.ztst == ZTST_ALWAYS)
-					m_conf.ds = nullptr;
-			}
 		}
 
 		// Replace current draw with a fullscreen sprite
@@ -3070,6 +3090,41 @@ void GSRendererHW::EmulateTextureSampler(const GSTextureCache::Source* tex)
 	if (!m_channel_shuffle)
 		m_conf.tex = tex->m_texture;
 	m_conf.pal = tex->m_palette;
+
+	// Detect framebuffer read that will need special handling
+	if (m_tex_is_fb)
+	{
+		if (m_conf.tex == m_conf.rt)
+		{
+			// This pattern is used by several games to emulate a stencil (shadow)
+			// Ratchet & Clank, Jak do alpha integer multiplication (tfx) which is mostly equivalent to +1/-1
+			// Tri-Ace (Star Ocean 3/RadiataStories/VP2) uses a palette to handle the +1/-1
+			GL_DBG("Source and Target are the same! Let's sample the framebuffer");
+			m_conf.tex = nullptr;
+			m_conf.ps.tex_is_fb = true;
+			if (m_prim_overlap == PRIM_OVERLAP_NO || !g_gs_device->Features().texture_barrier)
+				m_conf.require_one_barrier = true;
+			else
+				m_conf.require_full_barrier = true;
+		}
+		else if (m_conf.tex == m_conf.ds)
+		{
+			// if depth testing is disabled, we don't need to copy, and can just unbind the depth buffer
+			// no need for a barrier for GL either, since it's not bound to depth and texture concurrently
+			// otherwise, the backend should recognise the hazard, and copy the buffer (D3D/Vulkan).
+			if (m_conf.depth.ztst == ZTST_ALWAYS)
+			{
+				m_conf.ds = nullptr;
+				m_tex_is_fb = false;
+			}
+		}
+		else
+		{
+			// weird... we detected a fb read, but didn't end up using it?
+			DevCon.WriteLn("Tex-is-FB set but not used?");
+			m_tex_is_fb = false;
+		}
+	}
 }
 
 void GSRendererHW::EmulateATST(float& AREF, GSHWDrawConfig::PSSelector& ps, bool pass_2)
@@ -3172,27 +3227,6 @@ void GSRendererHW::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sourc
 	else
 		m_prim_overlap = PRIM_OVERLAP_UNKNOW;
 
-	// Detect framebuffer read that will need special handling
-	if (features.texture_barrier && (m_context->FRAME.Block() == m_context->TEX0.TBP0) && PRIM->TME && GSConfig.AccurateBlendingUnit != AccBlendLevel::Minimum)
-	{
-		const u32 fb_mask = GSLocalMemory::m_psm[m_context->FRAME.PSM].fmsk;
-		if (((m_context->FRAME.FBMSK & fb_mask) == (fb_mask & 0x00FFFFFF)) && (m_vt.m_primclass == GS_TRIANGLE_CLASS))
-		{
-			// This pattern is used by several games to emulate a stencil (shadow)
-			// Ratchet & Clank, Jak do alpha integer multiplication (tfx) which is mostly equivalent to +1/-1
-			// Tri-Ace (Star Ocean 3/RadiataStories/VP2) uses a palette to handle the +1/-1
-			GL_DBG("Source and Target are the same! Let's sample the framebuffer");
-			m_conf.ps.tex_is_fb = 1;
-			m_conf.require_full_barrier = !features.framebuffer_fetch;
-		}
-		else if (m_prim_overlap != PRIM_OVERLAP_NO)
-		{
-			// Note: It is fine if the texture fits in a single GS page. First access will cache
-			// the page in the GS texture buffer.
-			GL_INS("ERROR: Source and Target are the same!");
-		}
-	}
-
 	EmulateTextureShuffleAndFbmask();
 
 	// DATE: selection of the algorithm. Must be done before blending because GL42 is not compatible with blending
@@ -3635,13 +3669,6 @@ void GSRendererHW::DrawPrims(GSTexture* rt, GSTexture* ds, GSTextureCache::Sourc
 	g_gs_device->RenderHW(m_conf);
 }
 
-bool GSRendererHW::IsDummyTexture() const
-{
-	// Texture is actually the frame buffer. Stencil emulation to compute shadow (Jak series/tri-ace game)
-	// Will hit the "m_ps_sel.tex_is_fb = 1" path in the draw
-	return g_gs_device->Features().texture_barrier && (m_context->FRAME.Block() == m_context->TEX0.TBP0) && PRIM->TME && GSConfig.AccurateBlendingUnit != AccBlendLevel::Minimum && m_vt.m_primclass == GS_TRIANGLE_CLASS && (m_context->FRAME.FBMSK == 0x00FFFFFF);
-}
-
 bool GSRendererHW::CanUseSwPrimRender(bool no_rt, bool no_ds, bool draw_sprite_tex)
 {
 	// Master enable.

pcsx2/GS/Renderers/HW/GSRendererHW.h

@@ -151,6 +151,7 @@ private:
 	GSTextureCache::Source* m_src;
 
 	bool m_reset;
+	bool m_tex_is_fb;
 	bool m_channel_shuffle;
 	bool m_userhacks_tcoffset;
 	float m_userhacks_tcoffset_x;
@@ -201,8 +202,8 @@ public:
 	void PurgeTextureCache() override;
 
 	// Called by the texture cache to know if current texture is useful
-	bool IsDummyTexture() const;
+	bool UpdateTexIsFB(GSTextureCache::Target* src, const GIFRegTEX0& TEX0);
 
 	// Called by the texture cache when optimizing the copy range for sources
-	bool IsPossibleTextureShuffle(GSTextureCache::Source* src) const;
+	bool IsPossibleTextureShuffle(GSTextureCache::Target* dst, const GIFRegTEX0& TEX0) const;
 };

pcsx2/GS/Renderers/HW/GSTextureCache.cpp

@@ -1458,7 +1458,7 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 			AttachPaletteToSource(src, psm.pal, true);
 		}
 	}
-	else if (dst && GSRendererHW::GetInstance()->IsDummyTexture())
+	else if (dst && GSRendererHW::GetInstance()->UpdateTexIsFB(dst, TEX0))
 	{
 		// This shortcut is a temporary solution. It isn't a good solution
 		// as it won't work with Channel Shuffle/Texture Shuffle pattern
@@ -1470,21 +1470,23 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 		// Be aware that you can't use StrechRect between BeginScene/EndScene.
 		// So it could be tricky to put in the middle of the DrawPrims
 
-		// Texture is created to keep code compatibility
-		GSTexture* dTex = g_gs_device->CreateTexture(tw, th, false, GSTexture::Format::Color, true);
-
 		// Keep a trace of origin of the texture
-		src->m_texture = dTex;
+		src->m_texture = dst->m_texture;
 		src->m_target = true;
+		src->m_shared_texture = true;
 		src->m_from_target = &dst->m_texture;
 		src->m_from_target_TEX0 = dst->m_TEX0;
 		src->m_end_block = dst->m_end_block;
 		src->m_texture->SetScale(dst->m_texture->GetScale());
+		src->m_32_bits_fmt = dst->m_32_bits_fmt;
 
 		// Even if we sample the framebuffer directly we might need the palette
 		// to handle the format conversion on GPU
 		if (psm.pal > 0)
 			AttachPaletteToSource(src, psm.pal, true);
+
+		// This will get immediately invalidated.
+		m_temporary_source = src;
 	}
 	else if (dst)
 	{
@@ -1626,7 +1628,7 @@ GSTextureCache::Source* GSTextureCache::CreateSource(const GIFRegTEX0& TEX0, con
 			{
 				// if it looks like a texture shuffle, we might read up to +/- 8 pixels on either side.
 				GSVector4 adjusted_src_range(*src_range);
-				if (GSRendererHW::GetInstance()->IsPossibleTextureShuffle(src))
+				if (GSRendererHW::GetInstance()->IsPossibleTextureShuffle(dst, TEX0))
 					adjusted_src_range += GSVector4(-8.0f, 0.0f, 8.0f, 0.0f);
 
 				// don't forget to scale the copy range

pcsx2/GS/Renderers/Vulkan/GSDeviceVK.cpp

@@ -1956,7 +1956,6 @@ VkShaderModule GSDeviceVK::GetTFXFragmentShader(const GSHWDrawConfig::PSSelector
 	AddMacro(ss, "PS_CHANNEL_FETCH", sel.channel);
 	AddMacro(ss, "PS_URBAN_CHAOS_HLE", sel.urban_chaos_hle);
 	AddMacro(ss, "PS_TALES_OF_ABYSS_HLE", sel.tales_of_abyss_hle);
-	AddMacro(ss, "PS_TEX_IS_FB", sel.tex_is_fb);
 	AddMacro(ss, "PS_INVALID_TEX0", sel.invalid_tex0);
 	AddMacro(ss, "PS_AEM", sel.aem);
 	AddMacro(ss, "PS_TFX", sel.tfx);
@@ -2979,22 +2978,32 @@ void GSDeviceVK::RenderHW(GSHWDrawConfig& config)
 		}
 	}
 
-	if (config.tex && config.tex == config.ds)
+	if (config.tex)
 	{
-		// requires a copy of the depth buffer. this is mainly for ico.
+		if (config.tex == config.ds)
-		copy_ds = static_cast<GSTextureVK*>(CreateDepthStencil(rtsize.x, rtsize.y, GSTexture::Format::DepthStencil, false));
-		if (copy_ds)
 		{
-			EndRenderPass();
+			// requires a copy of the depth buffer. this is mainly for ico.
+			copy_ds = static_cast<GSTextureVK*>(CreateDepthStencil(rtsize.x, rtsize.y, GSTexture::Format::DepthStencil, false));
+			if (copy_ds)
+			{
+				EndRenderPass();
 
-			GL_PUSH("Copy depth to temp texture for shuffle {%d,%d %dx%d}",
+				GL_PUSH("Copy depth to temp texture for shuffle {%d,%d %dx%d}",
-				config.drawarea.left, config.drawarea.top,
+					config.drawarea.left, config.drawarea.top,
-				config.drawarea.width(), config.drawarea.height());
+					config.drawarea.width(), config.drawarea.height());
 
-			CopyRect(config.ds, copy_ds, config.drawarea, config.drawarea.left, config.drawarea.top);
+				CopyRect(config.ds, copy_ds, config.drawarea, config.drawarea.left, config.drawarea.top);
-			PSSetShaderResource(0, copy_ds, true);
+				PSSetShaderResource(0, copy_ds, true);
+			}
 		}
 	}
+	// clear texture binding when it's bound to RT or DS
+	if (!config.tex && m_tfx_textures[0] &&
+		((!pipe.feedback_loop && config.rt && static_cast<GSTextureVK*>(config.rt)->GetView() == m_tfx_textures[0]) ||
+			(config.ds && static_cast<GSTextureVK*>(config.ds)->GetView() == m_tfx_textures[0])))
+	{
+		PSSetShaderResource(0, nullptr, false);
+	}
 
 	const bool render_area_okay =
 		(!hdr_rt && DATE_rp != DATE_RENDER_PASS_STENCIL_ONE && CheckRenderPassArea(render_area));