GPU: Skip alpha stencil upload when safe.

Otherwise, let's not do things that give different behavior at 1x or are
inconsistent.

Common/GPU/OpenGL/GLFeatures.cpp

@@ -370,7 +370,7 @@ void CheckGLExtensions() {
 	gl_extensions.ARB_depth_clamp = g_set_gl_extensions.count("GL_ARB_depth_clamp") != 0;
 	gl_extensions.ARB_uniform_buffer_object = g_set_gl_extensions.count("GL_ARB_uniform_buffer_object") != 0;
 	gl_extensions.ARB_explicit_attrib_location = g_set_gl_extensions.count("GL_ARB_explicit_attrib_location") != 0;
-	gl_extensions.ARB_texture_non_power_of_two = g_set_gl_extensions.count("GL_ARB_texture_non_power_of_two");
+	gl_extensions.ARB_texture_non_power_of_two = g_set_gl_extensions.count("GL_ARB_texture_non_power_of_two") != 0;
 	if (gl_extensions.IsGLES) {
 		gl_extensions.EXT_blend_func_extended = g_set_gl_extensions.count("GL_EXT_blend_func_extended") != 0;
 		gl_extensions.OES_texture_npot = g_set_gl_extensions.count("GL_OES_texture_npot") != 0;

Core/HLE/ReplaceTables.cpp

@@ -708,7 +708,7 @@ static int Hook_hexyzforce_monoclome_thread() {
 static int Hook_starocean_write_stencil() {
 	const u32 fb_address = currentMIPS->r[MIPS_REG_T7];
 	if (Memory::IsVRAMAddress(fb_address)) {
-		gpu->PerformStencilUpload(fb_address, 0x00088000);
+		gpu->PerformStencilUpload(fb_address, 0x00088000, StencilUpload::IGNORE_ALPHA);
 	}
 	return 0;
 }

GPU/Common/FramebufferManagerCommon.h

@@ -171,11 +171,6 @@ inline DrawTextureFlags operator | (const DrawTextureFlags &lhs, const DrawTextu
 	return DrawTextureFlags((u32)lhs | (u32)rhs);
 }
 
-enum class StencilUpload {
-	NEEDS_CLEAR,
-	STENCIL_IS_ZERO,
-};
-
 enum class TempFBO {
 	DEPAL,
 	BLIT,
@@ -261,7 +256,7 @@ public:
 	void NotifyVideoUpload(u32 addr, int size, int width, GEBufferFormat fmt);
 	void UpdateFromMemory(u32 addr, int size, bool safe);
 	void ApplyClearToMemory(int x1, int y1, int x2, int y2, u32 clearColor);
-	bool PerformStencilUpload(u32 addr, int size, StencilUpload flags = StencilUpload::NEEDS_CLEAR);
+	bool PerformStencilUpload(u32 addr, int size, StencilUpload flags);
 	// Returns true if it's sure this is a direct FBO->FBO transfer and it has already handle it.
 	// In that case we hardly need to actually copy the bytes in VRAM, they will be wrong anyway (unless
 	// read framebuffers is on, in which case this should always return false).

GPU/Common/StencilCommon.cpp

@@ -182,17 +182,17 @@ bool FramebufferManagerCommon::PerformStencilUpload(u32 addr, int size, StencilU
 			return false;
 		}
 
-		shaderManager_->DirtyLastShader();
+		// Otherwise, we can skip alpha in many cases, in which case we don't even use a shader.
+		if (flags == StencilUpload::IGNORE_ALPHA) {
+			shaderManager_->DirtyLastShader();
 
-		// Let's not bother with the shader if it's just zero.
-		if (dstBuffer->fbo) {
-			draw_->BindFramebufferAsRenderTarget(dstBuffer->fbo, { Draw::RPAction::KEEP, Draw::RPAction::KEEP, Draw::RPAction::CLEAR }, "PerformStencilUpload_Clear");
+			if (dstBuffer->fbo) {
+				draw_->BindFramebufferAsRenderTarget(dstBuffer->fbo, { Draw::RPAction::KEEP, Draw::RPAction::KEEP, Draw::RPAction::CLEAR }, "PerformStencilUpload_Clear");
+			}
+
+			gstate_c.Dirty(DIRTY_BLEND_STATE | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_DEPTHSTENCIL_STATE);
+			return true;
 		}
-
-		// Here we might want to clear destination alpha by using a draw, but we haven't found a need for this yet.
-		// Will implement when needed...
-		gstate_c.Dirty(DIRTY_BLEND_STATE | DIRTY_VIEWPORTSCISSOR_STATE | DIRTY_DEPTHSTENCIL_STATE);
-		return true;
 	}
 
 	shaderManager_->DirtyLastShader();
@@ -263,10 +263,13 @@ bool FramebufferManagerCommon::PerformStencilUpload(u32 addr, int size, StencilU
 	// Our fragment shader (and discard) is slow.  Since the source is 1x, we can stencil to 1x.
 	// Then after we're done, we'll just blit it across and stretch it there. Not worth doing
 	// if already at 1x size though, of course.
-	// TODO: This path means that we don't write color alpha... Ugh.
 	if (dstBuffer->width == dstBuffer->renderWidth || !dstBuffer->fbo) {
 		useBlit = false;
 	}
+	// The blit path doesn't set alpha, so we can't use it if that's needed.
+	if (flags != StencilUpload::IGNORE_ALPHA) {
+		useBlit = false;
+	}
 
 	u16 w = useBlit ? dstBuffer->width : dstBuffer->renderWidth;
 	u16 h = useBlit ? dstBuffer->height : dstBuffer->renderHeight;

GPU/GPUCommon.cpp

@@ -2931,9 +2931,9 @@ void GPUCommon::NotifyVideoUpload(u32 addr, int size, int width, int format) {
 	InvalidateCache(addr, size, GPU_INVALIDATE_SAFE);
 }
 
-bool GPUCommon::PerformStencilUpload(u32 dest, int size) {
+bool GPUCommon::PerformStencilUpload(u32 dest, int size, StencilUpload flags) {
 	if (framebufferManager_->MayIntersectFramebuffer(dest)) {
-		framebufferManager_->PerformStencilUpload(dest, size);
+		framebufferManager_->PerformStencilUpload(dest, size, flags);
 		return true;
 	}
 	return false;

GPU/GPUCommon.h

@@ -124,7 +124,7 @@ public:
 
 	void InvalidateCache(u32 addr, int size, GPUInvalidationType type) override;
 	void NotifyVideoUpload(u32 addr, int size, int width, int format) override;
-	bool PerformStencilUpload(u32 dest, int size) override;
+	bool PerformStencilUpload(u32 dest, int size, StencilUpload flags) override;
 
 	void Execute_OffsetAddr(u32 op, u32 diff);
 	void Execute_Vaddr(u32 op, u32 diff);

GPU/GPUInterface.h

@@ -106,6 +106,12 @@ enum GPUSyncType {
 	GPU_SYNC_LIST,
 };
 
+enum class StencilUpload {
+	NEEDS_CLEAR,
+	STENCIL_IS_ZERO,
+	IGNORE_ALPHA,
+};
+
 // Used for debug
 struct FramebufferInfo {
 	u32 fb_address;
@@ -216,7 +222,7 @@ public:
 	virtual bool PerformMemorySet(u32 dest, u8 v, int size) = 0;
 	virtual bool PerformMemoryDownload(u32 dest, int size) = 0;
 	virtual bool PerformMemoryUpload(u32 dest, int size) = 0;
-	virtual bool PerformStencilUpload(u32 dest, int size) = 0;
+	virtual bool PerformStencilUpload(u32 dest, int size, StencilUpload flags = StencilUpload::NEEDS_CLEAR) = 0;
 
 	// Will cause the texture cache to be cleared at the start of the next frame.
 	virtual void ClearCacheNextFrame() = 0;

GPU/Software/SoftGpu.cpp

@@ -1161,7 +1161,7 @@ bool SoftGPU::PerformMemoryUpload(u32 dest, int size)
 	return false;
 }
 
-bool SoftGPU::PerformStencilUpload(u32 dest, int size)
+bool SoftGPU::PerformStencilUpload(u32 dest, int size, StencilUpload flags)
 {
 	return false;
 }

GPU/Software/SoftGpu.h

@@ -141,7 +141,7 @@ public:
 	bool PerformMemorySet(u32 dest, u8 v, int size) override;
 	bool PerformMemoryDownload(u32 dest, int size) override;
 	bool PerformMemoryUpload(u32 dest, int size) override;
-	bool PerformStencilUpload(u32 dest, int size) override;
+	bool PerformStencilUpload(u32 dest, int size, StencilUpload flags) override;
 	void ClearCacheNextFrame() override {}
 
 	void DeviceLost() override;