GPU: Hook Gods Eater Burst avatar read

Core/HLE/ReplaceTables.cpp

@@ -691,6 +691,35 @@ static int Hook_godseaterburst_blit_texture() {
 	return 0;
 }
 
+static int Hook_godseaterburst_depthmask_5551() {
+	// This function copies the 5551 framebuffer to a temporary, generating alpha based on depth.
+	// Depth is optional, in which case all pixels get full alpha.
+	// Called when your avatar changes to screenshot for save data.
+	uint32_t colorBuffer = currentMIPS->r[MIPS_REG_A1];
+	uint32_t depthBuffer = currentMIPS->r[MIPS_REG_T2];
+	uint32_t byteStride = currentMIPS->r[MIPS_REG_A2];
+	uint32_t height = currentMIPS->r[MIPS_REG_T1];
+	uint32_t size = byteStride * height;
+
+	if (!Memory::IsVRAMAddress(colorBuffer) || !Memory::IsValidRange(colorBuffer, size))
+		return 0;
+	if (depthBuffer != 0) {
+		if (!Memory::IsVRAMAddress(colorBuffer) || !Memory::IsValidRange(depthBuffer, size))
+			return 0;
+
+		// This is added to read from the linearized mirror.
+		uint32_t depthMirror = depthBuffer + 0x00200000;
+		// Depth download required, or it won't work and will be transparent.
+		gpu->PerformMemoryCopy(depthMirror, depthMirror, size, GPUCopyFlag::FORCE_DST_MEM | GPUCopyFlag::DEPTH_REQUESTED);
+		NotifyMemInfo(MemBlockFlags::WRITE, depthMirror, size, "godseaterburst_depthmask_5551");
+	}
+
+	gpu->PerformMemoryDownload(colorBuffer, size);
+	NotifyMemInfo(MemBlockFlags::WRITE, colorBuffer, size, "godseaterburst_depthmask_5551");
+
+	return 0;
+}
+
 static int Hook_hexyzforce_monoclome_thread() {
 	u32 fb_info;
 	if (!GetMIPSStaticAddress(fb_info, -4, 0)) {
@@ -1414,6 +1443,7 @@ static const ReplacementTableEntry entries[] = {
 	// { "vmmul_q_transp", &Replace_vmmul_q_transp, 0, REPFLAG_DISABLED },
 
 	{ "godseaterburst_blit_texture", &Hook_godseaterburst_blit_texture, 0, REPFLAG_HOOKENTER },
+	{ "godseaterburst_depthmask_5551", &Hook_godseaterburst_depthmask_5551, 0, REPFLAG_HOOKENTER },
 	{ "hexyzforce_monoclome_thread", &Hook_hexyzforce_monoclome_thread, 0, REPFLAG_HOOKENTER, 0x58 },
 	{ "starocean_write_stencil", &Hook_starocean_write_stencil, 0, REPFLAG_HOOKENTER, 0x260 },
 	{ "topx_create_saveicon", &Hook_topx_create_saveicon, 0, REPFLAG_HOOKENTER, 0x34 },

Core/MIPS/MIPSAnalyst.cpp

@@ -289,6 +289,7 @@ static const HardHashTableEntry hardcodedHashes[] = {
 	{ 0x70a6152b265228e8, 296, "unendingbloodycall_download_frame", }, // unENDing Bloody Call
 	{ 0x7245b74db370ae72, 64, "vmmul_q_transp3", },
 	{ 0x7259d52b21814a5a, 40, "vtfm_t_transp", },
+	{ 0x730f59cc6c0f5732, 452, "godseaterburst_depthmask_5551", }, // Gods Eater Burst (US)
 	{ 0x7354fd206796d817, 864, "flowers_download_frame", }, // Flowers
 	{ 0x736b34ebc702d873, 104, "vmmul_q_transp", },
 	{ 0x73a614c08f777d52, 792, "danganronpa2_2_download_frame", }, // Danganronpa 2

GPU/Common/FramebufferManagerCommon.cpp

@@ -1657,13 +1657,14 @@ bool FramebufferManagerCommon::NotifyFramebufferCopy(u32 src, u32 dst, int size,
 	VirtualFramebuffer *srcBuffer = nullptr;
 	bool ignoreDstBuffer = flags & GPUCopyFlag::FORCE_DST_MEM;
 	bool ignoreSrcBuffer = flags & (GPUCopyFlag::FORCE_SRC_MEM | GPUCopyFlag::MEMSET);
+	RasterChannel channel = flags & GPUCopyFlag::DEPTH_REQUESTED ? RASTER_DEPTH : RASTER_COLOR;
 
 	u32 dstY = (u32)-1;
 	u32 dstH = 0;
 	u32 srcY = (u32)-1;
 	u32 srcH = 0;
 	for (auto vfb : vfbs_) {
-		if (vfb->fb_stride == 0) {
+		if (vfb->fb_stride == 0 || channel != RASTER_COLOR) {
 			continue;
 		}
 
@@ -1715,14 +1716,36 @@ bool FramebufferManagerCommon::NotifyFramebufferCopy(u32 src, u32 dst, int size,
 		}
 	}
 
+	if (channel == RASTER_DEPTH) {
+		srcBuffer = nullptr;
+		dstBuffer = nullptr;
+		// Let's assume exact matches only for simplicity.
+		for (auto vfb : vfbs_) {
+			if (!ignoreDstBuffer && dst == vfb->z_address && size == vfb->z_stride * 2 * vfb->height) {
+				if (!dstBuffer || dstBuffer->depthBindSeq < vfb->depthBindSeq) {
+					dstBuffer = vfb;
+					dstY = 0;
+					dstH = vfb->height;
+				}
+			}
+			if (!ignoreSrcBuffer && src == vfb->z_address && size == vfb->z_stride * 2 * vfb->height) {
+				if (!srcBuffer || srcBuffer->depthBindSeq < vfb->depthBindSeq) {
+					srcBuffer = vfb;
+					srcY = 0;
+					srcH = vfb->height;
+				}
+			}
+		}
+	}
+
 	if (!useBufferedRendering_) {
 		// If we're copying into a recently used display buf, it's probably destined for the screen.
-		if (srcBuffer || (dstBuffer != displayFramebuf_ && dstBuffer != prevDisplayFramebuf_)) {
+		if (channel == RASTER_DEPTH || srcBuffer || (dstBuffer != displayFramebuf_ && dstBuffer != prevDisplayFramebuf_)) {
 			return false;
 		}
 	}
 
-	if (!dstBuffer && srcBuffer) {
+	if (!dstBuffer && srcBuffer && channel != RASTER_DEPTH) {
 		// Note - if we're here, we're in a memcpy, not a block transfer. Not allowing IntraVRAMBlockTransferAllowCreateFB.
 		// Technically, that makes BlockTransferAllowCreateFB a bit of a misnomer.
 		if (PSP_CoreParameter().compat.flags().BlockTransferAllowCreateFB) {
@@ -1740,7 +1763,7 @@ bool FramebufferManagerCommon::NotifyFramebufferCopy(u32 src, u32 dst, int size,
 		} else {
 			WARN_LOG_ONCE(dstnotsrccpy, G3D, "Inter-buffer memcpy %08x -> %08x (size: %x)", src, dst, size);
 			// Just do the blit!
-			BlitFramebuffer(dstBuffer, 0, dstY, srcBuffer, 0, srcY, srcBuffer->width, srcH, 0, RASTER_COLOR, "Blit_InterBufferMemcpy");
+			BlitFramebuffer(dstBuffer, 0, dstY, srcBuffer, 0, srcY, srcBuffer->width, srcH, 0, channel, "Blit_InterBufferMemcpy");
 			SetColorUpdated(dstBuffer, skipDrawReason);
 			RebindFramebuffer("RebindFramebuffer - Inter-buffer memcpy");
 		}
@@ -1752,7 +1775,9 @@ bool FramebufferManagerCommon::NotifyFramebufferCopy(u32 src, u32 dst, int size,
 		WARN_LOG_ONCE(btucpy, G3D, "Memcpy fbo upload %08x -> %08x (size: %x)", src, dst, size);
 		FlushBeforeCopy();
 		const u8 *srcBase = Memory::GetPointerUnchecked(src);
-		DrawPixels(dstBuffer, 0, dstY, srcBase, dstBuffer->fb_format, dstBuffer->fb_stride, dstBuffer->width, dstH, RASTER_COLOR, "MemcpyFboUpload_DrawPixels");
+		GEBufferFormat srcFormat = channel == RASTER_DEPTH ? GE_FORMAT_DEPTH16 : dstBuffer->fb_format;
+		int srcStride = channel == RASTER_DEPTH ? dstBuffer->z_stride : dstBuffer->fb_stride;
+		DrawPixels(dstBuffer, 0, dstY, srcBase, srcFormat, srcStride, dstBuffer->width, dstH, channel, "MemcpyFboUpload_DrawPixels");
 		SetColorUpdated(dstBuffer, skipDrawReason);
 		RebindFramebuffer("RebindFramebuffer - Memcpy fbo upload");
 		// This is a memcpy, let's still copy just in case.
@@ -1762,8 +1787,8 @@ bool FramebufferManagerCommon::NotifyFramebufferCopy(u32 src, u32 dst, int size,
 		FlushBeforeCopy();
 		if (srcH == 0 || srcY + srcH > srcBuffer->bufferHeight) {
 			WARN_LOG_ONCE(btdcpyheight, G3D, "Memcpy fbo download %08x -> %08x skipped, %d+%d is taller than %d", src, dst, srcY, srcH, srcBuffer->bufferHeight);
-		} else if (g_Config.bBlockTransferGPU && !srcBuffer->memoryUpdated) {
-			ReadFramebufferToMemory(srcBuffer, 0, srcY, srcBuffer->width, srcH, RASTER_COLOR);
+		} else if (g_Config.bBlockTransferGPU && (!srcBuffer->memoryUpdated || channel == RASTER_DEPTH)) {
+			ReadFramebufferToMemory(srcBuffer, 0, srcY, srcBuffer->width, srcH, channel);
 			srcBuffer->usageFlags = (srcBuffer->usageFlags | FB_USAGE_DOWNLOAD) & ~FB_USAGE_DOWNLOAD_CLEAR;
 		}
 		return false;
@@ -2580,7 +2605,7 @@ void FramebufferManagerCommon::PackFramebufferSync(VirtualFramebuffer *vfb, int
 	const int dstByteOffset = (y * stride + x) * dstBpp;
 	// Leave the gap between the end of the last line and the full stride.
 	// This is only used for the NotifyMemInfo range.
-	const int dstSize = (h * stride + w - 1) * dstBpp;
+	const int dstSize = ((h - 1) * stride + w) * dstBpp;
 
 	if (!Memory::IsValidRange(fb_address + dstByteOffset, dstSize)) {
 		ERROR_LOG_REPORT(G3D, "PackFramebufferSync would write outside of memory, ignoring");
@@ -2594,7 +2619,11 @@ void FramebufferManagerCommon::PackFramebufferSync(VirtualFramebuffer *vfb, int
 	DEBUG_LOG(G3D, "Reading framebuffer to mem, fb_address = %08x, ptr=%p", fb_address, destPtr);
 
 	if (destPtr) {
-		draw_->CopyFramebufferToMemorySync(vfb->fbo, channel == RASTER_COLOR ? Draw::FB_COLOR_BIT : Draw::FB_DEPTH_BIT, x, y, w, h, destFormat, destPtr, vfb->fb_stride, "PackFramebufferSync");
+		if (channel == RASTER_DEPTH)
+			PackDepthbuffer(vfb, x, y, w, h);
+		else
+			draw_->CopyFramebufferToMemorySync(vfb->fbo, channel == RASTER_COLOR ? Draw::FB_COLOR_BIT : Draw::FB_DEPTH_BIT, x, y, w, h, destFormat, destPtr, vfb->fb_stride, "PackFramebufferSync");
+
 		char tag[128];
 		size_t len = snprintf(tag, sizeof(tag), "FramebufferPack/%08x_%08x_%dx%d_%s", vfb->fb_address, vfb->z_address, w, h, GeBufferFormatToString(vfb->fb_format));
 		NotifyMemInfo(MemBlockFlags::WRITE, fb_address + dstByteOffset, dstSize, tag, len);
@@ -2605,6 +2634,17 @@ void FramebufferManagerCommon::PackFramebufferSync(VirtualFramebuffer *vfb, int
 	gpuStats.numReadbacks++;
 }
 
+void FramebufferManagerCommon::PackDepthbuffer(VirtualFramebuffer *vfb, int x, int y, int w, int h) {
+	_assert_msg_(vfb && vfb->z_address != 0 && vfb->z_stride != 0, "Depth buffer invalid");
+
+	Draw::DataFormat destFormat = GEFormatToThin3D(GE_FORMAT_DEPTH16);
+	const int dstByteOffset = (y * vfb->z_stride + x) * 2;
+	u8 *destPtr = Memory::GetPointerWriteUnchecked(vfb->z_address + dstByteOffset);
+	if (!draw_->CopyFramebufferToMemorySync(vfb->fbo, Draw::FB_DEPTH_BIT, x, y, w, h, destFormat, destPtr, vfb->fb_stride, "PackDepthbuffer")) {
+		WARN_LOG(G3D, "PackDepthbuffer failed");
+	}
+}
+
 void FramebufferManagerCommon::ReadFramebufferToMemory(VirtualFramebuffer *vfb, int x, int y, int w, int h, RasterChannel channel) {
 	// Clamp to bufferWidth. Sometimes block transfers can cause this to hit.
 	if (x + w >= vfb->bufferWidth) {
@@ -2623,7 +2663,8 @@ void FramebufferManagerCommon::ReadFramebufferToMemory(VirtualFramebuffer *vfb,
 			vfb->usageFlags |= FB_USAGE_DOWNLOAD;
 		} else if (x == 0 && y == 0 && w == vfb->width && h == vfb->height) {
 			// Mark it as fully downloaded until next render to it.
-			vfb->memoryUpdated = true;
+			if (channel == RASTER_COLOR)
+				vfb->memoryUpdated = true;
 			vfb->usageFlags |= FB_USAGE_DOWNLOAD;
 		} else {
 			// Let's try to set the flag eventually, if the game copies a lot.

GPU/Common/FramebufferManagerCommon.h

@@ -442,6 +442,8 @@ class FramebufferManagerCommon {
 
 protected:
 	virtual void PackFramebufferSync(VirtualFramebuffer *vfb, int x, int y, int w, int h, RasterChannel channel);
+	// Used for when a shader is required, such as GLES.
+	virtual void PackDepthbuffer(VirtualFramebuffer *vfb, int x, int y, int w, int h);
 	void SetViewport2D(int x, int y, int w, int h);
 	Draw::Texture *MakePixelTexture(const u8 *srcPixels, GEBufferFormat srcPixelFormat, int srcStride, int width, int height);
 	void DrawActiveTexture(float x, float y, float w, float h, float destW, float destH, float u0, float v0, float u1, float v1, int uvRotation, int flags);

GPU/Directx9/FramebufferManagerDX9.h

@@ -49,10 +49,10 @@ class FramebufferManagerDX9 : public FramebufferManagerCommon {
 
 protected:
 	void DecimateFBOs() override;
+	void PackDepthbuffer(VirtualFramebuffer *vfb, int x, int y, int w, int h) override;
 
 private:
 	void PackFramebufferSync(VirtualFramebuffer *vfb, int x, int y, int w, int h, RasterChannel channel) override;
-	void PackDepthbuffer(VirtualFramebuffer *vfb, int x, int y, int w, int h);
 	bool GetRenderTargetFramebuffer(LPDIRECT3DSURFACE9 renderTarget, LPDIRECT3DSURFACE9 offscreen, int w, int h, GPUDebugBuffer &buffer);
 
 	LPDIRECT3DDEVICE9 device_;

GPU/GLES/FramebufferManagerGLES.h

@@ -38,10 +38,9 @@ class FramebufferManagerGLES : public FramebufferManagerCommon {
 
 protected:
 	void UpdateDownloadTempBuffer(VirtualFramebuffer *nvfb) override;
+	void PackDepthbuffer(VirtualFramebuffer *vfb, int x, int y, int w, int h) override;
 
 private:
-	void PackDepthbuffer(VirtualFramebuffer *vfb, int x, int y, int w, int h);
-
 	u8 *convBuf_ = nullptr;
 	u32 convBufSize_ = 0;
 

GPU/GPUInterface.h

@@ -120,7 +120,8 @@ enum class GPUCopyFlag {
 	FORCE_DST_MEM = 2,
 	// Note: implies src == dst and FORCE_SRC_MEM.
 	MEMSET = 4,
-	DEBUG_NOTIFIED = 8,
+	DEPTH_REQUESTED = 8,
+	DEBUG_NOTIFIED = 16,
 };
 ENUM_CLASS_BITOPS(GPUCopyFlag);