Merge pull request #8441 from unknownbrackets/gpu-mpeg

Use video format when uploading directly to framebuffers

Core/HLE/sceMpeg.cpp

@@ -1094,7 +1094,7 @@ static u32 sceMpegAvcDecode(u32 mpeg, u32 auAddr, u32 frameWidth, u32 bufferAddr
 			// playing all pmp_queue frames
 			ctx->mediaengine->m_pFrameRGB = pmp_queue.front();
 			int bufferSize = ctx->mediaengine->writeVideoImage(buffer, frameWidth, ctx->videoPixelMode);
-			gpu->InvalidateCache(buffer, bufferSize, GPU_INVALIDATE_SAFE);
+			gpu->NotifyVideoUpload(buffer, bufferSize, frameWidth, ctx->videoPixelMode);
 			ctx->avc.avcFrameStatus = 1;
 			ctx->videoFrameCount++;
 			
@@ -1105,7 +1105,7 @@ static u32 sceMpegAvcDecode(u32 mpeg, u32 auAddr, u32 frameWidth, u32 bufferAddr
 	}
 	else if(ctx->mediaengine->stepVideo(ctx->videoPixelMode)) {
 		int bufferSize = ctx->mediaengine->writeVideoImage(buffer, frameWidth, ctx->videoPixelMode);
-		gpu->InvalidateCache(buffer, bufferSize, GPU_INVALIDATE_SAFE);
+		gpu->NotifyVideoUpload(buffer, bufferSize, frameWidth, ctx->videoPixelMode);
 		ctx->avc.avcFrameStatus = 1;
 		ctx->videoFrameCount++;
 	} else {
@@ -1872,7 +1872,8 @@ static u32 sceMpegAvcCsc(u32 mpeg, u32 sourceAddr, u32 rangeAddr, int frameWidth
 	int height = Memory::Read_U32(rangeAddr + 12);
 	int destSize = ctx->mediaengine->writeVideoImageWithRange(destAddr, frameWidth, ctx->videoPixelMode, x, y, width, height);
 
-	gpu->InvalidateCache(destAddr, destSize, GPU_INVALIDATE_SAFE);
+	gpu->NotifyVideoUpload(destAddr, destSize, frameWidth, ctx->videoPixelMode);
+
 	// Do not use avcDecodeDelayMs 's value
 	// Will cause video 's screen dislocation in Bleach heat of soul 6
 	// https://github.com/hrydgard/ppsspp/issues/5535

Core/HLE/scePsmf.cpp

@@ -1540,7 +1540,7 @@ static int scePsmfPlayerGetVideoData(u32 psmfPlayer, u32 videoDataAddr)
 	int bufw = videoData->frameWidth == 0 ? 512 : videoData->frameWidth & ~1;
 	// Always write the video frame, even after the video has ended.
 	int displaybufSize = psmfplayer->mediaengine->writeVideoImage(videoData->displaybuf, bufw, videoPixelMode);
-	gpu->InvalidateCache(videoData->displaybuf, displaybufSize, GPU_INVALIDATE_SAFE);
+	gpu->NotifyVideoUpload(videoData->displaybuf, displaybufSize, bufw, videoPixelMode);
 	__PsmfUpdatePts(psmfplayer, videoData);
 
 	_PsmfPlayerFillRingbuffer(psmfplayer);

Core/HW/MediaEngine.cpp

@@ -799,7 +799,8 @@ int MediaEngine::writeVideoImageWithRange(u32 bufferPtr, int frameWidth, int vid
 		delete [] imgbuf;
 	}
 
-	return videoImageSize;
+	// Account for the y offset as well.
+	return videoImageSize + videoLineSize * ypos;
 #endif // USE_FFMPEG
 	return 0;
 }

GPU/Common/FramebufferCommon.cpp

@@ -487,6 +487,26 @@ VirtualFramebuffer *FramebufferManagerCommon::DoSetRenderFrameBuffer(const Frame
 	return vfb;
 }
 
+void FramebufferManagerCommon::NotifyVideoUpload(u32 addr, int size, int width, GEBufferFormat fmt) {
+	// Note: UpdateFromMemory() is still called later.
+	// This is a special case where we have extra information prior to the invalidation.
+
+	// TODO: Could possibly be an offset...
+	VirtualFramebuffer *vfb = GetVFBAt(addr);
+	if (vfb) {
+		if (vfb->format != fmt || vfb->drawnFormat != fmt) {
+			DEBUG_LOG(ME, "Changing format for %08x from %d to %d", addr, vfb->drawnFormat, fmt);
+			vfb->format = fmt;
+			vfb->drawnFormat = fmt;
+
+			// Let's count this as a "render".  This will also force us to use the correct format.
+			vfb->last_frame_render = gpuStats.numFlips;
+		}
+
+		// TODO: Check width?
+	}
+}
+
 void FramebufferManagerCommon::UpdateFromMemory(u32 addr, int size, bool safe) {
 	addr &= ~0x40000000;
 	// TODO: Could go through all FBOs, but probably not important?

GPU/Common/FramebufferCommon.h

@@ -160,6 +160,7 @@ public:
 	virtual void RebindFramebuffer() = 0;
 
 	bool NotifyFramebufferCopy(u32 src, u32 dest, int size, bool isMemset, u32 skipDrawReason);
+	void NotifyVideoUpload(u32 addr, int size, int width, GEBufferFormat fmt);
 	void UpdateFromMemory(u32 addr, int size, bool safe);
 	virtual bool NotifyStencilUpload(u32 addr, int size, bool skipZero = false) = 0;
 	// Returns true if it's sure this is a direct FBO->FBO transfer and it has already handle it.

GPU/Directx9/FramebufferDX9.cpp

@@ -862,6 +862,9 @@ namespace DX9 {
 			int w = std::min(pixels % vfb->fb_stride, (int)vfb->width);
 			int h = std::min((pixels + vfb->fb_stride - 1) / vfb->fb_stride, (int)vfb->height);
 
+			// We might still have a pending draw to the fb in question, flush if so.
+			FlushBeforeCopy();
+
 			// No need to download if we already have it.
 			if (!vfb->memoryUpdated && vfb->clutUpdatedBytes < loadBytes) {
 				// We intentionally don't call OptimizeDownloadRange() here - we don't want to over download.

GPU/Directx9/GPU_DX9.cpp

@@ -1952,6 +1952,13 @@ void DIRECTX9_GPU::InvalidateCacheInternal(u32 addr, int size, GPUInvalidationTy
 	}
 }
 
+void DIRECTX9_GPU::NotifyVideoUpload(u32 addr, int size, int width, int format) {
+	if (Memory::IsVRAMAddress(addr)) {
+		framebufferManager_.NotifyVideoUpload(addr, size, width, (GEBufferFormat)format);
+	}
+	InvalidateCache(addr, size, GPU_INVALIDATE_SAFE);
+}
+
 void DIRECTX9_GPU::PerformMemoryCopyInternal(u32 dest, u32 src, int size) {
 	if (!framebufferManager_.NotifyFramebufferCopy(src, dest, size, false, gstate_c.skipDrawReason)) {
 		// We use a little hack for Download/Upload using a VRAM mirror.

GPU/Directx9/GPU_DX9.h

@@ -48,6 +48,7 @@ public:
 	void BeginFrame() override;
 	void UpdateStats() override;
 	void InvalidateCache(u32 addr, int size, GPUInvalidationType type) override;
+	void NotifyVideoUpload(u32 addr, int size, int width, int format) override;
 	bool PerformMemoryCopy(u32 dest, u32 src, int size) override;
 	bool PerformMemorySet(u32 dest, u8 v, int size) override;
 	bool PerformMemoryDownload(u32 dest, int size) override;

GPU/GLES/Framebuffer.cpp

@@ -1173,6 +1173,9 @@ void FramebufferManager::DownloadFramebufferForClut(u32 fb_address, u32 loadByte
 		int w = std::min(pixels % vfb->fb_stride, (int)vfb->width);
 		int h = std::min((pixels + vfb->fb_stride - 1) / vfb->fb_stride, (int)vfb->height);
 
+		// We might still have a pending draw to the fb in question, flush if so.
+		FlushBeforeCopy();
+
 		// No need to download if we already have it.
 		if (!vfb->memoryUpdated && vfb->clutUpdatedBytes < loadBytes) {
 			// We intentionally don't call OptimizeDownloadRange() here - we don't want to over download.

GPU/GLES/GLES_GPU.cpp

@@ -2201,6 +2201,13 @@ void GLES_GPU::InvalidateCacheInternal(u32 addr, int size, GPUInvalidationType t
 	}
 }
 
+void GLES_GPU::NotifyVideoUpload(u32 addr, int size, int width, int format) {
+	if (Memory::IsVRAMAddress(addr)) {
+		framebufferManager_.NotifyVideoUpload(addr, size, width, (GEBufferFormat)format);
+	}
+	InvalidateCache(addr, size, GPU_INVALIDATE_SAFE);
+}
+
 void GLES_GPU::PerformMemoryCopyInternal(u32 dest, u32 src, int size) {
 	if (!framebufferManager_.NotifyFramebufferCopy(src, dest, size, false, gstate_c.skipDrawReason)) {
 		// We use a little hack for Download/Upload using a VRAM mirror.

GPU/GLES/GLES_GPU.h

@@ -52,6 +52,7 @@ public:
 	void BeginFrame() override;
 	void UpdateStats() override;
 	void InvalidateCache(u32 addr, int size, GPUInvalidationType type) override;
+	void NotifyVideoUpload(u32 addr, int size, int width, int format) override;
 	bool PerformMemoryCopy(u32 dest, u32 src, int size) override;
 	bool PerformMemorySet(u32 dest, u8 v, int size) override;
 	bool PerformMemoryDownload(u32 dest, int size) override;

GPU/GPUInterface.h

@@ -253,6 +253,7 @@ public:
 	// Invalidate any cached content sourced from the specified range.
 	// If size = -1, invalidate everything.
 	virtual void InvalidateCache(u32 addr, int size, GPUInvalidationType type) = 0;
+	virtual void NotifyVideoUpload(u32 addr, int size, int width, int format) = 0;
 	// Update either RAM from VRAM, or VRAM from RAM... or even VRAM from VRAM.
 	virtual bool PerformMemoryCopy(u32 dest, u32 src, int size) = 0;
 	virtual bool PerformMemorySet(u32 dest, u8 v, int size) = 0;

GPU/Null/NullGpu.cpp

@@ -687,6 +687,10 @@ void NullGPU::InvalidateCache(u32 addr, int size, GPUInvalidationType type) {
 	// Nothing to invalidate.
 }
 
+void NullGPU::NotifyVideoUpload(u32 addr, int size, int width, int format) {
+	// Nothing to do.
+}
+
 bool NullGPU::PerformMemoryCopy(u32 dest, u32 src, int size) {
 	// Nothing to update.
 	InvalidateCache(dest, size, GPU_INVALIDATE_HINT);

GPU/Null/NullGpu.h

@@ -34,6 +34,7 @@ public:
 	void CopyDisplayToOutput() override {}
 	void UpdateStats() override;
 	void InvalidateCache(u32 addr, int size, GPUInvalidationType type) override;
+	void NotifyVideoUpload(u32 addr, int size, int width, int format) override;
 	bool PerformMemoryCopy(u32 dest, u32 src, int size) override;
 	bool PerformMemorySet(u32 dest, u8 v, int size) override;
 	bool PerformMemoryDownload(u32 dest, int size) override;

GPU/Software/SoftGpu.cpp

@@ -873,6 +873,11 @@ void SoftGPU::InvalidateCache(u32 addr, int size, GPUInvalidationType type)
 	// Nothing to invalidate.
 }
 
+void SoftGPU::NotifyVideoUpload(u32 addr, int size, int width, int format)
+{
+	// Ignore.
+}
+
 bool SoftGPU::PerformMemoryCopy(u32 dest, u32 src, int size)
 {
 	// Nothing to update.

GPU/Software/SoftGpu.h

@@ -58,6 +58,7 @@ public:
 	void CopyDisplayToOutput() override;
 	void UpdateStats() override;
 	void InvalidateCache(u32 addr, int size, GPUInvalidationType type) override;
+	void NotifyVideoUpload(u32 addr, int size, int width, int format) override;
 	bool PerformMemoryCopy(u32 dest, u32 src, int size) override;
 	bool PerformMemorySet(u32 dest, u8 v, int size) override;
 	bool PerformMemoryDownload(u32 dest, int size) override;

ext/native/gfx/gl_common.h

@@ -55,7 +55,7 @@ extern PFNGLDRAWTEXTURENVPROC glDrawTextureNV;
 typedef void (EGLAPIENTRYP PFNGLBLITFRAMEBUFFERNVPROC) (
 	GLint srcX0, GLint srcY0, GLint srcX1, GLuint srcY1,
 	GLint dstX0, GLint dstY0, GLint dstX1, GLuint dstY1,
-	GLint mask, GLenum filter);
+	GLbitfield mask, GLenum filter);
 #endif
 extern PFNGLBLITFRAMEBUFFERNVPROC glBlitFramebufferNV;