Merge pull request #17565 from hrydgard/breakout-vcache-vulkan

Vulkan: Breakout the vertex cache logic from DoFlush()

Common/GPU/Vulkan/VulkanMemory.h

@@ -135,6 +135,8 @@ class VulkanPushPool : public GPUMemoryManager {
 		return blocks_[curBlockIndex_].writePtr;
 	}
 
+	// NOTE: If you can avoid this by writing the data directly into memory returned from Allocate,
+	// do so. Savings from avoiding memcpy can be significant.
 	VkDeviceSize Push(const void *data, VkDeviceSize numBytes, int alignment, VkBuffer *vkbuf) {
 		uint32_t bindOffset;
 		uint8_t *ptr = Allocate(numBytes, alignment, vkbuf, &bindOffset);

GPU/Common/DrawEngineCommon.cpp

@@ -74,17 +74,18 @@ VertexDecoder *DrawEngineCommon::GetVertexDecoder(u32 vtype) {
 
 int DrawEngineCommon::ComputeNumVertsToDecode() const {
 	int vertsToDecode = 0;
+	int numDrawCalls = numDrawCalls_;
 	if (drawCalls_[0].indexType == GE_VTYPE_IDX_NONE >> GE_VTYPE_IDX_SHIFT) {
-		for (int i = 0; i < numDrawCalls_; i++) {
+		for (int i = 0; i < numDrawCalls; i++) {
 			const DeferredDrawCall &dc = drawCalls_[i];
 			vertsToDecode += dc.vertexCount;
 		}
 	} else {
 		// TODO: Share this computation with DecodeVertsStep?
-		for (int i = 0; i < numDrawCalls_; i++) {
+		for (int i = 0; i < numDrawCalls; i++) {
 			const DeferredDrawCall &dc = drawCalls_[i];
 			int lastMatch = i;
-			const int total = numDrawCalls_;
+			const int total = numDrawCalls;
 			int indexLowerBound = dc.indexLowerBound;
 			int indexUpperBound = dc.indexUpperBound;
 			for (int j = i + 1; j < total; ++j) {
@@ -642,7 +643,7 @@ void DrawEngineCommon::DecodeVertsStep(u8 *dest, int &i, int &decodedVerts, cons
 		for (int j = i + 1; j < total; ++j) {
 			if (drawCalls_[j].verts != dc.verts)
 				break;
-
+			// TODO: What if UV scale/offset changes between drawcalls here?
 			indexLowerBound = std::min(indexLowerBound, (int)drawCalls_[j].indexLowerBound);
 			indexUpperBound = std::max(indexUpperBound, (int)drawCalls_[j].indexUpperBound);
 			lastMatch = j;
@@ -779,16 +780,6 @@ uint64_t DrawEngineCommon::ComputeHash() {
 	return fullhash;
 }
 
-// Cheap bit scrambler from https://nullprogram.com/blog/2018/07/31/
-inline uint32_t lowbias32_r(uint32_t x) {
-	x ^= x >> 16;
-	x *= 0x43021123U;
-	x ^= x >> 15 ^ x >> 30;
-	x *= 0x1d69e2a5U;
-	x ^= x >> 16;
-	return x;
-}
-
 // vertTypeID is the vertex type but with the UVGen mode smashed into the top bits.
 void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimitiveType prim, int vertexCount, u32 vertTypeID, int cullMode, int *bytesRead) {
 	if (!indexGen.PrimCompatible(prevPrim_, prim) || numDrawCalls_ >= MAX_DEFERRED_DRAW_CALLS || vertexCountInDrawCalls_ + vertexCount > VERTEX_BUFFER_MAX) {
@@ -818,15 +809,6 @@ void DrawEngineCommon::SubmitPrim(const void *verts, const void *inds, GEPrimiti
 	if ((vertexCount < 2 && prim > 0) || (vertexCount < 3 && prim > GE_PRIM_LINE_STRIP && prim != GE_PRIM_RECTANGLES))
 		return;
 
-	if (g_Config.bVertexCache) {
-		u32 dhash = dcid_;
-		dhash = __rotl(dhash ^ (u32)(uintptr_t)verts, 13);
-		dhash = __rotl(dhash ^ (u32)(uintptr_t)inds, 19);
-		dhash = __rotl(dhash ^ (u32)vertTypeID, 7);
-		dhash = __rotl(dhash ^ (u32)vertexCount, 11);
-		dcid_ = lowbias32_r(dhash ^ (u32)prim);
-	}
-
 	DeferredDrawCall &dc = drawCalls_[numDrawCalls_];
 	dc.verts = verts;
 	dc.inds = inds;
@@ -870,6 +852,29 @@ bool DrawEngineCommon::CanUseHardwareTessellation(GEPatchPrimType prim) {
 	return false;
 }
 
+// Cheap bit scrambler from https://nullprogram.com/blog/2018/07/31/
+inline uint32_t lowbias32_r(uint32_t x) {
+	x ^= x >> 16;
+	x *= 0x43021123U;
+	x ^= x >> 15 ^ x >> 30;
+	x *= 0x1d69e2a5U;
+	x ^= x >> 16;
+	return x;
+}
+
+uint32_t DrawEngineCommon::ComputeDrawcallsHash() const {
+	uint32_t dcid = 0;
+	for (int i = 0; i < numDrawCalls_; i++) {
+		u32 dhash = dcid;
+		dhash = __rotl(dhash ^ (u32)(uintptr_t)drawCalls_[i].verts, 13);
+		dhash = __rotl(dhash ^ (u32)(uintptr_t)drawCalls_[i].inds, 19);
+		dhash = __rotl(dhash ^ (u32)drawCalls_[i].indexType, 7);
+		dhash = __rotl(dhash ^ (u32)drawCalls_[i].vertexCount, 11);
+		dcid = lowbias32_r(dhash ^ (u32)drawCalls_[i].prim);
+	}
+	return dcid;
+}
+
 void TessellationDataTransfer::CopyControlPoints(float *pos, float *tex, float *col, int posStride, int texStride, int colStride, const SimpleVertex *const *points, int size, u32 vertType) {
 	bool hasColor = (vertType & GE_VTYPE_COL_MASK) != 0;
 	bool hasTexCoord = (vertType & GE_VTYPE_TC_MASK) != 0;

GPU/Common/DrawEngineCommon.h

@@ -175,6 +175,8 @@ class DrawEngineCommon {
 		}
 	}
 
+	uint32_t ComputeDrawcallsHash() const;
+
 	bool useHWTransform_ = false;
 	bool useHWTessellation_ = false;
 	// Used to prevent unnecessary flushing in softgpu.
@@ -218,7 +220,6 @@ class DrawEngineCommon {
 
 	int decimationCounter_ = 0;
 	int decodeCounter_ = 0;
-	u32 dcid_ = 0;
 
 	// Vertex collector state
 	IndexGenerator indexGen;

GPU/D3D11/DrawEngineD3D11.cpp

@@ -363,12 +363,13 @@ void DrawEngineD3D11::DoFlush() {
 			useCache = false;
 
 		if (useCache) {
-			u32 id = dcid_ ^ gstate.getUVGenMode();  // This can have an effect on which UV decoder we need to use! And hence what the decoded data will look like. See #9263
+			// getUVGenMode can have an effect on which UV decoder we need to use! And hence what the decoded data will look like. See #9263
+			u32 dcid = (u32)XXH3_64bits(&drawCalls_, sizeof(DeferredDrawCall) * numDrawCalls_) ^ gstate.getUVGenMode();
 
-			VertexArrayInfoD3D11 *vai = vai_.Get(id);
+			VertexArrayInfoD3D11 *vai = vai_.Get(dcid);
 			if (!vai) {
 				vai = new VertexArrayInfoD3D11();
-				vai_.Insert(id, vai);
+				vai_.Insert(dcid, vai);
 			}
 
 			switch (vai->status) {
@@ -724,7 +725,6 @@ void DrawEngineD3D11::DoFlush() {
 	numDrawCalls_ = 0;
 	vertexCountInDrawCalls_ = 0;
 	decodeCounter_ = 0;
-	dcid_ = 0;
 	gstate_c.vertexFullAlpha = true;
 	framebufferManager_->SetColorUpdated(gstate_c.skipDrawReason);
 

GPU/Directx9/DrawEngineDX9.cpp

@@ -345,11 +345,12 @@ void DrawEngineDX9::DoFlush() {
 			useCache = false;
 
 		if (useCache) {
-			u32 id = dcid_ ^ gstate.getUVGenMode();  // This can have an effect on which UV decoder we need to use! And hence what the decoded data will look like. See #9263
-			VertexArrayInfoDX9 *vai = vai_.Get(id);
+			// getUVGenMode can have an effect on which UV decoder we need to use! And hence what the decoded data will look like. See #9263
+			u32 dcid = (u32)XXH3_64bits(&drawCalls_, sizeof(DeferredDrawCall) * numDrawCalls_) ^ gstate.getUVGenMode();
+			VertexArrayInfoDX9 *vai = vai_.Get(dcid);
 			if (!vai) {
 				vai = new VertexArrayInfoDX9();
-				vai_.Insert(id, vai);
+				vai_.Insert(dcid, vai);
 			}
 
 			switch (vai->status) {
@@ -666,7 +667,6 @@ void DrawEngineDX9::DoFlush() {
 	numDrawCalls_ = 0;
 	vertexCountInDrawCalls_ = 0;
 	decodeCounter_ = 0;
-	dcid_ = 0;
 	gstate_c.vertexFullAlpha = true;
 	framebufferManager_->SetColorUpdated(gstate_c.skipDrawReason);
 

GPU/GLES/DrawEngineGLES.cpp

@@ -248,7 +248,6 @@ void DrawEngineGLES::DoFlush() {
 		numDrawCalls_ = 0;
 		vertexCountInDrawCalls_ = 0;
 		decodeCounter_ = 0;
-		dcid_ = 0;
 		return;
 	}
 
@@ -483,7 +482,6 @@ void DrawEngineGLES::DoFlush() {
 	numDrawCalls_ = 0;
 	vertexCountInDrawCalls_ = 0;
 	decodeCounter_ = 0;
-	dcid_ = 0;
 	gstate_c.vertexFullAlpha = true;
 	framebufferManager_->SetColorUpdated(gstate_c.skipDrawReason);