Merge pull request #16734 from hrydgard/remove-nonindexed-optimization-vk

Vulkan: Don't use non-indexed draws for pure tristrips and fans (only PowerVR for now)

Common/GPU/Vulkan/thin3d_vulkan.cpp

@@ -861,6 +861,11 @@ VKContext::VKContext(VulkanContext *vulkan)
 		break;
 	}
 
+	if (caps_.vendor == GPUVendor::VENDOR_IMGTEC) {
+		// Enable some things that cut down pipeline counts but may have other costs.
+		caps_.verySlowShaderCompiler = true;
+	}
+
 	// Hide D3D9 when we know it likely won't work well.
 #if PPSSPP_PLATFORM(WINDOWS)
 	caps_.supportsD3D9 = true;

Common/GPU/thin3d.h

@@ -577,6 +577,8 @@ struct DeviceCaps {
 	bool isTilingGPU;  // This means that it benefits from correct store-ops, msaa without backing memory, etc.
 	bool sampleRateShadingSupported;
 
+	bool verySlowShaderCompiler;
+
 	// From the other backends, we can detect if D3D9 support is known bad (like on Xe) and disable it.
 	bool supportsD3D9;
 

GPU/Common/IndexGenerator.h

@@ -47,6 +47,15 @@ public:
 	}
 
 	GEPrimitiveType Prim() const { return prim_; }
+	GEPrimitiveType GeneralPrim() const {
+		switch (prim_) {
+		case GE_PRIM_LINE_STRIP: return GE_PRIM_LINES; break;
+		case GE_PRIM_TRIANGLE_STRIP:
+		case GE_PRIM_TRIANGLE_FAN: return GE_PRIM_TRIANGLES; break;
+		default:
+			return prim_;
+		}
+	}
 
 	void AddPrim(int prim, int vertexCount, bool clockwise);
 	void TranslatePrim(int prim, int numInds, const u8 *inds, int indexOffset, bool clockwise);

GPU/GLES/ShaderManagerGLES.cpp

@@ -956,7 +956,7 @@ enum class CacheDetectFlags {
 };
 
 #define CACHE_HEADER_MAGIC 0x83277592
-#define CACHE_VERSION 25
+#define CACHE_VERSION 26
 
 struct CacheHeader {
 	uint32_t magic;

GPU/Vulkan/DrawEngineVulkan.cpp

@@ -582,6 +582,10 @@ void DrawEngineVulkan::DoFlush() {
 	uint32_t ibOffset;
 	uint32_t vbOffset;
 	
+	// The optimization to avoid indexing isn't really worth it on Vulkan since it means creating more pipelines.
+	// This could be avoided with the new dynamic state extensions, but not available enough on mobile.
+	const bool forceIndexed = draw_->GetDeviceCaps().verySlowShaderCompiler;
+
 	if (useHWTransform) {
 		int vertexCount = 0;
 		bool useElements = true;
@@ -668,13 +672,19 @@ void DrawEngineVulkan::DoFlush() {
 					DecodeVertsToPushBuffer(vertexCache_, &vai->vbOffset, &vai->vb);
 					_dbg_assert_msg_(gstate_c.vertBounds.minV >= gstate_c.vertBounds.maxV, "Should not have checked UVs when caching.");
 					vai->numVerts = indexGen.VertexCount();
-					vai->prim = indexGen.Prim();
 					vai->maxIndex = indexGen.MaxIndex();
 					vai->flags = gstate_c.vertexFullAlpha ? VAIVULKAN_FLAG_VERTEXFULLALPHA : 0;
-					useElements = !indexGen.SeenOnlyPurePrims();
-					if (!useElements && indexGen.PureCount()) {
-						vai->numVerts = indexGen.PureCount();
+					if (forceIndexed) {
+						vai->prim = indexGen.GeneralPrim();
+						useElements = true;
+					} else {
+						vai->prim = indexGen.Prim();
+						useElements = !indexGen.SeenOnlyPurePrims();
+						if (!useElements && indexGen.PureCount()) {
+							vai->numVerts = indexGen.PureCount();
+						}
 					}
+
 					if (useElements) {
 						u32 size = sizeof(uint16_t) * indexGen.VertexCount();
 						void *dest = vertexCache_->Push(size, &vai->ibOffset, &vai->ib);
@@ -743,12 +753,18 @@ void DrawEngineVulkan::DoFlush() {
 
 	rotateVBO:
 			gpuStats.numUncachedVertsDrawn += indexGen.VertexCount();
-			useElements = !indexGen.SeenOnlyPurePrims();
+
 			vertexCount = indexGen.VertexCount();
-			if (!useElements && indexGen.PureCount()) {
-				vertexCount = indexGen.PureCount();
+			if (forceIndexed) {
+				useElements = true;
+				prim = indexGen.GeneralPrim();
+			} else {
+				useElements = !indexGen.SeenOnlyPurePrims();
+				if (!useElements && indexGen.PureCount()) {
+					vertexCount = indexGen.PureCount();
+				}
+				prim = indexGen.Prim();
 			}
-			prim = indexGen.Prim();
 		}
 
 		bool hasColor = (lastVType_ & GE_VTYPE_COL_MASK) != GE_VTYPE_COL_NONE;

GPU/Vulkan/ShaderManagerVulkan.cpp

@@ -511,7 +511,7 @@ enum class VulkanCacheDetectFlags {
 };
 
 #define CACHE_HEADER_MAGIC 0xff51f420 
-#define CACHE_VERSION 39
+#define CACHE_VERSION 40
 
 struct VulkanCacheHeader {
 	uint32_t magic;