From 7a28985454a4e2620dbad595bd8fd0468e4cb0b2 Mon Sep 17 00:00:00 2001
From: Samuliak <samuliak77@gmail.com>
Date: Thu, 29 Aug 2024 09:23:34 +0200
Subject: [PATCH] properly implement 0 stride vertex buffers

---
 .../HW/Latte/Renderer/Metal/LatteToMtl.cpp    | 51 +++++++++++++++++++
 src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.h |  2 +
 .../Renderer/Metal/MetalPipelineCache.cpp     | 33 +++++++-----
 3 files changed, 74 insertions(+), 12 deletions(-)

diff --git a/src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.cpp b/src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.cpp
index 51885e51..6f54272c 100644
--- a/src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.cpp
+++ b/src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.cpp
@@ -368,6 +368,57 @@ MTL::VertexFormat GetMtlVertexFormat(uint8 format)
 	}
 }
 
+uint32 GetMtlVertexFormatSize(uint8 format)
+{
+    switch (format)
+	{
+	case FMT_32_32_32_32_FLOAT:
+		return 16;
+	case FMT_32_32_32_FLOAT:
+		return 12;
+	case FMT_32_32_FLOAT:
+		return 8;
+	case FMT_32_FLOAT:
+		return 4;
+	case FMT_8_8_8_8:
+		return 4;
+	case FMT_8_8_8:
+		return 3;
+	case FMT_8_8:
+		return 2;
+	case FMT_8:
+		return 1;
+	case FMT_32_32_32_32:
+		return 16;
+	case FMT_32_32_32:
+		return 12;
+	case FMT_32_32:
+		return 8;
+	case FMT_32:
+		return 4;
+	case FMT_16_16_16_16:
+		return 8;
+	case FMT_16_16_16:
+		return 6;
+	case FMT_16_16:
+		return 4;
+	case FMT_16:
+		return 2;
+	case FMT_16_16_16_16_FLOAT:
+		return 8;
+	case FMT_16_16_16_FLOAT:
+		return 6;
+	case FMT_16_16_FLOAT:
+		return 4;
+	case FMT_16_FLOAT:
+		return 2;
+	case FMT_2_10_10_10:
+		return 4;
+	default:
+		return 0;
+	}
+}
+
 MTL::IndexType GetMtlIndexType(Renderer::INDEX_TYPE indexType)
 {
     switch (indexType)
diff --git a/src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.h b/src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.h
index cc0c5e02..2c805527 100644
--- a/src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.h
+++ b/src/Cafe/HW/Latte/Renderer/Metal/LatteToMtl.h
@@ -42,6 +42,8 @@ MTL::PrimitiveType GetMtlPrimitiveType(LattePrimitiveMode primitiveMode);
 
 MTL::VertexFormat GetMtlVertexFormat(uint8 format);
 
+uint32 GetMtlVertexFormatSize(uint8 format);
+
 MTL::IndexType GetMtlIndexType(Renderer::INDEX_TYPE indexType);
 
 MTL::BlendOperation GetMtlBlendOp(Latte::LATTE_CB_BLENDN_CONTROL::E_COMBINEFUNC combineFunc);
diff --git a/src/Cafe/HW/Latte/Renderer/Metal/MetalPipelineCache.cpp b/src/Cafe/HW/Latte/Renderer/Metal/MetalPipelineCache.cpp
index ade541ba..1d068426 100644
--- a/src/Cafe/HW/Latte/Renderer/Metal/MetalPipelineCache.cpp
+++ b/src/Cafe/HW/Latte/Renderer/Metal/MetalPipelineCache.cpp
@@ -298,6 +298,7 @@ MTL::RenderPipelineState* MetalPipelineCache::GetRenderPipelineState(const Latte
 	{
 		std::optional<LatteConst::VertexFetchType2> fetchType;
 
+		uint32 minBufferStride = 0;
 		for (sint32 j = 0; j < bufferGroup.attribCount; ++j)
 		{
 			auto& attr = bufferGroup.attrib[j];
@@ -311,6 +312,8 @@ MTL::RenderPipelineState* MetalPipelineCache::GetRenderPipelineState(const Latte
 			attribute->setBufferIndex(GET_MTL_VERTEX_BUFFER_INDEX(attr.attributeBufferIndex));
 			attribute->setFormat(GetMtlVertexFormat(attr.format));
 
+			minBufferStride = std::max(minBufferStride, attr.offset + GetMtlVertexFormatSize(attr.format));
+
 			if (fetchType.has_value())
 				cemu_assert_debug(fetchType == attr.fetchType);
 			else
@@ -327,24 +330,30 @@ MTL::RenderPipelineState* MetalPipelineCache::GetRenderPipelineState(const Latte
 		uint32 bufferStride = (lcr.GetRawView()[bufferBaseRegisterIndex + 2] >> 11) & 0xFFFF;
 		bufferStride = Align(bufferStride, 4);
 
-		// HACK
+		auto layout = vertexDescriptor->layouts()->object(GET_MTL_VERTEX_BUFFER_INDEX(bufferIndex));
 		if (bufferStride == 0)
 		{
-		    debug_printf("vertex buffer %u has a vertex stride of 0 bytes, using 4 bytes instead\n", bufferIndex);
-			bufferStride = 4;
-		}
+		    // Buffer stride cannot be zero, let's use the minimum stride
+			bufferStride = minBufferStride;
+		    debug_printf("vertex buffer %u has a vertex stride of 0 bytes, using %u bytes instead\n", bufferIndex, bufferStride);
 
-		auto layout = vertexDescriptor->layouts()->object(GET_MTL_VERTEX_BUFFER_INDEX(bufferIndex));
-		layout->setStride(bufferStride);
-		if (!fetchType.has_value() || fetchType == LatteConst::VertexFetchType2::VERTEX_DATA)
-			layout->setStepFunction(MTL::VertexStepFunctionPerVertex);
-		else if (fetchType == LatteConst::VertexFetchType2::INSTANCE_DATA)
-			layout->setStepFunction(MTL::VertexStepFunctionPerInstance);
+			// Additionally, constant vertex function must be used
+			layout->setStepFunction(MTL::VertexStepFunctionConstant);
+			layout->setStepRate(0);
+		}
 		else
 		{
-		    debug_printf("unimplemented vertex fetch type %u\n", (uint32)fetchType.value());
-			cemu_assert(false);
+    		if (!fetchType.has_value() || fetchType == LatteConst::VertexFetchType2::VERTEX_DATA)
+    			layout->setStepFunction(MTL::VertexStepFunctionPerVertex);
+    		else if (fetchType == LatteConst::VertexFetchType2::INSTANCE_DATA)
+    			layout->setStepFunction(MTL::VertexStepFunctionPerInstance);
+    		else
+    		{
+    		    debug_printf("unimplemented vertex fetch type %u\n", (uint32)fetchType.value());
+    			cemu_assert(false);
+    		}
 		}
+		layout->setStride(bufferStride);
 	}
 
 	auto mtlVertexShader = static_cast<RendererShaderMtl*>(vertexShader->shader);