diff --git a/src/d3d9/d3d9_device.cpp b/src/d3d9/d3d9_device.cpp
index 7f385d373..f7723723d 100644
--- a/src/d3d9/d3d9_device.cpp
+++ b/src/d3d9/d3d9_device.cpp
@@ -2399,10 +2399,11 @@ namespace dxvk {
 
     auto drawInfo = GenerateDrawInfo(PrimitiveType, PrimitiveCount, 0);
 
-    const uint32_t upSize = drawInfo.vertexCount * VertexStreamZeroStride;
+    const uint32_t dataSize = GetUPDataSize(drawInfo.vertexCount, VertexStreamZeroStride);
+    const uint32_t bufferSize = GetUPBufferSize(drawInfo.vertexCount, VertexStreamZeroStride);
 
-    auto upSlice = AllocTempBuffer<true>(upSize);
-    std::memcpy(upSlice.mapPtr, pVertexStreamZeroData, upSize);
+    auto upSlice = AllocTempBuffer<true>(bufferSize);
+    FillUPVertexBuffer(upSlice.mapPtr, pVertexStreamZeroData, dataSize, bufferSize);
 
     EmitCs([this,
       cBufferSlice  = std::move(upSlice.slice),
@@ -2445,21 +2446,21 @@ namespace dxvk {
 
     auto drawInfo = GenerateDrawInfo(PrimitiveType, PrimitiveCount, 0);
 
-    const uint32_t vertexSize  = (MinVertexIndex + NumVertices) * VertexStreamZeroStride;
+    const uint32_t vertexDataSize = GetUPDataSize(MinVertexIndex + NumVertices, VertexStreamZeroStride);
+    const uint32_t vertexBufferSize = GetUPBufferSize(MinVertexIndex + NumVertices, VertexStreamZeroStride);
 
     const uint32_t indexSize = IndexDataFormat == D3DFMT_INDEX16 ? 2 : 4;
     const uint32_t indicesSize = drawInfo.vertexCount * indexSize;
 
-    const uint32_t upSize = vertexSize + indicesSize;
+    const uint32_t upSize = vertexBufferSize + indicesSize;
 
     auto upSlice = AllocTempBuffer<true>(upSize);
     uint8_t* data = reinterpret_cast<uint8_t*>(upSlice.mapPtr);
-
-    std::memcpy(data, pVertexStreamZeroData, vertexSize);
-    std::memcpy(data + vertexSize, pIndexData, indicesSize);
+    FillUPVertexBuffer(data, pVertexStreamZeroData, vertexDataSize, vertexBufferSize);
+    std::memcpy(data + vertexBufferSize, pIndexData, indicesSize);
 
     EmitCs([this,
-      cVertexSize   = vertexSize,
+      cVertexSize   = vertexBufferSize,
       cBufferSlice  = std::move(upSlice.slice),
       cPrimType     = PrimitiveType,
       cPrimCount    = PrimitiveCount,
diff --git a/src/d3d9/d3d9_device.h b/src/d3d9/d3d9_device.h
index d79f0955e..0a318a2a0 100644
--- a/src/d3d9/d3d9_device.h
+++ b/src/d3d9/d3d9_device.h
@@ -1085,6 +1085,31 @@ namespace dxvk {
       const DWORD*                pShaderBytecode,
       const DxsoModuleInfo*       pModuleInfo);
 
+    inline uint32_t GetUPDataSize(uint32_t vertexCount, uint32_t stride) {
+      return vertexCount * stride;
+    }
+
+    inline uint32_t GetUPBufferSize(uint32_t vertexCount, uint32_t stride) {
+      return (vertexCount - 1) * stride + m_state.vertexDecl->GetSize();
+    }
+
+    inline void FillUPVertexBuffer(void* buffer, const void* userData, uint32_t dataSize, uint32_t bufferSize) {
+      uint8_t* data = reinterpret_cast<uint8_t*>(buffer);
+      // Don't copy excess data if we don't end up needing it.
+      dataSize = std::min(dataSize, bufferSize);
+      std::memcpy(data, userData, dataSize);
+      // Pad out with 0 to make buffer range checks happy
+      // Some games have components out of range in the vertex decl
+      // that they don't read from the shader.
+      // My tests show that these are read back as 0 always if out of range of
+      // the dataSize.
+      //
+      // So... make the actual buffer the range that satisfies the range of the vertex
+      // declaration and pad with 0s outside of it.
+      if (dataSize < bufferSize)
+        std::memset(data + dataSize, 0, bufferSize - dataSize);
+    }
+
     // So we don't do OOB.
     template <DxsoProgramType  ProgramType,
               D3D9ConstantType ConstantType>