diff --git a/DirectXTex/d3dx12.h b/DirectXTex/d3dx12.h
index 3970b09..ed0eec4 100644
--- a/DirectXTex/d3dx12.h
+++ b/DirectXTex/d3dx12.h
@@ -2011,6 +2011,29 @@ inline void MemcpySubresource(
     }
 }
 
+//------------------------------------------------------------------------------------------------
+// Row-by-row memcpy
+inline void MemcpySubresource(
+    _In_ const D3D12_MEMCPY_DEST* pDest,
+    _In_ const void* pResourceData,
+    _In_ const D3D12_SUBRESOURCE_INFO* pSrc,
+    SIZE_T RowSizeInBytes,
+    UINT NumRows,
+    UINT NumSlices) noexcept
+{
+    for (UINT z = 0; z < NumSlices; ++z)
+    {
+        auto pDestSlice = static_cast<BYTE*>(pDest->pData) + pDest->SlicePitch * z;
+        auto pSrcSlice = (static_cast<const BYTE*>(pResourceData) + pSrc->Offset) + pSrc->DepthPitch * LONG_PTR(z);
+        for (UINT y = 0; y < NumRows; ++y)
+        {
+            memcpy(pDestSlice + pDest->RowPitch * y,
+                pSrcSlice + pSrc->RowPitch * LONG_PTR(y),
+                RowSizeInBytes);
+        }
+    }
+}
+
 //------------------------------------------------------------------------------------------------
 // Returns required size of a buffer to be used for data upload
 inline UINT64 GetRequiredIntermediateSize(
@@ -2087,6 +2110,65 @@ inline UINT64 UpdateSubresources(
     return RequiredSize;
 }
 
+//------------------------------------------------------------------------------------------------
+// All arrays must be populated (e.g. by calling GetCopyableFootprints)
+inline UINT64 UpdateSubresources(
+    _In_ ID3D12GraphicsCommandList* pCmdList,
+    _In_ ID3D12Resource* pDestinationResource,
+    _In_ ID3D12Resource* pIntermediate,
+    _In_range_(0,D3D12_REQ_SUBRESOURCES) UINT FirstSubresource,
+    _In_range_(0,D3D12_REQ_SUBRESOURCES-FirstSubresource) UINT NumSubresources,
+    UINT64 RequiredSize,
+    _In_reads_(NumSubresources) const D3D12_PLACED_SUBRESOURCE_FOOTPRINT* pLayouts,
+    _In_reads_(NumSubresources) const UINT* pNumRows,
+    _In_reads_(NumSubresources) const UINT64* pRowSizesInBytes,
+    _In_ const void* pResourceData,
+    _In_reads_(NumSubresources) const D3D12_SUBRESOURCE_INFO* pSrcData) noexcept
+{
+    // Minor validation
+    auto IntermediateDesc = pIntermediate->GetDesc();
+    auto DestinationDesc = pDestinationResource->GetDesc();
+    if (IntermediateDesc.Dimension != D3D12_RESOURCE_DIMENSION_BUFFER ||
+        IntermediateDesc.Width < RequiredSize + pLayouts[0].Offset ||
+        RequiredSize > SIZE_T(-1) ||
+        (DestinationDesc.Dimension == D3D12_RESOURCE_DIMENSION_BUFFER &&
+            (FirstSubresource != 0 || NumSubresources != 1)))
+    {
+        return 0;
+    }
+
+    BYTE* pData;
+    HRESULT hr = pIntermediate->Map(0, nullptr, reinterpret_cast<void**>(&pData));
+    if (FAILED(hr))
+    {
+        return 0;
+    }
+
+    for (UINT i = 0; i < NumSubresources; ++i)
+    {
+        if (pRowSizesInBytes[i] > SIZE_T(-1)) return 0;
+        D3D12_MEMCPY_DEST DestData = { pData + pLayouts[i].Offset, pLayouts[i].Footprint.RowPitch, SIZE_T(pLayouts[i].Footprint.RowPitch) * SIZE_T(pNumRows[i]) };
+        MemcpySubresource(&DestData, pResourceData, &pSrcData[i], static_cast<SIZE_T>(pRowSizesInBytes[i]), pNumRows[i], pLayouts[i].Footprint.Depth);
+    }
+    pIntermediate->Unmap(0, nullptr);
+
+    if (DestinationDesc.Dimension == D3D12_RESOURCE_DIMENSION_BUFFER)
+    {
+        pCmdList->CopyBufferRegion(
+            pDestinationResource, 0, pIntermediate, pLayouts[0].Offset, pLayouts[0].Footprint.Width);
+    }
+    else
+    {
+        for (UINT i = 0; i < NumSubresources; ++i)
+        {
+            CD3DX12_TEXTURE_COPY_LOCATION Dst(pDestinationResource, i + FirstSubresource);
+            CD3DX12_TEXTURE_COPY_LOCATION Src(pIntermediate, pLayouts[i]);
+            pCmdList->CopyTextureRegion(&Dst, 0, 0, 0, &Src, nullptr);
+        }
+    }
+    return RequiredSize;
+}
+
 //------------------------------------------------------------------------------------------------
 // Heap-allocating UpdateSubresources implementation
 inline UINT64 UpdateSubresources(
@@ -2099,7 +2181,7 @@ inline UINT64 UpdateSubresources(
     _In_reads_(NumSubresources) const D3D12_SUBRESOURCE_DATA* pSrcData) noexcept
 {
     UINT64 RequiredSize = 0;
-    UINT64 MemToAlloc = static_cast<UINT64>(sizeof(D3D12_PLACED_SUBRESOURCE_FOOTPRINT) + sizeof(UINT) + sizeof(UINT64)) * NumSubresources;
+    auto MemToAlloc = static_cast<UINT64>(sizeof(D3D12_PLACED_SUBRESOURCE_FOOTPRINT) + sizeof(UINT) + sizeof(UINT64)) * NumSubresources;
     if (MemToAlloc > SIZE_MAX)
     {
        return 0;
@@ -2110,8 +2192,8 @@ inline UINT64 UpdateSubresources(
        return 0;
     }
     auto pLayouts = static_cast<D3D12_PLACED_SUBRESOURCE_FOOTPRINT*>(pMem);
-    UINT64* pRowSizesInBytes = reinterpret_cast<UINT64*>(pLayouts + NumSubresources);
-    UINT* pNumRows = reinterpret_cast<UINT*>(pRowSizesInBytes + NumSubresources);
+    auto pRowSizesInBytes = reinterpret_cast<UINT64*>(pLayouts + NumSubresources);
+    auto pNumRows = reinterpret_cast<UINT*>(pRowSizesInBytes + NumSubresources);
 
     auto Desc = pDestinationResource->GetDesc();
     ID3D12Device* pDevice = nullptr;
@@ -2124,6 +2206,44 @@ inline UINT64 UpdateSubresources(
     return Result;
 }
 
+//------------------------------------------------------------------------------------------------
+// Heap-allocating UpdateSubresources implementation
+inline UINT64 UpdateSubresources(
+    _In_ ID3D12GraphicsCommandList* pCmdList,
+    _In_ ID3D12Resource* pDestinationResource,
+    _In_ ID3D12Resource* pIntermediate,
+    UINT64 IntermediateOffset,
+    _In_range_(0,D3D12_REQ_SUBRESOURCES) UINT FirstSubresource,
+    _In_range_(0,D3D12_REQ_SUBRESOURCES-FirstSubresource) UINT NumSubresources,
+    _In_ const void* pResourceData,
+    _In_reads_(NumSubresources) D3D12_SUBRESOURCE_INFO* pSrcData) noexcept
+{
+    UINT64 RequiredSize = 0;
+    auto MemToAlloc = static_cast<UINT64>(sizeof(D3D12_PLACED_SUBRESOURCE_FOOTPRINT) + sizeof(UINT) + sizeof(UINT64)) * NumSubresources;
+    if (MemToAlloc > SIZE_MAX)
+    {
+        return 0;
+    }
+    void* pMem = HeapAlloc(GetProcessHeap(), 0, static_cast<SIZE_T>(MemToAlloc));
+    if (pMem == nullptr)
+    {
+        return 0;
+    }
+    auto pLayouts = reinterpret_cast<D3D12_PLACED_SUBRESOURCE_FOOTPRINT*>(pMem);
+    auto pRowSizesInBytes = reinterpret_cast<UINT64*>(pLayouts + NumSubresources);
+    auto pNumRows = reinterpret_cast<UINT*>(pRowSizesInBytes + NumSubresources);
+
+    auto Desc = pDestinationResource->GetDesc();
+    ID3D12Device* pDevice = nullptr;
+    pDestinationResource->GetDevice(IID_ID3D12Device, reinterpret_cast<void**>(&pDevice));
+    pDevice->GetCopyableFootprints(&Desc, FirstSubresource, NumSubresources, IntermediateOffset, pLayouts, pNumRows, pRowSizesInBytes, &RequiredSize);
+    pDevice->Release();
+
+    UINT64 Result = UpdateSubresources(pCmdList, pDestinationResource, pIntermediate, FirstSubresource, NumSubresources, RequiredSize, pLayouts, pNumRows, pRowSizesInBytes, pResourceData, pSrcData);
+    HeapFree(GetProcessHeap(), 0, pMem);
+    return Result;
+}
+
 //------------------------------------------------------------------------------------------------
 // Stack-allocating UpdateSubresources implementation
 template <UINT MaxSubresources>
@@ -2132,8 +2252,8 @@ inline UINT64 UpdateSubresources(
     _In_ ID3D12Resource* pDestinationResource,
     _In_ ID3D12Resource* pIntermediate,
     UINT64 IntermediateOffset,
-    _In_range_(0, MaxSubresources) UINT FirstSubresource,
-    _In_range_(1, MaxSubresources - FirstSubresource) UINT NumSubresources,
+    _In_range_(0,MaxSubresources) UINT FirstSubresource,
+    _In_range_(1,MaxSubresources-FirstSubresource) UINT NumSubresources,
     _In_reads_(NumSubresources) const D3D12_SUBRESOURCE_DATA* pSrcData) noexcept
 {
     UINT64 RequiredSize = 0;
@@ -2150,6 +2270,33 @@ inline UINT64 UpdateSubresources(
     return UpdateSubresources(pCmdList, pDestinationResource, pIntermediate, FirstSubresource, NumSubresources, RequiredSize, Layouts, NumRows, RowSizesInBytes, pSrcData);
 }
 
+//------------------------------------------------------------------------------------------------
+// Stack-allocating UpdateSubresources implementation
+template <UINT MaxSubresources>
+inline UINT64 UpdateSubresources(
+    _In_ ID3D12GraphicsCommandList* pCmdList,
+    _In_ ID3D12Resource* pDestinationResource,
+    _In_ ID3D12Resource* pIntermediate,
+    UINT64 IntermediateOffset,
+    _In_range_(0,MaxSubresources) UINT FirstSubresource,
+    _In_range_(1,MaxSubresources-FirstSubresource) UINT NumSubresources,
+    _In_ const void* pResourceData,
+    _In_reads_(NumSubresources) D3D12_SUBRESOURCE_INFO* pSrcData) noexcept
+{
+    UINT64 RequiredSize = 0;
+    D3D12_PLACED_SUBRESOURCE_FOOTPRINT Layouts[MaxSubresources];
+    UINT NumRows[MaxSubresources];
+    UINT64 RowSizesInBytes[MaxSubresources];
+
+    auto Desc = pDestinationResource->GetDesc();
+    ID3D12Device* pDevice = nullptr;
+    pDestinationResource->GetDevice(IID_ID3D12Device, reinterpret_cast<void**>(&pDevice));
+    pDevice->GetCopyableFootprints(&Desc, FirstSubresource, NumSubresources, IntermediateOffset, Layouts, NumRows, RowSizesInBytes, &RequiredSize);
+    pDevice->Release();
+
+    return UpdateSubresources(pCmdList, pDestinationResource, pIntermediate, FirstSubresource, NumSubresources, RequiredSize, Layouts, NumRows, RowSizesInBytes, pResourceData, pSrcData);
+}
+
 //------------------------------------------------------------------------------------------------
 inline constexpr bool D3D12IsLayoutOpaque( D3D12_TEXTURE_LAYOUT Layout ) noexcept
 { return Layout == D3D12_TEXTURE_LAYOUT_UNKNOWN || Layout == D3D12_TEXTURE_LAYOUT_64KB_UNDEFINED_SWIZZLE; }