diff --git a/rocclr/runtime/device/gpu/gpublit.cpp b/rocclr/runtime/device/gpu/gpublit.cpp
index 4d638e5e31..32e1c9e5a7 100644
--- a/rocclr/runtime/device/gpu/gpublit.cpp
+++ b/rocclr/runtime/device/gpu/gpublit.cpp
@@ -613,24 +613,36 @@ DmaBlitManager::copyBufferRect(
         size_t  srcOffset;
         size_t  dstOffset;
 
+        uint bytesPerElement = 16;
+        bool optimalElementSize = false;
         bool subWindowRectCopy = dev().settings().rectLinearDMA_;
-        // 19 bit limit in HW in SI and 16 bit limit in CI+(we adjust the ElementSize to 4bytes but the packet still has 14bits)
-        size_t pitchLimit = dev().settings().ciPlus_ ? 0xFFFF : 0x7FFFF;
 
         srcOffset   = srcRect.offset(0, 0, 0);
         dstOffset   = dstRect.offset(0, 0, 0);
 
+        while (bytesPerElement >= 1) {
+            if (((srcOffset % 4) == 0) &&
+                ((dstOffset % 4) == 0) &&
+                ((size[0] % bytesPerElement) == 0) &&
+                ((srcRect.rowPitch_ % bytesPerElement) == 0) &&
+                ((srcRect.slicePitch_ % bytesPerElement) == 0) &&
+                ((dstRect.rowPitch_ % bytesPerElement) == 0) &&
+                ((dstRect.slicePitch_ % bytesPerElement) == 0)) {
+                    optimalElementSize = true;
+                    break;
+            }
+            bytesPerElement = bytesPerElement >> 1;
+        }
+
+        // 19 bit limit in HW in SI and 16 bit limit in CI+(we adjust the ElementSize to 4bytes but the packet still has 14bits)
+        size_t pitchLimit = dev().settings().ciPlus_ ? (0x3FFF * bytesPerElement) | 0xF : 0x7FFFF;
+        size_t sizeLimit = dev().settings().ciPlus_ ? (0x3FFF * bytesPerElement) | 0xF : 0x3FFF;
+
         if (subWindowRectCopy &&
-            (((srcOffset % 4) != 0) ||
-            ((dstOffset % 4) != 0) ||
-            ((size[0] % 4) != 0) ||
-            ((srcRect.rowPitch_ % 4) != 0) ||
-            ((srcRect.slicePitch_ % 4) != 0) ||
-            ((dstRect.rowPitch_ % 4) != 0) ||
-            ((dstRect.slicePitch_ % 4) != 0) ||
+            (!optimalElementSize ||
             (srcRect.rowPitch_ > pitchLimit) ||
             (dstRect.rowPitch_ > pitchLimit) ||
-            (size[0] > 0x3fff) ||   // 14 bits limit in HW
+            (size[0] > sizeLimit) ||    // See above
             (size[1] > 0x3fff) ||   // 14 bits limit in HW
             (size[2] > 0x7ff))) {    // 11 bits limit in HW
             // Restriction with rectLinearDRMDMA packet
@@ -642,7 +654,7 @@ DmaBlitManager::copyBufferRect(
             if (!gpuMem(srcMemory).partialMemCopyTo(gpu(),
                 amd::Coord3D(srcOffset, srcRect.rowPitch_, srcRect.slicePitch_),
                 amd::Coord3D(dstOffset, dstRect.rowPitch_, dstRect.slicePitch_),
-                size, gpuMem(dstMemory), true)) {
+                size, gpuMem(dstMemory), true, false, bytesPerElement)) {
                 LogError("copyBufferRect failed!");
                 return false;
             }
diff --git a/rocclr/runtime/device/gpu/gpuresource.cpp b/rocclr/runtime/device/gpu/gpuresource.cpp
index 9ae11fba57..7035246240 100644
--- a/rocclr/runtime/device/gpu/gpuresource.cpp
+++ b/rocclr/runtime/device/gpu/gpuresource.cpp
@@ -1158,7 +1158,8 @@ Resource::partialMemCopyTo(
     const amd::Coord3D& size,
     Resource& dstResource,
     bool enableCopyRect,
-    bool flushDMA) const
+    bool flushDMA,
+    uint bytesPerElement) const
 {
     GpuEvent    event;
     bool        result;
@@ -1198,7 +1199,7 @@ Resource::partialMemCopyTo(
     result = gpu.copyPartial(event,
         gslResource(), calSrcOrigin,
         dstResource.gslResource(), calDstOrigin,
-        calSize, static_cast<CALmemcopyflags>(syncFlags), enableCopyRect);
+        calSize, static_cast<CALmemcopyflags>(syncFlags), enableCopyRect, bytesPerElement);
 
     if (result) {
         // Mark source and destination as busy
diff --git a/rocclr/runtime/device/gpu/gpuresource.hpp b/rocclr/runtime/device/gpu/gpuresource.hpp
index c7dcc635cf..0c891c2dde 100644
--- a/rocclr/runtime/device/gpu/gpuresource.hpp
+++ b/rocclr/runtime/device/gpu/gpuresource.hpp
@@ -232,7 +232,8 @@ public:
         const amd::Coord3D& size,       //!< Size of the region to copy
         Resource& dstResource,          //!< Destination resource
         bool enableRectCopy = false,    //!< Rectangular DMA support
-        bool flushDMA = false           //!< Flush DMA if requested
+        bool flushDMA = false,          //!< Flush DMA if requested
+        uint bytesPerElement = 1        //!< Bytes Per Element
         ) const;
 
     /*! \brief Copies size/4 DWORD of memory to a surface
diff --git a/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp b/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp
index 6c657344e0..dfd1f7dc84 100644
--- a/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp
+++ b/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.cpp
@@ -513,7 +513,8 @@ CALGSLContext::copyPartial(GpuEvent&      event,
                            size_t*          destOffset,
                            size_t*          size,
                            CALmemcopyflags  flags,
-                           bool             enableRectCopy)
+                           bool             enableRectCopy,
+                           uint32           bytesPerElement)
 {
     uint64      surfaceSize;
     uint32      mode = GSL_SYNCUPLOAD_IGNORE_ELEMENTSIZE;
@@ -563,7 +564,7 @@ CALGSLContext::copyPartial(GpuEvent&      event,
             }
             m_cs->syncUploadRawRect(srcMem, srcOffset[0], (uint32)srcOffset[1], (uint32)srcOffset[2],
                                     destMem, destOffset[0], (uint32)destOffset[1], (uint32)destOffset[2],
-                                    size[0], (uint32)size[1], (uint32)size[2], mode);
+                                    size[0], (uint32)size[1], (uint32)size[2], mode, bytesPerElement);
         }
         else
         {
diff --git a/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.h b/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.h
index cd02a46bc3..ecd04024d9 100644
--- a/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.h
+++ b/rocclr/runtime/device/gpu/gslbe/src/rt/GSLContext.h
@@ -71,7 +71,7 @@ public:
     void             destroyProgramObject(gslProgramObject func);
 
     bool             copyPartial(GpuEvent& event, gslMemObject srcMem, size_t* srcOffset,
-                        gslMemObject destMem, size_t* destOffset, size_t* size, CALmemcopyflags flags, bool enableCopyRect);
+                        gslMemObject destMem, size_t* destOffset, size_t* size, CALmemcopyflags flags, bool enableCopyRect, uint32 bytesPerElement);
 
     void             setSamplerParameter(uint32 sampler, gslTexParameterPname param, CALvoid* vals);