In CopyTensorContentsGeneric coalesce inner dimensions where possible

This reduces the number of function calls in the inner loop, and
allows for optimised implementations of memcpy to improve bandwidth

Signed-off-by: Matthew Bentham <matthew.bentham@arm.com>
Change-Id: I7458b45c075c87805242e92e54448b9dd762227f

1 files changed, 21 insertions, 0 deletions

diff --git a/src/backends/backendsCommon/WorkloadUtils.hpp b/src/backends/backendsCommon/WorkloadUtils.hpp
index 3e0c40d890..cb614ea5b9 100644
--- a/src/backends/backendsCommon/WorkloadUtils.hpp
+++ b/src/backends/backendsCommon/WorkloadUtils.hpp
@@ -125,6 +125,27 @@ void CopyTensorContentsGeneric(const ITensorHandle* srcTensor, ITensorHandle* ds
     size_t copyBatches = std::min(srcBatches, dstBatches);
     size_t copyDepth   = std::min(srcDepth, dstDepth);
 
+    // Coalesce inner dimensions where possible
+    // to reduce overheard calling copy() and to
+    // allow for memory bandwidth optimisations
+    if (copyLength == srcWidthStride &&
+        copyLength == dstWidthStride)
+    {
+        // There is no special padding between rows,
+        // and sizes are compatible, so copy whole rows
+        copyLength *= copyWidth;
+        copyWidth = 1;
+
+        if (copyLength == srcHeightStride &&
+            copyLength == dstHeightStride)
+        {
+            // There is no special padding between batches
+            // and sizes are compatible so copy whole batches
+            copyLength *= copyHeight;
+            copyHeight = 1;
+        }
+    }
+
     for (unsigned int d = 0; d < copyDepth; ++d)
     {
         auto srcPtrDepth = srcData;