Re-use auxiliary memory withing CpuWinogradConv2d operators

Input/Output transformation operations are independent and done in
different time-steps of the algorithm, this memory can be re-used
between this transformation stages.

Moreover, reduce the allocation when extracting workspace sizes for
Winograd trasformations. There is a mix return of sizes in bytes and
elements, thus ensure the correct is in place. storage_size() member
functions return elements while working_space() function bytes.

Resolves: COMPMID-4781

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I705445ba7ca818cead48369db3cacd49684c7192
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6145
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>

1 files changed, 13 insertions, 14 deletions

diff --git a/src/runtime/cpu/operators/CpuWinogradConv2d.h b/src/runtime/cpu/operators/CpuWinogradConv2d.h
index ae705ac86b..b5b9c3f2e3 100644
--- a/src/runtime/cpu/operators/CpuWinogradConv2d.h
+++ b/src/runtime/cpu/operators/CpuWinogradConv2d.h
@@ -93,20 +93,19 @@ public:
 private:
     enum AuxTensorIdx
     {
-        GemmWorkspace = 0,
-        Pretranspose,
-        InterleavedLHS,
-        TransposedRHS,
-        TempResult,
-        PermutedInput,
-        InputTransformed,
-        InputWorkspace,
-        PermutedOutput,
-        PermutedWeights,
-        WeightsTransformed,
-        OutputTransformed,
-        OutputWorkspace,
-        Count
+        GemmWorkspace      = 0,
+        Pretranspose       = 1,
+        InterleavedLHS     = 2,
+        TransposedRHS      = 3,
+        TempResult         = 4,
+        TransformedInput   = 5,
+        TransformedOutput  = 6,
+        WorkspaceIO        = 7,
+        TransformedWeights = 8,
+        PermutedWeights    = 9,
+        PermutedInput      = TransformedOutput,
+        PermutedOutput     = TransformedInput,
+        Count              = 10
     };
 
     std::unique_ptr<CpuGemm>       _gemm_function;