MLBEDSW-6260: Add support for using DMA to copy feature maps

- Reshape ops can be bypassed and there is no need to process them by the NPU.
There are use cases when the IFM must be preserved so a memcpy is needed.
This is implemented by an AvgPool.
- In order to reduce the cost of the AvgPool the IFM can be copied by DMA.
This is faster and also it can be turned into a real NOP in cases where
the IFM and the OFM can use the same memory space.
- Added new memcpy op. Only NHWC format supported since DMA can not change
the format on the fly.
- Allow ofm to reuse ifm for memcpy op
- Make sure the DMA copy size is 16 byte aligned

Change-Id: I3605a48d47646ff60d2bb3644dd3a23f872235a7
Signed-off-by: Johan Alfven <johan.alfven@arm.com>

1 files changed, 5 insertions, 0 deletions

diff --git a/ethosu/vela/operation.py b/ethosu/vela/operation.py
index 19b00b31..6be9dc25 100644
--- a/ethosu/vela/operation.py
+++ b/ethosu/vela/operation.py
@@ -51,6 +51,7 @@ class NpuBlockType(Enum):
     ConvolutionDepthWise = 4
     ElementWise = 5
     ReduceSum = 6
+    Dma = 7
 
 
 class Kernel:
@@ -174,6 +175,7 @@ class Op(Enum):
     )
     Dequantize = OperatorInfo(indices=NNG_IFM_INDICES)
     Div = OperatorInfo()
+    Memcpy = OperatorInfo(block_type=NpuBlockType.Dma, indices=NNG_IFM_INDICES)
     Elu = OperatorInfo()
     EmbeddingLookup = OperatorInfo()
     EmbeddingLookupSparse = OperatorInfo()
@@ -373,6 +375,9 @@ class Op(Enum):
     def is_resize_op(self):
         return self in (Op.ResizeBilinear, Op.ResizeNearestNeighbor)
 
+    def is_memcpy_op(self):
+        return self.info.block_type == NpuBlockType.Dma
+
     def needs_bias(self):
         return bool(self.info.indices.biases)