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>
diff --git a/ethosu/vela/high_level_command_to_npu_op.py b/ethosu/vela/high_level_command_to_npu_op.py
index 2c62c6f..7634fe1 100644
--- a/ethosu/vela/high_level_command_to_npu_op.py
+++ b/ethosu/vela/high_level_command_to_npu_op.py
@@ -54,6 +54,7 @@
from .high_level_command_stream import Box
from .high_level_command_stream import Command
from .high_level_command_stream import DMA
+from .high_level_command_stream import NOP
from .high_level_command_stream import NpuStripe
from .numeric_util import quantise_float32
from .numeric_util import round_up
@@ -627,7 +628,8 @@
else:
src_addr = cmd.in_tensor.address_for_coordinate(cmd.box.start_coord)
dest_addr = cmd.out_tensor.address_for_coordinate(cmd.box.start_coord)
- sz = cmd.in_tensor.address_for_coordinate(cmd.box.end_coord, is_top_box=True) - src_addr
+ # DMA must use 16 bytes alignment (tensors are always aligned but the sz calculation uses actual size)
+ sz = round_up(cmd.in_tensor.address_for_coordinate(cmd.box.end_coord, is_top_box=True) - src_addr, 16)
src = NpuAddressRange(src_region, int(src_addr), int(sz))
dest = NpuAddressRange(dest_region, int(dest_addr), int(sz))
return NpuDmaOperation(src, dest)
@@ -663,6 +665,9 @@
for cmd in sg.high_level_command_stream:
if isinstance(cmd, NpuStripe) and cmd.ps.npu_block_type == NpuBlockType.Default:
print("Warning: Skipping register command stream generation for", cmd.ps)
+ elif isinstance(cmd, NOP):
+ # NOP should not generate anything
+ continue
else:
npu_op = convert_command_to_npu_op(cmd, arch)
npu_op_list.append(npu_op)