MLBEDSW-5582: MLCE: memory corruption with zero concat
Fixed problem when ofm is produced by different NPU nodes by
making sure that output is always in NHWC format.
Signed-off-by: Johan Alfven <johan.alfven@arm.com>
Change-Id: I00e55c989d5860499fbaf4f4318661b17b4bda7e
diff --git a/ethosu/vela/extract_npu_subgraphs.py b/ethosu/vela/extract_npu_subgraphs.py
index ac24e43..015634c 100644
--- a/ethosu/vela/extract_npu_subgraphs.py
+++ b/ethosu/vela/extract_npu_subgraphs.py
@@ -106,9 +106,6 @@
# Deal with output tensors for the NPU graph. These are special.
npu_subgraph.output_tensors = [new_tens if tens == orig_tens else tens for tens in npu_subgraph.output_tensors]
- for tens in npu_subgraph.output_tensors:
- # Enforce output tensor from NPU graph to use normal NHWC output
- tens.needs_linear_format = True
def rewrite_tensor_npu_producer_cpu_consumers(
@@ -241,6 +238,11 @@
tens, call_pass[curr_sg], startup_init_passes[curr_sg], curr_sg, orig_sg, subgraph_for_pass
)
+ for tens in curr_sg.output_tensors:
+ # ofm can depend on multiple ops. These ops can be divided into different NPU
+ # nodes due to CPU nodes. If that is the case the ofm must be NHWC.
+ tens.needs_linear_format = True
+
return new_subgraphs