From 628928d32ceae5c95abd9b3a2cb7333b1e5de059 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johan=20Alfv=C3=A9n?= Date: Thu, 27 Jan 2022 06:47:26 +0100 Subject: MLBEDSW-6060: Revert patch for MLBEDSW-5582 - Issue was due to a previous patch to fix MLBEDSW-5582 - Revert fix for MLBEDSW-5582 commit 849ff81f82c10a68898e5101930b92372bec5565, - Made new fix for MLBEDSW-5582 that enforce output tensor from NPU graphs to be in NHWC format. This information is otherwise lost in the case when parts of a concatenation are placed in different custom operators resulting in mismatch bewteen NHWC and NHCWB16. Signed-off-by: Johan Alfven Change-Id: Iab3ba29d348353c854f357836e6aa7c338ae1572 --- ethosu/vela/extract_npu_subgraphs.py | 3 +++ ethosu/vela/pass_packing.py | 17 ++++++----------- 2 files changed, 9 insertions(+), 11 deletions(-) diff --git a/ethosu/vela/extract_npu_subgraphs.py b/ethosu/vela/extract_npu_subgraphs.py index d802b513..f46f031a 100644 --- a/ethosu/vela/extract_npu_subgraphs.py +++ b/ethosu/vela/extract_npu_subgraphs.py @@ -98,6 +98,9 @@ def rewrite_tensor_cpu_producer_npu_consumers( # 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( diff --git a/ethosu/vela/pass_packing.py b/ethosu/vela/pass_packing.py index b84e4559..1fefdf42 100644 --- a/ethosu/vela/pass_packing.py +++ b/ethosu/vela/pass_packing.py @@ -179,7 +179,7 @@ for (operation_set, incompatible_pack_flags, flags_to_set, flags_to_clear) in te def pack_into_passes(nng, arch, verbose_packing=False): - def visit_op(op, multiple_ops=None): + def visit_op(op, ignored): visit_op_refcount[op] += 1 if visit_op_refcount[op] == 1: # First-time visit, go and fix up unused output tensors @@ -187,6 +187,7 @@ def pack_into_passes(nng, arch, verbose_packing=False): if len(tens.consumers()) == 0: visit_op_refcount[op] += 1 + assert visit_op_refcount[op] <= len(op.outputs) if visit_op_refcount[op] == len(op.outputs): if op.type in startup_init_ops: @@ -197,9 +198,9 @@ def pack_into_passes(nng, arch, verbose_packing=False): ofm_tensor = op.outputs[0] ofm_shape = op.ofm_shapes[0] if op.run_on_npu else None - build_pass((op,), ofm_tensor, ofm_shape, multiple_ops) + build_pass((op,), ofm_tensor, ofm_shape) - def build_pass(start_ops_to_process, ofm_tensor=None, ofm_shape=None, multiple_ops=None): + def build_pass(start_ops_to_process, ofm_tensor=None, ofm_shape=None): reverse_ops_list = [] curr_flags = PassFlags.Empty npu_block_type = NpuBlockType.Default @@ -372,10 +373,6 @@ def pack_into_passes(nng, arch, verbose_packing=False): reverse_pass_list.append(ps) - if multiple_ops: - multiple_op_next = multiple_ops.pop(0) - visit_op(multiple_op_next, multiple_ops) - for inp, refcount in input_refcounts.items(): for _ in range(refcount): visit_tensor(inp) @@ -386,10 +383,8 @@ def pack_into_passes(nng, arch, verbose_packing=False): visit_tensor_refcount[tens] += 1 assert visit_tensor_refcount[tens] <= len(tens.consumers()) if visit_tensor_refcount[tens] == len(tens.consumers()): - if tens.ops: - op = tens.ops[0] - multiple_ops = [o for o in tens.ops if o != op] - visit_op(op, multiple_ops) + for op in reversed(tens.ops): + visit_op(op, tens) def create_primary_op(op_list): if any(op.type in (npu_post_ops | npu_post_fuse_limited_ops) and op.run_on_npu for op in op_list): -- cgit v1.2.1