commit 21d9031e6d0a423b2e6c02bddb98334a909804d7
author Rickard Bolin <rickard.bolin@arm.com> Thu Aug 11 14:26:20 2022 +0000
committer Rickard Bolin <rickard.bolin@arm.com> Tue Aug 16 12:17:10 2022 +0000
tree 9177290b32e8a88e991880e39e00b598b4b88ff1
parent dd49a7205b76d668308262b2ea4cc38968ff576c

MLBEDSW-6640: Modify elementwise block size selection

Limit relative cost to 1 for elementwise operations since increasing
block size when the full ofm already fits gives no additional benefits.

Signed-off-by: Rickard Bolin <rickard.bolin@arm.com>
Change-Id: Ib6128f6346834fd916efa59adbe07a069dbda0ae

ethosu/vela/architecture_allocator.py

diff --git a/ethosu/vela/architecture_allocator.py b/ethosu/vela/architecture_allocator.py
index b5edcab..9cc22bb 100644
--- a/ethosu/vela/architecture_allocator.py
+++ b/ethosu/vela/architecture_allocator.py
@@ -330,7 +330,7 @@
 
                     # Scale relative to every output OFM element
                     if npu_op_type == NpuBlockType.ElementWise:
-                        relative_cost = ofm_shape.elements() / (height * width * depth)
+                        relative_cost = max(ofm_shape.elements() / (height * width * depth), 1)
                     else:
                         relative_cost = (ifm_fetch + weight_fetch) / ofm_shape.elements()