commit 5cc4c76988eb72c89fe92e96ae670b1828fedf88
author Fredrik Svedberg <fredrik.svedberg@arm.com> Thu Jun 16 13:14:52 2022 +0200
committer Fredrik Svedberg <fredrik.svedberg@arm.com> Fri Jun 17 09:34:57 2022 +0200
tree 429bc452a6a2e846a9610500e67e82a674e750ae
parent 7ce6b32b815b6eb01bf3b348a3df37bb783c3a94

MLBEDSW-6614 Improve elementwise block size selection

Improved block size selection by favouring larger
block sizes for elementwise operations.

Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
Change-Id: I5b30b358d84fcd672935b863c2154bd8f4ccd928

ethosu/vela/architecture_allocator.py

diff --git a/ethosu/vela/architecture_allocator.py b/ethosu/vela/architecture_allocator.py
index e7e7471..b5edcab 100644
--- a/ethosu/vela/architecture_allocator.py
+++ b/ethosu/vela/architecture_allocator.py
@@ -329,7 +329,10 @@
                         ifm_fetch *= full_blocks.depth
 
                     # Scale relative to every output OFM element
-                    relative_cost = (ifm_fetch + weight_fetch) / ofm_shape.elements()
+                    if npu_op_type == NpuBlockType.ElementWise:
+                        relative_cost = ofm_shape.elements() / (height * width * depth)
+                    else:
+                        relative_cost = (ifm_fetch + weight_fetch) / ofm_shape.elements()
 
                     # If the entire IFM can be encompassed by both buffers, bias to prefer this configuration
                     if ifm_shape.elements() < ifm_block.elements() * 2: