diff options
| author | Fredrik Svedberg <fredrik.svedberg@arm.com> | 2020-08-25 11:31:47 +0200 |
|---|---|---|
| committer | Fredrik Svedberg <fredrik.svedberg@arm.com> | 2020-08-27 16:39:29 +0200 |
| commit | 880e73543120648f08886365a45e8b2ce32d5ff1 (patch) | |
| tree | 420627fb8e7e5000f23f82c52ba6569f4a696813 /ethosu/vela/architecture_features.py | |
| parent | a41cd4de2af1e43b76a2a33d78eeb2d90a88b757 (diff) | |
| download | ethos-u-vela-880e73543120648f08886365a45e8b2ce32d5ff1.tar.gz | |
[MLBEDSW-2846] Do not use NHCWB16 for reduce_sum int32
Added checks for not using NHCWB16 for reduce_sum int32 which makes
int8/uint8 softmax work.
Also enabled softmax graph rewrite by default and fixed a saturation
problem.
Change-Id: Ic01bd9ece7e5c3edb2900b7915cc747efe9e5760
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
Diffstat (limited to 'ethosu/vela/architecture_features.py')
| -rw-r--r-- | ethosu/vela/architecture_features.py | 3 |
1 files changed, 1 insertions, 2 deletions
diff --git a/ethosu/vela/architecture_features.py b/ethosu/vela/architecture_features.py index 8b968a3e..fd0e5c06 100644 --- a/ethosu/vela/architecture_features.py +++ b/ethosu/vela/architecture_features.py @@ -183,7 +183,6 @@ Note the difference between ArchitectureFeatures and CompilerOptions block_config_limit, global_memory_clock_scale, max_blockdep, - softmax_support, weight_estimation_scaling, ): accelerator_config = accelerator_config.lower() @@ -332,7 +331,7 @@ Note the difference between ArchitectureFeatures and CompilerOptions self.generate_block_config_map(Block(ifm_block_max.width, ifm_block_max.height, 128)) # Setup supported operators and restriction checkers class - self.supported_operators = SupportedOperators(softmax_support) + self.supported_operators = SupportedOperators() # Returns available number of SHRAM banks depending on activation lookup table # being used or not |