diff options
| author | Rickard Bolin <rickard.bolin@arm.com> | 2024-06-26 13:17:17 +0000 |
|---|---|---|
| committer | Rickard Bolin <rickard.bolin@arm.com> | 2024-06-26 14:46:27 +0000 |
| commit | 86ba3fd1eef6b6ad9047ef4aaed3461e0b250a1a (patch) | |
| tree | cf1f2b5e47828b5f13842754ec007de10c6c16ab | |
| parent | c8450913d7f557a3cf6d547caebe691f41d2a00d (diff) | |
| download | ethos-u-vela-86ba3fd1eef6b6ad9047ef4aaed3461e0b250a1a.tar.gz | |
MLBEDSW-9222: Fix rescale precision errors
After updating to NumPy 2.0, some variables that were previously
implicitly promoted to doubles instead remained floats, causing
precision loss. They are now explicitly promoted to doubles instead.
Change-Id: Ibc1838e2a3a05116e291a6320525b38972f0478e
Signed-off-by: Rickard Bolin <rickard.bolin@arm.com>
| -rw-r--r-- | ethosu/vela/register_command_stream_generator.py | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/ethosu/vela/register_command_stream_generator.py b/ethosu/vela/register_command_stream_generator.py index 01654697..d96072bf 100644 --- a/ethosu/vela/register_command_stream_generator.py +++ b/ethosu/vela/register_command_stream_generator.py @@ -730,7 +730,7 @@ def generate_ofm_scaling_for_pooling(emit: CommandStreamEmitter, pool_op: NpuPoo rescale = pool_op.rescale rescale_bits = len(bin(round_up_to_int(rescale))) - 2 + 1 scale, shift = scaling.quantise_pooling_scale(kernel.height * kernel.width, rescale_bits) - scale = int(round_away_zero(scale * rescale)) + scale = int(round_away_zero(scale * np.double(rescale))) else: # In case avg pool fused with concat or other memory operation, rescaling might be needed. # kernel height == kernel width == 1 is always true in this case @@ -745,7 +745,7 @@ def generate_ofm_scaling_for_pooling(emit: CommandStreamEmitter, pool_op: NpuPoo elif rescale < 1: rescale_bits = -(len(bin(round_up_to_int(1 / rescale))) - 2 - 1) scale, shift = scaling.quantise_pooling_scale(kernel.height * kernel.width, rescale_bits) - scale = int(round_away_zero(scale * rescale)) + scale = int(round_away_zero(scale * np.double(rescale))) else: scale = 1 shift = 0 |