diff options
| author | Fredrik Svedberg <fredrik.svedberg@arm.com> | 2020-03-30 13:15:28 +0200 |
|---|---|---|
| committer | Tim Hall <tim.hall@arm.com> | 2020-06-18 17:53:52 +0100 |
| commit | d67c0aaccd91f4be3ea76f69fa063301ffc73aa3 (patch) | |
| tree | a9ad54b32d18685347847fcc59db8374bbe36233 /ethosu/vela/weight_compressor.py | |
| parent | fed918bfb26dc330a5f066ea5947bc5eb2db4651 (diff) | |
| download | ethos-u-vela-d67c0aaccd91f4be3ea76f69fa063301ffc73aa3.tar.gz | |
MLBEDSW-819: make int16 changes
Enabled int16 support quantization to match the reference.
Change-Id: Ib369640241a9a491f2b0bc52d7f6cb025e30344b
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
Diffstat (limited to 'ethosu/vela/weight_compressor.py')
| -rw-r--r-- | ethosu/vela/weight_compressor.py | 11 |
1 files changed, 7 insertions, 4 deletions
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py index 0b4ac696..92197248 100644 --- a/ethosu/vela/weight_compressor.py +++ b/ethosu/vela/weight_compressor.py @@ -25,7 +25,7 @@ import math import numpy as np from collections import namedtuple from .numeric_util import round_up -from .scaling import quantise_scale +from .scaling import quantise_scale, reduced_quantise_scale from .tensor import TensorPurpose, TensorSubPurpose, TensorFormat, TensorBlockTraversal from .operation import NpuBlockType from .architecture_features import Block @@ -287,7 +287,7 @@ def calc_scales_and_pack_biases(tens, arch, oc_quantum, rescale_for_faf=False): if not rescale_for_faf: if ifm_dtype == DataType.uint8: scales = [np.double(ifm_scale * weight_scale) / np.double(ofm_scale) for weight_scale in weight_scales] - elif ifm_dtype == DataType.int8: + elif ifm_dtype == DataType.int8 or ifm_dtype == DataType.int16: scales = [ (np.double(ifm_scale) * np.double(weight_scale)) / np.double(ofm_scale) for weight_scale in weight_scales @@ -297,13 +297,16 @@ def calc_scales_and_pack_biases(tens, arch, oc_quantum, rescale_for_faf=False): else: if ifm_dtype == DataType.uint8: scales = [np.double(ifm_scale * weight_scale * 0x3000) for weight_scale in weight_scales] - elif ifm_dtype == DataType.int8: + elif ifm_dtype == DataType.int8 or ifm_dtype == DataType.int16: scales = [(np.double(ifm_scale * 0x3000) * np.double(weight_scale)) for weight_scale in weight_scales] else: assert False, str(ifm_dtype) + " not implemented" # quantise all of the weight scales into (scale_factor, shift) - quantised_scales = [quantise_scale(scale) for scale in scales] + if ifm_dtype == DataType.int16: + quantised_scales = [reduced_quantise_scale(scale) for scale in scales] + else: + quantised_scales = [quantise_scale(scale) for scale in scales] for _, shift in quantised_scales: assert shift >= 16 |