diff options
| author | Andreas Nevalainen <andreas.nevalainen@arm.com> | 2020-10-28 15:42:08 +0100 |
|---|---|---|
| committer | patrik.gustavsson <patrik.gustavsson@arm.com> | 2020-11-11 08:34:16 +0000 |
| commit | 897cc14968e017b1f48f376f7f7cefc515c5fe88 (patch) | |
| tree | 27d17a59f1529c1ed0b1cc59e60438ee7f5d29d2 /ethosu/vela/weight_compressor.py | |
| parent | 73320a48dfa711f5938b0e3d8e03b9858558b899 (diff) | |
| download | ethos-u-vela-897cc14968e017b1f48f376f7f7cefc515c5fe88.tar.gz | |
MLBEDSW-3222: Bias tensors in fast storage
For IFM streamed cascades bias tensors are read several times.
Moves these tensors to fast storage and add DMA commands.
Change-Id: I630f6275986c1b5e3f126c925b11e22500fb1128
Signed-off-by: Andreas Nevalainen <andreas.nevalainen@arm.com>
Diffstat (limited to 'ethosu/vela/weight_compressor.py')
| -rw-r--r-- | ethosu/vela/weight_compressor.py | 8 |
1 files changed, 7 insertions, 1 deletions
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py index 94535216..b0187b65 100644 --- a/ethosu/vela/weight_compressor.py +++ b/ethosu/vela/weight_compressor.py @@ -404,7 +404,7 @@ def compress_weights(arch, nng, tens, npu_block_type, ofm_block_depth, ofm_depth def calc_scales_and_pack_biases(tens, arch, ofm_depth_step, rescale_for_faf=False): - assert tens.purpose == TensorPurpose.FeatureMap + assert tens.purpose in [TensorPurpose.FeatureMap, TensorPurpose.FSBias] assert tens.format == TensorFormat.NHWC # the connected operator should expect a bias input unless it is a FullyConnected assert tens.consumer_list[0].type.needs_bias() @@ -531,3 +531,9 @@ def update_pass_weight_and_scale_tensors(nng, arch): if (ps.ops[-1].type in activation_ops) and (ps.npu_block_type != NpuBlockType.ElementWise): rescale_for_faf = True calc_scales_and_pack_biases(ps.scale_tensor, arch, ofm_depth_step, rescale_for_faf) + if ps.scale_tensor.ops[0].type == Op.DMA: + src_tens = ps.scale_tensor.get_dma_src_tensor() + src_tens.shape = ps.scale_tensor.shape + src_tens.quant_values = ps.scale_tensor.quant_values + src_tens.element_size_bytes = ps.scale_tensor.element_size_bytes + src_tens.copy_compressed_weight_info(ps.scale_tensor) |