diff options
| author | Louis Verhaard <louis.verhaard@arm.com> | 2020-09-23 10:27:11 +0200 |
|---|---|---|
| committer | Louis Verhaard <louis.verhaard@arm.com> | 2020-09-23 15:05:21 +0200 |
| commit | 9db529a502fa51c5a2b82e1981c4cdfefa3f85d7 (patch) | |
| tree | ae50227abaf70dad255e428145d23da7ed0de766 /ethosu/vela/weight_compressor.py | |
| parent | d8c032d4be2a641946507b63023456312e333cb8 (diff) | |
| download | ethos-u-vela-9db529a502fa51c5a2b82e1981c4cdfefa3f85d7.tar.gz | |
MLBEDSW-3070: Fix addressing of weights
Assign different equivalence ids to weights with same values but
different compression, to ensure correct addressing.
Signed-off-by: Louis Verhaard <louis.verhaard@arm.com>
Change-Id: I13aabad71520e4f4a78fb2d6a81740bdd4d1256c
Diffstat (limited to 'ethosu/vela/weight_compressor.py')
| -rw-r--r-- | ethosu/vela/weight_compressor.py | 9 |
1 files changed, 6 insertions, 3 deletions
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py index 2374cd42..c5a3f3fd 100644 --- a/ethosu/vela/weight_compressor.py +++ b/ethosu/vela/weight_compressor.py @@ -30,6 +30,7 @@ from .numeric_util import round_up_divide from .operation import NpuBlockType from .scaling import quantise_scale from .scaling import reduced_quantise_scale +from .tensor import create_equivalence_id from .tensor import TensorBlockTraversal from .tensor import TensorFormat from .tensor import TensorPurpose @@ -40,7 +41,7 @@ from ethosu import mlw_codec # Contains meta info for a weight compression. If two tensors have identical weight compression config, # then they also will have identical compressed weights. WeightCompressionConfig = namedtuple( - "WeightCompressionConfig", ["npu_block_type", "ofm_block_depth", "ofm_depth_step", "dilation", "equivalence_id"] + "WeightCompressionConfig", ["npu_block_type", "ofm_block_depth", "ofm_depth_step", "dilation", "value_id"] ) @@ -136,7 +137,7 @@ def create_weight_compression_config(tens, npu_block_type, ofm_block_depth, ofm_ # Note: for an ofm block only its depth is used in weight compression. # And block depth > ofm depth gives same result as block depth == ofm depth block_depth = min(ofm_block_depth, tens.quant_values.shape[-1]) - return WeightCompressionConfig(npu_block_type, block_depth, ofm_depth_step, dilation, tens.equivalence_id) + return WeightCompressionConfig(npu_block_type, block_depth, ofm_depth_step, dilation, tens.value_id) def set_storage_shape(tens): @@ -286,13 +287,15 @@ def compress_weights(arch, nng, tens, npu_block_type, ofm_block_depth, ofm_depth nng.weight_cache = CompressedWeightCache() wcc = create_weight_compression_config(tens, npu_block_type, ofm_block_depth, ofm_depth_step, dilation) tens.weight_compression_config = wcc + # Reassign equivalence id such that tensors with same weight compression get identical equivalence ids, + # but tensors with the same values but different compression get different equivalence ids + tens.equivalence_id = create_equivalence_id(wcc) tens_cached = nng.weight_cache.get_tensor_with_same_compression(wcc) if tens_cached is not None: # Cache hit, copy weights from the cache tens.copy_compressed_weight_info(tens_cached) set_storage_shape(tens) return - # No cache hit, perform the compression assert tens.quantization is not None assert tens.quantization.scale_f32 is not None |