diff options
| author | Louis Verhaard <louis.verhaard@arm.com> | 2022-03-01 11:26:58 +0100 |
|---|---|---|
| committer | Fredrik Svedberg <fredrik.svedberg@arm.com> | 2022-03-30 13:00:15 +0000 |
| commit | cc5f4de1c35ba44fca7ff6295c6ae846f8242344 (patch) | |
| tree | 68c4f8124a3ee6ec6f7fceb32a1d8aec11ac9a86 /ethosu/vela/weight_compressor.py | |
| parent | a19b4671dd0594181a2789930cc98bf5dc41ded4 (diff) | |
| download | ethos-u-vela-cc5f4de1c35ba44fca7ff6295c6ae846f8242344.tar.gz | |
MLBEDSW-6263: Use separate tensors for double buffering
Uses separate tensors for the individual weight buffers
in case of weight double buffering.
Each weight buffer tensor gets its own individual live range.
Change-Id: I724a8c61a7045615fbd2ed9535663076ac8edd13
Signed-off-by: Louis Verhaard <louis.verhaard@arm.com>
Diffstat (limited to 'ethosu/vela/weight_compressor.py')
| -rw-r--r-- | ethosu/vela/weight_compressor.py | 15 |
1 files changed, 11 insertions, 4 deletions
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py index 22fe512e..cdac6416 100644 --- a/ethosu/vela/weight_compressor.py +++ b/ethosu/vela/weight_compressor.py @@ -67,12 +67,19 @@ class NpuWeightTensor(Tensor): def __init__(self, name): Tensor.__init__(self, None, None, name + "_npu_encoded_weights") self.buffer = [] - self.max_range_bytes = 0 + self.double_buffer_sizes = [0, 0] # Required sizes if double buffering is used self.encoded_ranges = OrderedDict() self.hw_traversal = NpuBlockTraversal.DEPTH_FIRST self.dtype = DataType.uint8 self.scale_compression_config = None + def max_range_bytes(self): + return max(self.double_buffer_sizes) + + def double_buffer_size(self): + """Return total required size for double buffering""" + return sum(self.double_buffer_sizes) + class CompressedWeightCache: """Global tensor weight compression cache""" @@ -356,7 +363,7 @@ def encode_weight_and_scale_tensor( weights = np.flip(weights, axis=(0, 1)) encoded_stream = bytearray() - max_single_buffer_len = 0 + double_buffer_sizes = [0, 0] is_depthwise = npu_block_type == NpuBlockType.ConvolutionDepthWise # Bias & scale @@ -434,11 +441,11 @@ def encode_weight_and_scale_tensor( npu_tensor.encoded_ranges[key] = weight_range # Remember maximum encoded length for DoubleBuffering - max_single_buffer_len = max(max_single_buffer_len, len(encoded_stream) - buffer_start_offset) + double_buffer_sizes[idx % 2] = max(double_buffer_sizes[idx % 2], len(encoded_stream) - buffer_start_offset) # Attach buffer to tensor npu_tensor.buffer = encoded_stream - npu_tensor.max_range_bytes = max_single_buffer_len + npu_tensor.double_buffer_sizes = double_buffer_sizes npu_tensor.set_all_shapes([1, 1, 1, len(encoded_stream)]) npu_tensor.format = TensorFormat.WeightsCompressed |