[EXTAPI] exposing encode of biases to be consumed by an external API
*Renamed pack_bias_and_scale to encode_bias to be consumed externally
*added unit test for the API
Change-Id: I71829f3fcb390c475795848f0be3d132d3e158ee
Signed-off-by: Manupa Karunaratne <manupa.karunaratne@arm.com>
diff --git a/ethosu/vela/test/extapi/test_extapi_encode_bias.py b/ethosu/vela/test/extapi/test_extapi_encode_bias.py
new file mode 100644
index 0000000..59b8587
--- /dev/null
+++ b/ethosu/vela/test/extapi/test_extapi_encode_bias.py
@@ -0,0 +1,43 @@
+# Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# Contains unit tests for encode_biases API for an external consumer
+import random
+
+import numpy as np
+
+from ethosu.vela.weight_compressor import encode_bias
+
+
+def test_encode_bias():
+ bias_lower_limit = -(1 << (40 - 1))
+ bias_upper_limit = (1 << (40 - 1)) - 1
+ scale_lower_limit = 0
+ scale_upper_limit = (1 << 32) - 1
+ shift_lower_limit = 0
+ shift_upper_limit = (1 << 6) - 1
+
+ for _ in range(30):
+ bias = np.int64(random.randint(bias_lower_limit, bias_upper_limit))
+ scale = int(random.randint(scale_lower_limit, scale_upper_limit))
+ shift = int(random.randint(shift_lower_limit, shift_upper_limit))
+ biases_enc = encode_bias(bias, scale, shift)
+ assert isinstance(biases_enc, bytearray)
+ assert len(biases_enc) == 10
+
+
+if __name__ == "__main__":
+ test_encode_bias()
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py
index 687a080..fa7d2d1 100644
--- a/ethosu/vela/weight_compressor.py
+++ b/ethosu/vela/weight_compressor.py
@@ -94,6 +94,33 @@
return encoded_stream
+@typecheck
+def encode_bias(bias: np.int64, scale: int, shift: int):
+ """
+ Public facing API to pack bias and scale values as required by the hardware
+ :param bias: 64bit signed number that includes 40bit signed bias
+ :param scale: 32bit scale value
+ :param shift: 6bit shift value
+ :return: packed 80bit [0(2-bits),shift(6-bits),scale(32-bits),bias(40-bits)]
+ """
+ assert -(1 << (40 - 1)) <= bias < (1 << (40 - 1)) # signed 40-bit range
+ assert 0 <= scale < (1 << 32) # unsigned 32-bit range
+ assert 0 <= shift < (1 << 6) # unsigned 6-bit range
+
+ data = bytearray(10)
+ data[0] = (bias >> (0 * 8)) & 0xFF
+ data[1] = (bias >> (1 * 8)) & 0xFF
+ data[2] = (bias >> (2 * 8)) & 0xFF
+ data[3] = (bias >> (3 * 8)) & 0xFF
+ data[4] = (bias >> (4 * 8)) & 0xFF
+ data[5] = (scale >> (0 * 8)) & 0xFF
+ data[6] = (scale >> (1 * 8)) & 0xFF
+ data[7] = (scale >> (2 * 8)) & 0xFF
+ data[8] = (scale >> (3 * 8)) & 0xFF
+ data[9] = shift & 0x3F
+ return data
+
+
def create_weight_compression_config(tens, npu_block_type, ofm_block_depth, ofm_depth_step, dilation):
# 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
@@ -376,27 +403,6 @@
# the operator should only have a single output
assert len(tens.consumer_list[0].outputs) == 1
-
- def pack_bias_and_scale(bias, scale, shift):
- bias = np.int64(bias)
- assert -(1 << (40 - 1)) <= bias < (1 << (40 - 1)) # signed 40-bit range
- assert 0 <= scale < (1 << 32) # unsigned 32-bit range
- assert 0 <= shift < (1 << 6) # unsigned 6-bit range
-
- # pack the 80 bit value = [0(2-bits),shift(6-bits),scale(32-bits),bias(40-bits)]
- data = bytearray(10)
- data[0] = (bias >> (0 * 8)) & 0xFF
- data[1] = (bias >> (1 * 8)) & 0xFF
- data[2] = (bias >> (2 * 8)) & 0xFF
- data[3] = (bias >> (3 * 8)) & 0xFF
- data[4] = (bias >> (4 * 8)) & 0xFF
- data[5] = (scale >> (0 * 8)) & 0xFF
- data[6] = (scale >> (1 * 8)) & 0xFF
- data[7] = (scale >> (2 * 8)) & 0xFF
- data[8] = (scale >> (3 * 8)) & 0xFF
- data[9] = shift & 0x3F
- return data
-
biases = tens.quant_values
first_consumer_op = tens.consumer_list[0]
@@ -470,7 +476,7 @@
core_scales = quantised_scales[i + core : i + core + max_len : arch.ncores]
core_biases = biases[i + core : i + core + max_len : arch.ncores]
for j, core_bias in enumerate(core_biases):
- stream.extend(pack_bias_and_scale(core_bias, *core_scales[j]))
+ stream.extend(encode_bias(np.int64(core_bias), *core_scales[j]))
# Align to 16 for start for next substream
remainder = (len(stream)) % 16