blob: f6b03f6702cc2cad69fbe6211c14cb7aebe699de [file] [log] [blame]
Tim Hall79d07d22020-04-27 18:20:16 +01001# Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved.
2#
3# SPDX-License-Identifier: Apache-2.0
4#
5# Licensed under the Apache License, Version 2.0 (the License); you may
6# not use this file except in compliance with the License.
7# You may obtain a copy of the License at
8#
9# www.apache.org/licenses/LICENSE-2.0
10#
11# Unless required by applicable law or agreed to in writing, software
12# distributed under the License is distributed on an AS IS BASIS, WITHOUT
13# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14# See the License for the specific language governing permissions and
15# limitations under the License.
Tim Hall79d07d22020-04-27 18:20:16 +010016# Description:
17# Early optimisation of the network graph, using the rewrite_graph module to do the traversal of the graph. These are
18# split into two parts optimise_graph_a and optimise_graph_b.
Tim Hall79d07d22020-04-27 18:20:16 +010019import math
Diego Russoea6111a2020-04-14 18:41:58 +010020
21import numpy as np
22
Louis Verhaardd7911c42020-08-25 13:36:41 +020023from . import fp_math
Louis Verhaardb9fc33c2020-08-13 11:47:36 +020024from . import lut
Diego Russoea6111a2020-04-14 18:41:58 +010025from . import rewrite_graph
Louis Verhaardd7911c42020-08-25 13:36:41 +020026from . import scaling
Diego Russoea6111a2020-04-14 18:41:58 +010027from .data_type import DataType
Louis Verhaard7db78962020-05-25 15:05:26 +020028from .errors import UnsupportedFeatureError
Dwight Lidman42fed942020-05-29 09:37:03 +020029from .ethos_u55_regs.ethos_u55_regs import resampling_mode
Louis Verhaarde0ef2732020-06-03 08:56:44 +020030from .numeric_util import full_shape
Michael McGeagh8dbf8cf2020-09-08 11:09:48 +010031from .operation import create_avgpool_nop
Diego Russoe8a10452020-04-21 17:39:10 +010032from .operation import NpuBlockType
33from .operation import Operation
Fredrik Svedberga0c36242020-06-03 15:43:31 +020034from .softmax import SoftMax
Michael McGeaghc5b549b2020-08-07 11:54:28 +010035from .tensor import create_const_tensor
36from .tensor import create_reshape_tensor
Charles Xu9a03fdf2020-07-02 15:12:40 +020037from .tensor import QuantizationParameters
Diego Russoe8a10452020-04-21 17:39:10 +010038from .tensor import Tensor
Tim Hall79d07d22020-04-27 18:20:16 +010039
40passthrough_nodes = set(("Identity",))
41
Michael McGeagh11b0bdb2020-09-08 11:07:35 +010042conv_op = set(("Conv2D", "QuantizedConv2D", "Conv2DBackpropInputSwitchedBias", "Conv2DBiasAct"))
43fc_op = set(
44 (
45 "MatMul",
46 "QuantizedMatMul",
47 "BlockLSTM",
48 "RnnAct",
49 "UnidirectionalSequenceRnnAct",
50 "BidirectionalSequenceRnnAct",
51 "LstmAct",
52 "UnidirectionalSequenceLstmAct",
53 "BidirectionalSequenceLstmAct",
54 "FullyConnectedAct",
55 )
56)
57depthwise_op = set(("DepthwiseConv2dNative", "DepthwiseConv2dBiasAct",))
58pool_op = set(
59 ("AvgPool", "MaxPool", "QuantizedAvgPool", "QuantizedMaxPool", "AvgPoolAct", "MaxPoolAct", "ResizeBilinear")
60)
61reduce_sum_ops = set(("ReduceSum",))
62binary_elementwise_op = set(("AddAct", "MulAct", "SubAct", "Maximum", "Minimum"))
63elementwise_op = set(("LeakyRelu", "Abs", "CLZ", "SHL", "SHR")) | binary_elementwise_op
64relu_ops = set(("Relu", "Relu6", "ReluN1To1"))
65activation_ops = set(("Sigmoid", "Tanh")) | relu_ops
66memory_only_ops = set(("Reshape",))
67
Tim Hall79d07d22020-04-27 18:20:16 +010068
69def remove_passthrough_tensor(tens, arch):
70 if len(tens.ops) == 1 and tens.ops[0].type in passthrough_nodes:
71 assert len(tens.ops[0].inputs) == 1
72 tens = tens.ops[0].inputs[0]
73 return tens
74
75
76def rewrite_concat(tens, arch):
77 if len(tens.ops) == 1 and tens.ops[0].is_concat_op():
78 concat_op = tens.ops[0]
79 if tens != concat_op.outputs[0]:
80 return tens # don't attempt to rewrite the min/max outputs of QuantizedConcat
81
82 # Not supported so leave it and run on CPU
83 if not concat_op.run_on_npu:
84 return tens
85
86 inputs, axis = concat_op.get_concat_inputs_axis()
87
88 tens.ops = []
89 offset = 0
90 for idx, inp in enumerate(inputs):
91 new_op = Operation("ConcatSliceWrite", concat_op.name + str(idx))
92 new_op.inputs = [inp]
93 new_op.outputs = [tens]
94 new_op.attrs["concat_axis"] = axis
95 new_op.attrs["concat_start"] = offset
96 offset += inp.shape[axis]
97 new_op.attrs["concat_end"] = offset
98 new_op.run_on_npu = True
99 tens.ops.append(new_op)
100 assert tens.shape[axis] == offset
101
Patrik Gustavsson29d568e2020-08-18 10:11:21 +0200102 # If axis corresponds to C-dimension, NHCWB16 can only be used in the output if all the concat_start's are a
103 # multiple of 16. This as, it is only then the address offset for the ofm, for all operations, will be 16 byte
104 # aligned. For other values of axis the address offsets will be 16 byte aligned, as they are all based on c = 0
Patrik Gustavsson458a2082020-08-13 13:41:05 +0200105 # and those addresses are always 16 byte aligned due to the NHCWB16 format.
Patrik Gustavsson29d568e2020-08-18 10:11:21 +0200106 if axis == (len(tens.shape) - 1):
Patrik Gustavsson458a2082020-08-13 13:41:05 +0200107 for op in tens.ops:
108 if op.attrs["concat_start"] % 16 != 0:
109 tens.avoid_NHCWB16 = True
110 break
111
Tim Hall79d07d22020-04-27 18:20:16 +0100112 return tens
113
114
115def rewrite_split(tens, arch):
116
117 if len(tens.ops) == 1 and tens.ops[0].is_split_op():
118 split_op = tens.ops[0]
119
120 # Not supported so leave it and run on CPU
121 if not split_op.run_on_npu:
122 return tens
123
124 inp, outputs, axis, offset_start, offset_end = split_op.get_split_inputs_axis()
125
126 tens.ops = []
127 new_op = Operation("SplitSliceRead", split_op.name)
128 new_op.inputs = [inp]
Tim Hall79d07d22020-04-27 18:20:16 +0100129
130 # For Split the offset cannot be extracted from the tensor so it has to
131 # be calculated from the index of the output tensor
Diego Russoea6111a2020-04-14 18:41:58 +0100132 if axis is not None:
Tim Hall79d07d22020-04-27 18:20:16 +0100133 # Get the start and end of the split
134 offset_start = [0] * len(tens.shape)
135 offset_end = [0] * len(tens.shape)
136 for out in outputs:
137 if out == tens:
138 break
139 offset_start[axis] += out.shape[axis]
140
Patrik Gustavssoneebb1c22020-08-18 15:03:04 +0200141 # If start offset is not a multiple of 16 in the C-dimension, NHCWB16 need to be avoided in the input
142 if (offset_start[-1] % 16) != 0:
143 inp.avoid_NHCWB16 = True
144
Tim Hall79d07d22020-04-27 18:20:16 +0100145 offset_end[axis] = offset_start[axis] + tens.shape[axis]
146
147 new_op.attrs["split_start"] = offset_start
148 new_op.attrs["split_end"] = offset_end
149 new_op.run_on_npu = True
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100150 new_op.set_output_tensor(tens)
Tim Hall79d07d22020-04-27 18:20:16 +0100151
152 return tens
153
154
155def needed_total_padding(input_size, stride, filter_size):
156 out_size = (input_size + stride - 1) // stride
157 needed_input = (out_size - 1) * stride + filter_size
158 total_padding = max(0, needed_input - input_size)
159 return total_padding
160
161
162def calc_padding_and_skirt(padding_type, kernel_size, stride, input_dims):
163 ypad = needed_total_padding(int(input_dims[1]), int(stride[1]), int(kernel_size[0]))
164 xpad = needed_total_padding(int(input_dims[2]), int(stride[2]), int(kernel_size[1]))
165 if padding_type == b"SAME":
166 left_pad = (xpad + 0) // 2
167 right_pad = (xpad + 1) // 2
168 top_pad = (ypad + 0) // 2
169 bottom_pad = (ypad + 1) // 2
170 elif padding_type == b"VALID":
171 left_pad = 0
172 right_pad = 0
173 top_pad = 0
174 bottom_pad = 0
175 else:
Louis Verhaard7db78962020-05-25 15:05:26 +0200176 raise UnsupportedFeatureError("Unknown padding {}".format(str(padding_type)))
Tim Hall79d07d22020-04-27 18:20:16 +0100177 padding = (top_pad, left_pad, bottom_pad, right_pad)
178 skirt = (top_pad, left_pad, ypad - top_pad, xpad - left_pad)
179 return padding, skirt
180
Tim Hallc30f4952020-06-15 20:47:35 +0100181
Jacob Bohlin9b64ba02020-07-07 17:15:22 +0200182def calc_upscaled_padding_and_skirt(padding_type, kernel_size, stride, input_dims, upscaling_factor):
183 kernel_height, kernel_width = kernel_size[0], kernel_size[1]
Jacob Bohlincf7da102020-05-20 09:03:40 +0200184 if padding_type == b"SAME":
Jacob Bohlin9b64ba02020-07-07 17:15:22 +0200185 ypad = needed_total_padding(int(input_dims[1]) * upscaling_factor, int(stride[1]), int(kernel_height))
186 xpad = needed_total_padding(int(input_dims[2]) * upscaling_factor, int(stride[2]), int(kernel_width))
187
Jacob Bohlind47cc272020-08-24 11:42:14 +0200188 right_pad = max(((xpad + 1) // upscaling_factor) - 1, 0)
189 bottom_pad = max(((ypad + 1) // upscaling_factor) - 1, 0)
Jacob Bohlin9b64ba02020-07-07 17:15:22 +0200190 left_pad = max(kernel_width - 1 - right_pad, 0)
191 top_pad = max(kernel_height - 1 - bottom_pad, 0)
192
Jacob Bohlincf7da102020-05-20 09:03:40 +0200193 elif padding_type == b"VALID":
Jacob Bohlin9b64ba02020-07-07 17:15:22 +0200194 right_pad = max(kernel_width - 2, 0)
195 bottom_pad = max(kernel_height - 2, 0)
196 left_pad = kernel_width - 1
197 top_pad = kernel_height - 1
Jacob Bohlincf7da102020-05-20 09:03:40 +0200198 else:
199 assert 0, "Unknown padding"
200
201 padding = (top_pad, left_pad, bottom_pad, right_pad)
Jacob Bohlin9b64ba02020-07-07 17:15:22 +0200202 skirt = padding
Jacob Bohlincf7da102020-05-20 09:03:40 +0200203 return padding, skirt
204
Tim Hall79d07d22020-04-27 18:20:16 +0100205
206def fixup_conv2d_backprop(op, arch):
207 if op.type == "Conv2DBackpropInput":
208 # flip the inputs
209 op.inputs[0], op.inputs[2] = op.inputs[2], op.inputs[0]
Jacob Bohlincf7da102020-05-20 09:03:40 +0200210 op.type = "Conv2DBackpropInputSwitchedBias"
Jacob Bohlincf7da102020-05-20 09:03:40 +0200211
212 # Update strides
Tim Hallc30f4952020-06-15 20:47:35 +0100213 op.attrs.update({"stride_w": 1, "stride_h": 1, "strides": (1, 1, 1, 1)})
Tim Hall79d07d22020-04-27 18:20:16 +0100214
215 return op
216
217
Charles Xu9a03fdf2020-07-02 15:12:40 +0200218# Convert the op to an elementwise add
219def convert_resizebilinear_1x1_to_add(op):
220 op.type = "AddAct"
221 op.name = op.name + "_add"
222 op.attrs.update({"npu_block_type": NpuBlockType.ElementWise})
223 op.attrs["resizebilinear"] = True
224 # Create an input tensor filled with zeros
225 shape = op.outputs[0].shape
226 tens = Tensor(shape, op.inputs[0].dtype, op.inputs[1].name + "_add")
227 tens.values = np.zeros(shape)
228 tens.quant_values = np.zeros(shape, np.uint8)
229 tens.quantization = QuantizationParameters(0.0, 255.0)
230 tens.quantization.scale_f32 = 1.0
231 tens.quantization.zero_point = 0
232 tens.consumer_list = [op]
233 tens_op = op.inputs[1].ops[0]
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100234 tens_op.set_output_tensor(tens)
Charles Xu9a03fdf2020-07-02 15:12:40 +0200235 # Set the add inputs
236 op.inputs[1] = op.inputs[0]
237 op.inputs[0] = tens
238
239 return op
240
241
Charles Xu87c13502020-08-06 12:17:26 +0200242# Convert ResizeBilinear to a number of 2x2 pool ops
243def convert_resizebilinear_to_2x2_pool(op):
244 count = 0
245 pre_op = op
246 outputs = op.outputs
247
248 op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)})
249 if op.attrs["align_corners"]:
250 shape_modifier = 1
251 op.attrs["padding"] = b"VALID"
252 else:
253 shape_modifier = 0
254 op.attrs["padding"] = b"SAME"
255 op.inputs[0].resampling_mode = resampling_mode.NEAREST
256
257 upscaled_shape = np.array(op.inputs[0].shape[1:3])
258 out_shape = np.array(op.outputs[0].shape[1:3])
259 if (upscaled_shape == upscaled_shape * 2 - shape_modifier).all():
260 return op
261
262 while (upscaled_shape < out_shape).all():
263 if count == 0:
264 scaled_op = pre_op
265 else:
266 scaled_op = op.clone("_{}".format(count))
267 scaled_op.inputs[0] = pre_op.outputs[0]
268
269 upscaled_shape = upscaled_shape * 2 - shape_modifier
270
271 if (upscaled_shape == out_shape).all():
272 scaled_op.outputs = outputs
273 scaled_op.outputs[0].ops = [scaled_op]
274 else:
275 shape = outputs[0].shape.copy()
276 shape[1:3] = upscaled_shape[0:2]
277 out_tens = Tensor(shape, DataType.int16, "{}_{}".format(op.outputs[0].name, count))
278 out_tens.quantization = op.outputs[0].quantization.clone()
279 out_tens.quantization.quant_min = np.iinfo(np.int16).min
280 out_tens.quantization.quant_max = np.iinfo(np.int16).max
281 scaled_op.set_output_tensor(out_tens)
282 pre_op = scaled_op
283 count += 1
284
285 # Setup the scale value
286 if scaled_op.inputs[0].dtype.bits == 8 and scaled_op.outputs[0].dtype.bits == 16:
287 scaled_op.attrs["rescale"] = 128
288 elif scaled_op.inputs[0].dtype.bits == 16 and scaled_op.outputs[0].dtype.bits == 8:
289 scaled_op.attrs["rescale"] = 1 / 128
290 elif "rescale" in scaled_op.attrs:
291 del scaled_op.attrs["rescale"]
292
293 return op
294
295
Charles Xu9a03fdf2020-07-02 15:12:40 +0200296def fixup_resizebilinear(op, arch):
Charles Xu87c13502020-08-06 12:17:26 +0200297 if op.type == "ResizeBilinear" and op.run_on_npu:
298 if op.inputs[0].shape == op.outputs[0].shape:
Charles Xu36ffaf32020-08-05 15:40:44 +0200299 # Bypass nop resizebilinear
300 op.inputs = op.inputs[:1]
301 op.type = "Identity"
Charles Xu87c13502020-08-06 12:17:26 +0200302 elif op.inputs[0].shape[1] == 1 and op.inputs[0].shape[2] == 1:
303 convert_resizebilinear_1x1_to_add(op)
304 else:
305 convert_resizebilinear_to_2x2_pool(op)
Charles Xu9a03fdf2020-07-02 15:12:40 +0200306
307 return op
308
309
Dwight Lidmanc3862c22020-09-14 15:22:33 +0200310def convert_nop_split_to_identity(op, arch):
311 if op.type == "Split" and op.attrs.get("num_splits") == 1:
312 # the list comprehension should return a list with a single tensor
313 # if it shouldn't, remove_passthrough_tensor will fail appropriately
314 op.inputs = [i for i in op.inputs if i.shape == op.outputs[0].shape]
315 op.type = "Identity"
316 return op
317
318
Tim Hall79d07d22020-04-27 18:20:16 +0100319def fixup_fully_connected_input(op, arch):
320 if op.type == "FullyConnectedAct":
321 inp = op.inputs[0]
322 weights = op.inputs[1]
323
324 n_in_elems = weights.shape[-2]
325 elms = inp.elements()
326 batch_size = elms // n_in_elems
327 assert batch_size * n_in_elems == elms
328
329 desired_shape = [batch_size, n_in_elems]
330 if inp.shape != desired_shape:
331 # mismatch, insert a reshape to fix this.
Patrik Gustavssoncb337042020-09-16 14:55:40 +0200332 op.set_input_tensor(create_reshape_tensor(inp, desired_shape), 0)
Tim Hall79d07d22020-04-27 18:20:16 +0100333
334 return op
335
336
Patrik Gustavssoncb337042020-09-16 14:55:40 +0200337def convert_batched_fc_to_conv(op, arch):
338 if op.type == "FullyConnectedAct":
339 ifm = op.inputs[0]
340 ofm = op.outputs[0]
341 # Check if the FC is 2D and first dimension indicates batching
342 if len(ifm.shape) == len(ofm.shape) == 2 and ifm.shape[0] != 1:
343 n = ifm.shape[0]
344 batching_split = {4: (2, 2), 8: (2, 4), 16: (4, 4)}
345 h, w = batching_split.get(n, (1, n))
346
347 # Convert to convolution
348 op.name += "_conv"
349 op.type = "Conv2DBiasAct"
350 faf = op.attrs.get("fused_activation_function", None)
351 op.attrs = {
352 "dilation": (1, 1, 1, 1),
353 "dilation_h_factor": 1,
354 "dilation_w_factor": 1,
355 "fused_activation_function": faf,
356 "npu_block_type": NpuBlockType.ConvolutionMxN,
357 "padding": b"SAME",
358 "stride_h": 1,
359 "stride_w": 1,
360 "strides": (1, 1, 1, 1),
361 }
362
363 prev_op = ifm.ops[0]
364 desired_shape = [1, h, w, ifm.shape[-1]]
365 if len(ifm.consumer_list) == 1 and prev_op is not None and prev_op.type == "Reshape":
366 # There is a preceding Reshape
367 # Compare input of prev_op and input of op, to see if prev_op can be removed
368 ifm_prev_op = prev_op.inputs[0]
369 if ifm_prev_op.shape == ifm.shape and ifm_prev_op.quantization.is_scaling_equal(ifm.quantization):
370 # prev_op can be removed
371 op.set_input_tensor(ifm_prev_op, 0)
372 else:
373 op.inputs[0].set_all_shapes(desired_shape)
374 prev_op.set_input_tensor(
375 create_const_tensor(prev_op.inputs[1].name, [1], DataType.int32, desired_shape), 1
376 )
377 prev_op.attrs["new_shape"] = desired_shape
378 else:
379 # Add reshape op to the input if there is no preceding reshape
380 ifm.consumer_list.remove(op)
381 op.set_input_tensor(create_reshape_tensor(ifm, desired_shape), 0)
382
383 # Reshape Weights to be 4D. IO becomes HWIO
384 weight_tensor = op.inputs[1]
385 weight_tensor.quant_values = np.expand_dims(np.expand_dims(weight_tensor.quant_values, axis=0), axis=0)
386 weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape))
387
388 desired_shape = [1, h, w, ofm.shape[-1]]
389 if (
390 len(ofm.consumer_list) == 1
391 and ofm.consumer_list[0] is not None
392 and ofm.consumer_list[0].type == "Reshape"
393 ):
394 # There is a subsequent Reshape
395 # Compare desired shape and output of consumer op, to see if consumer op can be removed
396 ofm_cons_op = ofm.consumer_list[0].outputs[0]
397 if desired_shape == ofm_cons_op.shape and ofm.quantization.is_scaling_equal(ofm_cons_op.quantization):
398 op.outputs[0] = ofm_cons_op
399 op.outputs[0].ops = [op]
400 else:
401 op.outputs[0].set_all_shapes(desired_shape)
402 else:
403 # Add rehape op to the output
404 op.set_output_tensor(create_reshape_tensor(ofm, desired_shape, False))
405 return op
406
407
Tim Hall79d07d22020-04-27 18:20:16 +0100408def fixup_pack_input(op, arch):
409 if op.type == "Pack":
410 # Pack is also referred to as Stack
411 # Requires the rewrite_concat function to be called on the op afterwards
412 axis = int(op.attrs["axis"])
413 desired_shape = op.inputs[0].shape[:axis] + [1] + op.inputs[0].shape[axis:]
414
415 # Construct 1 shape tensor to be used by all inserted reshape ops
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100416 new_shape_tens = create_const_tensor(op.name + "_reshape_shape", [1], DataType.int32, desired_shape)
Tim Hall79d07d22020-04-27 18:20:16 +0100417
418 for idx, inp in enumerate(op.inputs):
Tim Hall79d07d22020-04-27 18:20:16 +0100419 reshape_out = inp.clone("_reshaped")
Michael McGeagh6a8d4242020-07-28 12:17:59 +0100420 reshape_out.set_all_shapes(desired_shape)
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100421
422 reshape_op = Operation("Reshape", "{}{}_reshape".format(op.name, idx))
423 reshape_op.attrs["new_shape"] = desired_shape
424 reshape_op.inputs = [inp, new_shape_tens]
425 reshape_op.set_output_tensor(reshape_out)
Tim Hall79d07d22020-04-27 18:20:16 +0100426
427 op.inputs[idx] = reshape_out
428
429 op.type = "PackReshaped"
430
431 return op
432
433
434def fixup_unpack_output(tens, arch):
435 op = tens.ops[0]
436 if op.type in set(("Unpack", "StridedSlice")):
437 # Unpack is also referred to as Unstack
438 # Requires the rewrite_split function to be called on the op afterwards
Patrik Gustavssoncf728902020-04-30 08:57:23 +0200439
440 reshape_input_shape = tens.shape
Tim Hall79d07d22020-04-27 18:20:16 +0100441 if op.type == "StridedSlice":
Patrik Gustavssoncf728902020-04-30 08:57:23 +0200442 new_axis_mask = op.attrs["new_axis_mask"]
Tim Hall79d07d22020-04-27 18:20:16 +0100443 shrink_axis_mask = op.attrs["shrink_axis_mask"]
Louis Verhaard7db78962020-05-25 15:05:26 +0200444 ellipsis_mask = op.attrs["ellipsis_mask"]
Patrik Gustavssoncf728902020-04-30 08:57:23 +0200445
446 if (new_axis_mask != 0 and shrink_axis_mask != 0) or ellipsis_mask != 0:
447 # Not supported, will be put on CPU
448 return tens
449 if shrink_axis_mask == 0 and new_axis_mask == 0:
Tim Hall79d07d22020-04-27 18:20:16 +0100450 # Equal Rank StridedSlice, no need to insert reshape
451 return tens
Patrik Gustavssoncf728902020-04-30 08:57:23 +0200452 elif shrink_axis_mask != 0:
453 n = 0
454 axis = 0
455 while shrink_axis_mask:
456 prev_mask = shrink_axis_mask
457 n += 1
458 shrink_axis_mask &= shrink_axis_mask - 1
459 axis = int(math.log2(prev_mask - shrink_axis_mask))
460 reshape_input_shape = reshape_input_shape[:axis] + [1] + reshape_input_shape[axis:]
Tim Hall79d07d22020-04-27 18:20:16 +0100461
Patrik Gustavssoncf728902020-04-30 08:57:23 +0200462 assert len(tens.shape) == (len(op.inputs[0].shape) - n)
463 op.attrs["shrink_axis_mask"] = 0
Tim Hall79d07d22020-04-27 18:20:16 +0100464
Patrik Gustavssoncf728902020-04-30 08:57:23 +0200465 elif new_axis_mask != 0:
466 n = 0
467 axis = 0
468 while new_axis_mask:
469 prev_mask = new_axis_mask
470 n += 1
471 new_axis_mask &= new_axis_mask - 1
472 axis = int(math.log2(prev_mask - new_axis_mask))
Louis Verhaard7db78962020-05-25 15:05:26 +0200473 reshape_input_shape = reshape_input_shape[:axis] + reshape_input_shape[(axis + 1) :]
Patrik Gustavssoncf728902020-04-30 08:57:23 +0200474 new_axis_mask >>= 1
475
476 assert len(tens.shape) == (len(op.inputs[0].shape) + n)
477 op.attrs["new_axis_mask"] = 0
Tim Hall79d07d22020-04-27 18:20:16 +0100478 else:
479 axis = int(op.attrs["axis"])
480 op.type = "UnpackReshaped"
Patrik Gustavssoncf728902020-04-30 08:57:23 +0200481 reshape_input_shape = tens.shape[:axis] + [1] + tens.shape[axis:]
Tim Hall79d07d22020-04-27 18:20:16 +0100482
483 # Construct 1 shape tensor to be used by all inserted reshape ops
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100484 new_shape_tens = create_const_tensor(op.name + "_reshape_shape", [1], DataType.int32, tens.shape)
Tim Hall79d07d22020-04-27 18:20:16 +0100485
486 for idx, out_tens in enumerate(op.outputs):
Tim Hall79d07d22020-04-27 18:20:16 +0100487 reshape_in = out_tens.clone("_reshaped")
Michael McGeagh6a8d4242020-07-28 12:17:59 +0100488 reshape_in.set_all_shapes(reshape_input_shape)
Tim Hall79d07d22020-04-27 18:20:16 +0100489 reshape_in.ops = [op]
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100490
491 reshape_op = Operation("Reshape", "{}{}_reshape".format(op.name, idx))
492 reshape_op.attrs["new_shape"] = reshape_input_shape
Tim Hall79d07d22020-04-27 18:20:16 +0100493 reshape_op.inputs = [reshape_in, new_shape_tens]
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100494 reshape_op.set_output_tensor(out_tens)
Tim Hall79d07d22020-04-27 18:20:16 +0100495
496 op.outputs[idx] = reshape_in
497
498 return tens
499
500
501def add_padding_fields(op, arch):
Jacob Bohlin90033f32020-08-28 15:45:44 +0200502 if op.run_on_npu:
503 if "padding" in op.attrs:
Michael McGeagh11b0bdb2020-09-08 11:07:35 +0100504 if op.type in conv_op | depthwise_op:
Jacob Bohlin90033f32020-08-28 15:45:44 +0200505 kernel_size = op.inputs[1].shape[:2]
506 input_shape = op.inputs[0].shape
Michael McGeagh11b0bdb2020-09-08 11:07:35 +0100507 elif op.type in pool_op | reduce_sum_ops:
Jacob Bohlin90033f32020-08-28 15:45:44 +0200508 kernel_size = op.attrs["ksize"][1:3]
509 input_shape = op.inputs[0].shape
510 elif op.type == "ExtractImagePatches":
511 kernel_size = op.attrs["ksizes"][1:3]
512 input_shape = op.inputs[0].shape
513 else:
514 raise UnsupportedFeatureError("Unknown operation that uses padding: {}".format(op.type))
Tim Hall79d07d22020-04-27 18:20:16 +0100515
Jacob Bohlin90033f32020-08-28 15:45:44 +0200516 if op.type == "Conv2DBackpropInputSwitchedBias":
517 upscaling_factor = op.outputs[0].shape[1] // input_shape[1]
518 padding, skirt = calc_upscaled_padding_and_skirt(
519 op.attrs["padding"], kernel_size, op.attrs["strides"], input_shape, upscaling_factor
520 )
521 else:
522 dilation_h, dilation_w = op.get_dilation_h_w()
523 dilated_kernel_size = [dilation_h * (kernel_size[0] - 1) + 1, dilation_w * (kernel_size[1] - 1) + 1]
524 padding, skirt = calc_padding_and_skirt(
525 op.attrs["padding"], dilated_kernel_size, op.attrs["strides"], input_shape
526 )
Jacob Bohlincf7da102020-05-20 09:03:40 +0200527
Jacob Bohlin90033f32020-08-28 15:45:44 +0200528 op.attrs["explicit_padding"] = padding
529 op.attrs["skirt"] = skirt
Jacob Bohlincf7da102020-05-20 09:03:40 +0200530
Tim Hall79d07d22020-04-27 18:20:16 +0100531 return op
532
533
Tim Hall79d07d22020-04-27 18:20:16 +0100534# Check if the op can be reordered
535def get_prepend_op(op):
536 inp = op.inputs[0]
537 # The op should be reordered between prev_op and prep_op
538 prev_op = inp.ops[-1]
539 prep_op = None
540 while prev_op.type in memory_only_ops and len(prev_op.outputs) == 1 and len(prev_op.outputs[0].consumers()) == 1:
541 prep_op = prev_op
542 inp = prev_op.inputs[0]
543 prev_op = inp.ops[-1]
Diego Russoea6111a2020-04-14 18:41:58 +0100544 if prev_op is not None and len(prev_op.outputs) == 1 and len(prev_op.outputs[0].consumers()) == 1:
Tim Hall79d07d22020-04-27 18:20:16 +0100545 return prep_op
546
547 return None
548
549
550def mark_npu_block_type(op, arch):
551 npu_block_type = NpuBlockType.Default
552 if op.type in conv_op:
553 npu_block_type = NpuBlockType.ConvolutionMxN
554 elif op.type in fc_op:
555 npu_block_type = NpuBlockType.VectorProduct
556 elif op.type in depthwise_op:
557 npu_block_type = NpuBlockType.ConvolutionDepthWise
558 elif op.type in pool_op:
559 npu_block_type = NpuBlockType.Pooling
560 elif op.type in elementwise_op:
561 npu_block_type = NpuBlockType.ElementWise
Fredrik Svedberga0c36242020-06-03 15:43:31 +0200562 elif op.type in reduce_sum_ops:
563 npu_block_type = NpuBlockType.ReduceSum
Tim Hall79d07d22020-04-27 18:20:16 +0100564
565 op.attrs["npu_block_type"] = npu_block_type
566 return op
567
568
569def convert_depthwise_to_conv(op, arch):
570 # Depthwise is equivalent to a single conv2d if the ifm depth is 1 and
571 # the ofm depth equals the depth multipler.
572 # If those conditions are true, then we can perform a simple
573 # switch of the operator type (and weight order)
574
Michael McGeagh11b0bdb2020-09-08 11:07:35 +0100575 if (op.type in depthwise_op) and (op.attrs["depth_multiplier"] != 1):
Tim Hall79d07d22020-04-27 18:20:16 +0100576 ifm_tensor = op.inputs[0]
577 weight_tensor = op.inputs[1]
578 ofm_tensor = op.outputs[0]
579 if (ifm_tensor.shape[3] == 1) and (ofm_tensor.shape[3] == op.attrs["depth_multiplier"]):
580 # Change op type to Conv2d
581 op.type = op.type.replace("DepthwiseConv2d", "Conv2D")
582 del op.attrs["channel_multiplier"]
583 del op.attrs["depth_multiplier"]
584
585 weight_tensor.quant_values = np.transpose(weight_tensor.quant_values, (0, 1, 3, 2))
Michael McGeagh6a8d4242020-07-28 12:17:59 +0100586 weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape))
Tim Hall79d07d22020-04-27 18:20:16 +0100587 else:
Louis Verhaard7db78962020-05-25 15:05:26 +0200588 raise UnsupportedFeatureError(
589 "Unsupported DepthwiseConv2d with depth_multiplier = {}, ifm channels = {}, ofm channels = {}".format(
Tim Hall79d07d22020-04-27 18:20:16 +0100590 op.attrs["depth_multiplier"], ifm_tensor.shape[3], ofm_tensor.shape[3]
591 )
592 )
Tim Hall79d07d22020-04-27 18:20:16 +0100593 return op
594
595
Jacob Bohline843d332020-06-23 12:12:56 +0200596def reorder_depthwise_weights(op, arch):
Michael McGeagh11b0bdb2020-09-08 11:07:35 +0100597 if op.type in depthwise_op:
Jacob Bohline843d332020-06-23 12:12:56 +0200598 weight_tensor = op.inputs[1]
599 weight_tensor.quant_values = np.transpose(weight_tensor.quant_values, (0, 1, 3, 2))
Michael McGeagh6a8d4242020-07-28 12:17:59 +0100600 weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape))
Jacob Bohline843d332020-06-23 12:12:56 +0200601 weight_tensor.weight_transpose_depthwise = True
602
603 return op
604
605
Michael McGeagh8d939c02020-07-29 13:11:43 +0100606def convert_conv_to_fc(op, arch):
607 # Conv 1x1 can be equivalent to Fully Connected.
608 # By representing certain convs as fully connected layers, Vela can better determine wether or not to use
609 # caching/double buffering for the weights.
610 # (Weights dont need to be reloaded for convs when IFM H and W are 1)
611 if op.type == "Conv2DBiasAct":
612 _, h, w, _ = op.inputs[0].shape
613 kh, kw, _, _ = op.inputs[1].shape
614 if h == 1 and w == 1 and kh == 1 and kw == 1:
615 # Overwrite this op as a Fully Connected Op
616 op.name += "_fc"
617 op.type = "FullyConnectedAct"
618 faf = op.attrs.get("fused_activation_function", None)
619 op.attrs = {
620 "fused_activation_function": faf,
621 "weights_format": 0,
622 "npu_block_type": NpuBlockType.VectorProduct,
623 }
624 # Reshape Weights to be 2D. HWIO becomes just IO (as H and W are 1, they can just be dropped)
625 weight_tensor = op.inputs[1]
626 weight_tensor.quant_values = weight_tensor.quant_values.squeeze(axis=(0, 1))
627 weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape))
628 # The output from a fully connected is expected to be 2D so we need to add a reshape layer to convert it
629 # back to 4D afterwards as the next layer is expecting that shape
630 orig_ofm_tensor = op.outputs[0]
631 # Reshape this ops output to be 2D: {(N*H*W), C} (We know N H and W are all 1 so this becomes {1, C})
632 fc_ofm_tensor = orig_ofm_tensor.clone("_fc")
633 fc_ofm_tensor.set_all_shapes([1, fc_ofm_tensor.shape[-1]])
634 fc_ofm_tensor.ops = [op]
635 # Add a reshape after the new OFM to convert it back to the original 4D shape
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100636 reshape_name = op.name + "_reshape"
637 new_shape_tens = create_const_tensor(reshape_name + "_shape", [1], DataType.int32, orig_ofm_tensor.shape)
Michael McGeagh8d939c02020-07-29 13:11:43 +0100638 reshape_op = Operation("Reshape", reshape_name)
Michael McGeagh8d939c02020-07-29 13:11:43 +0100639 reshape_op.attrs["new_shape"] = orig_ofm_tensor.shape
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100640 reshape_op.inputs = [fc_ofm_tensor, new_shape_tens]
641 reshape_op.set_output_tensor(orig_ofm_tensor)
Michael McGeagh8d939c02020-07-29 13:11:43 +0100642 # Replace this ops OFM to point to the 2D tensor
643 op.outputs[0] = fc_ofm_tensor
644 return op
645
646
Michael McGeagh8dbf8cf2020-09-08 11:09:48 +0100647def fixup_relus_with_differing_ifm_ofm_scaling(op, arch):
648 if op.run_on_npu and op.type in relu_ops:
649 ifm = op.inputs[0]
650 ofm = op.outputs[0]
651 # Relu with differing IFM and OFM scaling cannot be fused with another primary op
652 # and requires its own to be inserted
653 if not ifm.is_scaling_equal(ofm):
654 # Override this op with its own primary op (avgpool)
655 relu_fused_op = create_avgpool_nop(op.name + "_avgpool")
656 # And fuse the original activation function to it
657 relu_fused_op.attrs["fused_activation_function"] = op.type
658 # Tidy up and assign the ifm and ofm to the new op
659 ifm.consumer_list.remove(op)
660 relu_fused_op.add_input_tensor(ifm)
661 relu_fused_op.set_output_tensor(ofm)
662 op = relu_fused_op
663 return op
664
665
Tim Hall79d07d22020-04-27 18:20:16 +0100666# Reorder activation op if it's after the memory only operations
667def fixup_act_reorder(op, arch):
668 if op.type in activation_ops:
669 prep_op = get_prepend_op(op)
Diego Russoea6111a2020-04-14 18:41:58 +0100670 if prep_op is not None:
Tim Hall79d07d22020-04-27 18:20:16 +0100671 act_op = op.clone("_reordered")
Patrik Gustavssoncb337042020-09-16 14:55:40 +0200672
673 # There is only one input tensor, overwrite it
674 act_op.set_input_tensor(prep_op.inputs[0], 0)
675
Tim Hall79d07d22020-04-27 18:20:16 +0100676 act_op_out = act_op.inputs[0].clone("_acted")
677 act_op_out.quantization = op.outputs[0].quantization.clone()
Michael McGeaghc5b549b2020-08-07 11:54:28 +0100678 act_op.set_output_tensor(act_op_out)
Patrik Gustavssoncb337042020-09-16 14:55:40 +0200679
680 # Update the consumer list
681 act_op_out.consumer_list = op.outputs[0].consumer_list.copy()
682 act_op_out.consumer_list.append(prep_op)
683
Tim Hall79d07d22020-04-27 18:20:16 +0100684 prep_op.inputs[0] = act_op_out
685 prep_op.outputs[0].quantization = act_op_out.quantization.clone()
686
687 # Mark the op so that it will be removed as passthrough later on
688 op.type = "Identity"
689 return op
690
Louis Verhaarde0ef2732020-06-03 08:56:44 +0200691
Charles Xu78792222020-05-13 10:15:26 +0200692def fixup_elementwise_with_scalars(op, arch):
693 if op.type in binary_elementwise_op:
Louis Verhaarde0ef2732020-06-03 08:56:44 +0200694 ifm_tensor, ifm2_tensor, _, _ = op.get_ifm_ifm2_weights_ofm()
Charles Xu78792222020-05-13 10:15:26 +0200695 if ifm2_tensor.shape != [] and ifm_tensor.shape != []:
696 diff = len(ifm_tensor.shape) - len(ifm2_tensor.shape)
697 if diff > 0:
698 ifm2_tensor.shape = full_shape(len(ifm_tensor.shape), ifm2_tensor.shape, 1)
699 elif diff < 0:
700 ifm_tensor.shape = full_shape(len(ifm2_tensor.shape), ifm_tensor.shape, 1)
Louis Verhaarde0ef2732020-06-03 08:56:44 +0200701 elif ifm_tensor.shape == [] and ifm_tensor.quant_values is None:
702 # IFM is marked as a scalar, but is a result of an operation; change it to a shape of size 1
703 ifm_tensor.shape = len(ifm2_tensor.shape) * [1]
704 ifm_tensor.storage_shape = ifm_tensor.shape
705 elif ifm2_tensor.shape == [] and ifm2_tensor.quant_values is None:
706 # IFM2 is marked as a scalar, but is a result of an operation; change it to a shape of size 1
707 ifm2_tensor.shape = len(ifm_tensor.shape) * [1]
708 ifm2_tensor.storage_shape = ifm2_tensor.shape
Charles Xu78792222020-05-13 10:15:26 +0200709 return op
Tim Hall79d07d22020-04-27 18:20:16 +0100710
Louis Verhaarde0ef2732020-06-03 08:56:44 +0200711
Tim Hall4e127762020-05-15 16:05:49 +0100712# Set input/output tensor equivalence to the same id for memory operations
713def set_tensor_equivalence(op, arch):
Michael McGeagh11b0bdb2020-09-08 11:07:35 +0100714 if op.type in memory_only_ops:
Tim Hall4e127762020-05-15 16:05:49 +0100715 eid = op.outputs[0].equivalence_id
716 for inp in op.inputs:
717 inp.equivalence_id = eid
718 return op
719
720
Fredrik Svedberga0c36242020-06-03 15:43:31 +0200721def convert_softmax(op, arch):
722 if op.type == "Softmax" and op.run_on_npu:
723 softmax = SoftMax(op)
724 op = softmax.get_graph()
725 return op
726
727
Tim Hall79d07d22020-04-27 18:20:16 +0100728def convert_mul_max_to_abs_or_lrelu(op, arch):
Diego Russoea6111a2020-04-14 18:41:58 +0100729 r"""Whenever there is a subgraph with this topology:
Tim Hall79d07d22020-04-27 18:20:16 +0100730
731 Input X For X = -1 or X > 0
732 | \ / This subgraph can be replaced with either
733 | Mul an Abs (if X = -1) or a LeakyReLU (if X > 0)
734 | /
735 Max
736 """
737
738 if op.type == "Maximum":
739 # finds the Mul input(s) to the Max
740 muls = [i for i in op.inputs if i.ops[0].type == "MulAct"]
741 if len(muls) == 1:
742 mul = muls[0].ops[0]
743 elif len(muls) == 2:
744 # In the case both inputs are Muls, find the one with the same input as the Max
745 mul = [m for m in muls if len(set(op.inputs + m.ops[0].inputs)) == 1][0].ops[0]
746 else:
747 # No Mul inputs
748 return op
749
750 # make sure the Mul doesn't have any other consumers
Louis Verhaardd7911c42020-08-25 13:36:41 +0200751 mul_ofm = mul.outputs[0]
752 if len(mul_ofm.consumers()) != 1:
Tim Hall79d07d22020-04-27 18:20:16 +0100753 return op
754 # make sure the Mul doesn't have a faf
755 if mul.attrs["fused_activation_function"]:
756 return op
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200757 ifm, _, _, ofm = op.get_ifm_weights_biases_ofm()
758 if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype:
759 return op
Louis Verhaardd7911c42020-08-25 13:36:41 +0200760 if not ifm.is_scaling_equal(ofm) or not ifm.is_scaling_equal(mul_ofm):
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200761 # rewrite to LeakyRelu currently only makes sense if the quantization is identical
762 return op
Tim Hall79d07d22020-04-27 18:20:16 +0100763
764 # finds the branched input that goes to both the Max and the Mul
765 shared = set(op.inputs) & set(mul.inputs)
766 if len(shared) == 1:
767 shared_in = shared.pop()
768 # find the constant scalar input to the Mul
769 const_tens = (set(mul.inputs) - {shared_in}).pop()
770 # check that it is a scalar
771 if const_tens.shape != []:
772 return op
773 const = const_tens.ops[0]
774 # check that it is a constant
775 if const.type != "Const":
776 return op
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200777 # Remove the Mul from the shared input's consumers
778 shared_in.consumer_list.remove(mul)
Tim Hall79d07d22020-04-27 18:20:16 +0100779 else:
780 return op
781
782 val = const.outputs[0].values
783 if val >= 0:
784 new_op = "LeakyRelu"
785 op.attrs["alpha"] = val
Louis Verhaardd7911c42020-08-25 13:36:41 +0200786 # to produce bit exact results, the alpha is not enough;
787 # save additional scaling info in attr "alpha_scale", to be used as input
788 # to the LUT construction
789 alpha_scalar = const_tens.quant_values - const_tens.quantization.zero_point
790 mul_ifm_scale = np.double(ifm.quantization.scale_f32)
791 mul_ifm2_scale = np.double(const_tens.quantization.scale_f32)
792 mul_ofm_scale = np.double(mul_ofm.quantization.scale_f32)
793 alpha_scale, alpha_shift = scaling.elementwise_mul_scale(mul_ifm_scale, mul_ifm2_scale, mul_ofm_scale)
794 op.attrs["alpha_scaling"] = (alpha_scalar, alpha_scale, alpha_shift)
Tim Hall79d07d22020-04-27 18:20:16 +0100795 elif val == -1:
796 new_op = "Abs"
797 else:
798 return op
799
800 op.type = op.type.replace("Maximum", new_op)
801 op.name = op.name.replace("Maximum", new_op)
802 op.outputs[0].name = op.outputs[0].name.replace("Maximum", new_op)
803 op.inputs = [shared_in]
804 return op
805
806
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200807def convert_lrelu_to_mul_max(op, arch):
808 # Converts LeakyRelu to Max(alpha * IFM, identity * IFM)
809 # (the opposite of convert_mul_max_to_abs_or_lrelu)
810 ifm, _, _, ofm = op.get_ifm_weights_biases_ofm()
811
812 # Add multiplication with alpha
813 mul_alpha = Operation("MulAct", op.name + "_mul_alpha")
814 mul_alpha.add_input_tensor(ifm)
815 # Create const tensor containing alpha as scalar
816 alpha = op.attrs["alpha"]
817 quantization = ifm.quantization.clone()
818 quantization.min = 0
819 quantization.max = alpha * (quantization.quant_max - quantization.quant_min)
820 quantization.scale_f32 = alpha
821 quantization.zero_point = 0
822 alpha_tens = create_const_tensor(op.name + "_alpha_scalar", [], ifm.dtype, [1], np.int8, quantization=quantization)
823 mul_alpha.add_input_tensor(alpha_tens)
824 fm_alpha = ofm.clone(op.name + "_alpha")
825 mul_alpha.set_output_tensor(fm_alpha)
826
827 if ifm.is_scaling_equal(ofm):
828 # No identity multiplication is needed
829 fm_id = ifm
830 else:
831 # Add multiplication with identity
832 mul_identity = Operation("MulAct", op.name + "_mul_identity")
833 mul_identity.add_input_tensor(ifm)
834 # Create const tensor containing identity as scalar
835 quantization = ifm.quantization.clone()
836 quantization.min = 0
837 quantization.max = quantization.quant_max - quantization.quant_min
838 quantization.scale_f32 = 1
839 quantization.zero_point = 0
840 identity_tens = create_const_tensor(
841 op.name + "_id_scalar", [], ifm.dtype, [1], np.uint8, quantization=quantization
842 )
843 mul_identity.add_input_tensor(identity_tens)
844 fm_id = ofm.clone(op.name + "_id")
845 mul_identity.set_output_tensor(fm_id)
846
847 # Convert LeakyRelu to Max, add the results of the multiplication(s) as inputs
848 op.type = "Maximum"
849 op.name = op.name.replace("LeakyRelu", "Maximum")
850 op.inputs = []
851 ifm.consumer_list.remove(op)
852 op.add_input_tensor(fm_alpha)
853 op.add_input_tensor(fm_id)
854 return op
855
856
857def convert_lrelu_to_lut(op, arch):
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200858 # Rewrite LeakyRelu by Add with scalar 0 + LUT activation
Louis Verhaard58520b92020-08-24 16:45:38 +0200859 ifm, _, _, ofm = op.get_ifm_weights_biases_ofm()
860 assert ifm.dtype.size_in_bytes() == 1
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200861 op.type = "AddAct"
862 op.name = op.name + "_add"
863 op.attrs.update({"npu_block_type": NpuBlockType.ElementWise})
864 # Mark as no-op to enable potential fusing optimizations
865 op.attrs["is_nop"] = True
866 # Create an input tensor containing scalar zero
867 quantization = QuantizationParameters(0.0, 255.0)
Louis Verhaardd7911c42020-08-25 13:36:41 +0200868 quantization.scale_f32 = ifm.quantization.scale_f32
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200869 quantization.zero_point = 0
870 tens = create_const_tensor(op.inputs[0].name + "_add", [], ifm.dtype, [0], np.uint8, quantization=quantization)
871 op.add_input_tensor(tens)
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200872 # Generate the LUT
Louis Verhaardd7911c42020-08-25 13:36:41 +0200873 alpha = op.attrs["alpha"]
874 ifm_scale = np.double(ifm.quantization.scale_f32)
875 ofm_scale = np.double(ofm.quantization.scale_f32)
876 zp_in = ifm.quantization.zero_point
877 zp_out = ofm.quantization.zero_point
878 identity_scale, identity_shift = scaling.elementwise_mul_scale(ifm_scale, 1, ofm_scale)
879 alpha_scalar = 1
880 alpha_scale, alpha_shift = scaling.elementwise_mul_scale(ifm_scale, alpha, ofm_scale)
881 if "alpha_scaling" in op.attrs:
882 # The LeakyRelu was the result from convert_mul_max_to_abs_or_lrelu
883 alpha_scalar, alpha_scale, alpha_shift = op.attrs["alpha_scaling"]
884 values = []
Louis Verhaard58520b92020-08-24 16:45:38 +0200885 ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128)
Louis Verhaardd7911c42020-08-25 13:36:41 +0200886 quantized_min = min(ix)
887 quantized_max = max(ix)
888 for x in ix:
889 if x < zp_in:
890 lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(
891 alpha_scalar * (x - zp_in), alpha_scale, alpha_shift
892 )
893 else:
894 lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(x - zp_in, identity_scale, identity_shift)
895 lut_result = min(quantized_max, max(quantized_min, lut_result))
896 values.append(lut_result)
897 # The LUT must be applied without any preceding rescaling (the LUT itself performs the rescale),
898 # so even if the OFM has a different scale than the IFM, the generated OFM scale instructions
899 # should be the same as the IFM
900 op.attrs["forced_output_quantization"] = ifm.quantization
Louis Verhaard58520b92020-08-24 16:45:38 +0200901 lut_tensor = lut.create_lut_tensor(op.name + "_lut", values, DataType.int8)
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200902 op.set_activation_lut(lut_tensor)
903 return op
904
905
906def convert_lrelu(op, arch):
907 # Converts LeakyRelu to a LUT based solution if possible, otherwise a mul + max
908 if op.type != "LeakyRelu":
909 return op
910 ifm, _, _, ofm = op.get_ifm_weights_biases_ofm()
Louis Verhaardd7911c42020-08-25 13:36:41 +0200911 if ifm.dtype in (DataType.uint8, DataType.int8) and ifm.dtype == ofm.dtype:
912 # use LUT for int8/uint8
913 return convert_lrelu_to_lut(op, arch)
914 if ifm.is_scaling_equal(ofm) and ifm.dtype == ofm.dtype and ifm.dtype == DataType.int16:
915 # use LeakyRelu unmodified for int16 with equal input/output scaling
916 return op
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200917 return convert_lrelu_to_mul_max(op, arch)
918
919
Patrik Gustavssonfa4cb292020-09-10 08:19:36 +0200920def remove_unwanted_reshapes(op, arch):
921 # Try to remove reshapes enclosing ElementWise operator with only one non-constant input
922 if not op.run_on_npu or op.attrs["npu_block_type"] != NpuBlockType.ElementWise:
923 return op
924
925 # Check if the ElementWise operator only have one non-constant input
926 non_const_tens = [x for x in op.inputs if x.ops[0].type != "Const"]
927 if len(non_const_tens) != 1:
928 return op
929 ifm = non_const_tens[0]
930
931 # Check if operation is enclosed by Reshapes that can be removed
932 ofm = op.outputs[0]
933 prev_op = ifm.ops[0]
934 if (
935 len(ifm.consumer_list) == 1
936 and prev_op.type == "Reshape"
937 and len(ofm.consumer_list) == 1
938 and ofm.consumer_list[0].type == "Reshape"
939 ):
940 # Operation is enclosed by reshapes, check if they can be removed
941 prev_op_ifm, _, _, prev_op_ofm = prev_op.get_ifm_weights_biases_ofm()
942 cons_op = ofm.consumer_list[0]
943 cons_op_ifm = ofm
944 cons_op_ofm = cons_op.outputs[0]
945 if len(prev_op_ifm.shape) == len(cons_op_ofm.shape):
946 # Check if quantization is the same in the input and output for the reshape ops
947 if prev_op_ifm.quantization.is_scaling_equal(
948 prev_op_ofm.quantization
949 ) and cons_op_ifm.quantization.is_scaling_equal(cons_op_ofm.quantization):
950 op.inputs[0] = prev_op_ifm
951 op.outputs[0] = cons_op_ofm
952 return op
953
954
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200955def fuse_activation_function_with_prev(op, arch):
956 # if op is a no-op: attempts to move the activation function to the preceding op
957 if not op.attrs.get("is_nop", False) or op.attrs.get("fused_activation_function", None) is None:
958 return op
959 ifm, _, _, ofm = op.get_ifm_weights_biases_ofm()
960 # finds the input(s) to the operation
961 prev_op = ifm.ops[0]
962 # Note: the below checks on prev_op require that a first optimize pass on the full graph has been performed
963 fuse = (
964 prev_op.run_on_npu
965 and prev_op.attrs["npu_block_type"] != NpuBlockType.Default
966 and len(ifm.ops) == 1
967 and len(prev_op.outputs[0].consumers()) == 1
968 and prev_op.attrs.get("fused_activation_function", None) is None
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200969 )
970 if op.activation_lut is not None and arch.shram_reserved_unused_banks == 0:
971 # TODO: if SHRAM LUT space is shared with SHRAM ACC (32, 64 MAC),
972 # LUT currently only works correctly for elementwise ops
973 fuse = False
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200974 if not fuse:
975 return op
976 # Move the fused activation function + corresponding info to prev_op
Louis Verhaard98a34992020-09-01 10:39:04 +0200977 for attr in ("fused_activation_function", "forced_output_quantization"):
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200978 if attr in op.attrs:
979 prev_op.attrs[attr] = op.attrs[attr]
980 if op.activation_lut is not None:
981 prev_op.set_activation_lut(op.activation_lut)
982 # Bypass op
Louis Verhaard98a34992020-09-01 10:39:04 +0200983 prev_op.set_output_tensor(ofm)
Louis Verhaardb9fc33c2020-08-13 11:47:36 +0200984 return op
985
986
Dwight Lidman42fed942020-05-29 09:37:03 +0200987def add_attrs_to_resizebilinear(op, arch):
Tim Hallc30f4952020-06-15 20:47:35 +0100988 if op.type == "ResizeBilinear" and op.run_on_npu:
Dwight Lidman42fed942020-05-29 09:37:03 +0200989 input_tensor = op.inputs[0]
990 upscaled_shape = [input_tensor.shape[1] * 2, input_tensor.shape[2] * 2]
991 out_shape = op.outputs[0].shape[1:3]
992 if not op.attrs["align_corners"] and out_shape == upscaled_shape:
993 # this means the output is supposed to be a x2 upscale,
994 # so we need to do SAME padding
995 op.attrs["padding"] = b"SAME"
996 elif op.attrs["align_corners"] and out_shape == [upscaled_shape[0] - 1, upscaled_shape[1] - 1]:
997 # here we can just run the avg pool without padding and
998 # produce a (M * 2 - 1, N * 2 - 1) sized output
999 op.attrs["padding"] = b"VALID"
1000 else:
Charles Xu9a03fdf2020-07-02 15:12:40 +02001001 return op
Dwight Lidman42fed942020-05-29 09:37:03 +02001002 input_tensor.resampling_mode = resampling_mode.NEAREST
Tim Hallc30f4952020-06-15 20:47:35 +01001003 op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)})
Dwight Lidman42fed942020-05-29 09:37:03 +02001004 return op
1005
1006
Jacob Bohlina41cd4d2020-08-26 18:21:28 +02001007def fixup_bias_tensors(op, arch):
1008 if op.needs_bias() and not op.inputs[-1]:
1009 # Op has no bias, add bias tensor filled with zeros
1010 nr_biases = op.inputs[1].shape[-1]
1011 bias_values = [0] * nr_biases
1012 bias_tensor = create_const_tensor(op.name + "_bias", [nr_biases], DataType.int32, bias_values)
1013 bias_tensor.quant_values = bias_tensor.values
1014 op.set_input_tensor(bias_tensor, -1)
Jacob Bohlin67e0d8f2020-08-20 10:53:02 +02001015
1016 return op
1017
1018
Tim Hall79d07d22020-04-27 18:20:16 +01001019def supported_operator_check(op, arch):
1020 op.run_on_npu = arch.supported_operators.is_operator_supported(op)
1021 return op
1022
1023
1024def optimise_graph_a(nng, arch, verbose_graph=False):
1025 if verbose_graph:
1026 nng.print_graph()
1027
1028 op_rewrite_list = [
1029 # mark block type and check if the operations are supported
1030 mark_npu_block_type,
Tim Hall4e127762020-05-15 16:05:49 +01001031 set_tensor_equivalence,
Tim Hall79d07d22020-04-27 18:20:16 +01001032 supported_operator_check,
1033 # then do any rewrites of supported operators
1034 convert_depthwise_to_conv,
Michael McGeagh8d939c02020-07-29 13:11:43 +01001035 convert_conv_to_fc,
Fredrik Svedberga0c36242020-06-03 15:43:31 +02001036 convert_softmax,
Tim Hall79d07d22020-04-27 18:20:16 +01001037 fixup_fully_connected_input,
Patrik Gustavssoncb337042020-09-16 14:55:40 +02001038 convert_batched_fc_to_conv,
Tim Hall79d07d22020-04-27 18:20:16 +01001039 fixup_pack_input,
1040 fixup_conv2d_backprop,
Michael McGeagh8dbf8cf2020-09-08 11:09:48 +01001041 fixup_relus_with_differing_ifm_ofm_scaling,
Tim Hall79d07d22020-04-27 18:20:16 +01001042 fixup_act_reorder,
Tim Hall79d07d22020-04-27 18:20:16 +01001043 mark_npu_block_type,
Charles Xu78792222020-05-13 10:15:26 +02001044 fixup_elementwise_with_scalars,
Jacob Bohline843d332020-06-23 12:12:56 +02001045 reorder_depthwise_weights,
Charles Xu9a03fdf2020-07-02 15:12:40 +02001046 fixup_resizebilinear,
Jacob Bohlina41cd4d2020-08-26 18:21:28 +02001047 fixup_bias_tensors,
Dwight Lidmanc3862c22020-09-14 15:22:33 +02001048 convert_nop_split_to_identity,
Louis Verhaardb9fc33c2020-08-13 11:47:36 +02001049 convert_mul_max_to_abs_or_lrelu,
Patrik Gustavssonfa4cb292020-09-10 08:19:36 +02001050 remove_unwanted_reshapes,
Louis Verhaardb9fc33c2020-08-13 11:47:36 +02001051 convert_lrelu,
Tim Hall79d07d22020-04-27 18:20:16 +01001052 ]
1053
1054 for idx, sg in enumerate(nng.subgraphs):
1055 # rewrite graph pass
1056 nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
Diego Russoea6111a2020-04-14 18:41:58 +01001057 sg, arch, [fixup_unpack_output], op_rewrite_list, rewrite_unsupported=False
Tim Hall79d07d22020-04-27 18:20:16 +01001058 )
1059
1060 for idx, sg in enumerate(nng.subgraphs):
Louis Verhaardb9fc33c2020-08-13 11:47:36 +02001061 # remove passthrough tensors and attempt further optimizations
1062 nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
Charles Xu87c13502020-08-06 12:17:26 +02001063 sg, arch, [remove_passthrough_tensor], [fuse_activation_function_with_prev, add_padding_fields]
Louis Verhaardb9fc33c2020-08-13 11:47:36 +02001064 )
Tim Hall79d07d22020-04-27 18:20:16 +01001065
1066 if verbose_graph:
1067 nng.print_graph()
1068 return nng
1069
Diego Russoea6111a2020-04-14 18:41:58 +01001070
Tim Hall79d07d22020-04-27 18:20:16 +01001071def optimise_graph_b(nng, arch, verbose_graph=False):
1072 if verbose_graph:
1073 nng.print_graph()
1074
1075 for idx, sg in enumerate(nng.subgraphs):
1076 # combined rewrite graph pass
Diego Russoea6111a2020-04-14 18:41:58 +01001077 nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(sg, arch, [rewrite_concat, rewrite_split], [])
Tim Hall79d07d22020-04-27 18:20:16 +01001078
1079 if verbose_graph:
1080 nng.print_graph()
1081 return nng