Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1 | #!/usr/bin/env python3 |
| 2 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 3 | # Copyright (c) 2020-2021, ARM Limited. |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 4 | # |
| 5 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | # you may not use this file except in compliance with the License. |
| 7 | # You may obtain a copy of the License at |
| 8 | # |
| 9 | # http://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, |
| 13 | # WITHOUT 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. |
| 16 | |
| 17 | |
| 18 | import numpy as np |
| 19 | import argparse |
| 20 | import sys |
| 21 | import re |
| 22 | import os |
| 23 | import subprocess |
| 24 | import shlex |
| 25 | import json |
| 26 | import glob |
| 27 | import math |
| 28 | import queue |
| 29 | import threading |
| 30 | import traceback |
| 31 | import math |
| 32 | |
| 33 | from enum import IntEnum, Enum, unique |
| 34 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 35 | # Include the ../thirdparty/serialization_lib/python directory in PYTHONPATH |
| 36 | parent_dir = os.path.dirname(os.path.realpath(__file__)) |
| 37 | sys.path.append( |
| 38 | os.path.join(parent_dir, "..", "thirdparty", "serialization_lib", "python") |
| 39 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 40 | import tosa_serializer as ts |
| 41 | from tosa_serializer import * |
| 42 | import tosa |
| 43 | |
| 44 | # Convenience variables to the flatc-generated types that should be enums, but aren't |
| 45 | DType = tosa.DType.DType() |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 46 | Op = tosa.Op.Op() |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 47 | ResizeMode = tosa.ResizeMode.ResizeMode() |
| 48 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 49 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 50 | class TosaQuantGen: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 51 | """QuantizedInfo random generator helper functions. Specify with 'qgen': in the operator defintion""" |
| 52 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 53 | def __init__(self): |
| 54 | pass |
| 55 | |
| 56 | @staticmethod |
| 57 | def needsQinfo(op, dtype): |
Jared Smolens | 2a76ad2 | 2021-03-04 11:18:54 -0800 | [diff] [blame] | 58 | if dtype == DType.INT8 or dtype == DType.INT16: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 59 | return True |
| 60 | return False |
| 61 | |
| 62 | @staticmethod |
| 63 | def qgUnary(testGen, op, dtype): |
| 64 | qinfo = ts.TosaSerializerQuantInfo() |
| 65 | if TosaQuantGen.needsQinfo(op, dtype): |
| 66 | qinfo.UnaryQuantInfo(testGen.randInt(), testGen.randInt()) |
| 67 | else: |
| 68 | qinfo.UnaryQuantInfo(0, 0) |
| 69 | return qinfo |
| 70 | |
| 71 | @staticmethod |
| 72 | def qgConv(testGen, op, dtype): |
| 73 | qinfo = ts.TosaSerializerQuantInfo() |
| 74 | if TosaQuantGen.needsQinfo(op, dtype): |
| 75 | qinfo.ConvQuantInfo(testGen.randInt(), testGen.randInt()) |
| 76 | else: |
| 77 | qinfo.ConvQuantInfo(0, 0) |
| 78 | return qinfo |
| 79 | |
| 80 | @staticmethod |
| 81 | def qgMatmul(testGen, op, dtype): |
| 82 | qinfo = ts.TosaSerializerQuantInfo() |
| 83 | if TosaQuantGen.needsQinfo(op, dtype): |
| 84 | qinfo.MatMulQuantInfo(testGen.randInt(), testGen.randInt()) |
| 85 | else: |
| 86 | qinfo.MatMulQuantInfo(0, 0) |
| 87 | return qinfo |
| 88 | |
| 89 | @staticmethod |
| 90 | def qgPad(testGen, op, dtype): |
| 91 | qinfo = ts.TosaSerializerQuantInfo() |
| 92 | if TosaQuantGen.needsQinfo(op, dtype): |
| 93 | qinfo.PadQuantInfo(testGen.randInt()) |
| 94 | else: |
| 95 | qinfo.PadQuantInfo(0) |
| 96 | return qinfo |
| 97 | |
| 98 | @staticmethod |
| 99 | def computeMultiplierAndShift(scaleFp, scale32): |
| 100 | # Derived from computeMultiplierAndShiftTosaScale32 |
| 101 | # Provide a floating-point scaling factor and the scale32 parameter |
| 102 | # to compute the multiplier and shift |
| 103 | |
| 104 | if scale32: |
| 105 | scaleBits = 31 |
| 106 | else: |
| 107 | scaleBits = 15 |
| 108 | |
| 109 | m, shift = math.frexp(scaleFp) |
| 110 | |
| 111 | if scaleFp < 0.0: |
| 112 | m = -m |
| 113 | |
| 114 | multiplier = round(m * (1 << scaleBits)) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 115 | assert multiplier <= (1 << scaleBits) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 116 | |
| 117 | if multiplier == (1 << scaleBits): |
| 118 | multiplier = multiplier // 2 |
| 119 | shift = shift + 1 |
| 120 | |
| 121 | shift = (-shift) + scaleBits |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 122 | # print('scalefp {} scaleBits {} m {} mult {} shift {}'.format(scaleFp, scaleBits, m, multiplier, shift)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 123 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 124 | assert multiplier <= (1 << scaleBits) |
| 125 | assert shift >= 0 and shift <= 63 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 126 | |
| 127 | return multiplier, shift |
| 128 | |
| 129 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 130 | class TosaTensorGen: |
| 131 | """Tensor generators create a shape list for the placeholder and const tensor |
| 132 | data operands for the operator. The actual random data is generated separately for each test.""" |
| 133 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 134 | def __init__(self): |
| 135 | pass |
| 136 | |
| 137 | @staticmethod |
| 138 | def tgBasic(testGen, opName, rank): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 139 | pl, const = opName["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 140 | shape = testGen.makeShape(rank) |
| 141 | |
| 142 | shape_list = [] |
| 143 | for i in range(pl + const): |
| 144 | shape_list.append(shape.copy()) |
| 145 | |
| 146 | return shape_list |
| 147 | |
| 148 | @staticmethod |
| 149 | def tgNHWC(testGen, opName, rank): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 150 | pl, const = opName["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 151 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 152 | assert rank == 4 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 153 | |
| 154 | shape = testGen.makeShape(rank) |
| 155 | |
| 156 | # Constrict the batch size? |
| 157 | if testGen.args.max_batch_size: |
| 158 | shape[0] = (shape[0] % testGen.args.max_batch_size) + 1 |
| 159 | |
| 160 | shape_list = [] |
| 161 | for i in range(pl + const): |
| 162 | shape_list.append(shape.copy()) |
| 163 | |
| 164 | return shape_list |
| 165 | |
| 166 | @staticmethod |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 167 | def tgScatter(testGen, opName, rank): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 168 | pl, const = opName["operands"] |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 169 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 170 | assert pl == 2 |
| 171 | assert const == 0 |
| 172 | assert rank == 3 |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 173 | |
| 174 | values_in_shape = testGen.makeShape(rank) |
| 175 | |
| 176 | # Constrict the batch size? |
| 177 | if testGen.args.max_batch_size: |
| 178 | values_in_shape[0] = (values_in_shape[0] % testGen.args.max_batch_size) + 1 |
| 179 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 180 | W = testGen.randInt( |
| 181 | testGen.args.tensor_shape_range[0], testGen.args.tensor_shape_range[1] |
| 182 | ) |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 183 | input_shape = [values_in_shape[0], W, values_in_shape[2]] |
| 184 | |
| 185 | shape_list = [] |
| 186 | shape_list.append(values_in_shape.copy()) |
| 187 | shape_list.append(input_shape.copy()) |
| 188 | |
| 189 | return shape_list |
| 190 | |
| 191 | @staticmethod |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 192 | def tgBroadcastFuzz(testGen, op, rank): |
| 193 | shape = testGen.makeShape(rank) |
| 194 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 195 | pl, const = op["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 196 | |
| 197 | shape_list = [] |
| 198 | |
| 199 | # Choose one of the inputs to broadcast |
| 200 | bcast_idx = testGen.randInt(0, pl + const) |
| 201 | for i in range(pl + const): |
| 202 | shape_bcast = shape.copy() |
| 203 | |
| 204 | # If the chosen input, pick a random index to broadcast |
| 205 | if i == bcast_idx: |
| 206 | fuzz_idx = testGen.randInt(0, rank) |
| 207 | shape_bcast[fuzz_idx] = 1 |
| 208 | |
| 209 | shape_list.append(shape_bcast) |
| 210 | |
| 211 | return shape_list |
| 212 | |
| 213 | @staticmethod |
| 214 | def tgConv2D(testGen, op, rank): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 215 | pl, const = op["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 216 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 217 | assert rank == 4 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 218 | |
| 219 | # IFM dimensions are NHWC |
| 220 | ifm_shape = testGen.makeShape(rank) |
| 221 | |
| 222 | # Constrict the batch size? |
| 223 | if testGen.args.max_batch_size: |
| 224 | ifm_shape[0] = (ifm_shape[0] % testGen.args.max_batch_size) + 1 |
| 225 | |
| 226 | # Get the filter height/width from the operator parameters |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 227 | filter_hw = op["filter"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 228 | |
| 229 | # Generate a random OFM depth |
| 230 | ofm_depth = testGen.makeShape(1)[0] |
| 231 | |
| 232 | # The filter dimensions are OHWI |
| 233 | filter_shape = np.asarray([ofm_depth, filter_hw[0], filter_hw[1], ifm_shape[3]]) |
| 234 | |
| 235 | # The bias is OC |
| 236 | bias_shape = np.asarray([ofm_depth]) |
| 237 | |
| 238 | return [ifm_shape, filter_shape, bias_shape] |
| 239 | |
| 240 | @staticmethod |
| 241 | def tgTransposeConv2D(testGen, op, rank): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 242 | pl, const = op["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 243 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 244 | assert rank == 4 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 245 | |
| 246 | # IFM dimensions are NHWC |
| 247 | ifm_shape = testGen.makeShape(rank) |
| 248 | |
| 249 | # Constrict the batch size? |
| 250 | if testGen.args.max_batch_size: |
| 251 | ifm_shape[0] = (ifm_shape[0] % testGen.args.max_batch_size) + 1 |
| 252 | |
| 253 | # Get the filter height/width from the operator parameters |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 254 | filter_hw = op["filter"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 255 | |
| 256 | # Generate a random OFM depth |
| 257 | ofm_depth = testGen.makeShape(1)[0] |
| 258 | |
| 259 | # The filter dimensions are OHWI |
| 260 | filter_shape = np.asarray([ofm_depth, filter_hw[0], filter_hw[1], ifm_shape[3]]) |
| 261 | |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 262 | # The bias is OC |
| 263 | bias_shape = np.asarray([ofm_depth]) |
| 264 | |
| 265 | return [ifm_shape, filter_shape, bias_shape] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 266 | |
| 267 | @staticmethod |
| 268 | def tgDepthwiseConv2D(testGen, op, rank): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 269 | pl, const = op["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 270 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 271 | assert rank == 4 |
| 272 | assert pl == 1 and const == 2 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 273 | |
| 274 | # IFM dimensions are NHWC |
| 275 | ifm_shape = testGen.makeShape(rank) |
| 276 | |
| 277 | # Constrict the batch size? |
| 278 | if testGen.args.max_batch_size: |
| 279 | ifm_shape[0] = (ifm_shape[0] % testGen.args.max_batch_size) + 1 |
| 280 | |
| 281 | # Get the filter height/width from the operator parameters |
| 282 | # Filter is KH, HW, C, M |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 283 | filter_hw = op["filter"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 284 | |
| 285 | # Generate a random OFM depth, but don't let it get too big because |
| 286 | # the output depth is M * C |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 287 | filter_m = ( |
| 288 | testGen.makeShape(1)[0] % (testGen.args.tensor_shape_range[1] // 4) |
| 289 | ) + 1 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 290 | |
| 291 | # The filter dimensions are HWCM |
| 292 | filter_shape = np.asarray([filter_hw[0], filter_hw[1], ifm_shape[3], filter_m]) |
| 293 | |
| 294 | # The bias is M * C |
| 295 | bias_shape = np.asarray([ifm_shape[3] * filter_m]) |
| 296 | |
| 297 | return [ifm_shape, filter_shape, bias_shape] |
| 298 | |
| 299 | @staticmethod |
| 300 | def tgFullyConnected(testGen, op, rank): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 301 | pl, const = op["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 302 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 303 | assert rank == 2 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 304 | |
| 305 | input_shape = testGen.makeShape(rank) |
| 306 | filter_oc = testGen.makeShape(1)[0] |
| 307 | filter_shape = np.asarray([filter_oc, input_shape[1]]) |
| 308 | |
| 309 | bias_shape = np.asarray([filter_oc]) |
| 310 | |
| 311 | return [input_shape, filter_shape, bias_shape] |
| 312 | |
| 313 | @staticmethod |
| 314 | def tgMatmul(testGen, op, rank): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 315 | pl, const = op["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 316 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 317 | assert rank == 2 |
| 318 | assert pl == 2 and const == 0 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 319 | |
| 320 | a_shape = testGen.makeShape(rank) |
| 321 | b_oc = testGen.makeShape(1)[0] |
| 322 | b_shape = np.asarray([a_shape[1], b_oc]) |
| 323 | |
| 324 | return [a_shape, b_shape] |
| 325 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 326 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 327 | class TosaArgGen: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 328 | """Argument generators create exhaustive or random lists of attributes for operators that take |
| 329 | attributes or other parameters. The return value is a list of (descriptive_name, [arglist]) |
| 330 | tuples where the descriptive_name is appended to the test name and the arglist is expanded |
| 331 | as arguments to the operator build function.""" |
| 332 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 333 | def __init__(self): |
| 334 | pass |
| 335 | |
| 336 | @staticmethod |
| 337 | def agNone(testGen, opName, shapeList, dtype): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 338 | """A trivial argument generator for operators that don't take any |
| 339 | non-tensor arguments""" |
| 340 | return [("", [])] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 341 | |
| 342 | @staticmethod |
| 343 | def agAxis(testGen, opName, shapeList, dtype): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 344 | """Build the axis argument for operators that take a single axis""" |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 345 | axes = [] |
| 346 | |
| 347 | shape = shapeList[0] |
| 348 | |
| 349 | for a in range(0, len(shape)): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 350 | axes.append(("axis_{}".format(a), [a])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 351 | return axes |
| 352 | |
| 353 | @staticmethod |
| 354 | def agConv2D(testGen, opName, shapeList, dtype): |
| 355 | arg_list = [] |
| 356 | |
| 357 | ifm_shape = shapeList[0] |
| 358 | filter_shape = shapeList[1] |
| 359 | |
| 360 | # Must be rank 4 |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 361 | assert len(ifm_shape) == 4 |
| 362 | assert len(filter_shape) == 4 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 363 | |
| 364 | maxStride = testGen.args.max_conv_stride |
| 365 | maxPadding = testGen.args.max_conv_padding + 1 |
| 366 | maxDilation = testGen.args.max_conv_dilation |
| 367 | |
| 368 | # Strides, padding, dilations |
| 369 | for stride in range(0, maxStride ** 2): |
| 370 | for padding in range(0, (maxPadding) ** 4): |
| 371 | for dilation in range(0, maxDilation ** 2): |
| 372 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 373 | s = [stride // maxStride + 1, stride % maxStride + 1] |
| 374 | p = [ |
| 375 | (padding // (maxPadding * 4)) % maxPadding, |
| 376 | (padding // (maxPadding * 2)) % maxPadding, |
| 377 | (padding // (maxPadding * 1)) % maxPadding, |
| 378 | padding % maxPadding, |
| 379 | ] |
| 380 | d = [dilation // maxDilation + 1, dilation % maxDilation + 1] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 381 | |
| 382 | # 4 padding parameters for regular conv2d |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 383 | arg_list.append( |
| 384 | ( |
| 385 | "st{}{}_pad{}{}{}{}_dilat{}{}".format( |
| 386 | s[0], s[1], p[0], p[1], p[2], p[3], d[0], d[1] |
| 387 | ), |
| 388 | [s, p, d], |
| 389 | ) |
| 390 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 391 | return arg_list |
| 392 | |
| 393 | @staticmethod |
| 394 | def agTransposeConv2D(testGen, opName, shapeList, dtype): |
| 395 | arg_list = [] |
| 396 | |
| 397 | ifm_shape = shapeList[0] |
| 398 | filter_shape = shapeList[1] |
| 399 | |
| 400 | # Must be rank 4 |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 401 | assert len(ifm_shape) == 4 |
| 402 | assert len(filter_shape) == 4 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 403 | |
| 404 | maxStride = testGen.args.max_conv_stride |
| 405 | maxPadding = testGen.args.max_conv_padding + 1 |
| 406 | maxDilation = testGen.args.max_conv_dilation |
| 407 | |
| 408 | # Strides, padding, dilations |
| 409 | for stride in range(0, maxStride ** 2): |
| 410 | for out_padding in range(0, (maxPadding) ** 2): |
| 411 | for dilation in range(0, maxDilation ** 2): |
| 412 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 413 | s = [stride // maxStride + 1, stride % maxStride + 1] |
| 414 | p = [ |
| 415 | (out_padding // (maxPadding * 1)) % maxPadding, |
| 416 | out_padding % maxPadding, |
| 417 | ] |
| 418 | d = [dilation // maxDilation + 1, dilation % maxDilation + 1] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 419 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 420 | oh = ( |
| 421 | ifm_shape[1] |
| 422 | - filter_shape[1] |
| 423 | - (filter_shape[1] - 1) * (d[0] - 1) |
| 424 | + 2 * p[0] |
| 425 | ) // s[0] + 1 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 426 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 427 | ow = ( |
| 428 | ifm_shape[2] |
| 429 | - filter_shape[2] |
| 430 | - (filter_shape[2] - 1) * (d[1] - 1) |
| 431 | + 2 * p[1] |
| 432 | ) // s[1] + 1 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 433 | |
| 434 | # Output shape |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 435 | os = [ifm_shape[0], oh, ow, filter_shape[0]] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 436 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 437 | arg_list.append( |
| 438 | ( |
| 439 | "st{}{}_outpad{}{}_dilat{}{}_os{}x{}x{}x{}".format( |
| 440 | s[0], |
| 441 | s[1], |
| 442 | p[0], |
| 443 | p[1], |
| 444 | d[0], |
| 445 | d[1], |
| 446 | os[0], |
| 447 | os[1], |
| 448 | os[2], |
| 449 | os[3], |
| 450 | ), |
| 451 | [s, p, d, os], |
| 452 | ) |
| 453 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 454 | |
| 455 | return arg_list |
| 456 | |
| 457 | @staticmethod |
| 458 | def agPad(testGen, opName, shapeList, dtype): |
| 459 | arg_list = [] |
| 460 | rank = len(shapeList[0]) |
| 461 | |
| 462 | # Exhaustively test combinations of 0/1 padding on each side of each dimension |
| 463 | # This process might need some revision for >1 padding, but use rank**2 as a bitmask |
| 464 | # for now |
| 465 | for v in range(rank ** 2): |
| 466 | |
| 467 | # Create a flat arraypadding4D |
| 468 | paddings = np.zeros((rank * 2), dtype=np.int32) |
| 469 | |
| 470 | # Fill in the 1's |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 471 | for r in range(rank * 2): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 472 | if (v >> r) & 1: |
| 473 | paddings[r] = 1 |
| 474 | |
| 475 | # Reshape back to a 2D array |
| 476 | paddings = paddings.reshape((rank, 2)) |
| 477 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 478 | arg_list.append(("pad{0:b}".format(v), [paddings])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 479 | |
| 480 | return arg_list |
| 481 | |
| 482 | @staticmethod |
| 483 | def agPooling(testGen, opName, shapeList, dtype): |
| 484 | arg_list = [] |
| 485 | |
| 486 | shape = shapeList[0] |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 487 | assert len(shape) == 4 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 488 | |
| 489 | maxStride = testGen.args.max_pooling_stride |
| 490 | maxKernel = testGen.args.max_pooling_kernel |
| 491 | maxPadding = testGen.args.max_pooling_padding + 1 |
| 492 | |
| 493 | for kernel in range(0, maxKernel ** 2): |
| 494 | for stride in range(0, maxStride ** 2): |
| 495 | for padding in range(0, maxPadding ** 4): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 496 | s = [stride // maxStride + 1, stride % maxStride + 1] |
| 497 | k = [(kernel // maxKernel) + 2, (kernel % maxKernel) + 2] |
| 498 | p = [ |
| 499 | (padding // (maxPadding * 4)) % maxPadding, |
| 500 | (padding // (maxPadding * 2)) % maxPadding, |
| 501 | (padding // (maxPadding * 1)) % maxPadding, |
| 502 | padding % maxPadding, |
| 503 | ] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 504 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 505 | arg_list.append( |
| 506 | ( |
| 507 | "st{}{}_kern{}{}_pad{}{}{}{}".format( |
| 508 | s[0], s[1], k[0], k[1], p[0], p[1], p[2], p[3] |
| 509 | ), |
| 510 | [k, s, p], |
| 511 | ) |
| 512 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 513 | return arg_list |
| 514 | |
| 515 | @staticmethod |
| 516 | def agCast(testGen, opName, shapeList, inDtype): |
| 517 | arg_list = [] |
| 518 | |
| 519 | # Enumerate the output types here |
| 520 | if inDtype == DType.INT8: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 521 | dtypeList = [DType.BOOL, DType.INT16, DType.INT32, DType.FLOAT] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 522 | elif inDtype == DType.INT16: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 523 | dtypeList = [DType.BOOL, DType.INT8, DType.INT32, DType.FLOAT] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 524 | elif inDtype == DType.INT32: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 525 | dtypeList = [DType.BOOL, DType.INT8, DType.INT16, DType.FLOAT] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 526 | elif inDtype == DType.BOOL: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 527 | dtypeList = [DType.INT8, DType.INT16, DType.INT32] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 528 | elif inDtype == DType.FLOAT: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 529 | dtypeList = [DType.INT8, DType.INT16, DType.INT32] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 530 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 531 | raise Exception("Unexpected input dtype: {}".format(inDtype)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 532 | |
| 533 | for dtype in dtypeList: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 534 | arg_list.append(("out{}".format(DTypeNames[dtype]), [dtype])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 535 | |
| 536 | return arg_list |
| 537 | |
| 538 | @staticmethod |
| 539 | def agRescale(testGen, opName, shapeList, inDtype): |
| 540 | arg_list = [] |
| 541 | |
| 542 | # Enumerate the output types here |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 543 | for dtype in [DType.INT8, DType.INT16, DType.INT32]: |
| 544 | for scale32 in [False, True]: |
| 545 | for double_round in [False, True]: |
| 546 | for per_channel in [False, True]: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 547 | |
| 548 | if inDtype == DType.INT48 and scale32: |
| 549 | # Illegal condition. Must be scale32=False |
| 550 | continue |
| 551 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 552 | arg_list.append( |
| 553 | ( |
| 554 | "out{}_sc{}_dr{}_pc{}".format( |
| 555 | DTypeNames[dtype], |
| 556 | int(scale32), |
| 557 | int(double_round), |
| 558 | int(per_channel), |
| 559 | ), |
| 560 | [dtype, scale32, double_round, per_channel], |
| 561 | ) |
| 562 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 563 | |
| 564 | return arg_list |
| 565 | |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 566 | @staticmethod |
| 567 | def agMul(testGen, opName, shapeList, dtype): |
| 568 | arg_list = [] |
| 569 | |
| 570 | if dtype is DType.INT32: |
| 571 | for p in range(testGen.args.num_rand_permutations): |
| 572 | |
| 573 | shift = testGen.randInt(0, 32) |
| 574 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 575 | arg_list.append(("perm{}_shift{}".format(p, shift), [shift])) |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 576 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 577 | arg_list.append(("shift0", [0])) |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 578 | |
| 579 | return arg_list |
| 580 | |
| 581 | @staticmethod |
| 582 | def agArithmeticRightShift(testGen, opName, shapeList, dtype): |
| 583 | arg_list = [] |
| 584 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 585 | arg_list.append(("roundTrue", [True])) |
| 586 | arg_list.append(("roundFalse", [False])) |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 587 | |
| 588 | return arg_list |
| 589 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 590 | # Helper function for reshape. Gets some factors of a larger number. |
| 591 | @staticmethod |
| 592 | def getFactors(val, start=1): |
| 593 | factors = [] |
| 594 | |
| 595 | for i in range(start, int(np.sqrt(val))): |
| 596 | if (val % i) == 0: |
| 597 | factors.append(i) |
| 598 | |
| 599 | return factors |
| 600 | |
| 601 | @staticmethod |
| 602 | def agReshape(testGen, opName, shapeList, dtype): |
| 603 | arg_list = [] |
| 604 | |
| 605 | origShape = shapeList[0] |
| 606 | |
| 607 | totalElements = 1 |
| 608 | for s in origShape: |
| 609 | totalElements *= s |
| 610 | |
| 611 | # This code is NOT fast. Fortunately, the numbers are fairly small. |
| 612 | factors = TosaArgGen.getFactors(totalElements) |
| 613 | |
| 614 | for p in range(testGen.args.num_rand_permutations): |
| 615 | newRank = testGen.randInt(1, 6) |
| 616 | newShape = [] |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 617 | if len(factors) < newRank: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 618 | continue |
| 619 | |
| 620 | remainingElements = totalElements |
| 621 | shuffledFactors = testGen.rng.permutation(factors) |
| 622 | for i in range(newRank): |
| 623 | # pick rank-1 factors |
| 624 | newShape.append(shuffledFactors[0]) |
| 625 | remainingElements = remainingElements // shuffledFactors[0] |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 626 | shuffledFactors = testGen.rng.permutation( |
| 627 | TosaArgGen.getFactors(remainingElements) |
| 628 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 629 | newShape.append(remainingElements) |
| 630 | |
| 631 | # Toss in a -1 sometimes |
| 632 | minusOne = testGen.randInt(0, newRank * 4) |
| 633 | if minusOne < newRank: |
| 634 | newShape[minusOne] = -1 |
| 635 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 636 | arg_list.append(("perm{}_rank{}".format(p, newRank), [newShape])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 637 | |
| 638 | return arg_list |
| 639 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 640 | @staticmethod |
| 641 | def agTranspose(testGen, opName, shapeList, dtype): |
| 642 | arg_list = [] |
| 643 | |
| 644 | ifm_shape = shapeList[0] |
| 645 | |
| 646 | perms = range(len(ifm_shape)) |
| 647 | for p in range(testGen.args.num_rand_permutations): |
| 648 | perms = np.int32(testGen.rng.permutation(perms)).tolist() |
| 649 | |
| 650 | # Avoid duplicates |
| 651 | found = False |
| 652 | for name, other_perm in arg_list: |
| 653 | if other_perm[0] == perms: |
| 654 | found = True |
| 655 | break |
| 656 | |
| 657 | if not found: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 658 | arg_list.append(("perm{}".format(p), [perms])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 659 | |
| 660 | return arg_list |
| 661 | |
| 662 | @staticmethod |
| 663 | def agSlice(testGen, opName, shapeList, dtype): |
| 664 | arg_list = [] |
| 665 | |
| 666 | ifm_shape = shapeList[0] |
| 667 | rank = len(ifm_shape) |
| 668 | |
| 669 | for p in range(testGen.args.num_rand_permutations): |
| 670 | begin = [] |
| 671 | size = [] |
| 672 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 673 | valid = True |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 674 | |
| 675 | for i in range(rank): |
| 676 | if ifm_shape[i] > 1: |
| 677 | begin.append(testGen.randInt(0, ifm_shape[i])) |
| 678 | size.append(testGen.randInt(0, ifm_shape[i] - begin[i])) |
| 679 | |
| 680 | # Invalid slice size? |
| 681 | if size[i] == 0: |
| 682 | valid = False |
| 683 | else: |
| 684 | begin.append(0) |
| 685 | size.append(1) |
| 686 | |
| 687 | if valid: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 688 | arg_list.append(("perm{}".format(p), [begin, size])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 689 | return arg_list |
| 690 | |
| 691 | @staticmethod |
| 692 | def agTile(testGen, opName, shapeList, dtype): |
| 693 | arg_list = [] |
| 694 | |
| 695 | ifm_shape = shapeList[0] |
| 696 | rank = len(ifm_shape) |
| 697 | |
| 698 | for p in range(testGen.args.num_rand_permutations): |
| 699 | |
| 700 | # Pick a few random, but small multiple values |
| 701 | # because otherwise this has a tendency to generate |
| 702 | # enormous tensors |
| 703 | multiples = [] |
| 704 | for i in range(rank): |
| 705 | multiples.append(testGen.randInt(1, 4)) |
| 706 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 707 | arg_list.append(("perm{}".format(p), [multiples])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 708 | |
| 709 | return arg_list |
| 710 | |
| 711 | @staticmethod |
| 712 | def agResize(testGen, opName, shapeList, dtype): |
| 713 | arg_list = [] |
| 714 | |
| 715 | ifm_shape = shapeList[0] |
| 716 | |
| 717 | for m in [ResizeMode.NEAREST, ResizeMode.BILINEAR]: |
| 718 | |
| 719 | # Exclude illegal {mode, type} configurations. Pick legal output types |
| 720 | if m == ResizeMode.NEAREST and dtype == DType.INT8: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 721 | outputDTypeList = [DType.INT32] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 722 | elif m == ResizeMode.NEAREST and dtype == DType.INT16: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 723 | outputDTypeList = [DType.INT16] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 724 | elif m == ResizeMode.BILINEAR and dtype == DType.INT8: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 725 | outputDTypeList = [DType.INT8] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 726 | elif m == ResizeMode.BILINEAR and dtype == DType.INT16: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 727 | outputDTypeList = [DType.INT48] |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 728 | elif dtype == DType.FLOAT: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 729 | outputDTypeList = [DType.FLOAT] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 730 | else: |
| 731 | continue |
| 732 | |
| 733 | for outputDType in outputDTypeList: |
| 734 | for perm in range(testGen.args.num_rand_permutations): |
| 735 | |
| 736 | # Randomly generate legal output dimensions and shift |
| 737 | # and then compute the stride and offset based on them |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 738 | output_dims = [testGen.randInt(1), testGen.randInt(1)] |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 739 | in_center_h = (ifm_shape[1] - 1) / 2.0 |
| 740 | in_center_w = (ifm_shape[2] - 1) / 2.0 |
| 741 | out_center_h = (output_dims[0] - 1) / 2.0 |
| 742 | out_center_w = (output_dims[1] - 1) / 2.0 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 743 | |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 744 | fp_stride_y = float(ifm_shape[1]) / float(output_dims[0]) |
| 745 | fp_stride_x = float(ifm_shape[2]) / float(output_dims[1]) |
| 746 | fp_offset_y = in_center_h - fp_stride_y * out_center_h |
| 747 | fp_offset_x = in_center_w - fp_stride_x * out_center_w |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 748 | |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 749 | if outputDType == DType.FLOAT: |
| 750 | shift = 0 |
| 751 | stride = [0, 0] |
| 752 | offset = [0, 0] |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 753 | stride_fp = [fp_stride_y, fp_stride_x] |
| 754 | offset_fp = [fp_offset_y, fp_offset_x] |
| 755 | arg_list.append( |
| 756 | ( |
| 757 | "mode{}_odim{}x{}_out{}_st{:.2f}x{:.2f}_off{:.2f}x{:.2f}".format( |
| 758 | m, |
| 759 | output_dims[0], |
| 760 | output_dims[1], |
| 761 | testGen.typeStr(outputDType), |
| 762 | stride_fp[0], |
| 763 | stride_fp[1], |
| 764 | offset_fp[0], |
| 765 | offset_fp[1], |
| 766 | ), |
| 767 | [ |
| 768 | m, |
| 769 | stride, |
| 770 | offset, |
| 771 | shift, |
| 772 | stride_fp, |
| 773 | offset_fp, |
| 774 | output_dims, |
| 775 | dtype, |
| 776 | outputDType, |
| 777 | ], |
| 778 | ) |
| 779 | ) |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 780 | else: |
| 781 | shift = 11 |
| 782 | unit = float(1 << shift) |
| 783 | stride_y = int(round(fp_stride_y * unit)) |
| 784 | stride_x = int(round(fp_stride_x * unit)) |
| 785 | offset_y = int(round(fp_offset_y * unit)) |
| 786 | offset_x = int(round(fp_offset_x * unit)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 787 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 788 | while ( |
| 789 | stride_y >= 32768 |
| 790 | or stride_x >= 32768 |
| 791 | or offset_y >= 32768 |
| 792 | or offset_x >= 32768 |
| 793 | or offset_y < -32768 |
| 794 | or offset_x < -32768 |
| 795 | ): |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 796 | shift = shift - 1 |
| 797 | unit = float(1 << shift) |
| 798 | stride_y = int(round(fp_stride_y * unit)) |
| 799 | stride_x = int(round(fp_stride_x * unit)) |
| 800 | offset_y = int(round(fp_offset_y * unit)) |
| 801 | offset_x = int(round(fp_offset_x * unit)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 802 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 803 | stride = [stride_y, stride_x] |
| 804 | offset = [offset_y, offset_x] |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 805 | |
| 806 | stride_fp = [0.0, 0.0] |
| 807 | offset_fp = [0.0, 0.0] |
| 808 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 809 | arg_list.append( |
| 810 | ( |
| 811 | "mode{}_shift{}_odim{}x{}_out{}_st{}x{}_off{}x{}".format( |
| 812 | m, |
| 813 | shift, |
| 814 | output_dims[0], |
| 815 | output_dims[1], |
| 816 | testGen.typeStr(outputDType), |
| 817 | stride[0], |
| 818 | stride[1], |
| 819 | offset[0], |
| 820 | offset[1], |
| 821 | ), |
| 822 | [ |
| 823 | m, |
| 824 | stride, |
| 825 | offset, |
| 826 | shift, |
| 827 | stride_fp, |
| 828 | offset_fp, |
| 829 | output_dims, |
| 830 | dtype, |
| 831 | outputDType, |
| 832 | ], |
| 833 | ) |
| 834 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 835 | |
| 836 | return arg_list |
| 837 | |
| 838 | def agCondIf(testGen, opName, shapeList, dtype): |
| 839 | # CondIf generates the condition values here. |
| 840 | # Convert to tensors in the build function, along with the |
| 841 | # then and else blocks |
| 842 | arg_list = [] |
| 843 | |
| 844 | for c in [False, True]: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 845 | arg_list.append(("cond{}".format(int(c)), [c])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 846 | |
| 847 | return arg_list |
| 848 | |
| 849 | def agWhileLoop(testGen, opName, shapeList, dtype): |
| 850 | # While loop: 0 iterations, 1, more than 1 |
| 851 | arg_list = [] |
| 852 | |
| 853 | for iter in [0, 1, 4]: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 854 | arg_list.append(("iter{}".format(iter), [iter])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 855 | |
| 856 | return arg_list |
| 857 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 858 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 859 | class TosaTestGen: |
| 860 | def __init__(self, args): |
| 861 | self.args = args |
| 862 | self.basePath = args.output_dir |
| 863 | self.random_seed = args.random_seed |
| 864 | self.ser = None |
| 865 | self.rng = np.random.default_rng(self.random_seed) |
| 866 | self.createDynamicOpLists() |
| 867 | self.initOpListDefaults() |
| 868 | self.quantGen = TosaQuantGen() |
| 869 | # Force makeShape to do a specific starting shape |
| 870 | self.targetted_shape = None |
| 871 | |
| 872 | def createSerializer(self, opName, testPath): |
| 873 | self.testPath = os.path.join(opName, testPath) |
| 874 | |
| 875 | fullPath = os.path.join(self.basePath, self.testPath) |
| 876 | os.makedirs(fullPath, exist_ok=True) |
| 877 | self.ser = ts.TosaSerializer(fullPath) |
| 878 | |
| 879 | def getSerializer(self): |
| 880 | return self.ser |
| 881 | |
| 882 | def serialize(self, testName): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 883 | with open( |
| 884 | os.path.join(self.basePath, self.testPath, "{}.tosa".format(testName)), "wb" |
| 885 | ) as fd: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 886 | fd.write(self.ser.serialize()) |
| 887 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 888 | with open(os.path.join(self.basePath, self.testPath, "desc.json"), "w") as fd: |
| 889 | fd.write(self.ser.writeJson("{}.tosa".format(testName))) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 890 | |
| 891 | def getRandTensor(self, shape, dtype): |
| 892 | RAND_SHIFT_FACTOR = 0.5 |
| 893 | RAND_SCALE_FACTOR = 4.0 |
| 894 | |
| 895 | if dtype == DType.BOOL: |
| 896 | np_dt = np.bool |
| 897 | return np.bool_(self.rng.choice(a=[False, True], size=shape)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 898 | elif dtype == DType.INT4: |
| 899 | return np.int32(self.rng.integers(low=-7, high=8, size=shape)) |
| 900 | elif dtype == DType.INT8: |
| 901 | return np.int32(self.rng.integers(low=-127, high=128, size=shape)) |
| 902 | elif dtype == DType.INT16: |
| 903 | return np.int32(self.rng.integers(low=-32768, high=32768, size=shape)) |
| 904 | elif dtype == DType.INT32: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 905 | return np.int32( |
| 906 | self.rng.integers(low=-(1 << 31), high=(1 << 31), size=shape) |
| 907 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 908 | elif dtype == DType.INT48: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 909 | return np.int64( |
| 910 | self.rng.integers(low=-(1 << 47), high=(1 << 47), size=shape) |
| 911 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 912 | elif dtype == DType.FLOAT: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 913 | return np.float32( |
| 914 | self.rng.random(size=shape) - RAND_SHIFT_FACTOR * RAND_SCALE_FACTOR |
| 915 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 916 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 917 | raise Exception("Unrecognized Dtype: {}".format(dtype)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 918 | |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 919 | def buildPlaceholderTensors(self, shape_list, dtype_list): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 920 | placeholders = [] |
| 921 | |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 922 | assert len(shape_list) == len(dtype_list) |
| 923 | |
| 924 | for idx, shape in enumerate(shape_list): |
| 925 | arr = self.getRandTensor(shape, dtype_list[idx]) |
| 926 | placeholders.append(self.ser.addPlaceholder(shape, dtype_list[idx], arr)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 927 | |
| 928 | return placeholders |
| 929 | |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 930 | def buildConstTensors(self, shape_list, dtype_list): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 931 | consts = [] |
| 932 | |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 933 | assert len(shape_list) == len(dtype_list) |
| 934 | |
| 935 | for idx, shape in enumerate(shape_list): |
| 936 | arr = self.getRandTensor(shape, dtype_list[idx]) |
| 937 | consts.append(self.ser.addConst(shape, dtype_list[idx], arr)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 938 | |
| 939 | return consts |
| 940 | |
| 941 | def makeShape(self, rank): |
| 942 | if self.targetted_shape: |
| 943 | return np.int32(self.targetted_shape) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 944 | return np.int32( |
| 945 | self.rng.integers( |
| 946 | low=self.args.tensor_shape_range[0], |
| 947 | high=self.args.tensor_shape_range[1], |
| 948 | size=rank, |
| 949 | ) |
| 950 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 951 | |
| 952 | def setTargetShape(self, shape): |
| 953 | self.targetted_shape = shape |
| 954 | |
| 955 | def randInt(self, low=0, high=256): |
| 956 | return np.int32(self.rng.integers(low=low, high=high, size=1))[0] |
| 957 | |
| 958 | def getRandNumberDType(self, dtype): |
| 959 | if dtype == DType.FLOAT: |
| 960 | return self.rng.random() |
| 961 | elif dtype == DType.BOOL: |
| 962 | return self.rng.choice([False, True]) |
| 963 | elif dtype == DType.INT4: |
| 964 | low, high = (-7, 8) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 965 | elif dtype == DType.INT8: |
| 966 | low, high = (-127, 128) |
| 967 | elif dtype == DType.INT16: |
| 968 | low, high = (-32768, 32768) |
| 969 | elif dtype == DType.INT32: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 970 | low, high = (-(1 << 31), (1 << 31)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 971 | elif dtype == DType.INT48: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 972 | low, high = (-(1 << 47), (1 << 47)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 973 | # Special size |
| 974 | return np.int64(self.rng.integers(low, high, size=1))[0] |
| 975 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 976 | raise Exception("Unknown dtype: {}".format(dtype)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 977 | |
| 978 | return np.int32(self.rng.integers(low, high, size=1))[0] |
| 979 | |
| 980 | def shapeStr(self, shape): |
| 981 | |
| 982 | sStr = [] |
| 983 | # Convert to strings |
| 984 | for i in shape: |
| 985 | sStr.append(str(i)) |
| 986 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 987 | return "x".join(sStr) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 988 | |
| 989 | def typeStr(self, t): |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 990 | if isinstance(t, list): |
| 991 | assert len(t) >= 2 |
| 992 | return "{}x{}".format(self.typeStr(t[0]), self.typeStr(t[1])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 993 | else: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 994 | if t == DType.BOOL: |
| 995 | return "b" |
| 996 | elif t == DType.INT4: |
| 997 | return "i4" |
| 998 | elif t == DType.INT8: |
| 999 | return "i8" |
| 1000 | elif t == DType.UINT8: |
| 1001 | return "u8" |
| 1002 | elif t == DType.INT16: |
| 1003 | return "i16" |
| 1004 | elif t == DType.INT32: |
| 1005 | return "i32" |
| 1006 | elif t == DType.INT48: |
| 1007 | return "i48" |
| 1008 | elif t == DType.FLOAT: |
| 1009 | return "float" |
| 1010 | else: |
| 1011 | raise Exception("Unknown dtype, cannot convert to string: {}".format(t)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1012 | |
| 1013 | def typeWidth(self, t): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1014 | """ Get the datatype width for integer types""" |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1015 | if t == DType.INT4: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1016 | return 4 |
| 1017 | elif t == DType.INT8: |
| 1018 | return 8 |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1019 | elif t == DType.UINT8: |
| 1020 | return 8 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1021 | elif t == DType.INT16: |
| 1022 | return 16 |
| 1023 | elif t == DType.INT32: |
| 1024 | return 32 |
| 1025 | elif t == DType.INT48: |
| 1026 | return 48 |
| 1027 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1028 | raise Exception("Unknown dtype, cannot convert to string: {}".format(t)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1029 | |
| 1030 | # Argument generators |
| 1031 | # Returns a list of tuples (stringDescriptor, [build_fcn_arg_list]) |
| 1032 | # Where the string descriptor is used to generate the test name and |
| 1033 | # The build_fcn_arg_list is expanded and passed to the operator test |
| 1034 | # build function |
| 1035 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1036 | def build_unary(self, op, a, qinfo=None): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1037 | result_tens = OutputShaper.unaryOp(self.ser, a) |
| 1038 | self.ser.addOperator(op, [a.name], [result_tens.name], None, qinfo) |
| 1039 | return result_tens |
| 1040 | |
| 1041 | def build_binary_broadcast(self, op, a, b): |
| 1042 | result_tens = OutputShaper.binaryBroadcastOp(self.ser, a, b) |
| 1043 | self.ser.addOperator(op, [a.name, b.name], [result_tens.name]) |
| 1044 | return result_tens |
| 1045 | |
| 1046 | def build_binary_nonbroadcast(self, op, a, b): |
| 1047 | result_tens = OutputShaper.binaryNonBroadcastOp(self.ser, a, b) |
| 1048 | self.ser.addOperator(op, [a.name, b.name], [result_tens.name]) |
| 1049 | return result_tens |
| 1050 | |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1051 | def build_arithmetic_right_shift(self, op, a, b, round): |
| 1052 | result_tens = OutputShaper.binaryBroadcastOp(self.ser, a, b) |
| 1053 | |
| 1054 | attr = ts.TosaSerializerAttribute() |
| 1055 | attr.ArithmeticRightShiftAttribute(round) |
| 1056 | |
| 1057 | self.ser.addOperator(op, [a.name, b.name], [result_tens.name], attr) |
| 1058 | return result_tens |
| 1059 | |
| 1060 | def build_mul(self, op, a, b, shift): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1061 | result_tens = OutputShaper.binaryBroadcastOp(self.ser, a, b) |
| 1062 | |
| 1063 | # Special for multiply: |
| 1064 | # Force the result to INT32 for INT types |
| 1065 | if a.dtype != DType.FLOAT: |
| 1066 | result_tens.setDtype(DType.INT32) |
| 1067 | |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1068 | attr = ts.TosaSerializerAttribute() |
| 1069 | attr.MulAttribute(shift) |
| 1070 | |
| 1071 | self.ser.addOperator(op, [a.name, b.name], [result_tens.name], attr) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1072 | return result_tens |
| 1073 | |
| 1074 | def build_table(self, op, a): |
| 1075 | # Constant size, random values |
| 1076 | table_arr = self.getRandTensor([513], DType.INT16) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1077 | table_tens = self.ser.addConst(table_arr.shape, DType.INT16, table_arr) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1078 | |
| 1079 | result_tens = OutputShaper.tableOp(self.ser, a, table_tens) |
| 1080 | self.ser.addOperator(op, [a.name, table_tens.name], [result_tens.name], None) |
| 1081 | |
| 1082 | return result_tens |
| 1083 | |
| 1084 | def build_select(self, op, cond, a, b): |
| 1085 | |
| 1086 | # Replace the cond tensor with a boolean tensor since it probably |
| 1087 | # has the wrong dtype |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1088 | t = self.buildPlaceholderTensors([cond.shape], [DType.BOOL]) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1089 | cond = t[0] |
| 1090 | |
| 1091 | result_tens = OutputShaper.selectOp(self.ser, cond, a, b) |
| 1092 | self.ser.addOperator(op, [cond.name, a.name, b.name], [result_tens.name]) |
| 1093 | |
| 1094 | return result_tens |
| 1095 | |
| 1096 | def build_comparison(self, op, a, b): |
| 1097 | result_tens = OutputShaper.binaryComparisonOp(self.ser, a, b) |
| 1098 | self.ser.addOperator(op, [a.name, b.name], [result_tens.name]) |
| 1099 | return result_tens |
| 1100 | |
| 1101 | def build_argmax(self, op, a, axis): |
| 1102 | result_tens = OutputShaper.argmaxOp(self.ser, a, axis) |
| 1103 | |
| 1104 | attr = ts.TosaSerializerAttribute() |
| 1105 | attr.AxisAttribute(axis) |
| 1106 | |
| 1107 | self.ser.addOperator(op, [a.name], [result_tens.name], attr) |
| 1108 | return result_tens |
| 1109 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1110 | def build_pool2d(self, op, input, kernel, stride, pad, qinfo=None): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1111 | result_tens = OutputShaper.pool2dOp(self.ser, input, kernel, stride, pad) |
| 1112 | |
| 1113 | attr = ts.TosaSerializerAttribute() |
| 1114 | attr.Pool2dAttribute(kernel, stride, pad) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1115 | |
| 1116 | self.ser.addOperator(op, [input.name], [result_tens.name], attr, qinfo) |
| 1117 | return result_tens |
| 1118 | |
| 1119 | def build_conv2d(self, op, ifm, filter, bias, strides, padding, dilations, qinfo): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1120 | assert len(padding) == 4 |
| 1121 | result_tens = OutputShaper.conv2dOp( |
| 1122 | self.ser, ifm, filter, strides, padding, dilations |
| 1123 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1124 | |
| 1125 | attr = ts.TosaSerializerAttribute() |
| 1126 | attr.Conv2dAttribute(padding, strides, dilations) |
| 1127 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1128 | self.ser.addOperator( |
| 1129 | op, [ifm.name, filter.name, bias.name], [result_tens.name], attr, qinfo |
| 1130 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1131 | return result_tens |
| 1132 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1133 | def build_transpose_conv2d( |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1134 | self, op, ifm, filter, bias, stride, outpad, dilation, output_shape, qinfo |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1135 | ): |
| 1136 | assert len(outpad) == 2 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1137 | result_tens = OutputShaper.transposeConv2DOp(self.ser, ifm, output_shape) |
| 1138 | |
| 1139 | attr = ts.TosaSerializerAttribute() |
| 1140 | attr.TransposeConv2DAttribute(outpad, stride, dilation, output_shape) |
| 1141 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1142 | self.ser.addOperator( |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1143 | op, [ifm.name, filter.name, bias.name], [result_tens.name], attr, qinfo |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1144 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1145 | return result_tens |
| 1146 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1147 | def build_depthwise_conv2d( |
| 1148 | self, op, ifm, filter, bias, strides, padding, dilations, qinfo |
| 1149 | ): |
| 1150 | result_tens = OutputShaper.depthwiseConv2dOp( |
| 1151 | self.ser, ifm, filter, strides, padding, dilations |
| 1152 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1153 | |
| 1154 | attr = ts.TosaSerializerAttribute() |
| 1155 | attr.Conv2dAttribute(padding, strides, dilations) |
| 1156 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1157 | self.ser.addOperator( |
| 1158 | op, [ifm.name, filter.name, bias.name], [result_tens.name], attr, qinfo |
| 1159 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1160 | return result_tens |
| 1161 | |
| 1162 | def build_fully_connected(self, op, ifm, filter, bias, qinfo): |
| 1163 | result_tens = OutputShaper.fullyConnectedOp(self.ser, ifm, filter) |
| 1164 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1165 | self.ser.addOperator( |
| 1166 | op, [ifm.name, filter.name, bias.name], [result_tens.name], None, qinfo |
| 1167 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1168 | return result_tens |
| 1169 | |
| 1170 | def build_matmul(self, op, a, b, qinfo): |
| 1171 | result_tens = OutputShaper.matmulOp(self.ser, a, b) |
| 1172 | self.ser.addOperator(op, [a.name, b.name], [result_tens.name], None, qinfo) |
| 1173 | return result_tens |
| 1174 | |
| 1175 | def build_reduce(self, op, a, axis): |
| 1176 | result_tens = OutputShaper.reduceOp(self.ser, a, axis) |
| 1177 | |
| 1178 | attr = ts.TosaSerializerAttribute() |
| 1179 | attr.AxisAttribute(axis) |
| 1180 | |
| 1181 | self.ser.addOperator(op, [a.name], result_tens.name, attr) |
| 1182 | return result_tens |
| 1183 | |
| 1184 | def build_clamp(self, op, a): |
| 1185 | result_tens = OutputShaper.unaryOp(self.ser, a) |
| 1186 | |
| 1187 | attr = ts.TosaSerializerAttribute() |
| 1188 | |
| 1189 | # Get two random ints |
| 1190 | v = [self.randInt(), self.randInt()] |
| 1191 | |
| 1192 | if a.dtype == DType.FLOAT: |
| 1193 | attr.ClampAttribute(0, 0, min(v), max(v)) |
| 1194 | else: |
| 1195 | attr.ClampAttribute(min(v), max(v), 0, 0) |
| 1196 | |
| 1197 | self.ser.addOperator(op, [a.name], [result_tens.name], attr) |
| 1198 | return result_tens |
| 1199 | |
| 1200 | def build_leaky_relu(self, op, a): |
| 1201 | result_tens = OutputShaper.unaryOp(self.ser, a) |
| 1202 | attr = ts.TosaSerializerAttribute() |
| 1203 | |
| 1204 | attr.LeakyReluAttribute(self.getRandNumberDType(DType.FLOAT)) |
| 1205 | |
| 1206 | self.ser.addOperator(op, [a.name], [result_tens.name], attr) |
| 1207 | return result_tens |
| 1208 | |
| 1209 | # Needs an additional type/input |
| 1210 | def build_prelu(self, op, a): |
| 1211 | result_tens = OutputShaper.unaryOp(self.ser, a) |
| 1212 | |
| 1213 | self.ser.addOperator(op, [a.name], [result_tens.name]) |
| 1214 | return result_tens |
| 1215 | |
| 1216 | def build_relun(self, op, a): |
| 1217 | result_tens = OutputShaper.unaryOp(self.ser, a) |
| 1218 | |
| 1219 | attr = ts.TosaSerializerAttribute() |
| 1220 | |
| 1221 | if a.dtype == DType.FLOAT: |
| 1222 | attr.ReluNAttribute(0, self.getRandNumberDType(a.dtype)) |
| 1223 | else: |
| 1224 | attr.ReluNAttribute(self.getRandNumberDType(a.dtype), 0) |
| 1225 | |
| 1226 | self.ser.addOperator(op, [a.name], [result_tens.name], attr) |
| 1227 | return result_tens |
| 1228 | |
| 1229 | def build_sigmoid(self, op, a): |
| 1230 | result_tens = OutputShaper.unaryOp(self.ser, a) |
| 1231 | self.ser.addOperator(op, [a.name], [result_tens.name]) |
| 1232 | return result_tens |
| 1233 | |
| 1234 | def build_tanh(self, op, a): |
| 1235 | result_tens = OutputShaper.unaryOp(self.ser, a) |
| 1236 | self.ser.addOperator(op, [a.name], [result_tens.name]) |
| 1237 | return result_tens |
| 1238 | |
| 1239 | def build_concat(self, op, a, b, axis): |
| 1240 | result_tens = OutputShaper.concatOp(self.ser, a, b, axis) |
| 1241 | |
| 1242 | attr = ts.TosaSerializerAttribute() |
| 1243 | attr.AxisAttribute(axis) |
| 1244 | |
| 1245 | self.ser.addOperator(op, [a.name, b.name], [result_tens.name], attr) |
| 1246 | |
| 1247 | def build_pad(self, op, a, padding, qinfo): |
| 1248 | result_tens = OutputShaper.padOp(self.ser, a, padding) |
| 1249 | |
| 1250 | # Need to turn the padding array into a TOSA tensor here. |
| 1251 | # This is one of the few tensor operands that does not get |
| 1252 | # randomly generated |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1253 | padding_tens = self.ser.addConst(padding.shape, DType.INT32, padding) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1254 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1255 | self.ser.addOperator( |
| 1256 | op, [a.name, padding_tens.name], [result_tens.name], None, qinfo |
| 1257 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1258 | |
| 1259 | def build_reshape(self, op, a, newShape): |
| 1260 | result_tens = OutputShaper.reshapeOp(self.ser, a, newShape) |
| 1261 | |
| 1262 | attr = ts.TosaSerializerAttribute() |
| 1263 | attr.ReshapeAttribute(newShape) |
| 1264 | |
| 1265 | self.ser.addOperator(op, [a.name], [result_tens.name], attr) |
| 1266 | return result_tens |
| 1267 | |
| 1268 | def build_reverse(self, op, a, axis): |
| 1269 | result_tens = OutputShaper.unaryOp(self.ser, a) |
| 1270 | |
| 1271 | attr = ts.TosaSerializerAttribute() |
| 1272 | attr.AxisAttribute(axis) |
| 1273 | |
| 1274 | self.ser.addOperator(op, [a.name], [result_tens.name], attr) |
| 1275 | return result_tens |
| 1276 | |
| 1277 | def build_transpose(self, op, a, perms): |
| 1278 | result_tens = OutputShaper.transposeOp(self.ser, a, perms) |
| 1279 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1280 | perms_tens = self.ser.addConst([len(perms)], DType.INT32, np.int32(perms)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1281 | |
| 1282 | self.ser.addOperator(op, [a.name, perms_tens.name], [result_tens.name]) |
| 1283 | return result_tens |
| 1284 | |
| 1285 | def build_slice(self, op, a, begin, size): |
| 1286 | result_tens = OutputShaper.sliceOp(self.ser, a, begin, size) |
| 1287 | |
| 1288 | attr = ts.TosaSerializerAttribute() |
| 1289 | attr.SliceAttribute(begin, size) |
| 1290 | |
| 1291 | self.ser.addOperator(op, [a.name], [result_tens.name], attr) |
| 1292 | return result_tens |
| 1293 | |
| 1294 | def build_tile(self, op, a, multiples): |
| 1295 | result_tens = OutputShaper.tileOp(self.ser, a, multiples) |
| 1296 | |
| 1297 | attr = ts.TosaSerializerAttribute() |
| 1298 | attr.TileAttribute(multiples) |
| 1299 | |
| 1300 | self.ser.addOperator(op, [a.name], [result_tens.name], attr) |
| 1301 | return result_tens |
| 1302 | |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 1303 | def build_gather(self, op, values): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1304 | |
| 1305 | # Create a new indicies tensor |
| 1306 | # here with data that doesn't exceed the dimensions of the values tensor |
| 1307 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1308 | K = values.shape[1] # K |
| 1309 | W = self.randInt( |
| 1310 | self.args.tensor_shape_range[0], self.args.tensor_shape_range[1] |
| 1311 | ) # W |
| 1312 | indicies_arr = np.int32( |
| 1313 | self.rng.integers(low=0, high=K, size=[values.shape[0], W]) |
| 1314 | ) # (N, W) |
| 1315 | indicies = self.ser.addConst(indicies_arr.shape, DType.INT32, indicies_arr) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1316 | |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 1317 | result_tens = OutputShaper.gatherOp(self.ser, values, indicies) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1318 | |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 1319 | self.ser.addOperator(op, [values.name, indicies.name], [result_tens.name]) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1320 | |
| 1321 | return result_tens |
| 1322 | |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 1323 | def build_scatter(self, op, values_in, input): |
| 1324 | |
| 1325 | # Create a new indicies tensor |
| 1326 | # here with data that doesn't exceed the dimensions of the values_in tensor |
| 1327 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1328 | K = values_in.shape[1] # K |
| 1329 | W = input.shape[1] # W |
| 1330 | indicies_arr = np.int32( |
| 1331 | self.rng.integers(low=0, high=K, size=[values_in.shape[0], W]) |
| 1332 | ) # (N, W) |
| 1333 | indicies = self.ser.addConst(indicies_arr.shape, DType.INT32, indicies_arr) |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 1334 | |
| 1335 | result_tens = OutputShaper.scatterOp(self.ser, values_in, indicies, input) |
| 1336 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1337 | self.ser.addOperator( |
| 1338 | op, [values_in.name, indicies.name, input.name], [result_tens.name] |
| 1339 | ) |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 1340 | |
| 1341 | return result_tens |
| 1342 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1343 | def build_resize( |
| 1344 | self, |
| 1345 | op, |
| 1346 | input, |
| 1347 | mode, |
| 1348 | stride, |
| 1349 | offset, |
| 1350 | shift, |
| 1351 | stride_fp, |
| 1352 | offset_fp, |
| 1353 | output_dims, |
| 1354 | input_dtype, |
| 1355 | output_dtype, |
| 1356 | ): |
| 1357 | result_tens = OutputShaper.resizeOp( |
| 1358 | self.ser, |
| 1359 | input, |
| 1360 | mode, |
| 1361 | stride, |
| 1362 | offset, |
| 1363 | shift, |
| 1364 | stride_fp, |
| 1365 | offset_fp, |
| 1366 | output_dims, |
| 1367 | input_dtype, |
| 1368 | output_dtype, |
| 1369 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1370 | |
| 1371 | attr = ts.TosaSerializerAttribute() |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 1372 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1373 | attr.ResizeAttribute( |
| 1374 | output_dims, stride, offset, shift, stride_fp, offset_fp, mode |
| 1375 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1376 | |
| 1377 | self.ser.addOperator(op, [input.name], [result_tens.name], attr) |
| 1378 | return result_tens |
| 1379 | |
| 1380 | def build_identityn(self, op, val, val2): |
| 1381 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1382 | result_tens = OutputShaper.unaryOp(self.ser, val) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1383 | result_tens2 = OutputShaper.unaryOp(self.ser, val2) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1384 | self.ser.addOperator( |
| 1385 | op, [val.name, val2.name], [result_tens.name, result_tens2.name] |
| 1386 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1387 | return result_tens |
| 1388 | |
| 1389 | def build_placeholder(self, op, val): |
| 1390 | # Add an identity op to avoid warning in the reference model |
| 1391 | return self.build_unary(Op.IDENTITY, val) |
| 1392 | |
| 1393 | # Type Conversion |
| 1394 | def build_cast(self, op, val, out_dtype): |
| 1395 | result_tens = OutputShaper.typeConversionOp(self.ser, val, out_dtype) |
| 1396 | self.ser.addOperator(op, [val.name], [result_tens.name]) |
| 1397 | return result_tens |
| 1398 | |
| 1399 | def build_rescale(self, op, val, out_dtype, scale32, double_round, per_channel): |
| 1400 | result_tens = OutputShaper.typeConversionOp(self.ser, val, out_dtype) |
| 1401 | |
| 1402 | if per_channel: |
| 1403 | nc = val.shape[-1] |
| 1404 | else: |
| 1405 | nc = 1 |
| 1406 | |
| 1407 | in_type_width = self.typeWidth(val.dtype) |
| 1408 | out_type_width = self.typeWidth(out_dtype) |
| 1409 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1410 | if val.dtype == DType.INT8: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1411 | input_zp = self.randInt(-128, 127) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1412 | in_type_width = in_type_width + 1 |
| 1413 | else: |
| 1414 | input_zp = 0 |
| 1415 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1416 | if out_dtype == DType.INT8: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1417 | output_zp = self.randInt(-128, 127) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1418 | out_type_width = out_type_width + 1 |
| 1419 | else: |
| 1420 | output_zp = 0 |
| 1421 | |
| 1422 | # Calculate scale based on: |
| 1423 | # scale = a *(2^output_width)/(2^input_width)) |
| 1424 | |
| 1425 | a = np.float32(self.rng.random(size=[nc])) |
| 1426 | scale_arr = a * np.float32((1 << out_type_width) / (1 << in_type_width)) |
| 1427 | |
| 1428 | if scale32: |
| 1429 | pass |
| 1430 | # Cap the scaling at 2^15 - 1 for scale16 |
| 1431 | scale_arr = np.clip(scale_arr, 1.0 / (1 << 31), (1 << 31) - 1) |
| 1432 | else: |
| 1433 | # Cap the scaling at 2^15 - 1 for scale16 |
| 1434 | scale_arr = np.clip(scale_arr, 1.0 / (1 << 31), 32767.0) |
| 1435 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1436 | # print('{} {} -> {}'.format(out_type_width, in_type_width, scale_arr)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1437 | |
| 1438 | multiplier_arr = np.int32(np.zeros(shape=[nc])) |
| 1439 | shift_arr = np.int32(np.zeros(shape=[nc])) |
| 1440 | |
| 1441 | for i in range(nc): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1442 | multiplier_arr[i], shift_arr[i] = TosaQuantGen.computeMultiplierAndShift( |
| 1443 | scale_arr[i], scale32 |
| 1444 | ) |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1445 | if shift_arr[i] < 2 or shift_arr[i] > 62: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1446 | self.ser.setExpectedFailure(True, "OpRescale: invalid shift value") |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1447 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1448 | # print('multiplier {} shift {} inzp {} outzp {}'.format(multiplier_arr, shift_arr, input_zp, output_zp)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1449 | |
| 1450 | attr = ts.TosaSerializerAttribute() |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1451 | attr.RescaleAttribute( |
| 1452 | input_zp, |
| 1453 | output_zp, |
| 1454 | multiplier_arr, |
| 1455 | shift_arr, |
| 1456 | scale32, |
| 1457 | double_round, |
| 1458 | per_channel, |
| 1459 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1460 | |
| 1461 | self.ser.addOperator(op, [val.name], [result_tens.name], attr) |
| 1462 | return result_tens |
| 1463 | |
| 1464 | def build_cond_if_const(self, op, then_tens, else_tens, cond): |
| 1465 | # For cond_if with constants, we're supplied with then/else tensors that we ignore |
| 1466 | # (except for the generated shap) and the condition. Build Then/Else blocks |
| 1467 | # and fill them with const nodes for the body. |
| 1468 | |
| 1469 | # Condition tensor |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1470 | cond_tens = self.ser.addConst([], DType.BOOL, [cond]) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1471 | |
| 1472 | # Make then/else tensors |
| 1473 | out_shape = then_tens.shape |
| 1474 | then_arr = np.int32(self.rng.integers(0, 255, size=out_shape)) |
| 1475 | else_arr = np.int32(self.rng.integers(0, 255, size=out_shape)) |
| 1476 | |
| 1477 | # And the result tensor based on any of the outputs |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1478 | result_tens = self.ser.addOutput(out_shape, DType.INT32) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1479 | |
| 1480 | # Create the attribute with the names of the then/else blocks |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1481 | then_block = "THEN_BLOCK" |
| 1482 | else_block = "ELSE_BLOCK" |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1483 | attr = ts.TosaSerializerAttribute() |
| 1484 | attr.CondIfAttribute(then_block, else_block) |
| 1485 | |
| 1486 | # Finally, build the op and the two blocks |
| 1487 | self.ser.addOperator(op, [cond_tens.name], [result_tens.name], attr) |
| 1488 | |
| 1489 | self.ser.startBasicBlock(then_block) |
| 1490 | # Build the actual then/else tensors inside their blocks |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1491 | then_tens = self.ser.addConst(out_shape, DType.INT32, then_arr) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1492 | self.ser.addOutputTensor(then_tens) |
| 1493 | |
| 1494 | self.ser.startBasicBlock(else_block) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1495 | else_tens = self.ser.addConst(out_shape, DType.INT32, else_arr) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1496 | self.ser.addOutputTensor(else_tens) |
| 1497 | |
| 1498 | return result_tens |
| 1499 | |
| 1500 | def build_cond_if_binary(self, op, a, b, cond): |
| 1501 | # For cond_if with a binary op in the then/else blocks, take a and b and |
| 1502 | # alternately add or subtract them based on the condition |
| 1503 | |
| 1504 | # Condition tensor |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1505 | cond_tens = self.ser.addConst([], DType.BOOL, [cond]) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1506 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1507 | result_tens = self.ser.addOutput(a.shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1508 | self.ser.currBasicBlock.addOutput(result_tens.name) |
| 1509 | |
| 1510 | # Create the attribute with the names of the then/else blocks |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1511 | then_block = "THEN_BLOCK" |
| 1512 | else_block = "ELSE_BLOCK" |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1513 | attr = ts.TosaSerializerAttribute() |
| 1514 | attr.CondIfAttribute(then_block, else_block) |
| 1515 | |
| 1516 | # Finally, build the op and the two blocks |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1517 | self.ser.addOperator( |
| 1518 | op, [cond_tens.name, a.name, b.name], [result_tens.name], attr |
| 1519 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1520 | |
| 1521 | self.ser.startBasicBlock(then_block) |
| 1522 | self.ser.addInputTensor(a) |
| 1523 | self.ser.addInputTensor(b) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1524 | then_tens = self.ser.addOutput(a.shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1525 | self.ser.addOperator(Op.ADD, [a.name, b.name], [then_tens.name]) |
| 1526 | |
| 1527 | self.ser.startBasicBlock(else_block) |
| 1528 | self.ser.addInputTensor(a) |
| 1529 | self.ser.addInputTensor(b) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1530 | else_tens = self.ser.addOutput(a.shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1531 | self.ser.addOperator(Op.SUB, [a.name, b.name], [else_tens.name]) |
| 1532 | |
| 1533 | return result_tens |
| 1534 | |
| 1535 | def build_while_loop(self, op, a, iter_val): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1536 | iter = self.ser.addPlaceholder([], DType.INT32, [np.int32(iter_val)]) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1537 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1538 | cond_block = "COND_BLOCK" |
| 1539 | body_block = "BODY_BLOCK" |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1540 | |
| 1541 | attr = ts.TosaSerializerAttribute() |
| 1542 | attr.WhileLoopAttribute(cond_block, body_block) |
| 1543 | |
| 1544 | # Accumulator tensor |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1545 | # acc = self.ser.addOutput(a.shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1546 | acc_init_val = np.int32(np.zeros(a.shape)) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1547 | acc = self.ser.addPlaceholder(a.shape, a.dtype, acc_init_val) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1548 | |
| 1549 | # Intermediate/output tensors for everything going through the loop |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1550 | iter_out = self.ser.addIntermediate(iter.shape, iter.dtype) |
| 1551 | a_out = self.ser.addIntermediate(a.shape, a.dtype) |
| 1552 | acc_out = self.ser.addIntermediate(acc.shape, acc.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1553 | |
| 1554 | # While_loop operator |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1555 | self.ser.addOperator( |
| 1556 | op, |
| 1557 | [iter.name, a.name, acc.name], |
| 1558 | [iter_out.name, a_out.name, acc_out.name], |
| 1559 | attr, |
| 1560 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1561 | |
| 1562 | # COND block (input: iter, output: cond_tens ) |
| 1563 | self.ser.startBasicBlock(cond_block) |
| 1564 | self.ser.addInputTensor(iter) |
| 1565 | self.ser.addInputTensor(a) |
| 1566 | self.ser.addInputTensor(acc) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1567 | zero_tens = self.ser.addConst([], DType.INT32, [np.int32(0)]) |
| 1568 | cond_tens = self.ser.addOutput([], DType.BOOL) |
| 1569 | self.ser.addOperator(Op.GREATER, [iter.name, zero_tens.name], [cond_tens.name]) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1570 | |
| 1571 | # BODY block (input: a, acc, iter, output: a, acc, iter) |
| 1572 | # Note that local intermediate tensors need to be declared here for the outputs |
| 1573 | self.ser.startBasicBlock(body_block) |
| 1574 | self.ser.addInputTensor(iter) |
| 1575 | self.ser.addInputTensor(a) |
| 1576 | self.ser.addInputTensor(acc) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1577 | one_tens = self.ser.addConst([], DType.INT32, [np.int32(1)]) |
| 1578 | iter_body_out = self.ser.addIntermediate(iter.shape, iter.dtype) |
| 1579 | acc_body_out = self.ser.addIntermediate(acc.shape, acc.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1580 | self.ser.addOperator(Op.ADD, [a.name, acc.name], [acc_body_out.name]) |
| 1581 | self.ser.addOperator(Op.SUB, [iter.name, one_tens.name], [iter_body_out.name]) |
| 1582 | self.ser.addOutputTensor(iter_body_out) |
| 1583 | self.ser.addOutputTensor(a) |
| 1584 | self.ser.addOutputTensor(acc_body_out) |
| 1585 | |
| 1586 | return acc_out |
| 1587 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1588 | def genOpTestList( |
| 1589 | self, opName, shapeFilter=[None], rankFilter=None, dtypeFilter=None |
| 1590 | ): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1591 | |
| 1592 | try: |
| 1593 | op = self.TOSA_OP_LIST[opName] |
| 1594 | except KeyError as e: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1595 | raise Exception("Cannot find op with name {}".format(opName)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1596 | |
| 1597 | # Initialize a new random number generator |
| 1598 | self.rng = np.random.default_rng(self.random_seed) |
| 1599 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1600 | build_fcn, tgen_fcn, agen_fcn = op["build_fcn"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1601 | |
| 1602 | # Generate the lists of arguments |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1603 | rmin, rmax = op["rank"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1604 | |
| 1605 | # Test list consists of a tuple of: |
| 1606 | # (opName, testNameStr, dtype, shapeList, argumentsList) |
| 1607 | testList = [] |
| 1608 | |
| 1609 | if not shapeFilter: |
| 1610 | shapeFilter = [None] |
| 1611 | |
| 1612 | for r in range(rmin, rmax + 1): |
| 1613 | |
| 1614 | # Filter out the rank? |
| 1615 | if rankFilter is not None and r not in rankFilter: |
| 1616 | continue |
| 1617 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1618 | for t in op["types"]: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1619 | |
| 1620 | # Filter tests based on dtype? |
| 1621 | if dtypeFilter is not None: |
| 1622 | if t not in dtypeFilter: |
| 1623 | continue |
| 1624 | |
| 1625 | # Create the placeholder and const tensors |
| 1626 | for shape in shapeFilter: |
| 1627 | # A None shape chooses a random shape of a given rank |
| 1628 | |
| 1629 | # Filter out by rank |
| 1630 | if shape is not None and len(shape) != r: |
| 1631 | continue |
| 1632 | |
| 1633 | self.setTargetShape(shape) |
| 1634 | shapeList = tgen_fcn(self, op, r) |
| 1635 | |
| 1636 | shapeStr = self.shapeStr(shapeList[0]) |
| 1637 | typeStr = self.typeStr(t) |
| 1638 | |
| 1639 | # Argument lists consists of tuples of the (str, []) string representation and the build function argument list |
| 1640 | argList = [] |
| 1641 | if agen_fcn: |
| 1642 | argList = agen_fcn(self, opName, shapeList, t) |
| 1643 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1644 | argList = [("", [])] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1645 | |
| 1646 | for argStr, args in argList: |
| 1647 | if argStr: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1648 | testStr = "{}_{}_{}_{}".format( |
| 1649 | opName, shapeStr, typeStr, argStr |
| 1650 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1651 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1652 | testStr = "{}_{}_{}".format(opName, shapeStr, typeStr) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1653 | |
| 1654 | testList.append((opName, testStr, t, shapeList, args)) |
| 1655 | |
| 1656 | return testList |
| 1657 | |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1658 | def serializeTest(self, opName, testStr, dtype_or_dtypeList, shapeList, testArgs): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1659 | try: |
| 1660 | op = self.TOSA_OP_LIST[opName] |
| 1661 | except KeyError as e: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1662 | raise Exception("Cannot find op with name {}".format(opName)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1663 | |
| 1664 | # Create a serializer |
| 1665 | self.createSerializer(opName, testStr) |
| 1666 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1667 | build_fcn, tgen_fcn, agen_fcn = op["build_fcn"] |
| 1668 | pCount, cCount = op["operands"] |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1669 | num_operands = pCount + cCount |
| 1670 | |
| 1671 | if isinstance(dtype_or_dtypeList, list): |
| 1672 | dtypeList = dtype_or_dtypeList |
| 1673 | else: |
| 1674 | dtypeList = [dtype_or_dtypeList] * (num_operands) |
| 1675 | |
| 1676 | assert ( |
| 1677 | len(shapeList) == num_operands |
| 1678 | ), "shapeList length {} must match number of operands {}".format( |
| 1679 | len(shapeList), num_operands |
| 1680 | ) |
| 1681 | assert ( |
| 1682 | len(dtypeList) == num_operands |
| 1683 | ), "dtypeList length {} must match number of operands {}".format( |
| 1684 | len(dtypeList), num_operands |
| 1685 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1686 | |
| 1687 | try: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1688 | qgen = op["qgen"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1689 | except KeyError: |
| 1690 | qgen = None |
| 1691 | |
| 1692 | # Build the random tensor operands and the test |
| 1693 | tens = [] |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1694 | |
| 1695 | # If test is ArithmeticRightShift, force value of operand[1] to be within [0, num_bits] |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1696 | if op["op"] == Op.ARITHMETIC_RIGHT_SHIFT: |
| 1697 | assert ( |
| 1698 | pCount == 2 and cCount == 0 |
| 1699 | ), "Op.ArithmeticRightShift must have 2 placeholders, 0 consts" |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1700 | |
| 1701 | placeholders = [] |
| 1702 | for idx, shape in enumerate(shapeList[:]): |
| 1703 | if idx == 1: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1704 | if dtypeList[idx] == DType.INT8: |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1705 | arr = np.int32(self.rng.integers(low=0, high=8, size=shape)) |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1706 | elif dtypeList[idx] == DType.INT16: |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1707 | arr = np.int32(self.rng.integers(low=0, high=16, size=shape)) |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1708 | elif dtypeList[idx] == DType.INT32: |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1709 | arr = np.int32(self.rng.integers(low=0, high=32, size=shape)) |
| 1710 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1711 | raise Exception("OpArithmeticRightShift: invalid input dtype") |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1712 | else: |
Kevin Cheng | 14d7f7a | 2021-05-12 10:44:49 -0700 | [diff] [blame] | 1713 | arr = self.getRandTensor(shape, dtypeList[idx]) |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1714 | placeholders.append(self.ser.addPlaceholder(shape, dtypeList[idx], arr)) |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1715 | |
| 1716 | tens.extend(placeholders) |
Kevin Cheng | 14d7f7a | 2021-05-12 10:44:49 -0700 | [diff] [blame] | 1717 | elif op["op"] == Op.DIV: |
| 1718 | assert ( |
| 1719 | pCount == 2 and cCount == 0 |
| 1720 | ), "Op.Div must have 2 placeholders, 0 consts" |
| 1721 | |
| 1722 | placeholders = [] |
| 1723 | |
| 1724 | # Two invalid cases for Op.DIV: |
| 1725 | # 1. divisor == 0 |
Kevin Cheng | 47315e1 | 2021-05-13 17:41:28 -0700 | [diff] [blame] | 1726 | # 2. dividend == -(1<<31) and divisor == -1 |
Kevin Cheng | 14d7f7a | 2021-05-12 10:44:49 -0700 | [diff] [blame] | 1727 | while True: |
| 1728 | dividend_arr = self.getRandTensor(shapeList[0], dtypeList[0]) |
| 1729 | divisor_arr = self.getRandTensor(shapeList[1], dtypeList[1]) |
| 1730 | |
| 1731 | if (divisor_arr == 0).any(): |
| 1732 | continue |
| 1733 | |
Kevin Cheng | 47315e1 | 2021-05-13 17:41:28 -0700 | [diff] [blame] | 1734 | if (dividend_arr == -(2 ** 31)).any() and (divisor_arr == -1).any(): |
Kevin Cheng | 14d7f7a | 2021-05-12 10:44:49 -0700 | [diff] [blame] | 1735 | continue |
| 1736 | |
| 1737 | break |
| 1738 | |
| 1739 | placeholders.append( |
| 1740 | self.ser.addPlaceholder(shapeList[0], dtypeList[0], dividend_arr) |
| 1741 | ) |
| 1742 | placeholders.append( |
| 1743 | self.ser.addPlaceholder(shapeList[1], dtypeList[1], divisor_arr) |
| 1744 | ) |
| 1745 | |
| 1746 | tens.extend(placeholders) |
| 1747 | elif op["op"] == Op.MUL: |
| 1748 | assert ( |
| 1749 | pCount == 2 and cCount == 0 |
| 1750 | ), "Op.MUL must have 2 placeholders, 0 consts" |
| 1751 | |
| 1752 | if dtypeList[0] == DType.FLOAT: |
| 1753 | tens.extend(self.buildPlaceholderTensors(shapeList[:], dtypeList[:])) |
| 1754 | else: |
| 1755 | placeholders = [] |
| 1756 | |
| 1757 | # Make sure multiply result in int32 range |
| 1758 | shift = testArgs[0] |
| 1759 | if dtypeList[0] == DType.INT8: |
| 1760 | num_bits = 8 |
| 1761 | elif dtypeList[0] == DType.INT16: |
| 1762 | num_bits = 16 |
| 1763 | elif dtypeList[0] == DType.INT32: |
| 1764 | num_bits = 32 |
| 1765 | else: |
| 1766 | raise Exception("OpMul: invalid input dtype") |
| 1767 | |
| 1768 | for idx, shape in enumerate(shapeList[:]): |
| 1769 | low = -(2 ** (num_bits - 1)) |
| 1770 | high = (2 ** (num_bits - 1)) - 1 |
| 1771 | |
| 1772 | a_arr = np.int32( |
| 1773 | self.rng.integers(low=low, high=high, size=shapeList[0]) |
| 1774 | ) |
| 1775 | b_arr = np.int32( |
| 1776 | self.rng.integers(low=low, high=high, size=shapeList[1]) |
| 1777 | ) |
| 1778 | |
| 1779 | i = 0 |
| 1780 | while True: |
| 1781 | |
| 1782 | a_arr_64 = a_arr.astype(np.int64) |
| 1783 | b_arr_64 = b_arr.astype(np.int64) |
| 1784 | |
| 1785 | if shift > 0: |
| 1786 | rounding = 1 << (shift - 1) |
| 1787 | result_arr = ((a_arr_64 * b_arr_64) + rounding) >> shift |
| 1788 | else: |
| 1789 | result_arr = a_arr_64 * b_arr_64 |
| 1790 | |
| 1791 | if (result_arr > -(2 ** 31)).all() and ( |
| 1792 | result_arr <= ((2 ** 31) - 1) |
| 1793 | ).all(): |
| 1794 | break |
| 1795 | |
| 1796 | i = i + 1 |
| 1797 | a_arr = a_arr // 2 |
| 1798 | b_arr = b_arr // 2 |
| 1799 | |
| 1800 | placeholders.append( |
| 1801 | self.ser.addPlaceholder(shapeList[0], dtypeList[0], a_arr) |
| 1802 | ) |
| 1803 | placeholders.append( |
| 1804 | self.ser.addPlaceholder(shapeList[1], dtypeList[1], b_arr) |
| 1805 | ) |
| 1806 | |
| 1807 | tens.extend(placeholders) |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 1808 | else: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1809 | tens.extend( |
| 1810 | self.buildPlaceholderTensors(shapeList[0:pCount], dtypeList[0:pCount]) |
| 1811 | ) |
| 1812 | tens.extend(self.buildConstTensors(shapeList[pCount:], dtypeList[pCount:])) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1813 | |
| 1814 | if qgen is not None: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1815 | qinfo = qgen(self, op, dtypeList[0]) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1816 | else: |
| 1817 | qinfo = None |
| 1818 | |
| 1819 | try: |
| 1820 | if qinfo is not None: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1821 | resultName = build_fcn(self, op["op"], *tens, *testArgs, qinfo) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1822 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1823 | resultName = build_fcn(self, op["op"], *tens, *testArgs) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1824 | except TypeError as e: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1825 | print( |
| 1826 | "build_fcn: {}\nTensors: {}\nArgs: {}\n".format( |
| 1827 | build_fcn, tens, testArgs |
| 1828 | ) |
| 1829 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1830 | raise e |
| 1831 | |
| 1832 | # Save the serialized test |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1833 | self.serialize("test") |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1834 | |
| 1835 | def createDynamicOpLists(self): |
| 1836 | |
| 1837 | # Dynamically create op lists for convolutions with a list of kernel sizes |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1838 | KERNELS = [[1, 1], [2, 2], [3, 3], [5, 5], [3, 1], [1, 3]] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1839 | |
| 1840 | for k in KERNELS: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1841 | testName = "conv2d_{}x{}".format(k[0], k[1]) |
| 1842 | self.TOSA_OP_LIST[testName] = self.TOSA_OP_LIST["conv2d_TEMPLATE"].copy() |
| 1843 | self.TOSA_OP_LIST[testName]["filter"] = k |
| 1844 | self.TOSA_OP_LIST[testName]["template"] = False |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1845 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1846 | testName = "depthwise_conv2d_{}x{}".format(k[0], k[1]) |
| 1847 | self.TOSA_OP_LIST[testName] = self.TOSA_OP_LIST[ |
| 1848 | "depthwise_conv2d_TEMPLATE" |
| 1849 | ].copy() |
| 1850 | self.TOSA_OP_LIST[testName]["filter"] = k |
| 1851 | self.TOSA_OP_LIST[testName]["template"] = False |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1852 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1853 | testName = "transpose_conv2d_{}x{}".format(k[0], k[1]) |
| 1854 | self.TOSA_OP_LIST[testName] = self.TOSA_OP_LIST[ |
| 1855 | "transpose_conv2d_TEMPLATE" |
| 1856 | ].copy() |
| 1857 | self.TOSA_OP_LIST[testName]["filter"] = k |
| 1858 | self.TOSA_OP_LIST[testName]["template"] = False |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1859 | |
| 1860 | # Delete any templates after having created any dynamic ops |
| 1861 | # This is a two-pass operation because it's bad practice to delete |
| 1862 | # keys from dictionaries while iterating |
| 1863 | keyList = [] |
| 1864 | for k in self.TOSA_OP_LIST: |
| 1865 | try: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1866 | if self.TOSA_OP_LIST[k]["template"] == True: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1867 | keyList.append(k) |
| 1868 | continue |
| 1869 | except KeyError: |
| 1870 | pass |
| 1871 | |
| 1872 | for k in keyList: |
| 1873 | del self.TOSA_OP_LIST[k] |
| 1874 | |
| 1875 | def initOpListDefaults(self): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1876 | """Fill in default fields for ops if they aren't already specified. |
| 1877 | Look for missing required fields (datastructure linting).""" |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1878 | for op in self.TOSA_OP_LIST: |
| 1879 | |
| 1880 | # Required fields |
| 1881 | try: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1882 | pl, c = self.TOSA_OP_LIST[op]["operands"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1883 | except (KeyError, ValueError, TypeError): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1884 | raise Exception( |
| 1885 | "Op {} is missing a valid operand tuple in TOSA_OP_LIST".format(op) |
| 1886 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1887 | |
| 1888 | try: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1889 | fcn, tgen, arggen = self.TOSA_OP_LIST[op]["build_fcn"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1890 | except (KeyError, ValueError, TypeError): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1891 | raise Exception( |
| 1892 | "Op {} is missing a valid build_fcn tuple in TOSA_OP_LIST".format( |
| 1893 | op |
| 1894 | ) |
| 1895 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1896 | |
| 1897 | try: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1898 | types = self.TOSA_OP_LIST[op]["types"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1899 | except KeyError as e: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1900 | raise Exception( |
| 1901 | "Op {} is missing a valid type list in TOSA_OP_LIST".format(op) |
| 1902 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1903 | |
| 1904 | try: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1905 | opcode = self.TOSA_OP_LIST[op]["op"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1906 | except KeyError as e: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1907 | raise Exception( |
| 1908 | "Op {} is missing the Op field in TOSA_OP_LIST".format(op) |
| 1909 | ) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1910 | |
| 1911 | # Put in default rank range, if missing |
| 1912 | try: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1913 | rank = self.TOSA_OP_LIST[op]["rank"] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1914 | except KeyError: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1915 | self.TOSA_OP_LIST[op]["rank"] = self.DEFAULT_RANK_RANGE |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1916 | |
| 1917 | # Tensor operator list |
| 1918 | # 'op': op name |
| 1919 | # 'operands': tuple of (placeholder, const) operands |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 1920 | # 'rank': optional, restricts rank to tuple inclusive of (min, max), |
| 1921 | # if not specified, defaults to (1, 4) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1922 | # 'build_fcn': tuple of the function to (build_operator(), TensorGen function, ArgGen enum) |
| 1923 | # 'types': array of datatypes to be tested |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1924 | TYPE_FP = [DType.FLOAT] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1925 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1926 | TYPE_INT = [DType.INT8, DType.INT16, DType.INT32] # Excludes INT4 |
| 1927 | TYPE_INT_FP = [DType.INT8, DType.INT16, DType.INT32, DType.FLOAT] # Excludes INT4 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1928 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1929 | TYPE_BOOL = [DType.BOOL] |
| 1930 | TYPE_FI32 = [DType.FLOAT, DType.INT32] |
| 1931 | TYPE_FIB = [DType.FLOAT, DType.INT8, DType.INT16, DType.INT32, DType.BOOL] |
| 1932 | TYPE_FI16 = [DType.FLOAT, DType.INT16] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1933 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1934 | TYPE_NARROW_INT_FP = [DType.INT8, DType.INT16, DType.FLOAT] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1935 | |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1936 | TYPE_CONV2D = [ |
| 1937 | [DType.INT8, DType.INT8, DType.INT32], |
| 1938 | [DType.INT16, DType.INT8, DType.INT48], |
| 1939 | DType.FLOAT, |
| 1940 | ] |
| 1941 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1942 | DEFAULT_RANK_RANGE = (1, 4) |
| 1943 | |
| 1944 | TOSA_OP_LIST = { |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 1945 | # Tensor operators |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1946 | "argmax": { |
| 1947 | "op": Op.ARGMAX, |
| 1948 | "operands": (1, 0), |
| 1949 | "build_fcn": (build_argmax, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 1950 | "types": TYPE_NARROW_INT_FP, |
| 1951 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 1952 | "avg_pool2d": { |
| 1953 | "op": Op.AVG_POOL2D, |
| 1954 | "operands": (1, 0), |
| 1955 | "rank": (4, 4), |
| 1956 | "build_fcn": (build_pool2d, TosaTensorGen.tgNHWC, TosaArgGen.agPooling), |
| 1957 | "qgen": TosaQuantGen.qgUnary, |
| 1958 | "types": TYPE_NARROW_INT_FP, |
| 1959 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1960 | # Templated operator. Filled in by createDynamicOpLists |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1961 | "conv2d_TEMPLATE": { |
| 1962 | "op": Op.CONV2D, |
| 1963 | "operands": (1, 2), |
| 1964 | "rank": (4, 4), |
| 1965 | "build_fcn": (build_conv2d, TosaTensorGen.tgConv2D, TosaArgGen.agConv2D), |
| 1966 | "qgen": TosaQuantGen.qgConv, |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1967 | "types": TYPE_CONV2D, |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1968 | "template": True, |
| 1969 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 1970 | # Conv3d TBD |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 1971 | # Templated operator. Filled in by createDynamicOpLists |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1972 | "depthwise_conv2d_TEMPLATE": { |
| 1973 | "op": Op.DEPTHWISE_CONV2D, |
| 1974 | "operands": (1, 2), |
| 1975 | "filter": [1, 1], |
| 1976 | "rank": (4, 4), |
| 1977 | "build_fcn": ( |
| 1978 | build_depthwise_conv2d, |
| 1979 | TosaTensorGen.tgDepthwiseConv2D, |
| 1980 | TosaArgGen.agConv2D, |
| 1981 | ), |
| 1982 | "qgen": TosaQuantGen.qgConv, |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 1983 | "types": TYPE_CONV2D, |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 1984 | "template": True, |
| 1985 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 1986 | "fully_connected": { |
| 1987 | "op": Op.FULLY_CONNECTED, |
| 1988 | "operands": (1, 2), |
| 1989 | "rank": (2, 2), |
| 1990 | "build_fcn": (build_fully_connected, TosaTensorGen.tgFullyConnected, None), |
| 1991 | "qgen": TosaQuantGen.qgConv, |
| 1992 | "types": TYPE_CONV2D, |
| 1993 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 1994 | "matmul": { |
| 1995 | "op": Op.MATMUL, |
| 1996 | "operands": (2, 0), |
| 1997 | "rank": (2, 2), |
| 1998 | "build_fcn": (build_matmul, TosaTensorGen.tgMatmul, None), |
| 1999 | "qgen": TosaQuantGen.qgMatmul, |
| 2000 | "types": TYPE_NARROW_INT_FP, |
| 2001 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2002 | "max_pool2d": { |
| 2003 | "op": Op.MAX_POOL2D, |
| 2004 | "operands": (1, 0), |
| 2005 | "rank": (4, 4), |
| 2006 | "build_fcn": (build_pool2d, TosaTensorGen.tgNHWC, TosaArgGen.agPooling), |
| 2007 | "types": TYPE_NARROW_INT_FP, |
| 2008 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2009 | # Templated operator. Filled in by createDynamicOpLists |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2010 | "transpose_conv2d_TEMPLATE": { |
| 2011 | "op": Op.TRANSPOSE_CONV2D, |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2012 | "operands": (1, 2), |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2013 | "rank": (4, 4), |
| 2014 | "build_fcn": ( |
| 2015 | build_transpose_conv2d, |
| 2016 | TosaTensorGen.tgTransposeConv2D, |
| 2017 | TosaArgGen.agTransposeConv2D, |
| 2018 | ), |
| 2019 | "qgen": TosaQuantGen.qgConv, |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2020 | "types": TYPE_CONV2D, |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2021 | "template": True, |
| 2022 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2023 | # Activation functions |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2024 | "clamp": { |
| 2025 | "op": Op.CLAMP, |
| 2026 | "operands": (1, 0), |
| 2027 | "build_fcn": (build_clamp, TosaTensorGen.tgBasic, None), |
| 2028 | "types": TYPE_NARROW_INT_FP, |
| 2029 | }, |
| 2030 | "relun": { |
| 2031 | "op": Op.RELUN, |
| 2032 | "operands": (1, 0), |
| 2033 | "build_fcn": (build_relun, TosaTensorGen.tgBasic, None), |
| 2034 | "types": TYPE_FI32, |
| 2035 | }, |
| 2036 | "sigmoid": { |
| 2037 | "op": Op.SIGMOID, |
| 2038 | "operands": (1, 0), |
| 2039 | "build_fcn": (build_sigmoid, TosaTensorGen.tgBasic, None), |
| 2040 | "types": TYPE_FP, |
| 2041 | }, |
| 2042 | "tanh": { |
| 2043 | "op": Op.TANH, |
| 2044 | "operands": (1, 0), |
| 2045 | "build_fcn": (build_tanh, TosaTensorGen.tgBasic, None), |
| 2046 | "types": TYPE_FP, |
| 2047 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2048 | # Elementwise Binary Operators |
| 2049 | "add": { |
| 2050 | "op": Op.ADD, |
| 2051 | "operands": (2, 0), |
| 2052 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2053 | "types": TYPE_FI32, |
| 2054 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2055 | "arithmetic_right_shift": { |
| 2056 | "op": Op.ARITHMETIC_RIGHT_SHIFT, |
| 2057 | "operands": (2, 0), |
| 2058 | "build_fcn": ( |
| 2059 | build_arithmetic_right_shift, |
| 2060 | TosaTensorGen.tgBroadcastFuzz, |
| 2061 | TosaArgGen.agArithmeticRightShift, |
| 2062 | ), |
| 2063 | "types": TYPE_INT, |
| 2064 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2065 | "bitwise_and": { |
| 2066 | "op": Op.BITWISE_AND, |
| 2067 | "operands": (2, 0), |
| 2068 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2069 | "types": TYPE_INT, |
| 2070 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2071 | "bitwise_or": { |
| 2072 | "op": Op.BITWISE_OR, |
| 2073 | "operands": (2, 0), |
| 2074 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2075 | "types": TYPE_INT, |
| 2076 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2077 | "bitwise_xor": { |
| 2078 | "op": Op.BITWISE_XOR, |
| 2079 | "operands": (2, 0), |
| 2080 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2081 | "types": TYPE_INT, |
| 2082 | }, |
Kevin Cheng | 14d7f7a | 2021-05-12 10:44:49 -0700 | [diff] [blame] | 2083 | "div": { |
| 2084 | "op": Op.DIV, |
| 2085 | "operands": (2, 0), |
| 2086 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2087 | "types": [DType.INT32], |
| 2088 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2089 | "logical_and": { |
| 2090 | "op": Op.LOGICAL_AND, |
| 2091 | "operands": (2, 0), |
| 2092 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2093 | "types": TYPE_BOOL, |
| 2094 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2095 | "logical_left_shift": { |
| 2096 | "op": Op.LOGICAL_LEFT_SHIFT, |
| 2097 | "operands": (2, 0), |
| 2098 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2099 | "types": TYPE_INT, |
| 2100 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2101 | "logical_right_shift": { |
| 2102 | "op": Op.LOGICAL_RIGHT_SHIFT, |
| 2103 | "operands": (2, 0), |
| 2104 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2105 | "types": TYPE_INT, |
| 2106 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2107 | "logical_or": { |
| 2108 | "op": Op.LOGICAL_OR, |
| 2109 | "operands": (2, 0), |
| 2110 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2111 | "types": TYPE_BOOL, |
| 2112 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2113 | "logical_xor": { |
| 2114 | "op": Op.LOGICAL_XOR, |
| 2115 | "operands": (2, 0), |
| 2116 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2117 | "types": TYPE_BOOL, |
| 2118 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2119 | "maximum": { |
| 2120 | "op": Op.MAXIMUM, |
| 2121 | "operands": (2, 0), |
| 2122 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2123 | "types": TYPE_FI32, |
| 2124 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2125 | "minimum": { |
| 2126 | "op": Op.MINIMUM, |
| 2127 | "operands": (2, 0), |
| 2128 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2129 | "types": TYPE_FI32, |
| 2130 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2131 | "mul": { |
| 2132 | "op": Op.MUL, |
| 2133 | "operands": (2, 0), |
| 2134 | "build_fcn": (build_mul, TosaTensorGen.tgBroadcastFuzz, TosaArgGen.agMul), |
| 2135 | "types": TYPE_INT_FP, |
| 2136 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2137 | "pow": { |
| 2138 | "op": Op.POW, |
| 2139 | "operands": (2, 0), |
| 2140 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBasic, None), |
| 2141 | "types": TYPE_FP, |
| 2142 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2143 | "sub": { |
| 2144 | "op": Op.SUB, |
| 2145 | "operands": (2, 0), |
| 2146 | "build_fcn": (build_binary_broadcast, TosaTensorGen.tgBroadcastFuzz, None), |
| 2147 | "types": TYPE_FI32, |
| 2148 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2149 | "table": { |
| 2150 | "op": Op.TABLE, |
| 2151 | # Use the automatic generation functions to create the input array |
| 2152 | # but create the table tensor in the build function, as it may be |
| 2153 | # a different type from the input |
| 2154 | "operands": (1, 0), |
| 2155 | "build_fcn": (build_table, TosaTensorGen.tgBasic, None), |
| 2156 | "types": [DType.INT16], |
| 2157 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2158 | # Elementwise Unary operators |
| 2159 | "abs": { |
| 2160 | "op": Op.ABS, |
| 2161 | "operands": (1, 0), |
| 2162 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2163 | "types": TYPE_FI32, |
| 2164 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2165 | "bitwise_not": { |
| 2166 | "op": Op.BITWISE_NOT, |
| 2167 | "operands": (1, 0), |
| 2168 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2169 | "types": TYPE_INT, |
| 2170 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2171 | "ceil": { |
| 2172 | "op": Op.CEIL, |
| 2173 | "operands": (1, 0), |
| 2174 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2175 | "types": TYPE_FP, |
| 2176 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2177 | "clz": { |
| 2178 | "op": Op.CLZ, |
| 2179 | "operands": (1, 0), |
| 2180 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2181 | "types": [DType.INT32], |
| 2182 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2183 | "exp": { |
| 2184 | "op": Op.EXP, |
| 2185 | "operands": (1, 0), |
| 2186 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2187 | "types": TYPE_FP, |
| 2188 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2189 | "floor": { |
| 2190 | "op": Op.FLOOR, |
| 2191 | "operands": (1, 0), |
| 2192 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2193 | "types": TYPE_FP, |
| 2194 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2195 | "log": { |
| 2196 | "op": Op.LOG, |
| 2197 | "operands": (1, 0), |
| 2198 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2199 | "types": TYPE_FP, |
| 2200 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2201 | "logical_not": { |
| 2202 | "op": Op.LOGICAL_NOT, |
| 2203 | "operands": (1, 0), |
| 2204 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2205 | "types": TYPE_BOOL, |
| 2206 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2207 | "negate": { |
| 2208 | "op": Op.NEGATE, |
| 2209 | "operands": (1, 0), |
| 2210 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2211 | "qgen": TosaQuantGen.qgUnary, |
| 2212 | "types": TYPE_INT_FP, |
| 2213 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2214 | "reciprocal": { |
| 2215 | "op": Op.RECIPROCAL, |
| 2216 | "operands": (1, 0), |
| 2217 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2218 | "types": TYPE_FP, |
| 2219 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2220 | "rsqrt": { |
| 2221 | "op": Op.RSQRT, |
| 2222 | "operands": (1, 0), |
| 2223 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2224 | "types": TYPE_FP, |
| 2225 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2226 | # Elementwise Ternary operators |
| 2227 | "select": { |
| 2228 | "op": Op.SELECT, |
| 2229 | "operands": (3, 0), |
| 2230 | "build_fcn": (build_select, TosaTensorGen.tgBroadcastFuzz, None), |
| 2231 | "types": TYPE_FIB, |
| 2232 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2233 | # Comparison operators |
| 2234 | "equal": { |
| 2235 | "op": Op.EQUAL, |
| 2236 | "operands": (2, 0), |
| 2237 | "build_fcn": (build_comparison, TosaTensorGen.tgBroadcastFuzz, None), |
| 2238 | "types": TYPE_FI32, |
| 2239 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2240 | "greater_equal": { |
| 2241 | "op": Op.GREATER_EQUAL, |
| 2242 | "operands": (2, 0), |
| 2243 | "build_fcn": (build_comparison, TosaTensorGen.tgBroadcastFuzz, None), |
| 2244 | "types": TYPE_FI32, |
| 2245 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2246 | "greater": { |
| 2247 | "op": Op.GREATER, |
| 2248 | "operands": (2, 0), |
| 2249 | "build_fcn": (build_comparison, TosaTensorGen.tgBroadcastFuzz, None), |
| 2250 | "types": TYPE_FI32, |
| 2251 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2252 | # Reduction operators |
| 2253 | "reduce_all": { |
| 2254 | "op": Op.REDUCE_ALL, |
| 2255 | "operands": (1, 0), |
| 2256 | "build_fcn": (build_reduce, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 2257 | "types": TYPE_BOOL, |
| 2258 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2259 | "reduce_any": { |
| 2260 | "op": Op.REDUCE_ANY, |
| 2261 | "operands": (1, 0), |
| 2262 | "build_fcn": (build_reduce, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 2263 | "types": TYPE_BOOL, |
| 2264 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2265 | "reduce_max": { |
| 2266 | "op": Op.REDUCE_MAX, |
| 2267 | "operands": (1, 0), |
| 2268 | "build_fcn": (build_reduce, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 2269 | "types": TYPE_INT_FP, |
| 2270 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2271 | "reduce_min": { |
| 2272 | "op": Op.REDUCE_MAX, |
| 2273 | "operands": (1, 0), |
| 2274 | "build_fcn": (build_reduce, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 2275 | "types": TYPE_INT_FP, |
| 2276 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2277 | "reduce_product": { |
| 2278 | "op": Op.REDUCE_PRODUCT, |
| 2279 | "operands": (1, 0), |
| 2280 | "build_fcn": (build_reduce, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 2281 | "types": TYPE_FP, |
| 2282 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2283 | "reduce_sum": { |
| 2284 | "op": Op.REDUCE_SUM, |
| 2285 | "operands": (1, 0), |
| 2286 | "build_fcn": (build_reduce, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 2287 | "types": TYPE_FI32, |
| 2288 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2289 | # Data layout operators |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2290 | "concat": { |
| 2291 | "op": Op.CONCAT, |
| 2292 | "operands": (2, 0), |
| 2293 | "build_fcn": (build_concat, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 2294 | "types": TYPE_FIB, |
| 2295 | }, |
| 2296 | "pad": { |
| 2297 | "op": Op.PAD, |
| 2298 | "operands": (1, 0), |
| 2299 | "build_fcn": (build_pad, TosaTensorGen.tgBasic, TosaArgGen.agPad), |
| 2300 | "qgen": TosaQuantGen.qgPad, |
| 2301 | "types": TYPE_FIB, |
| 2302 | }, |
| 2303 | "reshape": { |
| 2304 | "op": Op.RESHAPE, |
| 2305 | "operands": (1, 0), |
| 2306 | "build_fcn": (build_reshape, TosaTensorGen.tgBasic, TosaArgGen.agReshape), |
| 2307 | "types": TYPE_FIB, |
| 2308 | }, |
| 2309 | "reverse": { |
| 2310 | "op": Op.REVERSE, |
| 2311 | "operands": (1, 0), |
| 2312 | "build_fcn": (build_reverse, TosaTensorGen.tgBasic, TosaArgGen.agAxis), |
| 2313 | "types": TYPE_FIB, |
| 2314 | }, |
| 2315 | "slice": { |
| 2316 | "op": Op.SLICE, |
| 2317 | "operands": (1, 0), |
| 2318 | "build_fcn": (build_slice, TosaTensorGen.tgBasic, TosaArgGen.agSlice), |
| 2319 | "types": TYPE_FIB, |
| 2320 | }, |
| 2321 | "tile": { |
| 2322 | "op": Op.TILE, |
| 2323 | "operands": (1, 0), |
| 2324 | "build_fcn": (build_tile, TosaTensorGen.tgBasic, TosaArgGen.agTile), |
| 2325 | "types": TYPE_FIB, |
| 2326 | }, |
| 2327 | "transpose": { |
| 2328 | "op": Op.TRANSPOSE, |
| 2329 | "operands": (1, 0), |
| 2330 | "rank": (2, 4), # Do not allow tranpose on rank=1 |
| 2331 | "build_fcn": ( |
| 2332 | build_transpose, |
| 2333 | TosaTensorGen.tgBasic, |
| 2334 | TosaArgGen.agTranspose, |
| 2335 | ), |
| 2336 | "types": TYPE_FIB, |
| 2337 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2338 | # Data nodes |
| 2339 | "const": { |
| 2340 | "op": Op.CONST, |
| 2341 | "operands": (1, 0), |
| 2342 | "build_fcn": (build_placeholder, TosaTensorGen.tgBasic, None), |
| 2343 | "types": TYPE_FIB, |
| 2344 | }, |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2345 | "identity": { |
| 2346 | "op": Op.IDENTITY, |
| 2347 | "operands": (1, 0), |
| 2348 | "build_fcn": (build_unary, TosaTensorGen.tgBasic, None), |
| 2349 | "types": TYPE_FIB, |
| 2350 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2351 | # Scatter/Gather |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2352 | "gather": { |
| 2353 | "op": Op.GATHER, |
| 2354 | # Only specify 'values' tensor here. 'indices' is generated in op building stage |
| 2355 | "operands": (1, 0), |
| 2356 | "rank": (3, 3), |
| 2357 | "build_fcn": (build_gather, TosaTensorGen.tgBasic, None), |
| 2358 | "types": TYPE_INT_FP, |
| 2359 | }, |
| 2360 | "scatter": { |
| 2361 | "op": Op.SCATTER, |
| 2362 | # Only specify 'values_in' tensor here. |
| 2363 | #'indices' and 'input' are generated in op building stage |
| 2364 | "operands": (2, 0), |
| 2365 | "rank": (3, 3), |
| 2366 | "build_fcn": (build_scatter, TosaTensorGen.tgScatter, None), |
| 2367 | "types": TYPE_INT_FP, |
| 2368 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2369 | # Image operations |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2370 | "resize": { |
| 2371 | "op": Op.RESIZE, |
| 2372 | "operands": (1, 0), |
| 2373 | "rank": (4, 4), |
| 2374 | "build_fcn": (build_resize, TosaTensorGen.tgNHWC, TosaArgGen.agResize), |
| 2375 | "types": [DType.INT8, DType.INT16, DType.FLOAT], |
| 2376 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2377 | # Type conversion |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2378 | "cast": { |
| 2379 | "op": Op.CAST, |
| 2380 | "operands": (1, 0), |
| 2381 | "build_fcn": (build_cast, TosaTensorGen.tgBasic, TosaArgGen.agCast), |
| 2382 | "types": [DType.FLOAT, DType.INT8, DType.INT16, DType.INT32, DType.BOOL], |
| 2383 | }, |
| 2384 | "rescale": { |
| 2385 | "op": Op.RESCALE, |
| 2386 | "operands": (1, 0), |
| 2387 | "build_fcn": (build_rescale, TosaTensorGen.tgBasic, TosaArgGen.agRescale), |
| 2388 | "types": [DType.INT8, DType.INT16, DType.INT32, DType.INT48], |
| 2389 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2390 | # Custom |
| 2391 | # Not implemented. |
Jared Smolens | 573ecd4 | 2021-03-04 15:24:10 -0800 | [diff] [blame] | 2392 | # Control flow operators |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2393 | # Two varients of cond_if, one that generates one of two constant tensors (no |
| 2394 | # inputs to the basic blocks, one output) and another that either adds or subtracts two tensors |
| 2395 | # (two inputs to the basic blocks, one output) |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2396 | "cond_if_const": { |
| 2397 | "op": Op.COND_IF, |
| 2398 | "operands": (0, 2), |
| 2399 | "build_fcn": ( |
| 2400 | build_cond_if_const, |
| 2401 | TosaTensorGen.tgBasic, |
| 2402 | TosaArgGen.agCondIf, |
| 2403 | ), |
| 2404 | "types": [DType.BOOL], |
| 2405 | }, |
| 2406 | "cond_if_binary": { |
| 2407 | "op": Op.COND_IF, |
| 2408 | "operands": (2, 0), |
| 2409 | "build_fcn": ( |
| 2410 | build_cond_if_binary, |
| 2411 | TosaTensorGen.tgBasic, |
| 2412 | TosaArgGen.agCondIf, |
| 2413 | ), |
| 2414 | "types": TYPE_FI32, |
| 2415 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2416 | # while_loop |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2417 | "while_loop": { |
| 2418 | "op": Op.WHILE_LOOP, |
| 2419 | "operands": (0, 1), |
| 2420 | "build_fcn": ( |
| 2421 | build_while_loop, |
| 2422 | TosaTensorGen.tgBasic, |
| 2423 | TosaArgGen.agWhileLoop, |
| 2424 | ), |
| 2425 | "types": [DType.INT32], |
| 2426 | }, |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2427 | } |
| 2428 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2429 | |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2430 | class OutputShaper: |
| 2431 | # Methods in this class compute the expected output shape and datatype |
| 2432 | # for common classes of operations |
| 2433 | def __init__(self): |
| 2434 | pass |
| 2435 | |
| 2436 | # These methods return arguments that can be used for |
| 2437 | # creating a new output tensor |
| 2438 | @staticmethod |
| 2439 | def binaryBroadcastOp(ser, a, b): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2440 | assert len(a.shape) == len(b.shape) |
| 2441 | assert a.dtype == b.dtype |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2442 | |
| 2443 | shape = [] |
| 2444 | for i in range(len(a.shape)): |
| 2445 | if a.shape[i] == 1: |
| 2446 | shape.append(b.shape[i]) |
| 2447 | else: |
| 2448 | shape.append(a.shape[i]) |
| 2449 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2450 | return ser.addOutput(shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2451 | |
| 2452 | @staticmethod |
| 2453 | def binaryNonBroadcastOp(ser, a, b): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2454 | assert len(a.shape) == len(b.shape) |
| 2455 | assert a.dtype == b.dtype |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2456 | |
| 2457 | shape = [] |
| 2458 | for i in range(len(a.shape)): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2459 | assert a.shape[i] == b.shape[i] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2460 | shape.append(a.shape[i]) |
| 2461 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2462 | return ser.addOutput(shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2463 | |
| 2464 | @staticmethod |
| 2465 | def unaryOp(ser, a): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2466 | return ser.addOutput(a.shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2467 | |
| 2468 | @staticmethod |
| 2469 | def selectOp(ser, cond, a, b): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2470 | assert len(a.shape) == len(b.shape) and len(a.shape) == len(cond.shape) |
| 2471 | assert a.dtype == b.dtype |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2472 | |
| 2473 | shape = [] |
| 2474 | for i in range(len(a.shape)): |
| 2475 | shape.append(max(cond.shape[i], a.shape[i], b.shape[i])) |
| 2476 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2477 | return ser.addOutput(shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2478 | |
| 2479 | @staticmethod |
| 2480 | def binaryComparisonOp(ser, a, b): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2481 | assert len(a.shape) == len(b.shape) |
| 2482 | assert a.dtype == b.dtype |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2483 | |
| 2484 | # Do broadcast |
| 2485 | shape = [] |
| 2486 | for i in range(len(a.shape)): |
| 2487 | if a.shape[i] == 1: |
| 2488 | shape.append(b.shape[i]) |
| 2489 | else: |
| 2490 | shape.append(a.shape[i]) |
| 2491 | |
| 2492 | # Force the output type to bool |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2493 | return ser.addOutput(shape, DType.BOOL) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2494 | |
| 2495 | @staticmethod |
| 2496 | def reduceOp(ser, a, axis): |
| 2497 | |
| 2498 | shape = a.shape.copy() |
| 2499 | |
| 2500 | shape[axis] = 1 |
| 2501 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2502 | return ser.addOutput(shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2503 | |
| 2504 | @staticmethod |
| 2505 | def argmaxOp(ser, a, axis): |
| 2506 | shape = a.shape.copy() |
| 2507 | del shape[axis] |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2508 | return ser.addOutput(shape, DType.INT32) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2509 | |
| 2510 | @staticmethod |
| 2511 | def conv2dOp(ser, ifm, filter, strides, padding, dilations): |
| 2512 | |
| 2513 | # IFM: NHWC |
| 2514 | # Filter: OHWI |
| 2515 | # OFM: NHWC |
| 2516 | |
| 2517 | if len(padding) == 2: |
| 2518 | # Expand padding to 4 parameters in the case of transpose_conv2d |
| 2519 | # From H,W to T,B,L,R |
| 2520 | padding = [padding[0], padding[0], padding[1], padding[1]] |
| 2521 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2522 | h = ( |
| 2523 | ifm.shape[1] |
| 2524 | - filter.shape[1] |
| 2525 | - (filter.shape[1] - 1) * (dilations[0] - 1) |
| 2526 | + padding[0] |
| 2527 | + padding[1] |
| 2528 | ) // strides[0] + 1 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2529 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2530 | w = ( |
| 2531 | ifm.shape[2] |
| 2532 | - filter.shape[2] |
| 2533 | - (filter.shape[2] - 1) * (dilations[1] - 1) |
| 2534 | + padding[2] |
| 2535 | + padding[3] |
| 2536 | ) // strides[1] + 1 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2537 | |
| 2538 | if h <= 0 or w <= 0: |
| 2539 | # Invalid test parameters? |
| 2540 | h = 0 |
| 2541 | w = 0 |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2542 | ser.setExpectedFailure(True, "Invalid combination of conv2d parameters") |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2543 | |
| 2544 | ofm_shape = [ifm.shape[0], h, w, filter.shape[0]] |
| 2545 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 2546 | if ifm.dtype == DType.INT8: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2547 | out_dtype = DType.INT32 |
| 2548 | elif ifm.dtype == DType.INT16: |
| 2549 | out_dtype = DType.INT48 |
| 2550 | elif ifm.dtype == DType.FLOAT: |
| 2551 | out_dtype = DType.FLOAT |
| 2552 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2553 | raise Exception("Unsupported input dtype: {}".format(ifm.dtype)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2554 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2555 | return ser.addOutput(ofm_shape, out_dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2556 | |
| 2557 | @staticmethod |
| 2558 | def depthwiseConv2dOp(ser, ifm, filter, strides, padding, dilations): |
| 2559 | # IFM: NHWC |
| 2560 | # Filter: HWCM |
| 2561 | # OFM: NHW C*M |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2562 | h = ( |
| 2563 | ifm.shape[1] |
| 2564 | - filter.shape[0] |
| 2565 | - (filter.shape[0] - 1) * (dilations[0] - 1) |
| 2566 | + padding[0] |
| 2567 | + padding[1] |
| 2568 | ) // strides[0] + 1 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2569 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2570 | w = ( |
| 2571 | ifm.shape[2] |
| 2572 | - filter.shape[1] |
| 2573 | - (filter.shape[1] - 1) * (dilations[1] - 1) |
| 2574 | + padding[2] |
| 2575 | + padding[3] |
| 2576 | ) // strides[1] + 1 |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2577 | |
| 2578 | if h <= 0 or w <= 0: |
| 2579 | # Invalid test parameters? |
| 2580 | h = 0 |
| 2581 | w = 0 |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2582 | ser.setExpectedFailure(True, "Invalid combination of conv2d parameters") |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2583 | |
| 2584 | ofm_shape = [ifm.shape[0], h, w, filter.shape[2] * filter.shape[3]] |
| 2585 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 2586 | if ifm.dtype == DType.INT8: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2587 | out_dtype = DType.INT32 |
| 2588 | elif ifm.dtype == DType.INT16: |
| 2589 | out_dtype = DType.INT48 |
| 2590 | elif ifm.dtype == DType.FLOAT: |
| 2591 | out_dtype = DType.FLOAT |
| 2592 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2593 | raise Exception("Unsupported input dtype: {}".format(ifm.dtype)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2594 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2595 | return ser.addOutput(ofm_shape, out_dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2596 | |
| 2597 | @staticmethod |
| 2598 | def pool2dOp(ser, ifm, kernel, stride, pad): |
| 2599 | # input: NHWC |
| 2600 | h = (ifm.shape[1] + pad[0] + pad[1] + stride[0] - kernel[0]) // stride[0] |
| 2601 | w = (ifm.shape[2] + pad[2] + pad[3] + stride[1] - kernel[1]) // stride[1] |
| 2602 | |
| 2603 | if h <= 0 or w <= 0: |
| 2604 | # Invalid test parameters? |
| 2605 | h = 0 |
| 2606 | w = 0 |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2607 | ser.setExpectedFailure(True, "Invalid combination of pooling parameters") |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2608 | |
| 2609 | ofm_shape = [ifm.shape[0], h, w, ifm.shape[3]] |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2610 | return ser.addOutput(ofm_shape, ifm.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2611 | |
| 2612 | @staticmethod |
| 2613 | def fullyConnectedOp(ser, input, filter): |
| 2614 | # input: N, IC |
| 2615 | # filter: OC, IC |
| 2616 | # output: N, OC |
| 2617 | |
| 2618 | output_shape = [input.shape[0], filter.shape[0]] |
| 2619 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 2620 | if input.dtype == DType.INT8: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2621 | out_dtype = DType.INT32 |
| 2622 | elif input.dtype == DType.INT16: |
| 2623 | out_dtype = DType.INT48 |
| 2624 | elif input.dtype == DType.FLOAT: |
| 2625 | out_dtype = DType.FLOAT |
| 2626 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2627 | raise Exception("Unsupported input dtype: {}".format(input.dtype)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2628 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2629 | return ser.addOutput(output_shape, out_dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2630 | |
| 2631 | @staticmethod |
| 2632 | def matmulOp(ser, a, b): |
| 2633 | # a: M, K |
| 2634 | # b: K, N |
| 2635 | # out: M, N |
| 2636 | |
| 2637 | output_shape = [a.shape[0], b.shape[1]] |
| 2638 | |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 2639 | if a.dtype == DType.INT8: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2640 | out_dtype = DType.INT32 |
| 2641 | elif a.dtype == DType.INT16: |
| 2642 | out_dtype = DType.INT48 |
| 2643 | elif a.dtype == DType.FLOAT: |
| 2644 | out_dtype = DType.FLOAT |
| 2645 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2646 | raise Exception("UNsupported input dtype for matmul: {}".format(a.dtype)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2647 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2648 | return ser.addOutput(output_shape, out_dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2649 | |
| 2650 | @staticmethod |
| 2651 | def concatOp(ser, a, b, axis): |
| 2652 | |
| 2653 | output_shape = a.shape.copy() |
| 2654 | output_shape[axis] = a.shape[axis] + b.shape[axis] |
| 2655 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2656 | return ser.addOutput(output_shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2657 | |
| 2658 | @staticmethod |
| 2659 | def padOp(ser, a, padding): |
| 2660 | |
| 2661 | output_shape = a.shape.copy() |
| 2662 | |
| 2663 | for i in range(len(output_shape)): |
| 2664 | output_shape[i] = padding[i][0] + padding[i][1] + output_shape[i] |
| 2665 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2666 | return ser.addOutput(output_shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2667 | |
| 2668 | @staticmethod |
| 2669 | def reshapeOp(ser, a, shape): |
| 2670 | output_shape = shape.copy() |
| 2671 | |
| 2672 | totalElements = 1 |
| 2673 | for i in a.shape: |
| 2674 | totalElements *= i |
| 2675 | |
| 2676 | # If there are any -1 elements, figure out what that dimension must be |
| 2677 | totalOutputElements = 1 |
| 2678 | for i in output_shape: |
| 2679 | if i != -1: |
| 2680 | totalOutputElements *= i |
| 2681 | |
| 2682 | # And fill it in |
| 2683 | for i in range(len(output_shape)): |
| 2684 | if output_shape[i] == -1: |
| 2685 | output_shape[i] = totalElements // totalOutputElements |
| 2686 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2687 | return ser.addOutput(output_shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2688 | |
| 2689 | @staticmethod |
| 2690 | def sliceOp(ser, a, begin, size): |
| 2691 | |
| 2692 | output_shape = size.copy() |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2693 | return ser.addOutput(output_shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2694 | |
| 2695 | @staticmethod |
| 2696 | def tileOp(ser, a, multiples): |
| 2697 | |
| 2698 | output_shape = a.shape.copy() |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2699 | assert len(multiples) == len(output_shape) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2700 | |
| 2701 | for i in range(len(output_shape)): |
| 2702 | output_shape[i] = a.shape[i] * multiples[i] |
| 2703 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2704 | return ser.addOutput(output_shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2705 | |
| 2706 | @staticmethod |
| 2707 | def transposeOp(ser, a, perms): |
| 2708 | output_shape = a.shape.copy() |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2709 | assert len(perms) == len(output_shape) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2710 | |
| 2711 | for i in range(len(output_shape)): |
| 2712 | output_shape[i] = a.shape[perms[i]] |
| 2713 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2714 | return ser.addOutput(output_shape, a.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2715 | |
| 2716 | @staticmethod |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 2717 | def gatherOp(ser, values, indices): |
| 2718 | assert len(values.shape) == 3 |
| 2719 | assert len(indices.shape) == 2 |
| 2720 | assert values.shape[0] == indices.shape[0] |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2721 | |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 2722 | output_shape = [values.shape[0], indices.shape[1], values.shape[2]] |
| 2723 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2724 | return ser.addOutput(output_shape, values.dtype) |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 2725 | |
| 2726 | @staticmethod |
| 2727 | def scatterOp(ser, values_in, indices, input): |
| 2728 | assert len(values_in.shape) == 3 |
| 2729 | assert len(indices.shape) == 2 |
| 2730 | assert len(input.shape) == 3 |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2731 | assert values_in.shape[0] == indices.shape[0] # N |
| 2732 | assert input.shape[1] == indices.shape[1] # W |
| 2733 | assert values_in.shape[2] == input.shape[2] # C |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 2734 | |
| 2735 | output_shape = values_in.shape |
| 2736 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2737 | return ser.addOutput(output_shape, values_in.dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2738 | |
| 2739 | @staticmethod |
| 2740 | def tableOp(ser, input, table): |
| 2741 | # Same shape as the input, but with the type of the table. |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2742 | return ser.addOutput(input.shape, DType.INT32) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2743 | |
| 2744 | @staticmethod |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2745 | def resizeOp( |
| 2746 | ser, |
| 2747 | input, |
| 2748 | mode, |
| 2749 | stride, |
| 2750 | offset, |
| 2751 | shift, |
| 2752 | stride_fp, |
| 2753 | offset_fp, |
| 2754 | output_dims, |
| 2755 | input_dtype, |
| 2756 | output_dtype, |
| 2757 | ): |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2758 | |
| 2759 | output_dims = [input.shape[0], output_dims[0], output_dims[1], input.shape[3]] |
| 2760 | |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 2761 | if input_dtype == DType.FLOAT: |
| 2762 | if stride_fp[0] <= 0 or stride_fp[1] <= 0: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2763 | ser.setExpectedFailure(True, "Negative or zero stride") |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 2764 | else: |
| 2765 | if stride[0] <= 0 or stride[1] <= 0: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2766 | ser.setExpectedFailure(True, "Negative or zero stride") |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2767 | |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 2768 | if mode == ResizeMode.BILINEAR: |
| 2769 | if input_dtype == DType.INT8: |
| 2770 | if output_dtype != DType.INT32: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2771 | ser.setExpectedFailure(True, "Invalid output data type") |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 2772 | elif input_dtype == DType.INT16: |
| 2773 | if output_dtype != DType.INT48: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2774 | ser.setExpectedFailure(true, "Invalid output data type") |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 2775 | elif input_dtype == DType.FLOAT: |
| 2776 | if output_dtype != DType.FLOAT: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2777 | ser.setExpectedFailure(true, "Invalid output data type") |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 2778 | else: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2779 | ser.setExpectedFailure(true, "Invalid input data type") |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 2780 | |
| 2781 | elif mode == ResizeMode.NEAREST: |
| 2782 | if input_dtype == DType.INT8: |
| 2783 | if output_dtype != DType.INT8: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2784 | ser.setExpectedFailure(True, "Invalid output data type") |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 2785 | elif input_dtype == DType.INT16: |
| 2786 | if output_dtype != DType.INT16: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2787 | ser.setExpectedFailure(true, "Invalid output data type") |
Kevin Cheng | 77d0f76 | 2020-11-24 10:26:32 -0800 | [diff] [blame] | 2788 | elif input_dtype == DType.FLOAT: |
| 2789 | if output_dtype != DType.FLOAT: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2790 | ser.setExpectedFailure(true, "Invalid output data type") |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 2791 | else: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2792 | ser.setExpectedFailure(true, "Invalid input data type") |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 2793 | |
| 2794 | else: |
Kevin Cheng | 989cb05 | 2021-04-28 16:29:44 -0700 | [diff] [blame] | 2795 | ser.setExpectedFailure(true, "Invalid resize mode") |
Kevin Cheng | aee1fac | 2020-11-11 13:54:06 -0800 | [diff] [blame] | 2796 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2797 | return ser.addOutput(output_dims, output_dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2798 | |
| 2799 | @staticmethod |
| 2800 | def typeConversionOp(ser, val, out_dtype): |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2801 | return ser.addOutput(val.shape, out_dtype) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2802 | |
| 2803 | @staticmethod |
| 2804 | def transposeConv2DOp(ser, ifm, output_shape): |
Kevin Cheng | 3a47857 | 2021-01-22 17:21:02 -0800 | [diff] [blame] | 2805 | if ifm.dtype == DType.INT8: |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2806 | out_dtype = DType.INT32 |
| 2807 | elif ifm.dtype == DType.INT16: |
| 2808 | out_dtype = DType.INT48 |
| 2809 | elif ifm.dtype == DType.FLOAT: |
| 2810 | out_dtype = DType.FLOAT |
| 2811 | else: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2812 | raise Exception("Unsupported input dtype: {}".format(ifm.dtype)) |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2813 | |
| 2814 | if output_shape[1] <= 0 or output_shape[2] <= 0: |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2815 | ser.setExpectedFailure(True, "Negative output shape") |
Eric Kunze | e5e2676 | 2020-10-13 16:11:07 -0700 | [diff] [blame] | 2816 | |
Kevin Cheng | 550ccc5 | 2021-03-03 11:21:43 -0800 | [diff] [blame] | 2817 | return ser.addOutput(output_shape, out_dtype) |