Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1 | # Copyright (c) 2021-2022, ARM Limited. |
| 2 | # SPDX-License-Identifier: Apache-2.0 |
| 3 | import itertools |
| 4 | import math |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 5 | import warnings |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 6 | |
| 7 | import numpy as np |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 8 | from generator.tosa_error_if import ErrorIf |
| 9 | from generator.tosa_error_if import TosaErrorIfArgGen |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 10 | from generator.tosa_utils import get_accum_dtype_from_tgTypes |
| 11 | from generator.tosa_utils import get_wrong_output_type |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 12 | from generator.tosa_utils import MAX_RESIZE_DIMENSION |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 13 | from serializer.tosa_serializer import DTypeNames |
| 14 | from tosa.DType import DType |
| 15 | from tosa.Op import Op |
| 16 | from tosa.ResizeMode import ResizeMode |
| 17 | |
| 18 | # DTypeNames, DType, Op and ResizeMode are convenience variables to the |
| 19 | # flatc-generated types that should be enums, but aren't |
| 20 | |
| 21 | |
| 22 | class TosaQuantGen: |
| 23 | """QuantizedInfo random generator helper functions. |
| 24 | |
| 25 | Specify with 'qgen': in the operator defintion. |
| 26 | """ |
| 27 | |
| 28 | def __init__(self): |
| 29 | pass |
| 30 | |
| 31 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 32 | def getZeroPoint(testGen, dtype, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 33 | |
| 34 | if dtype == DType.INT8: |
Jeremy Johnson | 0042343 | 2022-09-12 17:27:37 +0100 | [diff] [blame] | 35 | if testGen.args.zeropoint is not None: |
| 36 | return min(127, max(-128, testGen.args.zeropoint)) |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 37 | return testGen.randInt(-128, 128) |
| 38 | elif dtype == DType.UINT8: |
Jeremy Johnson | 0042343 | 2022-09-12 17:27:37 +0100 | [diff] [blame] | 39 | if testGen.args.zeropoint is not None: |
| 40 | return min(255, max(0, testGen.args.zeropoint)) |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 41 | return testGen.randInt(0, 256) |
| 42 | elif error_name in [ |
| 43 | ErrorIf.InputZeroPointNotZero, |
| 44 | ErrorIf.WeightZeroPointNotZero, |
| 45 | ErrorIf.OutputZeroPointNotZero, |
| 46 | ]: |
| 47 | zero_point = testGen.randInt(-128, 128) |
| 48 | if zero_point == 0: |
| 49 | zero_point = 1 |
| 50 | return zero_point |
| 51 | return 0 |
| 52 | |
| 53 | @staticmethod |
| 54 | def qgUnary(testGen, op, dtype, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 55 | if error_name == ErrorIf.InputZeroPointNotZero: |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 56 | qinfo = [ |
| 57 | TosaQuantGen.getZeroPoint(testGen, dtype, error_name), |
| 58 | TosaQuantGen.getZeroPoint(testGen, dtype), |
| 59 | ] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 60 | elif error_name == ErrorIf.OutputZeroPointNotZero: |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 61 | qinfo = [ |
| 62 | TosaQuantGen.getZeroPoint(testGen, dtype), |
| 63 | TosaQuantGen.getZeroPoint(testGen, dtype, error_name), |
| 64 | ] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 65 | else: |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 66 | qinfo = [ |
| 67 | TosaQuantGen.getZeroPoint(testGen, dtype), |
| 68 | TosaQuantGen.getZeroPoint(testGen, dtype), |
| 69 | ] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 70 | return qinfo |
| 71 | |
| 72 | @staticmethod |
| 73 | def qgConv(testGen, op, dtype_or_dtypeList, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 74 | if isinstance(dtype_or_dtypeList, list): |
| 75 | # a list of [input, weights, accumulator] dtypes |
| 76 | dtypeList = dtype_or_dtypeList |
| 77 | else: |
| 78 | # an int, [input, weights, accumulator] dtypes are the same |
| 79 | dtypeList = [dtype_or_dtypeList] * 3 |
| 80 | |
| 81 | if error_name == ErrorIf.InputZeroPointNotZero: |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 82 | qinfo = [ |
| 83 | TosaQuantGen.getZeroPoint(testGen, dtypeList[0], error_name), |
| 84 | TosaQuantGen.getZeroPoint(testGen, dtypeList[1]), |
| 85 | ] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 86 | elif error_name == ErrorIf.WeightZeroPointNotZero: |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 87 | qinfo = [ |
| 88 | TosaQuantGen.getZeroPoint(testGen, dtypeList[0]), |
| 89 | TosaQuantGen.getZeroPoint(testGen, dtypeList[1], error_name), |
| 90 | ] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 91 | else: |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 92 | qinfo = [ |
| 93 | TosaQuantGen.getZeroPoint(testGen, dtypeList[0]), |
| 94 | TosaQuantGen.getZeroPoint(testGen, dtypeList[1]), |
| 95 | ] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 96 | return qinfo |
| 97 | |
| 98 | @staticmethod |
| 99 | def qgMatmul(testGen, op, dtype, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 100 | if error_name == ErrorIf.InputZeroPointNotZero: |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 101 | qinfo = [ |
| 102 | TosaQuantGen.getZeroPoint(testGen, dtype, error_name), |
| 103 | TosaQuantGen.getZeroPoint(testGen, dtype, error_name), |
| 104 | ] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 105 | else: |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 106 | qinfo = [ |
| 107 | TosaQuantGen.getZeroPoint(testGen, dtype), |
| 108 | TosaQuantGen.getZeroPoint(testGen, dtype), |
| 109 | ] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 110 | return qinfo |
| 111 | |
| 112 | @staticmethod |
| 113 | def computeMultiplierAndShift(scaleFp, scale32): |
| 114 | # Derived from computeMultiplierAndShiftTosaScale32 |
| 115 | # Provide a floating-point scaling factor and the scale32 parameter |
| 116 | # to compute the multiplier and shift |
| 117 | |
| 118 | if scale32: |
| 119 | scaleBits = 31 |
| 120 | else: |
| 121 | scaleBits = 15 |
| 122 | |
| 123 | m, shift = math.frexp(scaleFp) |
| 124 | |
| 125 | if scaleFp < 0.0: |
| 126 | m = -m |
| 127 | |
| 128 | multiplier = round(m * (1 << scaleBits)) |
| 129 | assert multiplier <= (1 << scaleBits) |
| 130 | |
| 131 | if multiplier == (1 << scaleBits): |
| 132 | multiplier = multiplier // 2 |
| 133 | shift = shift + 1 |
| 134 | |
| 135 | shift = (-shift) + scaleBits |
| 136 | # print('scalefp {} scaleBits {} m {} mult {} shift {}'.format( |
| 137 | # scaleFp, scaleBits, m, multiplier, shift)) |
| 138 | |
| 139 | # Adjust multiplier such that shift is in allowed value range. |
| 140 | if shift == 0: |
| 141 | multiplier = multiplier // 4 |
| 142 | shift = shift + 2 |
| 143 | elif shift == 1: |
| 144 | multiplier = multiplier // 2 |
| 145 | shift = shift + 1 |
| 146 | elif shift == 63: |
| 147 | multiplier = multiplier * 2 |
| 148 | shift = shift - 1 |
| 149 | |
| 150 | assert multiplier <= (1 << scaleBits) |
| 151 | assert shift >= 2 and shift <= 62 |
| 152 | |
| 153 | return multiplier, shift |
| 154 | |
| 155 | |
| 156 | class TosaTensorGen: |
| 157 | """Tensor generators create a shape list for the placeholder and const tensor |
| 158 | data operands for the operator. |
| 159 | |
| 160 | The actual random data is generated separately for each test. |
| 161 | """ |
| 162 | |
| 163 | def __init__(self): |
| 164 | pass |
| 165 | |
| 166 | @staticmethod |
| 167 | def tgBasic(testGen, opName, rank, error_name=None): |
| 168 | pl, const = opName["operands"] |
| 169 | shape = testGen.makeShape(rank) |
| 170 | |
| 171 | # Constrict the overall size of the shape when creating ERROR_IF tests |
| 172 | if error_name: |
| 173 | shape = TosaErrorIfArgGen.eiRestrictDimensions(shape) |
| 174 | |
| 175 | shape_list = [] |
| 176 | for i in range(pl + const): |
| 177 | shape_list.append(shape.copy()) |
| 178 | |
| 179 | if error_name == ErrorIf.RankMismatch: |
| 180 | if rank == 1 and i != 1: |
| 181 | shape = testGen.makeShape(rank + testGen.rng.choice([1, 2, 3])) |
| 182 | elif i != 1: |
| 183 | shape = testGen.makeShape(rank + testGen.rng.choice([-1, 1])) |
| 184 | |
| 185 | return shape_list |
| 186 | |
| 187 | @staticmethod |
| 188 | def tgNHWC(testGen, opName, rank, error_name=None): |
| 189 | pl, const = opName["operands"] |
| 190 | |
| 191 | if error_name != ErrorIf.WrongRank: |
| 192 | assert rank == 4 |
| 193 | |
| 194 | shape = testGen.makeShape(rank) |
| 195 | |
| 196 | # Constrict the batch size? |
| 197 | if testGen.args.max_batch_size: |
| 198 | shape[0] = (shape[0] % testGen.args.max_batch_size) + 1 |
| 199 | |
| 200 | # Constrict the overall size of the shape when creating ERROR_IF tests |
| 201 | if error_name and error_name != ErrorIf.MaxDimExceeded: |
| 202 | shape = TosaErrorIfArgGen.eiRestrictDimensions(shape) |
| 203 | |
| 204 | shape_list = [] |
| 205 | for i in range(pl + const): |
| 206 | shape_list.append(shape.copy()) |
| 207 | |
| 208 | return shape_list |
| 209 | |
| 210 | @staticmethod |
| 211 | def tgScatter(testGen, opName, rank, error_name=None): |
| 212 | pl, const = opName["operands"] |
| 213 | |
| 214 | assert pl == 2 |
| 215 | assert const == 0 |
| 216 | if error_name != ErrorIf.WrongRank: |
| 217 | assert rank == 3 |
| 218 | |
| 219 | values_in_shape = testGen.makeShape(rank) |
| 220 | |
| 221 | # ignore max batch size if target shape is set |
| 222 | if testGen.args.max_batch_size and not testGen.args.target_shapes: |
| 223 | values_in_shape[0] = (values_in_shape[0] % testGen.args.max_batch_size) + 1 |
| 224 | |
| 225 | W = testGen.randInt( |
| 226 | testGen.args.tensor_shape_range[0], testGen.args.tensor_shape_range[1] |
| 227 | ) |
| 228 | # Constrict W if one dimension is too large to keep tensor size reasonable |
| 229 | if max(values_in_shape) > 5000: |
| 230 | W = testGen.randInt(0, 16) |
| 231 | |
| 232 | input_shape = [values_in_shape[0], W, values_in_shape[2]] |
| 233 | |
| 234 | shape_list = [] |
| 235 | shape_list.append(values_in_shape.copy()) |
| 236 | shape_list.append(input_shape.copy()) |
| 237 | |
| 238 | return shape_list |
| 239 | |
| 240 | @staticmethod |
| 241 | def tgBroadcastFuzz(testGen, op, rank, error_name=None): |
| 242 | shape = testGen.makeShape(rank) |
| 243 | |
| 244 | pl, const = op["operands"] |
| 245 | |
| 246 | shape_list = [] |
| 247 | |
| 248 | # Choose one of the inputs to broadcast |
| 249 | # Note: Simplifies OutputShaper code if we don't change first shape for errors |
| 250 | bcast_idx = testGen.randInt(0 if error_name is None else 1, pl + const) |
| 251 | for i in range(pl + const): |
| 252 | shape_bcast = shape.copy() |
| 253 | |
| 254 | # If the chosen input, pick a random index to broadcast |
| 255 | if i == bcast_idx: |
| 256 | fuzz_idx = testGen.randInt(0, rank) |
| 257 | if error_name == ErrorIf.DimensionMismatch: |
| 258 | shape_bcast[fuzz_idx] += 1 |
| 259 | elif error_name == ErrorIf.RankMismatch: |
| 260 | # Add one rank to the shape (or more for rank of 1) |
| 261 | extra_ranks = testGen.rng.choice([1, 2, 3]) if rank == 1 else 1 |
| 262 | shape_bcast = np.concatenate( |
| 263 | (shape_bcast, testGen.makeShape(extra_ranks)) |
| 264 | ) |
| 265 | if rank != 1: |
| 266 | # Either keep the extra rank, or remove it |
| 267 | new_len = testGen.rng.choice([-2, len(shape_bcast)]) |
| 268 | shape_bcast = shape_bcast[:new_len] |
| 269 | else: |
| 270 | shape_bcast[fuzz_idx] = 1 |
| 271 | |
| 272 | shape_list.append(shape_bcast) |
| 273 | |
| 274 | return shape_list |
| 275 | |
| 276 | @staticmethod |
| 277 | def tgConv2D(testGen, op, rank, error_name=None): |
| 278 | pl, const = op["operands"] |
| 279 | |
| 280 | if error_name != ErrorIf.WrongRank: |
| 281 | assert rank == 4 |
| 282 | |
| 283 | # IFM dimensions are NHWC |
| 284 | ifm_shape = testGen.makeShape(rank) |
| 285 | |
| 286 | # Constrict the batch size? |
| 287 | if testGen.args.max_batch_size: |
| 288 | ifm_shape[0] = (ifm_shape[0] % testGen.args.max_batch_size) + 1 |
| 289 | |
| 290 | # Constrict the overall size of the shape when creating ERROR_IF tests |
| 291 | if error_name: |
| 292 | ifm_shape = TosaErrorIfArgGen.eiRestrictDimensions( |
| 293 | ifm_shape, max_dim=24, max_items=10000 |
| 294 | ) |
| 295 | |
| 296 | # Get the filter height/width from the operator parameters |
| 297 | filter_hw = op["filter"] |
| 298 | |
| 299 | # Generate a random OFM depth |
| 300 | ofm_depth = testGen.makeShape(1)[0] |
| 301 | |
| 302 | # The filter dimensions are OHWI |
| 303 | filter_shape = np.asarray([ofm_depth, filter_hw[0], filter_hw[1], ifm_shape[3]]) |
| 304 | |
| 305 | # The bias is OC |
| 306 | bias_shape = np.asarray([ofm_depth]) |
| 307 | |
| 308 | return [ifm_shape, filter_shape, bias_shape] |
| 309 | |
| 310 | @staticmethod |
| 311 | def tgConv3D(testGen, op, rank, error_name=None): |
| 312 | pl, const = op["operands"] |
| 313 | |
| 314 | if error_name != ErrorIf.WrongRank: |
| 315 | assert rank == 5 |
| 316 | |
| 317 | # IFM dimensions are NDHWC |
| 318 | ifm_shape = testGen.makeShape(rank) |
| 319 | |
| 320 | # Constrict the batch size? |
| 321 | if testGen.args.max_batch_size: |
| 322 | ifm_shape[0] = (ifm_shape[0] % testGen.args.max_batch_size) + 1 |
| 323 | |
| 324 | # Constrict the overall size of the shape when creating ERROR_IF tests |
| 325 | if error_name: |
| 326 | ifm_shape = TosaErrorIfArgGen.eiRestrictDimensions( |
| 327 | ifm_shape, max_dim=24, max_items=10000 |
| 328 | ) |
| 329 | |
| 330 | # Get the filter depth/height/width from the operator parameters |
| 331 | filter_dhw = op["filter"] |
| 332 | |
| 333 | # Generate a random OFM channel |
| 334 | ofm_channel = testGen.makeShape(1)[0] |
| 335 | |
| 336 | # The filter dimensions are ODHWI |
| 337 | filter_shape = np.asarray( |
| 338 | [ofm_channel, filter_dhw[0], filter_dhw[1], filter_dhw[2], ifm_shape[4]] |
| 339 | ) |
| 340 | |
| 341 | # The bias is OC |
| 342 | bias_shape = np.asarray([ofm_channel]) |
| 343 | |
| 344 | return [ifm_shape, filter_shape, bias_shape] |
| 345 | |
| 346 | @staticmethod |
| 347 | def tgTransposeConv2D(testGen, op, rank, error_name=None): |
| 348 | pl, const = op["operands"] |
| 349 | |
| 350 | if error_name != ErrorIf.WrongRank: |
| 351 | assert rank == 4 |
| 352 | |
| 353 | # IFM dimensions are NHWC |
| 354 | ifm_shape = testGen.makeShape(rank) |
| 355 | |
| 356 | # Constrict the batch size? |
| 357 | if testGen.args.max_batch_size: |
| 358 | ifm_shape[0] = (ifm_shape[0] % testGen.args.max_batch_size) + 1 |
| 359 | |
| 360 | # Constrict the overall size of the shape when creating ERROR_IF tests |
| 361 | if error_name: |
| 362 | ifm_shape = TosaErrorIfArgGen.eiRestrictDimensions( |
| 363 | ifm_shape, max_dim=24, max_items=10000 |
| 364 | ) |
| 365 | |
| 366 | # Get the filter height/width from the operator parameters |
| 367 | filter_hw = op["filter"] |
| 368 | |
| 369 | # Generate a random OFM depth |
| 370 | ofm_depth = testGen.makeShape(1)[0] |
| 371 | |
| 372 | # The filter dimensions are OHWI |
| 373 | filter_shape = np.asarray([ofm_depth, filter_hw[0], filter_hw[1], ifm_shape[3]]) |
| 374 | |
| 375 | # The bias is OC |
| 376 | bias_shape = np.asarray([ofm_depth]) |
| 377 | |
| 378 | return [ifm_shape, filter_shape, bias_shape] |
| 379 | |
| 380 | @staticmethod |
| 381 | def tgDepthwiseConv2D(testGen, op, rank, error_name=None): |
| 382 | pl, const = op["operands"] |
| 383 | |
| 384 | if error_name != ErrorIf.WrongRank: |
| 385 | assert rank == 4 |
| 386 | assert pl == 1 and const == 2 |
| 387 | |
| 388 | # IFM dimensions are NHWC |
| 389 | ifm_shape = testGen.makeShape(rank) |
| 390 | |
| 391 | # Constrict the batch size? |
| 392 | if testGen.args.max_batch_size: |
| 393 | ifm_shape[0] = (ifm_shape[0] % testGen.args.max_batch_size) + 1 |
| 394 | |
| 395 | # Constrict the overall size of the shape when creating ERROR_IF tests |
| 396 | if error_name: |
| 397 | ifm_shape = TosaErrorIfArgGen.eiRestrictDimensions( |
| 398 | ifm_shape, max_dim=24, max_items=10000 |
| 399 | ) |
| 400 | |
| 401 | # Get the filter height/width from the operator parameters |
| 402 | # Filter is KH, HW, C, M |
| 403 | filter_hw = op["filter"] |
| 404 | |
| 405 | # Generate a random OFM depth, but don't let it get too big because |
| 406 | # the output depth is M * C |
| 407 | filter_m = ( |
| 408 | testGen.makeShape(1)[0] % (testGen.args.tensor_shape_range[1] // 4) |
| 409 | ) + 1 |
| 410 | |
| 411 | # The filter dimensions are HWCM |
| 412 | filter_shape = np.asarray([filter_hw[0], filter_hw[1], ifm_shape[3], filter_m]) |
| 413 | |
| 414 | # The bias is M * C |
| 415 | bias_shape = np.asarray([ifm_shape[3] * filter_m]) |
| 416 | |
| 417 | return [ifm_shape, filter_shape, bias_shape] |
| 418 | |
| 419 | @staticmethod |
| 420 | def tgFullyConnected(testGen, op, rank, error_name=None): |
| 421 | pl, const = op["operands"] |
| 422 | |
| 423 | if error_name != ErrorIf.WrongRank: |
| 424 | assert rank == 2 |
| 425 | |
| 426 | input_shape = testGen.makeShape(rank) |
| 427 | |
| 428 | # Constrict the overall size of the shape when creating ERROR_IF tests |
| 429 | if error_name: |
| 430 | input_shape = TosaErrorIfArgGen.eiRestrictDimensions(input_shape) |
| 431 | |
| 432 | filter_oc = testGen.rng.integers( |
| 433 | low=testGen.args.tensor_shape_range[0], |
| 434 | high=testGen.args.tensor_shape_range[1], |
| 435 | size=1, |
| 436 | )[0] |
| 437 | filter_shape = np.asarray([filter_oc, input_shape[1]]) |
| 438 | |
| 439 | bias_shape = np.asarray([filter_oc]) |
| 440 | |
| 441 | return [input_shape, filter_shape, bias_shape] |
| 442 | |
| 443 | @staticmethod |
| 444 | def tgMatmul(testGen, op, rank, error_name=None): |
| 445 | pl, const = op["operands"] |
| 446 | |
| 447 | if error_name != ErrorIf.WrongRank: |
| 448 | assert rank == 3 |
| 449 | assert pl == 2 and const == 0 |
| 450 | |
| 451 | a_shape = testGen.makeShape(rank) |
| 452 | |
| 453 | # Constrict the overall size of the shape when creating ERROR_IF tests |
| 454 | if error_name: |
| 455 | a_shape = TosaErrorIfArgGen.eiRestrictDimensions(a_shape) |
| 456 | |
| 457 | # Get a random number for b_oc even if target shape is defined |
| 458 | b_oc = np.int32( |
| 459 | testGen.rng.integers( |
| 460 | low=testGen.args.tensor_shape_range[0], |
| 461 | high=testGen.args.tensor_shape_range[1], |
| 462 | size=1, |
| 463 | ) |
| 464 | )[0] |
| 465 | # If N or H is large let b_oc be 1 to reduce output tensor size |
| 466 | if max(a_shape) > 1000: |
| 467 | b_oc = 1 |
| 468 | |
| 469 | b_shape = np.asarray([a_shape[0], a_shape[2], b_oc]) |
| 470 | return [a_shape, b_shape] |
| 471 | |
| 472 | @staticmethod |
| 473 | def tgConcat(testGen, opName, rank, error_name=None): |
| 474 | pl, const = opName["operands"] |
| 475 | shape = testGen.makeShape(rank) |
| 476 | |
| 477 | # Create extra tensors to concat. |
| 478 | # Take into account value of pl when getting maximum number of concats |
| 479 | num_tensors = testGen.randInt(0, 4) |
| 480 | shape_list = [] |
| 481 | for i in range(pl + const + num_tensors): |
| 482 | if error_name == ErrorIf.ConcatInputRankMismatch and i != 0: |
| 483 | remove = testGen.rng.choice([True, False]) |
| 484 | wrongShape = shape.copy() |
| 485 | |
| 486 | if remove and len(shape) > 1: |
| 487 | wrongShape = wrongShape[1:] |
| 488 | else: |
| 489 | wrongShape = list(wrongShape) |
| 490 | wrongShape.append(testGen.rng.integers(1, 10)) |
| 491 | |
| 492 | shape_list.append(wrongShape) |
| 493 | else: |
| 494 | shape_list.append(shape.copy()) |
| 495 | |
| 496 | return shape_list |
| 497 | |
| 498 | @staticmethod |
| 499 | def tgConcatConstInput(testGen, shapeList, axis, error_name=None): |
| 500 | if error_name in [ |
| 501 | ErrorIf.AxisSmallerZero, |
| 502 | ErrorIf.AxisLargerRank, |
| 503 | ErrorIf.ConcatInputRankMismatch, |
| 504 | ]: |
| 505 | return shapeList |
| 506 | |
| 507 | # Split concat shape along axis to allow for multiple const inputs |
| 508 | # without making too many large tensors |
| 509 | if len(shapeList) == 2 or shapeList[0][axis] < len(shapeList): |
| 510 | # If axis can't be split we still need to invalidate other dimensions |
| 511 | if error_name == ErrorIf.ConcatInputDimMismatch: |
| 512 | for shape in shapeList[1:]: |
| 513 | # Negative test shapeLists are created individually for each test, |
| 514 | # so no need to copy the shape before altering it. |
| 515 | shape[(axis + 1) % len(shape)] += testGen.rng.integers(5, 10) |
| 516 | return shapeList |
| 517 | |
| 518 | # Create copy of shape we are going to split (so we don't alter shapeList) |
| 519 | shape = shapeList[0].copy() |
| 520 | # Add original shape as first input |
| 521 | new_shapeList = [shape.copy()] |
| 522 | length_on_axis = shape[axis] |
| 523 | remaining_length = length_on_axis |
| 524 | for i in range(len(shapeList) - 2): |
| 525 | # Calculate split on axis and remaining value |
| 526 | split_shape_val = int(shape[axis] / 2) |
| 527 | remaining_length = remaining_length - split_shape_val |
| 528 | |
| 529 | # Append new shape, and set remaining shape |
| 530 | shape[axis] = split_shape_val |
| 531 | new_shapeList.append(shape.copy()) |
| 532 | |
| 533 | # invalidate dimensions |
| 534 | if error_name == ErrorIf.ConcatInputDimMismatch: |
| 535 | shape[(axis + 1) % len(shape)] += testGen.rng.integers(5, 10) |
| 536 | else: |
| 537 | shape[axis] = remaining_length |
| 538 | |
| 539 | if i == len(shapeList) - 3: |
| 540 | new_shapeList.append(shape.copy()) |
| 541 | |
| 542 | return new_shapeList |
| 543 | |
| 544 | |
| 545 | class TosaTensorValuesGen: |
| 546 | """Tensor Value generators create the random data for each test.""" |
| 547 | |
| 548 | def __init__(self): |
| 549 | pass |
| 550 | |
| 551 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 552 | def tvgDefault(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 553 | pCount, cCount = op["operands"] |
| 554 | |
| 555 | tens = [] |
| 556 | tens.extend( |
| 557 | testGen.buildPlaceholderTensors(shapeList[0:pCount], dtypeList[0:pCount]) |
| 558 | ) |
| 559 | tens.extend(testGen.buildConstTensors(shapeList[pCount:], dtypeList[pCount:])) |
| 560 | |
| 561 | return tens |
| 562 | |
| 563 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 564 | def tvgNegate(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 0e46364 | 2022-05-03 12:10:23 +0100 | [diff] [blame] | 565 | if dtypeList[0] == DType.INT32 and error_name is None: |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 566 | pCount, cCount = op["operands"] |
| 567 | assert ( |
| 568 | pCount == 1 and cCount == 0 |
| 569 | ), "Op.NEGATE must have 1 placeholders, 0 consts" |
Jeremy Johnson | 0e46364 | 2022-05-03 12:10:23 +0100 | [diff] [blame] | 570 | # Must create tensors with values within accumulator (int32) negatable |
| 571 | # range |
| 572 | max_val = (1 << 31) - 1 |
| 573 | min_val = -max_val |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 574 | arr = np.int32( |
| 575 | testGen.rng.integers(low=min_val, high=(max_val + 1), size=shapeList[0]) |
| 576 | ) |
| 577 | placeholders = [] |
| 578 | placeholders.append( |
| 579 | testGen.ser.addPlaceholder(shapeList[0], dtypeList[0], arr) |
| 580 | ) |
| 581 | return placeholders |
| 582 | else: |
| 583 | return TosaTensorValuesGen.tvgDefault( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 584 | testGen, op, dtypeList, shapeList, testArgs, error_name |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 585 | ) |
| 586 | |
| 587 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 588 | def tvgAddSub(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 589 | if dtypeList[0] == DType.INT32 and error_name is None: |
| 590 | # Make sure the operation does not cause value saturation - where |
| 591 | # the number wraps due to limited number of bits to store the answer |
| 592 | pCount, cCount = op["operands"] |
| 593 | assert ( |
| 594 | pCount == 2 and cCount == 0 |
| 595 | ), "Op.ADD / Op.SUB must have 2 placeholders, 0 consts" |
| 596 | placeholders = [] |
| 597 | add = op["op"] == Op.ADD |
| 598 | a_arr = testGen.getRandTensor(shapeList[0], dtypeList[0]) |
| 599 | b_arr = testGen.getRandTensor(shapeList[1], dtypeList[1]) |
| 600 | if add: |
| 601 | res_arr = np.add(a_arr, b_arr, dtype=np.int64) |
| 602 | else: |
| 603 | res_arr = np.subtract(a_arr, b_arr, dtype=np.int64) |
| 604 | |
| 605 | # Work out the saturation limits |
| 606 | max_i32 = (1 << 31) - 1 |
| 607 | min_i32 = -(1 << 31) |
| 608 | max_arr = np.full(shapeList[1], max_i32) |
| 609 | min_arr = np.full(shapeList[1], min_i32) |
| 610 | |
| 611 | # Find how much values exceed the maximum/minimums |
| 612 | sat_max_arr = np.maximum(res_arr - max_arr, 0) |
| 613 | sat_min_arr = np.minimum(res_arr - min_arr, 0) |
| 614 | |
| 615 | if not add: |
| 616 | # Swap saturation values and negate values as we need to perform opposite operations |
| 617 | sat_max_arr, sat_min_arr = -sat_min_arr, -sat_max_arr |
| 618 | |
| 619 | # Create new array of unsaturated values by clipping values as needed |
| 620 | b_unsat_arr = b_arr |
| 621 | if (sat_max_arr != 0).any(): |
| 622 | # Clip values that cause saturation |
| 623 | b_unsat_arr = np.subtract(b_unsat_arr, sat_max_arr, dtype=np.int32) |
| 624 | # Reduce axes in unsaturated tensor to match original tensor |
| 625 | for axis, dim in enumerate(b_arr.shape): |
| 626 | if dim != b_unsat_arr.shape[axis]: |
| 627 | assert ( |
| 628 | dim == 1 |
| 629 | ), "Op.ADD / SUB dimension must be 1 or matching to be broadcastable" |
| 630 | b_unsat_arr = np.amin(b_unsat_arr, axis=axis, keepdims=True) |
| 631 | |
| 632 | if (sat_min_arr != 0).any(): |
| 633 | # Clip values that cause saturation |
| 634 | b_unsat_arr = np.subtract(b_unsat_arr, sat_min_arr, dtype=np.int32) |
| 635 | # Reduce axes in unsaturated tensor to match original tensor |
| 636 | for axis, dim in enumerate(b_arr.shape): |
| 637 | if dim != b_unsat_arr.shape[axis]: |
| 638 | assert ( |
| 639 | dim == 1 |
| 640 | ), "Op.ADD / SUB dimension must be 1 or matching to be broadcastable" |
| 641 | b_unsat_arr = np.amax(b_unsat_arr, axis=axis, keepdims=True) |
| 642 | |
| 643 | placeholders.append( |
| 644 | testGen.ser.addPlaceholder(shapeList[0], dtypeList[0], a_arr) |
| 645 | ) |
| 646 | placeholders.append( |
| 647 | testGen.ser.addPlaceholder(shapeList[1], dtypeList[1], b_unsat_arr) |
| 648 | ) |
| 649 | |
| 650 | return placeholders |
| 651 | else: |
| 652 | return TosaTensorValuesGen.tvgDefault( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 653 | testGen, op, dtypeList, shapeList, testArgs, error_name |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 654 | ) |
| 655 | |
| 656 | @staticmethod |
| 657 | def tvgCondIfWhileLoop( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 658 | testGen, op, dtypeList, shapeList, testArgs, error_name=None |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 659 | ): |
| 660 | if dtypeList[0] in ( |
| 661 | DType.INT32, |
| 662 | DType.INT16, |
| 663 | DType.INT8, |
| 664 | ): |
| 665 | # Limit input tensors with cond_if_binary or while_loop to stop |
| 666 | # saturation of add/sub ops with int32 and keep all logical shift |
| 667 | # values between 0 to 31 for int16 or int8 |
| 668 | pCount, cCount = op["operands"] |
| 669 | pRemain = pCount |
| 670 | placeholders = [] |
| 671 | for idx, shape in enumerate(shapeList[:]): |
| 672 | if dtypeList[0] == DType.INT32: |
| 673 | arr = testGen.getRandTensor(shapeList[idx], DType.INT16) |
| 674 | else: |
| 675 | arr = np.int32( |
| 676 | testGen.rng.integers(low=0, high=32, size=shapeList[idx]) |
| 677 | ) |
| 678 | if pRemain > 0: |
| 679 | placeholders.append( |
| 680 | testGen.ser.addPlaceholder(shape, dtypeList[idx], arr) |
| 681 | ) |
| 682 | pRemain -= 1 |
| 683 | else: |
| 684 | placeholders.append( |
| 685 | testGen.ser.addConst(shape, dtypeList[idx], arr) |
| 686 | ) |
| 687 | |
| 688 | return placeholders |
| 689 | else: |
| 690 | return TosaTensorValuesGen.tvgDefault( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 691 | testGen, op, dtypeList, shapeList, testArgs, error_name |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 692 | ) |
| 693 | |
| 694 | @staticmethod |
| 695 | def tvgArithmeticRightShift( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 696 | testGen, op, dtypeList, shapeList, testArgs, error_name=None |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 697 | ): |
| 698 | pCount, cCount = op["operands"] |
| 699 | # Force value of operand[1] to be within [0, num_bits] |
| 700 | assert ( |
| 701 | pCount == 2 and cCount == 0 |
| 702 | ), "Op.ArithmeticRightShift must have 2 placeholders, 0 consts" |
| 703 | |
| 704 | placeholders = [] |
| 705 | for idx, shape in enumerate(shapeList[:]): |
| 706 | if idx == 1: |
| 707 | if dtypeList[idx] == DType.INT8: |
| 708 | arr = np.int32(testGen.rng.integers(low=0, high=8, size=shape)) |
| 709 | elif dtypeList[idx] == DType.INT16: |
| 710 | arr = np.int32(testGen.rng.integers(low=0, high=16, size=shape)) |
| 711 | elif dtypeList[idx] == DType.INT32: |
| 712 | arr = np.int32(testGen.rng.integers(low=0, high=32, size=shape)) |
| 713 | elif error_name == ErrorIf.WrongInputType: |
| 714 | arr = np.int32(testGen.rng.integers(low=0, high=8, size=shape)) |
| 715 | else: |
| 716 | raise Exception("OpArithmeticRightShift: invalid input dtype") |
| 717 | else: |
| 718 | arr = testGen.getRandTensor(shape, dtypeList[idx]) |
| 719 | placeholders.append(testGen.ser.addPlaceholder(shape, dtypeList[idx], arr)) |
| 720 | |
| 721 | return placeholders |
| 722 | |
| 723 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 724 | def tvgSelect(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 725 | # Set datatype of condition tensor to boolean |
| 726 | dtypeList[0] = DType.BOOL |
| 727 | |
| 728 | return TosaTensorValuesGen.tvgDefault( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 729 | testGen, op, dtypeList, shapeList, testArgs, error_name |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 730 | ) |
| 731 | |
| 732 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 733 | def tvgIntDiv(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 734 | if error_name is None: |
| 735 | pCount, cCount = op["operands"] |
| 736 | assert ( |
| 737 | pCount == 2 and cCount == 0 |
| 738 | ), "Op.INTDIV must have 2 placeholders, 0 consts" |
| 739 | |
| 740 | placeholders = [] |
| 741 | |
| 742 | # Two invalid cases for Op.INTDIV: |
| 743 | # 1. divisor == 0 |
| 744 | # 2. dividend == -(1<<31) and divisor == -1 |
| 745 | while True: |
| 746 | dividend_arr = testGen.getRandTensor(shapeList[0], dtypeList[0]) |
| 747 | divisor_arr = testGen.getRandTensor(shapeList[1], dtypeList[1]) |
| 748 | |
| 749 | if (divisor_arr == 0).any(): |
| 750 | continue |
| 751 | |
| 752 | if (dividend_arr == -(2**31)).any() and (divisor_arr == -1).any(): |
| 753 | continue |
| 754 | |
| 755 | break |
| 756 | |
| 757 | placeholders.append( |
| 758 | testGen.ser.addPlaceholder(shapeList[0], dtypeList[0], dividend_arr) |
| 759 | ) |
| 760 | placeholders.append( |
| 761 | testGen.ser.addPlaceholder(shapeList[1], dtypeList[1], divisor_arr) |
| 762 | ) |
| 763 | |
| 764 | return placeholders |
| 765 | else: |
| 766 | return TosaTensorValuesGen.tvgDefault( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 767 | testGen, op, dtypeList, shapeList, testArgs, error_name |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 768 | ) |
| 769 | |
| 770 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 771 | def tvgMul(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 772 | if error_name is None: |
| 773 | pCount, cCount = op["operands"] |
| 774 | assert ( |
| 775 | pCount == 2 and cCount == 0 |
| 776 | ), "Op.MUL must have 2 placeholders, 0 consts" |
| 777 | |
| 778 | tens = [] |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 779 | if dtypeList[0] in (DType.FP16, DType.FP32): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 780 | tens.extend(testGen.buildPlaceholderTensors(shapeList[:], dtypeList[:])) |
| 781 | else: |
| 782 | placeholders = [] |
| 783 | |
| 784 | # Make sure multiply result in int32 range |
| 785 | shift = testArgs[0] |
| 786 | if dtypeList[0] == DType.INT8: |
| 787 | num_bits = 8 |
| 788 | elif dtypeList[0] == DType.INT16: |
| 789 | num_bits = 16 |
| 790 | elif dtypeList[0] == DType.INT32: |
| 791 | num_bits = 32 |
| 792 | elif error_name == ErrorIf.WrongInputType: |
| 793 | num_bits = 8 |
| 794 | else: |
| 795 | raise Exception("OpMul: invalid input dtype") |
| 796 | |
| 797 | for idx, shape in enumerate(shapeList[:]): |
| 798 | low = -(2 ** (num_bits - 1)) |
| 799 | high = (2 ** (num_bits - 1)) - 1 |
| 800 | |
| 801 | a_arr = np.int32( |
| 802 | testGen.rng.integers(low=low, high=high, size=shapeList[0]) |
| 803 | ) |
| 804 | b_arr = np.int32( |
| 805 | testGen.rng.integers(low=low, high=high, size=shapeList[1]) |
| 806 | ) |
| 807 | |
| 808 | i = 0 |
| 809 | while True: |
| 810 | |
| 811 | a_arr_64 = a_arr.astype(np.int64) |
| 812 | b_arr_64 = b_arr.astype(np.int64) |
| 813 | |
| 814 | if shift > 0: |
| 815 | rounding = 1 << (shift - 1) |
| 816 | result_arr = ((a_arr_64 * b_arr_64) + rounding) >> shift |
| 817 | else: |
| 818 | result_arr = a_arr_64 * b_arr_64 |
| 819 | |
| 820 | if (result_arr > -(2**31)).all() and ( |
| 821 | result_arr <= ((2**31) - 1) |
| 822 | ).all(): |
| 823 | break |
| 824 | |
| 825 | i = i + 1 |
| 826 | a_arr = a_arr // 2 |
| 827 | b_arr = b_arr // 2 |
| 828 | |
| 829 | placeholders.append( |
| 830 | testGen.ser.addPlaceholder(shapeList[0], dtypeList[0], a_arr) |
| 831 | ) |
| 832 | placeholders.append( |
| 833 | testGen.ser.addPlaceholder(shapeList[1], dtypeList[1], b_arr) |
| 834 | ) |
| 835 | |
| 836 | tens.extend(placeholders) |
| 837 | |
| 838 | return tens |
| 839 | else: |
| 840 | return TosaTensorValuesGen.tvgDefault( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 841 | testGen, op, dtypeList, shapeList, testArgs, error_name |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 842 | ) |
| 843 | |
| 844 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 845 | def tvgConcat(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 846 | count = len(shapeList) - testGen.args.num_const_inputs_concat |
| 847 | if count < 1: |
| 848 | count = 1 |
| 849 | if testGen.args.num_const_inputs_concat == 0: |
| 850 | count = len(shapeList) |
| 851 | |
| 852 | # Ensure axis is an int |
| 853 | testArgs[0] = int(testArgs[0]) |
| 854 | |
| 855 | shapeList = TosaTensorGen.tgConcatConstInput( |
| 856 | testGen, shapeList, testArgs[0], error_name |
| 857 | ) |
| 858 | |
| 859 | tens = [] |
| 860 | tens.extend( |
| 861 | testGen.buildPlaceholderTensors(shapeList[0:count], dtypeList[0:count]) |
| 862 | ) |
| 863 | tens.extend(testGen.buildConstTensors(shapeList[count:], dtypeList[count:])) |
| 864 | |
| 865 | return tens |
| 866 | |
| 867 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 868 | def tvgLogicalShift(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 869 | pCount, cCount = op["operands"] |
| 870 | assert ( |
| 871 | pCount == 2 and cCount == 0 |
| 872 | ), "Op.LOGICAL_LEFT_SHIFT or Op.LOGICAL_RIGHT_SHIFT must have 2 placeholders, 0 consts" |
| 873 | values_arr = testGen.getRandTensor(shapeList[0], dtypeList[0]) |
| 874 | shift_arr = np.int32(testGen.rng.integers(low=0, high=32, size=shapeList[1])) |
| 875 | placeholders = [] |
| 876 | placeholders.append( |
| 877 | testGen.ser.addPlaceholder(shapeList[0], dtypeList[0], values_arr) |
| 878 | ) |
| 879 | placeholders.append( |
| 880 | testGen.ser.addPlaceholder(shapeList[1], dtypeList[1], shift_arr) |
| 881 | ) |
| 882 | |
| 883 | return placeholders |
| 884 | |
| 885 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 886 | def tvgEqual(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 887 | if error_name is None: |
| 888 | pCount, cCount = op["operands"] |
| 889 | assert ( |
| 890 | pCount == 2 and cCount == 0 |
| 891 | ), "Op.EQUAL must have 2 placeholders, 0 consts" |
| 892 | a_arr = testGen.getRandTensor(shapeList[0], dtypeList[0]) |
| 893 | b_arr = testGen.getRandTensor(shapeList[1], dtypeList[1]) |
| 894 | # Using random numbers means that it will be very unlikely that |
| 895 | # there are any matching (equal) values, therefore force that |
| 896 | # there are twice the number of matching values as the tensor rank |
| 897 | for num in range(0, len(shapeList[0]) * 2): |
| 898 | a_index = [] |
| 899 | b_index = [] |
| 900 | # Choose an index in each axis for the whole shape |
| 901 | for axis in range(0, len(shapeList[0])): |
| 902 | # Index can be up to the largest dimension in both shapes |
| 903 | index = np.int32( |
| 904 | testGen.rng.integers( |
| 905 | 0, max(shapeList[0][axis], shapeList[1][axis]) |
| 906 | ) |
| 907 | ) |
| 908 | # Reduce the index down to a shape's dim for broadcasting |
| 909 | a_index.append(min(shapeList[0][axis] - 1, index)) |
| 910 | b_index.append(min(shapeList[1][axis] - 1, index)) |
| 911 | |
| 912 | a_arr[tuple(a_index)] = b_arr[tuple(b_index)] |
| 913 | |
| 914 | placeholders = [] |
| 915 | placeholders.append( |
| 916 | testGen.ser.addPlaceholder(shapeList[0], dtypeList[0], a_arr) |
| 917 | ) |
| 918 | placeholders.append( |
| 919 | testGen.ser.addPlaceholder(shapeList[1], dtypeList[1], b_arr) |
| 920 | ) |
| 921 | return placeholders |
| 922 | else: |
| 923 | return TosaTensorValuesGen.tvgDefault( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 924 | testGen, op, dtypeList, shapeList, testArgs, error_name |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 925 | ) |
| 926 | |
| 927 | @staticmethod |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 928 | def tvgReduceSum(testGen, op, dtypeList, shapeList, testArgs, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 929 | if dtypeList[0] == DType.INT32: |
| 930 | pCount, cCount = op["operands"] |
| 931 | assert ( |
| 932 | pCount == 1 and cCount == 0 |
| 933 | ), "Op.REDUCE_SUM must have 1 placeholders, 0 consts" |
| 934 | # Limit values so that the sum cannot exceed the range of an int32 during |
| 935 | # summation of any axis |
| 936 | range_val = int((1 << 31) / max(shapeList[0])) |
| 937 | values_arr = np.int32( |
| 938 | testGen.rng.integers(low=-range_val, high=range_val, size=shapeList[0]) |
| 939 | ) |
| 940 | placeholders = [] |
| 941 | placeholders.append( |
| 942 | testGen.ser.addPlaceholder(shapeList[0], dtypeList[0], values_arr) |
| 943 | ) |
| 944 | return placeholders |
| 945 | else: |
| 946 | return TosaTensorValuesGen.tvgDefault( |
Eric Kunze | b5fabec | 2022-06-07 05:20:44 +0000 | [diff] [blame] | 947 | testGen, op, dtypeList, shapeList, testArgs, error_name |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 948 | ) |
| 949 | |
| 950 | |
| 951 | class TosaArgGen: |
| 952 | """Argument generators create exhaustive or random lists of attributes for |
| 953 | operators that take attributes or other parameters. |
| 954 | |
| 955 | The return value is a list of (descriptive_name, [arglist]) tuples where |
| 956 | the descriptive_name is appended to the test name and the arglist is expanded |
| 957 | as arguments to the operator build function. |
| 958 | """ |
| 959 | |
| 960 | def __init__(self): |
| 961 | pass |
| 962 | |
| 963 | @staticmethod |
| 964 | def agNone(testGen, opName, shapeList, dtype, error_name=None): |
| 965 | """A trivial argument generator for operators that don't take any |
| 966 | non-tensor arguments""" |
| 967 | return [("", [])] |
| 968 | |
| 969 | @staticmethod |
| 970 | def agAxis(testGen, opName, shapeList, dtype, error_name=None): |
| 971 | """Build the axis argument for operators that take a single axis""" |
| 972 | axes = [] |
| 973 | shape = shapeList[0] |
| 974 | |
| 975 | if error_name == ErrorIf.AxisSmallerZero: |
| 976 | small_axis = testGen.rng.integers(-5, 0) |
| 977 | axes.append(("axis{}".format(small_axis), [small_axis])) |
| 978 | elif error_name == ErrorIf.AxisLargerRank: |
| 979 | large_axis = testGen.rng.integers(len(shape) + 1, len(shape) + 10) |
| 980 | axes.append(("axis{}".format(large_axis), [large_axis])) |
| 981 | else: |
| 982 | for a in range(0, len(shape)): |
| 983 | axes.append(("axis{}".format(a), [a])) |
| 984 | |
| 985 | return axes |
| 986 | |
| 987 | @staticmethod |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 988 | def agConv(testGen, opName, shapeList, dtypes, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 989 | arg_list = [] |
| 990 | |
| 991 | ifm_shape = shapeList[0] |
| 992 | filter_shape = shapeList[1] |
| 993 | # determine the kernel shape from operator name (e.g. "conv2d_3x3" => [3,3]) |
| 994 | k = [int(x) for x in opName.split("_")[-1].split("x")] |
| 995 | |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 996 | accum_dtype = get_accum_dtype_from_tgTypes(dtypes) |
| 997 | |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 998 | # Check the rank |
| 999 | rank = 5 if opName.startswith("conv3d") else 4 |
| 1000 | if error_name != ErrorIf.WrongRank: |
| 1001 | assert len(ifm_shape) == rank |
| 1002 | assert len(filter_shape) == rank |
| 1003 | |
| 1004 | # kernel rank omits batch and channels |
| 1005 | k_rank = rank - 2 |
| 1006 | assert len(k) == k_rank |
| 1007 | |
| 1008 | # Generate comprehensive argument lists |
| 1009 | # - except for named errors, which use specific invalid value(s) |
| 1010 | if error_name == ErrorIf.PadSmallerZero: |
| 1011 | p_vals = [testGen.rng.choice(range(-5, 0))] |
| 1012 | else: |
| 1013 | p_vals = [x for x in range(0, testGen.args.max_conv_padding + 1)] |
| 1014 | paddings = {x for x in itertools.product(*([p_vals] * k_rank * 2))} |
| 1015 | if error_name == ErrorIf.StrideSmallerOne: |
| 1016 | # Can't use stride=0, as it is used to derive output shape, as a divisor |
| 1017 | s_vals = [testGen.rng.choice(range(-5, 0))] |
| 1018 | else: |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1019 | # Stride must be greater than 1 to force non-integer error |
Jeremy Johnson | 93d4390 | 2022-09-27 12:26:14 +0100 | [diff] [blame] | 1020 | startStride = 1 if error_name != ErrorIf.ConvOutputShapeNonInteger else 2 |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1021 | s_vals = [x for x in range(startStride, testGen.args.max_conv_stride + 1)] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1022 | strides = {x for x in itertools.product(*([s_vals] * k_rank))} |
| 1023 | if error_name == ErrorIf.DilationSmallerOne: |
| 1024 | d_vals = [testGen.rng.choice(range(-5, 1))] |
| 1025 | else: |
| 1026 | d_vals = [x for x in range(1, testGen.args.max_conv_dilation + 1)] |
| 1027 | dilations = {x for x in itertools.product(*([d_vals] * k_rank))} |
| 1028 | |
| 1029 | if not error_name and testGen.args.oversize: |
| 1030 | # add some oversize argument values |
| 1031 | if max(ifm_shape) < 64: |
| 1032 | bigPadding = 9 |
| 1033 | paddings.update( |
| 1034 | {x for x in itertools.product(*([[0, bigPadding]] * (k_rank * 2)))} |
| 1035 | ) |
| 1036 | bigStride = 8 |
| 1037 | strides.update({x for x in itertools.product(*([[1, bigStride]] * k_rank))}) |
| 1038 | bigDilation = 7 |
| 1039 | dilations.update( |
| 1040 | {x for x in itertools.product(*([[1, bigDilation]] * k_rank))} |
| 1041 | ) |
| 1042 | |
| 1043 | # There are too many parameter combinations, so generate them sparsely, |
| 1044 | # very sparse for negative tests |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1045 | sparsity_factor = 2 if error_name else 120 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1046 | sparsity = len(paddings) * len(strides) * len(dilations) // sparsity_factor + 1 |
| 1047 | # If there are only a small number of tests, just select them all |
| 1048 | if sparsity < 13: |
| 1049 | sparsity = 1 |
| 1050 | # To get a variety of parameter combinations sparsity should not be a |
| 1051 | # multiple of 2, 3 or 5 |
| 1052 | while sparsity % 2 == 0 or sparsity % 3 == 0 or sparsity % 5 == 0: |
| 1053 | sparsity += 1 |
| 1054 | |
| 1055 | n = 0 |
| 1056 | for s in sorted(list(strides)): |
| 1057 | for p in sorted(list(paddings)): |
| 1058 | for d in sorted(list(dilations)): |
| 1059 | if ( |
| 1060 | n % sparsity == 0 |
Jeremy Johnson | 93d4390 | 2022-09-27 12:26:14 +0100 | [diff] [blame] | 1061 | # the padded shape must exceed the dilation * kernel to get a positive |
| 1062 | # sized output shape |
| 1063 | and (ifm_shape[1] - 1 + p[0] + p[1]) > d[0] * (k[0] - 1) |
| 1064 | and (ifm_shape[2] - 1 + p[2] + p[3]) > d[1] * (k[1] - 1) |
| 1065 | and ( |
| 1066 | k_rank < 3 |
| 1067 | or ((ifm_shape[3] - 1 + p[4] + p[5]) > d[2] * (k[2] - 1)) |
| 1068 | ) |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1069 | ): |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1070 | remainders = [] |
| 1071 | for index in range(k_rank): |
| 1072 | pad_offset = index * 2 |
| 1073 | remainders.append( |
| 1074 | ( |
| 1075 | ifm_shape[index + 1] |
| 1076 | - 1 |
| 1077 | + p[pad_offset] |
| 1078 | + p[pad_offset + 1] |
| 1079 | - (k[index] - 1) * d[index] |
| 1080 | ) |
| 1081 | % s[index] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1082 | ) |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1083 | if ( |
| 1084 | # the parameters must produce integer exact output |
| 1085 | error_name != ErrorIf.ConvOutputShapeNonInteger |
| 1086 | and max(remainders) == 0 |
| 1087 | ) or ( |
| 1088 | error_name == ErrorIf.ConvOutputShapeNonInteger |
| 1089 | and max(remainders) > 0 |
| 1090 | ): |
| 1091 | arg_list.append( |
| 1092 | ( |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1093 | "acc{}_st{}_pad{}_dilat{}".format( |
| 1094 | testGen.typeStr(accum_dtype), |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1095 | "".join([str(x) for x in s]), |
| 1096 | "".join([str(x) for x in p]), |
| 1097 | "".join([str(x) for x in d]), |
| 1098 | ), |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1099 | [accum_dtype, s, p, d], |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1100 | ) |
| 1101 | ) |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1102 | n += 1 |
| 1103 | |
| 1104 | return arg_list |
| 1105 | |
| 1106 | @staticmethod |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1107 | def agFullyConnected(testGen, opName, shapeList, dtypes, error_name=None): |
| 1108 | |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1109 | assert isinstance(dtypes, list) or isinstance( |
| 1110 | dtypes, tuple |
| 1111 | ), f"{dtypes} unexpected" |
| 1112 | input_dtype = dtypes[0] |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1113 | |
| 1114 | if error_name == ErrorIf.WrongOutputType: |
| 1115 | accum_dtype = get_wrong_output_type(opName, testGen.rng, input_dtype) |
| 1116 | elif error_name == ErrorIf.WrongInputType: |
| 1117 | # Pick some potentially correct output dtype if input type is incorrect |
| 1118 | accum_dtype = DType.INT32 |
| 1119 | else: |
| 1120 | accum_dtype = get_accum_dtype_from_tgTypes(dtypes) |
| 1121 | |
| 1122 | return [(f"acc{testGen.typeStr(accum_dtype)}", [accum_dtype])] |
| 1123 | |
| 1124 | @staticmethod |
| 1125 | def agMatMul(testGen, opName, shapeList, dtype, error_name=None): |
| 1126 | # Get valid accumulate type(s) |
| 1127 | if dtype == DType.INT8: |
| 1128 | accum_dtypes = [DType.INT32] |
| 1129 | elif dtype == DType.INT16: |
| 1130 | accum_dtypes = [DType.INT48] |
| 1131 | elif dtype == DType.FP16: |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1132 | accum_dtypes = [DType.FP16, DType.FP32] |
| 1133 | elif dtype == DType.FP32: |
| 1134 | accum_dtypes = [DType.FP32] |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1135 | elif error_name is None: |
| 1136 | assert False, f"Invalid I/O DType for MatMul: {DTypeNames[dtype]}" |
| 1137 | |
| 1138 | if error_name == ErrorIf.WrongOutputType: |
| 1139 | # Get incorrect output dtype for ErrorIf case |
| 1140 | accum_dtypes = [get_wrong_output_type(opName, testGen.rng, dtype)] |
| 1141 | elif error_name == ErrorIf.WrongInputType: |
| 1142 | # Pick some potentially correct output dtype if input type is incorrect |
| 1143 | accum_dtypes = [DType.INT32] |
| 1144 | |
| 1145 | return [(f"acc{testGen.typeStr(a)}", [a]) for a in accum_dtypes] |
| 1146 | |
| 1147 | @staticmethod |
| 1148 | def agTransposeConv2D(testGen, opName, shapeList, dtypes, error_name=None): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1149 | arg_list = [] |
| 1150 | |
| 1151 | ifm_shape = shapeList[0] |
| 1152 | filter_shape = shapeList[1] |
| 1153 | |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1154 | accum_dtype = get_accum_dtype_from_tgTypes(dtypes) |
| 1155 | |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1156 | # Must be rank 4 |
| 1157 | if error_name != ErrorIf.WrongRank: |
| 1158 | assert len(ifm_shape) == 4 |
| 1159 | assert len(filter_shape) == 4 |
| 1160 | |
| 1161 | # Generate comprehensive argument lists |
| 1162 | # - except for named errors, which use specific invalid value(s) |
Eric Kunze | c1a9783 | 2022-07-01 16:56:09 -0700 | [diff] [blame] | 1163 | smallest_padding_size = -min(filter_shape[1], filter_shape[2]) + 1 |
| 1164 | if error_name == ErrorIf.PadLargerEqualKernel: |
| 1165 | max_filter_size = -max(filter_shape[1], filter_shape[2]) |
| 1166 | p_vals = [testGen.rng.choice(range(max_filter_size - 10, max_filter_size))] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1167 | else: |
Eric Kunze | c1a9783 | 2022-07-01 16:56:09 -0700 | [diff] [blame] | 1168 | p_vals = [ |
| 1169 | x |
| 1170 | for x in range(smallest_padding_size, testGen.args.max_conv_padding + 1) |
| 1171 | ] |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1172 | paddings = {x for x in itertools.product(*([p_vals] * 4))} |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1173 | if error_name == ErrorIf.StrideSmallerOne: |
| 1174 | # Can't use stride=0, as it is used to derive output shape, as a divisor |
| 1175 | s_vals = [testGen.rng.choice(range(-5, 0))] |
| 1176 | else: |
| 1177 | s_vals = [x for x in range(1, testGen.args.max_conv_stride + 1)] |
| 1178 | strides = {x for x in itertools.product(*([s_vals] * 2))} |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1179 | |
Jeremy Johnson | 5860df6 | 2022-05-04 15:30:58 +0100 | [diff] [blame] | 1180 | if not error_name and testGen.args.oversize: |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1181 | # add some oversize argument values |
| 1182 | if max(ifm_shape) < 64: |
| 1183 | bigPadding = 9 |
| 1184 | paddings.update( |
Eric Kunze | c1a9783 | 2022-07-01 16:56:09 -0700 | [diff] [blame] | 1185 | { |
| 1186 | x |
| 1187 | for x in itertools.product( |
| 1188 | *([[smallest_padding_size, bigPadding]] * 4) |
| 1189 | ) |
| 1190 | } |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1191 | ) |
| 1192 | bigStride = 8 |
| 1193 | strides.update({x for x in itertools.product(*([[1, bigStride]] * 2))}) |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1194 | |
| 1195 | # There are too many parameter combinations, so generate them sparsely, |
| 1196 | # very sparse for negative tests |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1197 | sparsity_factor = 2 if error_name else 10 |
TatWai Chong | 24594f5 | 2022-06-08 00:48:04 -0700 | [diff] [blame] | 1198 | sparsity = len(paddings) * len(strides) // sparsity_factor + 1 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1199 | # If there are only a small number of tests, just select them all |
| 1200 | if sparsity < 13: |
| 1201 | sparsity = 1 |
| 1202 | # To get a variety of parameter combinations sparsity should not be a |
| 1203 | # multiple of 2, 3 or 5 |
| 1204 | while sparsity % 2 == 0 or sparsity % 3 == 0 or sparsity % 5 == 0: |
| 1205 | sparsity += 1 |
| 1206 | |
| 1207 | n = 0 |
| 1208 | for s in sorted(list(strides)): |
| 1209 | for p in sorted(list(paddings)): |
TatWai Chong | 24594f5 | 2022-06-08 00:48:04 -0700 | [diff] [blame] | 1210 | if n % sparsity == 0: |
| 1211 | # Determine the output shape |
Eric Kunze | c1a9783 | 2022-07-01 16:56:09 -0700 | [diff] [blame] | 1212 | oh = (ifm_shape[1] - 1) * s[0] + p[0] + p[1] + filter_shape[1] |
| 1213 | ow = (ifm_shape[2] - 1) * s[1] + p[2] + p[3] + filter_shape[2] |
TatWai Chong | 24594f5 | 2022-06-08 00:48:04 -0700 | [diff] [blame] | 1214 | os = [ifm_shape[0], oh, ow, filter_shape[0]] |
| 1215 | arg_list.append( |
| 1216 | ( |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1217 | "acc{}_st{}_pad{}_os{}".format( |
| 1218 | testGen.typeStr(accum_dtype), |
TatWai Chong | 24594f5 | 2022-06-08 00:48:04 -0700 | [diff] [blame] | 1219 | "".join([str(x) for x in s]), |
| 1220 | "".join([str(x) for x in p]), |
| 1221 | "x".join([str(x) for x in os]), |
| 1222 | ), |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1223 | [accum_dtype, s, p, os], |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1224 | ) |
TatWai Chong | 24594f5 | 2022-06-08 00:48:04 -0700 | [diff] [blame] | 1225 | ) |
| 1226 | n += 1 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1227 | |
| 1228 | return arg_list |
| 1229 | |
| 1230 | @staticmethod |
| 1231 | def agPad(testGen, opName, shapeList, dtype, error_name=None): |
| 1232 | arg_list = [] |
| 1233 | rank = len(shapeList[0]) |
| 1234 | |
| 1235 | # Exhaustively test combinations of padding on each side of each dimension |
| 1236 | # - the range of padding values is defined by pad_min and pad_max |
| 1237 | # - for padding >9, the name format needs to be more distinctive |
| 1238 | pad_min, pad_max = 0, 1 |
| 1239 | pad_values = [x for x in range(pad_min, pad_max + 1)] |
| 1240 | if error_name == ErrorIf.PadSmallerZero: |
| 1241 | pad_values = [x for x in range(-2, 0)] |
| 1242 | axis_pad_values = [x for x in itertools.product(pad_values, pad_values)] |
| 1243 | shape_pad_values = itertools.product(*([axis_pad_values] * rank)) |
| 1244 | |
| 1245 | if dtype in [DType.BOOL, DType.INT8, DType.INT16, DType.INT32]: |
| 1246 | pad_const_int = testGen.getRandNumberDType(dtype) |
| 1247 | pad_const_fp = 0 |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1248 | elif dtype in (DType.FP16, DType.FP32): |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1249 | pad_const_int = 0 |
| 1250 | pad_const_fp = testGen.getRandNumberDType(dtype) |
| 1251 | else: |
| 1252 | return [] |
| 1253 | |
| 1254 | for paddings in shape_pad_values: |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1255 | paddings = list(paddings) |
| 1256 | args_valid = True |
| 1257 | |
| 1258 | if error_name == ErrorIf.PadSmallerZero: |
| 1259 | # Prevent negative output shapes while ensuring still testing for negative padding |
| 1260 | for i in range(rank): |
| 1261 | dim_after_padding = ( |
| 1262 | paddings[i][0] + paddings[i][1] + shapeList[0][i] |
| 1263 | ) |
| 1264 | if dim_after_padding < 1: |
| 1265 | paddings[i] = (0, 0) |
| 1266 | if all([p > -1 for p in paddings[i]]): |
| 1267 | args_valid = False |
| 1268 | |
| 1269 | if args_valid: |
| 1270 | name = "pad" |
| 1271 | for r in range(rank): |
| 1272 | before, after = paddings[r] |
| 1273 | name = f"{name}{before}{after}" |
| 1274 | arg_list.append( |
| 1275 | (name, [np.array(paddings), pad_const_int, pad_const_fp]) |
| 1276 | ) |
| 1277 | |
| 1278 | if error_name == ErrorIf.PadSmallerZero and len(arg_list) == 0: |
| 1279 | warnings.warn(f"No ErrorIf test created for input shape: {shapeList[0]}") |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1280 | |
| 1281 | return arg_list |
| 1282 | |
| 1283 | @staticmethod |
| 1284 | def agPooling(testGen, opName, shapeList, dtype, error_name=None): |
| 1285 | arg_list = [] |
| 1286 | |
| 1287 | shape = shapeList[0] |
| 1288 | if error_name != ErrorIf.WrongRank: |
| 1289 | assert len(shape) == 4 |
| 1290 | |
| 1291 | # Generate comprehensive argument lists |
| 1292 | p_vals = [x for x in range(0, testGen.args.max_pooling_padding + 1)] |
| 1293 | paddings = {x for x in itertools.product(*([p_vals] * 4))} |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1294 | # Stride must be greater than 1 to force non-integer error |
| 1295 | startStride = 1 if error_name != ErrorIf.PoolingOutputShapeNonInteger else 2 |
| 1296 | s_vals = [x for x in range(startStride, testGen.args.max_pooling_stride + 1)] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1297 | strides = {x for x in itertools.product(*([s_vals] * 2))} |
| 1298 | k_vals = [x for x in range(2, testGen.args.max_pooling_kernel + 1)] |
| 1299 | kernels = {x for x in itertools.product(*([k_vals] * 2))} |
| 1300 | |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1301 | if opName == "max_pool2d": |
| 1302 | accum_dtypes = [None] # max_pool has no accumulate dtype |
| 1303 | elif dtype == DType.INT8 or dtype == DType.INT16: |
| 1304 | accum_dtypes = [DType.INT32] |
| 1305 | elif dtype == DType.FP16: |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1306 | accum_dtypes = [DType.FP16, DType.FP32] |
| 1307 | elif dtype == DType.FP32: |
| 1308 | accum_dtypes = [DType.FP32] |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1309 | elif error_name is None: |
| 1310 | assert False, f"Invalid I/O DType for pooling: {DTypeNames[dtype]}" |
| 1311 | else: |
| 1312 | # Set to something for the ErrorIf case which has |
| 1313 | # incorrect input data-type |
| 1314 | accum_dtypes = [DType.INT32] |
| 1315 | |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1316 | if testGen.args.oversize: |
| 1317 | # add some oversize argument values |
| 1318 | bigStride = 7 |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1319 | strides.update( |
| 1320 | {x for x in itertools.product(*([[startStride, bigStride]] * 2))} |
| 1321 | ) |
| 1322 | bigKernel = 9 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1323 | kernels.update({x for x in itertools.product(*([[2, bigKernel]] * 2))}) |
| 1324 | if max(shape) < 64: |
| 1325 | # padding must be less than the kernel size |
| 1326 | bigPadding = bigKernel - 1 |
| 1327 | paddings.update( |
| 1328 | {x for x in itertools.product(*([[0, bigPadding]] * 4))} |
| 1329 | ) |
| 1330 | |
| 1331 | # There are too many parameter combinations, so generate them sparsely, |
| 1332 | # very sparse for negative tests |
| 1333 | sparsity_factor = 2 if error_name else 500 |
| 1334 | sparsity = len(paddings) * len(strides) * len(kernels) // sparsity_factor + 1 |
| 1335 | |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1336 | arg_str = ( |
| 1337 | "acc{}_st{}_kern{}_pad{}" |
| 1338 | if accum_dtypes[0] is not None |
| 1339 | else "st{}_kern{}_pad{}" |
| 1340 | ) |
| 1341 | |
| 1342 | def get_arg_list_element(accum, stride, pad, kern): |
| 1343 | # Return tuple containing the formatted argument string and |
| 1344 | # the corresponding argument values |
| 1345 | arg_str_elems = [ |
| 1346 | "".join([str(x) for x in stride]), |
| 1347 | "".join([str(x) for x in kern]), |
| 1348 | "".join([str(x) for x in pad]), |
| 1349 | ] |
| 1350 | # Note: different order to string |
| 1351 | arg_val_elems = [stride, pad, kern] |
| 1352 | |
| 1353 | if accum is not None: |
| 1354 | arg_str_elems.insert(0, testGen.typeStr(accum)) |
| 1355 | arg_val_elems.insert(0, accum) |
| 1356 | return (arg_str.format(*arg_str_elems), arg_val_elems) |
| 1357 | |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1358 | n = 0 |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1359 | for a in accum_dtypes: |
| 1360 | for s in sorted(list(strides)): |
| 1361 | for p in sorted(list(paddings)): |
| 1362 | for k in sorted(list(kernels)): |
| 1363 | if error_name in [ |
| 1364 | ErrorIf.StrideSmallerOne, |
| 1365 | ErrorIf.KernelSmallerOne, |
| 1366 | ErrorIf.PadSmallerZero, |
| 1367 | ErrorIf.PadLargerEqualKernel, |
| 1368 | ]: |
| 1369 | sNew, pNew, kNew = TosaErrorIfArgGen.eiPoolingErrorIf( |
| 1370 | testGen, error_name, s, p, k |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1371 | ) |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1372 | if None not in [sNew, pNew, kNew] and n % sparsity == 0: |
| 1373 | arg_vals = [a, sNew, pNew, kNew] |
| 1374 | arg_list.append(get_arg_list_element(*arg_vals)) |
| 1375 | elif ( |
| 1376 | n % sparsity == 0 |
| 1377 | # padding must not exceed the kernel size |
| 1378 | and p[0] < k[0] |
| 1379 | and p[1] < k[0] |
| 1380 | and p[2] < k[1] |
| 1381 | and p[3] < k[1] |
| 1382 | # the padded shape must exceed the kernel size |
| 1383 | and (shape[1] + p[0] + p[1]) > k[0] |
| 1384 | and (shape[2] + p[2] + p[3]) > k[1] |
Jeremy Johnson | 4a6fb9b | 2022-04-26 15:47:21 +0100 | [diff] [blame] | 1385 | ): |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1386 | remainder_h = (shape[1] + p[0] + p[1] - k[0]) % s[0] |
| 1387 | remainder_w = (shape[2] + p[2] + p[3] - k[1]) % s[1] |
| 1388 | if ( |
| 1389 | # the parameters must produce integer exact output |
| 1390 | error_name != ErrorIf.PoolingOutputShapeNonInteger |
| 1391 | and remainder_h == 0 |
| 1392 | and remainder_w == 0 |
| 1393 | ) or ( |
| 1394 | error_name == ErrorIf.PoolingOutputShapeNonInteger |
| 1395 | and (remainder_h != 0 or remainder_w != 0) |
| 1396 | ): |
| 1397 | arg_vals = [a, s, p, k] |
| 1398 | arg_list.append(get_arg_list_element(*arg_vals)) |
| 1399 | n += 1 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1400 | |
| 1401 | return arg_list |
| 1402 | |
| 1403 | @staticmethod |
| 1404 | def agCast(testGen, opName, shapeList, inDtype, error_name=None): |
| 1405 | arg_list = [] |
| 1406 | |
| 1407 | # Enumerate the output types here |
| 1408 | if error_name == ErrorIf.WrongOutputType: |
| 1409 | dtypeList = TosaErrorIfArgGen.eiCastErrorIf(testGen, inDtype) |
| 1410 | elif inDtype == DType.INT8: |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1411 | dtypeList = [DType.BOOL, DType.INT16, DType.INT32, DType.FP32] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1412 | elif inDtype == DType.INT16: |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1413 | dtypeList = [DType.BOOL, DType.INT8, DType.INT32, DType.FP32] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1414 | elif inDtype == DType.INT32: |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1415 | dtypeList = [DType.BOOL, DType.INT8, DType.INT16, DType.FP32] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1416 | elif inDtype == DType.BOOL: |
| 1417 | dtypeList = [DType.INT8, DType.INT16, DType.INT32] |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1418 | elif inDtype == DType.FP16: |
| 1419 | dtypeList = [DType.INT8, DType.INT16, DType.INT32] |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1420 | elif inDtype == DType.FP32: |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1421 | dtypeList = [DType.INT8, DType.INT16, DType.INT32] |
| 1422 | elif error_name == ErrorIf.WrongInputType: |
| 1423 | # Pick some potentially correct output type for incorrect input type |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1424 | dtypeList = [DType.BOOL, DType.INT8, DType.INT16, DType.FP32] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1425 | else: |
| 1426 | raise Exception("Unexpected input dtype: {}".format(inDtype)) |
| 1427 | |
| 1428 | for dtype in dtypeList: |
| 1429 | arg_list.append(("out{}".format(DTypeNames[dtype]), [dtype])) |
| 1430 | |
| 1431 | return arg_list |
| 1432 | |
| 1433 | @staticmethod |
| 1434 | def agRescale(testGen, opName, shapeList, inDtype, error_name=None): |
| 1435 | arg_list = [] |
| 1436 | |
| 1437 | # Enumerate the output types here |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1438 | for outDtype in [ |
| 1439 | DType.UINT8, |
| 1440 | DType.INT8, |
| 1441 | DType.INT16, |
| 1442 | DType.INT32, |
| 1443 | DType.UINT16, |
| 1444 | ]: |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1445 | if ( |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1446 | outDtype in [DType.UINT8, DType.INT8, DType.UINT16] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1447 | and error_name == ErrorIf.OutputZeroPointNotZero |
| 1448 | ): |
| 1449 | continue |
| 1450 | if ( |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1451 | outDtype != DType.UINT16 |
| 1452 | and error_name == ErrorIf.U16OutputZeroPointNotValid |
| 1453 | ) or ( |
| 1454 | inDtype != DType.UINT16 |
| 1455 | and error_name == ErrorIf.U16InputZeroPointNotValid |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1456 | ): |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1457 | # ErrorIfs only valid with UINT16 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1458 | continue |
| 1459 | if ( |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1460 | inDtype == DType.UINT8 |
| 1461 | and outDtype not in [DType.INT8, DType.INT16] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1462 | and error_name != ErrorIf.WrongOutputType |
| 1463 | ): |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1464 | # The only output dtypes for UINT8 are INT8/INT16, skip all others |
| 1465 | continue |
| 1466 | if ( |
| 1467 | inDtype not in [DType.INT8, DType.INT16] |
| 1468 | and outDtype == DType.UINT8 |
| 1469 | and error_name != ErrorIf.WrongOutputType |
| 1470 | ): |
| 1471 | # The only input dtypes for UINT8 are INT8/INT16, skip all others |
| 1472 | continue |
| 1473 | if ( |
| 1474 | inDtype == DType.UINT16 |
| 1475 | and outDtype != DType.INT16 |
| 1476 | and error_name != ErrorIf.WrongOutputType |
| 1477 | ): |
| 1478 | # The only output dtype for UINT16 is INT16, skip all others |
| 1479 | continue |
| 1480 | if ( |
| 1481 | inDtype != DType.INT16 |
| 1482 | and outDtype == DType.UINT16 |
| 1483 | and error_name != ErrorIf.WrongOutputType |
| 1484 | ): |
| 1485 | # The only input dtype for UINT16 is INT16, skip all others |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1486 | continue |
| 1487 | if ( |
| 1488 | error_name == ErrorIf.WrongOutputType |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1489 | and not TosaErrorIfArgGen.eiRescaleWrongOutputType(inDtype, outDtype) |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1490 | ): |
| 1491 | continue |
| 1492 | |
| 1493 | for scale32 in [False, True]: |
| 1494 | if error_name == ErrorIf.ScaleTrue and not scale32: |
| 1495 | continue |
| 1496 | elif error_name == ErrorIf.ScaleNotTrue and scale32: |
| 1497 | continue |
| 1498 | for double_round in [False, True]: |
| 1499 | if error_name == ErrorIf.ScaleNotTrue and not double_round: |
| 1500 | continue |
| 1501 | for per_channel in [False, True]: |
| 1502 | |
| 1503 | if ( |
| 1504 | inDtype == DType.INT48 |
| 1505 | and scale32 |
| 1506 | and error_name != ErrorIf.ScaleTrue |
| 1507 | ): |
| 1508 | # Illegal condition. Must be scale32=False |
| 1509 | continue |
| 1510 | if ( |
| 1511 | double_round |
| 1512 | and not scale32 |
| 1513 | and error_name != ErrorIf.ScaleNotTrue |
| 1514 | ): |
| 1515 | # Illegal condition. ERROR_IF(!scale32 && double_round) |
| 1516 | continue |
| 1517 | |
| 1518 | arg_list.append( |
| 1519 | ( |
| 1520 | "out{}_sc{}_dr{}_pc{}".format( |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1521 | DTypeNames[outDtype], |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1522 | int(scale32), |
| 1523 | int(double_round), |
| 1524 | int(per_channel), |
| 1525 | ), |
Jeremy Johnson | f7f78ae | 2022-05-25 15:26:38 +0100 | [diff] [blame] | 1526 | [outDtype, scale32, double_round, per_channel], |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1527 | ) |
| 1528 | ) |
| 1529 | |
| 1530 | return arg_list |
| 1531 | |
| 1532 | @staticmethod |
| 1533 | def agMul(testGen, opName, shapeList, dtype, error_name=None): |
| 1534 | arg_list = [] |
| 1535 | |
| 1536 | if dtype is DType.INT32: |
| 1537 | for p in range(testGen.args.num_rand_permutations): |
| 1538 | |
| 1539 | shift = testGen.randInt(0, 32) |
| 1540 | |
| 1541 | arg_list.append(("perm{}_shift{}".format(p, shift), [shift])) |
| 1542 | else: |
| 1543 | arg_list.append(("perm0_shift0", [0])) |
| 1544 | |
| 1545 | return arg_list |
| 1546 | |
| 1547 | @staticmethod |
| 1548 | def agArithmeticRightShift(testGen, opName, shapeList, dtype, error_name=None): |
| 1549 | arg_list = [] |
| 1550 | |
| 1551 | arg_list.append(("roundTrue", [True])) |
| 1552 | arg_list.append(("roundFalse", [False])) |
| 1553 | |
| 1554 | return arg_list |
| 1555 | |
| 1556 | # Helper function for reshape. Gets some factors of a larger number. |
| 1557 | @staticmethod |
| 1558 | def getFactors(val, start=1): |
| 1559 | factors = [] |
| 1560 | |
| 1561 | for i in range(start, int(np.sqrt(val)) + 1): |
| 1562 | if (val % i) == 0: |
| 1563 | factors.append(i) |
| 1564 | |
| 1565 | return factors |
| 1566 | |
| 1567 | @staticmethod |
| 1568 | def agReshape(testGen, opName, shapeList, dtype, error_name=None): |
| 1569 | arg_list = [] |
| 1570 | |
| 1571 | origShape = shapeList[0] |
| 1572 | |
| 1573 | totalElements = 1 |
| 1574 | for s in origShape: |
| 1575 | totalElements *= s |
| 1576 | |
| 1577 | # This code is NOT fast. Fortunately, the numbers are fairly small. |
| 1578 | factors = TosaArgGen.getFactors(totalElements) |
| 1579 | |
| 1580 | for p in range(testGen.args.num_rand_permutations): |
| 1581 | newRank = testGen.randInt(1, 7) |
| 1582 | if len(factors) < newRank: |
| 1583 | continue |
| 1584 | |
| 1585 | found = True |
| 1586 | # escape_counter breaks while loop if it continues on for too long |
| 1587 | escape_counter = 0 |
| 1588 | while found: |
| 1589 | newShape = [] |
| 1590 | # Generate newShape ensuring it isn't a duplicate |
| 1591 | remainingElements = totalElements |
| 1592 | shuffledFactors = testGen.rng.permutation(factors) |
| 1593 | for i in range(1, newRank): |
| 1594 | # pick rank-1 factors |
| 1595 | newShape.append(shuffledFactors[0]) |
| 1596 | remainingElements = remainingElements // shuffledFactors[0] |
| 1597 | shuffledFactors = testGen.rng.permutation( |
| 1598 | TosaArgGen.getFactors(remainingElements) |
| 1599 | ) |
| 1600 | newShape.append(remainingElements) |
| 1601 | |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1602 | # Check for duplicates |
| 1603 | found = False |
| 1604 | for name, other_shape in arg_list: |
| 1605 | if other_shape[0] == newShape: |
| 1606 | found = True |
| 1607 | break |
| 1608 | |
| 1609 | escape_counter += 1 |
| 1610 | if escape_counter >= 100: |
| 1611 | break |
| 1612 | |
| 1613 | if not found: |
| 1614 | arg_list.append(("perm{}_rank{}".format(p, newRank), [newShape])) |
| 1615 | |
| 1616 | return arg_list |
| 1617 | |
| 1618 | @staticmethod |
| 1619 | def agTranspose(testGen, opName, shapeList, dtype, error_name=None): |
| 1620 | arg_list = [] |
| 1621 | |
| 1622 | ifm_shape = shapeList[0] |
| 1623 | |
| 1624 | if error_name == ErrorIf.IndexOutsideBounds: |
| 1625 | incorrect_large_index = range(len(ifm_shape) + 1, 2 * len(ifm_shape) + 1) |
| 1626 | incorrect_small_index = range(-len(ifm_shape), 0) |
| 1627 | permutations = [p for p in itertools.permutations(incorrect_large_index)] |
| 1628 | permutations.extend( |
| 1629 | [p for p in itertools.permutations(incorrect_small_index)] |
| 1630 | ) |
| 1631 | elif error_name == ErrorIf.IndexUsedTwice: |
| 1632 | # Create list with a duplicated index |
| 1633 | perm_range = list(range(len(ifm_shape))) |
| 1634 | index_choice = testGen.rng.choice(range(len(perm_range))) |
| 1635 | perm_range[(index_choice + 1) % len(perm_range)] = perm_range[index_choice] |
| 1636 | permutations = [p for p in itertools.permutations(perm_range)] |
| 1637 | |
| 1638 | else: |
| 1639 | # Get all permutations |
| 1640 | permutations = [p for p in itertools.permutations(range(len(ifm_shape)))] |
| 1641 | |
| 1642 | # Limit to possible permutations from shape dimension or argument setting |
| 1643 | limit = min(len(permutations), testGen.args.num_rand_permutations) |
| 1644 | |
| 1645 | # Get random permutation generator that uses all permutations |
| 1646 | random_permutations = testGen.rng.permutation(permutations) |
| 1647 | |
| 1648 | # Create list of required amount of permutations |
| 1649 | arg_list = [ |
| 1650 | ("perm{}".format(p), [random_permutations[p].tolist()]) |
| 1651 | for p in range(limit) |
| 1652 | ] |
| 1653 | return arg_list |
| 1654 | |
| 1655 | @staticmethod |
| 1656 | def agSlice(testGen, opName, shapeList, dtype, error_name=None): |
| 1657 | arg_list = [] |
| 1658 | |
| 1659 | ifm_shape = shapeList[0] |
| 1660 | rank = len(ifm_shape) |
| 1661 | |
| 1662 | for p in range(testGen.args.num_rand_permutations): |
| 1663 | start = [] |
| 1664 | size = [] |
| 1665 | |
| 1666 | valid = True |
| 1667 | |
| 1668 | for i in range(rank): |
| 1669 | if ifm_shape[i] > 1: |
| 1670 | start.append(testGen.randInt(0, ifm_shape[i])) |
| 1671 | size.append(testGen.randInt(0, ifm_shape[i] - start[i])) |
| 1672 | |
| 1673 | # Invalid slice size? |
| 1674 | if size[i] == 0: |
| 1675 | valid = False |
| 1676 | else: |
| 1677 | start.append(0) |
| 1678 | size.append(1) |
| 1679 | |
| 1680 | if valid: |
| 1681 | # If ERROR_IF test required then incorrect start, size will be returned |
| 1682 | start, size = TosaErrorIfArgGen.eiSliceErrorIf( |
| 1683 | testGen, error_name, ifm_shape, start, size |
| 1684 | ) |
| 1685 | arg_list.append(("perm{}".format(p), [start, size])) |
| 1686 | return arg_list |
| 1687 | |
| 1688 | @staticmethod |
| 1689 | def agTile(testGen, opName, shapeList, dtype, error_name=None): |
| 1690 | arg_list = [] |
| 1691 | |
| 1692 | ifm_shape = shapeList[0] |
| 1693 | rank = len(ifm_shape) |
| 1694 | |
| 1695 | for p in range(testGen.args.num_rand_permutations): |
| 1696 | |
| 1697 | # Pick a few random, but small multiple values |
| 1698 | # because otherwise this has a tendency to generate |
| 1699 | # enormous tensors |
| 1700 | multiples = [] |
| 1701 | for i in range(rank): |
| 1702 | if ifm_shape[i] > 1000: |
| 1703 | # Multiple of 1 if ifm_shape dimension is large to reduce |
| 1704 | # tensor size |
| 1705 | multiples.append(1) |
| 1706 | elif max(ifm_shape) > 1000: |
| 1707 | multiples.append(2) |
| 1708 | else: |
| 1709 | multiples.append(testGen.randInt(1, 4)) |
| 1710 | arg_list.append(("perm{}".format(p), [multiples])) |
| 1711 | |
| 1712 | return arg_list |
| 1713 | |
| 1714 | @staticmethod |
| 1715 | def agResize(testGen, opName, shapeList, dtype, error_name=None): |
| 1716 | arg_list = [] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1717 | ifm_shape = shapeList[0] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1718 | |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1719 | def get_aspect_ratio_resize_params(): |
| 1720 | common_aspect_ratios = ((3, 2), (16, 9), (4, 3)) |
| 1721 | aspect_ratio = testGen.rng.choice(common_aspect_ratios) |
| 1722 | invert = testGen.rng.choice((False, True)) |
| 1723 | letterbox = testGen.rng.choice((False, True)) |
| 1724 | |
| 1725 | scale_y_n = aspect_ratio[0] if invert else aspect_ratio[1] |
| 1726 | scale_x_n = aspect_ratio[1] if invert else aspect_ratio[0] |
| 1727 | scale_y_d = scale_x_d = 1 |
| 1728 | offset_x = offset_y = 0 |
| 1729 | |
| 1730 | if letterbox: |
| 1731 | max_border = scale_y_n |
| 1732 | border_y = testGen.randInt(low=0, high=max_border) |
| 1733 | border_x = 0 |
| 1734 | else: |
| 1735 | # Pillarboxing |
| 1736 | border_y = 0 |
| 1737 | max_border = scale_x_n |
| 1738 | border_x = testGen.randInt(low=0, high=max_border) |
| 1739 | |
| 1740 | scale = (scale_y_n, scale_y_d, scale_x_n, scale_x_d) |
| 1741 | offset = (offset_y, offset_x) |
| 1742 | border = (border_y, border_x) |
| 1743 | |
| 1744 | return scale, offset, border |
| 1745 | |
| 1746 | def get_upscale_downscale_params(): |
| 1747 | valid_params = False |
| 1748 | while not valid_params: |
| 1749 | upscale = testGen.rng.choice((False, True)) |
| 1750 | |
| 1751 | # True if sampling begins from (0,0). Otherwise (-0.5,-0.5) |
| 1752 | origin_sampling = testGen.rng.choice((False, True)) |
| 1753 | |
| 1754 | if upscale: |
| 1755 | shift = testGen.randInt(low=1, high=4) |
| 1756 | scale_x_d = scale_y_d = 1 |
| 1757 | scale_x_n = scale_y_n = ( |
| 1758 | 1 << shift if origin_sampling else 2 << shift |
| 1759 | ) |
| 1760 | border_x = border_y = 0 if origin_sampling else (1 << shift) - 1 |
| 1761 | offset_x = offset_y = 0 if origin_sampling else -(1 << shift) + 1 |
| 1762 | else: |
| 1763 | scale_x_n = 1 |
| 1764 | scale_y_n = 1 |
| 1765 | |
| 1766 | # Return list of valid scale_*_d values (max value 4) given input dim shape |
| 1767 | def get_valid_denom(ifm_dim): |
| 1768 | return [x for x in range(1, 5) if ifm_dim % x == 1] |
| 1769 | |
| 1770 | # Generate list of valid downscale values and choose one randomly |
| 1771 | valid_scale_y_ds = get_valid_denom(ifm_shape[1]) |
| 1772 | valid_scale_x_ds = get_valid_denom(ifm_shape[2]) |
| 1773 | |
| 1774 | if not valid_scale_y_ds and not valid_scale_x_ds: |
| 1775 | # Bad parameters, skip |
| 1776 | continue |
| 1777 | |
| 1778 | if not valid_scale_y_ds: |
| 1779 | scale_y_d = 1 |
| 1780 | else: |
| 1781 | scale_y_d = testGen.rng.choice(valid_scale_y_ds) |
| 1782 | |
| 1783 | if not valid_scale_x_ds: |
| 1784 | scale_x_d = 1 |
| 1785 | else: |
| 1786 | scale_x_d = testGen.rng.choice(valid_scale_x_ds) |
| 1787 | |
| 1788 | border_x = border_y = 0 |
| 1789 | offset_y = testGen.randInt(0, 16 * scale_y_n) |
| 1790 | offset_x = testGen.randInt(0, 16 * scale_x_n) |
| 1791 | valid_params = True |
| 1792 | |
| 1793 | scale = (scale_y_n, scale_y_d, scale_x_n, scale_x_d) |
| 1794 | offset = (offset_y, offset_x) |
| 1795 | border = (border_y, border_x) |
| 1796 | return scale, offset, border |
| 1797 | |
| 1798 | def get_rand_params(): |
| 1799 | # Scale |
| 1800 | scale_y_n = testGen.randInt(low=1, high=(1 << 11)) |
| 1801 | scale_x_n = testGen.randInt(low=1, high=(1 << 11)) |
| 1802 | |
| 1803 | scale_y_d = testGen.randInt(low=1, high=(16 * scale_y_n)) |
| 1804 | scale_x_d = testGen.randInt(low=1, high=(16 * scale_x_n)) |
| 1805 | |
| 1806 | # Offsets and border within the scale |
| 1807 | offset_y = testGen.randInt(low=-scale_y_n, high=(16 * scale_y_n)) |
| 1808 | offset_x = testGen.randInt(low=-scale_x_n, high=(16 * scale_x_n)) |
| 1809 | border_y = testGen.randInt(low=(-16 * scale_y_n), high=scale_y_n) |
| 1810 | border_x = testGen.randInt(low=(-16 * scale_x_n), high=scale_x_n) |
| 1811 | |
| 1812 | scale = (scale_y_n, scale_y_d, scale_x_n, scale_x_d) |
| 1813 | offset = (offset_y, offset_x) |
| 1814 | border = (border_y, border_x) |
| 1815 | return scale, offset, border |
| 1816 | |
| 1817 | for mode in [ResizeMode.NEAREST, ResizeMode.BILINEAR]: |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1818 | # Exclude illegal {mode, type} configurations. Pick legal output types |
| 1819 | if mode == ResizeMode.NEAREST and dtype == DType.INT8: |
| 1820 | outputDTypeList = [DType.INT8] |
| 1821 | elif mode == ResizeMode.NEAREST and dtype == DType.INT16: |
| 1822 | outputDTypeList = [DType.INT16] |
| 1823 | elif mode == ResizeMode.BILINEAR and dtype == DType.INT8: |
| 1824 | outputDTypeList = [DType.INT32] |
| 1825 | elif mode == ResizeMode.BILINEAR and dtype == DType.INT16: |
| 1826 | outputDTypeList = [DType.INT48] |
James Ward | 8b39043 | 2022-08-12 20:48:56 +0100 | [diff] [blame] | 1827 | elif dtype == DType.FP16: |
| 1828 | outputDTypeList = [DType.FP16] |
Jeremy Johnson | bc2a3db | 2022-09-27 13:50:00 +0100 | [diff] [blame] | 1829 | elif dtype == DType.FP32: |
| 1830 | outputDTypeList = [DType.FP32] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1831 | elif error_name == ErrorIf.WrongInputType: |
| 1832 | # If an incorrect input type is used then we set a 'correct' |
| 1833 | # output type to avoid other errors |
| 1834 | outputDTypeList = [DType.INT8, DType.INT16, DType.INT32] |
| 1835 | else: |
| 1836 | continue |
| 1837 | |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1838 | arg_str = "mode{}_out{}_sc{}x{}x{}x{}_off{}x{}_bor{}x{}" |
| 1839 | |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1840 | for outputDType in outputDTypeList: |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1841 | perm = 0 |
| 1842 | while perm < testGen.args.num_rand_permutations: |
| 1843 | # Random choice of type of params we are testing |
| 1844 | _rnd_param_fn = testGen.rng.choice( |
| 1845 | ( |
| 1846 | get_rand_params, |
| 1847 | get_upscale_downscale_params, |
| 1848 | get_aspect_ratio_resize_params, |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1849 | ) |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1850 | ) |
| 1851 | scale, offset, border = _rnd_param_fn() |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1852 | |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1853 | # Expand params for bounds-checking |
| 1854 | (scale_y_n, scale_y_d, scale_x_n, scale_x_d) = scale |
| 1855 | (offset_y, offset_x) = offset |
| 1856 | (border_y, border_x) = border |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1857 | |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1858 | # Make sure output dimensions OH and OW are integers |
| 1859 | partial_output_y = ( |
| 1860 | (ifm_shape[1] - 1) * scale_y_n - offset_y + border_y |
| 1861 | ) |
| 1862 | partial_output_x = ( |
| 1863 | (ifm_shape[2] - 1) * scale_x_n - offset_x + border_x |
| 1864 | ) |
| 1865 | if error_name == ErrorIf.ResizeOutputShapeNonInteger: |
| 1866 | if ( |
| 1867 | partial_output_y % scale_y_d == 0 |
| 1868 | and partial_output_x % scale_x_d == 0 |
| 1869 | ): |
| 1870 | # Skip this test as it doesn't produce NonInteger output |
| 1871 | perm += 1 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1872 | continue |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1873 | else: |
| 1874 | while partial_output_y % scale_y_d != 0: |
| 1875 | scale_y_d -= 1 |
| 1876 | while partial_output_x % scale_x_d != 0: |
| 1877 | scale_x_d -= 1 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1878 | |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1879 | output_y = partial_output_y // scale_y_d + 1 |
| 1880 | output_x = partial_output_x // scale_x_d + 1 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1881 | |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1882 | if ( |
| 1883 | output_y >= testGen.args.max_resize_output_dim |
| 1884 | or output_x >= testGen.args.max_resize_output_dim |
| 1885 | ) and error_name is None: |
| 1886 | # Skip positive test if output dim will be too high |
| 1887 | # Avoid high test latency and OOM issues |
| 1888 | perm += 1 |
| 1889 | continue |
| 1890 | |
| 1891 | if ( |
| 1892 | output_y <= 0 |
| 1893 | or output_y >= MAX_RESIZE_DIMENSION |
| 1894 | or output_x <= 0 |
| 1895 | or output_x >= MAX_RESIZE_DIMENSION |
| 1896 | ): |
| 1897 | # Output dimensions out of scope |
| 1898 | if error_name is not None and perm > 0: |
| 1899 | # As long as we have one ERROR_IF test, don't worry |
| 1900 | # about creating all the other permutations |
| 1901 | perm += 1 |
| 1902 | continue |
| 1903 | |
| 1904 | if error_name == ErrorIf.ResizeOutputShapeMismatch and ( |
| 1905 | ( |
| 1906 | output_y + scale_y_d >= MAX_RESIZE_DIMENSION |
| 1907 | and output_y - scale_y_d < 1 |
| 1908 | ) |
| 1909 | or ( |
| 1910 | output_x + scale_x_d >= MAX_RESIZE_DIMENSION |
| 1911 | and output_x - scale_x_d < 1 |
| 1912 | ) |
| 1913 | ): |
| 1914 | # Can't create a negative test with these params as it |
| 1915 | # will create invalid output size |
| 1916 | if perm > 0: |
| 1917 | perm += 1 |
| 1918 | continue |
| 1919 | |
| 1920 | scale = [scale_y_n, scale_y_d, scale_x_n, scale_x_d] |
| 1921 | offset = [offset_y, offset_x] |
| 1922 | border = [border_y, border_x] |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1923 | |
| 1924 | # Common for all data types |
| 1925 | if error_name is not None: |
| 1926 | ( |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1927 | scale, |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1928 | offset, |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1929 | border, |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1930 | outputDTypeNew, |
| 1931 | ) = TosaErrorIfArgGen.eiResizeErrorIf( |
| 1932 | testGen, |
| 1933 | error_name, |
| 1934 | mode, |
| 1935 | dtype, |
| 1936 | shapeList, |
| 1937 | outputDType, |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1938 | scale, |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1939 | offset, |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1940 | border, |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1941 | ) |
| 1942 | else: |
| 1943 | outputDTypeNew = outputDType |
| 1944 | |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1945 | arg_to_append = ( |
| 1946 | arg_str.format( |
| 1947 | "N" if mode == ResizeMode.NEAREST else "B", |
| 1948 | testGen.typeStr(outputDTypeNew), |
| 1949 | scale[0], |
| 1950 | scale[1], |
| 1951 | scale[2], |
| 1952 | scale[3], |
| 1953 | offset[0], |
| 1954 | offset[1], |
| 1955 | border[0], |
| 1956 | border[1], |
| 1957 | ), |
| 1958 | [ |
| 1959 | mode, |
| 1960 | scale, |
| 1961 | offset, |
| 1962 | border, |
| 1963 | dtype, |
| 1964 | outputDTypeNew, |
| 1965 | ], |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1966 | ) |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1967 | if arg_to_append in arg_list: |
| 1968 | # Skip already generated test params |
| 1969 | continue |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1970 | |
Jeremy Johnson | a0e03f3 | 2022-06-13 17:48:09 +0100 | [diff] [blame] | 1971 | # Valid permutation |
| 1972 | perm += 1 |
| 1973 | arg_list.append(arg_to_append) |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1974 | return arg_list |
| 1975 | |
| 1976 | @staticmethod |
| 1977 | def agTable(testGen, opName, shapeList, dtype, error_name=None): |
| 1978 | arg_list = [] |
| 1979 | |
| 1980 | if dtype == DType.INT8: |
| 1981 | table = np.int32( |
| 1982 | testGen.rng.integers(low=-128, high=128, size=[256]) |
| 1983 | ).tolist() |
| 1984 | else: # INT16 |
| 1985 | table = np.int32( |
| 1986 | testGen.rng.integers(low=-32768, high=32768, size=[513]) |
| 1987 | ).tolist() |
Jerry Ge | d511f9e | 2022-08-12 16:12:40 -0700 | [diff] [blame] | 1988 | # Make sure all slopes are within REQUIRE min/max 16-bit int |
| 1989 | for idx in range(len(table) - 1): |
| 1990 | slope = table[idx + 1] - table[idx] |
| 1991 | # Alter the next table entry to force the slope to be ok |
| 1992 | if slope > 32767: |
| 1993 | table[idx + 1] -= slope - 32767 |
| 1994 | if slope < -32768: |
| 1995 | table[idx + 1] -= slope + 32768 |
| 1996 | slope = table[idx + 1] - table[idx] |
| 1997 | assert slope <= 32767 and slope >= -32768 |
Jeremy Johnson | 9a66abb | 2022-04-07 11:29:20 +0100 | [diff] [blame] | 1998 | arg_list.append( |
| 1999 | ( |
| 2000 | "", |
| 2001 | [table], |
| 2002 | ) |
| 2003 | ) |
| 2004 | return arg_list |
| 2005 | |
| 2006 | def agCondIf(testGen, opName, shapeList, dtype, error_name=None): |
| 2007 | # CondIf generates the condition values here. |
| 2008 | # Convert to tensors in the build function, along with the |
| 2009 | # then and else blocks |
| 2010 | arg_list = [] |
| 2011 | |
| 2012 | for c in [False, True]: |
| 2013 | arg_list.append(("cond{}".format(int(c)), [c])) |
| 2014 | |
| 2015 | return arg_list |
| 2016 | |
| 2017 | def agWhileLoop(testGen, opName, shapeList, dtype, error_name=None): |
| 2018 | # While loop: 0 iterations, 1, more than 1 |
| 2019 | arg_list = [] |
| 2020 | |
| 2021 | for iter in [0, 1, 4]: |
| 2022 | arg_list.append(("iter{}".format(iter), [iter])) |
| 2023 | |
| 2024 | return arg_list |