topi.cc 23.5 KB
Newer Older
1 2 3 4 5 6 7 8
/*
 * Licensed to the Apache Software Foundation (ASF) under one
 * or more contributor license agreements.  See the NOTICE file
 * distributed with this work for additional information
 * regarding copyright ownership.  The ASF licenses this file
 * to you under the Apache License, Version 2.0 (the
 * "License"); you may not use this file except in compliance
 * with the License.  You may obtain a copy of the License at
9
 *
10
 *   http://www.apache.org/licenses/LICENSE-2.0
11
 *
12 13 14 15 16 17 18 19
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
 * KIND, either express or implied.  See the License for the
 * specific language governing permissions and limitations
 * under the License.
 */

20 21 22 23 24
/*!
*  Copyright (c) 2017 by Contributors
* \brief Registration of TVM operators and schedules
* \file topi.cc
*/
25 26
#define TOPI_REDUCE_ATLEAST1D 0

27 28 29 30 31 32 33 34 35 36 37 38
#include <tvm/runtime/packed_func.h>
#include <tvm/runtime/module.h>
#include <tvm/runtime/registry.h>
#include <tvm/packed_func_ext.h>
#include <tvm/build_module.h>

#include <topi/broadcast.h>
#include <topi/elemwise.h>
#include <topi/nn.h>
#include <topi/reduction.h>
#include <topi/transform.h>

39
#include <topi/nn/bias_add.h>
40 41 42 43 44 45 46
#include <topi/nn/bnn.h>
#include <topi/nn/dense.h>
#include <topi/nn/dilate.h>
#include <topi/nn/flatten.h>
#include <topi/nn/mapping.h>
#include <topi/nn/pooling.h>
#include <topi/nn/softmax.h>
47
#include <topi/nn/upsampling.h>
48 49
#include <topi/nn/l2_normalize.h>
#include <topi/nn/local_response_norm.h>
50
#include <topi/nn/batch_matmul.h>
51

52
#include <topi/vision/reorg.h>
53
#include <topi/image/resize.h>
54 55 56 57 58 59 60 61 62 63
#include <topi/generic/default.h>
#include <topi/generic/extern.h>
#include <topi/generic/injective.h>

#include <topi/cuda/dense.h>
#include <topi/cuda/extern.h>
#include <topi/cuda/injective.h>
#include <topi/cuda/pooling.h>
#include <topi/cuda/reduction.h>
#include <topi/cuda/softmax.h>
64
#include <topi/cuda/normalization.h>
65 66 67 68 69 70

#include <topi/x86/bnn.h>
#include <topi/x86/default.h>
#include <topi/x86/injective.h>

#include <topi/rocm/dense.h>
71
#include <topi/rocm/normalization.h>
72 73

namespace topi {
74

75 76 77 78
using namespace tvm;
using namespace tvm::runtime;

/*! \brief Canonicalize an argument that may be Array<Expr> or int to Array<Expr> */
79
Array<Integer> ArrayOrInt(TVMArgValue arg) {
80
  if (arg.type_code() == kDLInt || arg.type_code() == kDLUInt) {
81
    Array<Integer> result;
82 83 84 85 86 87 88
    result.push_back(arg.operator int());
    return result;
  } else {
    return arg;
  }
}

89 90 91 92 93 94
inline bool IsTensorType(TVMArgValue arg) {
  return (arg.type_code() == kNodeHandle &&
          arg.node_sptr()->is_type<tvm::TensorNode>());
}


95 96
TVM_REGISTER_GLOBAL("topi.TEST_create_target")
.set_body([](TVMArgs args, TVMRetValue *rv) {
97
  *rv = tvm::Target::Create(args[0]);
98 99 100
  });

/* Ops from broadcast.h */
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120
#define TOPI_REGISTER_BCAST_OP(OpName, Op)                              \
  TVM_REGISTER_GLOBAL(OpName)                                           \
  .set_body([](TVMArgs args, TVMRetValue *rv) {                         \
      bool lhs_is_tensor = IsTensorType(args[0]);                       \
      bool rhs_is_tensor = IsTensorType(args[1]);                       \
      if (lhs_is_tensor && rhs_is_tensor) {                             \
        *rv = Op(args[0].operator tvm::Tensor(),                        \
                 args[1].operator tvm::Tensor());                       \
      } else if (!lhs_is_tensor && rhs_is_tensor) {                     \
        *rv = Op(args[0].operator tvm::Expr(),                          \
                 args[1].operator tvm::Tensor());                       \
      } else if (lhs_is_tensor && !rhs_is_tensor) {                     \
        *rv = Op(args[0].operator tvm::Tensor(),                        \
                 args[1].operator tvm::Expr());                         \
      } else if (!lhs_is_tensor && !rhs_is_tensor) {                    \
        *rv = Op(args[0].operator tvm::Expr(),                          \
                 args[1].operator tvm::Expr());                         \
      }                                                                 \
    });                                                                 \

121 122 123 124 125
TVM_REGISTER_GLOBAL("topi.broadcast_to")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = broadcast_to(args[0], args[1]);
  });

126 127 128 129 130 131 132 133 134
TOPI_REGISTER_BCAST_OP("topi.add", topi::add);
TOPI_REGISTER_BCAST_OP("topi.subtract", topi::subtract);
TOPI_REGISTER_BCAST_OP("topi.multiply", topi::multiply);
TOPI_REGISTER_BCAST_OP("topi.divide", topi::divide);
TOPI_REGISTER_BCAST_OP("topi.mod", topi::mod);
TOPI_REGISTER_BCAST_OP("topi.maximum", topi::maximum);
TOPI_REGISTER_BCAST_OP("topi.minimum", topi::minimum);
TOPI_REGISTER_BCAST_OP("topi.power", topi::power);
TOPI_REGISTER_BCAST_OP("topi.left_shift", topi::left_shift);
135 136
TOPI_REGISTER_BCAST_OP("topi.logical_and", topi::logical_and);
TOPI_REGISTER_BCAST_OP("topi.logical_or", topi::logical_or);
137 138 139
TOPI_REGISTER_BCAST_OP("topi.right_shift", topi::right_shift);
TOPI_REGISTER_BCAST_OP("topi.greater", topi::greater);
TOPI_REGISTER_BCAST_OP("topi.less", topi::less);
140 141 142 143
TOPI_REGISTER_BCAST_OP("topi.equal", topi::equal);
TOPI_REGISTER_BCAST_OP("topi.not_equal", topi::not_equal);
TOPI_REGISTER_BCAST_OP("topi.greater_equal", topi::greater_equal);
TOPI_REGISTER_BCAST_OP("topi.less_equal", topi::less_equal);
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165

/* Ops from elemwise.h */
TVM_REGISTER_GLOBAL("topi.exp")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = exp(args[0]);
  });

TVM_REGISTER_GLOBAL("topi.tanh")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = tanh(args[0]);
  });

TVM_REGISTER_GLOBAL("topi.sigmoid")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = sigmoid(args[0]);
  });

TVM_REGISTER_GLOBAL("topi.sqrt")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = sqrt(args[0]);
  });

166 167 168 169 170
TVM_REGISTER_GLOBAL("topi.rsqrt")
.set_body([](TVMArgs args, TVMRetValue *rv) {
*rv = rsqrt(args[0]);
});

171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195
TVM_REGISTER_GLOBAL("topi.log")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = log(args[0]);
  });

TVM_REGISTER_GLOBAL("topi.identity")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = identity(args[0]);
  });

TVM_REGISTER_GLOBAL("topi.negative")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = negative(args[0]);
  });

TVM_REGISTER_GLOBAL("topi.clip")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = clip(args[0], args[1], args[2]);
  });

TVM_REGISTER_GLOBAL("topi.cast")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = cast(args[0], args[1]);
  });

196 197 198 199 200
TVM_REGISTER_GLOBAL("topi.elemwise_sum")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = elemwise_sum(args[0]);
  });

201 202 203 204 205
TVM_REGISTER_GLOBAL("topi.sign")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = sign(args[0]);
  });

206 207 208 209 210 211 212 213 214 215
TVM_REGISTER_GLOBAL("topi.full")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = full(args[0], args[1], args[2]);
  });

TVM_REGISTER_GLOBAL("topi.full_like")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = full_like(args[0], args[1]);
  });

216 217 218 219 220 221 222 223
/* Ops from nn.h */
TVM_REGISTER_GLOBAL("topi.nn.relu")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = relu<float>(args[0]);
  });

TVM_REGISTER_GLOBAL("topi.nn.leaky_relu")
.set_body([](TVMArgs args, TVMRetValue *rv) {
224
  *rv = leaky_relu(args[0], args[1]);
225 226
  });

227 228
TVM_REGISTER_GLOBAL("topi.nn.prelu")
.set_body([](TVMArgs args, TVMRetValue *rv) {
229
  *rv = prelu(args[0], args[1], args[2]);
230 231
  });

232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262
TVM_REGISTER_GLOBAL("topi.nn.pad")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = pad(args[0], args[1], args[2], args[3]);
  });

/* Ops from reduction.h */
TVM_REGISTER_GLOBAL("topi.sum")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::sum(args[0], ArrayOrInt(args[1]), args[2]);
  });

TVM_REGISTER_GLOBAL("topi.min")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::min(args[0], ArrayOrInt(args[1]), args[2]);
  });

TVM_REGISTER_GLOBAL("topi.max")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::max(args[0], ArrayOrInt(args[1]), args[2]);
  });

TVM_REGISTER_GLOBAL("topi.argmin")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::argmin(args[0], ArrayOrInt(args[1]), args[2]);
  });

TVM_REGISTER_GLOBAL("topi.argmax")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::argmax(args[0], ArrayOrInt(args[1]), args[2]);
  });

263 264 265 266 267
TVM_REGISTER_GLOBAL("topi.prod")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::prod(args[0], ArrayOrInt(args[1]), args[2]);
  });

268 269 270 271 272
TVM_REGISTER_GLOBAL("topi.all")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::all(args[0], ArrayOrInt(args[1]), args[2]);
  });

273 274 275 276 277 278 279 280 281 282 283
/* Ops from transform.h */
TVM_REGISTER_GLOBAL("topi.expand_dims")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = expand_dims(args[0], args[1], args[2]);
  });

TVM_REGISTER_GLOBAL("topi.transpose")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = transpose(args[0], args[1]);
  });

284 285 286 287 288
TVM_REGISTER_GLOBAL("topi.flip")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = flip(args[0], args[1]);
  });

289 290 291 292 293 294 295 296 297 298 299 300 301 302 303
TVM_REGISTER_GLOBAL("topi.reshape")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = reshape(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.squeeze")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = squeeze(args[0], ArrayOrInt(args[1]));
  });

TVM_REGISTER_GLOBAL("topi.concatenate")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = concatenate(args[0], args[1]);
  });

304 305 306 307 308
TVM_REGISTER_GLOBAL("topi.stack")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = stack(args[0], args[1]);
});

309 310 311 312 313
TVM_REGISTER_GLOBAL("topi.shape")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = shape(args[0], args[1]);
});

314 315 316 317 318 319 320
TVM_REGISTER_GLOBAL("topi.split")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  if (args[1].type_code() == kDLInt || args[1].type_code() == kDLUInt) {
    *rv = split_sections(args[0], args[1], args[2]);
  } else {
    *rv = split(args[0], args[1], args[2]);
  }
321
});
322

323 324 325 326 327
TVM_REGISTER_GLOBAL("topi.layout_transform")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = layout_transform(args[0], args[1], args[2]);
});

328 329
TVM_REGISTER_GLOBAL("topi.take")
.set_body([](TVMArgs args, TVMRetValue *rv) {
330 331 332
  if (args.size() == 3) {
    std::string mode = args[2];
    *rv = take(args[0], args[1], mode);
333 334
  } else {
    int axis = args[2];
335 336
    std::string mode = args[3];
    *rv = take(args[0], args[1], axis, mode);
337 338 339
  }
  });

340 341 342 343 344 345 346 347
TVM_REGISTER_GLOBAL("topi.sequence_mask")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  double pad_val = args[2];
  int axis = args[3];
  *rv = sequence_mask(args[0], args[1], pad_val, axis);
});


348 349 350 351
TVM_REGISTER_GLOBAL("topi.where")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = where(args[0], args[1], args[2]);
});
352

353 354 355 356 357
TVM_REGISTER_GLOBAL("topi.arange")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = arange(args[0], args[1], args[2], args[3]);
});

358 359 360 361 362 363 364 365 366 367
TVM_REGISTER_GLOBAL("topi.repeat")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = repeat(args[0], args[1], args[2]);
});

TVM_REGISTER_GLOBAL("topi.tile")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = tile(args[0], args[1]);
});

368 369 370 371
TVM_REGISTER_GLOBAL("topi.gather_nd")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = gather_nd(args[0], args[1]);
});
372

373 374 375 376 377 378 379 380 381
TVM_REGISTER_GLOBAL("topi.matmul")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  switch ( args.size() ) {
    case 2: *rv = matmul(args[0], args[1]); break;
    case 3: *rv = matmul(args[0], args[1], args[2]); break;
    case 4: *rv = matmul(args[0], args[1], args[2], args[3]); break;
    default: CHECK(0) << "topi.matmul expects 2, 3 or 4 arguments";
  }});

382 383 384 385 386 387 388 389 390 391 392 393
TVM_REGISTER_GLOBAL("topi.tensordot")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  if (args.size() == 2) {
    *rv = tensordot(args[0], args[1]);
  } else if (args.size() == 3) {
    *rv = tensordot(args[0], args[1], args[2]);
  } else {
    Array<Expr> axes = args[3];
    *rv = tensordot(args[0], args[1], args[2], axes);
  }
  });

394 395 396 397 398
TVM_REGISTER_GLOBAL("topi.strided_slice")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = strided_slice(args[0], args[1], args[2], args[3]);
  });

399 400 401 402 403 404
/* Ops from nn/upsampling.h */
TVM_REGISTER_GLOBAL("topi.nn.upsampling")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::upsampling(args[0], args[1], args[2], args[3]);
  });

405 406 407 408 409 410 411 412 413 414 415 416 417 418
/* Ops from nn/bnn.h */
TVM_REGISTER_GLOBAL("topi.nn.binarize_pack")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::binarize_pack(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.nn.binary_dense")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::binary_dense(args[0], args[1]);
  });

/* Ops from nn/dense.h */
TVM_REGISTER_GLOBAL("topi.nn.dense")
.set_body([](TVMArgs args, TVMRetValue *rv) {
419
  *rv = nn::dense(args[0], args[1], args[2], args[3]);
420 421
  });

422 423 424 425 426 427
/* Ops from nn/bias_add.h */
TVM_REGISTER_GLOBAL("topi.nn.bias_add")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::bias_add(args[0], args[1], args[2]);
  });

428 429 430 431 432 433
/* Ops from nn/batch_matmul.h */
TVM_REGISTER_GLOBAL("topi.nn.batch_matmul")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::batch_matmul(args[0], args[1]);
  });

434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461
/* Ops from nn/dilate.h */
TVM_REGISTER_GLOBAL("topi.nn.dilate")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::dilate(args[0], args[1]);
  });

/* Ops from nn/flatten.h */
TVM_REGISTER_GLOBAL("topi.nn.flatten")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::flatten(args[0]);
  });

/* Ops from nn/mapping.h */
TVM_REGISTER_GLOBAL("topi.nn.scale_shift_nchw")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::scale_shift_nchw(args[0], args[1], args[2]);
  });

TVM_REGISTER_GLOBAL("topi.nn.scale_shift_nhwc")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::scale_shift_nhwc(args[0], args[1], args[2]);
  });

/* Ops from nn/pooling.h */
TVM_REGISTER_GLOBAL("topi.nn.pool")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::pool(args[0], args[1], args[2], args[3],
                 static_cast<nn::PoolType>(static_cast<int>(args[4])),
462
                 args[5], args[6], args[7]);
463 464 465 466 467 468 469 470
  });

TVM_REGISTER_GLOBAL("topi.nn.global_pool")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::global_pool(args[0],
                        static_cast<nn::PoolType>(static_cast<int>(args[1])));
  });

471 472 473 474 475 476 477
TVM_REGISTER_GLOBAL("topi.nn.adaptive_pool")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::adaptive_pool(args[0], args[1],
                          static_cast<nn::PoolType>(static_cast<int>(args[2])),
                          args[3]);
});

478 479 480
/* Ops from nn/softmax.h */
TVM_REGISTER_GLOBAL("topi.nn.softmax")
.set_body([](TVMArgs args, TVMRetValue *rv) {
481
  *rv = nn::softmax(args[0], args[1]);
482 483 484 485 486 487 488
  });

TVM_REGISTER_GLOBAL("topi.nn.log_softmax")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::log_softmax(args[0]);
  });

489 490 491 492 493 494 495 496 497 498 499 500 501 502
/* Ops from nn/l2_normalize.h */
TVM_REGISTER_GLOBAL("topi.nn.l2_normalize")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::l2_normalize(args[0], static_cast<double>(args[1]), args[2]);
  });

TVM_REGISTER_GLOBAL("topi.nn.lrn")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = nn::lrn(args[0], args[1], args[2],
                static_cast<double>(args[3]),
                static_cast<double>(args[4]),
                static_cast<double>(args[5]));
  });

503 504 505 506
TVM_REGISTER_GLOBAL("topi.vision.reorg")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = vision::reorg(args[0], args[1]);
  });
507 508

/* Ops from image/resize.h */
509 510 511 512 513
TVM_REGISTER_GLOBAL("topi.image.bilinear_sample_nchw")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = image::bilinear_sample_nchw(args[0], args[1], args[2], args[3]);
  });

514 515 516 517 518
TVM_REGISTER_GLOBAL("topi.image.resize")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = image::resize(args[0], args[1], args[2], args[3], args[4]);
  });

519 520 521
/* Generic schedules */
TVM_REGISTER_GLOBAL("topi.generic.default_schedule")
.set_body([](TVMArgs args, TVMRetValue *rv) {
522 523 524 525 526
  if (args[2]) {
    *rv = topi::generic::default_schedule_auto_inline(args[0], args[1]);
  } else {
    *rv = topi::generic::default_schedule(args[0], args[1]);
  }
527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551
  });

TVM_REGISTER_GLOBAL("topi.generic.schedule_extern")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::generic::schedule_extern(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.generic.schedule_injective")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::generic::schedule_injective(args[0], args[1]);
  });

/* x86 schedules */
TVM_REGISTER_GLOBAL("topi.x86.schedule_binarize_pack")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::x86::schedule_binarize_pack(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.x86.schedule_binary_dense")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::x86::schedule_binary_dense(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.x86.default_schedule")
.set_body([](TVMArgs args, TVMRetValue *rv) {
552 553 554 555 556
  if (args[2]) {
    *rv = topi::x86::default_schedule_auto_inline(args[0], args[1]);
  } else {
    *rv = topi::x86::default_schedule(args[0], args[1]);
  }
557 558 559 560 561 562 563 564 565 566
  });

TVM_REGISTER_GLOBAL("topi.x86.schedule_injective")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::x86::schedule_injective(args[0], args[1]);
  });

/* ROCm schedules */
TVM_REGISTER_GLOBAL("topi.rocm.dense_cuda")
.set_body([](TVMArgs args, TVMRetValue *rv) {
567
  *rv = rocm::dense_rocm(args[0], args[1], args[2], args[3], args[4]);
568 569 570 571 572 573 574
  });

TVM_REGISTER_GLOBAL("topi.rocm.schedule_dense")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::rocm::schedule_dense(args[0], args[1]);
  });

575 576 577 578 579 580 581 582 583 584
TVM_REGISTER_GLOBAL("topi.rocm.schedule_lrn")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::rocm::schedule_lrn(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.rocm.schedule_l2_normalize")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::rocm::schedule_l2_normalize(args[0], args[1]);
  });

585 586 587
/* CUDA schedules */
TVM_REGISTER_GLOBAL("topi.cuda.dense_cuda")
.set_body([](TVMArgs args, TVMRetValue *rv) {
588
  *rv = cuda::dense_cuda(args[0], args[1], args[2], args[3], args[4]);
589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625
  });

TVM_REGISTER_GLOBAL("topi.cuda.schedule_dense")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_dense(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.cuda.schedule_extern")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_extern(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.cuda.schedule_injective")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_injective(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.cuda.schedule_pool")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_pool(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.cuda.schedule_global_pool")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_global_pool(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.cuda.schedule_reduce")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_reduce(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.cuda.schedule_softmax")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_softmax(args[0], args[1]);
  });

626 627 628 629 630 631 632 633 634 635
TVM_REGISTER_GLOBAL("topi.cuda.schedule_lrn")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_lrn(args[0], args[1]);
  });

TVM_REGISTER_GLOBAL("topi.cuda.schedule_l2_normalize")
.set_body([](TVMArgs args, TVMRetValue *rv) {
  *rv = topi::cuda::schedule_l2_normalize(args[0], args[1]);
  });

636 637 638 639 640 641 642 643 644 645 646 647 648 649 650
/*! \brief Builder function for instantiating schedules. */
using FTVMScheduleBuilder = std::function<
  tvm::Schedule(const tvm::Target& target, const tvm::Array<tvm::Tensor>& outs)>;

/*!
 * \brief Helper function for registering generic functions matching the
 * FTVMScheduleBuilder signature. The schedule builder function is wrapped
 * with a PackedFunc suitable for passing to a tvm::GenericFunc.
 *
 * \param builder The schedule builder to wrap.
 *
 * \return The wrapped schedule builder
 */
inline PackedFunc WrapSchedule(FTVMScheduleBuilder builder) {
  return PackedFunc([builder](TVMArgs args, TVMRetValue* ret) {
651
    auto target = Target::Current(false);
652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678
    Array<Tensor> outs;
    NodeRef argNodeRef = args[0];
    if (argNodeRef->type_index() == outs->type_index()) {
      outs = args[0];
    } else {
      outs = Array<Tensor> { args[0] };
    }

    *ret = builder(target, outs);
  });
}

TVM_REGISTER_GENERIC_FUNC(schedule_injective)
.set_default(WrapSchedule(topi::generic::schedule_injective))
.register_func({ "cpu" }, WrapSchedule(topi::x86::schedule_injective))
.register_func({ "cuda", "gpu" }, WrapSchedule(topi::cuda::schedule_injective));

TVM_REGISTER_GENERIC_FUNC(schedule_softmax)
.set_default(WrapSchedule(topi::generic::default_schedule))
.register_func({ "cpu" }, WrapSchedule(topi::x86::default_schedule))
.register_func({ "cuda", "gpu" }, WrapSchedule(topi::cuda::schedule_softmax));

TVM_REGISTER_GENERIC_FUNC(schedule_dense)
.set_default(WrapSchedule(topi::generic::default_schedule))
.register_func({ "cuda", "gpu" }, WrapSchedule(topi::cuda::schedule_dense))
.register_func({ "rocm" }, WrapSchedule(topi::rocm::schedule_dense));

679 680 681
TVM_REGISTER_GENERIC_FUNC(schedule_batch_matmul)
.set_default(WrapSchedule(topi::generic::default_schedule));

682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708
TVM_REGISTER_GENERIC_FUNC(schedule_pool)
.set_default(WrapSchedule(topi::generic::default_schedule))
.register_func({ "cpu" }, WrapSchedule(topi::x86::default_schedule))
.register_func({ "cuda", "gpu" }, WrapSchedule(topi::cuda::schedule_pool));

TVM_REGISTER_GENERIC_FUNC(schedule_global_pool)
.set_default(WrapSchedule(topi::generic::default_schedule))
.register_func({ "cpu" }, WrapSchedule(topi::x86::default_schedule))
.register_func({ "cuda", "gpu" }, WrapSchedule(topi::cuda::schedule_global_pool));

TVM_REGISTER_GENERIC_FUNC(schedule_reduce)
.set_default(WrapSchedule(topi::generic::default_schedule_auto_inline))
.register_func({ "cpu" }, WrapSchedule(topi::x86::default_schedule_auto_inline))
.register_func({ "cuda", "gpu" }, WrapSchedule(topi::cuda::schedule_reduce));

TVM_REGISTER_GENERIC_FUNC(schedule_binarize_pack)
.set_default(WrapSchedule(topi::generic::default_schedule))
.register_func({ "cpu" }, WrapSchedule(topi::x86::schedule_binarize_pack));

TVM_REGISTER_GENERIC_FUNC(schedule_binary_dense)
.set_default(WrapSchedule(topi::generic::default_schedule))
.register_func({ "cpu" }, WrapSchedule(topi::x86::schedule_binary_dense));

/*! \brief Builder function for instantiating dense ops. */
using FTVMDenseOpBuilder = std::function<tvm::Tensor(const Target& target,
                                                     const tvm::Tensor& data,
                                                     const tvm::Tensor& weight,
709 710
                                                     const tvm::Tensor& bias,
                                                     const Type& out_dtype)>;
711 712 713 714 715 716 717 718 719 720 721 722

/*!
* \brief Helper function for registering dense ops matching the
* FTVMDenseOpBuilder signature. The op builder function is wrapped
* with a PackedFunc suitable for passing to a tvm::GenericFunc.
*
* \param builder The op builder to wrap.
*
* \return The wrapped op builder
*/
inline PackedFunc WrapDenseOp(FTVMDenseOpBuilder builder) {
  return PackedFunc([builder](TVMArgs args, TVMRetValue* ret) {
723
    auto target = Target::Current(false);
724 725 726
    Tensor data = args[0];
    Tensor weight = args[1];
    Tensor bias = args[2];
727
    Type out_dtype = args[3];
728

729
    *ret = builder(target, data, weight, bias, out_dtype);
730 731 732 733 734 735 736
  });
}

TVM_REGISTER_GENERIC_FUNC(dense)
.set_default(WrapDenseOp([](const Target& target,
                            const tvm::Tensor& data,
                            const tvm::Tensor& weight,
737 738 739
                            const tvm::Tensor& bias,
                            const Type& out_dtype) {
  return topi::nn::dense(data, weight, bias, out_dtype);
740 741 742 743
}))
.register_func({ "cuda", "gpu" }, WrapDenseOp(topi::cuda::dense_cuda))
.register_func({ "rocm" }, WrapDenseOp(topi::rocm::dense_rocm));

744
}  // namespace topi