tensorize.cc 20.5 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 25 26 27 28
/*!
 *  Copyright (c) 2017 by Contributors
 * \brief Logics related to tensorize, used by ComputeOpNode.
 * \file tensorize.cc
 */
#include <tvm/ir.h>
#include <tvm/ir_mutator.h>
#include <tvm/ir_pass.h>
#include <tvm/api_registry.h>
29 30
#include "op_util.h"
#include "compute_op.h"
31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
#include "../schedule/message_passing.h"

namespace tvm {

using namespace ir;
using namespace op;

// Detect the region of input and output to be tensrized.
// out_dom: the domain of root iter vars in output op
// in_region: region of each input tensor.
// return The location of the tensorized scope start.
size_t InferTensorizeRegion(
    const ComputeOpNode* self,
    const Stage& stage,
    const std::unordered_map<IterVar, Range>& dom_map,
    std::unordered_map<IterVar, Range>* out_dom,
    std::unordered_map<Tensor, Array<Range> >* in_region) {
  // Get the bound of the tensorized scope.
  bool found_point = false;
  size_t loc_scope = 0;
  std::unordered_map<IterVar, IntSet> up_state;
  // Loop over the leafs
  for (size_t i = stage->leaf_iter_vars.size(); i != 0; --i) {
    IterVar iv = stage->leaf_iter_vars[i - 1];
    CHECK(iv->iter_type == kDataPar ||
          iv->iter_type == kCommReduce);
    auto vit = dom_map.find(iv);
    CHECK(vit != dom_map.end());
    const Range& vrange = vit->second;
    if (is_one(vrange->extent)) {
      up_state[iv] = IntSet::single_point(vrange->min);
    } else if (found_point) {
      CHECK(is_zero(vrange->min));
      up_state[iv] = IntSet::single_point(iv->var);
    } else {
      up_state[iv] = IntSet::range(vrange);
    }
    auto iit = stage->iter_var_attrs.find(iv);
    if (iit != stage->iter_var_attrs.end()) {
      const IterVarAttr& attr = (*iit).second;
      if (!found_point) {
        CHECK(!attr->bind_thread.defined())
Siju committed
73
            << "Do not allow thread in tensorize scope";
74 75
      }
      if (attr->iter_type == kTensorized) {
Siju committed
76
        CHECK(!found_point) << "Do not allow two tensorized point";
77 78 79 80 81 82 83 84 85 86 87
        found_point = true;
        loc_scope = i - 1;
      }
    }
  }
  CHECK(found_point);
  // Get domain of the tensorized scope.
  schedule::PassUpDomain(stage, dom_map, &up_state);
  // Get domains if inputs
  std::unordered_map<Tensor, TensorDom> in_dom;
  std::unordered_map<const Variable*, IntSet> temp_dmap;
88
  arith::Analyzer analyzer;
89 90 91 92 93 94
  Array<Tensor> inputs = self->InputTensors();
  for (Tensor t : inputs) {
    in_dom.emplace(t, TensorDom(t.ndim()));
  }
  for (IterVar iv : self->root_iter_vars()) {
    IntSet iset = up_state.at(iv);
95 96 97
    Range iv_range = iset.cover_range(dom_map.at(iv));
    (*out_dom)[iv] = iv_range;
    analyzer.Bind(iv->var, iv_range);
98 99 100
    temp_dmap[iv->var.get()] = iset;
  }
  // Input domains
101
  self->PropBoundToInputs(stage->op, &analyzer, temp_dmap, &in_dom);
102 103 104 105
  Range none;
  for (const auto& kv : in_dom) {
    Array<Range> vec;
    const Tensor& t = kv.first;
106
    for (size_t i = 0; i < t.ndim(); ++i) {
107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166
      Range r = arith::Union(kv.second.data.at(i)).cover_range(none);
      CHECK(r.defined()) << "cannot deduce region of tensorized scope for input " << t;
      vec.push_back(std::move(r));
    }
    (*in_region)[t] = std::move(vec);
  }
  return loc_scope;
}

void VerifyTensorizeLoopNest(const ComputeOpNode* self,
                             const Stage& stage,
                             const ComputeLoopNest& n,
                             size_t tloc) {
  // Veirfication step.
  std::unordered_set<const Variable*> banned;
  CHECK_EQ(n.main_nest.size(), stage->leaf_iter_vars.size() + 1);
  CHECK(n.init_nest.size() == stage->leaf_iter_vars.size() + 1 ||
        n.init_nest.size() == 0);
  auto f_push_banned = [&banned](const Stmt& s) {
    if (const For* op = s.as<For>()) {
        banned.insert(op->loop_var.get());
    } else if (const AttrStmt* op = s.as<AttrStmt>()) {
      if (const IterVarNode* iv = op->node.as<IterVarNode>()) {
        banned.insert(iv->var.get());
      }
    } else if (const LetStmt* op = s.as<LetStmt>()) {
      banned.insert(op->var.get());
    }
  };
  for (size_t i = tloc; i < stage->leaf_iter_vars.size(); ++i) {
    for (const Stmt& s : n.main_nest[i + 1]) {
      f_push_banned(s);
    }
    if (n.init_nest.size() != 0) {
      for (const Stmt& s : n.init_nest[i + 1]) {
        f_push_banned(s);
      }
    }
  }
  for (const Expr& pred : n.main_predicates) {
    if (ir::ExprUseVar(pred, banned)) {
      LOG(FATAL) << "Tensorize failed, split condition "
                 << pred << " relies on var defined inside tensorize scope";
    }
  }
  for (const Expr& pred : n.init_predicates) {
    if (ir::ExprUseVar(pred, banned)) {
      LOG(FATAL) << "Tensorize failed, split condition "
                 << pred << " relies on var defined inside tensorize scope";
    }
  }
}

// Remap the tensor placeholder, index and inline things.
class TensorIntrinMatcher final : public IRMutator {
 public:
  Expr Mutate_(const Call* op, const Expr& e) final {
    Expr expr = IRMutator::Mutate_(op, e);
    op = expr.as<Call>();
    if (op->call_type == Call::Halide) {
167
      Tensor t = Downcast<Operation>(op->func).output(op->value_index);
168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195
      auto it = in_remap_.find(t);
      if (it != in_remap_.end()) {
        const InputEntry& e = it->second;
        CHECK_EQ(op->args.size(), e.region.size());
        Array<Expr> args;
        for (size_t i = e.start; i < e.region.size(); ++i) {
          args.push_back(op->args[i] - e.region[i]->min);
        }
        return Call::make(
            op->type, e.tensor->op->name, args,
            op->call_type, e.tensor->op, e.tensor->value_index);
      }
    }
    return expr;
  }

  Expr Mutate_(const Variable* op, const Expr& e) final {
    auto it = var_remap_.find(op);
    if (it != var_remap_.end()) {
      return it->second;
    } else {
      return e;
    }
  }

  Expr Mutate_(const Reduce* op, const Expr& e) final {
    Expr expr = IRMutator::Mutate_(op, e);
    op = expr.as<Reduce>();
196
    Array<IterVar> axis;
197 198 199
    for (size_t i = 0; i < op->axis.size(); ++i) {
      auto it = axis_remap_.find(op->axis[i]);
      if (it != axis_remap_.end()) {
200
        axis.push_back(it->second);
201 202
      }
    }
203 204
    return Reduce::make(
        op->combiner, op->source, axis, op->condition, op->value_index);
205 206 207 208
  }

  void Init(const ComputeOpNode* self,
            const Stage& stage,
209
            const std::unordered_map<IterVar, Range>& dom_map,
210 211
            const std::unordered_map<IterVar, Range>& out_dom,
            const std::unordered_map<Tensor, Array<Range> >& in_region,
212 213
            const TensorIntrin& intrin,
            Map<Var, Range>* compute_intrin_iter_space) {
214
    CHECK(self == stage->op.get());
215 216 217 218 219 220 221 222 223 224

    for (size_t i = 0; i < stage->leaf_iter_vars.size(); ++i) {
      IterVar iv = stage->leaf_iter_vars[i];
      auto vit = dom_map.find(iv);
      if (vit != dom_map.end()) {
        const Range vrange = vit->second;
        compute_intrin_iter_space->Set(iv->var, vrange);
      }
    }

225 226 227 228 229 230 231 232 233 234
    // input remap.
    Array<Tensor> inputs = self->InputTensors();
    CHECK_EQ(inputs.size(), intrin->inputs.size());
    for (size_t i = 0; i < inputs.size(); ++i) {
      InputEntry e;
      e.tensor = intrin->inputs[i];
      e.region = Array<Range>(in_region.at(inputs[i]));
      CHECK_GE(e.region.size(), e.tensor.ndim());
      // Enable fuzzy matching, to match [1, n, m] to [n, m]
      e.start = e.region.size() - e.tensor.ndim();
235 236 237
      for (size_t j = 0; j < e.start; ++j) {
        auto canonical_extent = Simplify(e.region[j]->extent, *compute_intrin_iter_space);
        CHECK(is_one(canonical_extent))
238 239
            << "Tensorize " << intrin->name << ":"
            << " Input dimension mismatch with tensor intrin "
240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257
            << " expected shape=" << e.tensor->shape
            << ", given region=" << e.region;
      }
      in_remap_[inputs[i]] = e;
    }
    // output remap
    const ComputeOpNode* intrin_compute = intrin->op.as<ComputeOpNode>();
    CHECK(intrin_compute) << "Only support compute intrinsic for now";
    CHECK_GE(self->axis.size(), intrin_compute->axis.size())
        << "Tensorize: Output mismatch with tensor intrin ";
    // Enable fuzzy matching, to match [1, n, m] to [n, m]
    size_t axis_start = self->axis.size() - intrin_compute->axis.size();
    for (size_t i = 0; i < axis_start; ++i) {
      Range r = out_dom.at(self->axis[i]);
      CHECK(is_one(r->extent))
          << "Tensorize: Output mismatch with tensor intrin "
          << " intrin-dim=" << intrin_compute->axis.size()
          << ", tensorize-dim=" << self->axis.size();
258
      var_remap_[self->axis[i]->var.get()] = r->min;
259 260 261 262 263 264 265 266 267 268
    }
    // Assume we tensorize at regin axis i [min, min + extent)
    // The corresponding intrinsic axis is j [0, extent)
    // Remap index i to j + min
    for (size_t i = axis_start; i < self->axis.size(); ++i) {
      IterVar iv = self->axis[i];
      IterVar target_iv = intrin_compute->axis[i - axis_start];
      Range r = out_dom.at(iv);
      var_remap_[iv->var.get()] = target_iv->var + r->min;
      axis_remap_[iv] = target_iv;
269
      compute_intrin_iter_space->Set(target_iv->var, target_iv->dom);
270 271 272 273 274 275 276 277 278 279 280
    }
    // Remap reduction axis
    CHECK_GE(self->reduce_axis.size(), intrin_compute->reduce_axis.size())
        << "Tensorize: Reduction dimension mismatch with tensor intrin";
    axis_start = self->reduce_axis.size() - intrin_compute->reduce_axis.size();
    for (size_t i = 0; i < axis_start; ++i) {
      Range r = out_dom.at(self->reduce_axis[i]);
      CHECK(is_one(r->extent))
          << "Tensorize: Reduction mismatch with tensor intrin "
          << " intrin-dim=" << intrin_compute->reduce_axis.size()
          << ", tensorize-dim=" << self->reduce_axis.size();
281
      var_remap_[self->reduce_axis[i]->var.get()] = r->min;
282 283 284 285 286 287 288
    }
    for (size_t i = axis_start; i < self->reduce_axis.size(); ++i) {
      IterVar iv = self->reduce_axis[i];
      IterVar target_iv = intrin_compute->reduce_axis[i - axis_start];
      Range r = out_dom.at(iv);
      var_remap_[iv->var.get()] = target_iv->var + r->min;
      axis_remap_[iv] = target_iv;
289
      compute_intrin_iter_space->Set(target_iv->var, target_iv->dom);
290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311
    }
  }

 private:
  // Input entry
  struct InputEntry {
    Tensor tensor;
    size_t start;
    Array<Range> region;
  };
  // input data remap
  std::unordered_map<Tensor, InputEntry> in_remap_;
  // variable remap.
  std::unordered_map<const Variable*, Expr> var_remap_;
  // IterVar remap.
  std::unordered_map<IterVar, IterVar> axis_remap_;
};

// Try to match tensor dataflow of the stage with the intrinsic
Array<Expr> MatchTensorizeBody(
    const ComputeOpNode* self,
    const Stage& stage,
312
    const std::unordered_map<IterVar, Range>& dom_map,
313 314
    const std::unordered_map<IterVar, Range>& out_dom,
    const std::unordered_map<Tensor, Array<Range> >& in_region,
315 316
    const TensorIntrin& intrin,
    Map<Var, Range>* compute_intrin_iter_space) {
317
  TensorIntrinMatcher matcher;
318
  matcher.Init(self, stage, dom_map, out_dom, in_region, intrin, compute_intrin_iter_space);
319 320 321 322 323 324 325 326 327 328
  Array<Expr> ret;
  for (Expr expr : self->body) {
    ret.push_back(matcher.Mutate(expr));
  }
  return ret;
}

void VerifyTensorizeBody(
    const ComputeOpNode* self,
    const Stage& stage,
329
    const std::unordered_map<IterVar, Range>& dom_map,
330 331 332
    const std::unordered_map<IterVar, Range>& out_dom,
    const std::unordered_map<Tensor, Array<Range> >& in_region,
    const TensorIntrin& intrin) {
333
  Map<Var, Range> compute_intrin_iter_space;
334
  Array<Expr> body = MatchTensorizeBody(self, stage, dom_map, out_dom, in_region, intrin,
335
                                        &compute_intrin_iter_space);
336 337 338 339 340
  const ComputeOpNode* intrin_compute = intrin->op.as<ComputeOpNode>();
  CHECK(intrin_compute) << "Only support compute intrinsic for now";
  CHECK_EQ(body.size(), intrin_compute->body.size())
      << "Tensorize failed: body size mismatch";
  for (size_t i = 0; i < body.size(); ++i) {
341 342 343 344
    Expr lhs = Simplify(body[i], compute_intrin_iter_space);
    lhs = CanonicalSimplify(lhs, compute_intrin_iter_space);
    Expr rhs = Simplify(intrin_compute->body[i], compute_intrin_iter_space);
    rhs = CanonicalSimplify(rhs, compute_intrin_iter_space);
345 346 347 348 349 350 351
    if (lhs.type() != rhs.type()) {
      LOG(FATAL)
          << "Failed to match the data type with TensorIntrin "
          << intrin->name << "'s declaration "
          << " provided=" << lhs.type()
          << ", intrin=" << rhs.type();
    }
352
    CHECK(Equal(lhs, rhs))
353 354 355 356
        << "Failed to match the compute with TensorIntrin "
        << intrin->name << "'s declaration "
        << " provided= " << lhs
        << ", intrin=  " << rhs;
357 358 359 360 361
  }
}

Stmt MakeTensorize(const ComputeOpNode* self,
                   const Stage& stage,
362
                   const std::unordered_map<IterVar, Range>& dom_map,
363
                   bool debug_keep_trivial_loop) {
364 365 366 367 368 369
  std::unordered_map<IterVar, Range> out_dom;
  std::unordered_map<Tensor, Array<Range> > in_region;
  size_t tloc = InferTensorizeRegion(self, stage, dom_map, &out_dom, &in_region);
  TensorIntrin intrin = stage->iter_var_attrs.at(
      stage->leaf_iter_vars[tloc])->tensor_intrin;
  CHECK(intrin.defined());
370
  ComputeLoopNest n = ComputeLoopNest::make(self, stage, dom_map, debug_keep_trivial_loop);
371
  VerifyTensorizeLoopNest(self, stage, n, tloc);
372
  VerifyTensorizeBody(self, stage, dom_map, out_dom, in_region, intrin);
373 374
  // Start bind data.
  Stmt nop = Evaluate::make(0);
375
  std::vector<Stmt> input_bind_nest, output_bind_nest;
376 377 378 379
  Array<Tensor> inputs = self->InputTensors();
  CHECK_EQ(inputs.size(), intrin->inputs.size())
      << "Tensorize failed: input size mismatch ";
  // input binding
380
  for (size_t i = 0; i < intrin->inputs.size(); ++i) {
381 382 383 384 385 386 387 388 389 390 391
    Tensor tensor = inputs[i];
    Buffer buffer = intrin->buffers[i];
    Array<NodeRef> bind_spec{buffer, tensor};
    auto it = in_region.find(tensor);
    CHECK(it != in_region.end());
    const Array<Range>& region = it->second;
    Array<Expr> tuple;
    for (const Range r : region) {
      tuple.push_back(r->min);
      tuple.push_back(r->extent);
    }
392
    input_bind_nest.emplace_back(AttrStmt::make(
393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411
        bind_spec, ir::attr::buffer_bind_scope,
        Call::make(Handle(), ir::intrinsic::tvm_tuple, tuple, Call::Intrinsic), nop));
  }
  // output binding
  const ComputeOpNode* intrin_compute = intrin->op.as<ComputeOpNode>();
  CHECK(intrin_compute) << "Only support compute intrinsic for now";
  CHECK_EQ(intrin->inputs.size() + intrin_compute->body.size(), intrin->buffers.size());
  CHECK_EQ(intrin_compute->body.size(), self->body.size());
  Array<Expr> tuple;
  for (IterVar iv : self->axis) {
    auto it = out_dom.find(iv);
    CHECK(it != out_dom.end());
    tuple.push_back(it->second->min);
    tuple.push_back(it->second->extent);
  }
  for (size_t i = intrin->inputs.size(); i < intrin->buffers.size(); ++i) {
    Tensor tensor = stage->op.output(i - intrin->inputs.size());
    Buffer buffer = intrin->buffers[i];
    Array<NodeRef> bind_spec{buffer, tensor};
412
    output_bind_nest.emplace_back(AttrStmt::make(
413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446
        bind_spec, ir::attr::buffer_bind_scope,
        Call::make(Handle(), ir::intrinsic::tvm_tuple, tuple, Call::Intrinsic), nop));
  }
  // Check variable remap
  std::unordered_map<const Variable*, Expr> vmap;
  ir::ArgBinder binder(&vmap);
  CHECK_GE(self->reduce_axis.size(), intrin_compute->reduce_axis.size())
      << "Tensorization fail: reduction axis size do not match";
  size_t start = self->reduce_axis.size() - intrin_compute->reduce_axis.size();
  for (size_t i = 0; i < start; ++i) {
    IterVar iv = self->reduce_axis[i];
    auto it = out_dom.find(iv);
    CHECK(it != out_dom.end());
    CHECK(is_one(it->second->extent))
        << "Tensorization fail: reduction axis size do not match";
  }
  for (size_t i = start; i < self->reduce_axis.size(); ++i) {
    IterVar iv = self->reduce_axis[i];
    IterVar target = intrin_compute->reduce_axis[i - start];
    auto it = out_dom.find(iv);
    CHECK(it != out_dom.end());
    binder.Bind(target->dom->min, make_const(iv->dom->min.type(), 0),
                "tensir_intrin.reduction.min");
    binder.Bind(target->dom->extent, it->second->extent,
                "tensir_intrin.reduction.extent");
  }
  if (tloc <= n.num_common_loop) {
    // Do no need to split reduction
    std::vector<std::vector<Stmt> > nest(
        n.main_nest.begin(), n.main_nest.begin() + tloc + 1);
    nest.emplace_back(op::MakeIfNest(n.main_predicates));
    CHECK_EQ(n.init_predicates.size(), 0U);
    CHECK(intrin->body.defined())
        << "Normal store op for intrin " << intrin << " is not defined";
447 448
    Stmt body = MergeNest(output_bind_nest, intrin->body);
    body = MergeNest(input_bind_nest, body);
449
    body = Substitute(body, vmap);
450
    body = MergeNest(binder.asserts(), body);
451
    body = Substitute(body, n.main_vmap);
452
    return MergeNest(nest, body);
453 454 455 456 457 458 459 460 461 462 463
  } else {
    // Need to split reduction
    CHECK(intrin->reduce_update.defined())
        << "Reduction update op for intrin " << intrin << " is not defined";
    // Need init and update steps
    CHECK_NE(self->reduce_axis.size(), 0U);
    std::vector<std::vector<Stmt> > common(
        n.main_nest.begin(), n.main_nest.begin() + n.num_common_loop + 1);
    std::vector<std::vector<Stmt> > update_nest(
        n.main_nest.begin() + n.num_common_loop + 1, n.main_nest.begin() + tloc + 1);
    update_nest.emplace_back(op::MakeIfNest(n.main_predicates));
464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495

    if (intrin->reduce_init.defined()) {
      // init nest
      std::vector<std::vector<Stmt> > init_nest(
          n.init_nest.begin(), n.init_nest.begin() + tloc + 1);
      init_nest.emplace_back(op::MakeIfNest(n.init_predicates));
      Stmt init = MergeNest(output_bind_nest, intrin->reduce_init);
      init = Substitute(init, n.init_vmap);
      init = MergeNest(init_nest, init);
      // The update
      Stmt update = MergeNest(output_bind_nest, intrin->reduce_update);
      update = MergeNest(input_bind_nest, update);
      update = Substitute(update, vmap);
      update = MergeNest(binder.asserts(), update);
      update = Substitute(update, n.main_vmap);
      update = MergeNest(update_nest, update);
      return MergeNest(common, Block::make(init, update));
    } else {
      // When init op is not available, use body op for reset in the first iter.
      CHECK(intrin->body.defined())
          << "Normal body op for intrin " << intrin << " is not defined";
      Stmt update = TransformUpdate(stage, dom_map, n,
                                    intrin->body,
                                    intrin->reduce_update);
      update = MergeNest(output_bind_nest, update);
      update = MergeNest(input_bind_nest, update);
      update = Substitute(update, vmap);
      update = MergeNest(binder.asserts(), update);
      update = Substitute(update, n.main_vmap);
      update = MergeNest(update_nest, update);
      return MergeNest(common, update);
    }
496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520
  }
}

// Register functions for unittests
TVM_REGISTER_API("test.op.InferTensorizeRegion")
.set_body([](TVMArgs args, TVMRetValue* ret) {
    Stage stage = args[0];
    Map<IterVar, Range> dmap = args[1];
    std::unordered_map<IterVar, Range> out_dom;
    std::unordered_map<Tensor, Array<Range> > in_region;
    CHECK(stage->op.as<ComputeOpNode>());
    InferTensorizeRegion(stage->op.as<ComputeOpNode>(),
                         stage,
                         as_unordered_map(dmap),
                         &out_dom, &in_region);
    *ret = Array<NodeRef>{Map<IterVar, Range>(out_dom),
                          Map<Tensor, Array<Range> >(in_region)};
  });

TVM_REGISTER_API("test.op.MatchTensorizeBody")
.set_body([](TVMArgs args, TVMRetValue* ret) {
    Stage stage = args[0];
    Map<IterVar, Range> out_dom = args[1];
    Map<Tensor, Array<Range> > in_region = args[2];
    TensorIntrin intrin = args[3];
521
    Map<Var, Range> vrange;
522 523 524
    CHECK(stage->op.as<ComputeOpNode>());
    *ret = MatchTensorizeBody(stage->op.as<ComputeOpNode>(),
                              stage,
525
                              {{}},
526 527
                              as_unordered_map(out_dom),
                              as_unordered_map(in_region),
528 529
                              intrin,
                              &vrange);
530 531
  });
}  // namespace tvm