loop_vectorization_utils.h 28 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
/*
 * 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
 *
 *   http://www.apache.org/licenses/LICENSE-2.0
 *
 * 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.
 */

/*!
 * \file common.h
 * \brief Common utilities for TL transforms
 */

#include <tvm/tir/builtin.h>
#include <tvm/tir/op.h>
#include <tvm/tir/stmt_functor.h>
#include <tvm/tir/transform.h>
#include <tvm/tir/utils.h>

#include <queue>

#include "../../op/parallel.h"
#include "../loop_partition.h"
#include "../loop_vectorize.h"
36
#include "arith/ir_mutator_with_analyzer.h"
37
38
39
40
41
42
43
44
45
46

namespace tvm {
namespace tl {

using namespace tir;

// Vectorize Part
// Use the same code as tir.transform.vectorize_loop
inline PrimExpr CreateNewLanes(bool is_scalable, int lanes_or_vscale_factor) {
  if (is_scalable) {
47
48
    return Mul(Call(DataType::Int(32), builtin::vscale(), {}),
               lanes_or_vscale_factor);
49
50
51
52
53
54
55
56
57
58
59
  } else {
    return lanes_or_vscale_factor;
  }
}

inline PrimExpr BroadcastTo(PrimExpr e, int lanes, bool is_scalable) {
  // Check if e is already in the expected form
  if (e.dtype().get_lanes_or_vscale_factor() == lanes &&
      e.dtype().is_scalable_vector() == is_scalable)
    return e;

60
  if (const BroadcastNode *op = e.as<BroadcastNode>()) {
61
62
63
64
65
66
67
68
69
    ICHECK(op->dtype.is_scalable_vector() == is_scalable)
        << "Can't broadcast between scalable and fixed length vectors.";
    int e_lanes = op->dtype.get_lanes_or_vscale_factor();

    if (lanes % e_lanes == 0) {
      return Broadcast(op->value, CreateNewLanes(is_scalable, lanes));
    }
  }

70
71
72
  ICHECK(e.dtype().is_scalar())
      << "Cannot broadcast lanes=" << e.dtype().get_lanes_or_vscale_factor()
      << " is_scalable=" << e.dtype().is_scalable_vector() << " to " << lanes;
73
74
75
76
77

  return Broadcast(e, CreateNewLanes(is_scalable, lanes));
}

// Rewrite vectorized allocation access
78
79
// This is necessary for making each vector component containing its own
// workspace. Originates from Halide's loop vectorizer
80
81
82
//
// s[i] = s[i * lanes + var]
//
83
84
// The same principle applies when using one thread to simulate multiple
// context.
85
86
//
class VecAllocAccess : public StmtExprMutator {
87
88
public:
  VecAllocAccess(const VarNode *buf, Var var, PrimExpr var_lanes)
89
90
      : buf_(buf), var_(var), var_lanes_(var_lanes) {}

91
  PrimExpr VisitExpr_(const BufferLoadNode *op) final {
92
93
94
95
    auto load = Downcast<BufferLoad>(StmtExprMutator::VisitExpr_(op));
    return UpdateBufferAccess(load);
  }

96
  Stmt VisitStmt_(const BufferStoreNode *op) final {
97
98
99
100
    auto store = Downcast<BufferStore>(StmtExprMutator::VisitStmt_(op));
    return UpdateBufferAccess(store);
  }

101
102
private:
  template <typename Node> Node UpdateBufferAccess(Node node) {
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
    // Only update the buffer that's being replaced.
    if (node->buffer->data.get() != buf_) {
      return node;
    }

    // Find/make a Buffer object with the correct updated shape.
    Buffer buf;
    auto it = buffer_map_.find(node->buffer.get());
    if (it != buffer_map_.end()) {
      buf = it->second;
    } else {
      // Extend the least significant dimension by a factor of
      // var_lanes_.  Typically, this will be a 1-d index into a flat
      // memory space.
      Array<PrimExpr> shape = node->buffer->shape;
118
119
      shape.Set(shape.size() - 1,
                analyzer_.Simplify(shape[shape.size() - 1] * var_lanes_));
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

      // TODO(Lunderberg): Move this pass to be prior to
      // StorageFlatten/FlattenBuffer, implement by appending a
      // dimension to the buffer.  Since it is currently after the
      // flattening, the strides are not technically necessary, but
      // are updated for consistency.

      // Update strides if defined.
      Array<PrimExpr> strides;
      for (size_t i = 0; i < strides.size(); i++) {
        PrimExpr stride = strides[i];
        if (i != strides.size() - 1) {
          stride *= var_lanes_;
        }
        strides.push_back(analyzer_.Simplify(stride));
      }

      // Copy everything into the new buffer.
      buf = node->buffer;
      auto buf_writer = buf.CopyOnWrite();
      buf_writer->shape = shape;
      buf_writer->strides = strides;
      buffer_map_[buf.get()] = buf;
    }

    // Extend the last index by the number of lanes in the vectorized
    // variable.
    Array<PrimExpr> indices = node->indices;
148
149
150
    indices.Set(
        indices.size() - 1,
        analyzer_.Simplify(indices[indices.size() - 1] * var_lanes_ + var_));
151
152
153
154
155
156
157
158

    auto writer = node.CopyOnWrite();
    writer->buffer = buf;
    writer->indices = indices;
    return node;
  }

  // buffer var
159
  const VarNode *buf_;
160
  // Updated buffer objects.
161
  std::unordered_map<const BufferNode *, Buffer> buffer_map_;
162
163
164
165
166
167
168
169
170
171
172
  // variable to be replaced
  Var var_;
  // the lanes.
  PrimExpr var_lanes_;
  // Analyzer for simplifications
  arith::Analyzer analyzer_;
};

// We use ExprFunctor directly instead of StmtExprMutator
// This is because the transformation can change the dtype of the Expr
// The existing ExprMutator transformation rules may not be well defined.
173
174
175
class Vectorizer : public StmtMutator,
                   public ExprFunctor<PrimExpr(const PrimExpr &)> {
public:
176
177
178
179
180
181
182
  using ExprFunctor::VisitExpr;
  using StmtMutator::operator();

  Vectorizer(Var var, PrimExpr var_lanes) : var_(var), var_lanes_(var_lanes) {
    ramp_ = Ramp(IntImm(var->dtype, 0), IntImm(var->dtype, 1), var_lanes);
  }

183
  Stmt VisitStmt(const Stmt &stmt) final {
184
185
186
187
188
189
190
191
192
193
    ICHECK(!need_scalarize_);
    Stmt ret = StmtMutator::VisitStmt(stmt);
    if (need_scalarize_) {
      need_scalarize_ = false;
      return Scalarize(stmt);
    } else {
      return ret;
    }
  }

194
195
196
  PrimExpr VisitExpr(const PrimExpr &e) final {
    return ExprFunctor::VisitExpr(e);
  }
197

198
  PrimExpr VisitExpr_(const AddNode *op) final {
199
200
201
    return AddSubVec(op, [](PrimExpr a, PrimExpr b) { return a + b; });
  }

202
  PrimExpr VisitExpr_(const SubNode *op) final {
203
204
205
    return AddSubVec(op, [](PrimExpr a, PrimExpr b) { return a - b; });
  }

206
  PrimExpr VisitExpr_(const MulNode *op) final {
207
208
209
210
211
212
213
214
215
216
    PrimExpr a = this->VisitExpr(op->a);
    PrimExpr b = this->VisitExpr(op->b);
    if (a.same_as(op->a) && b.same_as(op->b)) {
      return GetRef<PrimExpr>(op);
    } else {
      bool is_vec_a = a.dtype().is_scalable_or_fixed_length_vector();
      bool is_vec_b = b.dtype().is_scalable_or_fixed_length_vector();
      if (is_vec_a && is_vec_b) {
        // Let's not multiply scalable and fixed length vectors
        ICHECK(a.dtype().is_scalable_vector() == b.dtype().is_scalable_vector())
217
218
            << "Fixed length and scalable vectors can't be mixed in "
               "multiplication.";
219
220
      }
      if (is_vec_a || is_vec_b) {
221
222
        const RampNode *b_ramp = b.as<RampNode>();
        const RampNode *a_ramp = a.as<RampNode>();
223
224
225
226
227
228
229
230
231
232
233
        if (a_ramp && b.dtype().is_scalar() && analyzer_.CanProve(b > 0)) {
          PrimExpr lanes = a_ramp->lanes;
          return Ramp(a_ramp->base * b, a_ramp->stride * b, lanes);
        }
        if (b_ramp && a.dtype().is_scalar() && analyzer_.CanProve(a > 0)) {
          PrimExpr lanes = b_ramp->lanes;
          return Ramp(b_ramp->base * a, b_ramp->stride * a, lanes);
        }
        int a_lanes = a.dtype().get_lanes_or_vscale_factor();
        int b_lanes = b.dtype().get_lanes_or_vscale_factor();
        int max_lanes = std::max(a_lanes, b_lanes);
234
235
236
237
        bool is_scalable =
            a.dtype().is_scalable_vector() || b.dtype().is_scalable_vector();
        return Mul(BroadcastTo(a, max_lanes, is_scalable),
                   BroadcastTo(b, max_lanes, is_scalable));
238
239
240
241
      }
    }
    return BinaryVec<Mul>(op);
  }
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
  PrimExpr VisitExpr_(const DivNode *op) final { return BinaryVec<Div>(op); }
  PrimExpr VisitExpr_(const ModNode *op) final { return BinaryVec<Mod>(op); }
  PrimExpr VisitExpr_(const FloorDivNode *op) final {
    return BinaryVec<FloorDiv>(op);
  }
  PrimExpr VisitExpr_(const FloorModNode *op) final {
    return BinaryVec<FloorMod>(op);
  }
  PrimExpr VisitExpr_(const MinNode *op) final { return BinaryVec<Min>(op); }
  PrimExpr VisitExpr_(const MaxNode *op) final { return BinaryVec<Max>(op); }
  PrimExpr VisitExpr_(const EQNode *op) final { return BinaryVec<EQ>(op); }
  PrimExpr VisitExpr_(const NENode *op) final { return BinaryVec<NE>(op); }
  PrimExpr VisitExpr_(const LTNode *op) final { return BinaryVec<LT>(op); }
  PrimExpr VisitExpr_(const LENode *op) final { return BinaryVec<LE>(op); }
  PrimExpr VisitExpr_(const GTNode *op) final { return BinaryVec<GT>(op); }
  PrimExpr VisitExpr_(const GENode *op) final { return BinaryVec<GE>(op); }
  PrimExpr VisitExpr_(const AndNode *op) final { return BinaryVec<And>(op); }
  PrimExpr VisitExpr_(const OrNode *op) final { return BinaryVec<Or>(op); }

  PrimExpr VisitExpr_(const NotNode *op) final {
262
263
264
265
266
267
268
269
    PrimExpr a = this->VisitExpr(op->a);
    if (a.same_as(op->a)) {
      return GetRef<PrimExpr>(op);
    } else {
      return !(a);
    }
  }

270
  PrimExpr VisitExpr_(const RampNode *op) final {
271
272
273
274
275
276
277
278
279
    PrimExpr base = this->VisitExpr(op->base);
    PrimExpr stride = this->VisitExpr(op->stride);
    ICHECK(!base.dtype().is_scalable_vector())
        << "Creating scalable vectors from existing vectors is not supported.";
    ICHECK(!stride.dtype().is_scalable_vector())
        << "Ramp stride with scalable dtype is not supported";
    if (base.dtype().is_fixed_length_vector() && stride.dtype().is_scalar()) {
      ICHECK(op->lanes->IsInstance<IntImmNode>())
          << "Vectorizing over existing scalable vectors is not supported.";
280
      const RampNode *base_ramp = base.as<RampNode>();
281
      int op_lanes = static_cast<int>(Downcast<IntImm>(op->lanes)->value);
282
283
      int base_ramp_lanes =
          static_cast<int>(Downcast<IntImm>(base_ramp->lanes)->value);
284
      if (analyzer_.CanProve(base_ramp->stride ==
285
286
                             stride *
                                 make_const(stride.dtype(), base_ramp_lanes))) {
287
288
289
290
291
292
293
294
        return Ramp(base_ramp->base, stride, op_lanes * base_ramp_lanes);
      }
    }
    int lanes = std::max(base.dtype().lanes(), stride.dtype().lanes());
    base = BroadcastTo(base, lanes, false);
    stride = BroadcastTo(stride, lanes, false);
    Array<PrimExpr> elems;
    for (int i = 0; i < lanes; ++i) {
295
296
      elems.push_back(Ramp(Shuffle::ExtractElement(base, i),
                           Shuffle::ExtractElement(stride, i), op->lanes));
297
298
299
300
    }
    return Shuffle::Concat(elems);
  }

301
  PrimExpr VisitExpr_(const BroadcastNode *op) final {
302
303
304
305
306
307
308
309
310
311
312
313
    PrimExpr value = this->VisitExpr(op->value);
    if (value.dtype().is_scalable_or_fixed_length_vector()) {
      need_scalarize_ = true;
      return GetRef<PrimExpr>(op);
    }
    if (value.same_as(op->value)) {
      return GetRef<PrimExpr>(op);
    } else {
      return Broadcast(op->value, op->lanes);
    }
  }

314
  PrimExpr VisitExpr_(const SelectNode *op) final {
315
316
317
    PrimExpr cond = this->VisitExpr(op->condition);
    PrimExpr t = this->VisitExpr(op->true_value);
    PrimExpr f = this->VisitExpr(op->false_value);
318
319
    if (cond.same_as(op->condition) && t.same_as(op->true_value) &&
        f.same_as(op->false_value)) {
320
321
322
323
324
325
      return GetRef<PrimExpr>(op);
    } else {
      int cond_lanes = cond.dtype().get_lanes_or_vscale_factor();
      int t_lanes = t.dtype().get_lanes_or_vscale_factor();
      int f_lanes = f.dtype().get_lanes_or_vscale_factor();
      int lanes = std::max(std::max(cond_lanes, t_lanes), f_lanes);
326
327
      bool is_scalable = cond.dtype().is_scalable_vector() ||
                         t.dtype().is_scalable_vector() ||
328
                         f.dtype().is_scalable_vector();
329
330
      return Select(BroadcastTo(cond, lanes, is_scalable),
                    BroadcastTo(t, lanes, is_scalable),
331
332
333
334
                    BroadcastTo(f, lanes, is_scalable));
    }
  }

335
  PrimExpr VisitExpr_(const CastNode *op) final {
336
337
338
339
340
    PrimExpr value = this->VisitExpr(op->value);
    if (value.same_as(op->value)) {
      return GetRef<PrimExpr>(op);
    } else {
      if (value.dtype().is_scalable_vector()) {
341
342
343
        return Cast(op->dtype.with_scalable_vscale_factor(
                        value.dtype().vscale_factor()),
                    value);
344
345
346
347
348
349
      } else {
        return Cast(op->dtype.with_lanes(value.dtype().lanes()), value);
      }
    }
  }

350
351
352
  PrimExpr VisitExpr_(const FloatImmNode *op) final {
    return GetRef<PrimExpr>(op);
  }
353

354
355
356
  PrimExpr VisitExpr_(const IntImmNode *op) final {
    return GetRef<PrimExpr>(op);
  }
357

358
359
360
  PrimExpr VisitExpr_(const StringImmNode *op) final {
    return GetRef<PrimExpr>(op);
  }
361
362

  // Variable
363
  PrimExpr VisitExpr_(const VarNode *op) final {
364
365
366
367
368
369
370
371
372
373
374
375
376
    Var var = GetRef<Var>(op);

    if (var.same_as(var_)) {
      return ramp_;
    }
    auto it = let_binding_.find(var);
    if (it != let_binding_.end()) {
      return it->second;
    } else {
      return std::move(var);
    }
  }
  // IfThenElse expr
377
  PrimExpr MutateIfThenElseExpr_(const CallNode *op) {
378
379
380
381
382
383
384
    PrimExpr cond = this->VisitExpr(op->args[0]);
    if (cond.dtype().is_scalable_or_fixed_length_vector()) {
      need_scalarize_ = true;
      return GetRef<PrimExpr>(op);
    }
    PrimExpr t = this->VisitExpr(op->args[1]);
    PrimExpr f = this->VisitExpr(op->args[2]);
385
386
    if (cond.same_as(op->args[0]) && t.same_as(op->args[1]) &&
        f.same_as(op->args[2])) {
387
388
389
390
391
      return GetRef<PrimExpr>(op);
    } else {
      int t_lanes = t.dtype().get_lanes_or_vscale_factor();
      int f_lanes = f.dtype().get_lanes_or_vscale_factor();
      int lanes = std::max(t_lanes, f_lanes);
392
393
      bool is_scalable =
          t.dtype().is_scalable_vector() || f.dtype().is_scalable_vector();
394
395
396
      t = BroadcastTo(t, lanes, is_scalable);
      f = BroadcastTo(f, lanes, is_scalable);
      if (is_scalable) {
397
398
        return Call(op->dtype.with_scalable_vscale_factor(lanes), op->op,
                    {cond, t, f});
399
400
401
402
403
404
      } else {
        return Call(op->dtype.with_lanes(lanes), op->op, {cond, t, f});
      }
    }
  }
  // Reinterpret expr
405
  PrimExpr MutateReinterpretExpr_(const CallNode *op) {
406
407
408
409
410
411
412
    ICHECK(op->op.same_as(builtin::reinterpret()));
    PrimExpr value = this->VisitExpr(op->args[0]);
    if (value.same_as(op->args[0])) {
      return GetRef<PrimExpr>(op);
    } else {
      int lanes = value.dtype().get_lanes_or_vscale_factor();
      if (value.dtype().is_scalable_vector()) {
413
414
        return Call(op->dtype.with_scalable_vscale_factor(lanes), op->op,
                    {value});
415
416
417
418
419
420
      } else {
        return Call(op->dtype.with_lanes(lanes), op->op, {value});
      }
    }
  }
  // Call
421
  PrimExpr VisitExpr_(const CallNode *op) final {
422
423
424
425
426
427
428
429
430
431
432
433
434
435
    if (op->op.same_as(builtin::if_then_else())) {
      return MutateIfThenElseExpr_(op);
    } else if (op->op.same_as(builtin::texture2d_load())) {
      int lane = 0;
      Array<PrimExpr> fcd = MutateArray({op->args.back()}, &lane);
      auto new_args = op->args;
      new_args.pop_back();
      new_args.push_back(fcd[0]);
      return Call(op->dtype.with_lanes(4), op->op, new_args);
    } else if (op->op.same_as(builtin::texture2d_store())) {
      int lane = 0;
      // Vectorize the value to store
      Array<PrimExpr> value{op->args.back()};
      Array<PrimExpr> mutated_value = MutateArray(value, &lane);
436
437
      Array<PrimExpr> new_args{op->args[0], op->args[1], op->args[2],
                               mutated_value[0]};
438
439
440
441
442
      return Call(op->dtype.with_lanes(lane), op->op, new_args);
    } else if (op->op.same_as(builtin::reinterpret())) {
      return MutateReinterpretExpr_(op);
    }
    auto optional_op = op->op.as<Op>();
443
444
    bool vectorizable = optional_op &&
                        op_vectorizable_.get(optional_op.value(), false) &&
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
                        !op->dtype.is_scalable_vector();

    if (!vectorizable) {
      // Cannot vectorize this op
      Array<PrimExpr> new_args;
      for (auto arg : op->args) {
        auto new_arg = this->VisitExpr(arg);
        if (new_arg.dtype().is_scalable_or_fixed_length_vector()) {
          need_scalarize_ = true;
          return GetRef<PrimExpr>(op);
        }
        new_args.push_back(new_arg);
      }
      if (op->args.same_as(new_args)) {
        return GetRef<PrimExpr>(op);
      } else {
        return Call(op->dtype, op->op, new_args);
      }
    } else {
      int lane = 0;
      Array<PrimExpr> new_args = MutateArray(op->args, &lane);
      // normal code path.
      if (op->args.same_as(new_args)) {
        return GetRef<PrimExpr>(op);
      } else {
        return Call(op->dtype.with_lanes(lane), op->op, new_args);
      }
    }
  }
  // BufferLoad
475
  PrimExpr VisitExpr_(const BufferLoadNode *op) final {
476
477
    auto load = GetRef<BufferLoad>(op);

478
479
480
    auto fmutate = [this](const PrimExpr &index) {
      return this->VisitExpr(index);
    };
481
482
483
484
485
486
487
488
489
490
    Array<PrimExpr> indices = op->indices.Map(fmutate);

    if (!indices.same_as(op->indices)) {
      auto writer = load.CopyOnWrite();
      writer->indices = indices;
    }

    return std::move(load);
  }
  // Let
491
  PrimExpr VisitExpr_(const LetNode *op) final {
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
    PrimExpr value = this->VisitExpr(op->value);
    // Weaker SSA condition
    // A single var can be binded in multiple lets
    // but they have to bind to the same value.
    // This is used to allow cases when we reuse a single let
    // expression to construct a nested expr.
    // (let x = 1 in x + 1) * (let x = 1 in x + 1)
    auto it = let_binding_.find(op->var);
    if (it != let_binding_.end()) {
      ICHECK(deep_equal_(it->second, value))
          << "Let cannot bind the same var to two different values";
    }
    if (value.dtype().get_lanes_or_vscale_factor() !=
        op->value.dtype().get_lanes_or_vscale_factor()) {
      Var new_var(op->var->name_hint, value.dtype());
      let_binding_[op->var] = new_var;
      return Let(new_var, value, this->VisitExpr(op->body));
    } else {
      let_binding_[op->var] = op->var;
      PrimExpr body = this->VisitExpr(op->body);
      if (value.same_as(op->value) && body.same_as(op->body)) {
        return GetRef<PrimExpr>(op);
      } else {
        return Let(op->var, value, body);
      }
    }
  }
  // BufferStore
520
  Stmt VisitStmt_(const BufferStoreNode *op) final {
521
522
    auto store = GetRef<BufferStore>(op);

523
524
525
    auto fmutate = [this](const PrimExpr &index) {
      return this->VisitExpr(index);
    };
526
527
528
529
530
531
    Array<PrimExpr> indices = op->indices.Map(fmutate);

    PrimExpr value = this->VisitExpr(op->value);

    if (!indices.same_as(op->indices) || !value.same_as(op->value)) {
      ICHECK(!op->buffer->dtype.is_scalable_vector())
532
533
          << "Vectorizing over scalable buffer elements is not supported in "
             "vectorizer.";
534
535
536
537
538
539
      // How many lanes of indexing are present in the index and
      // buffer element type, excluding the last index.
      int other_index_lanes = op->buffer->dtype.lanes();
      for (size_t i = 0; i < indices.size() - 1; i++) {
        other_index_lanes *= indices[i].dtype().lanes();
        // Only allow the last index to be scalable
540
541
        ICHECK(!indices[i].dtype().is_scalable_vector())
            << "Only the last index can be scalable.";
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
      }

      // The total number of lanes of indexing, including the last index.
      auto last_index_dtype = indices[indices.size() - 1].dtype();
      int lanes_in_last_index = last_index_dtype.get_lanes_or_vscale_factor();
      int index_lanes = other_index_lanes * lanes_in_last_index;

      // The total number of lanes in this store operation.  Either
      // the index or the value will be broadcast out to this number
      // of lanes, depending on which has more lanes.
      int value_dtype_lanes = value.dtype().get_lanes_or_vscale_factor();
      bool is_last_index_scalable = last_index_dtype.is_scalable_vector();
      int total_lanes = std::max(index_lanes, value_dtype_lanes);

      ICHECK_EQ(total_lanes % other_index_lanes, 0)
557
558
          << "When storing to buffer " << op->buffer->name
          << ", cannot produce " << total_lanes
559
560
561
562
563
          << " lanes of storage location by changing the last index.";
      int last_index_lanes = total_lanes / other_index_lanes;

      // Broadcast the last index such that the total number of index
      // lanes matches the desired number.
564
565
566
      indices.Set(indices.size() - 1,
                  BroadcastTo(indices[indices.size() - 1], last_index_lanes,
                              is_last_index_scalable));
567
568
569
570
571
572
573
574
575

      auto writer = store.CopyOnWrite();
      writer->indices = indices;
      writer->value = BroadcastTo(value, total_lanes, is_last_index_scalable);
    }

    return std::move(store);
  }
  // For
576
  Stmt VisitStmt_(const ForNode *op) final {
577
578
579
580
581
582
583
584
585
586
587
588
589
    if (op->kind == ForKind::kVectorized) {
      LOG(WARNING) << "Detect vectorize inside vectorized loop, ignoring...";
    }
    ICHECK(is_zero(op->min));
    ICHECK(!op->extent.dtype().is_scalable_or_fixed_length_vector());
    PrimExpr extent = this->VisitExpr(op->extent);
    if (extent.dtype().is_scalable_or_fixed_length_vector()) {
      return Scalarize(GetRef<Stmt>(op));
    }
    Stmt body = this->VisitStmt(op->body);
    if (extent.same_as(op->extent) && body.same_as(op->body)) {
      return GetRef<Stmt>(op);
    } else {
590
591
      return For(op->loop_var, op->min, extent, op->kind, body,
                 op->thread_binding, op->annotations);
592
593
594
    }
  }
  // IfThenElse
595
  Stmt VisitStmt_(const IfThenElseNode *op) final {
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
    ICHECK(!op->condition.dtype().is_scalable_or_fixed_length_vector());
    PrimExpr condition = this->VisitExpr(op->condition);
    if (condition.dtype().is_scalable_or_fixed_length_vector()) {
      return Scalarize(GetRef<Stmt>(op));
    }
    Stmt then_case = this->VisitStmt(op->then_case);
    Optional<Stmt> else_case = NullOpt;
    if (op->else_case) {
      else_case = this->VisitStmt(op->else_case.value());
    }
    if (condition.same_as(op->condition) && then_case.same_as(op->then_case) &&
        else_case.same_as(op->else_case)) {
      return GetRef<Stmt>(op);
    } else {
      return IfThenElse(condition, then_case, else_case);
    }
  }
  // While
614
  Stmt VisitStmt_(const WhileNode *op) final {
615
616
617
    LOG(FATAL) << "A while loop inside a vectorized loop not supported.";
  }
  // LetStmt
618
  Stmt VisitStmt_(const LetStmtNode *op) final {
619
    PrimExpr value = this->VisitExpr(op->value);
620
621
    ICHECK(!let_binding_.count(op->var))
        << "SSA violation, a single var is binded twice";
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
    let_binding_[op->var] = value;

    if (value.dtype().get_lanes_or_vscale_factor() !=
        op->value.dtype().get_lanes_or_vscale_factor()) {
      Var new_var(op->var->name_hint, value.dtype());
      let_binding_[op->var] = new_var;
      return LetStmt(new_var, value, this->VisitStmt(op->body));
    } else {
      let_binding_[op->var] = op->var;
      Stmt body = this->VisitStmt(op->body);
      if (value.same_as(op->value) && body.same_as(op->body)) {
        return GetRef<Stmt>(op);
      } else {
        return LetStmt(op->var, value, body);
      }
    }
  }
  // Allocate
640
  Stmt VisitStmt_(const AllocateNode *op) final {
641
642
643
    // Mutate the condition
    PrimExpr condition = this->VisitExpr(op->condition);
    if (condition.dtype().is_scalable_or_fixed_length_vector()) {
644
645
      LOG(WARNING) << "Cannot handle vector extent in alloc of "
                   << op->buffer_var->name_hint;
646
647
648
649
650
      return Scalarize(GetRef<Stmt>(op));
    }

    // Mutate the extents
    Array<PrimExpr> extents;
651
    for (const auto &extent : op->extents) {
652
653
      PrimExpr new_ext = this->VisitExpr(extent);
      if (new_ext.dtype().is_scalable_or_fixed_length_vector()) {
654
655
        LOG(WARNING) << "Cannot handle vector extent in alloc of "
                     << op->buffer_var->name_hint;
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
        return Scalarize(GetRef<Stmt>(op));
      }
      extents.push_back(new_ext);
    }

    // TODO(Lunderberg): Move this pass to be prior to
    // StorageFlatten/FlattenBuffer.  That will allow this pass to be
    // implemented as adding a new buffer dimension, which is later
    // flattened.

    // Extend the least significant dimension by a factor of
    // var_lanes_.  Typically, this will be a 1-d index into a flat
    // memory space.
    extents.Set(extents.size() - 1, extents[extents.size() - 1] * var_lanes_);

    // Rewrite access to the buffer in the body.
672
673
    Stmt body =
        VecAllocAccess(op->buffer_var.get(), var_, var_lanes_)(op->body);
674
675
676
677
678
679
680
681
682
683
684
    body = this->VisitStmt(body);
    return Allocate(op->buffer_var, op->dtype, extents, condition, body);
  }

  // scalarize the statement
  Stmt Scalarize(Stmt stmt) {
    Var idx(var_->name_hint + ".s", var_->dtype);
    stmt = Substitute(stmt, {{var_, idx}});
    return For(idx, IntImm(var_->dtype, 0), var_lanes_, ForKind::kSerial, stmt);
  }
  // ProducerStore
685
  Stmt VisitStmt_(const ProducerStoreNode *op) final {
686
687
688
    LOG(FATAL) << "ProducerProvide cannot appear in a TIR PrimFunc";
  }

689
private:
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
  // analyzer
  arith::Analyzer analyzer_;
  // deep equal
  ExprDeepEqual deep_equal_;
  // variable to be replaced
  Var var_;
  // the lanes.
  PrimExpr var_lanes_;
  // ramp representing the var.
  PrimExpr ramp_;
  // flag to mark requirement of scalarization.
  bool need_scalarize_{false};
  // Let binding
  std::unordered_map<Var, PrimExpr, ObjectPtrHash, ObjectPtrEqual> let_binding_;
  // vectorizable property
705
706
  OpAttrMap<TVectorizable> op_vectorizable_ =
      Op::GetAttrMap<TVectorizable>("TVectorizable");
707
708
709

  // mutate array, with given lane requirement
  // when finished, p_lane updates the lane requirement.
710
711
712
713
  Array<PrimExpr> MutateArray(Array<PrimExpr> arr, int *p_lanes) {
    if (arr.size() == 0)
      return arr;
    int &lanes = *p_lanes;
714
715
716
717
718
    bool changed = false;
    std::vector<PrimExpr> new_arr(arr.size());
    for (size_t i = 0; i < arr.size(); i++) {
      PrimExpr old_elem = arr[i];
      PrimExpr new_elem = this->VisitExpr(old_elem);
719
720
      if (!new_elem.same_as(old_elem))
        changed = true;
721
722
723
724
725
726
727
728
729
730
      new_arr[i] = new_elem;
      lanes = std::max(lanes, new_elem.dtype().lanes());
    }

    for (size_t i = 0; i < arr.size(); ++i) {
      if (new_arr[i].dtype().lanes() != lanes) {
        new_arr[i] = BroadcastTo(new_arr[i], lanes, false);
        changed = true;
      }
    }
731
732
    if (!changed)
      return arr;
733
734
    return Array<PrimExpr>(new_arr);
  }
735
736
737
  template <typename TOp, typename T> PrimExpr BinaryVec(const T *op) {
    static_assert(std::is_same<typename TOp::ContainerType, T>::value,
                  "constraint");
738
739
740
741
742
743
744
745
    PrimExpr a = this->VisitExpr(op->a);
    PrimExpr b = this->VisitExpr(op->b);
    if (a.same_as(op->a) && b.same_as(op->b)) {
      return GetRef<PrimExpr>(op);
    } else {
      int a_lanes = a.dtype().get_lanes_or_vscale_factor();
      int b_lanes = b.dtype().get_lanes_or_vscale_factor();
      int lanes = std::max(a_lanes, b_lanes);
746
747
748
749
      bool is_scalable =
          a.dtype().is_scalable_vector() || b.dtype().is_scalable_vector();
      return TOp(BroadcastTo(a, lanes, is_scalable),
                 BroadcastTo(b, lanes, is_scalable));
750
751
752
    }
  }
  template <typename T, typename FCompute>
753
  PrimExpr AddSubVec(const T *op, FCompute fcompute) {
754
755
756
757
758
759
760
761
762
    PrimExpr a = this->VisitExpr(op->a);
    PrimExpr b = this->VisitExpr(op->b);
    if (a.same_as(op->a) && b.same_as(op->b)) {
      return GetRef<PrimExpr>(op);
    } else {
      int a_lanes = a.dtype().get_lanes_or_vscale_factor();
      int b_lanes = b.dtype().get_lanes_or_vscale_factor();
      int lanes = std::max(a_lanes, b_lanes);
      if (lanes != 1) {
763
764
        const RampNode *b_ramp = b.as<RampNode>();
        const RampNode *a_ramp = a.as<RampNode>();
765
        if (a.dtype().is_scalar() && b_ramp) {
766
767
768
769
          return Ramp(
              fcompute(a, b_ramp->base),
              fcompute(make_zero(b_ramp->stride.dtype()), b_ramp->stride),
              b_ramp->lanes);
770
771
772
773
774
        }
        if (b.dtype().is_scalar() && a_ramp) {
          return Ramp(fcompute(a_ramp->base, b), a_ramp->stride, a_ramp->lanes);
        }
      }
775
776
777
778
      bool is_scalable =
          a.dtype().is_scalable_vector() || b.dtype().is_scalable_vector();
      return fcompute(BroadcastTo(a, lanes, is_scalable),
                      BroadcastTo(b, lanes, is_scalable));
779
780
781
782
    }
  }
};

783
784
} // namespace tl
} // namespace tvm