simplify_algebra.cpp 53.9 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
/*
 * The MIT License (MIT)
 *
 * Copyright (c) 2015-2022 Advanced Micro Devices, Inc. All rights reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to deal
 * in the Software without restriction, including without limitation the rights
 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 * THE SOFTWARE.
 */
Paul's avatar
Paul committed
24
#include <migraphx/simplify_algebra.hpp>
Paul's avatar
Paul committed
25
#include <migraphx/dead_code_elimination.hpp>
Paul's avatar
Paul committed
26
#include <migraphx/program.hpp>
27
#include <migraphx/op/concat.hpp>
28
#include <migraphx/op/slice.hpp>
29
#include <migraphx/op/convolution.hpp>
Paul's avatar
Paul committed
30
#include <migraphx/op/broadcast.hpp>
31
32
#include <migraphx/op/reshape.hpp>
#include <migraphx/op/transpose.hpp>
Paul's avatar
Paul committed
33
#include <migraphx/matcher.hpp>
34
#include <migraphx/common.hpp>
Paul's avatar
Paul committed
35
#include <migraphx/literal.hpp>
36
37
38
#include <migraphx/make_op.hpp>
#include <migraphx/serialize.hpp>

39
#include <migraphx/algorithm.hpp>
Shucai Xiao's avatar
Shucai Xiao committed
40
#include <unordered_set>
Paul's avatar
Paul committed
41

Paul's avatar
Paul committed
42
namespace migraphx {
Paul's avatar
Paul committed
43
inline namespace MIGRAPHX_INLINE_NS {
Paul's avatar
Paul committed
44

Paul's avatar
Paul committed
45
auto lit_broadcast() { return match::any_of(match::is_constant(), match::name("broadcast")); }
Paul's avatar
Paul committed
46
auto not_lit_broadcast() { return match::none_of(match::is_constant(), match::name("broadcast")); }
Paul's avatar
Paul committed
47
48
auto op_lit_broadcast(std::string op, std::string x, std::string y)
{
Paul's avatar
Paul committed
49
50
    return match::name(std::move(op))(match::either_arg(0, 1)(
        lit_broadcast().bind(std::move(x)), not_lit_broadcast().bind(std::move(y))));
Paul's avatar
Paul committed
51
52
}

Paul's avatar
Paul committed
53
54
auto conv_const_weights()
{
55
56
57
    return match::name("convolution")(
        match::used_once(),
        match::args(match::none_of(match::is_constant()), match::is_constant().bind("w")));
Paul's avatar
Paul committed
58
59
}

Shucai Xiao's avatar
Shucai Xiao committed
60
61
auto reduction() { return match::name_contains("reduce"); }

62
// conv(x, w) * a => conv(x, a * w)
Paul's avatar
Paul committed
63
64
65
struct find_mul_conv
{
    auto matcher() const
Paul's avatar
Paul committed
66
    {
67
68
69
        return match::name("mul")(
            match::either_arg(0, 1)(conv_const_weights().bind("conv"),
                                    match::name("broadcast", "multibroadcast").bind("a")));
Paul's avatar
Paul committed
70
    }
Paul's avatar
Paul committed
71

72
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
73
    {
Paul's avatar
Paul committed
74
        auto ins      = r.result;
Paul's avatar
Paul committed
75
        auto conv_ins = r.instructions["conv"];
Paul's avatar
Paul committed
76
77
78
        auto a_ins    = r.instructions["a"];
        auto w_ins    = r.instructions["w"];

79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
        const auto& a_input_lens = a_ins->inputs().front()->get_shape().lens();

        std::size_t num_not_one_dims = std::count_if(
            a_input_lens.cbegin(), a_input_lens.cend(), [](auto dim) { return dim != 1; });
        if(num_not_one_dims > 1)
            return;

        // check broadcasted along channels
        const auto& a_lens    = a_ins->get_shape().lens();
        const auto& a_strides = a_ins->get_shape().strides();

        auto is_broadcasted_axis = [](auto len, auto stride) { return len == 1 or stride == 0; };

        if(a_strides.at(1) != 1)
            return;

        if(not is_broadcasted_axis(a_lens.front(), a_strides.front()))
            return;

        if(not std::equal(a_lens.begin() + 2,
                          a_lens.end(),
                          a_strides.begin() + 2,
                          a_strides.end(),
                          is_broadcasted_axis))
Paul's avatar
Paul committed
103
104
            return;

105
        auto sq    = m.insert_instruction(ins, make_op("squeeze"), a_ins->inputs().front());
106
        auto new_a = m.insert_instruction(
107
            ins, make_op("broadcast", {{"axis", 0}, {"out_lens", w_ins->get_shape().lens()}}), sq);
108
109
        auto new_mul  = m.insert_instruction(ins, make_op("mul"), new_a, w_ins);
        auto new_conv = m.insert_instruction(
Paul's avatar
Paul committed
110
            ins, conv_ins->get_operator(), conv_ins->inputs().front(), new_mul);
111
        m.replace_instruction(ins, new_conv);
Paul's avatar
Paul committed
112
    }
Paul's avatar
Paul committed
113
114
};

115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
struct find_mul_slice_conv
{
    static auto conv()
    {
        return match::name("convolution")(
            match::all_of[match::outputs()](match::name("slice")),
            match::args(match::any(), match::is_constant().bind("w")));
    }
    auto matcher() const
    {
        return match::name("mul")(match::either_arg(0, 1)(
            match::name("slice")(match::used_once(), match::arg(0)(conv().bind("conv")))
                .bind("slice"),
            match::name("broadcast")(match::is_constant()).bind("a")));
    }

131
    void apply(module& m, const match::matcher_result& r) const
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
    {
        auto ins       = r.result;
        auto slice_ins = r.instructions["slice"];
        auto conv_ins  = r.instructions["conv"];
        auto a_ins     = r.instructions["a"];
        auto w_ins     = r.instructions["w"];

        auto broadcast_op = any_cast<op::broadcast>(a_ins->get_operator());
        if(broadcast_op.axis != 1)
            return;

        auto slice_op = any_cast<op::slice>(slice_ins->get_operator());
        if(slice_op.axes.size() != 1)
            return;
        if(slice_op.axes.front() != 1)
            return;

        auto slice_idx = std::distance(conv_ins, slice_ins);
        if(std::any_of(conv_ins->outputs().begin(), conv_ins->outputs().end(), [&](auto i) {
               if(i == slice_ins)
                   return false;
               if(std::distance(conv_ins, i) < slice_idx)
                   return true;
               auto sop = any_cast<op::slice>(i->get_operator());
               if(sop.axes != slice_op.axes)
                   return true;
               if(std::max(sop.starts.front(), slice_op.starts.front()) <
                  std::min(sop.ends.front(), slice_op.ends.front()))
                   return true;
               return false;
           }))
            return;

        auto w_slice_op  = slice_op;
        w_slice_op.axes  = {0};
167
        auto slice_w_ins = m.insert_instruction(ins, w_slice_op, w_ins);
168

169
        auto new_a = m.insert_instruction(
170
            ins,
171
            make_op("broadcast", {{"axis", 0}, {"out_lens", slice_w_ins->get_shape().lens()}}),
172
            a_ins->inputs().front());
173
        auto new_mul = m.insert_instruction(ins, make_op("mul"), new_a, slice_w_ins);
174
175
176

        std::vector<instruction_ref> sliced_weights;
        if(slice_op.starts.front() != 0)
177
            sliced_weights.push_back(m.insert_instruction(
178
179
180
                ins,
                make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", slice_op.starts}}),
                w_ins));
181
182
183
        sliced_weights.push_back(new_mul);
        int64_t end_axis = w_ins->get_shape().lens().at(0);
        if(slice_op.ends.front() != end_axis)
184
            sliced_weights.push_back(m.insert_instruction(
185
186
187
                ins,
                make_op("slice", {{"axes", {0}}, {"starts", slice_op.ends}, {"ends", {end_axis}}}),
                w_ins));
188

189
        auto new_weights =
190
            m.insert_instruction(ins, make_op("concat", {{"axis", 0}}), sliced_weights);
191

192
        auto new_conv = m.insert_instruction(
193
194
195
            ins, conv_ins->get_operator(), conv_ins->inputs().front(), new_weights);
        assert(conv_ins->get_shape() == new_conv->get_shape());

196
        auto slice1 = m.insert_instruction(ins, slice_op, new_conv);
197
        assert(ins->get_shape().lens() == slice1->get_shape().lens());
198
        m.replace_instruction(ins, slice1);
199
        // TODO: Check each slice doesn't overlap and that it occurs after slice_ins
200
201
        auto outputs = conv_ins->outputs();
        for(auto output : outputs)
202
203
204
205
206
            if(output != slice_ins)
                instruction::replace_argument(output, conv_ins, new_conv);
    }
};

207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
struct find_mul_dot
{
    auto matcher() const
    {
        auto is_dot_const_inputs =
            match::name("dot")(match::any_of[match::inputs()](match::is_constant()));
        return match::name("mul")(match::either_arg(0, 1)(
            is_dot_const_inputs.bind("dot"), match::name("broadcast", "multibroadcast").bind("c")));
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins     = r.result;
        auto dot_ins = r.instructions["dot"];
        auto a_ins   = dot_ins->inputs()[0];
        auto b_ins   = dot_ins->inputs()[1];
        auto c_ins   = r.instructions["c"];

        const auto& c_strides = c_ins->get_shape().strides();

        // There should only be one stride that is not zero
        if(std::count_if(c_strides.begin(), c_strides.end(), [](auto s) { return s != 0; }) > 1)
            return;

        auto add_mul_const = [&](instruction_ref x_ins) {
            if(not x_ins->can_eval())
                return m.end();
            auto broadcast_v        = c_ins->get_operator().to_value();
            broadcast_v["out_lens"] = x_ins->get_shape().lens();

            auto cb_ins =
                m.insert_instruction(ins, make_op(c_ins->name(), broadcast_v), c_ins->inputs());
            return m.insert_instruction(ins, make_op("mul"), x_ins, cb_ins);
        };

        if(c_strides.back() == 1)
        {
            b_ins = add_mul_const(b_ins);
        }
        else if(c_strides[c_strides.size() - 2] == 1)
        {
            a_ins = add_mul_const(a_ins);
        }
        else if(c_ins->get_shape().scalar())
        {
            if(a_ins->can_eval())
                a_ins = add_mul_const(a_ins);
            else
                b_ins = add_mul_const(b_ins);
        }
        else
        {
            return;
        }

        if(contains({a_ins, b_ins}, m.end()))
            return;

        m.replace_instruction(ins, make_op("dot"), a_ins, b_ins);
    }
};

struct find_dot_mul
{
    auto matcher() const
    {
        auto const_broadcast = match::name("broadcast", "multibroadcast")(match::is_constant());
        auto mul             = match::name("mul")(
            match::used_once(),
            match::either_arg(0, 1)(const_broadcast.bind("d"),
                                    match::none_of(match::is_constant()).bind("z")));
        return match::name("dot")(match::either_arg(0, 1)(mul, match::is_constant().bind("c")));
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins   = r.result;
        auto a_ins = ins->inputs()[0];
        auto b_ins = ins->inputs()[1];
        auto d_ins = r.instructions["d"];
        auto c_ins = r.instructions["c"];
        auto z_ins = r.instructions["z"];

        const auto& d_strides = d_ins->get_shape().strides();

        // There should only be one stride that is not zero
        if(std::count_if(d_strides.begin(), d_strides.end(), [](auto s) { return s != 0; }) > 1)
            return;

        if(not d_ins->get_shape().scalar())
        {
            if(d_strides.back() == 1 and not b_ins->can_eval())
                return;
            if(d_strides[d_strides.size() - 2] == 1 and not a_ins->can_eval())
                return;
        }

        auto broadcast_v = d_ins->get_operator().to_value();
        auto c_lens      = c_ins->get_shape().lens();
        std::vector<int64_t> permutation(c_lens.size());
        std::iota(permutation.begin(), permutation.end(), 0);
        std::swap(permutation.back(), permutation[permutation.size() - 2]);
        c_lens                  = reorder_dims(c_lens, permutation);
        broadcast_v["out_lens"] = c_lens;
        auto db_ins =
            m.insert_instruction(ins, make_op(d_ins->name(), broadcast_v), d_ins->inputs());
        auto db_transpose_ins =
            m.insert_instruction(ins, make_op("transpose", {{"permutation", permutation}}), db_ins);
        auto cd_ins = m.insert_instruction(ins, make_op("mul"), c_ins, db_transpose_ins);

        if(c_ins == b_ins)
        {
            a_ins = z_ins;
            b_ins = cd_ins;
        }
        else
        {
            a_ins = cd_ins;
            b_ins = z_ins;
        }

        m.replace_instruction(ins, make_op("dot"), a_ins, b_ins);
    }
};

332
333
334
335
336
337
// ******************************
//  a * (x + b) => a * x + a * b
// ******************************
// When a * (x + b) is followed by another add of constant, then the
// additional add can be const folded. Also, better fusions can be applied
// when the add comes after.
Paul's avatar
Paul committed
338
339
340
341
342
struct find_mul_add
{
    auto matcher() const
    {
        return match::name("mul")(match::either_arg(0, 1)(
Paul's avatar
Paul committed
343
344
345
            match::name("add")(
                match::either_arg(0, 1)(
                    match::any().bind("x"),
Paul's avatar
Paul committed
346
                    match::any_of(conv_const_weights(), match::is_constant()).bind("b")),
Paul's avatar
Paul committed
347
                match::none_of(match::args(match::is_constant(), match::is_constant())),
Paul's avatar
Paul committed
348
                match::used_once()),
Paul's avatar
Paul committed
349
            match::is_constant().bind("a")));
Paul's avatar
Paul committed
350
351
    }

352
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
353
    {
Paul's avatar
Paul committed
354
        auto ins   = r.result;
Paul's avatar
Paul committed
355
        auto a_ins = r.instructions["a"];
Paul's avatar
Paul committed
356
        auto b_ins = r.instructions["b"];
Paul's avatar
Paul committed
357
        auto x_ins = r.instructions["x"];
Paul's avatar
Paul committed
358
        assert(x_ins != b_ins);
Paul's avatar
Paul committed
359

360
361
362
        auto ax_ins = m.insert_instruction(ins, make_op("mul"), a_ins, x_ins);
        auto ab_ins = m.insert_instruction(ins, make_op("mul"), a_ins, b_ins);
        m.replace_instruction(ins, make_op("add"), ax_ins, ab_ins);
Paul's avatar
Paul committed
363
364
365
    }
};

366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
struct find_dot_add
{
    auto matcher() const
    {
        return match::name("dot")(match::either_arg(0, 1)(
            match::name("add")(
                match::either_arg(0, 1)(match::any().bind("x"),
                                        match::any_of(match::is_constant()).bind("b")),
                match::none_of(match::args(match::is_constant(), match::is_constant())),
                match::used_once()),
            match::is_constant().bind("a")));
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins   = r.result;
        auto a_ins = r.instructions["a"];
        auto b_ins = r.instructions["b"];
        auto x_ins = r.instructions["x"];
        assert(x_ins != b_ins);

        const bool flipped = a_ins == ins->inputs().back();

        auto insert_dot = [&](auto x, auto y) {
            if(flipped)
                return m.insert_instruction(ins, make_op("dot"), y, x);
            else
                return m.insert_instruction(ins, make_op("dot"), x, y);
        };

        auto ax_ins = insert_dot(a_ins, x_ins);
        auto ab_ins = insert_dot(a_ins, b_ins);
        m.replace_instruction(ins, make_op("add"), ax_ins, ab_ins);
    }
};

402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
struct find_conv_add
{
    auto matcher() const
    {
        auto add = match::name("add")(
            match::either_arg(0, 1)(match::any().bind("x"),
                                    match::any_of(match::is_constant()).bind("a")),
            match::used_once());
        return match::name("convolution")(match::used_once(),
                                          match::args(add, match::is_constant().bind("w")));
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins   = r.result;
        auto a_ins = r.instructions["a"];
        auto x_ins = r.instructions["x"];
        auto w_ins = r.instructions["w"];

        auto conv1 = m.insert_instruction(ins, ins->get_operator(), a_ins, w_ins);
        auto conv2 = m.insert_instruction(ins, ins->get_operator(), x_ins, w_ins);

        m.replace_instruction(ins, make_op("add"), conv1, conv2);
    }
};

Paul's avatar
Paul committed
428
struct find_add_lit_broadcast
Paul's avatar
Paul committed
429
430
431
432
433
434
435
{
    auto matcher() const
    {
        return match::name("add")(
            match::either_arg(0, 1)(op_lit_broadcast("add", "a", "x"), lit_broadcast().bind("b")));
    }

436
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
437
438
439
440
441
442
    {
        auto ins   = r.result;
        auto x_ins = r.instructions["x"];
        auto a_ins = r.instructions["a"];
        auto b_ins = r.instructions["b"];

443
444
        auto sumab = m.insert_instruction(ins, make_op("add"), a_ins, b_ins);
        m.replace_instruction(ins, make_op("add"), x_ins, sumab);
Paul's avatar
Paul committed
445
446
447
448
    }
};

struct find_double_add_lit_broadcast
Paul's avatar
Paul committed
449
{
Paul's avatar
Paul committed
450
451
    auto matcher() const
    {
Paul's avatar
Paul committed
452
        return match::name("add")(
Paul's avatar
Paul committed
453
            match::args(op_lit_broadcast("add", "a", "x"), op_lit_broadcast("add", "b", "y")));
Paul's avatar
Paul committed
454
455
    }

456
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
457
    {
Paul's avatar
Paul committed
458
459
460
461
462
        auto ins   = r.result;
        auto x_ins = r.instructions["x"];
        auto y_ins = r.instructions["y"];
        auto a_ins = r.instructions["a"];
        auto b_ins = r.instructions["b"];
Paul's avatar
Paul committed
463
464
465

        instruction_ref sumab;

Paul's avatar
Paul committed
466
        if(a_ins->name() == "broadcast" and b_ins->name() == "broadcast")
Paul's avatar
Paul committed
467
468
469
        {
            if(a_ins->inputs().at(0)->get_shape() != b_ins->inputs().at(0)->get_shape())
                return;
470
            auto op     = a_ins->get_operator();
471
            auto presum = m.insert_instruction(
472
                ins, make_op("add"), a_ins->inputs().at(0), b_ins->inputs().at(0));
473
            sumab = m.insert_instruction(ins, op, presum);
Paul's avatar
Paul committed
474
475
476
        }
        else
        {
477
            sumab = m.insert_instruction(ins, make_op("add"), a_ins, b_ins);
Paul's avatar
Paul committed
478
479
        }

480
481
        auto sumxy = m.insert_instruction(ins, make_op("add"), x_ins, y_ins);
        m.replace_instruction(ins, make_op("add"), sumxy, sumab);
Paul's avatar
Paul committed
482
483
484
    }
};

Paul's avatar
Paul committed
485
486
struct find_inner_broadcast
{
Paul's avatar
Paul committed
487
    // (match::none_of(match::name("quantizelinear",  "dequantizelinear")))
488
    auto matcher() const { return pointwise(match::all_of[match::inputs()](match::broadcast())); }
Paul's avatar
Paul committed
489

490
491
492
493
494
495
496
497
498
    static auto non_scalar_op(const std::string& name)
    {
        return [=](instruction_ref ins) {
            if(ins->get_shape().scalar())
                return false;
            return ins->name() == name;
        };
    }

499
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
500
    {
501
502
503
504
        auto ins        = r.result;
        auto broadcasts = ins->inputs();
        if(broadcasts.empty())
            return;
Paul's avatar
Paul committed
505
506
507
        // Skip if different data types are used
        if (any_of(broadcasts, [&](auto i) { return i->get_shape().type() != broadcasts.front()->get_shape().type(); }))
            return;
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
        bool mixed_broadcasts = any_of(broadcasts, non_scalar_op("broadcast")) and
                                any_of(broadcasts, non_scalar_op("multibroadcast"));
        // If the broadcast is not a single dimension, then dont perform inner_broadcast
        if(mixed_broadcasts and any_of(broadcasts, [&](instruction_ref i) {
               if(i->get_shape().scalar())
                   return false;
               if(i->name() == "multibroadcast")
                   return false;
               auto input       = i->inputs().at(0);
               const auto& lens = input->get_shape().lens();
               return std::count_if(lens.begin(), lens.end(), [&](std::size_t d) {
                          return d == 1;
                      }) < (lens.size() - 1);
           }))
            return;
523
524
525
526
        std::vector<instruction_ref> inputs;
        std::transform(broadcasts.begin(),
                       broadcasts.end(),
                       std::back_inserter(inputs),
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
                       [&](instruction_ref i) {
                           auto input = i->inputs().front();
                           if(mixed_broadcasts and not i->get_shape().scalar() and
                              i->get_shape().lens().size() > 1)
                               return m.insert_instruction(i, make_op("squeeze"), input);
                           return input;
                       });

        std::sort(broadcasts.begin(), broadcasts.end(), by(std::less<>{}, [](instruction_ref i) {
                      if(i->get_shape().scalar())
                          return 2;
                      else if(i->name() == "broadcast")
                          return 0;
                      if(i->name() == "multibroadcast")
                          return 1;
                      return 3;
                  }));
544
        auto op = insert_common_op(m, ins, ins->get_operator(), inputs);
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
        m.replace_instruction(ins, broadcasts.front()->get_operator(), op);
    }
};

struct find_dot_broadcast
{
    auto matcher() const
    {
        return match::name("dot")(match::all_of[match::inputs()](match::broadcast()));
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins = r.result;
        auto a   = ins->inputs()[0];
        auto b   = ins->inputs()[1];
        if(a->get_operator().name() != b->get_operator().name())
            return;
        if(ins->get_shape().lens().size() < 3)
            return;
        auto nbatch_axes      = ins->get_shape().lens().size() - 2;
        const auto& a_strides = a->get_shape().strides();
        const auto& b_strides = b->get_shape().strides();
        // Find leading batch axes that are broadcasted
        auto p =
            std::mismatch(a_strides.begin(),
                          a_strides.begin() + nbatch_axes,
                          b_strides.begin(),
                          b_strides.begin() + nbatch_axes,
                          [](auto astride, auto bstride) { return astride == 0 and bstride == 0; });
        auto naxes = p.first - a_strides.begin();
        assert(naxes <= nbatch_axes);
        std::vector<std::size_t> axes(naxes);
        std::iota(axes.begin(), axes.end(), 0);

        auto insert_broadcast = [&](instruction_ref b_ins) -> instruction_ref {
            auto input = b_ins->inputs()[0];
            std::vector<std::size_t> lens(b_ins->get_shape().lens().begin() + naxes,
                                          b_ins->get_shape().lens().end());
            if(b_ins->name() == "multibroadcast")
            {
                return m.insert_instruction(
                    ins, make_op("multibroadcast", {{"out_lens", lens}}), input);
            }
            else if(b_ins->name() == "broadcast")
            {
                auto v    = b_ins->get_operator().to_value();
                auto axis = v.at("axis").to<std::size_t>() - naxes;
                return m.insert_instruction(
                    ins, make_op("broadcast", {{"axis", axis}, {"out_lens", lens}}), input);
            }
            assert(false);
            return m.end();
        };
        auto a1        = insert_broadcast(a);
        auto b1        = insert_broadcast(b);
        auto dot       = m.insert_instruction(ins, make_op("dot"), a1, b1);
        auto broadcast = m.insert_instruction(
            ins, make_op("multibroadcast", {{"out_lens", ins->get_shape().lens()}}), dot);
        m.replace_instruction(ins, broadcast);
Paul's avatar
Paul committed
605
606
607
    }
};

608
struct find_concat_op
609
610
611
{
    auto matcher() const
    {
612
        return match::name("concat")(match::any_of[match::inputs()](
613
614
            match::any_of(match::pointwise(), match::name("broadcast", "multibroadcast")),
            match::used_once()));
615
616
    }

617
618
    template <class Iterator>
    static std::vector<std::size_t> get_output_lens(Iterator start, Iterator last, std::size_t axis)
619
    {
620
621
622
        assert(start != last);
        std::size_t dim = 0;
        for(auto ins : range(start, last))
623
        {
624
            dim += ins->get_shape().lens().at(axis);
625
        }
626
627
628
        auto lens  = (*start)->get_shape().lens();
        lens[axis] = dim;
        return lens;
629
630
    }

631
632
    static bool is_valid_op(const operation& op)
    {
633
634
        return contains({"broadcast", "multibroadcast"}, op.name()) or
               op.attributes().contains("pointwise");
635
636
    }

637
    void apply(module& m, const match::matcher_result& r) const
638
    {
639
640
        auto ins  = r.result;
        auto axis = any_cast<op::concat>(ins->get_operator()).axis;
641

642
643
644
645
646
647
        auto each = [&](auto start, auto last) -> std::vector<instruction_ref> {
            if(std::distance(start, last) < 2)
                return {start, last};
            auto x = *start;
            if(x->inputs().size() > 2 or x->inputs().empty() or x->outputs().size() > 1)
                return {start, last};
648
649
            auto op = x->get_operator();
            if(not is_valid_op(op))
650
651
652
653
654
655
656
657
658
659
660
661
                return {start, last};
            auto iaxis = axis;
            // Adjust broadcast lens
            if(op.name() == "broadcast")
            {
                auto b = any_cast<op::broadcast>(op);
                if(b.axis != iaxis)
                    return {start, last};
                b.broadcast_lens = get_output_lens(start, last, iaxis);
                op               = b;
                iaxis            = 0;
            }
662
663
664
665
666
667
668
669
670
671
            else if(op.name() == "multibroadcast")
            {
                shape bshape = (*start)->get_shape();
                auto input   = (*start)->inputs()[0];
                if(iaxis >= bshape.strides().size() or bshape.strides()[iaxis] == 0)
                    return {start, last};
                op.from_value({{"out_lens", get_output_lens(start, last, iaxis)}});
                auto delta = bshape.lens().size() - input->get_shape().lens().size();
                iaxis -= delta;
            }
672
673
674
675
676
677
678
679

            std::vector<instruction_ref> concats;
            for(std::size_t i = 0; i < x->inputs().size(); i++)
            {
                std::vector<instruction_ref> inputs;
                std::transform(start, last, std::back_inserter(inputs), [&](auto j) {
                    return j->inputs().at(i);
                });
680
                auto concat =
681
                    m.insert_instruction(ins, make_op("concat", {{"axis", iaxis}}), inputs);
682
683
                concats.push_back(concat);
            }
684
            auto y = m.insert_instruction(ins, op, concats);
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
            return {y};
        };

        std::vector<instruction_ref> args;
        auto update_args = [&](auto start, auto last) {
            auto x = each(start, last);
            args.insert(args.end(), x.begin(), x.end());
        };
        auto pred = [](auto i, auto j) {
            return i->get_operator() == j->get_operator() and
                   i->inputs().size() == i->inputs().size() and
                   i->outputs().size() == i->outputs().size();
        };
        group_unique(ins->inputs().begin(), ins->inputs().end(), update_args, pred);
        if(args.size() == 1)
700
            m.replace_instruction(ins, args.front());
701
        else
702
            m.replace_instruction(ins, make_op("concat", {{"axis", axis}}), args);
703
704
705
    }
};

706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
void move_instructions_back(module& m, instruction_ref pos, std::vector<instruction_ref> inss)
{
    auto start = range(m.begin(), pos);
    for(auto ins : iterator_for(start))
    {
        auto it = std::find(inss.begin(), inss.end(), ins);
        if(it != inss.end())
            inss.erase(it);
    }
    for(auto ins : inss)
    {
        if(not m.has_instruction(ins))
            continue;
        move_instructions_back(m, pos, ins->inputs());
        m.move_instruction(ins, pos);
    }
}

724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
std::vector<instruction_ref> get_splits(instruction_ref ins)
{
    std::vector<instruction_ref> result;
    std::copy_if(ins->outputs().begin(),
                 ins->outputs().end(),
                 std::back_inserter(result),
                 [&](auto i) { return i->name() == "slice"; });
    if(result.size() < 2)
        return {};
    auto get_slice = [](auto& i) -> auto& { return any_cast<op::slice>(i->get_operator()); };
    auto&& axes    = get_slice(result.front()).axes;
    if(std::any_of(result.begin(), result.end(), [&](auto i) { return get_slice(i).axes != axes; }))
        return {};
    auto get_start = [&](auto& i) -> auto& { return get_slice(i).starts; };
    auto get_end   = [&](auto& i) -> auto& { return get_slice(i).ends; };
    std::sort(
        result.begin(), result.end(), [&](auto x, auto y) { return get_start(x) < get_start(y); });
    if(std::any_of(get_start(result.front()).begin(), get_start(result.front()).end(), [&](auto i) {
           return i != 0;
       }))
        return {};
    auto it = std::adjacent_find(
        result.begin(), result.end(), [&](auto x, auto y) { return get_end(x) != get_start(y); });
    if(it != result.end())
        return {};
    for(std::size_t i = 0; i < axes.size(); i++)
    {
        auto axis = axes[i];
        if(ins->get_shape().lens()[axis] != get_slice(result.back()).ends[i])
            return {};
    }
    return result;
}

struct find_splits
{
    auto matcher() const
    {
762
763
764
        return match::any(
            match::any_of[match::outputs()](match::name("slice")(match::any_of[match::outputs()](
                match::pointwise(match::any_of(match::nargs(1), match::nargs(2))), reduction()))));
765
766
    }

Shucai Xiao's avatar
Shucai Xiao committed
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
    static bool is_dependent(const module& m, instruction_ref ins1, instruction_ref ins2)
    {

        std::unordered_set<instruction_ref> traversed;
        return fix<bool>([&](auto self, auto ins) -> bool {
            if(ins == ins2)
                return true;

            if(contains(traversed, ins))
                return false;

            traversed.insert(ins);
            const auto& inputs = ins->inputs();
            return std::any_of(inputs.begin(), inputs.end(), [&](auto in) {
                return m.has_instruction(in) and self(in);
            });
        })(ins1);
    }

786
    static std::vector<std::vector<instruction_ref>>
Shucai Xiao's avatar
Shucai Xiao committed
787
    get_split_groups(const module& m, const std::vector<instruction_ref>& splits)
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
    {
        std::vector<std::vector<instruction_ref>> groups;
        for(auto out : splits.front()->outputs())
        {
            if(out->name() == "slice")
                continue;
            std::vector<instruction_ref> group;
            for(auto split : splits)
            {
                auto it =
                    std::find_if(split->outputs().begin(), split->outputs().end(), [&](auto i) {
                        return i->get_operator() == out->get_operator();
                    });
                if(it == split->outputs().end())
                    break;
                assert((*it)->name() != "slice");
Shucai Xiao's avatar
Shucai Xiao committed
804

805
                // If there is a duplicate bail
Shucai Xiao's avatar
Shucai Xiao committed
806
807
808
809
810
                // there are should be no dependency between instructions in the group
                if(std::any_of(group.begin(), group.end(), [&](auto i) {
                       return is_dependent(m, *it, i) or is_dependent(m, i, *it);
                   }))
                {
811
                    return {};
Shucai Xiao's avatar
Shucai Xiao committed
812
813
                }

814
815
816
817
818
819
820
821
822
                group.push_back(*it);
            }
            if(group.size() != splits.size())
                continue;
            groups.push_back(group);
        }
        return groups;
    }

Shucai Xiao's avatar
Shucai Xiao committed
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
    bool is_fusable(instruction_ref start, instruction_ref split_front) const
    {
        auto op = start->get_operator();
        if(contains(op.name(), "reduce"))
        {
            auto slc         = any_cast<op::slice>(split_front->get_operator());
            auto slc_axes    = slc.axes;
            auto reduce_axes = start->get_operator().to_value()["axes"].to_vector<int64_t>();
            // axes of slice and reduce op cannot have overlap
            if(std::any_of(slc_axes.begin(), slc_axes.end(), [&](auto axis) {
                   return (std::find(reduce_axes.begin(), reduce_axes.end(), axis) !=
                           reduce_axes.end());
               }))
            {
                return false;
            }
        }
        else if(not op.attributes().contains("pointwise"))
        {
            return false;
        }

        return true;
    }

848
    void apply(module& m, const match::matcher_result& r) const
849
    {
Shucai Xiao's avatar
Shucai Xiao committed
850
        auto ins    = r.result;
851
852
853
        auto splits = get_splits(ins);
        if(splits.empty())
            return;
Shucai Xiao's avatar
Shucai Xiao committed
854

855
        for(const auto& group : get_split_groups(m, splits))
856
        {
Shucai Xiao's avatar
Shucai Xiao committed
857
858
859
860
861
            auto start       = group.front();
            auto split_front = splits.front();
            auto op          = start->get_operator();
            if(not is_fusable(start, split_front))
            {
862
                continue;
Shucai Xiao's avatar
Shucai Xiao committed
863
            }
864
865
866
867
868
869

            // Make sure there is no duplicates
            assert(std::none_of(
                std::next(group.begin()), group.end(), [&](auto i) { return i == start; }));

            auto split_idx    = 0;
870
            instruction_ref c = m.end();
871
872
            if(start->inputs().size() == 1)
            {
873
                c = m.insert_instruction(std::next(ins), op, ins);
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
            }
            else if(start->inputs().size() == 2)
            {
                assert(not std::none_of(start->inputs().begin(), start->inputs().end(), [](auto i) {
                    return i->name() == "slice";
                }) && "one argument must be a split");
                auto data_idx = 1;
                if(start->inputs().back()->name() == "slice")
                {
                    split_idx = 1;
                    data_idx  = 0;
                }

                std::vector<instruction_ref> data_args;
                std::transform(group.begin(),
                               group.end(),
                               std::back_inserter(data_args),
                               [&](auto i) { return i->inputs()[data_idx]; });

                // Data arguments must be a constant
                if(std::any_of(data_args.begin(), data_args.end(), [](auto i) {
                       return not i->can_eval();
                   }))
                    return;

899
                move_instructions_back(m, ins, data_args);
900
901
902
903
904
905
906

                auto slice_op = any_cast<op::slice>(splits.front()->get_operator());
                assert(not slice_op.axes.empty());
                if(slice_op.axes.size() > 1)
                    return;
                auto concat_axis = slice_op.axes.front();
                // TODO: Check if axises match
907
                auto concat = m.insert_instruction(
908
                    ins, make_op("concat", {{"axis", concat_axis}}), data_args);
909
910
911
912
913

                std::vector<instruction_ref> args;
                args.resize(2);
                args[split_idx] = ins;
                args[data_idx]  = concat;
914
                c               = m.insert_instruction(std::next(ins), op, args);
915
            }
916
            if(c != m.end())
917
918
919
920
921
922
            {
                for(auto i : group)
                {
                    auto split = i->inputs()[split_idx];
                    assert(split->name() == "slice");
                    // Insert contiguous for reshapes
923
924
                    auto outputs = i->outputs();
                    for(auto output : outputs)
925
                    {
926
                        if(output->name() != "reshape")
927
                            continue;
928
                        auto x = m.insert_instruction(output, make_op("contiguous"), i);
929
                        m.replace_instruction(output, output->get_operator(), x);
930
931
                    }

932
                    m.replace_instruction(i, split->get_operator(), c);
933
934
935
936
937
938
939
940
941
942
943
944
945
946
                }
            }
        }
    }
};

struct find_split_concat
{
    auto matcher() const
    {
        return match::any(match::any_of[match::outputs()](
            match::name("slice")(match::all_of[match::outputs()](match::name("concat")))));
    }

947
    void apply(module& m, const match::matcher_result& r) const
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
    {
        auto ins = r.result;

        auto splits = get_splits(ins);
        if(splits.empty())
            return;
        if(std::any_of(
               splits.begin(), splits.end(), [](auto i) { return i->outputs().size() != 1; }))
            return;
        // Check for concat operator
        auto concat = splits.front()->outputs().front();
        if(std::any_of(splits.begin(), splits.end(), [&](auto i) {
               return i->outputs().front() != concat;
           }))
            return;
        // Check axis match
        auto concat_op = any_cast<op::concat>(concat->get_operator());
        auto split_op  = any_cast<op::slice>(splits.front()->get_operator());
        if(split_op.axes.size() != 1)
            return;
        if(split_op.axes.front() != concat_op.axis)
            return;
        // Replace args
        auto args = concat->inputs();
        auto it =
            std::find_if(args.begin(), args.end(), [&](auto i) { return i == splits.front(); });
        if(std::distance(it, args.end()) < splits.size())
            return;
976
977
978
979
980
981
982
        // If the slices are not in order then stop
        if(not std::is_sorted(it, it + splits.size(), [](instruction_ref x, instruction_ref y) {
               auto xop = any_cast<op::slice>(x->get_operator());
               auto yop = any_cast<op::slice>(y->get_operator());
               return std::tie(xop.starts, xop.ends) < std::tie(yop.starts, yop.ends);
           }))
            return;
983
984
985
986
        *it = splits.front()->inputs().front();
        args.erase(std::next(it), it + splits.size());

        if(args.size() == 1)
987
            m.replace_instruction(concat, args.front());
988
        else
989
            m.replace_instruction(concat, concat->get_operator(), args);
990
991
992
    }
};

993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
bool axis_equal(const std::vector<std::size_t>& x,
                const std::vector<std::size_t>& y,
                std::size_t axis)
{
    return x.size() == y.size() and x.size() > axis and
           std::equal(x.begin(), x.begin() + axis, y.begin()) and
           std::equal(x.begin() + axis + 1, x.end(), y.begin() + axis + 1);
}

bool axis_shape_equal(const shape& x, const shape& y, std::size_t axis)
{
    // TODO: Check strides
    return axis_equal(x.lens(), y.lens(), axis);
}

struct find_add_convs
{
    auto matcher() const
    {
        return match::name("add")(
            match::args(conv_const_weights().bind("a"), conv_const_weights().bind("b")));
    }

    static bool symmetrical_strides(const op::convolution& op)
    {
        return op.stride[0] == op.stride[1];
    }

    static std::size_t compute_stride_factor(const op::convolution& x, const op::convolution& y)
    {
        if(not symmetrical_strides(x))
            return 0;
        if(not symmetrical_strides(y))
            return 0;
        if((x.stride[0] % y.stride[0]) != 0)
            return 0;
        return x.stride[0] / y.stride[0];
    }

1032
    void apply(module& m, const match::matcher_result& r) const
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
    {
        auto ins       = r.result;
        auto a_conv    = r.instructions["a"];
        auto a_input   = a_conv->inputs().at(0);
        auto a_weights = a_conv->inputs().at(1);
        auto b_conv    = r.instructions["b"];
        auto b_input   = b_conv->inputs().at(0);
        auto b_weights = b_conv->inputs().at(1);

        if(not axis_shape_equal(a_weights->get_shape(), b_weights->get_shape(), 1))
            return;

        auto a_op   = any_cast<op::convolution>(a_conv->get_operator());
        auto b_op   = any_cast<op::convolution>(b_conv->get_operator());
        auto new_op = a_op;

        if(a_op != b_op)
        {
            if(std::tie(a_op.padding, a_op.dilation, a_op.group) ==
                   std::tie(b_op.padding, b_op.dilation, b_op.group) and
               a_weights->get_shape().lens()[2] == 1 and a_weights->get_shape().lens()[3] == 1)
            {
                if(a_op.stride < b_op.stride)
                {
                    auto n = compute_stride_factor(b_op, a_op);
                    if(n == 0)
                        return;
                    new_op  = a_op;
1061
                    b_input = m.insert_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
1062
                        ins, make_op("step", {{"axes", {2, 3}}, {"steps", {n, n}}}), b_input);
1063
1064
1065
1066
1067
1068
1069
                }
                else if(b_op.stride < a_op.stride)
                {
                    auto n = compute_stride_factor(a_op, b_op);
                    if(n == 0)
                        return;
                    new_op  = b_op;
1070
                    a_input = m.insert_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
1071
                        ins, make_op("step", {{"axes", {2, 3}}, {"steps", {n, n}}}), a_input);
1072
1073
1074
1075
1076
1077
1078
1079
                }
                else
                    return;
            }
            else
                return;
        }

1080
        auto concat_input =
1081
            m.insert_instruction(ins, make_op("concat", {{"axis", 1}}), a_input, b_input);
1082
        auto concat_weights =
1083
1084
            m.insert_instruction(ins, make_op("concat", {{"axis", 1}}), a_weights, b_weights);
        m.replace_instruction(ins, new_op, concat_input, concat_weights);
1085
1086
1087
    }
};

1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
MIGRAPHX_PRED_MATCHER(horiz_conv_dot, instruction_ref ins)
{
    auto pred = [&](auto name) {
        return [=](auto i) {
            return i->name() == name and i->inputs().front() == ins and
                   i->inputs().at(1)->can_eval();
        };
    };
    auto dots  = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("dot"));
    auto convs = std::count_if(ins->outputs().begin(), ins->outputs().end(), pred("convolution"));
1098
    return (dots >= 2 or convs >= 2);
1099
1100
1101
1102
1103
1104
}

struct find_conv_dot_horiz_fusion
{
    auto matcher() const { return horiz_conv_dot(); }

1105
    void apply(module& m, const match::matcher_result& r) const
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
    {
        auto ins = r.result;

        auto pred = [](auto i, auto j) {
            if(i->get_operator() != j->get_operator())
                return false;
            if(not contains({"dot", "convolution"}, i->name()))
                return true;
            auto x = i->inputs()[1]->get_shape().lens();
            auto y = j->inputs()[1]->get_shape().lens();
            if(x.size() != y.size())
                return false;
1118
            // Check that non-axes match
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
            int axis = 1;
            if(i->name() == "dot")
            {
                axis = x.size() - 1;
            }
            return axis_equal(x, y, axis);
        };

        auto each = [&](auto start, auto last) {
            if(std::distance(start, last) < 2)
                return;
            auto&& name = (*start)->name();
            if(not contains({"dot", "convolution"}, name))
                return;
1133
1134
1135
1136
1137
1138
1139
            auto op   = (*start)->get_operator();
            int group = 1;
            if(name == "convolution")
                group = any_cast<op::convolution>(op).group;
            // Skip group convolution
            if(group != 1)
                return;
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
            auto input = (*start)->inputs().front();
            std::vector<instruction_ref> args;
            std::transform(
                start, last, std::back_inserter(args), [&](auto x) { return x->inputs().at(1); });
            int axis        = 1;
            int concat_axis = 0;
            if(name == "dot")
            {
                axis        = int(args.front()->get_shape().lens().size() - 1);
                concat_axis = axis;
            }

1152
            move_instructions_back(m, input, args);
1153
            // TODO: Check if axes match
1154
            auto concat =
1155
1156
                m.insert_instruction(input, make_op("concat", {{"axis", concat_axis}}), args);
            auto fused     = m.insert_instruction(std::next(input), op, input, concat);
1157
1158
1159
            int64_t offset = 0;
            for(auto arg : range(start, last))
            {
1160
1161
1162
1163
1164
1165
1166
1167
1168
                auto outputs = arg->outputs();
                for(auto output : outputs)
                {
                    if(output->name() != "reshape")
                        continue;
                    auto x = m.insert_instruction(output, make_op("contiguous"), arg);
                    m.replace_instruction(output, output->get_operator(), x);
                }

1169
                int64_t len = arg->get_shape().lens()[axis];
1170
                m.replace_instruction(
1171
1172
1173
1174
                    arg,
                    make_op("slice",
                            {{"axes", {axis}}, {"starts", {offset}}, {"ends", {offset + len}}}),
                    fused);
1175
1176
1177
1178
1179
1180
1181
1182
1183
                offset += len;
            }
        };

        auto outputs = ins->outputs();
        group_by(outputs.begin(), outputs.end(), each, pred);
    }
};

1184
1185
1186
1187
1188
1189
1190
struct find_div_const
{
    auto matcher() const
    {
        return match::name("div")(match::arg(1)(match::is_constant().bind("c")));
    }

1191
    void apply(module& m, const match::matcher_result& r) const
1192
1193
1194
1195
    {
        auto ins   = r.result;
        auto c_ins = r.instructions["c"];

1196
        auto recip = m.insert_instruction(std::next(c_ins), make_op("recip"), c_ins);
1197
1198
1199

        auto args = ins->inputs();

1200
        m.replace_instruction(ins, make_op("mul"), args.front(), recip);
1201
1202
1203
    }
};

1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
struct find_unit_ops
{
    auto matcher() const
    {
        auto mul_1 = match::name("mul")(
            match::either_arg(0, 1)(match::has_value(1.0f), match::any().bind("x")));
        auto div_1 =
            match::name("div")(match::args(match::any().bind("x"), match::has_value(1.0f)));
        auto add_0 = match::name("add")(
            match::either_arg(0, 1)(match::has_value(0.0f, 1e-12), match::any().bind("x")));
        auto sub_0 =
            match::name("sub")(match::args(match::any().bind("x"), match::has_value(0.0f)));
        return match::any_of(mul_1, div_1, add_0, sub_0);
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins  = r.result;
        auto c_in = r.instructions["x"];

        m.replace_instruction(ins, c_in);
    }
};

struct find_neg_unit_ops
{
    auto matcher() const
    {
        auto mul_neg_1 = match::name("mul")(
            match::either_arg(0, 1)(match::has_value(-1.0f), match::any().bind("x")));
        auto div_neg_1 =
            match::name("div")(match::args(match::any().bind("x"), match::has_value(-1.0f)));
        auto sub_0 =
            match::name("sub")(match::args(match::has_value(0.0f), match::any().bind("x")));
        return match::any_of(mul_neg_1, div_neg_1, sub_0);
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins  = r.result;
        auto c_in = r.instructions["x"];

1246
        auto neg = m.insert_instruction(ins, make_op("neg"), c_in);
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
        m.replace_instruction(ins, neg);
    }
};

struct find_zero_ops
{
    auto matcher() const
    {
        auto mul_zero = match::name("mul")(
            match::either_arg(0, 1)(match::has_value(0.0f).bind("x"), match::any()));
        auto div_zero =
            match::name("div")(match::args(match::has_value(0.0f).bind("x"), match::any()));
        return match::any_of(mul_zero, div_zero);
    }

    void apply(module& m, const match::matcher_result& r) const
    {
        auto ins      = r.result;
        auto zero_ins = r.instructions["x"];

        m.replace_instruction(ins, zero_ins);
    }
};

1271
1272
1273
1274
1275
1276
1277
struct find_sub_const
{
    auto matcher() const
    {
        return match::name("sub")(match::arg(1)(match::is_constant().bind("c")));
    }

1278
    void apply(module& m, const match::matcher_result& r) const
1279
1280
1281
1282
    {
        auto ins   = r.result;
        auto c_ins = r.instructions["c"];

1283
        auto neg = m.insert_instruction(std::next(c_ins), make_op("neg"), c_ins);
1284
1285
1286

        auto args = ins->inputs();

1287
        m.replace_instruction(ins, make_op("add"), args.front(), neg);
1288
1289
1290
    }
};

kahmed10's avatar
kahmed10 committed
1291
1292
1293
1294
1295
1296
1297
1298
struct find_rsqrt
{
    auto matcher() const
    {
        return match::name("recip")(match::args(
            match::name("sqrt")(match::used_once(), match::args(match::any().bind("x")))));
    }

1299
    void apply(module& m, const match::matcher_result& r) const
kahmed10's avatar
kahmed10 committed
1300
1301
1302
1303
    {
        auto ins   = r.result;
        auto x_ins = r.instructions["x"];

1304
        m.replace_instruction(ins, make_op("rsqrt"), x_ins);
kahmed10's avatar
kahmed10 committed
1305
1306
1307
    }
};

1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
static bool same_ops(const std::vector<instruction_ref>& vec_ins)
{
    return std::all_of(vec_ins.begin(), vec_ins.end(), [&](auto i) {
        return i->get_operator() == vec_ins.front()->get_operator();
    });
}

struct find_split_reshape
{
    auto matcher() const
    {
        return match::name("reshape")(match::arg(0)(match::name("contiguous")(
                                          match::arg(0)(match::name("slice").bind("slice")))))
            .bind("reshape");
    }

1324
    void apply(module& m, const match::matcher_result& r) const
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
    {
        auto slc = r.instructions["slice"];
        auto rsp = r.instructions["reshape"];

        auto input         = slc->inputs().front();
        auto split_outputs = get_splits(input);
        if(split_outputs.empty())
        {
            return;
        }

shivadbhavsar's avatar
shivadbhavsar committed
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
        // Only want to apply this optimization if each split output is followed by
        // a contiguous op and a reshape
        if(std::any_of(split_outputs.begin(), split_outputs.end(), [](auto i) {
               if(i->outputs().size() == 1)
               {
                   auto cont = i->outputs().front();
                   return cont->outputs().size() != 1;
               }
               return false;
           }))
        {
            return;
        }

1350
1351
1352
1353
1354
1355
1356
1357
        std::vector<instruction_ref> vec_rsp(split_outputs.size());
        std::transform(split_outputs.begin(), split_outputs.end(), vec_rsp.begin(), [](auto i) {
            auto cont = i->outputs().front();
            return cont->outputs().front();
        });

        // all outputs are reshape and of the same shape
        auto dims = any_cast<op::reshape>(rsp->get_operator()).dims;
1358
        if(not same_ops(vec_rsp))
1359
1360
1361
1362
1363
        {
            return;
        }

        // ensure reshape happens after the axis dimension
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
        auto axis         = any_cast<op::slice>(slc->get_operator()).axes[0];
        auto slc_lens     = slc->get_shape().lens();
        auto slc_dim_size = std::accumulate(
            slc_lens.begin() + axis, slc_lens.end(), 1, std::multiplies<std::size_t>());

        // search the reshape output (standard shape) to decide which axis are
        // in its output corresponding to the slc_dim_size
        auto rsp_lens    = rsp->get_shape().lens();
        auto rsp_strides = rsp->get_shape().strides();
        rsp_strides.insert(rsp_strides.begin(), rsp_strides[0] * rsp_lens[0]);
1374
1375
1376

        auto ait     = std::find(rsp_strides.begin(), rsp_strides.end(), slc_dim_size);
        int rsp_axis = -1;
1377
        if(ait == rsp_strides.end())
1378
1379
1380
        {
            return;
        }
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
        else if(ait == rsp_strides.end() - 1)
        {
            // edge case
            // slice_dim == 1, in that case it could match with last stride of 1.
            // it should accumulate lengths from last dim in that case. discount 1 to avoid going
            // out of bounds.
            assert(slc_dim_size == 1);
            rsp_axis = std::distance(rsp_strides.begin(), ait) - 1;
        }
        else
        {
            rsp_axis = std::distance(rsp_strides.begin(), ait);
        }
1394
        // calculate reshape output shape
1395
        std::vector<int64_t> vec_dims(vec_rsp.size());
1396

1397
1398
1399
1400
1401
        std::transform(vec_rsp.begin(), vec_rsp.end(), vec_dims.begin(), [&](auto is) {
            return is->get_shape().lens()[rsp_axis];
        });

        std::vector<int64_t> rsp_out_lens(rsp_lens.begin(), rsp_lens.end());
1402

1403
        rsp_out_lens[rsp_axis] = std::accumulate(vec_dims.begin(), vec_dims.end(), std::int64_t{0});
1404

1405
1406
1407
1408
1409
        // insert the reshape instruction and add contiguous if needed
        if(not input->get_shape().standard())
        {
            input = m.insert_instruction(std::next(input), make_op("contiguous"), input);
        }
1410
        auto rsp_ins = m.insert_instruction(
1411
            std::next(input), make_op("reshape", {{"dims", rsp_out_lens}}), input);
1412
1413

        // replace the original reshape with slice
1414
1415
        int64_t start = 0;
        for(std::size_t i = 0; i < vec_rsp.size(); ++i)
1416
        {
1417
            m.replace_instruction(
1418
1419
1420
1421
1422
                vec_rsp[i],
                make_op(
                    "slice",
                    {{"axes", {rsp_axis}}, {"starts", {start}}, {"ends", {start + vec_dims[i]}}}),
                rsp_ins);
1423
            start += vec_dims[i];
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
        }
    }
};

struct find_split_transpose
{
    auto matcher() const
    {
        return match::name("transpose")(match::arg(0)(match::name("slice").bind("slice")))
            .bind("trans");
    }

1436
    void apply(module& m, const match::matcher_result& r) const
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
    {
        auto slc   = r.instructions["slice"];
        auto trans = r.instructions["trans"];

        auto input         = slc->inputs().front();
        auto split_outputs = get_splits(input);
        if(split_outputs.empty())
        {
            return;
        }

        std::vector<instruction_ref> vec_trans(split_outputs.size());
        std::transform(split_outputs.begin(), split_outputs.end(), vec_trans.begin(), [](auto i) {
            assert(i->outputs().size() == 1);
            return i->outputs().front();
        });

        // all transpose are the same
        auto perm = any_cast<op::transpose>(trans->get_operator()).dims;
1456
        if(not same_ops(vec_trans))
1457
1458
1459
1460
1461
        {
            return;
        }

        // insert an transpose instruction
1462
        auto tr = m.insert_instruction(
1463
            std::next(input), make_op("transpose", {{"permutation", perm}}), input);
1464
1465
1466
1467
1468

        // compute the axis in the slice
        auto axis = any_cast<op::slice>(slc->get_operator()).axes.front();
        auto it   = std::find(perm.begin(), perm.end(), axis);
        assert(it != perm.end());
Paul Fultz II's avatar
Paul Fultz II committed
1469
        int64_t axis_new = std::distance(perm.begin(), it);
1470
1471
1472
1473
1474
1475
1476

        for(auto in : split_outputs)
        {
            auto oper    = any_cast<op::slice>(in->get_operator());
            auto starts  = oper.starts;
            auto ends    = oper.ends;
            auto tr_orig = in->outputs().front();
1477
            m.replace_instruction(
1478
1479
1480
                tr_orig,
                make_op("slice", {{"axes", {axis_new}}, {"starts", starts}, {"ends", ends}}),
                tr);
1481
1482
1483
1484
        }
    }
};

1485
void simplify_algebra::apply(module& m) const
Paul's avatar
Paul committed
1486
{
Paul's avatar
Paul committed
1487
    // Run simplifications multiple times
1488
    for(int i = 0; i < 8; i++)
Paul's avatar
Paul committed
1489
    {
1490
        match::find_matches(m,
Paul's avatar
Paul committed
1491
                            find_inner_broadcast{},
1492
                            find_dot_broadcast{},
Paul's avatar
Paul committed
1493
1494
                            find_double_add_lit_broadcast{},
                            find_add_lit_broadcast{},
1495
                            find_add_convs{},
1496
                            find_conv_dot_horiz_fusion{},
Paul's avatar
Paul committed
1497
                            find_mul_conv{},
1498
                            find_mul_slice_conv{},
1499
1500
                            find_mul_dot{},
                            find_dot_mul{},
1501
                            find_mul_add{},
1502
1503
1504
                            find_unit_ops{},
                            find_neg_unit_ops{},
                            find_zero_ops{},
1505
                            find_dot_add{},
1506
                            find_conv_add{},
1507
1508
                            find_div_const{},
                            find_sub_const{},
kahmed10's avatar
kahmed10 committed
1509
                            find_rsqrt{},
1510
                            find_concat_op{},
1511
                            find_split_concat{},
1512
1513
1514
                            find_splits{},
                            find_split_reshape{},
                            find_split_transpose{});
1515
        dead_code_elimination{}.apply(m);
Paul's avatar
Paul committed
1516
    }
Paul's avatar
Paul committed
1517
}
Paul's avatar
Paul committed
1518

Paul's avatar
Paul committed
1519
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1520
} // namespace migraphx