simplify_algebra.cpp 47.4 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
// ******************************
//  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
213
214
215
216
217
struct find_mul_add
{
    auto matcher() const
    {
        return match::name("mul")(match::either_arg(0, 1)(
Paul's avatar
Paul committed
218
219
220
            match::name("add")(
                match::either_arg(0, 1)(
                    match::any().bind("x"),
Paul's avatar
Paul committed
221
                    match::any_of(conv_const_weights(), match::is_constant()).bind("b")),
Paul's avatar
Paul committed
222
                match::none_of(match::args(match::is_constant(), match::is_constant())),
Paul's avatar
Paul committed
223
                match::used_once()),
Paul's avatar
Paul committed
224
            match::is_constant().bind("a")));
Paul's avatar
Paul committed
225
226
    }

227
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
228
    {
Paul's avatar
Paul committed
229
        auto ins   = r.result;
Paul's avatar
Paul committed
230
        auto a_ins = r.instructions["a"];
Paul's avatar
Paul committed
231
        auto b_ins = r.instructions["b"];
Paul's avatar
Paul committed
232
        auto x_ins = r.instructions["x"];
Paul's avatar
Paul committed
233
        assert(x_ins != b_ins);
Paul's avatar
Paul committed
234

235
236
237
        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
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
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);
    }
};

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
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
303
struct find_add_lit_broadcast
Paul's avatar
Paul committed
304
305
306
307
308
309
310
{
    auto matcher() const
    {
        return match::name("add")(
            match::either_arg(0, 1)(op_lit_broadcast("add", "a", "x"), lit_broadcast().bind("b")));
    }

311
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
312
313
314
315
316
317
    {
        auto ins   = r.result;
        auto x_ins = r.instructions["x"];
        auto a_ins = r.instructions["a"];
        auto b_ins = r.instructions["b"];

318
319
        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
320
321
322
323
    }
};

struct find_double_add_lit_broadcast
Paul's avatar
Paul committed
324
{
Paul's avatar
Paul committed
325
326
    auto matcher() const
    {
Paul's avatar
Paul committed
327
        return match::name("add")(
Paul's avatar
Paul committed
328
            match::args(op_lit_broadcast("add", "a", "x"), op_lit_broadcast("add", "b", "y")));
Paul's avatar
Paul committed
329
330
    }

331
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
332
    {
Paul's avatar
Paul committed
333
334
335
336
337
        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
338
339
340

        instruction_ref sumab;

Paul's avatar
Paul committed
341
        if(a_ins->name() == "broadcast" and b_ins->name() == "broadcast")
Paul's avatar
Paul committed
342
343
344
        {
            if(a_ins->inputs().at(0)->get_shape() != b_ins->inputs().at(0)->get_shape())
                return;
345
            auto op     = a_ins->get_operator();
346
            auto presum = m.insert_instruction(
347
                ins, make_op("add"), a_ins->inputs().at(0), b_ins->inputs().at(0));
348
            sumab = m.insert_instruction(ins, op, presum);
Paul's avatar
Paul committed
349
350
351
        }
        else
        {
352
            sumab = m.insert_instruction(ins, make_op("add"), a_ins, b_ins);
Paul's avatar
Paul committed
353
354
        }

355
356
        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
357
358
359
    }
};

Paul's avatar
Paul committed
360
361
struct find_inner_broadcast
{
362
    auto matcher() const { return pointwise(match::all_of[match::inputs()](match::broadcast())); }
Paul's avatar
Paul committed
363

364
    void apply(module& m, const match::matcher_result& r) const
Paul's avatar
Paul committed
365
    {
366
367
368
369
370
371
372
373
374
375
        auto ins        = r.result;
        auto broadcasts = ins->inputs();
        if(broadcasts.empty())
            return;
        std::vector<instruction_ref> inputs;
        std::transform(broadcasts.begin(),
                       broadcasts.end(),
                       std::back_inserter(inputs),
                       [](auto i) { return i->inputs().front(); });
        if(std::any_of(inputs.begin(), inputs.end(), [&](auto i) {
376
377
               return i->get_shape() != inputs.front()->get_shape() and
                      i->get_shape().elements() != 1;
378
           }))
Paul's avatar
Paul committed
379
380
            return;

381
382
383
384
385
386
387
        auto b_it = std::find_if(broadcasts.begin(), broadcasts.end(), [&](auto i) {
            return not i->get_shape().scalar();
        });
        if(b_it == broadcasts.end())
            b_it = broadcasts.begin();
        auto op = insert_common_op(m, ins, ins->get_operator(), inputs);
        m.replace_instruction(ins, (*b_it)->get_operator(), op);
Paul's avatar
Paul committed
388
389
390
    }
};

Paul's avatar
Paul committed
391
392
393
394
395
396
397
398
399
400
401
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
428
429
430
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;
        // Find leading batch axes that are broadcasted
        auto p = std::mismatch(a->get_shape().strides().begin(), a->get_shape().strides().begin()+nbatch_axes, b->get_shape().strides().begin(), b->get_shape().strides().begin()+nbatch_axes, [](auto astride, auto bstride) {
            return astride == 0 and bstride == 0;
        });
        auto naxes = p.first - a->get_shape().lens().begin();
        std::vector<std::size_t> axes(naxes);
        std::iota(axes.begin(), axes.end(), 0);

        auto insert_sqeeze = [&](instruction_ref b_ins) {
            auto input = b_ins->inputs()[0];
            auto delta = b_ins->get_shape().lens().size() - input->get_shape().lens().size();
            auto squeeze_axes = axes;
            squeeze_axes.erase(squeeze_axes.end() - delta, squeeze_axes.end());
            if (squeeze_axes.empty())
                return input;
            return m.insert_instruction(ins, make_op("squeeze", {{"axes", squeeze_axes}}), input);
        };
        auto a1 = insert_sqeeze(a);
        auto b1 = insert_sqeeze(b);
        auto dot = m.insert_instruction(ins, make_op("dot"), a1, b1);
        auto unsqueeze = m.insert_instruction(ins, make_op("unsqueeze", {{"axes", axes}}), dot);
        auto broadcast = m.insert_instruction(ins, make_op("multibroadcast", {{"out_lens", ins->get_shape().lens()}}), unsqueeze);
        m.replace_instruction(ins, broadcast);
    }
};

431
struct find_concat_op
432
433
434
{
    auto matcher() const
    {
435
        return match::name("concat")(match::any_of[match::inputs()](
Paul's avatar
Paul committed
436
            match::any_of(match::pointwise(), match::name("broadcast", "multibroadcast")), match::used_once()));
437
438
    }

439
440
    template <class Iterator>
    static std::vector<std::size_t> get_output_lens(Iterator start, Iterator last, std::size_t axis)
441
    {
442
443
444
        assert(start != last);
        std::size_t dim = 0;
        for(auto ins : range(start, last))
445
        {
446
            dim += ins->get_shape().lens().at(axis);
447
        }
448
449
450
        auto lens  = (*start)->get_shape().lens();
        lens[axis] = dim;
        return lens;
451
452
    }

453
454
    static bool is_valid_op(const operation& op)
    {
Paul's avatar
Paul committed
455
        return contains({"broadcast", "multibroadcast"}, op.name()) or op.attributes().contains("pointwise");
456
457
    }

458
    void apply(module& m, const match::matcher_result& r) const
459
    {
460
461
        auto ins  = r.result;
        auto axis = any_cast<op::concat>(ins->get_operator()).axis;
462

463
464
465
466
467
468
        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};
469
470
            auto op = x->get_operator();
            if(not is_valid_op(op))
471
472
473
474
475
476
477
478
479
480
481
482
                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;
            }
Paul's avatar
Paul committed
483
484
485
486
487
488
489
490
491
492
            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;
            }
493
494
495
496
497
498
499
500

            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);
                });
501
                auto concat =
502
                    m.insert_instruction(ins, make_op("concat", {{"axis", iaxis}}), inputs);
503
504
                concats.push_back(concat);
            }
505
            auto y = m.insert_instruction(ins, op, concats);
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
            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)
521
            m.replace_instruction(ins, args.front());
522
        else
523
            m.replace_instruction(ins, make_op("concat", {{"axis", axis}}), args);
524
525
526
    }
};

527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
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);
    }
}

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
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
    {
583
584
585
        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()))));
586
587
    }

Shucai Xiao's avatar
Shucai Xiao committed
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
    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);
    }

607
    static std::vector<std::vector<instruction_ref>>
Shucai Xiao's avatar
Shucai Xiao committed
608
    get_split_groups(const module& m, const std::vector<instruction_ref>& splits)
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
    {
        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
625

626
                // If there is a duplicate bail
Shucai Xiao's avatar
Shucai Xiao committed
627
628
629
630
631
                // 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);
                   }))
                {
632
                    return {};
Shucai Xiao's avatar
Shucai Xiao committed
633
634
                }

635
636
637
638
639
640
641
642
643
                group.push_back(*it);
            }
            if(group.size() != splits.size())
                continue;
            groups.push_back(group);
        }
        return groups;
    }

Shucai Xiao's avatar
Shucai Xiao committed
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
    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;
    }

669
    void apply(module& m, const match::matcher_result& r) const
670
    {
Shucai Xiao's avatar
Shucai Xiao committed
671
        auto ins    = r.result;
672
673
674
        auto splits = get_splits(ins);
        if(splits.empty())
            return;
Shucai Xiao's avatar
Shucai Xiao committed
675

676
        for(const auto& group : get_split_groups(m, splits))
677
        {
Shucai Xiao's avatar
Shucai Xiao committed
678
679
680
681
682
            auto start       = group.front();
            auto split_front = splits.front();
            auto op          = start->get_operator();
            if(not is_fusable(start, split_front))
            {
683
                continue;
Shucai Xiao's avatar
Shucai Xiao committed
684
            }
685
686
687
688
689
690

            // 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;
691
            instruction_ref c = m.end();
692
693
            if(start->inputs().size() == 1)
            {
694
                c = m.insert_instruction(std::next(ins), op, ins);
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
            }
            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;

720
                move_instructions_back(m, ins, data_args);
721
722
723
724
725
726
727

                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
728
                auto concat = m.insert_instruction(
729
                    ins, make_op("concat", {{"axis", concat_axis}}), data_args);
730
731
732
733
734

                std::vector<instruction_ref> args;
                args.resize(2);
                args[split_idx] = ins;
                args[data_idx]  = concat;
735
                c               = m.insert_instruction(std::next(ins), op, args);
736
            }
737
            if(c != m.end())
738
739
740
741
742
743
            {
                for(auto i : group)
                {
                    auto split = i->inputs()[split_idx];
                    assert(split->name() == "slice");
                    // Insert contiguous for reshapes
744
745
                    auto outputs = i->outputs();
                    for(auto output : outputs)
746
                    {
747
                        if(output->name() != "reshape")
748
                            continue;
749
                        auto x = m.insert_instruction(output, make_op("contiguous"), i);
750
                        m.replace_instruction(output, output->get_operator(), x);
751
752
                    }

753
                    m.replace_instruction(i, split->get_operator(), c);
754
755
756
757
758
759
760
761
762
763
764
765
766
767
                }
            }
        }
    }
};

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")))));
    }

768
    void apply(module& m, const match::matcher_result& r) const
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
    {
        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;
797
798
799
800
801
802
803
        // 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;
804
805
806
807
        *it = splits.front()->inputs().front();
        args.erase(std::next(it), it + splits.size());

        if(args.size() == 1)
808
            m.replace_instruction(concat, args.front());
809
        else
810
            m.replace_instruction(concat, concat->get_operator(), args);
811
812
813
    }
};

814
815
816
817
818
819
820
821
822
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
848
849
850
851
852
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];
    }

853
    void apply(module& m, const match::matcher_result& r) const
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
    {
        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;
882
                    b_input = m.insert_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
883
                        ins, make_op("step", {{"axes", {2, 3}}, {"steps", {n, n}}}), b_input);
884
885
886
887
888
889
890
                }
                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;
891
                    a_input = m.insert_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
892
                        ins, make_op("step", {{"axes", {2, 3}}, {"steps", {n, n}}}), a_input);
893
894
895
896
897
898
899
900
                }
                else
                    return;
            }
            else
                return;
        }

901
        auto concat_input =
902
            m.insert_instruction(ins, make_op("concat", {{"axis", 1}}), a_input, b_input);
903
        auto concat_weights =
904
905
            m.insert_instruction(ins, make_op("concat", {{"axis", 1}}), a_weights, b_weights);
        m.replace_instruction(ins, new_op, concat_input, concat_weights);
906
907
908
    }
};

909
910
911
912
913
914
915
916
917
918
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"));
919
    return (dots >= 2 or convs >= 2);
920
921
922
923
924
925
}

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

926
    void apply(module& m, const match::matcher_result& r) const
927
928
929
930
931
932
933
934
935
936
937
938
    {
        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;
939
            // Check that non-axes match
940
941
942
943
944
945
946
947
948
949
950
951
952
953
            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;
954
955
956
957
958
959
960
            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;
961
962
963
964
965
966
967
968
969
970
971
972
            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;
            }

973
            move_instructions_back(m, input, args);
974
            // TODO: Check if axes match
975
            auto concat =
976
977
                m.insert_instruction(input, make_op("concat", {{"axis", concat_axis}}), args);
            auto fused     = m.insert_instruction(std::next(input), op, input, concat);
978
979
980
            int64_t offset = 0;
            for(auto arg : range(start, last))
            {
981
982
983
984
985
986
987
988
989
                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);
                }

990
                int64_t len = arg->get_shape().lens()[axis];
991
                m.replace_instruction(
992
993
994
995
                    arg,
                    make_op("slice",
                            {{"axes", {axis}}, {"starts", {offset}}, {"ends", {offset + len}}}),
                    fused);
996
997
998
999
1000
1001
1002
1003
1004
                offset += len;
            }
        };

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

1005
1006
1007
1008
1009
1010
1011
struct find_div_const
{
    auto matcher() const
    {
        return match::name("div")(match::arg(1)(match::is_constant().bind("c")));
    }

1012
    void apply(module& m, const match::matcher_result& r) const
1013
1014
1015
1016
    {
        auto ins   = r.result;
        auto c_ins = r.instructions["c"];

1017
        auto recip = m.insert_instruction(std::next(c_ins), make_op("recip"), c_ins);
1018
1019
1020

        auto args = ins->inputs();

1021
        m.replace_instruction(ins, make_op("mul"), args.front(), recip);
1022
1023
1024
    }
};

1025
1026
1027
1028
1029
1030
1031
1032
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
1061
1062
1063
1064
1065
1066
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"];

1067
        auto neg = m.insert_instruction(ins, make_op("neg"), c_in);
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
        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);
    }
};

1092
1093
1094
1095
1096
1097
1098
struct find_sub_const
{
    auto matcher() const
    {
        return match::name("sub")(match::arg(1)(match::is_constant().bind("c")));
    }

1099
    void apply(module& m, const match::matcher_result& r) const
1100
1101
1102
1103
    {
        auto ins   = r.result;
        auto c_ins = r.instructions["c"];

1104
        auto neg = m.insert_instruction(std::next(c_ins), make_op("neg"), c_ins);
1105
1106
1107

        auto args = ins->inputs();

1108
        m.replace_instruction(ins, make_op("add"), args.front(), neg);
1109
1110
1111
    }
};

kahmed10's avatar
kahmed10 committed
1112
1113
1114
1115
1116
1117
1118
1119
struct find_rsqrt
{
    auto matcher() const
    {
        return match::name("recip")(match::args(
            match::name("sqrt")(match::used_once(), match::args(match::any().bind("x")))));
    }

1120
    void apply(module& m, const match::matcher_result& r) const
kahmed10's avatar
kahmed10 committed
1121
1122
1123
1124
    {
        auto ins   = r.result;
        auto x_ins = r.instructions["x"];

1125
        m.replace_instruction(ins, make_op("rsqrt"), x_ins);
kahmed10's avatar
kahmed10 committed
1126
1127
1128
    }
};

1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
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");
    }

1145
    void apply(module& m, const match::matcher_result& r) const
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
    {
        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
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
        // 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;
        }

1171
1172
1173
1174
1175
1176
1177
1178
        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;
1179
        if(not same_ops(vec_rsp))
1180
1181
1182
1183
1184
        {
            return;
        }

        // ensure reshape happens after the axis dimension
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
        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]);
1195
1196
1197

        auto ait     = std::find(rsp_strides.begin(), rsp_strides.end(), slc_dim_size);
        int rsp_axis = -1;
1198
        if(ait == rsp_strides.end())
1199
1200
1201
        {
            return;
        }
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
        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);
        }
1215
        // calculate reshape output shape
1216
        std::vector<int64_t> vec_dims(vec_rsp.size());
1217

1218
1219
1220
1221
1222
        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());
1223

1224
        rsp_out_lens[rsp_axis] = std::accumulate(vec_dims.begin(), vec_dims.end(), std::int64_t{0});
1225

1226
1227
1228
1229
1230
        // 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);
        }
1231
        auto rsp_ins = m.insert_instruction(
1232
            std::next(input), make_op("reshape", {{"dims", rsp_out_lens}}), input);
1233
1234

        // replace the original reshape with slice
1235
1236
        int64_t start = 0;
        for(std::size_t i = 0; i < vec_rsp.size(); ++i)
1237
        {
1238
            m.replace_instruction(
1239
1240
1241
1242
1243
                vec_rsp[i],
                make_op(
                    "slice",
                    {{"axes", {rsp_axis}}, {"starts", {start}}, {"ends", {start + vec_dims[i]}}}),
                rsp_ins);
1244
            start += vec_dims[i];
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
        }
    }
};

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

1257
    void apply(module& m, const match::matcher_result& r) const
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
    {
        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;
1277
        if(not same_ops(vec_trans))
1278
1279
1280
1281
1282
        {
            return;
        }

        // insert an transpose instruction
1283
        auto tr = m.insert_instruction(
1284
            std::next(input), make_op("transpose", {{"permutation", perm}}), input);
1285
1286
1287
1288
1289

        // 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
1290
        int64_t axis_new = std::distance(perm.begin(), it);
1291
1292
1293
1294
1295
1296
1297

        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();
1298
            m.replace_instruction(
1299
1300
1301
                tr_orig,
                make_op("slice", {{"axes", {axis_new}}, {"starts", starts}, {"ends", ends}}),
                tr);
1302
1303
1304
1305
        }
    }
};

1306
void simplify_algebra::apply(module& m) const
Paul's avatar
Paul committed
1307
{
Paul's avatar
Paul committed
1308
    // Run simplifications multiple times
1309
    for(int i = 0; i < 8; i++)
Paul's avatar
Paul committed
1310
    {
1311
        match::find_matches(m,
Paul's avatar
Paul committed
1312
                            find_inner_broadcast{},
Paul's avatar
Paul committed
1313
1314
                            find_double_add_lit_broadcast{},
                            find_add_lit_broadcast{},
1315
                            find_add_convs{},
1316
                            find_conv_dot_horiz_fusion{},
Paul's avatar
Paul committed
1317
                            find_mul_conv{},
1318
                            find_mul_slice_conv{},
1319
                            find_mul_add{},
1320
1321
1322
                            find_unit_ops{},
                            find_neg_unit_ops{},
                            find_zero_ops{},
1323
                            find_dot_add{},
1324
                            find_conv_add{},
1325
1326
                            find_div_const{},
                            find_sub_const{},
kahmed10's avatar
kahmed10 committed
1327
                            find_rsqrt{},
1328
                            find_concat_op{},
1329
                            find_split_concat{},
1330
1331
1332
                            find_splits{},
                            find_split_reshape{},
                            find_split_transpose{});
1333
        dead_code_elimination{}.apply(m);
Paul's avatar
Paul committed
1334
    }
Paul's avatar
Paul committed
1335
}
Paul's avatar
Paul committed
1336

Paul's avatar
Paul committed
1337
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1338
} // namespace migraphx