program.cpp 36.4 KB
Newer Older
1
#include "migraphx/instruction_ref.hpp"
2
#include <functional>
Paul's avatar
Paul committed
3
4
5
#include <migraphx/program.hpp>
#include <migraphx/stringutils.hpp>
#include <migraphx/instruction.hpp>
6
#include <migraphx/op/identity.hpp>
Paul's avatar
Paul committed
7
#include <migraphx/target.hpp>
Paul's avatar
Paul committed
8
9
10
#include <migraphx/env.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/time.hpp>
11
#include <migraphx/pass_manager.hpp>
12
#include <migraphx/register_target.hpp>
Shucai Xiao's avatar
Shucai Xiao committed
13
#include <migraphx/iterator_for.hpp>
14
#include <migraphx/iterator.hpp>
15
#include <migraphx/algorithm.hpp>
16
#include <migraphx/output_iterator.hpp>
Shucai Xiao's avatar
Shucai Xiao committed
17
#include <migraphx/make_op.hpp>
18
#include <migraphx/marker.hpp>
Paul's avatar
Paul committed
19
#include <iostream>
20
#include <numeric>
Paul's avatar
Paul committed
21
#include <sstream>
Paul's avatar
Paul committed
22
#include <algorithm>
23
#include <set>
Paul's avatar
Paul committed
24
#include <utility>
25
#include <iomanip>
26

27
#include <unordered_set>
Shucai Xiao's avatar
Shucai Xiao committed
28
29
#include <map>
#include <cassert>
Paul's avatar
Paul committed
30

Paul's avatar
Paul committed
31
namespace migraphx {
Paul's avatar
Paul committed
32
inline namespace MIGRAPHX_INLINE_NS {
Paul's avatar
Paul committed
33

34
35
using milliseconds = std::chrono::duration<double, std::milli>;

Paul's avatar
Paul committed
36
37
struct program_impl
{
Shucai Xiao's avatar
Shucai Xiao committed
38
    // A map is used to keep references to modules of the program
39
    std::unordered_map<std::string, module> modules;
Paul's avatar
Paul committed
40
    context ctx;
41
    std::string target_name;
Paul's avatar
Paul committed
42
43
};

44
program::program() : impl(std::make_unique<program_impl>()) { this->create_module("main"); }
Paul's avatar
Paul committed
45

Paul's avatar
Paul committed
46
program::program(program&&) noexcept = default;
Shucai Xiao's avatar
Shucai Xiao committed
47
program::~program() noexcept         = default;
Paul's avatar
Paul committed
48

49
// copy constructor
Shucai Xiao's avatar
Shucai Xiao committed
50
program::program(const program& p) { assign(p); }
51
52

// copy assignment operator
Shucai Xiao's avatar
Shucai Xiao committed
53
program& program::operator=(program p)
54
{
Shucai Xiao's avatar
Shucai Xiao committed
55
    std::swap(p.impl, this->impl);
56
57
58
    return *this;
}

Shucai Xiao's avatar
Shucai Xiao committed
59
void program::assign(const program& p)
60
{
Shucai Xiao's avatar
Shucai Xiao committed
61
    if(!impl)
62
63
64
    {
        impl = std::make_unique<program_impl>();
    }
Shucai Xiao's avatar
Shucai Xiao committed
65
    else if(!impl->modules.empty())
66
    {
Shucai Xiao's avatar
Shucai Xiao committed
67
        impl->modules.clear();
68
    }
Shucai Xiao's avatar
Shucai Xiao committed
69

70
    impl->ctx         = p.impl->ctx;
Shucai Xiao's avatar
Shucai Xiao committed
71
72
    impl->target_name = p.impl->target_name;
    impl->modules     = p.impl->modules;
Shucai Xiao's avatar
Shucai Xiao committed
73
74
75
76

    // build a map from old ins to new ins
    // Build a map from old module to new module
    std::unordered_map<module_ref, module_ref> mod_map;
Paul's avatar
Paul committed
77
78
79
80
81
    std::transform(
        impl->modules.begin(),
        impl->modules.end(),
        std::inserter(mod_map, mod_map.begin()),
        [&](auto&& xp) { return std::make_pair(&p.impl->modules.at(xp.first), &xp.second); });
Shucai Xiao's avatar
Shucai Xiao committed
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97

    std::unordered_map<instruction_ref, instruction_ref> ins_map;
    for(auto&& pp : mod_map)
    {
        auto old_ins = iterator_for(*pp.first);
        auto new_ins = iterator_for(*pp.second);
        std::transform(old_ins.begin(),
                       old_ins.end(),
                       new_ins.begin(),
                       std::inserter(ins_map, ins_map.begin()),
                       [](auto x, auto y) { return std::make_pair(x, y); });
    }

    // Update all references from all modules
    for(auto&& mp : impl->modules)
    {
98
        for(auto ins : iterator_for(mp.second))
Shucai Xiao's avatar
Shucai Xiao committed
99
100
            instruction::replace_refs(ins, ins_map, mod_map);
    }
101
102
}

Paul's avatar
Paul committed
103
shape program::get_parameter_shape(std::string name) const
Paul's avatar
Paul committed
104
{
Shucai Xiao's avatar
Shucai Xiao committed
105
106
    const auto* mm = this->get_main_module();
    return mm->get_parameter_shape(std::move(name));
Paul's avatar
Paul committed
107
108
}

109
110
std::vector<std::string> program::get_parameter_names() const
{
Shucai Xiao's avatar
Shucai Xiao committed
111
112
    const auto* mm = this->get_main_module();
    return mm->get_parameter_names();
113
114
}

mei-ye's avatar
mei-ye committed
115
116
instruction_ref program::get_parameter(std::string name) const
{
Shucai Xiao's avatar
Shucai Xiao committed
117
118
    const auto* mm = this->get_main_module();
    return mm->get_parameter(std::move(name));
mei-ye's avatar
mei-ye committed
119
120
}

Paul's avatar
Paul committed
121
122
std::unordered_map<std::string, shape> program::get_parameter_shapes() const
{
Shucai Xiao's avatar
Shucai Xiao committed
123
124
    const auto* mm = this->get_main_module();
    return mm->get_parameter_shapes();
Paul's avatar
Paul committed
125
126
}

Shucai Xiao's avatar
Shucai Xiao committed
127
std::size_t program::size() const { return impl->modules.size(); }
128

129
130
std::vector<shape> program::get_output_shapes() const
{
Shucai Xiao's avatar
Shucai Xiao committed
131
132
    const auto* mm = this->get_main_module();
    return mm->get_output_shapes();
133
}
Paul's avatar
Paul committed
134

Paul's avatar
Paul committed
135
136
context& program::get_context() const { return impl->ctx; }

Paul's avatar
Paul committed
137
138
instruction_ref program::validate() const
{
Shucai Xiao's avatar
Shucai Xiao committed
139
140
    const auto* mm = this->get_main_module();
    return mm->validate();
Paul's avatar
Paul committed
141
142
}

143
144
bool program::is_compiled() const { return not this->impl->target_name.empty(); }

145
void program::compile(const target& t, compile_options options)
Paul's avatar
Paul committed
146
{
147
148
149
    assert(not this->is_compiled());
    this->impl->target_name = t.name();
    this->impl->ctx         = t.get_context();
Paul's avatar
Paul committed
150
    if(enabled(MIGRAPHX_TRACE_COMPILE{}))
151
        options.trace = tracer{std::cout};
Shucai Xiao's avatar
Shucai Xiao committed
152

153
154
    options.trace(*this);
    options.trace();
Shucai Xiao's avatar
Shucai Xiao committed
155

Shucai Xiao's avatar
Shucai Xiao committed
156
    auto&& passes = t.get_passes(this->impl->ctx, options);
157
158
159
    run_passes(*this, passes, options.trace);

    auto mods = this->get_modules();
Shucai Xiao's avatar
Shucai Xiao committed
160

161
162
    // Validate and finalize
    for(const auto& mod : reverse(mods))
Paul's avatar
Paul committed
163
    {
Shucai Xiao's avatar
Shucai Xiao committed
164
165
166
167
168
169
        auto invalid = mod->validate();
        if(invalid != mod->end())
        {
            MIGRAPHX_THROW("Invalid module " + mod->name() + " from compilation at instruction " +
                           std::to_string(std::distance(mod->begin(), invalid)));
        }
170
171
172
173
174
        auto dangling = mod->find_dangling_reference();
        if(dangling != mod->end())
        {
            auto index = std::distance(mod->begin(), dangling);
            MIGRAPHX_THROW("Dangling reference in module " + mod->name() + " from instruction " +
175
                           std::to_string(index) + ", (" + dangling->name() + ")");
176
        }
Shucai Xiao's avatar
Shucai Xiao committed
177
        mod->finalize(this->impl->ctx);
Paul's avatar
Paul committed
178
    }
Paul's avatar
Paul committed
179
180
181
182
}

void program::finalize()
{
Shucai Xiao's avatar
Shucai Xiao committed
183
184
    auto* mm = this->get_main_module();
    mm->finalize(this->impl->ctx);
Paul's avatar
Paul committed
185
186
}

187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
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
template <class T>
std::string classify(T x)
{
    switch(std::fpclassify(x))
    {
    case FP_INFINITE: return "inf";
    case FP_NAN: return "nan";
    case FP_NORMAL: return "normal";
    case FP_SUBNORMAL: return "subnormal";
    case FP_ZERO: return "zero";
    default: return "unknown";
    }
}

std::unordered_set<std::string> classify_argument(const argument& a)
{
    std::unordered_set<std::string> result;
    a.visit(
        [&](auto t) {
            for(const auto& x : t)
                result.insert(classify(x));
        },
        [&](const auto& xs) {
            for(const auto& x : xs)
            {
                auto r = classify_argument(x);
                result.insert(r.begin(), r.end());
            }
        });
    return result;
}

void preview_argument(std::ostream& os, const argument& a)
{
    a.visit(
        [&](auto t) {
            if(t.size() <= 10)
            {
                os << t;
            }
            else
            {
                os << to_string_range(t.begin(), t.begin() + 5);
                os << ", ..., ";
                os << to_string_range(t.end() - 5, t.end());
            }
        },
        [&](const auto& xs) {
            for(const auto& x : xs)
            {
                os << '{';
                preview_argument(os, x);
                os << '}';
            }
        });
}

Paul's avatar
Paul committed
244
template <class F>
Shucai Xiao's avatar
Shucai Xiao committed
245
std::vector<argument> generic_eval(const module* mod,
246
247
                                   context& ctx,
                                   std::unordered_map<std::string, argument> params,
Shucai Xiao's avatar
Shucai Xiao committed
248
                                   std::unordered_map<instruction_ref, argument> results,
249
                                   F make_trace)
Paul's avatar
Paul committed
250
{
Shucai Xiao's avatar
Shucai Xiao committed
251
252
    assert(mod->validate() == mod->end());
    results.reserve(mod->size() * 2);
Paul's avatar
Paul committed
253
254
    std::vector<argument> values;
    values.reserve(16);
255
    auto trace = make_trace(mod);
Shucai Xiao's avatar
Shucai Xiao committed
256
    for(auto ins : iterator_for(*mod))
Paul's avatar
Paul committed
257
    {
258
        assert(results.find(ins) == results.end());
259
260
        const auto& name = ins->name();
        if(name == "@literal")
Paul's avatar
Paul committed
261
        {
Paul's avatar
Paul committed
262
            results.emplace(ins, trace(ins, [&] { return ins->get_literal().get_argument(); }));
Paul's avatar
Paul committed
263
        }
264
        else if(name == "@param")
Paul's avatar
Paul committed
265
        {
Paul's avatar
Paul committed
266
267
268
269
270
            results.emplace(
                ins, trace(ins, [&] {
                    auto param_name = any_cast<builtin::param>(ins->get_operator()).parameter;
                    if(not contains(params, param_name))
                        MIGRAPHX_THROW("Parameter not found: " + param_name);
271
                    auto param = params[param_name];
Paul's avatar
Paul committed
272
273
274
275
276
                    if(param.get_shape() != ins->get_shape())
                        MIGRAPHX_THROW("Incorrect shape {" + to_string(param.get_shape()) +
                                       "} for parameter: " + param_name);
                    return param;
                }));
Paul's avatar
Paul committed
277
        }
278
        else if(name == "@outline")
Paul's avatar
Paul committed
279
        {
Paul's avatar
Paul committed
280
            results.emplace(ins, trace(ins, [&] { return argument{ins->get_shape(), nullptr}; }));
Paul's avatar
Paul committed
281
        }
282
283
284
285
286
287
288
289
290
291
292
293
294
        else if(name == "@return")
        {
            std::vector<argument> prog_outputs;
            std::transform(ins->inputs().begin(),
                           ins->inputs().end(),
                           std::back_inserter(prog_outputs),
                           [&](instruction_ref i) {
                               assert(results.find(i) != results.end());
                               return results[i];
                           });

            return prog_outputs;
        }
Paul's avatar
Paul committed
295
296
        else
        {
Paul's avatar
Paul committed
297
            values.resize(ins->inputs().size());
Paul's avatar
Paul committed
298
299
300
301
302
            std::transform(
                ins->inputs().begin(), ins->inputs().end(), values.begin(), [&](instruction_ref i) {
                    assert(results.find(i) != results.end());
                    return results[i];
                });
Shucai Xiao's avatar
Shucai Xiao committed
303
304
305
306

            const auto& mod_args = ins->module_inputs();
            auto module_eval     = [&](module_ref smod,
                                   const std::unordered_map<std::string, argument>& inputs) {
Shucai Xiao's avatar
Shucai Xiao committed
307
308
                auto ssctx = ctx;
                return generic_eval(smod, ssctx, inputs, results, make_trace);
Shucai Xiao's avatar
Shucai Xiao committed
309
310
            };

Shucai Xiao's avatar
Shucai Xiao committed
311
312
313
314
            results.emplace(ins, trace(ins, [&] {
                                return ins->normalized_operator().compute(
                                    ctx, ins->get_shape(), values, mod_args, module_eval);
                            }));
Paul's avatar
Paul committed
315
        }
316
        assert(results.find(ins) != results.end());
317
        assert(results.at(ins).get_shape() == ins->get_shape());
Paul's avatar
Paul committed
318
    }
Shucai Xiao's avatar
Shucai Xiao committed
319
    return {results.at(std::prev(mod->end()))};
Paul's avatar
Paul committed
320
321
}

Shucai Xiao's avatar
Shucai Xiao committed
322
323
324
325
template <class F>
std::vector<argument> generic_eval(const program& p,
                                   context& ctx,
                                   std::unordered_map<std::string, argument> params,
326
                                   F make_trace)
Shucai Xiao's avatar
Shucai Xiao committed
327
{
Shucai Xiao's avatar
Shucai Xiao committed
328
    const module* mm = p.get_main_module();
329
    return generic_eval(mm, ctx, params, {}, make_trace);
Shucai Xiao's avatar
Shucai Xiao committed
330
331
}

332
std::vector<argument> program::eval(parameter_map params) const
Paul's avatar
Paul committed
333
{
Paul's avatar
Paul committed
334
335
    auto& ctx = this->impl->ctx;
#ifndef NDEBUG
336
337
338
339
340
341
342
343
344
345
346
    auto with_check_context = [&](auto f) {
        return [=, &ctx](auto&&) {
            auto sctx          = std::make_shared<context>(ctx);
            auto check_context = [=, &ctx](auto g) {
                assert(is_shared(ctx, *sctx));
                auto x = g();
                *sctx  = ctx;
                return x;
            };
            return [=](auto&&... xs) { return f(xs..., check_context); };
        };
Paul's avatar
Paul committed
347
348
    };
#else
349
350
351
352
353
    auto with_check_context = [](auto f) {
        return [=](auto&&) {
            return [=](auto&&... xs) { return f(xs..., [](auto g) { return g(); }); };
        };
    };
Paul's avatar
Paul committed
354
#endif
Paul's avatar
Paul committed
355
356
357
358

    auto trace_level = value_of(MIGRAPHX_TRACE_EVAL{});

    if(trace_level > 0)
Paul's avatar
Paul committed
359
    {
360
361
362
363
364
365
366
367
368
369
370
371
372
373
        return generic_eval(*this,
                            ctx,
                            std::move(params),
                            with_check_context([&](auto& ins, auto f, auto&& check_context) {
                                ctx.finish();
                                std::cout << "Run instruction: ";
                                this->debug_print(ins);
                                timer t{};
                                auto result = check_context(f);
                                double t1   = t.record<milliseconds>();
                                ctx.finish();
                                double t2 = t.record<milliseconds>();
                                std::cout << "Time: " << t1 << "ms, " << t2 << "ms" << std::endl;
                                if(trace_level > 1 and ins->name().front() != '@' and
Shucai Xiao's avatar
Shucai Xiao committed
374
375
                                   ins->name() != "load" and not result.empty())
                                {
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
                                    target tgt  = make_target(this->impl->target_name);
                                    auto buffer = tgt.copy_from(result);
                                    if(trace_level == 2)
                                    {
                                        std::cout << "Output has "
                                                  << to_string_range(classify_argument(buffer))
                                                  << std::endl;
                                        std::cout << "Output: ";
                                        preview_argument(std::cout, buffer);
                                        std::cout << std::endl;
                                    }
                                    else
                                    {
                                        std::cout << "Output: " << buffer << std::endl;
                                    }
Shucai Xiao's avatar
Shucai Xiao committed
391
                                }
392
393
                                return result;
                            }));
Paul's avatar
Paul committed
394
395
396
    }
    else
    {
397
398
399
400
401
402
        return generic_eval(*this,
                            ctx,
                            std::move(params),
                            with_check_context([&](auto&, auto f, auto&& check_context) {
                                return check_context(f);
                            }));
Paul's avatar
Paul committed
403
    }
Paul's avatar
Paul committed
404
405
}

406
const int program_file_version = 5;
407
408
409
410
411
412
413
414

value program::to_value() const
{
    value result;
    result["version"] = program_file_version;
    result["target"]  = this->impl->target_name;
    if(not this->impl->target_name.empty())
        result["context"] = this->impl->ctx.to_value();
Shucai Xiao's avatar
Shucai Xiao committed
415

416
    value module_vals = value::object{};
Shucai Xiao's avatar
Shucai Xiao committed
417
    std::unordered_map<instruction_ref, std::string> names;
418
    for(auto& mod : this->get_modules())
Shucai Xiao's avatar
Shucai Xiao committed
419
    {
Shucai Xiao's avatar
Shucai Xiao committed
420
421
        value mod_val;
        value nodes;
422
423
        mod_val["name"] = mod->name();
        names           = mod->print(
Shucai Xiao's avatar
Shucai Xiao committed
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
            [&](auto ins, auto ins_names) {
                value node;
                node["output"]     = ins_names.at(ins);
                node["name"]       = ins->name();
                node["shape"]      = migraphx::to_value(ins->get_shape());
                node["normalized"] = ins->is_normalized();
                if(ins->name() == "@literal")
                    node["literal"] = migraphx::to_value(ins->get_literal());
                node["operator"] = ins->get_operator().to_value();
                std::vector<std::string> inputs;
                std::transform(ins->inputs().begin(),
                               ins->inputs().end(),
                               std::back_inserter(inputs),
                               [&](auto i) {
                                   assert(contains(ins_names, i));
                                   return ins_names.at(i);
                               });
                node["inputs"]   = inputs;
                auto module_args = ins->module_inputs();
                if(not module_args.empty())
                {
                    std::vector<std::string> module_inputs;
                    std::transform(module_args.begin(),
                                   module_args.end(),
                                   std::back_inserter(module_inputs),
                                   [&](auto mod_ref) { return mod_ref->name(); });
                    node["module_inputs"] = module_inputs;
                }

                nodes.push_back(node);
            },
            names);
        mod_val["nodes"] = nodes;

458
        module_vals[mod->name()] = mod_val;
Shucai Xiao's avatar
Shucai Xiao committed
459
    }
Shucai Xiao's avatar
Shucai Xiao committed
460
461
462

    result["modules"] = module_vals;

463
464
    return result;
}
Shucai Xiao's avatar
Shucai Xiao committed
465

Shucai Xiao's avatar
Shucai Xiao committed
466
467
468
469
470
static void mod_from_val(module_ref mod,
                         const value& v,
                         std::unordered_map<std::string, instruction_ref>& instructions,
                         const std::unordered_map<std::string, module_ref>& map_mods)
{
471
    const auto& module_val = v.at(mod->name());
Shucai Xiao's avatar
Shucai Xiao committed
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
    for(const value& node : module_val.at("nodes"))
    {
        instruction_ref output;
        auto name       = node.at("name").to<std::string>();
        auto fields     = node.at("operator");
        auto normalized = node.at("normalized").to<bool>();

        if(name == "@param")
        {
            output = mod->add_parameter(fields["parameter"].to<std::string>(),
                                        migraphx::from_value<shape>(node.at("shape")));
        }
        else if(name == "@literal")
        {
            output = mod->add_literal(migraphx::from_value<literal>(node.at("literal")));
        }
        else
        {
            auto op = make_op(name, fields);
            std::vector<instruction_ref> inputs;
            std::transform(node.at("inputs").begin(),
                           node.at("inputs").end(),
                           std::back_inserter(inputs),
                           [&](const value& i) {
                               auto i_name = i.to<std::string>();
                               assert(contains(instructions, i_name));
                               return instructions.at(i_name);
                           });

            std::vector<module_ref> module_inputs;
            if(node.contains("module_inputs"))
            {
                std::transform(node.at("module_inputs").begin(),
                               node.at("module_inputs").end(),
                               std::back_inserter(module_inputs),
                               [&](const value& i) { return map_mods.at(i.to<std::string>()); });

                for(auto& smod : module_inputs)
                {
                    mod_from_val(smod, v, instructions, map_mods);
                }
            }

            if(name == "@return")
            {
                output = mod->add_return(inputs);
            }
            else if(module_inputs.empty())
            {
                output = mod->add_instruction(op, inputs);
            }
            else
            {
                output = mod->add_instruction(op, inputs, module_inputs);
            }
        }
        output->set_normalized(normalized);
        instructions[node.at("output").to<std::string>()] = output;
    }
}

533
534
535
536
void program::from_value(const value& v)
{
    auto version = v.at("version").to<int>();
    if(version != program_file_version)
Shucai Xiao's avatar
Shucai Xiao committed
537
538
539
540
    {
        MIGRAPHX_THROW("Warning: Program version mismatch");
    }

541
542
543
544
545
546
547
548
    this->impl->target_name = v.at("target").to<std::string>();
    if(not this->impl->target_name.empty())
    {
        target t        = make_target(this->impl->target_name);
        this->impl->ctx = t.get_context();
        this->impl->ctx.from_value(v.at("context"));
    }

Shucai Xiao's avatar
Shucai Xiao committed
549
550
    auto module_vals = v.at("modules");
    for(const auto& vv : module_vals)
551
    {
552
        const auto& name = vv.get_key();
Shucai Xiao's avatar
Shucai Xiao committed
553
554
        if(name == "main")
            continue;
555
        impl->modules.emplace(name, name);
556
    }
557
    std::unordered_map<std::string, module_ref> map_mods;
Paul's avatar
Paul committed
558
559
560
561
    std::transform(impl->modules.begin(),
                   impl->modules.end(),
                   std::inserter(map_mods, map_mods.end()),
                   [&](auto&& pp) { return std::make_pair(pp.first, &pp.second); });
Shucai Xiao's avatar
Shucai Xiao committed
562
563
564
565
566

    std::unordered_map<std::string, instruction_ref> map_insts;
    auto* mm = get_main_module();
    mod_from_val(mm, module_vals, map_insts, map_mods);

567
568
569
    this->finalize();
}

Paul's avatar
Paul committed
570
571
572
double common_average(const std::vector<double>& v)
{
    std::size_t n = v.size() / 4;
Paul's avatar
Paul committed
573
574
    double total  = std::accumulate(v.begin() + n, v.end() - n, 0.0);
    return total / std::distance(v.begin() + n, v.end() - n);
Paul's avatar
Paul committed
575
576
}

Paul Fultz II's avatar
Paul Fultz II committed
577
578
579
580
581
582
583
584
std::string perf_group(const operation& op)
{
    auto attr = op.attributes();
    if(attr.contains("group"))
        return attr.at("group").to<std::string>();
    return op.name();
}

585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
void program::mark(const parameter_map& params, marker&& m)
{
    auto& ctx = this->impl->ctx;
    // Run once by itself
    eval(params);
    ctx.finish();
    // Start marking
    m.mark_start(*this);
    generic_eval(*this, ctx, params, always([&](auto ins, auto f) {
        argument result;
        m.mark_start(ins);
        result = f();
        m.mark_stop(ins);
        return result;
    }));
    m.mark_stop(*this);
}

603
604
605
606
607
608
609
610
611
static void print_space(std::ostream& os, int n)
{
    for(int i = 0; i < n; ++i)
    {
        os << ' ';
    }
}

using op_flops = std::function<double(const std::vector<shape>& vec_ss)>;
612
auto& get_flops_funcs()
613
614
615
616
{
    static std::unordered_map<std::string, op_flops> op_funcs;
    op_funcs.emplace("gemm", [&](const std::vector<shape>& vec_ss) {
        assert(vec_ss.size() >= 2);
Shucai Xiao's avatar
Shucai Xiao committed
617
618
619
        auto sa     = vec_ss.front();
        auto sb     = vec_ss.at(1);
        auto batch  = 1;
620
        auto lens_a = sa.lens();
Shucai Xiao's avatar
Shucai Xiao committed
621
622
623
624
        batch =
            std::accumulate(lens_a.rbegin() + 2, lens_a.rend(), 1, std::multiplies<std::size_t>{});
        auto m      = lens_a[lens_a.size() - 2];
        auto k      = lens_a.back();
625
626
627
628
629
630
631
        auto lens_b = sb.lens();
        assert(k == lens_b[lens_b.size() - 2]);
        auto n = lens_b.back();

        return 2.0 * m * n * k * batch;
    });

632
633
    op_funcs.emplace("convolution", [&](const std::vector<shape>& vec_ss) {
        assert(vec_ss.size() >= 2);
Shucai Xiao's avatar
Shucai Xiao committed
634
635
636
637
638
639
640
641
642
        auto alens = vec_ss.front().lens();
        auto blens = vec_ss.at(1).lens();
        auto olens = vec_ss.back().lens();

        auto n  = alens.front();
        auto k  = blens.front();
        auto c  = alens.at(1);
        auto y  = blens.at(2);
        auto x  = blens.back();
643
644
645
646
647
648
        auto ho = olens.at(2);
        auto wo = olens.back();

        return 2.0 * n * k * ho * wo * c * y * x;
    });

649
650
651
    return op_funcs;
}

652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
int program::max_ins_length() const
{
    std::unordered_map<instruction_ref, std::string> names;
    int max_ins_len = 0;

    this->print(names, [&](auto ins, auto ins_names) {
        std::stringstream ss;
        instruction::print(ss, ins, ins_names);
        if(max_ins_len < ss.str().length())
        {
            max_ins_len = ss.str().length();
        }

        // skip return instruction
        if(ins->name() == "@return")
            return;
    });

    return max_ins_len;
}

static auto& get_titles()
{
Shucai Xiao's avatar
Shucai Xiao committed
675
676
677
    static std::vector<std::string> titles = {"Instructions",
                                              "Time(ms)    \t",
                                              "Percentage  \t",
678
                                              "(b, m, n, k)                    \t",
Shucai Xiao's avatar
Shucai Xiao committed
679
680
                                              "Flops(TFlops/s)  \t",
                                              "Throughput(GB/s)"};
681
682
683
684
685
686

    return titles;
}

static void print_title(std::ostream& os, std::size_t max_ins_len)
{
Shucai Xiao's avatar
Shucai Xiao committed
687
688
    auto titles      = get_titles();
    std::string& str = titles.front();
689
690
    str.append(max_ins_len + 1 - str.length(), ' ');
    str.append(1, '\t');
Shucai Xiao's avatar
Shucai Xiao committed
691
    for(auto& s : titles)
692
693
694
695
696
697
    {
        os << s;
    }
    os << std::endl;
}

Shucai Xiao's avatar
Shucai Xiao committed
698
699
700
701
702
static void print_ins_perf(std::ostream& os,
                           const std::vector<std::string>& titles,
                           instruction_ref ins,
                           double t,
                           double total_t)
703
{
Shucai Xiao's avatar
Shucai Xiao committed
704
705
706
    auto& time_str  = titles.at(1);
    auto& time_per  = titles.at(2);
    auto& size_str  = titles.at(3);
707
    auto& flops_str = titles.at(4);
708
    auto& thrpt_str = titles.at(5);
709
710

    auto& flops_funcs = get_flops_funcs();
Shucai Xiao's avatar
Shucai Xiao committed
711
    std::string tms   = std::to_string(t);
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
    tms.append(time_str.length() - tms.length(), ' ');
    tms.append(1, '\t');
    double percent   = 100.0 * t / total_t;
    std::string pers = std::to_string(percent);
    auto loc         = pers.find('.');
    if(loc != std::string::npos)
    {
        pers.erase(pers.begin() + loc + 6, pers.end());
    }
    pers.append(time_per.length() - pers.length(), ' ');
    pers.append(1, '\t');

    // calculate flops
    std::string szs;
    std::string flps;
    std::string op_name = ins->name();
    auto nloc           = op_name.find("::");
    op_name.erase(op_name.begin(), op_name.begin() + nloc + 2);
Shucai Xiao's avatar
Shucai Xiao committed
730
    auto inss = to_shapes(ins->inputs());
731
732
733
734
735
    if(contains(flops_funcs, op_name))
    {
        // print size
        auto alens = inss.front().lens();
        auto blens = inss.at(1).lens();
Shucai Xiao's avatar
Shucai Xiao committed
736
737
        auto mb =
            std::accumulate(alens.rbegin() + 2, alens.rend(), 1, std::multiplies<std::size_t>{});
738
739
740
741
742
743
744
745
746
747
748
749
750
        int mm = alens[alens.size() - 2];
        int mk = alens.back();
        int mn = blens.back();

        szs = "{";
        szs.append(std::to_string(mb));
        szs.append(1, ',');
        szs.append(std::to_string(mm));
        szs.append(1, ',');
        szs.append(std::to_string(mk));
        szs.append(1, ',');
        szs.append(std::to_string(mn));
        szs.append("}");
751
        szs.append(size_str.length() - szs.length(), ' ');
Shucai Xiao's avatar
Shucai Xiao committed
752

753
754
755
756
757
758
759
760
761
762
763
764
765
766
        auto op_flop_func = flops_funcs.at(op_name);
        double flops      = op_flop_func(inss);
        flops /= t;
        // convert to GFlops
        flops /= 1.0e9;
        flps      = std::to_string(flops);
        auto floc = flps.find('.');
        if(floc != std::string::npos)
        {
            flps.erase(flps.begin() + floc + 4, flps.end());
        }
    }
    szs.append(size_str.length() - szs.length(), ' ');
    flps.append(flops_str.length() - flps.length(), ' ');
767
768
769
770

    // print throughput for pointwise instruction
    auto alias_num = ins->get_operator().output_alias({});
    std::string thrpt;
Shucai Xiao's avatar
Shucai Xiao committed
771
    if(alias_num != 0)
772
    {
Shucai Xiao's avatar
Shucai Xiao committed
773
774
775
776
        auto size =
            std::accumulate(inss.begin(), inss.end(), std::size_t{0}, [&](auto init, auto s) {
                return init + s.bytes();
            });
777

Shucai Xiao's avatar
Shucai Xiao committed
778
        double throughput = size / t;
779
        // convert to GB/s
Shucai Xiao's avatar
Shucai Xiao committed
780
        throughput /= 1.0e6;
Shucai Xiao's avatar
Shucai Xiao committed
781
        thrpt     = std::to_string(throughput);
782
783
784
785
786
        auto floc = flps.find('.');
        if(floc != std::string::npos)
        {
            thrpt.erase(thrpt.begin() + floc + 4, thrpt.end());
        }
Shucai Xiao's avatar
Shucai Xiao committed
787
788
789
        // debugging
        thrpt.append(1, ',');
        thrpt.append(std::to_string(size));
790
    }
Shucai Xiao's avatar
Shucai Xiao committed
791
    thrpt.append(thrpt_str.length() + 7 - thrpt.length(), ' ');
792
793

    os << tms << pers << szs << flps << thrpt << std::endl;
794
795
}

796
797
798
799
void program::perf_report(std::ostream& os,
                          std::size_t n,
                          parameter_map params,
                          std::size_t batch) const
Paul's avatar
Paul committed
800
{
801
    auto& ctx = this->impl->ctx;
Paul's avatar
Paul committed
802
803
    // Run once by itself
    eval(params);
Paul's avatar
Paul committed
804
    ctx.finish();
Paul's avatar
Paul committed
805
    // Run and time entire program
Paul's avatar
Paul committed
806
807
    std::vector<double> total_vec;
    total_vec.reserve(n);
Paul's avatar
Paul committed
808
    for(std::size_t i = 0; i < n; i++)
Paul's avatar
Paul committed
809
    {
Paul's avatar
Paul committed
810
811
812
813
        total_vec.push_back(time<milliseconds>([&] {
            eval(params);
            ctx.finish();
        }));
Paul's avatar
Paul committed
814
    }
Paul's avatar
Paul committed
815
816
    std::sort(total_vec.begin(), total_vec.end());
    std::unordered_map<instruction_ref, std::vector<double>> ins_vec;
Paul's avatar
Paul committed
817
    // Fill the map
818
    generic_eval(*this, ctx, params, always([&](auto ins, auto) {
Paul's avatar
Paul committed
819
        ins_vec[ins].reserve(n);
820
        return argument{ins->get_shape(), nullptr};
821
    }));
822

Paul's avatar
Paul committed
823
    // Run and time each instruction
Paul's avatar
Paul committed
824
    for(std::size_t i = 0; i < n; i++)
Paul's avatar
Paul committed
825
    {
826
        generic_eval(*this, ctx, params, always([&](auto ins, auto f) {
827
            argument result;
Paul's avatar
Paul committed
828
829
830
831
            ins_vec[ins].push_back(time<milliseconds>([&] {
                result = f();
                ctx.finish();
            }));
832
            return result;
833
        }));
Paul's avatar
Paul committed
834
    }
Paul's avatar
Paul committed
835
836
    for(auto&& p : ins_vec)
        std::sort(p.second.begin(), p.second.end());
Paul's avatar
Paul committed
837
    // Run and time implicit overhead
Paul's avatar
Paul committed
838
839
    std::vector<double> overhead_vec;
    overhead_vec.reserve(n);
Paul's avatar
Paul committed
840
    for(std::size_t i = 0; i < n; i++)
Paul's avatar
Paul committed
841
    {
Paul's avatar
Paul committed
842
        overhead_vec.push_back(time<milliseconds>([&] { dry_run(params); }));
Paul's avatar
Paul committed
843
844
    }

Paul's avatar
Paul committed
845
    double total_time             = common_average(total_vec);
Paul's avatar
Paul committed
846
    double rate                   = 1000.0 / total_time;
Paul's avatar
Paul committed
847
    double overhead_time          = common_average(overhead_vec);
Paul's avatar
Paul committed
848
    double overhead_percent       = overhead_time * 100.0 / total_time;
Paul's avatar
Paul committed
849
    double total_instruction_time = 0.0;
Paul's avatar
Paul committed
850
    std::unordered_map<std::string, double> op_times;
Paul's avatar
Paul committed
851
    for(auto&& p : ins_vec)
Paul's avatar
Paul committed
852
853
    {
        double avg = common_average(p.second);
Paul Fultz II's avatar
Paul Fultz II committed
854
        op_times[perf_group(p.first->get_operator())] += avg;
Paul's avatar
Paul committed
855
856
        total_instruction_time += avg;
    }
Paul's avatar
Paul committed
857
858
    double calculate_overhead_time    = total_time - total_instruction_time;
    double calculate_overhead_percent = calculate_overhead_time * 100.0 / total_time;
Paul's avatar
Paul committed
859

Shucai Xiao's avatar
Shucai Xiao committed
860
    std::unordered_map<instruction_ref, std::string> names;
861
862

    // count max instruction length
Shucai Xiao's avatar
Shucai Xiao committed
863
    auto titles           = get_titles();
864
865
    const int max_ins_len = max_ins_length();
    print_title(os, max_ins_len);
866

Shucai Xiao's avatar
Shucai Xiao committed
867
    this->print(names, [&](auto ins, auto ins_names) {
868
869
870
        std::stringstream ss;
        instruction::print(ss, ins, ins_names);
        os << ss.str();
871
872
873
874
875

        // skip return instruction
        if(ins->name() == "@return")
            return;

876
877
        // insert space to align
        print_space(os, max_ins_len - ss.str().length());
Shucai Xiao's avatar
Shucai Xiao committed
878
        os << "\t";
Shucai Xiao's avatar
Shucai Xiao committed
879
        double avg = common_average(ins_vec[ins]);
880
        print_ins_perf(os, titles, ins, avg, total_instruction_time);
Paul's avatar
Paul committed
881
    });
Paul's avatar
Paul committed
882
883
884

    os << std::endl;
    os << "Summary:" << std::endl;
885
886
887
888
889
890
891
    std::vector<std::pair<double, std::string>> op_times_sorted;
    std::transform(op_times.begin(),
                   op_times.end(),
                   std::back_inserter(op_times_sorted),
                   [](auto p) { return std::make_pair(p.second, p.first); });
    std::sort(op_times_sorted.begin(), op_times_sorted.end(), std::greater<>{});
    for(auto&& p : op_times_sorted)
Paul's avatar
Paul committed
892
    {
893
894
        auto&& name    = p.second;
        double avg     = p.first;
Paul's avatar
Paul committed
895
896
897
898
899
        double percent = std::ceil(100.0 * avg / total_instruction_time);
        os << name << ": " << avg << "ms, " << percent << "%" << std::endl;
    }

    os << std::endl;
Paul's avatar
Paul committed
900

901
902
    os << "Batch size: " << batch << std::endl;
    os << "Rate: " << rate * batch << "/sec" << std::endl;
Paul's avatar
Paul committed
903
904
    os << "Total time: " << total_time << "ms" << std::endl;
    os << "Total instructions time: " << total_instruction_time << "ms" << std::endl;
Paul's avatar
Paul committed
905
906
907
908
    os << "Overhead time: " << overhead_time << "ms"
       << ", " << calculate_overhead_time << "ms" << std::endl;
    os << "Overhead: " << std::round(overhead_percent) << "%"
       << ", " << std::round(calculate_overhead_percent) << "%" << std::endl;
Paul's avatar
Paul committed
909
910
}

Paul's avatar
Paul committed
911
912
void program::debug_print() const { std::cout << *this << std::endl; }
void program::debug_print(instruction_ref ins) const
Paul's avatar
Paul committed
913
{
Shucai Xiao's avatar
Shucai Xiao committed
914
    std::unordered_map<instruction_ref, std::string> names;
915
    if(std::any_of(this->impl->modules.begin(), this->impl->modules.end(), [&](const auto& pp) {
916
           return is_end(pp.second.end(), ins);
Shucai Xiao's avatar
Shucai Xiao committed
917
       }))
Paul's avatar
Paul committed
918
919
920
921
    {
        std::cout << "End instruction" << std::endl;
        return;
    }
922
923
    else if(std::none_of(this->impl->modules.begin(),
                         this->impl->modules.end(),
924
                         [&](const auto& pp) { return pp.second.has_instruction(ins); }))
Paul's avatar
Paul committed
925
926
927
928
    {
        std::cout << "Instruction not part of program" << std::endl;
        return;
    }
Shucai Xiao's avatar
Shucai Xiao committed
929

Paul's avatar
Paul committed
930
    std::stringstream ss;
Shucai Xiao's avatar
Shucai Xiao committed
931
    this->print(names, [&](auto x, auto ins_names) {
Paul's avatar
Paul committed
932
        if(x == ins)
Paul's avatar
Paul committed
933
        {
Shucai Xiao's avatar
Shucai Xiao committed
934
            instruction::print(std::cout, x, ins_names);
Paul's avatar
Paul committed
935
936
937
938
939
            std::cout << std::endl;
        }
    });
}

Shucai Xiao's avatar
Shucai Xiao committed
940
941
942
943
void program::print(
    std::unordered_map<instruction_ref, std::string>& names,
    const std::function<void(instruction_ref, std::unordered_map<instruction_ref, std::string>)>&
        print_func) const
944
{
945
    for(const auto& pp : this->impl->modules)
946
    {
947
        names = pp.second.print(print_func, names);
948
949
950
    }
}

Shucai Xiao's avatar
Shucai Xiao committed
951
void program::print_graph(std::ostream& os, bool brief) const
952
{
Shucai Xiao's avatar
Shucai Xiao committed
953
954
    const auto* mm = this->get_main_module();
    mm->print_graph(os, brief);
955
956
957
958
}

void program::print_cpp(std::ostream& os) const
{
Shucai Xiao's avatar
Shucai Xiao committed
959
960
961
962
963
964
965
966
    auto vec_modules = this->get_modules();
    std::unordered_map<instruction_ref, std::string> names;
    for(auto& mod : vec_modules)
    {
        os << "module: \"" << mod->name() << "\"" << std::endl;
        names = mod->print_cpp(os, names);
        os << std::endl;
    }
967
968
}

Paul's avatar
Paul committed
969
970
void program::dry_run(std::unordered_map<std::string, argument> params) const
{
Paul's avatar
Paul committed
971
    auto& ctx = this->impl->ctx;
972
973
974
    generic_eval(*this, ctx, std::move(params), always([](auto ins, auto&&...) {
        return argument{ins->get_shape(), nullptr};
    }));
Paul's avatar
Paul committed
975
976
}

Shucai Xiao's avatar
Shucai Xiao committed
977
void program::annotate(std::ostream& os, const std::function<void(instruction_ref)>& a) const
Paul's avatar
Paul committed
978
{
979
    for(auto& pp : this->impl->modules)
Shucai Xiao's avatar
Shucai Xiao committed
980
    {
981
982
        std::cout << pp.first << ":" << std::endl;
        pp.second.annotate(os, a);
Shucai Xiao's avatar
Shucai Xiao committed
983
    }
Paul's avatar
Paul committed
984
985
}

Paul's avatar
Paul committed
986
const module* program::get_module(const std::string& name) const { return &impl->modules.at(name); }
Shucai Xiao's avatar
Shucai Xiao committed
987
988
989

module* program::create_module(const std::string& name)
{
990
    assert(not contains(impl->modules, name));
991
992
    auto r = impl->modules.emplace(name, name);
    return &(r.first->second);
Shucai Xiao's avatar
Shucai Xiao committed
993
994
}

Paul's avatar
Paul committed
995
module* program::get_module(const std::string& name) { return &impl->modules.at(name); }
Shucai Xiao's avatar
Shucai Xiao committed
996
997
998
999
1000

module* program::get_main_module() { return get_module("main"); }

const module* program::get_main_module() const { return get_module("main"); }

Paul's avatar
Paul committed
1001
template <class T>
1002
std::vector<T*> generic_get_modules(T* mm)
Shucai Xiao's avatar
Shucai Xiao committed
1003
{
1004
    std::vector<T*> vec_modules;
Shucai Xiao's avatar
Shucai Xiao committed
1005
1006
1007
1008
1009
    vec_modules.push_back(mm);
    auto sub_modules = mm->get_sub_modules();
    vec_modules.insert(vec_modules.end(), sub_modules.begin(), sub_modules.end());
    return vec_modules;
}
Shucai Xiao's avatar
Shucai Xiao committed
1010

Paul's avatar
Paul committed
1011
template <class Map, class T, class OutputIterator>
1012
void generic_get_unused_modules(Map& m, const std::vector<T*>& mods, OutputIterator out)
Shucai Xiao's avatar
Shucai Xiao committed
1013
{
1014
1015
1016
1017
    std::unordered_set<std::string> used;
    std::transform(mods.begin(), mods.end(), std::inserter(used, used.end()), [](auto&& mod) {
        return mod->name();
    });
Paul's avatar
Paul committed
1018
1019
1020
1021
1022
    transform_if(m.begin(),
                 m.end(),
                 out,
                 [&](auto&& pp) { return not contains(used, pp.first); },
                 [](auto&& pp) { return &pp.second; });
1023
}
Shucai Xiao's avatar
Shucai Xiao committed
1024

1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
std::vector<const module*> program::get_modules() const
{
    auto result = generic_get_modules(this->get_main_module());
    generic_get_unused_modules(impl->modules, result, std::back_inserter(result));
    return result;
}

std::vector<module*> program::get_modules()
{
    auto result = generic_get_modules(this->get_main_module());
    generic_get_unused_modules(impl->modules, result, std::back_inserter(result));
    return result;
Shucai Xiao's avatar
Shucai Xiao committed
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
1067
1068
1069
1070
1071
1072
1073
template <class Map, class T>
bool is_unused_module(Map& m, const std::vector<T*>& mods, const std::string& name)
{
    bool is_unused = false;
    generic_get_unused_modules(m, mods, make_function_output_iterator([&](auto* mod) {
                                   if(mod->name() == name)
                                       is_unused = true;
                               }));
    return is_unused;
}

template <class Map>
bool references_instruction(Map& m, const instruction& ins, const std::string& name)
{
    return std::any_of(m.begin(), m.end(), [&](auto&& p) {
        if(p.first == name)
            return false;
        return std::any_of(p.second.begin(), p.second.end(), [&](auto&& i) {
            return std::any_of(i.inputs().begin(), i.inputs().end(), [&](auto&& j) {
                return std::addressof(*j) == std::addressof(ins);
            });
        });
    });
}

void program::remove_module(const std::string& name)
{
    // cppcheck-suppress assertWithSideEffect
    assert(is_unused_module(impl->modules, generic_get_modules(this->get_main_module()), name) &&
           "Module used in program");
    assert(std::none_of(
               impl->modules.at(name).begin(),
               impl->modules.at(name).end(),
               [&](auto&& ins) { return references_instruction(impl->modules, ins, name); }) &&
           "Instruction referenced in another module");
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089

    // if an instruction has an input out side of the current module, need to remove
    // the instruction from its input's outputs
    auto& mod = impl->modules.at(name);
    for(auto ins : iterator_for(mod))
    {
        auto inputs = ins->inputs();
        for(auto in : inputs)
        {
            if(not mod.has_instruction(in))
            {
                in->remove_output(ins);
            }
        }
    }

1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
    impl->modules.erase(name);
}

void program::remove_unused_modules()
{
    std::vector<module*> unused;
    generic_get_unused_modules(
        impl->modules, generic_get_modules(this->get_main_module()), std::back_inserter(unused));
    for(auto* m : unused)
        this->remove_module(m->name());
}

1102
1103
program& program::sort()
{
1104
    for(auto& pp : this->impl->modules)
Shucai Xiao's avatar
Shucai Xiao committed
1105
    {
1106
        pp.second.sort();
Shucai Xiao's avatar
Shucai Xiao committed
1107
1108
    }

1109
1110
1111
    return *this;
}

Paul's avatar
Paul committed
1112
bool operator==(const program& x, const program& y) { return to_string(x) == to_string(y); }
Paul's avatar
Paul committed
1113

Paul's avatar
Paul committed
1114
std::ostream& operator<<(std::ostream& os, const program& p)
Paul's avatar
Paul committed
1115
{
Shucai Xiao's avatar
Shucai Xiao committed
1116
1117
1118
    auto vec_modules = p.get_modules();
    std::unordered_map<instruction_ref, std::string> names;
    for(auto& mod : vec_modules)
Shucai Xiao's avatar
Shucai Xiao committed
1119
    {
Shucai Xiao's avatar
Shucai Xiao committed
1120
1121
1122
1123
1124
1125
1126
        os << "module: \"" << mod->name() << "\"" << std::endl;
        names = mod->print(
            [&](auto ins, auto ins_names) {
                instruction::print(os, ins, ins_names);
                os << std::endl;
            },
            names);
1127
        os << std::endl;
Shucai Xiao's avatar
Shucai Xiao committed
1128
1129
    }

Paul's avatar
Paul committed
1130
    return os;
Paul's avatar
Paul committed
1131
}
Paul's avatar
Paul committed
1132

Paul's avatar
Paul committed
1133
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1134
} // namespace migraphx