onnx.cpp 47.7 KB
Newer Older
Paul's avatar
Paul committed
1
2
3
4
5
6
7
8
#include <google/protobuf/text_format.h>
#include <google/protobuf/io/zero_copy_stream_impl.h>
#include <onnx.pb.h>
#include <iostream>
#include <fstream>
#include <unordered_map>
#include <functional>
#include <array>
Paul's avatar
Paul committed
9
#include <utility>
10
#include <vector>
Paul's avatar
Paul committed
11

Paul's avatar
Paul committed
12
13
14
15
16
17
#include <migraphx/fallthrough.hpp>
#include <migraphx/program.hpp>
#include <migraphx/operators.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/config.hpp>
18
#include <migraphx/onnx.hpp>
Paul's avatar
Paul committed
19
20

namespace migraphx {
Paul's avatar
Paul committed
21
inline namespace MIGRAPHX_INLINE_NS {
Paul's avatar
Paul committed
22
23
24
25
26

struct onnx_parser
{
    using attribute_map = std::unordered_map<std::string, onnx::AttributeProto>;
    using node_map      = std::unordered_map<std::string, onnx::NodeProto>;
Paul's avatar
Paul committed
27
28
    using op_func =
        std::function<std::vector<instruction_ref>(attribute_map, std::vector<instruction_ref>)>;
Paul's avatar
Paul committed
29
30
    node_map nodes;
    std::unordered_map<std::string, instruction_ref> instructions;
Scott Thornton's avatar
Scott Thornton committed
31
    program prog    = program();
32
    bool is_pytorch = false;
Paul's avatar
Paul committed
33
34

    std::unordered_map<std::string, op_func> ops;
35
    std::unordered_map<std::string, operation> map_actv_funcs;
Paul's avatar
Paul committed
36
37
38

    onnx_parser()
    {
Shucai Xiao's avatar
Shucai Xiao committed
39
        add_generic_op("MatMul", op::dot{});
Khalique's avatar
Khalique committed
40
        add_generic_op("Relu", op::relu{});
Khalique's avatar
Khalique committed
41
42
        add_generic_op("Sigmoid", op::sigmoid{});
        add_generic_op("Abs", op::abs{});
Shucai Xiao's avatar
Shucai Xiao committed
43
44
        add_generic_op("Exp", op::exp{});
        add_generic_op("Log", op::log{});
Khalique's avatar
Khalique committed
45
46
        // disable dropout for inference
        add_generic_op("Dropout", op::identity{});
Khalique's avatar
Khalique committed
47
        add_generic_op("Identity", op::identity{});
Shucai Xiao's avatar
Shucai Xiao committed
48
49
50
        add_generic_op("Sin", op::sin{});
        add_generic_op("Cos", op::cos{});
        add_generic_op("Tan", op::tan{});
51
52
        add_generic_op("Sinh", op::sinh{});
        add_generic_op("Cosh", op::cosh{});
53
        add_generic_op("Tanh", op::tanh{});
54
55
56
        add_generic_op("Asin", op::asin{});
        add_generic_op("Acos", op::acos{});
        add_generic_op("Atan", op::atan{});
Paul's avatar
Paul committed
57

Khalique's avatar
Khalique committed
58
59
60
61
62
        add_binary_op("Add", op::add{});
        add_binary_op("Div", op::div{});
        add_binary_op("Mul", op::mul{});
        add_binary_op("Sub", op::sub{});

Khalique's avatar
Khalique committed
63
64
65
        add_variadic_op("Sum", op::add{});
        add_variadic_op("Max", op::max{});
        add_variadic_op("Min", op::min{});
Paul's avatar
Paul committed
66

Khalique's avatar
Khalique committed
67
        add_mem_op("LRN", &onnx_parser::parse_lrn);
Khalique's avatar
Khalique committed
68
        add_mem_op("ImageScaler", &onnx_parser::parse_imagescaler);
69
        add_mem_op("LeakyRelu", &onnx_parser::parse_leaky_relu);
Khalique's avatar
Khalique committed
70
        add_mem_op("Elu", &onnx_parser::parse_elu);
Paul's avatar
Paul committed
71
72
        add_mem_op("Constant", &onnx_parser::parse_constant);
        add_mem_op("Conv", &onnx_parser::parse_conv);
Paul's avatar
Paul committed
73
74
        add_mem_op("MaxPool", &onnx_parser::parse_pooling);
        add_mem_op("AveragePool", &onnx_parser::parse_pooling);
75
76
        add_mem_op("GlobalMaxPool", &onnx_parser::parse_pooling);
        add_mem_op("GlobalAveragePool", &onnx_parser::parse_pooling);
Paul's avatar
Paul committed
77
        add_mem_op("Reshape", &onnx_parser::parse_reshape);
Paul's avatar
Paul committed
78
79
        add_mem_op("Flatten", &onnx_parser::parse_flatten);
        add_mem_op("Gemm", &onnx_parser::parse_gemm);
80
        add_mem_op("BatchNormalization", &onnx_parser::parse_batchnorm);
Paul's avatar
Paul committed
81
        add_mem_op("Softmax", &onnx_parser::parse_softmax);
82
83
84
        add_mem_op("Squeeze", &onnx_parser::parse_squeeze);
        add_mem_op("Unsqueeze", &onnx_parser::parse_unsqueeze);
        add_mem_op("Slice", &onnx_parser::parse_slice);
Scott Thornton's avatar
Scott Thornton committed
85
        add_mem_op("Concat", &onnx_parser::parse_concat);
86
87
88
        add_mem_op("Gather", &onnx_parser::parse_gather);
        add_mem_op("Shape", &onnx_parser::parse_shape);
        add_mem_op("ConstantFill", &onnx_parser::parse_constant_fill);
Khalique's avatar
Khalique committed
89
        add_mem_op("Transpose", &onnx_parser::parse_transpose);
Shucai Xiao's avatar
Shucai Xiao committed
90
        add_mem_op("RNN", &onnx_parser::parse_rnn);
91
        add_mem_op("GRU", &onnx_parser::parse_gru);
Khalique's avatar
Khalique committed
92
        add_mem_op("Pad", &onnx_parser::parse_pad);
93
94
95
96
97
98
99

        // init the activation function map
        init_actv_func();
    }

    void init_actv_func()
    {
100
101
102
103
104
        map_actv_funcs.insert(std::make_pair("tanh", op::tanh{}));
        map_actv_funcs.insert(std::make_pair("relu", op::relu{}));
        map_actv_funcs.insert(std::make_pair("sigmoid", op::sigmoid{}));
        map_actv_funcs.insert(std::make_pair("leakyrelu", op::leaky_relu{}));
        map_actv_funcs.insert(std::make_pair("elu", op::elu{}));
Paul's avatar
Paul committed
105
106
107
108
    }

    template <class F>
    void add_op(std::string name, F f)
Paul's avatar
Paul committed
109
110
111
112
113
114
115
116
117
    {
        ops.emplace(name, [=](auto&&... xs) {
            return std::vector<instruction_ref>{f(std::forward<decltype(xs)>(xs)...)};
        });
    }

    // Multi output op
    template <class F>
    void add_multi_op(std::string name, F f)
Paul's avatar
Paul committed
118
119
120
121
122
123
124
    {
        ops.emplace(name, f);
    }

    template <class F>
    void add_mem_op(std::string name, F f)
    {
Paul's avatar
Paul committed
125
        add_op(name, [=](auto&&... xs) {
Paul's avatar
Paul committed
126
127
128
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }
Khalique's avatar
Khalique committed
129

130
    template <class T>
Khalique's avatar
Khalique committed
131
    void add_binary_op(std::string name, T x)
132
    {
Paul's avatar
Paul committed
133
        add_op(name, [this, x](attribute_map attributes, std::vector<instruction_ref> args) {
Scott Thornton's avatar
Scott Thornton committed
134
            if(args.size() != 2)
Paul's avatar
Paul committed
135
                MIGRAPHX_THROW("binary operators should have 2 operands");
136
            if(contains(attributes, "broadcast") and contains(attributes, "axis"))
137
138
139
140
            {
                uint64_t broadcasted = parse_value(attributes.at("broadcast")).at<uint64_t>();
                if(broadcasted != 0)
                {
141
                    uint64_t axis = parse_value(attributes.at("axis")).at<uint64_t>();
142
143
144
145
                    auto l =
                        prog.add_instruction(op::broadcast{axis, args[0]->get_shape()}, args[1]);
                    return prog.add_instruction(x, args[0], l);
                }
146
                return prog.add_instruction(x, args);
147
            }
Paul's avatar
Paul committed
148
            else
149
            {
Khalique's avatar
Khalique committed
150
                return add_broadcastable_binary_op(args[0], args[1], x);
151
152
153
154
            }
        });
    }

Khalique's avatar
Khalique committed
155
156
157
158
159
    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
        if(arg0->get_shape() != arg1->get_shape())
        {
Khalique's avatar
Khalique committed
160
161
162
163
164
165
166
167
168
169
170
171
172
            // Example:
            // s0 = (3,2,4,5) and s1 = (2,1,1)
            //
            // In this case we need to broadcast (:,1,1) portion of
            // s1 plus broadcast the 1st dimension of s1
            // giving output_lens = (3,2,4,5)
            //
            // Another example:
            // s0 = (3,2,1,5) and s1 = (2,7,5)
            // In this case we need to broadcast the (:,:,1:,:) axis
            // of s0 plus the 1st dimension of s1 giving
            // output_lens = (3,2,7,5)
            //
Khalique's avatar
Khalique committed
173
174
175
176
177
178
179
180
            // Get lengths for both arguments
            const std::vector<std::size_t>* s0 = &arg0->get_shape().lens();
            const std::vector<std::size_t>* s1 = &arg1->get_shape().lens();

            // Make sure s0 is the smaller size
            if(s0->size() > s1->size())
                std::swap(s0, s1);

Khalique's avatar
Khalique committed
181
            std::vector<std::size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
182
183
            auto offset = s1->size() - s0->size();
            std::transform(s0->begin(),
Khalique's avatar
Khalique committed
184
185
186
187
                           s0->end(),
                           s1->begin() + offset,
                           output_lens.begin() + offset,
                           [](auto a, auto b) { return std::max(a, b); });
Khalique's avatar
Khalique committed
188
189
190
191
192
193
194
195
196

            auto l0 = prog.add_instruction(op::multibroadcast{output_lens}, arg0);
            auto l1 = prog.add_instruction(op::multibroadcast{output_lens}, arg1);
            return prog.add_instruction(x, l0, l1);
        }
        else
        {
            return prog.add_instruction(x, {arg0, arg1});
        }
197
198
    }

Paul's avatar
Paul committed
199
    template <class T>
Paul's avatar
Paul committed
200
201
    void add_generic_op(std::string name, T x)
    {
Paul's avatar
Paul committed
202
        add_op(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
Paul's avatar
Paul committed
203
204
205
206
            return prog.add_instruction(x, args);
        });
    }

Khalique's avatar
Khalique committed
207
    template <class T>
Khalique's avatar
Khalique committed
208
    void add_variadic_op(std::string name, T x)
Khalique's avatar
Khalique committed
209
    {
Paul's avatar
Paul committed
210
        add_op(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
211
            return std::accumulate(std::next(args.begin()),
Khalique's avatar
Khalique committed
212
213
214
215
216
                                   args.end(),
                                   args.front(),
                                   [this, x](instruction_ref a, instruction_ref b) {
                                       return add_broadcastable_binary_op(a, b, x);
                                   });
Khalique's avatar
Khalique committed
217
        });
Khalique's avatar
Khalique committed
218
219
    }

Paul's avatar
Paul committed
220
    instruction_ref
Paul's avatar
Paul committed
221
    parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
222
223
    {
        auto dims = args.front()->get_shape().lens();
Scott Thornton's avatar
Scott Thornton committed
224
225
        auto r =
            prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, args.front());
226
227
        auto s = prog.add_instruction(op::softmax{}, r);
        return prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1])}}, s);
Paul's avatar
Paul committed
228
229
    }

Paul's avatar
Paul committed
230
    instruction_ref
Paul's avatar
Paul committed
231
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
232
    {
233
        op::convolution op;
234
        auto l0 = args[0];
Paul's avatar
Paul committed
235
236
        if(contains(attributes, "pads"))
        {
Scott Thornton's avatar
Scott Thornton committed
237
            if(contains(attributes, "auto_pad"))
238
            {
Paul's avatar
Paul committed
239
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
240
            }
241
242
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
243
            if(padding.size() != 4)
244
            {
Paul's avatar
Paul committed
245
                MIGRAPHX_THROW("padding should have 4 values");
246
            }
Scott Thornton's avatar
Scott Thornton committed
247
            if(padding[0] != padding[2] || padding[1] != padding[3])
248
            {
249
250
                // insert zeros for pad op (args[0] has 4 dims)
                padding = {0, 0, padding[0], padding[1], 0, 0, padding[2], padding[3]};
Khalique's avatar
Khalique committed
251
                l0      = prog.add_instruction(op::pad{padding}, l0);
252
            }
253
254
255
256
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
257
            }
Paul's avatar
Paul committed
258
        }
Paul's avatar
Paul committed
259
260
261
262
263
264
265
266
        if(contains(attributes, "strides"))
        {
            copy(attributes["strides"].ints(), op.stride.begin());
        }
        if(contains(attributes, "dilations"))
        {
            copy(attributes["dilations"].ints(), op.dilation.begin());
        }
Scott Thornton's avatar
Scott Thornton committed
267
        if(contains(attributes, "auto_pad"))
268
269
        {
            auto s = attributes["auto_pad"].s();
Scott Thornton's avatar
Scott Thornton committed
270
            if(contains(attributes, "pads") and to_upper(s) != "NOTSET")
271
            {
Paul's avatar
Paul committed
272
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
273
274
            }

wsttiger's avatar
fixes  
wsttiger committed
275
            if(s.find("SAME") != std::string::npos)
276
            {
277
                op.padding_mode = op::padding_mode_t::same;
278
279
            }
        }
Khalique's avatar
Khalique committed
280
281
282
283
        if(contains(attributes, "group"))
        {
            op.group = parse_value(attributes.at("group")).at<int>();
        }
Paul's avatar
Paul committed
284
285
286
287
        if(args.size() == 3)
        {
            uint64_t axis = 1;
            auto l1       = prog.add_instruction(op, args[0], args[1]);
Scott Thornton's avatar
Scott Thornton committed
288
            auto l2       = prog.add_instruction(op::broadcast{axis, l1->get_shape()}, args[2]);
289
            return prog.add_instruction(op::add{}, l1, l2);
Paul's avatar
Paul committed
290
        }
291
        return prog.add_instruction(op, l0, args[1]);
Paul's avatar
Paul committed
292
    }
Paul's avatar
Paul committed
293

Paul's avatar
Paul committed
294
295
296
    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
Paul's avatar
Paul committed
297
    {
Khalique's avatar
Khalique committed
298
        op::pooling op{ends_with(name, "MaxPool") ? "max" : "average"};
299
        auto l0 = args[0];
Khalique's avatar
Khalique committed
300
        if(starts_with(name, "Global"))
301
        {
Khalique's avatar
Khalique committed
302
303
            auto lens  = args.front()->get_shape().lens();
            op.lengths = {lens[2], lens[3]};
304
        }
Paul's avatar
Paul committed
305
306
        if(contains(attributes, "pads"))
        {
307
308
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
309
            if(padding.size() != 4)
310
            {
Paul's avatar
Paul committed
311
                MIGRAPHX_THROW("padding should have 4 values");
312
            }
Scott Thornton's avatar
Scott Thornton committed
313
            if(padding[0] != padding[2] || padding[1] != padding[3])
314
            {
315
316
                // insert zeros for pad op (args[0] has 4 dims)
                padding = {0, 0, padding[0], padding[1], 0, 0, padding[2], padding[3]};
Khalique's avatar
Khalique committed
317
                l0      = prog.add_instruction(op::pad{padding}, l0);
318
319
320
321
322
            }
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
323
            }
Paul's avatar
Paul committed
324
325
326
327
328
329
330
331
332
        }
        if(contains(attributes, "strides"))
        {
            copy(attributes["strides"].ints(), op.stride.begin());
        }
        if(contains(attributes, "kernel_shape"))
        {
            copy(attributes["kernel_shape"].ints(), op.lengths.begin());
        }
Scott Thornton's avatar
Scott Thornton committed
333
        if(contains(attributes, "auto_pad"))
334
335
        {
            auto s = attributes["auto_pad"].s();
336
            if(s.find("SAME_UPPER") == std::string::npos)
337
            {
338
                MIGRAPHX_THROW("auto_pad only supports SAME_UPPER for pooling");
339
            }
340
            op.padding_mode = op::padding_mode_t::same;
341
342
        }

343
        return prog.add_instruction(op, l0);
Paul's avatar
Paul committed
344
345
    }

Paul's avatar
Paul committed
346
    instruction_ref
Paul's avatar
Paul committed
347
    parse_reshape(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
348
    {
349
        op::reshape op;
Paul's avatar
Paul committed
350
351
352
353
354
355
356
        if(args.size() == 1)
        {
            literal s = parse_value(attributes.at("shape"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
        }
        if(args.size() == 2)
        {
Paul's avatar
Paul committed
357
            literal s = args[1]->get_literal();
Paul's avatar
Paul committed
358
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
Paul's avatar
Paul committed
359
        }
Paul's avatar
Paul committed
360
361
362
        return prog.add_instruction(op, args[0]);
    }

Paul's avatar
Paul committed
363
    instruction_ref
Paul's avatar
Paul committed
364
    parse_flatten(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
365
    {
366
        uint64_t axis = 1;
Paul's avatar
Paul committed
367
368
369
370
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
371
        return prog.add_instruction(op::flatten{axis}, args[0]);
Paul's avatar
Paul committed
372
373
    }

374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
    instruction_ref
    parse_squeeze(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::squeeze op;
        literal s = parse_value(attributes.at("axes"));
        s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        return prog.add_instruction(op, args[0]);
    }

    instruction_ref
    parse_unsqueeze(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::unsqueeze op;
        literal s = parse_value(attributes.at("axes"));
        s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        return prog.add_instruction(op, args[0]);
    }

Scott Thornton's avatar
Scott Thornton committed
392
393
394
395
396
397
398
    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        std::size_t axis = parse_value(attributes.at("axis")).at<int>();
        op::concat op{axis};
        return prog.add_instruction(op, std::move(args));
    }
399

400
401
402
    instruction_ref
    parse_gather(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
403
        int axis = 0;
404
405
406
407
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
408
        op::gather op{axis};
409
410
411
        return prog.add_instruction(op, std::move(args));
    }

412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
    instruction_ref
    parse_slice(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::slice op;
        if(contains(attributes, "axes"))
        {
            literal s = parse_value(attributes.at("axes"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.axes)); });
        }
        {
            literal s = parse_value(attributes.at("ends"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.ends)); });
        }
        {
            literal s = parse_value(attributes.at("starts"));
            s.visit([&](auto v) { copy(v, std::back_inserter(op.starts)); });
        }
        return prog.add_instruction(op, args[0]);
    }

Paul's avatar
Paul committed
432
433
434
    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
Paul's avatar
Paul committed
435
    {
Shucai Xiao's avatar
Shucai Xiao committed
436
        literal v     = parse_value(attributes.at("value"));
437
438
439
        auto dim_size = attributes.at("value").t().dims_size();
        // if dim_size is 0, it is a scalar
        if(dim_size == 0)
440
        {
441
            migraphx::shape scalar_shape{v.get_shape().type(), {1}, {0}};
442
443
444
            return prog.add_literal(migraphx::literal{scalar_shape, v.data()});
        }

Paul's avatar
Paul committed
445
446
        return prog.add_literal(v);
    }
Paul's avatar
Paul committed
447

Paul's avatar
Paul committed
448
    instruction_ref
Paul's avatar
Paul committed
449
    parse_gemm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
450
451
    {
        float alpha = 1.0f;
Khalique's avatar
Khalique committed
452
        float beta  = 1.0f;
Paul's avatar
Paul committed
453
454
455
456
457
458
459
460
        bool transa = false;
        bool transb = false;
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        if(contains(attributes, "beta"))
        {
461
            beta = parse_value(attributes.at("beta")).at<float>();
Paul's avatar
Paul committed
462
463
464
465
466
467
468
469
470
        }
        if(contains(attributes, "transA"))
        {
            transa = parse_value(attributes.at("transA")).at<bool>();
        }
        if(contains(attributes, "transB"))
        {
            transb = parse_value(attributes.at("transB")).at<bool>();
        }
471
472
473
474

        // beginning or end of both args have dimension 1, need to squeeze
        // before calling gemm, then doing unsqueeze after getting results
        std::size_t num_squeeze = args[0]->get_shape().lens().size();
Shucai Xiao's avatar
Shucai Xiao committed
475
        if(num_squeeze > 2)
476
477
478
479
480
481
482
        {
            std::vector<int64_t> vec_axises(num_squeeze - 2);
            std::iota(vec_axises.begin(), vec_axises.end(), 0);
            args[0] = prog.add_instruction(op::squeeze{vec_axises}, args[0]);
            args[1] = prog.add_instruction(op::squeeze{vec_axises}, args[1]);
        }

Paul's avatar
Paul committed
483
        std::vector<int64_t> perm = {1, 0};
484
485
        auto l1 = (transa) ? prog.add_instruction(op::transpose{perm}, args[0]) : args[0];
        auto l2 = (transb) ? prog.add_instruction(op::transpose{perm}, args[1]) : args[1];
Paul's avatar
Paul committed
486
487
        if(args.size() == 3)
        {
Khalique's avatar
Khalique committed
488
            if(beta != 0.f)
489
            {
Khalique's avatar
Khalique committed
490
                auto l3 = prog.add_instruction(op::dot{alpha}, l1, l2);
Shucai Xiao's avatar
Shucai Xiao committed
491
                if(num_squeeze > 2)
492
493
494
495
496
497
                {
                    std::vector<int64_t> vec_axises(num_squeeze - 2);
                    std::iota(vec_axises.begin(), vec_axises.end(), 0);
                    l3 = prog.add_instruction(op::unsqueeze{vec_axises}, l3);
                }

Khalique's avatar
Khalique committed
498
                auto l4 = args[2];
Khalique's avatar
Khalique committed
499
                if(l4->get_shape().scalar()) // ignore args[2] (no C value added to alpha*A*B)
Khalique's avatar
Khalique committed
500
                    return l3;
Khalique's avatar
Khalique committed
501
                if(beta != 1.f)
Khalique's avatar
Khalique committed
502
503
                {
                    auto beta_val = prog.add_literal(beta);
Khalique's avatar
Khalique committed
504
505
                    auto l5 = prog.add_instruction(op::scalar{args[2]->get_shape()}, beta_val);
                    l4      = prog.add_instruction(op::mul{}, args[2], l5);
Khalique's avatar
Khalique committed
506
507
                }
                return add_broadcastable_binary_op(l3, l4, op::add{});
508
            }
Paul's avatar
Paul committed
509
        }
510
511

        auto dot_res = prog.add_instruction(op::dot{alpha, beta}, l1, l2);
Shucai Xiao's avatar
Shucai Xiao committed
512
        if(num_squeeze > 2)
513
514
515
516
517
518
519
        {
            std::vector<int64_t> vec_axises(num_squeeze - 2);
            std::iota(vec_axises.begin(), vec_axises.end(), 0);
            dot_res = prog.add_instruction(op::unsqueeze{vec_axises}, dot_res);
        }

        return dot_res;
Paul's avatar
Paul committed
520
521
    }

522
    instruction_ref
Paul's avatar
Paul committed
523
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
524
    {
Scott Thornton's avatar
Scott Thornton committed
525
526
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
527
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
Scott Thornton's avatar
Scott Thornton committed
528
        bool is_test                                      = false;
529
530
531
532
533
534
        if(contains(attributes, "epsilon"))
        {
            epsilon = parse_value(attributes.at("epsilon")).at<float>();
        }
        if(contains(attributes, "momentum"))
        {
535
            momentum = parse_value(attributes.at("momentum")).at<float>();
536
537
538
        }
        if(contains(attributes, "is_test"))
        {
wsttiger's avatar
wsttiger committed
539
            is_test = parse_value(attributes.at("is_test")).at<uint64_t>() > 0;
540
541
542
        }
        if(contains(attributes, "spatial"))
        {
543
            bn_mode = (parse_value(attributes.at("spatial")).at<uint64_t>() > 0)
544
545
                          ? op::batch_norm_inference::spatial
                          : op::batch_norm_inference::per_activation;
546
        }
Paul's avatar
Paul committed
547
        (void)is_test;
Paul's avatar
Paul committed
548
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
Paul's avatar
Paul committed
549
        return prog.add_instruction(op, std::move(args));
550
551
    }

552
553
554
555
    instruction_ref parse_leaky_relu(const std::string&,
                                     attribute_map attributes,
                                     std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
556
        float alpha = 0.01; // default alpha val for leaky relu
557
558
559
560
561
562
563
564
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        op::leaky_relu op{alpha};
        return prog.add_instruction(op, args.front());
    }

Khalique's avatar
Khalique committed
565
566
    instruction_ref
    parse_elu(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
567
568
569
570
571
572
573
574
575
576
    {
        float alpha = 1.0; // default alpha val for elu
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        op::elu op{alpha};
        return prog.add_instruction(op, args.front());
    }

Khalique's avatar
Khalique committed
577
578
    instruction_ref
    parse_lrn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
579
580
    {
        float alpha = 0.0001;
Khalique's avatar
Khalique committed
581
582
583
        float beta  = 0.75;
        float bias  = 1.0;
        int size    = 1;
Khalique's avatar
Khalique committed
584
585
586
587
588
589
590
591
592
593
594
595
        if(contains(attributes, "alpha"))
            alpha = parse_value(attributes.at("alpha")).at<float>();
        if(contains(attributes, "beta"))
            beta = parse_value(attributes.at("beta")).at<float>();
        if(contains(attributes, "bias"))
            bias = parse_value(attributes.at("bias")).at<float>();
        if(contains(attributes, "size"))
            size = parse_value(attributes.at("size")).at<int>();
        op::lrn op{alpha, beta, bias, size};
        return prog.add_instruction(op, args.front());
    }

Khalique's avatar
Khalique committed
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
    instruction_ref parse_imagescaler(const std::string&,
                                      attribute_map attributes,
                                      std::vector<instruction_ref> args)
    {
        float scale = 1.0;
        std::vector<float> bias{};
        if(contains(attributes, "scale"))
        {
            scale = parse_value(attributes.at("scale")).at<float>();
        }

        if(contains(attributes, "bias"))
        {
            auto&& bias_floats = attributes["bias"].floats();
            bias               = std::vector<float>(bias_floats.begin(), bias_floats.end());
        }
        auto input_shape = args.front()->get_shape();
Khalique's avatar
Khalique committed
613

Khalique's avatar
Khalique committed
614
615
        auto scale_val = prog.add_literal(scale);
        auto bias_vals = prog.add_literal(
Paul's avatar
Paul committed
616
            migraphx::literal{migraphx::shape{migraphx::shape::float_type, {bias.size()}}, bias});
Khalique's avatar
Khalique committed
617

Paul's avatar
Paul committed
618
619
        auto scale_tensor = prog.add_instruction(migraphx::op::scalar{input_shape}, scale_val);
        auto img_scaled   = prog.add_instruction(migraphx::op::mul{}, args.front(), scale_tensor);
Paul's avatar
Paul committed
620
        auto bias_bcast = prog.add_instruction(migraphx::op::broadcast{1, input_shape}, bias_vals);
Paul's avatar
Paul committed
621
        return prog.add_instruction(migraphx::op::add{}, img_scaled, bias_bcast);
Khalique's avatar
Khalique committed
622
    }
Khalique's avatar
Khalique committed
623

Khalique's avatar
Khalique committed
624
625
    instruction_ref
    parse_transpose(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
626
627
628
629
630
631
632
    {
        std::vector<int64_t> perm{};
        if(contains(attributes, "perm"))
        {
            auto&& perm_vals = attributes["perm"].ints();
            perm             = std::vector<int64_t>(perm_vals.begin(), perm_vals.end());
        }
Paul's avatar
Paul committed
633
        return prog.add_instruction(migraphx::op::transpose{perm}, args.front());
Khalique's avatar
Khalique committed
634
635
    }

Khalique's avatar
Khalique committed
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
    instruction_ref
    parse_pad(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        std::vector<int64_t> pads{};
        float value = 0.0f;
        if(contains(attributes, "pads"))
        {
            auto&& pad_vals = attributes["pads"].ints();
            pads            = std::vector<int64_t>(pad_vals.begin(), pad_vals.end());
        }
        if(contains(attributes, "value"))
        {
            value = parse_value(attributes.at("value")).at<float>();
        }
        if(contains(attributes, "mode"))
        {
            auto mode = attributes.at("mode").s();
            if(mode != "constant")
                MIGRAPHX_THROW("migraphx currently only supports constant padding");
        }
        return prog.add_instruction(migraphx::op::pad{pads, value}, args.front());
    }
658
659
660
    // Use a literal instruction to replace the shape since, output of
    // shape operator are literals in migraphx
    instruction_ref
Shucai Xiao's avatar
Shucai Xiao committed
661
    parse_shape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
662
663
    {
        if(args.size() != 1)
664
            MIGRAPHX_THROW("Shape: operator should have 1 operand");
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
        std::vector<std::size_t> arg_shape = args[0]->get_shape().lens();
        std::vector<int64_t> vec_shape(arg_shape.size());
        migraphx::shape s(migraphx::shape::int64_type, {arg_shape.size()});
        std::transform(arg_shape.begin(), arg_shape.end(), vec_shape.begin(), [](auto i) {
            return int64_t(i);
        });
        return prog.add_literal(migraphx::literal{s, vec_shape});
    }

    // Use a literal instruction to replace the constantFill operator. In RNN, input shape
    // and value are fixed, so no need to do the actual computation for the constantFill
    // operator
    instruction_ref parse_constant_fill(const std::string&,
                                        attribute_map attributes,
                                        std::vector<instruction_ref> args)
    {
        int input_as_shape = 0;
        int dtype          = 1;
        float value        = 0.0f;

        if(contains(attributes, "dtype"))
        {
            dtype = parse_value(attributes.at("dtype")).at<int>();
        }
        migraphx::shape::type_t type = get_type(dtype);

        if(contains(attributes, "input_as_shape"))
        {
            input_as_shape = parse_value(attributes.at("input_as_shape")).at<int>();
        }

        if(contains(attributes, "value"))
        {
            value = parse_value(attributes.at("value")).at<float>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
701
702
        if(contains(attributes, "extra_shape"))
        {
703
            MIGRAPHX_THROW("ConstantFill: cannot handle extra shape attribute");
704
705
        }

706
707
        if(input_as_shape == 1)
        {
Shucai Xiao's avatar
Shucai Xiao committed
708
            if(args.size() != 1)
709
            {
710
                MIGRAPHX_THROW("ConstantFill: need an input argument as output shape");
711
712
            }

Shucai Xiao's avatar
Shucai Xiao committed
713
714
            if(contains(attributes, "shape"))
            {
715
                MIGRAPHX_THROW("ConstantFill: cannot set the shape argument and pass in an input "
Shucai Xiao's avatar
Shucai Xiao committed
716
                               "at the same time");
717
718
            }

719
720
721
            migraphx::argument in = args[0]->eval();
            if(in.empty())
            {
722
                MIGRAPHX_THROW("ConstantFill: cannot handle dynamic shape as input");
723
            }
724

725
726
727
            std::vector<std::size_t> dims;
            in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
            migraphx::shape s(type, dims);
728
729
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
730
731
732
        }
        else if(input_as_shape == 0)
        {
Shucai Xiao's avatar
Shucai Xiao committed
733
734
            if(!contains(attributes, "shape"))
            {
735
                MIGRAPHX_THROW("ConstantFill: attribute output shape is needed");
736
737
738
            }

            literal ls = parse_value(attributes.at("shape"));
739
            std::vector<std::size_t> dims;
Shucai Xiao's avatar
Shucai Xiao committed
740
            ls.visit([&](auto s) { dims.assign(s.begin(), s.end()); });
741
            migraphx::shape s{type, dims};
742
743
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
744
745
746
        }
        else
        {
747
            MIGRAPHX_THROW("ConstantFill: wrong value of attribute input_as_shape");
748
749
750
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
751
    std::vector<instruction_ref>
Shucai Xiao's avatar
Shucai Xiao committed
752
753
754
    parse_rnn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        migraphx::shape input_shape = args[0]->get_shape();
755
        std::size_t hidden_size     = args[1]->get_shape().lens()[1];
Shucai Xiao's avatar
Shucai Xiao committed
756
757
758

        if(contains(attributes, "hidden_size"))
        {
Shucai Xiao's avatar
Shucai Xiao committed
759
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
760
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
761
762
763
            {
                MIGRAPHX_THROW("RNN: hidden size mismatch in input and attribute");
            }
Shucai Xiao's avatar
Shucai Xiao committed
764
765
766
767
768
769
770
771
772
        }

        // Handling of direction to be added later
        std::string direction{"forward"};
        if(contains(attributes, "direction"))
        {
            direction = attributes.at("direction").s();
        }

773
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
774
775
        if(direction == "bidirectional")
        {
776
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
777
778
779
        }
        else if(direction == "reverse")
        {
780
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
781
782
        }

783
784
785
786
787
        std::vector<std::string> vec_names{"tanh"};
        if(contains(attributes, "activations"))
        {
            auto names = attributes.at("activations").strings();
            vec_names.clear();
Shucai Xiao's avatar
Shucai Xiao committed
788
            for_each(names.begin(), names.end(), [&](auto& fn) { vec_names.push_back(fn); });
789
790
        }

Shucai Xiao's avatar
Shucai Xiao committed
791
        for_each(vec_names.begin(), vec_names.end(), [&](auto& fn) {
792
793
            if(map_actv_funcs.count(fn) == 0)
            {
Shucai Xiao's avatar
Shucai Xiao committed
794
                MIGRAPHX_THROW("RNN: activation function " + std::string(fn) + " not supported");
795
796
797
            }
        });

Shucai Xiao's avatar
Shucai Xiao committed
798
        // bidirectional case should have two activation functions.
Shucai Xiao's avatar
Shucai Xiao committed
799
        // one is for forward, and the other is for reverse.
Shucai Xiao's avatar
Shucai Xiao committed
800
        // if only one actv function is provided, we use it in both
801
        // forward and reverse direction
802
        if(dirct == op::rnn_direction::bidirectional)
803
        {
Shucai Xiao's avatar
Shucai Xiao committed
804
            if(vec_names.size() == 1)
805
806
807
808
809
            {
                vec_names.push_back(vec_names.at(0));
            }
        }

Shucai Xiao's avatar
Shucai Xiao committed
810
811
812
        std::vector<operation> vec_actv_funcs(vec_names.size());
        std::transform(vec_names.begin(), vec_names.end(), vec_actv_funcs.begin(), [&](auto& fn) {
            return map_actv_funcs[fn];
813
        });
Shucai Xiao's avatar
Shucai Xiao committed
814

Shucai Xiao's avatar
Shucai Xiao committed
815
816
817
818
819
820
821
        // To be added later
        float clip = 0.0;
        if(contains(attributes, "clip"))
        {
            clip = parse_value(attributes.at("clip")).at<float>();
        }

822
823
        // if the number of arguments is less than 6, append
        // undefined operator to have 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
824
        if(args.size() < 6)
825
826
827
828
829
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), (6 - args.size()), ins);
        }

Shucai Xiao's avatar
Shucai Xiao committed
830
831
        // first output for the concatenation of hidden states
        auto hidden_states = prog.add_instruction(op::rnn{hidden_size, vec_actv_funcs, dirct, clip},
Shucai Xiao's avatar
Shucai Xiao committed
832
                                                  std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
833

834
        // second output for the last hidden state
Shucai Xiao's avatar
Shucai Xiao committed
835
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
836

Shucai Xiao's avatar
Shucai Xiao committed
837
        return {hidden_states, last_output};
Shucai Xiao's avatar
Shucai Xiao committed
838
839
    }

840
    std::vector<instruction_ref>
841
842
843
844
845
846
847
    parse_gru(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        migraphx::shape input_shape = args[0]->get_shape();
        std::size_t hidden_size     = args[2]->get_shape().lens()[2];

        if(contains(attributes, "hidden_size"))
        {
Shucai Xiao's avatar
Shucai Xiao committed
848
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
849
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
850
851
852
            {
                MIGRAPHX_THROW("GRU: hidden size mismatch in input and attribute");
            }
853
854
855
856
857
858
859
860
861
        }

        // Handling of direction to be added later
        std::string direction{"forward"};
        if(contains(attributes, "direction"))
        {
            direction = attributes.at("direction").s();
        }

862
        op::rnn_direction dirct = op::rnn_direction::forward;
863
864
        if(direction == "bidirectional")
        {
865
            dirct = op::rnn_direction::bidirectional;
866
867
868
        }
        else if(direction == "reverse")
        {
869
            dirct = op::rnn_direction::reverse;
870
871
        }

872
        std::vector<std::string> vec_names = {"sigmoid", "tanh"};
873
874
        if(contains(attributes, "activations"))
        {
875
            auto names = attributes.at("activations").strings();
876
            vec_names.clear();
Shucai Xiao's avatar
Shucai Xiao committed
877
            vec_names.resize(names.size());
Shucai Xiao's avatar
Shucai Xiao committed
878
879
            std::transform(
                names.begin(), names.end(), vec_names.begin(), [](auto& str) { return str; });
880
881
        }

882
        // need 4 activation functions
883
        if(dirct == op::rnn_direction::bidirectional)
884
        {
Shucai Xiao's avatar
Shucai Xiao committed
885
            // 4 activation functions are used in the bidirectional
886
            // scenario. No spec is provided in onnx::operator. we
Shucai Xiao's avatar
Shucai Xiao committed
887
888
            // use the algorithm that: if 1 actv function is provided,
            // repeat 1 four times. If 2 actv functins are provided,
889
890
            // assume forward and reverse use the same pair of actv
            // functions. For the case of 3 actv functions provided,
Shucai Xiao's avatar
Shucai Xiao committed
891
892
893
            // assume the 3rd one is repeated once and used by the
            // reverse direction.
            // This may need change later
894
            if(vec_names.size() == 1)
895
            {
896
                vec_names.insert(vec_names.end(), 3, vec_names.at(0));
897
            }
898
            else if(vec_names.size() == 2)
899
            {
900
901
902
                // repeat the activation functions
                vec_names.push_back(vec_names.at(0));
                vec_names.push_back(vec_names.at(1));
903
            }
904
            else if(vec_names.size() == 3)
905
            {
906
                vec_names.push_back(vec_names.at(2));
907
908
            }
        }
Shucai Xiao's avatar
Shucai Xiao committed
909
        else
910
        {
911
            if(vec_names.size() == 1)
912
            {
913
                vec_names.push_back(vec_names.at(0));
914
915
916
            }
        }

917
        for_each(vec_names.begin(), vec_names.end(), [&](auto& name) {
Shucai Xiao's avatar
Shucai Xiao committed
918
919
            if(map_actv_funcs.count(name) == 0)
            {
Shucai Xiao's avatar
Shucai Xiao committed
920
                MIGRAPHX_THROW("GRU: activation function " + std::string(name) + " not supported");
Shucai Xiao's avatar
Shucai Xiao committed
921
922
            }
        });
923

Shucai Xiao's avatar
Shucai Xiao committed
924
925
926
        std::vector<operation> vec_actv_funcs(vec_names.size());
        std::transform(vec_names.begin(), vec_names.end(), vec_actv_funcs.begin(), [&](auto& name) {
            return map_actv_funcs[name];
Shucai Xiao's avatar
Shucai Xiao committed
927
        });
928
929
930
931
932
933
934
935

        float clip = 0.0;
        if(contains(attributes, "clip"))
        {
            clip = parse_value(attributes.at("clip")).at<float>();
        }

        int linear_before_reset = 0;
Shucai Xiao's avatar
Shucai Xiao committed
936
        if(contains(attributes, "linear_before_reset"))
937
938
939
940
        {
            linear_before_reset = parse_value(attributes.at("linear_before_reset")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
941
        // append undefined opeator to make 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
942
        if(args.size() < 6)
Shucai Xiao's avatar
Shucai Xiao committed
943
944
945
946
947
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), 6 - args.size(), ins);
        }

948
949
        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
950
            op::gru{hidden_size, vec_actv_funcs, dirct, clip, linear_before_reset},
Shucai Xiao's avatar
Shucai Xiao committed
951
            std::move(args));
952
953

        // second output for last gru output
954
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
955

Shucai Xiao's avatar
Shucai Xiao committed
956
        return {hidden_states, last_output};
957
958
    }

Paul's avatar
Paul committed
959
960
961
962
963
964
965
966
967
968
969
970
    void parse_from(std::istream& is)
    {
        onnx::ModelProto model;
        if(model.ParseFromIstream(&is))
        {
            if(model.has_graph())
            {
                this->parse_graph(model.graph());
            }
        }
        else
        {
Paul's avatar
Paul committed
971
            MIGRAPHX_THROW("Failed reading onnx file.");
Paul's avatar
Paul committed
972
973
974
975
976
977
        }
    }

    void parse_graph(const onnx::GraphProto& graph)
    {
        nodes = get_nodes(graph);
978
979
980
981
982
        std::unordered_map<std::string, onnx::TensorProto> initializer_data;
        for(auto&& f : graph.initializer())
        {
            initializer_data[f.name()] = f;
        }
Paul's avatar
Paul committed
983
984
985
        for(auto&& input : graph.input())
        {
            const std::string& name = input.name();
986
987
988
989
990
991
992
993
994
995
996
997
            // Does the input have an initializer?
            if(contains(initializer_data, name))
            {
                auto t             = initializer_data[name];
                instructions[name] = prog.add_literal(parse_tensor(t));
            }
            else
            {
                // TODO: Get shape of input parameter
                shape s            = parse_type(input.type());
                instructions[name] = prog.add_parameter(name, s);
            }
Paul's avatar
Paul committed
998
999
1000
        }
        for(auto&& p : nodes)
        {
Paul's avatar
Paul committed
1001
            this->parse_node(p.first);
Paul's avatar
Paul committed
1002
1003
1004
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
1005
    void parse_undefined(const std::string& name)
1006
    {
Shucai Xiao's avatar
Shucai Xiao committed
1007
        auto ins           = prog.add_instruction(op::undefined{});
1008
1009
1010
        instructions[name] = ins;
    }

Paul's avatar
Paul committed
1011
    void parse_node(const std::string& name)
Paul's avatar
Paul committed
1012
    {
Paul's avatar
Paul committed
1013
        if(name.empty())
Paul's avatar
Paul committed
1014
            MIGRAPHX_THROW("Onnx node must have a name");
Paul's avatar
Paul committed
1015
1016
1017
1018
1019
1020
1021
1022
        if(instructions.count(name) == 0)
        {
            auto&& node = nodes.at(name);
            std::vector<instruction_ref> args;
            for(auto&& input : node.input())
            {
                if(nodes.count(input) > 0)
                {
Paul's avatar
Paul committed
1023
1024
                    assert(name != input);
                    this->parse_node(input);
Paul's avatar
Paul committed
1025
                }
Shucai Xiao's avatar
Shucai Xiao committed
1026
                else if(input.empty())
Paul's avatar
Paul committed
1027
                {
1028
                    this->parse_undefined(input);
Paul's avatar
Paul committed
1029
                }
1030
                args.push_back(instructions.at(input));
Paul's avatar
Paul committed
1031
            }
Paul's avatar
Paul committed
1032
            std::vector<instruction_ref> result;
Paul's avatar
Paul committed
1033
1034
            if(ops.count(node.op_type()) == 0)
            {
Paul's avatar
Paul committed
1035
                result.push_back(prog.add_instruction(unknown{node.op_type()}, args));
Paul's avatar
Paul committed
1036
1037
1038
            }
            else
            {
Paul's avatar
Paul committed
1039
                result = ops[node.op_type()](get_attributes(node), args);
Paul's avatar
Paul committed
1040
            }
Paul's avatar
Paul committed
1041
            // Even no output nodes produce output in migraphx
Paul's avatar
Paul committed
1042
            if(node.output().empty() and result.size() == 1)
Paul's avatar
Paul committed
1043
1044
            {
                instructions[name] = result.front();
Paul's avatar
Paul committed
1045
1046
1047
            }
            else
            {
Paul's avatar
Paul committed
1048
1049
1050
1051
1052
1053
                assert(node.output().size() >= result.size());
                std::transform(result.begin(),
                               result.end(),
                               node.output().begin(),
                               std::inserter(instructions, instructions.end()),
                               [](auto&& x, auto&& y) { return std::make_pair(y, x); });
Paul's avatar
Paul committed
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
            }
        }
    }

    static attribute_map get_attributes(const onnx::NodeProto& node)
    {
        std::unordered_map<std::string, onnx::AttributeProto> result;
        for(auto&& attr : node.attribute())
        {
            result[attr.name()] = attr;
        }
        return result;
    }

    static node_map get_nodes(const onnx::GraphProto& graph)
    {
        std::unordered_map<std::string, onnx::NodeProto> result;
Paul's avatar
Paul committed
1071
        std::size_t n = 0;
Paul's avatar
Paul committed
1072
1073
        for(auto&& node : graph.node())
        {
Paul's avatar
Paul committed
1074
            if(node.output().empty())
Paul's avatar
Paul committed
1075
            {
Paul's avatar
Paul committed
1076
                if(node.name().empty())
Paul's avatar
Paul committed
1077
1078
1079
1080
1081
1082
1083
1084
1085
                {
                    result["migraphx_unamed_node_" + std::to_string(n)] = node;
                    n++;
                }
                else
                {
                    result[node.name()] = node;
                }
            }
Paul's avatar
Paul committed
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
            for(auto&& output : node.output())
            {
                result[output] = node;
            }
        }
        return result;
    }

    template <class T>
    static literal from_repeated(shape::type_t t, const T& r)
    {
        std::size_t size = r.size();
        return literal{{t, {size}}, r.begin(), r.end()};
    }

    static literal parse_value(const onnx::AttributeProto& attr)
    {
        switch(attr.type())
        {
        case onnx::AttributeProto::UNDEFINED: return {};
        case onnx::AttributeProto::FLOAT: return literal{attr.f()};
        case onnx::AttributeProto::INT: return literal{attr.i()};
        case onnx::AttributeProto::STRING: return {};
        case onnx::AttributeProto::TENSOR: return parse_tensor(attr.t());
        case onnx::AttributeProto::GRAPH: return {};
Paul's avatar
Paul committed
1111
        case onnx::AttributeProto::FLOATS: return from_repeated(shape::float_type, attr.floats());
Paul's avatar
Paul committed
1112
1113
1114
1115
1116
        case onnx::AttributeProto::INTS: return from_repeated(shape::int64_type, attr.ints());
        case onnx::AttributeProto::STRINGS: return {};
        case onnx::AttributeProto::TENSORS: return {};
        case onnx::AttributeProto::GRAPHS: return {};
        }
Paul's avatar
Paul committed
1117
        MIGRAPHX_THROW("Invalid attribute type");
Paul's avatar
Paul committed
1118
1119
1120
1121
1122
    }

    static literal parse_tensor(const onnx::TensorProto& t)
    {
        std::vector<std::size_t> dims(t.dims().begin(), t.dims().end());
Khalique's avatar
Khalique committed
1123
        // in case of scalar constants in onnx file, use dims=1 to fill initializer data
1124
        if(dims.empty())
Khalique's avatar
Khalique committed
1125
1126
1127
        {
            dims = {1};
        }
1128
1129
        if(t.has_raw_data())
        {
wsttiger's avatar
wsttiger committed
1130
            const std::string& s = t.raw_data();
Scott Thornton's avatar
Scott Thornton committed
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
            switch(t.data_type())
            {
            case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
            case onnx::TensorProto::FLOAT: return literal{{shape::float_type, dims}, s.data()};
            case onnx::TensorProto::UINT8: throw std::runtime_error("");
            case onnx::TensorProto::INT8: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::UINT16: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT16: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT32: return literal{{shape::int32_type, dims}, s.data()};
            case onnx::TensorProto::INT64: return literal{{shape::int64_type, dims}, s.data()};
            case onnx::TensorProto::STRING: throw std::runtime_error("");
            case onnx::TensorProto::BOOL: return literal{{shape::int32_type, dims}, s.data()};
Paul's avatar
Paul committed
1143
            case onnx::TensorProto::FLOAT16: return literal{{shape::half_type, dims}, s.data()};
Scott Thornton's avatar
Scott Thornton committed
1144
1145
1146
1147
1148
1149
            case onnx::TensorProto::DOUBLE: return literal{{shape::double_type, dims}, s.data()};
            case onnx::TensorProto::UINT32: throw std::runtime_error("");
            case onnx::TensorProto::UINT64: throw std::runtime_error("");
            case onnx::TensorProto::COMPLEX64: throw std::runtime_error("");
            case onnx::TensorProto::COMPLEX128: throw std::runtime_error("");
            }
Paul's avatar
Paul committed
1150
            MIGRAPHX_THROW("Invalid tensor type");
1151
        }
Paul's avatar
Paul committed
1152
1153
1154
1155
        switch(t.data_type())
        {
        case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
        case onnx::TensorProto::FLOAT:
Paul's avatar
Paul committed
1156
            return literal{{shape::float_type, dims}, t.float_data().begin(), t.float_data().end()};
Paul's avatar
Paul committed
1157
1158
        case onnx::TensorProto::UINT8: throw std::runtime_error("");
        case onnx::TensorProto::INT8:
Paul's avatar
Paul committed
1159
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1160
        case onnx::TensorProto::UINT16:
Paul's avatar
Paul committed
1161
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1162
        case onnx::TensorProto::INT16:
Paul's avatar
Paul committed
1163
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1164
        case onnx::TensorProto::INT32:
Paul's avatar
Paul committed
1165
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1166
        case onnx::TensorProto::INT64:
Paul's avatar
Paul committed
1167
            return literal{{shape::int64_type, dims}, t.int64_data().begin(), t.int64_data().end()};
Paul's avatar
Paul committed
1168
1169
        case onnx::TensorProto::STRING: throw std::runtime_error("");
        case onnx::TensorProto::BOOL:
Paul's avatar
Paul committed
1170
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1171
        case onnx::TensorProto::FLOAT16:
Khalique's avatar
Khalique committed
1172
        {
Khalique's avatar
Khalique committed
1173
            std::vector<uint16_t> data_uint16(t.int32_data().begin(), t.int32_data().end());
1174
            std::vector<half> data_half;
Khalique's avatar
Khalique committed
1175
1176
1177
            std::transform(data_uint16.begin(),
                           data_uint16.end(),
                           std::back_inserter(data_half),
1178
                           [](uint16_t raw_val) { return *reinterpret_cast<half*>(&raw_val); });
1179
            return literal{{shape::half_type, dims}, data_half.begin(), data_half.end()};
Khalique's avatar
Khalique committed
1180
        }
Paul's avatar
Paul committed
1181
1182
1183
1184
1185
1186
1187
1188
        case onnx::TensorProto::DOUBLE:
            return literal{
                {shape::double_type, dims}, t.double_data().begin(), t.double_data().end()};
        case onnx::TensorProto::UINT32: throw std::runtime_error("");
        case onnx::TensorProto::UINT64: throw std::runtime_error("");
        case onnx::TensorProto::COMPLEX64: throw std::runtime_error("");
        case onnx::TensorProto::COMPLEX128: throw std::runtime_error("");
        }
Paul's avatar
Paul committed
1189
        MIGRAPHX_THROW("Invalid tensor type");
Paul's avatar
Paul committed
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
    }

    static shape parse_type(const onnx::TypeProto& t)
    {
        shape::type_t shape_type{};
        switch(t.tensor_type().elem_type())
        {
        case onnx::TensorProto::UNDEFINED:
            break; // throw std::runtime_error("Unsupported type UNDEFINED");
        case onnx::TensorProto::FLOAT: shape_type = shape::float_type; break;
        case onnx::TensorProto::UINT8:
            break; // throw std::runtime_error("Unsupported type UINT8");
        case onnx::TensorProto::INT8: shape_type = shape::int8_type; break;
        case onnx::TensorProto::UINT16: shape_type = shape::uint16_type; break;
        case onnx::TensorProto::INT16: shape_type = shape::int16_type; break;
        case onnx::TensorProto::INT32: shape_type = shape::int32_type; break;
        case onnx::TensorProto::INT64: shape_type = shape::int64_type; break;
        case onnx::TensorProto::STRING:
            break; // throw std::runtime_error("Unsupported type STRING");
        case onnx::TensorProto::BOOL:
            break; // throw std::runtime_error("Unsupported type BOOL");
Paul's avatar
Paul committed
1211
        case onnx::TensorProto::FLOAT16: shape_type = shape::half_type; break;
Paul's avatar
Paul committed
1212
1213
1214
1215
1216
1217
1218
1219
1220
        case onnx::TensorProto::DOUBLE: shape_type = shape::double_type; break;
        case onnx::TensorProto::UINT32: shape_type = shape::uint32_type; break;
        case onnx::TensorProto::UINT64: shape_type = shape::uint64_type; break;
        case onnx::TensorProto::COMPLEX64:
            break; // throw std::runtime_error("Unsupported type COMPLEX64");
        case onnx::TensorProto::COMPLEX128:
            break; // throw std::runtime_error("Unsupported type COMPLEX128");
        }
        std::vector<std::size_t> dims;
Paul's avatar
Paul committed
1221
        auto&& tensor_dims = t.tensor_type().shape().dim();
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
        std::transform(tensor_dims.begin(),
                       tensor_dims.end(),
                       std::back_inserter(dims),
                       [](auto&& d) -> std::size_t {
                           if(not d.has_dim_value())
                           {
                               long default_batch_size = 1; // FIXME
                               return default_batch_size;
                           }
                           return d.dim_value();
                       });
Paul's avatar
Paul committed
1233
1234
        return {shape_type, dims};
    }
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256

    shape::type_t get_type(int dtype)
    {
        switch(dtype)
        {
        case 1: return shape::float_type;
        case 2: return shape::uint8_type;
        case 3: return shape::int8_type;
        case 4: return shape::uint16_type;
        case 5: return shape::int16_type;
        case 6: return shape::int32_type;
        case 7: return shape::int64_type;
        case 10: return shape::half_type;
        case 11: return shape::double_type;
        case 12: return shape::uint32_type;
        case 13: return shape::uint64_type;
        default:
        {
            MIGRAPHX_THROW("Prototensor data type " + std::to_string(dtype) + " not supported");
        }
        }
    }
Paul's avatar
Paul committed
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
};

program parse_onnx(const std::string& name)
{
    std::fstream input(name.c_str(), std::ios::in | std::ios::binary);
    onnx_parser parser;
#ifndef NDEBUG
    // Log the program when it can't be parsed
    try
    {
        parser.parse_from(input);
    }
    catch(...)
    {
        std::cerr << parser.prog << std::endl;
        throw;
    }
#else
    parser.parse_from(input);
#endif
    return std::move(parser.prog);
}

Paul's avatar
Paul committed
1280
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1281
} // namespace migraphx