onnx.cpp 53.5 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);
Shucai Xiao's avatar
Shucai Xiao committed
82
        add_mem_op("LogSoftmax", &onnx_parser::parse_logsoftmax);
83
84
85
        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
86
        add_mem_op("Concat", &onnx_parser::parse_concat);
87
88
89
        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
90
        add_mem_op("Transpose", &onnx_parser::parse_transpose);
Shucai Xiao's avatar
Shucai Xiao committed
91
        add_mem_op("RNN", &onnx_parser::parse_rnn);
92
        add_mem_op("GRU", &onnx_parser::parse_gru);
Shucai Xiao's avatar
Shucai Xiao committed
93
        add_mem_op("LSTM", &onnx_parser::parse_lstm);
Khalique's avatar
Khalique committed
94
        add_mem_op("Pad", &onnx_parser::parse_pad);
95
96
97
98
99
100
101

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

    void init_actv_func()
    {
102
103
104
105
106
        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
107
108
109
110
    }

    template <class F>
    void add_op(std::string name, F f)
Paul's avatar
Paul committed
111
112
113
114
115
116
117
118
119
    {
        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
120
121
122
123
124
125
126
    {
        ops.emplace(name, f);
    }

    template <class F>
    void add_mem_op(std::string name, F f)
    {
Paul's avatar
Paul committed
127
        add_op(name, [=](auto&&... xs) {
Paul's avatar
Paul committed
128
129
130
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }
Khalique's avatar
Khalique committed
131

132
    template <class T>
Khalique's avatar
Khalique committed
133
    void add_binary_op(std::string name, T x)
134
    {
Paul's avatar
Paul committed
135
        add_op(name, [this, x](attribute_map attributes, std::vector<instruction_ref> args) {
Scott Thornton's avatar
Scott Thornton committed
136
            if(args.size() != 2)
Paul's avatar
Paul committed
137
                MIGRAPHX_THROW("binary operators should have 2 operands");
138
            if(contains(attributes, "broadcast") and contains(attributes, "axis"))
139
140
141
142
            {
                uint64_t broadcasted = parse_value(attributes.at("broadcast")).at<uint64_t>();
                if(broadcasted != 0)
                {
143
                    uint64_t axis = parse_value(attributes.at("axis")).at<uint64_t>();
144
145
146
147
                    auto l =
                        prog.add_instruction(op::broadcast{axis, args[0]->get_shape()}, args[1]);
                    return prog.add_instruction(x, args[0], l);
                }
148
                return prog.add_instruction(x, args);
149
            }
Paul's avatar
Paul committed
150
            else
151
            {
Khalique's avatar
Khalique committed
152
                return add_broadcastable_binary_op(args[0], args[1], x);
153
154
155
156
            }
        });
    }

Khalique's avatar
Khalique committed
157
158
159
160
161
    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
162
163
164
165
166
167
168
169
170
171
172
173
174
            // 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
175
176
177
178
179
180
181
182
            // 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
183
            std::vector<std::size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
184
185
            auto offset = s1->size() - s0->size();
            std::transform(s0->begin(),
Khalique's avatar
Khalique committed
186
187
188
189
                           s0->end(),
                           s1->begin() + offset,
                           output_lens.begin() + offset,
                           [](auto a, auto b) { return std::max(a, b); });
Khalique's avatar
Khalique committed
190
191
192
193
194
195
196
197
198

            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});
        }
199
200
    }

Paul's avatar
Paul committed
201
    template <class T>
Paul's avatar
Paul committed
202
203
    void add_generic_op(std::string name, T x)
    {
Paul's avatar
Paul committed
204
        add_op(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
Paul's avatar
Paul committed
205
206
207
208
            return prog.add_instruction(x, args);
        });
    }

Khalique's avatar
Khalique committed
209
    template <class T>
Khalique's avatar
Khalique committed
210
    void add_variadic_op(std::string name, T x)
Khalique's avatar
Khalique committed
211
    {
Paul's avatar
Paul committed
212
        add_op(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
213
            return std::accumulate(std::next(args.begin()),
Khalique's avatar
Khalique committed
214
215
216
217
218
                                   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
219
        });
Khalique's avatar
Khalique committed
220
221
    }

Paul's avatar
Paul committed
222
    instruction_ref
Paul's avatar
Paul committed
223
    parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
224
225
    {
        auto dims = args.front()->get_shape().lens();
Scott Thornton's avatar
Scott Thornton committed
226
227
        auto r =
            prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, args.front());
228
229
        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
230
231
    }

Shucai Xiao's avatar
Shucai Xiao committed
232
233
234
    instruction_ref parse_logsoftmax(const std::string&,
                                     const attribute_map& attributes,
                                     std::vector<instruction_ref> args)
Shucai Xiao's avatar
Shucai Xiao committed
235
236
237
238
239
240
241
242
243
244
    {
        int axis = 1;
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }

        return prog.add_instruction(op::logsoftmax{axis}, std::move(args));
    }

Paul's avatar
Paul committed
245
    instruction_ref
Paul's avatar
Paul committed
246
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
247
    {
248
        op::convolution op;
249
        auto l0 = args[0];
Paul's avatar
Paul committed
250
251
        if(contains(attributes, "pads"))
        {
Scott Thornton's avatar
Scott Thornton committed
252
            if(contains(attributes, "auto_pad"))
253
            {
Paul's avatar
Paul committed
254
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
255
            }
256
257
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
258
            if(padding.size() != 4)
259
            {
Paul's avatar
Paul committed
260
                MIGRAPHX_THROW("padding should have 4 values");
261
            }
Scott Thornton's avatar
Scott Thornton committed
262
            if(padding[0] != padding[2] || padding[1] != padding[3])
263
            {
264
265
                // 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
266
                l0      = prog.add_instruction(op::pad{padding}, l0);
267
            }
268
269
270
271
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
272
            }
Paul's avatar
Paul committed
273
        }
Paul's avatar
Paul committed
274
275
276
277
278
279
280
281
        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
282
        if(contains(attributes, "auto_pad"))
283
284
        {
            auto s = attributes["auto_pad"].s();
Scott Thornton's avatar
Scott Thornton committed
285
            if(contains(attributes, "pads") and to_upper(s) != "NOTSET")
286
            {
Paul's avatar
Paul committed
287
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
288
289
            }

wsttiger's avatar
fixes  
wsttiger committed
290
            if(s.find("SAME") != std::string::npos)
291
            {
292
                op.padding_mode = op::padding_mode_t::same;
293
294
            }
        }
Khalique's avatar
Khalique committed
295
296
297
298
        if(contains(attributes, "group"))
        {
            op.group = parse_value(attributes.at("group")).at<int>();
        }
Paul's avatar
Paul committed
299
300
301
302
        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
303
            auto l2       = prog.add_instruction(op::broadcast{axis, l1->get_shape()}, args[2]);
304
            return prog.add_instruction(op::add{}, l1, l2);
Paul's avatar
Paul committed
305
        }
306
        return prog.add_instruction(op, l0, args[1]);
Paul's avatar
Paul committed
307
    }
Paul's avatar
Paul committed
308

Paul's avatar
Paul committed
309
310
311
    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
Paul's avatar
Paul committed
312
    {
Khalique's avatar
Khalique committed
313
        op::pooling op{ends_with(name, "MaxPool") ? "max" : "average"};
314
        auto l0 = args[0];
Khalique's avatar
Khalique committed
315
        if(starts_with(name, "Global"))
316
        {
Khalique's avatar
Khalique committed
317
318
            auto lens  = args.front()->get_shape().lens();
            op.lengths = {lens[2], lens[3]};
319
        }
Paul's avatar
Paul committed
320
321
        if(contains(attributes, "pads"))
        {
322
323
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
324
            if(padding.size() != 4)
325
            {
Paul's avatar
Paul committed
326
                MIGRAPHX_THROW("padding should have 4 values");
327
            }
Scott Thornton's avatar
Scott Thornton committed
328
            if(padding[0] != padding[2] || padding[1] != padding[3])
329
            {
330
331
                // 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
332
                l0      = prog.add_instruction(op::pad{padding}, l0);
333
334
335
336
337
            }
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
338
            }
Paul's avatar
Paul committed
339
340
341
342
343
344
345
346
347
        }
        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
348
        if(contains(attributes, "auto_pad"))
349
350
        {
            auto s = attributes["auto_pad"].s();
351
            if(s.find("SAME_UPPER") == std::string::npos)
352
            {
353
                MIGRAPHX_THROW("auto_pad only supports SAME_UPPER for pooling");
354
            }
355
            op.padding_mode = op::padding_mode_t::same;
356
357
        }

358
        return prog.add_instruction(op, l0);
Paul's avatar
Paul committed
359
360
    }

Paul's avatar
Paul committed
361
    instruction_ref
Paul's avatar
Paul committed
362
    parse_reshape(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
363
    {
364
        op::reshape op;
Paul's avatar
Paul committed
365
366
367
368
369
370
371
        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
372
            auto s = args[1]->eval();
Paul's avatar
Paul committed
373
            if(s.empty())
Paul's avatar
Paul committed
374
                MIGRAPHX_THROW("Dynamic shape is not supported.");
Paul's avatar
Paul committed
375
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
Paul's avatar
Paul committed
376
        }
Paul's avatar
Paul committed
377
378
379
        return prog.add_instruction(op, args[0]);
    }

Paul's avatar
Paul committed
380
    instruction_ref
Paul's avatar
Paul committed
381
    parse_flatten(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
382
    {
383
        uint64_t axis = 1;
Paul's avatar
Paul committed
384
385
386
387
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
388
        return prog.add_instruction(op::flatten{axis}, args[0]);
Paul's avatar
Paul committed
389
390
    }

391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
    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
409
410
411
412
413
414
415
    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));
    }
416

417
418
419
    instruction_ref
    parse_gather(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
420
        int axis = 0;
421
422
423
424
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
425
        op::gather op{axis};
426
427
428
        return prog.add_instruction(op, std::move(args));
    }

429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
    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
449
450
451
    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
Paul's avatar
Paul committed
452
    {
Shucai Xiao's avatar
Shucai Xiao committed
453
        literal v     = parse_value(attributes.at("value"));
454
455
456
        auto dim_size = attributes.at("value").t().dims_size();
        // if dim_size is 0, it is a scalar
        if(dim_size == 0)
457
        {
458
            migraphx::shape scalar_shape{v.get_shape().type()};
459
460
461
            return prog.add_literal(migraphx::literal{scalar_shape, v.data()});
        }

Paul's avatar
Paul committed
462
463
        return prog.add_literal(v);
    }
Paul's avatar
Paul committed
464

Paul's avatar
Paul committed
465
    instruction_ref
Paul's avatar
Paul committed
466
    parse_gemm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
467
468
    {
        float alpha = 1.0f;
Khalique's avatar
Khalique committed
469
        float beta  = 1.0f;
Paul's avatar
Paul committed
470
471
472
473
474
475
476
477
        bool transa = false;
        bool transb = false;
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        if(contains(attributes, "beta"))
        {
478
            beta = parse_value(attributes.at("beta")).at<float>();
Paul's avatar
Paul committed
479
480
481
482
483
484
485
486
487
        }
        if(contains(attributes, "transA"))
        {
            transa = parse_value(attributes.at("transA")).at<bool>();
        }
        if(contains(attributes, "transB"))
        {
            transb = parse_value(attributes.at("transB")).at<bool>();
        }
488

489
490
491
492
493
        std::vector<int64_t> perm(args[0]->get_shape().lens().size());
        std::iota(perm.begin(), perm.end(), int64_t{0});
        // swap the last two elements
        std::swap(*perm.rbegin(), *(perm.rbegin() + 1));

494
495
        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
496
497
        if(args.size() == 3)
        {
Khalique's avatar
Khalique committed
498
            if(beta != 0.f)
499
            {
Khalique's avatar
Khalique committed
500
                auto l3 = prog.add_instruction(op::dot{alpha}, l1, l2);
Khalique's avatar
Khalique committed
501
                auto l4 = args[2];
Khalique's avatar
Khalique committed
502
                if(l4->get_shape().scalar()) // ignore args[2] (no C value added to alpha*A*B)
Khalique's avatar
Khalique committed
503
                    return l3;
Khalique's avatar
Khalique committed
504
                if(beta != 1.f)
Khalique's avatar
Khalique committed
505
506
                {
                    auto beta_val = prog.add_literal(beta);
Khalique's avatar
Khalique committed
507
508
                    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
509
510
                }
                return add_broadcastable_binary_op(l3, l4, op::add{});
511
            }
Paul's avatar
Paul committed
512
        }
513
        return prog.add_instruction(op::dot{alpha, beta}, l1, l2);
Paul's avatar
Paul committed
514
515
    }

516
    instruction_ref
Paul's avatar
Paul committed
517
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
518
    {
Scott Thornton's avatar
Scott Thornton committed
519
520
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
521
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
Scott Thornton's avatar
Scott Thornton committed
522
        bool is_test                                      = false;
523
524
525
526
527
528
        if(contains(attributes, "epsilon"))
        {
            epsilon = parse_value(attributes.at("epsilon")).at<float>();
        }
        if(contains(attributes, "momentum"))
        {
529
            momentum = parse_value(attributes.at("momentum")).at<float>();
530
531
532
        }
        if(contains(attributes, "is_test"))
        {
wsttiger's avatar
wsttiger committed
533
            is_test = parse_value(attributes.at("is_test")).at<uint64_t>() > 0;
534
535
536
        }
        if(contains(attributes, "spatial"))
        {
537
            bn_mode = (parse_value(attributes.at("spatial")).at<uint64_t>() > 0)
538
539
                          ? op::batch_norm_inference::spatial
                          : op::batch_norm_inference::per_activation;
540
        }
Paul's avatar
Paul committed
541
        (void)is_test;
Paul's avatar
Paul committed
542
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
Paul's avatar
Paul committed
543
        return prog.add_instruction(op, std::move(args));
544
545
    }

546
547
548
549
    instruction_ref parse_leaky_relu(const std::string&,
                                     attribute_map attributes,
                                     std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
550
        float alpha = 0.01; // default alpha val for leaky relu
551
552
553
554
555
556
557
558
        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
559
560
    instruction_ref
    parse_elu(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
561
562
563
564
565
566
567
568
569
570
    {
        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
571
572
    instruction_ref
    parse_lrn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
573
574
    {
        float alpha = 0.0001;
Khalique's avatar
Khalique committed
575
576
577
        float beta  = 0.75;
        float bias  = 1.0;
        int size    = 1;
Khalique's avatar
Khalique committed
578
579
580
581
582
583
584
585
586
587
588
589
        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
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
    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
607

Khalique's avatar
Khalique committed
608
609
        auto scale_val = prog.add_literal(scale);
        auto bias_vals = prog.add_literal(
Paul's avatar
Paul committed
610
            migraphx::literal{migraphx::shape{migraphx::shape::float_type, {bias.size()}}, bias});
Khalique's avatar
Khalique committed
611

Paul's avatar
Paul committed
612
613
        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
614
        auto bias_bcast = prog.add_instruction(migraphx::op::broadcast{1, input_shape}, bias_vals);
Paul's avatar
Paul committed
615
        return prog.add_instruction(migraphx::op::add{}, img_scaled, bias_bcast);
Khalique's avatar
Khalique committed
616
    }
Khalique's avatar
Khalique committed
617

Khalique's avatar
Khalique committed
618
619
    instruction_ref
    parse_transpose(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
620
621
622
623
624
625
626
    {
        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
627
        return prog.add_instruction(migraphx::op::transpose{perm}, args.front());
Khalique's avatar
Khalique committed
628
629
    }

Khalique's avatar
Khalique committed
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
    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());
    }
652
653
654
    // 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
655
    parse_shape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
656
657
    {
        if(args.size() != 1)
658
            MIGRAPHX_THROW("Shape: operator should have 1 operand");
659
660
661
662
663
664
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
        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
695
696
        if(contains(attributes, "extra_shape"))
        {
697
            MIGRAPHX_THROW("ConstantFill: cannot handle extra shape attribute");
698
699
        }

700
701
        if(input_as_shape == 1)
        {
Shucai Xiao's avatar
Shucai Xiao committed
702
            if(args.size() != 1)
703
            {
704
                MIGRAPHX_THROW("ConstantFill: need an input argument as output shape");
705
706
            }

Shucai Xiao's avatar
Shucai Xiao committed
707
708
            if(contains(attributes, "shape"))
            {
709
                MIGRAPHX_THROW("ConstantFill: cannot set the shape argument and pass in an input "
Shucai Xiao's avatar
Shucai Xiao committed
710
                               "at the same time");
711
712
            }

713
714
715
            migraphx::argument in = args[0]->eval();
            if(in.empty())
            {
716
                MIGRAPHX_THROW("ConstantFill: cannot handle dynamic shape as input");
717
            }
718

719
720
721
            std::vector<std::size_t> dims;
            in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
            migraphx::shape s(type, dims);
722
723
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
724
725
726
        }
        else if(input_as_shape == 0)
        {
Shucai Xiao's avatar
Shucai Xiao committed
727
728
            if(!contains(attributes, "shape"))
            {
729
                MIGRAPHX_THROW("ConstantFill: attribute output shape is needed");
730
731
732
            }

            literal ls = parse_value(attributes.at("shape"));
733
            std::vector<std::size_t> dims;
Shucai Xiao's avatar
Shucai Xiao committed
734
            ls.visit([&](auto s) { dims.assign(s.begin(), s.end()); });
735
            migraphx::shape s{type, dims};
736
737
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
738
739
740
        }
        else
        {
741
            MIGRAPHX_THROW("ConstantFill: wrong value of attribute input_as_shape");
742
743
744
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
745
    std::vector<instruction_ref>
Shucai Xiao's avatar
Shucai Xiao committed
746
747
748
    parse_rnn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        migraphx::shape input_shape = args[0]->get_shape();
749
        std::size_t hidden_size     = args[1]->get_shape().lens()[1];
Shucai Xiao's avatar
Shucai Xiao committed
750
751
752

        if(contains(attributes, "hidden_size"))
        {
Shucai Xiao's avatar
Shucai Xiao committed
753
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
754
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
755
756
757
            {
                MIGRAPHX_THROW("RNN: hidden size mismatch in input and attribute");
            }
Shucai Xiao's avatar
Shucai Xiao committed
758
759
760
761
762
763
764
765
766
        }

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

767
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
768
769
        if(direction == "bidirectional")
        {
770
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
771
772
773
        }
        else if(direction == "reverse")
        {
774
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
775
776
        }

777
778
779
780
781
        std::vector<std::string> vec_names{"tanh"};
        if(contains(attributes, "activations"))
        {
            auto names = attributes.at("activations").strings();
            vec_names.clear();
782
            vec_names.resize(names.size());
783
            std::copy(names.begin(), names.end(), vec_names.begin());
784
785
        }

786
787
788
        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
            return (map_actv_funcs.count(name) == 0);
        });
Shucai Xiao's avatar
Shucai Xiao committed
789
        if(name_it != vec_names.end())
790
791
792
        {
            MIGRAPHX_THROW("RNN: activation function " + std::string(*name_it) + " not supported");
        }
793

Shucai Xiao's avatar
Shucai Xiao committed
794
        // bidirectional case should have two activation functions.
Shucai Xiao's avatar
Shucai Xiao committed
795
        // one is for forward, and the other is for reverse.
Shucai Xiao's avatar
Shucai Xiao committed
796
        // if only one actv function is provided, we use it in both
797
        // forward and reverse direction
798
        if(dirct == op::rnn_direction::bidirectional)
799
        {
Shucai Xiao's avatar
Shucai Xiao committed
800
            if(vec_names.size() == 1)
801
802
803
804
805
            {
                vec_names.push_back(vec_names.at(0));
            }
        }

Shucai Xiao's avatar
Shucai Xiao committed
806
807
808
        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];
809
        });
Shucai Xiao's avatar
Shucai Xiao committed
810

Shucai Xiao's avatar
Shucai Xiao committed
811
812
813
814
815
816
817
        // To be added later
        float clip = 0.0;
        if(contains(attributes, "clip"))
        {
            clip = parse_value(attributes.at("clip")).at<float>();
        }

818
819
        // if the number of arguments is less than 6, append
        // undefined operator to have 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
820
        if(args.size() < 6)
821
822
823
824
825
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), (6 - args.size()), ins);
        }

Shucai Xiao's avatar
Shucai Xiao committed
826
827
        // 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
828
                                                  std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
829

830
        // second output for the last hidden state
Shucai Xiao's avatar
Shucai Xiao committed
831
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
832

Shucai Xiao's avatar
Shucai Xiao committed
833
        return {hidden_states, last_output};
Shucai Xiao's avatar
Shucai Xiao committed
834
835
    }

836
    std::vector<instruction_ref>
837
838
839
840
841
842
843
    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
844
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
845
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
846
847
848
            {
                MIGRAPHX_THROW("GRU: hidden size mismatch in input and attribute");
            }
849
850
851
852
853
854
855
856
857
        }

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

858
        op::rnn_direction dirct = op::rnn_direction::forward;
859
860
        if(direction == "bidirectional")
        {
861
            dirct = op::rnn_direction::bidirectional;
862
863
864
        }
        else if(direction == "reverse")
        {
865
            dirct = op::rnn_direction::reverse;
866
867
        }

868
        std::vector<std::string> vec_names = {"sigmoid", "tanh"};
869
870
        if(contains(attributes, "activations"))
        {
871
            auto names = attributes.at("activations").strings();
872
            vec_names.clear();
Shucai Xiao's avatar
Shucai Xiao committed
873
            vec_names.resize(names.size());
874
            std::copy(names.begin(), names.end(), vec_names.begin());
875
876
        }

877
        // need 4 activation functions
878
        if(dirct == op::rnn_direction::bidirectional)
879
        {
Shucai Xiao's avatar
Shucai Xiao committed
880
            // 4 activation functions are used in the bidirectional
881
            // scenario. No spec is provided in onnx::operator. we
Shucai Xiao's avatar
Shucai Xiao committed
882
883
            // use the algorithm that: if 1 actv function is provided,
            // repeat 1 four times. If 2 actv functins are provided,
884
885
            // 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
886
887
888
            // assume the 3rd one is repeated once and used by the
            // reverse direction.
            // This may need change later
889
            if(vec_names.size() == 1)
890
            {
891
                vec_names.insert(vec_names.end(), 3, vec_names.at(0));
892
            }
893
            else if(vec_names.size() == 2)
894
            {
895
896
897
                // repeat the activation functions
                vec_names.push_back(vec_names.at(0));
                vec_names.push_back(vec_names.at(1));
898
            }
899
            else if(vec_names.size() == 3)
900
            {
901
                vec_names.push_back(vec_names.at(2));
902
903
            }
        }
Shucai Xiao's avatar
Shucai Xiao committed
904
        else
905
        {
906
            if(vec_names.size() == 1)
907
            {
908
                vec_names.push_back(vec_names.at(0));
909
910
911
            }
        }

912
913
914
        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
            return (map_actv_funcs.count(name) == 0);
        });
Shucai Xiao's avatar
Shucai Xiao committed
915
        if(name_it != vec_names.end())
916
917
918
        {
            MIGRAPHX_THROW("GRU: activation function " + std::string(*name_it) + " not supported");
        }
919

Shucai Xiao's avatar
Shucai Xiao committed
920
921
922
        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
923
        });
924
925
926
927
928
929
930
931

        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
932
        if(contains(attributes, "linear_before_reset"))
933
934
935
936
        {
            linear_before_reset = parse_value(attributes.at("linear_before_reset")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
937
        // append undefined opeator to make 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
938
        if(args.size() < 6)
Shucai Xiao's avatar
Shucai Xiao committed
939
940
941
942
943
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), 6 - args.size(), ins);
        }

944
945
        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
946
            op::gru{hidden_size, vec_actv_funcs, dirct, clip, linear_before_reset},
Shucai Xiao's avatar
Shucai Xiao committed
947
            std::move(args));
948
949

        // second output for last gru output
950
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
951

Shucai Xiao's avatar
Shucai Xiao committed
952
        return {hidden_states, last_output};
953
954
    }

Shucai Xiao's avatar
Shucai Xiao committed
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
    std::vector<instruction_ref>
    parse_lstm(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"))
        {
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
            if(hidden_size != hidden_size_att)
            {
                MIGRAPHX_THROW("LSTM: hidden size mismatch in input and attribute");
            }
        }

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

Shucai Xiao's avatar
Shucai Xiao committed
977
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
978
979
        if(direction == "bidirectional")
        {
Shucai Xiao's avatar
Shucai Xiao committed
980
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
981
982
983
        }
        else if(direction == "reverse")
        {
Shucai Xiao's avatar
Shucai Xiao committed
984
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
985
        }
Shucai Xiao's avatar
Shucai Xiao committed
986
        else if(direction == "forward")
Shucai Xiao's avatar
Shucai Xiao committed
987
        {
Shucai Xiao's avatar
Shucai Xiao committed
988
            dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
989
990
991
992
993
994
995
996
997
998
999
1000
        }
        else
        {
            MIGRAPHX_THROW("LSTM: incorrect direction attribute");
        }

        std::vector<std::string> vec_names = {"sigmoid", "tanh", "tanh"};
        if(contains(attributes, "activations"))
        {
            auto names = attributes.at("activations").strings();
            vec_names.clear();
            vec_names.resize(names.size());
1001
            std::copy(names.begin(), names.end(), vec_names.begin());
Shucai Xiao's avatar
Shucai Xiao committed
1002
1003
1004
        }

        // need 6 activation functions for bidirectional directions
Shucai Xiao's avatar
Shucai Xiao committed
1005
        if(dirct == op::rnn_direction::bidirectional)
Shucai Xiao's avatar
Shucai Xiao committed
1006
1007
1008
1009
1010
1011
        {
            // 6 activation functions are used in the bidirectional
            // scenario. No spec is provided in onnx::operator. we
            // use the algorithm that: if 1 actv function is provided,
            // repeat 1st six times. If 2 actv functins are provided,
            // repeat 2nd once, then repeat all three once
Shucai Xiao's avatar
Shucai Xiao committed
1012
            // if 3 actv funcs are provide, repeat all three once.
Shucai Xiao's avatar
Shucai Xiao committed
1013
1014
1015
1016
            // the same algorithm is used for 4, 5, and 6 actv funcions
            // provided. This may need change later
            switch(vec_names.size())
            {
1017
            case 1:
Shucai Xiao's avatar
Shucai Xiao committed
1018
1019
1020
1021
1022
1023
                vec_names = {vec_names.at(0),
                             vec_names.at(0),
                             vec_names.at(0),
                             vec_names.at(0),
                             vec_names.at(0),
                             vec_names.at(0)};
1024
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1025
1026
1027

            case 2:
                // repeat the 2nd actv func once, then repeat all three another time
Shucai Xiao's avatar
Shucai Xiao committed
1028
1029
1030
1031
1032
1033
                vec_names = {vec_names.at(0),
                             vec_names.at(1),
                             vec_names.at(1),
                             vec_names.at(0),
                             vec_names.at(1),
                             vec_names.at(1)};
Shucai Xiao's avatar
Shucai Xiao committed
1034
1035
1036
1037
                break;

            case 3:
                // repeat all three actv funcs once
Shucai Xiao's avatar
Shucai Xiao committed
1038
1039
1040
1041
1042
1043
                vec_names = {vec_names.at(0),
                             vec_names.at(1),
                             vec_names.at(2),
                             vec_names.at(0),
                             vec_names.at(1),
                             vec_names.at(2)};
Shucai Xiao's avatar
Shucai Xiao committed
1044
1045
                break;

Shucai Xiao's avatar
Shucai Xiao committed
1046
1047
1048
1049
1050
1051
1052
            case 4:
                vec_names = {vec_names.at(0),
                             vec_names.at(1),
                             vec_names.at(2),
                             vec_names.at(3),
                             vec_names.at(3),
                             vec_names.at(3)};
1053
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1054

Shucai Xiao's avatar
Shucai Xiao committed
1055
1056
1057
1058
1059
1060
1061
            case 5:
                vec_names = {vec_names.at(0),
                             vec_names.at(1),
                             vec_names.at(2),
                             vec_names.at(3),
                             vec_names.at(4),
                             vec_names.at(4)};
1062
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1063

Shucai Xiao's avatar
Shucai Xiao committed
1064
            default: break;
Shucai Xiao's avatar
Shucai Xiao committed
1065
1066
1067
1068
1069
1070
            }
        }
        else
        {
            switch(vec_names.size())
            {
Shucai Xiao's avatar
Shucai Xiao committed
1071
            case 1: vec_names = {vec_names.at(0), vec_names.at(0), vec_names.at(0)}; break;
Shucai Xiao's avatar
Shucai Xiao committed
1072
1073
1074

            case 2:
                // repeat the 2nd actv func once, so we have 3 actv funcs
Shucai Xiao's avatar
Shucai Xiao committed
1075
                vec_names = {vec_names.at(0), vec_names.at(1), vec_names.at(1)};
Shucai Xiao's avatar
Shucai Xiao committed
1076
1077
                break;

Shucai Xiao's avatar
Shucai Xiao committed
1078
            default: break;
Shucai Xiao's avatar
Shucai Xiao committed
1079
1080
1081
            }
        }

1082
1083
1084
        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
            return (map_actv_funcs.count(name) == 0);
        });
Shucai Xiao's avatar
Shucai Xiao committed
1085
        if(name_it != vec_names.end())
1086
1087
1088
        {
            MIGRAPHX_THROW("LSTM: activation function " + std::string(*name_it) + " not supported");
        }
Shucai Xiao's avatar
Shucai Xiao committed
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110

        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];
        });

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

        int input_forget = 0;
        if(contains(attributes, "input_forget"))
        {
            input_forget = parse_value(attributes.at("input_forget")).at<int>();
        }

        // append undefined opeator to make 6 arguments
        if(args.size() < 8)
        {
            auto ins = prog.add_instruction(op::undefined{});
Shucai Xiao's avatar
Shucai Xiao committed
1111
            args.insert(args.end(), 8 - args.size(), ins);
Shucai Xiao's avatar
Shucai Xiao committed
1112
1113
1114
1115
        }

        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
1116
            op::lstm{hidden_size, vec_actv_funcs, dirct, clip, input_forget}, std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
1117
1118

        // second output for last lstm output
Shucai Xiao's avatar
Shucai Xiao committed
1119
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
1120
1121
1122
1123
1124
1125
1126

        // third output for last cell output
        auto last_cell_output = prog.add_instruction(op::lstm_last_cell_output{}, hidden_states);

        return {hidden_states, last_output, last_cell_output};
    }

Paul's avatar
Paul committed
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
    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
1139
            MIGRAPHX_THROW("Failed reading onnx file.");
Paul's avatar
Paul committed
1140
1141
1142
1143
1144
1145
        }
    }

    void parse_graph(const onnx::GraphProto& graph)
    {
        nodes = get_nodes(graph);
1146
1147
1148
1149
1150
        std::unordered_map<std::string, onnx::TensorProto> initializer_data;
        for(auto&& f : graph.initializer())
        {
            initializer_data[f.name()] = f;
        }
Paul's avatar
Paul committed
1151
1152
1153
        for(auto&& input : graph.input())
        {
            const std::string& name = input.name();
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
            // 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
1166
        }
Paul's avatar
Paul committed
1167
        for(auto&& output : graph.output())
Paul's avatar
Paul committed
1168
        {
Paul's avatar
Paul committed
1169
            this->parse_node(output.name());
Paul's avatar
Paul committed
1170
1171
1172
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
1173
    void parse_undefined(const std::string& name)
1174
    {
Shucai Xiao's avatar
Shucai Xiao committed
1175
        auto ins           = prog.add_instruction(op::undefined{});
1176
1177
1178
        instructions[name] = ins;
    }

Paul's avatar
Paul committed
1179
    void parse_node(const std::string& name)
Paul's avatar
Paul committed
1180
    {
Paul's avatar
Paul committed
1181
        if(name.empty())
Paul's avatar
Paul committed
1182
            MIGRAPHX_THROW("Onnx node must have a name");
Paul's avatar
Paul committed
1183
1184
1185
1186
1187
1188
1189
1190
        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
1191
1192
                    assert(name != input);
                    this->parse_node(input);
Paul's avatar
Paul committed
1193
                }
Shucai Xiao's avatar
Shucai Xiao committed
1194
                else if(input.empty())
Paul's avatar
Paul committed
1195
                {
1196
                    this->parse_undefined(input);
Paul's avatar
Paul committed
1197
                }
1198
                args.push_back(instructions.at(input));
Paul's avatar
Paul committed
1199
            }
Paul's avatar
Paul committed
1200
            std::vector<instruction_ref> result;
Paul's avatar
Paul committed
1201
1202
            if(ops.count(node.op_type()) == 0)
            {
Paul's avatar
Paul committed
1203
                result.push_back(prog.add_instruction(unknown{node.op_type()}, args));
Paul's avatar
Paul committed
1204
1205
1206
            }
            else
            {
Paul's avatar
Paul committed
1207
                result = ops[node.op_type()](get_attributes(node), args);
Paul's avatar
Paul committed
1208
            }
Paul's avatar
Paul committed
1209
            // Even no output nodes produce output in migraphx
Paul's avatar
Paul committed
1210
            if(node.output().empty() and result.size() == 1)
Paul's avatar
Paul committed
1211
1212
            {
                instructions[name] = result.front();
Paul's avatar
Paul committed
1213
1214
1215
            }
            else
            {
Paul's avatar
Paul committed
1216
1217
1218
1219
1220
1221
                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
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
            }
        }
    }

    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
1239
        std::size_t n = 0;
Paul's avatar
Paul committed
1240
1241
        for(auto&& node : graph.node())
        {
Paul's avatar
Paul committed
1242
            if(node.output().empty())
Paul's avatar
Paul committed
1243
            {
Paul's avatar
Paul committed
1244
                if(node.name().empty())
Paul's avatar
Paul committed
1245
1246
1247
1248
1249
1250
1251
1252
1253
                {
                    result["migraphx_unamed_node_" + std::to_string(n)] = node;
                    n++;
                }
                else
                {
                    result[node.name()] = node;
                }
            }
Paul's avatar
Paul committed
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
            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
1279
        case onnx::AttributeProto::FLOATS: return from_repeated(shape::float_type, attr.floats());
Paul's avatar
Paul committed
1280
1281
1282
1283
1284
        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
1285
        MIGRAPHX_THROW("Invalid attribute type");
Paul's avatar
Paul committed
1286
1287
1288
1289
1290
    }

    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
1291
        // in case of scalar constants in onnx file, use dims=1 to fill initializer data
1292
        if(dims.empty())
Khalique's avatar
Khalique committed
1293
1294
1295
        {
            dims = {1};
        }
1296
1297
        if(t.has_raw_data())
        {
wsttiger's avatar
wsttiger committed
1298
            const std::string& s = t.raw_data();
Scott Thornton's avatar
Scott Thornton committed
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
            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
1311
            case onnx::TensorProto::FLOAT16: return literal{{shape::half_type, dims}, s.data()};
Scott Thornton's avatar
Scott Thornton committed
1312
1313
1314
1315
1316
1317
            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
1318
            MIGRAPHX_THROW("Invalid tensor type");
1319
        }
Paul's avatar
Paul committed
1320
1321
1322
1323
        switch(t.data_type())
        {
        case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
        case onnx::TensorProto::FLOAT:
Paul's avatar
Paul committed
1324
            return literal{{shape::float_type, dims}, t.float_data().begin(), t.float_data().end()};
Paul's avatar
Paul committed
1325
1326
        case onnx::TensorProto::UINT8: throw std::runtime_error("");
        case onnx::TensorProto::INT8:
Paul's avatar
Paul committed
1327
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1328
        case onnx::TensorProto::UINT16:
Paul's avatar
Paul committed
1329
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1330
        case onnx::TensorProto::INT16:
Paul's avatar
Paul committed
1331
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1332
        case onnx::TensorProto::INT32:
Paul's avatar
Paul committed
1333
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1334
        case onnx::TensorProto::INT64:
Paul's avatar
Paul committed
1335
            return literal{{shape::int64_type, dims}, t.int64_data().begin(), t.int64_data().end()};
Paul's avatar
Paul committed
1336
1337
        case onnx::TensorProto::STRING: throw std::runtime_error("");
        case onnx::TensorProto::BOOL:
Paul's avatar
Paul committed
1338
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1339
        case onnx::TensorProto::FLOAT16:
Khalique's avatar
Khalique committed
1340
        {
Khalique's avatar
Khalique committed
1341
            std::vector<uint16_t> data_uint16(t.int32_data().begin(), t.int32_data().end());
1342
            std::vector<half> data_half;
Khalique's avatar
Khalique committed
1343
1344
1345
            std::transform(data_uint16.begin(),
                           data_uint16.end(),
                           std::back_inserter(data_half),
1346
                           [](uint16_t raw_val) { return *reinterpret_cast<half*>(&raw_val); });
1347
            return literal{{shape::half_type, dims}, data_half.begin(), data_half.end()};
Khalique's avatar
Khalique committed
1348
        }
Paul's avatar
Paul committed
1349
1350
1351
1352
1353
1354
1355
1356
        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
1357
        MIGRAPHX_THROW("Invalid tensor type");
Paul's avatar
Paul committed
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
    }

    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
1379
        case onnx::TensorProto::FLOAT16: shape_type = shape::half_type; break;
Paul's avatar
Paul committed
1380
1381
1382
1383
1384
1385
1386
1387
1388
        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
1389
        auto&& tensor_dims = t.tensor_type().shape().dim();
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
        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
1401
1402
        return {shape_type, dims};
    }
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424

    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
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
};

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
1448
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1449
} // namespace migraphx