onnx.cpp 52.5 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);
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);
Shucai Xiao's avatar
Shucai Xiao committed
92
        add_mem_op("LSTM", &onnx_parser::parse_lstm);
Khalique's avatar
Khalique committed
93
        add_mem_op("Pad", &onnx_parser::parse_pad);
94
95
96
97
98
99
100

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

    void init_actv_func()
    {
101
102
103
104
105
        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
106
107
108
109
    }

    template <class F>
    void add_op(std::string name, F f)
Paul's avatar
Paul committed
110
111
112
113
114
115
116
117
118
    {
        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
119
120
121
122
123
124
125
    {
        ops.emplace(name, f);
    }

    template <class F>
    void add_mem_op(std::string name, F f)
    {
Paul's avatar
Paul committed
126
        add_op(name, [=](auto&&... xs) {
Paul's avatar
Paul committed
127
128
129
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }
Khalique's avatar
Khalique committed
130

131
    template <class T>
Khalique's avatar
Khalique committed
132
    void add_binary_op(std::string name, T x)
133
    {
Paul's avatar
Paul committed
134
        add_op(name, [this, x](attribute_map attributes, std::vector<instruction_ref> args) {
Scott Thornton's avatar
Scott Thornton committed
135
            if(args.size() != 2)
Paul's avatar
Paul committed
136
                MIGRAPHX_THROW("binary operators should have 2 operands");
137
            if(contains(attributes, "broadcast") and contains(attributes, "axis"))
138
139
140
141
            {
                uint64_t broadcasted = parse_value(attributes.at("broadcast")).at<uint64_t>();
                if(broadcasted != 0)
                {
142
                    uint64_t axis = parse_value(attributes.at("axis")).at<uint64_t>();
143
144
145
146
                    auto l =
                        prog.add_instruction(op::broadcast{axis, args[0]->get_shape()}, args[1]);
                    return prog.add_instruction(x, args[0], l);
                }
147
                return prog.add_instruction(x, args);
148
            }
Paul's avatar
Paul committed
149
            else
150
            {
Khalique's avatar
Khalique committed
151
                return add_broadcastable_binary_op(args[0], args[1], x);
152
153
154
155
            }
        });
    }

Khalique's avatar
Khalique committed
156
157
158
159
160
    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
161
162
163
164
165
166
167
168
169
170
171
172
173
            // 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
174
175
176
177
178
179
180
181
            // 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
182
            std::vector<std::size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
183
184
            auto offset = s1->size() - s0->size();
            std::transform(s0->begin(),
Khalique's avatar
Khalique committed
185
186
187
188
                           s0->end(),
                           s1->begin() + offset,
                           output_lens.begin() + offset,
                           [](auto a, auto b) { return std::max(a, b); });
Khalique's avatar
Khalique committed
189
190
191
192
193
194
195
196
197

            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});
        }
198
199
    }

Paul's avatar
Paul committed
200
    template <class T>
Paul's avatar
Paul committed
201
202
    void add_generic_op(std::string name, T x)
    {
Paul's avatar
Paul committed
203
        add_op(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
Paul's avatar
Paul committed
204
205
206
207
            return prog.add_instruction(x, args);
        });
    }

Khalique's avatar
Khalique committed
208
    template <class T>
Khalique's avatar
Khalique committed
209
    void add_variadic_op(std::string name, T x)
Khalique's avatar
Khalique committed
210
    {
Paul's avatar
Paul committed
211
        add_op(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
212
            return std::accumulate(std::next(args.begin()),
Khalique's avatar
Khalique committed
213
214
215
216
217
                                   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
218
        });
Khalique's avatar
Khalique committed
219
220
    }

Paul's avatar
Paul committed
221
    instruction_ref
Paul's avatar
Paul committed
222
    parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
223
224
    {
        auto dims = args.front()->get_shape().lens();
Scott Thornton's avatar
Scott Thornton committed
225
226
        auto r =
            prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, args.front());
227
228
        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
229
230
    }

Paul's avatar
Paul committed
231
    instruction_ref
Paul's avatar
Paul committed
232
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
233
    {
234
        op::convolution op;
235
        auto l0 = args[0];
Paul's avatar
Paul committed
236
237
        if(contains(attributes, "pads"))
        {
Scott Thornton's avatar
Scott Thornton committed
238
            if(contains(attributes, "auto_pad"))
239
            {
Paul's avatar
Paul committed
240
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
241
            }
242
243
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
244
            if(padding.size() != 4)
245
            {
Paul's avatar
Paul committed
246
                MIGRAPHX_THROW("padding should have 4 values");
247
            }
Scott Thornton's avatar
Scott Thornton committed
248
            if(padding[0] != padding[2] || padding[1] != padding[3])
249
            {
250
251
                // 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
252
                l0      = prog.add_instruction(op::pad{padding}, l0);
253
            }
254
255
256
257
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
258
            }
Paul's avatar
Paul committed
259
        }
Paul's avatar
Paul committed
260
261
262
263
264
265
266
267
        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
268
        if(contains(attributes, "auto_pad"))
269
270
        {
            auto s = attributes["auto_pad"].s();
Scott Thornton's avatar
Scott Thornton committed
271
            if(contains(attributes, "pads") and to_upper(s) != "NOTSET")
272
            {
Paul's avatar
Paul committed
273
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
274
275
            }

wsttiger's avatar
fixes  
wsttiger committed
276
            if(s.find("SAME") != std::string::npos)
277
            {
278
                op.padding_mode = op::padding_mode_t::same;
279
280
            }
        }
Khalique's avatar
Khalique committed
281
282
283
284
        if(contains(attributes, "group"))
        {
            op.group = parse_value(attributes.at("group")).at<int>();
        }
Paul's avatar
Paul committed
285
286
287
288
        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
289
            auto l2       = prog.add_instruction(op::broadcast{axis, l1->get_shape()}, args[2]);
290
            return prog.add_instruction(op::add{}, l1, l2);
Paul's avatar
Paul committed
291
        }
292
        return prog.add_instruction(op, l0, args[1]);
Paul's avatar
Paul committed
293
    }
Paul's avatar
Paul committed
294

Paul's avatar
Paul committed
295
296
297
    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
Paul's avatar
Paul committed
298
    {
Khalique's avatar
Khalique committed
299
        op::pooling op{ends_with(name, "MaxPool") ? "max" : "average"};
300
        auto l0 = args[0];
Khalique's avatar
Khalique committed
301
        if(starts_with(name, "Global"))
302
        {
Khalique's avatar
Khalique committed
303
304
            auto lens  = args.front()->get_shape().lens();
            op.lengths = {lens[2], lens[3]};
305
        }
Paul's avatar
Paul committed
306
307
        if(contains(attributes, "pads"))
        {
308
309
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
310
            if(padding.size() != 4)
311
            {
Paul's avatar
Paul committed
312
                MIGRAPHX_THROW("padding should have 4 values");
313
            }
Scott Thornton's avatar
Scott Thornton committed
314
            if(padding[0] != padding[2] || padding[1] != padding[3])
315
            {
316
317
                // 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
318
                l0      = prog.add_instruction(op::pad{padding}, l0);
319
320
321
322
323
            }
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
324
            }
Paul's avatar
Paul committed
325
326
327
328
329
330
331
332
333
        }
        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
334
        if(contains(attributes, "auto_pad"))
335
336
        {
            auto s = attributes["auto_pad"].s();
337
            if(s.find("SAME_UPPER") == std::string::npos)
338
            {
339
                MIGRAPHX_THROW("auto_pad only supports SAME_UPPER for pooling");
340
            }
341
            op.padding_mode = op::padding_mode_t::same;
342
343
        }

344
        return prog.add_instruction(op, l0);
Paul's avatar
Paul committed
345
346
    }

Paul's avatar
Paul committed
347
    instruction_ref
Paul's avatar
Paul committed
348
    parse_reshape(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
349
    {
350
        op::reshape op;
Paul's avatar
Paul committed
351
352
353
354
355
356
357
        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
358
            literal s = args[1]->get_literal();
Paul's avatar
Paul committed
359
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
Paul's avatar
Paul committed
360
        }
Paul's avatar
Paul committed
361
362
363
        return prog.add_instruction(op, args[0]);
    }

Paul's avatar
Paul committed
364
    instruction_ref
Paul's avatar
Paul committed
365
    parse_flatten(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
366
    {
367
        uint64_t axis = 1;
Paul's avatar
Paul committed
368
369
370
371
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
372
        return prog.add_instruction(op::flatten{axis}, args[0]);
Paul's avatar
Paul committed
373
374
    }

375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
    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
393
394
395
396
397
398
399
    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));
    }
400

401
402
403
    instruction_ref
    parse_gather(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
404
        int axis = 0;
405
406
407
408
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
409
        op::gather op{axis};
410
411
412
        return prog.add_instruction(op, std::move(args));
    }

413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
    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
433
434
435
    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
Paul's avatar
Paul committed
436
437
438
439
    {
        literal v = parse_value(attributes.at("value"));
        return prog.add_literal(v);
    }
Paul's avatar
Paul committed
440

Paul's avatar
Paul committed
441
    instruction_ref
Paul's avatar
Paul committed
442
    parse_gemm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
443
444
    {
        float alpha = 1.0f;
Khalique's avatar
Khalique committed
445
        float beta  = 1.0f;
Paul's avatar
Paul committed
446
447
448
449
450
451
452
453
        bool transa = false;
        bool transb = false;
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        if(contains(attributes, "beta"))
        {
454
            beta = parse_value(attributes.at("beta")).at<float>();
Paul's avatar
Paul committed
455
456
457
458
459
460
461
462
463
464
        }
        if(contains(attributes, "transA"))
        {
            transa = parse_value(attributes.at("transA")).at<bool>();
        }
        if(contains(attributes, "transB"))
        {
            transb = parse_value(attributes.at("transB")).at<bool>();
        }
        std::vector<int64_t> perm = {1, 0};
465
466
        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
467
468
        if(args.size() == 3)
        {
Khalique's avatar
Khalique committed
469
            if(beta != 0.f)
470
            {
Khalique's avatar
Khalique committed
471
                auto l3 = prog.add_instruction(op::dot{alpha}, l1, l2);
Khalique's avatar
Khalique committed
472
                auto l4 = args[2];
Khalique's avatar
Khalique committed
473
                if(l4->get_shape().scalar()) // ignore args[2] (no C value added to alpha*A*B)
Khalique's avatar
Khalique committed
474
                    return l3;
Khalique's avatar
Khalique committed
475
                if(beta != 1.f)
Khalique's avatar
Khalique committed
476
477
                {
                    auto beta_val = prog.add_literal(beta);
Khalique's avatar
Khalique committed
478
479
                    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
480
481
                }
                return add_broadcastable_binary_op(l3, l4, op::add{});
482
            }
Paul's avatar
Paul committed
483
        }
Shucai Xiao's avatar
Shucai Xiao committed
484
        return prog.add_instruction(op::dot{alpha, beta}, l1, l2);
Paul's avatar
Paul committed
485
486
    }

487
    instruction_ref
Paul's avatar
Paul committed
488
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
489
    {
Scott Thornton's avatar
Scott Thornton committed
490
491
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
492
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
Scott Thornton's avatar
Scott Thornton committed
493
        bool is_test                                      = false;
494
495
496
497
498
499
        if(contains(attributes, "epsilon"))
        {
            epsilon = parse_value(attributes.at("epsilon")).at<float>();
        }
        if(contains(attributes, "momentum"))
        {
500
            momentum = parse_value(attributes.at("momentum")).at<float>();
501
502
503
        }
        if(contains(attributes, "is_test"))
        {
wsttiger's avatar
wsttiger committed
504
            is_test = parse_value(attributes.at("is_test")).at<uint64_t>() > 0;
505
506
507
        }
        if(contains(attributes, "spatial"))
        {
508
            bn_mode = (parse_value(attributes.at("spatial")).at<uint64_t>() > 0)
509
510
                          ? op::batch_norm_inference::spatial
                          : op::batch_norm_inference::per_activation;
511
        }
Paul's avatar
Paul committed
512
        (void)is_test;
Paul's avatar
Paul committed
513
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
Paul's avatar
Paul committed
514
        return prog.add_instruction(op, std::move(args));
515
516
    }

517
518
519
520
    instruction_ref parse_leaky_relu(const std::string&,
                                     attribute_map attributes,
                                     std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
521
        float alpha = 0.01; // default alpha val for leaky relu
522
523
524
525
526
527
528
529
        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
530
531
    instruction_ref
    parse_elu(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
532
533
534
535
536
537
538
539
540
541
    {
        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
542
543
    instruction_ref
    parse_lrn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
544
545
    {
        float alpha = 0.0001;
Khalique's avatar
Khalique committed
546
547
548
        float beta  = 0.75;
        float bias  = 1.0;
        int size    = 1;
Khalique's avatar
Khalique committed
549
550
551
552
553
554
555
556
557
558
559
560
        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
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
    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
578

Khalique's avatar
Khalique committed
579
580
        auto scale_val = prog.add_literal(scale);
        auto bias_vals = prog.add_literal(
Paul's avatar
Paul committed
581
            migraphx::literal{migraphx::shape{migraphx::shape::float_type, {bias.size()}}, bias});
Khalique's avatar
Khalique committed
582

Paul's avatar
Paul committed
583
584
        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
585
        auto bias_bcast = prog.add_instruction(migraphx::op::broadcast{1, input_shape}, bias_vals);
Paul's avatar
Paul committed
586
        return prog.add_instruction(migraphx::op::add{}, img_scaled, bias_bcast);
Khalique's avatar
Khalique committed
587
    }
Khalique's avatar
Khalique committed
588

Khalique's avatar
Khalique committed
589
590
    instruction_ref
    parse_transpose(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
591
592
593
594
595
596
597
    {
        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
598
        return prog.add_instruction(migraphx::op::transpose{perm}, args.front());
Khalique's avatar
Khalique committed
599
600
    }

Khalique's avatar
Khalique committed
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
    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());
    }
623
624
625
    // 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
626
    parse_shape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
627
628
    {
        if(args.size() != 1)
629
            MIGRAPHX_THROW("Shape: operator should have 1 operand");
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
        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
666
667
        if(contains(attributes, "extra_shape"))
        {
668
            MIGRAPHX_THROW("ConstantFill: cannot handle extra shape attribute");
669
670
        }

671
672
        if(input_as_shape == 1)
        {
Shucai Xiao's avatar
Shucai Xiao committed
673
            if(args.size() != 1)
674
            {
675
                MIGRAPHX_THROW("ConstantFill: need an input argument as output shape");
676
677
            }

Shucai Xiao's avatar
Shucai Xiao committed
678
679
            if(contains(attributes, "shape"))
            {
680
                MIGRAPHX_THROW("ConstantFill: cannot set the shape argument and pass in an input "
Shucai Xiao's avatar
Shucai Xiao committed
681
                               "at the same time");
682
683
            }

684
685
686
            migraphx::argument in = args[0]->eval();
            if(in.empty())
            {
687
                MIGRAPHX_THROW("ConstantFill: cannot handle dynamic shape as input");
688
            }
689

690
691
692
            std::vector<std::size_t> dims;
            in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
            migraphx::shape s(type, dims);
693
694
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
695
696
697
        }
        else if(input_as_shape == 0)
        {
Shucai Xiao's avatar
Shucai Xiao committed
698
699
            if(!contains(attributes, "shape"))
            {
700
                MIGRAPHX_THROW("ConstantFill: attribute output shape is needed");
701
702
703
            }

            literal ls = parse_value(attributes.at("shape"));
704
            std::vector<std::size_t> dims;
Shucai Xiao's avatar
Shucai Xiao committed
705
            ls.visit([&](auto s) { dims.assign(s.begin(), s.end()); });
706
            migraphx::shape s{type, dims};
707
708
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
709
710
711
        }
        else
        {
712
            MIGRAPHX_THROW("ConstantFill: wrong value of attribute input_as_shape");
713
714
715
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
716
    std::vector<instruction_ref>
Shucai Xiao's avatar
Shucai Xiao committed
717
718
719
    parse_rnn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        migraphx::shape input_shape = args[0]->get_shape();
720
        std::size_t hidden_size     = args[1]->get_shape().lens()[1];
Shucai Xiao's avatar
Shucai Xiao committed
721
722
723

        if(contains(attributes, "hidden_size"))
        {
Shucai Xiao's avatar
Shucai Xiao committed
724
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
725
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
726
727
728
            {
                MIGRAPHX_THROW("RNN: hidden size mismatch in input and attribute");
            }
Shucai Xiao's avatar
Shucai Xiao committed
729
730
731
732
733
734
735
736
737
        }

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

738
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
739
740
        if(direction == "bidirectional")
        {
741
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
742
743
744
        }
        else if(direction == "reverse")
        {
745
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
746
747
        }

748
749
750
751
752
        std::vector<std::string> vec_names{"tanh"};
        if(contains(attributes, "activations"))
        {
            auto names = attributes.at("activations").strings();
            vec_names.clear();
753
            vec_names.resize(names.size());
754
            std::copy(names.begin(), names.end(), vec_names.begin());
755
756
        }

757
758
759
760
        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
            return (map_actv_funcs.count(name) == 0);
        });
        if (name_it != vec_names.end())
761
762
763
        {
            MIGRAPHX_THROW("RNN: activation function " + std::string(*name_it) + " not supported");
        }
764

Shucai Xiao's avatar
Shucai Xiao committed
765
        // bidirectional case should have two activation functions.
Shucai Xiao's avatar
Shucai Xiao committed
766
        // one is for forward, and the other is for reverse.
Shucai Xiao's avatar
Shucai Xiao committed
767
        // if only one actv function is provided, we use it in both
768
        // forward and reverse direction
769
        if(dirct == op::rnn_direction::bidirectional)
770
        {
Shucai Xiao's avatar
Shucai Xiao committed
771
            if(vec_names.size() == 1)
772
773
774
775
776
            {
                vec_names.push_back(vec_names.at(0));
            }
        }

Shucai Xiao's avatar
Shucai Xiao committed
777
778
779
        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];
780
        });
Shucai Xiao's avatar
Shucai Xiao committed
781

Shucai Xiao's avatar
Shucai Xiao committed
782
783
784
785
786
787
788
        // To be added later
        float clip = 0.0;
        if(contains(attributes, "clip"))
        {
            clip = parse_value(attributes.at("clip")).at<float>();
        }

789
790
        // if the number of arguments is less than 6, append
        // undefined operator to have 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
791
        if(args.size() < 6)
792
793
794
795
796
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), (6 - args.size()), ins);
        }

Shucai Xiao's avatar
Shucai Xiao committed
797
798
        // 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
799
                                                  std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
800

801
        // second output for the last hidden state
Shucai Xiao's avatar
Shucai Xiao committed
802
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
803

Shucai Xiao's avatar
Shucai Xiao committed
804
        return {hidden_states, last_output};
Shucai Xiao's avatar
Shucai Xiao committed
805
806
    }

807
    std::vector<instruction_ref>
808
809
810
811
812
813
814
    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
815
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
816
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
817
818
819
            {
                MIGRAPHX_THROW("GRU: hidden size mismatch in input and attribute");
            }
820
821
822
823
824
825
826
827
828
        }

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

829
        op::rnn_direction dirct = op::rnn_direction::forward;
830
831
        if(direction == "bidirectional")
        {
832
            dirct = op::rnn_direction::bidirectional;
833
834
835
        }
        else if(direction == "reverse")
        {
836
            dirct = op::rnn_direction::reverse;
837
838
        }

839
        std::vector<std::string> vec_names = {"sigmoid", "tanh"};
840
841
        if(contains(attributes, "activations"))
        {
842
            auto names = attributes.at("activations").strings();
843
            vec_names.clear();
Shucai Xiao's avatar
Shucai Xiao committed
844
            vec_names.resize(names.size());
845
            std::copy(names.begin(), names.end(), vec_names.begin());
846
847
        }

848
        // need 4 activation functions
849
        if(dirct == op::rnn_direction::bidirectional)
850
        {
Shucai Xiao's avatar
Shucai Xiao committed
851
            // 4 activation functions are used in the bidirectional
852
            // scenario. No spec is provided in onnx::operator. we
Shucai Xiao's avatar
Shucai Xiao committed
853
854
            // use the algorithm that: if 1 actv function is provided,
            // repeat 1 four times. If 2 actv functins are provided,
855
856
            // 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
857
858
859
            // assume the 3rd one is repeated once and used by the
            // reverse direction.
            // This may need change later
860
            if(vec_names.size() == 1)
861
            {
862
                vec_names.insert(vec_names.end(), 3, vec_names.at(0));
863
            }
864
            else if(vec_names.size() == 2)
865
            {
866
867
868
                // repeat the activation functions
                vec_names.push_back(vec_names.at(0));
                vec_names.push_back(vec_names.at(1));
869
            }
870
            else if(vec_names.size() == 3)
871
            {
872
                vec_names.push_back(vec_names.at(2));
873
874
            }
        }
Shucai Xiao's avatar
Shucai Xiao committed
875
        else
876
        {
877
            if(vec_names.size() == 1)
878
            {
879
                vec_names.push_back(vec_names.at(0));
880
881
882
            }
        }

883
884
885
886
887
        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
            return (map_actv_funcs.count(name) == 0);
        });
        if (name_it != vec_names.end())
        {   
888
889
            MIGRAPHX_THROW("GRU: activation function " + std::string(*name_it) + " not supported");
        }
890

Shucai Xiao's avatar
Shucai Xiao committed
891
892
893
        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
894
        });
895
896
897
898
899
900
901
902

        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
903
        if(contains(attributes, "linear_before_reset"))
904
905
906
907
        {
            linear_before_reset = parse_value(attributes.at("linear_before_reset")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
908
        // append undefined opeator to make 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
909
        if(args.size() < 6)
Shucai Xiao's avatar
Shucai Xiao committed
910
911
912
913
914
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), 6 - args.size(), ins);
        }

915
916
        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
917
            op::gru{hidden_size, vec_actv_funcs, dirct, clip, linear_before_reset},
Shucai Xiao's avatar
Shucai Xiao committed
918
            std::move(args));
919
920

        // second output for last gru output
921
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
922

Shucai Xiao's avatar
Shucai Xiao committed
923
        return {hidden_states, last_output};
924
925
    }

Shucai Xiao's avatar
Shucai Xiao committed
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
    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
948
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
949
950
        if(direction == "bidirectional")
        {
Shucai Xiao's avatar
Shucai Xiao committed
951
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
952
953
954
        }
        else if(direction == "reverse")
        {
Shucai Xiao's avatar
Shucai Xiao committed
955
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
956
        }
Shucai Xiao's avatar
Shucai Xiao committed
957
        else if(direction == "forward")
Shucai Xiao's avatar
Shucai Xiao committed
958
        {
Shucai Xiao's avatar
Shucai Xiao committed
959
            dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
960
961
962
963
964
965
966
967
968
969
970
971
        }
        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());
972
            std::copy(names.begin(), names.end(), vec_names.begin());
Shucai Xiao's avatar
Shucai Xiao committed
973
974
975
        }

        // need 6 activation functions for bidirectional directions
Shucai Xiao's avatar
Shucai Xiao committed
976
        if(dirct == op::rnn_direction::bidirectional)
Shucai Xiao's avatar
Shucai Xiao committed
977
978
979
980
981
982
        {
            // 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
983
            // if 3 actv funcs are provide, repeat all three once.
Shucai Xiao's avatar
Shucai Xiao committed
984
985
986
987
            // the same algorithm is used for 4, 5, and 6 actv funcions
            // provided. This may need change later
            switch(vec_names.size())
            {
988
            case 1:
Shucai Xiao's avatar
Shucai Xiao committed
989
990
991
992
993
994
                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)};
995
                break;
Shucai Xiao's avatar
Shucai Xiao committed
996
997
998

            case 2:
                // repeat the 2nd actv func once, then repeat all three another time
Shucai Xiao's avatar
Shucai Xiao committed
999
1000
1001
1002
1003
1004
                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
1005
1006
1007
1008
                break;

            case 3:
                // repeat all three actv funcs once
Shucai Xiao's avatar
Shucai Xiao committed
1009
1010
1011
1012
1013
1014
                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
1015
1016
                break;

Shucai Xiao's avatar
Shucai Xiao committed
1017
1018
1019
1020
1021
1022
1023
            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)};
1024
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1025

Shucai Xiao's avatar
Shucai Xiao committed
1026
1027
1028
1029
1030
1031
1032
            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)};
1033
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1034

Shucai Xiao's avatar
Shucai Xiao committed
1035
            default: break;
Shucai Xiao's avatar
Shucai Xiao committed
1036
1037
1038
1039
1040
1041
            }
        }
        else
        {
            switch(vec_names.size())
            {
Shucai Xiao's avatar
Shucai Xiao committed
1042
            case 1: vec_names = {vec_names.at(0), vec_names.at(0), vec_names.at(0)}; break;
Shucai Xiao's avatar
Shucai Xiao committed
1043
1044
1045

            case 2:
                // repeat the 2nd actv func once, so we have 3 actv funcs
Shucai Xiao's avatar
Shucai Xiao committed
1046
                vec_names = {vec_names.at(0), vec_names.at(1), vec_names.at(1)};
Shucai Xiao's avatar
Shucai Xiao committed
1047
1048
                break;

Shucai Xiao's avatar
Shucai Xiao committed
1049
            default: break;
Shucai Xiao's avatar
Shucai Xiao committed
1050
1051
1052
            }
        }

1053
1054
1055
1056
        auto name_it = std::find_if(vec_names.begin(), vec_names.end(), [&](auto& name) {
            return (map_actv_funcs.count(name) == 0);
        });
        if (name_it !=  vec_names.end()) 
1057
1058
1059
        {
            MIGRAPHX_THROW("LSTM: activation function " + std::string(*name_it) + " not supported");
        }
Shucai Xiao's avatar
Shucai Xiao committed
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081

        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
1082
            args.insert(args.end(), 8 - args.size(), ins);
Shucai Xiao's avatar
Shucai Xiao committed
1083
1084
1085
1086
        }

        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
1087
            op::lstm{hidden_size, vec_actv_funcs, dirct, clip, input_forget}, std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
1088
1089

        // second output for last lstm output
Shucai Xiao's avatar
Shucai Xiao committed
1090
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
1091
1092
1093
1094
1095
1096
1097

        // 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
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
    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
1110
            MIGRAPHX_THROW("Failed reading onnx file.");
Paul's avatar
Paul committed
1111
1112
1113
1114
1115
1116
        }
    }

    void parse_graph(const onnx::GraphProto& graph)
    {
        nodes = get_nodes(graph);
1117
1118
1119
1120
1121
        std::unordered_map<std::string, onnx::TensorProto> initializer_data;
        for(auto&& f : graph.initializer())
        {
            initializer_data[f.name()] = f;
        }
Paul's avatar
Paul committed
1122
1123
1124
        for(auto&& input : graph.input())
        {
            const std::string& name = input.name();
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
            // 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
1137
1138
1139
        }
        for(auto&& p : nodes)
        {
Paul's avatar
Paul committed
1140
            this->parse_node(p.first);
Paul's avatar
Paul committed
1141
1142
1143
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
1144
    void parse_undefined(const std::string& name)
1145
    {
Shucai Xiao's avatar
Shucai Xiao committed
1146
        auto ins           = prog.add_instruction(op::undefined{});
1147
1148
1149
        instructions[name] = ins;
    }

Paul's avatar
Paul committed
1150
    void parse_node(const std::string& name)
Paul's avatar
Paul committed
1151
    {
Paul's avatar
Paul committed
1152
        if(name.empty())
Paul's avatar
Paul committed
1153
            MIGRAPHX_THROW("Onnx node must have a name");
Paul's avatar
Paul committed
1154
1155
1156
1157
1158
1159
1160
1161
        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
1162
1163
                    assert(name != input);
                    this->parse_node(input);
Paul's avatar
Paul committed
1164
                }
Shucai Xiao's avatar
Shucai Xiao committed
1165
                else if(input.empty())
Paul's avatar
Paul committed
1166
                {
1167
                    this->parse_undefined(input);
Paul's avatar
Paul committed
1168
                }
1169
                args.push_back(instructions.at(input));
Paul's avatar
Paul committed
1170
            }
Paul's avatar
Paul committed
1171
            std::vector<instruction_ref> result;
Paul's avatar
Paul committed
1172
1173
            if(ops.count(node.op_type()) == 0)
            {
Paul's avatar
Paul committed
1174
                result.push_back(prog.add_instruction(unknown{node.op_type()}, args));
Paul's avatar
Paul committed
1175
1176
1177
            }
            else
            {
Paul's avatar
Paul committed
1178
                result = ops[node.op_type()](get_attributes(node), args);
Paul's avatar
Paul committed
1179
            }
Paul's avatar
Paul committed
1180
            // Even no output nodes produce output in migraphx
Paul's avatar
Paul committed
1181
            if(node.output().empty() and result.size() == 1)
Paul's avatar
Paul committed
1182
1183
            {
                instructions[name] = result.front();
Paul's avatar
Paul committed
1184
1185
1186
            }
            else
            {
Paul's avatar
Paul committed
1187
1188
1189
1190
1191
1192
                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
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
            }
        }
    }

    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
1210
        std::size_t n = 0;
Paul's avatar
Paul committed
1211
1212
        for(auto&& node : graph.node())
        {
Paul's avatar
Paul committed
1213
            if(node.output().empty())
Paul's avatar
Paul committed
1214
            {
Paul's avatar
Paul committed
1215
                if(node.name().empty())
Paul's avatar
Paul committed
1216
1217
1218
1219
1220
1221
1222
1223
1224
                {
                    result["migraphx_unamed_node_" + std::to_string(n)] = node;
                    n++;
                }
                else
                {
                    result[node.name()] = node;
                }
            }
Paul's avatar
Paul committed
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
            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
1250
        case onnx::AttributeProto::FLOATS: return from_repeated(shape::float_type, attr.floats());
Paul's avatar
Paul committed
1251
1252
1253
1254
1255
        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
1256
        MIGRAPHX_THROW("Invalid attribute type");
Paul's avatar
Paul committed
1257
1258
1259
1260
1261
    }

    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
1262
        // in case of scalar constants in onnx file, use dims=1 to fill initializer data
1263
        if(dims.empty())
Khalique's avatar
Khalique committed
1264
1265
1266
        {
            dims = {1};
        }
1267
1268
        if(t.has_raw_data())
        {
wsttiger's avatar
wsttiger committed
1269
            const std::string& s = t.raw_data();
Scott Thornton's avatar
Scott Thornton committed
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
            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
1282
            case onnx::TensorProto::FLOAT16: return literal{{shape::half_type, dims}, s.data()};
Scott Thornton's avatar
Scott Thornton committed
1283
1284
1285
1286
1287
1288
            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
1289
            MIGRAPHX_THROW("Invalid tensor type");
1290
        }
Paul's avatar
Paul committed
1291
1292
1293
1294
        switch(t.data_type())
        {
        case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
        case onnx::TensorProto::FLOAT:
Paul's avatar
Paul committed
1295
            return literal{{shape::float_type, dims}, t.float_data().begin(), t.float_data().end()};
Paul's avatar
Paul committed
1296
1297
        case onnx::TensorProto::UINT8: throw std::runtime_error("");
        case onnx::TensorProto::INT8:
Paul's avatar
Paul committed
1298
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1299
        case onnx::TensorProto::UINT16:
Paul's avatar
Paul committed
1300
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1301
        case onnx::TensorProto::INT16:
Paul's avatar
Paul committed
1302
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1303
        case onnx::TensorProto::INT32:
Paul's avatar
Paul committed
1304
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1305
        case onnx::TensorProto::INT64:
Paul's avatar
Paul committed
1306
            return literal{{shape::int64_type, dims}, t.int64_data().begin(), t.int64_data().end()};
Paul's avatar
Paul committed
1307
1308
        case onnx::TensorProto::STRING: throw std::runtime_error("");
        case onnx::TensorProto::BOOL:
Paul's avatar
Paul committed
1309
            return literal{{shape::int32_type, dims}, t.int32_data().begin(), t.int32_data().end()};
Paul's avatar
Paul committed
1310
        case onnx::TensorProto::FLOAT16:
Khalique's avatar
Khalique committed
1311
        {
Khalique's avatar
Khalique committed
1312
            std::vector<uint16_t> data_uint16(t.int32_data().begin(), t.int32_data().end());
1313
            std::vector<half> data_half;
Khalique's avatar
Khalique committed
1314
1315
1316
            std::transform(data_uint16.begin(),
                           data_uint16.end(),
                           std::back_inserter(data_half),
1317
                           [](uint16_t raw_val) { return *reinterpret_cast<half*>(&raw_val); });
1318
            return literal{{shape::half_type, dims}, data_half.begin(), data_half.end()};
Khalique's avatar
Khalique committed
1319
        }
Paul's avatar
Paul committed
1320
1321
1322
1323
1324
1325
1326
1327
        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
1328
        MIGRAPHX_THROW("Invalid tensor type");
Paul's avatar
Paul committed
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
    }

    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
1350
        case onnx::TensorProto::FLOAT16: shape_type = shape::half_type; break;
Paul's avatar
Paul committed
1351
1352
1353
1354
1355
1356
1357
1358
1359
        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
1360
        auto&& tensor_dims = t.tensor_type().shape().dim();
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
        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
1372
1373
        return {shape_type, dims};
    }
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395

    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
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
};

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
1419
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1420
} // namespace migraphx