onnx.cpp 64.4 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()
    {
Khalique's avatar
Khalique committed
39
        add_generic_op("Relu", op::relu{});
Khalique's avatar
Khalique committed
40
41
        add_generic_op("Sigmoid", op::sigmoid{});
        add_generic_op("Abs", op::abs{});
Shucai Xiao's avatar
Shucai Xiao committed
42
        add_generic_op("Exp", op::exp{});
Shucai Xiao's avatar
Shucai Xiao committed
43
        add_generic_op("Erf", op::erf{});
Shucai Xiao's avatar
Shucai Xiao committed
44
        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
        add_binary_op("Add", op::add{});
        add_binary_op("Div", op::div{});
        add_binary_op("Mul", op::mul{});
        add_binary_op("Sub", op::sub{});
Shucai Xiao's avatar
Shucai Xiao committed
62
        add_binary_op("Pow", op::pow{});
Khalique's avatar
Khalique committed
63

Khalique's avatar
Khalique committed
64
65
66
        add_variadic_op("Sum", op::add{});
        add_variadic_op("Max", op::max{});
        add_variadic_op("Min", op::min{});
Paul's avatar
Paul committed
67

68
69
        add_mem_op("ArgMax", &onnx_parser::parse_argmax);
        add_mem_op("ArgMin", &onnx_parser::parse_argmin);
70
        add_mem_op("Cast", &onnx_parser::parse_cast);
Khalique's avatar
Khalique committed
71
        add_mem_op("Clip", &onnx_parser::parse_clip);
Khalique's avatar
Khalique committed
72
        add_mem_op("LRN", &onnx_parser::parse_lrn);
Khalique's avatar
Khalique committed
73
        add_mem_op("ImageScaler", &onnx_parser::parse_imagescaler);
74
        add_mem_op("LeakyRelu", &onnx_parser::parse_leaky_relu);
Khalique's avatar
Khalique committed
75
        add_mem_op("Elu", &onnx_parser::parse_elu);
76
        add_mem_op("Expand", &onnx_parser::parse_expand);
Paul's avatar
Paul committed
77
78
        add_mem_op("Constant", &onnx_parser::parse_constant);
        add_mem_op("Conv", &onnx_parser::parse_conv);
Paul's avatar
Paul committed
79
80
        add_mem_op("MaxPool", &onnx_parser::parse_pooling);
        add_mem_op("AveragePool", &onnx_parser::parse_pooling);
81
82
        add_mem_op("GlobalMaxPool", &onnx_parser::parse_pooling);
        add_mem_op("GlobalAveragePool", &onnx_parser::parse_pooling);
Paul's avatar
Paul committed
83
        add_mem_op("Reshape", &onnx_parser::parse_reshape);
Paul's avatar
Paul committed
84
85
        add_mem_op("Flatten", &onnx_parser::parse_flatten);
        add_mem_op("Gemm", &onnx_parser::parse_gemm);
86
        add_mem_op("MatMul", &onnx_parser::parse_matmul);
87
        add_mem_op("BatchNormalization", &onnx_parser::parse_batchnorm);
Paul's avatar
Paul committed
88
        add_mem_op("Softmax", &onnx_parser::parse_softmax);
Shucai Xiao's avatar
Shucai Xiao committed
89
        add_mem_op("LogSoftmax", &onnx_parser::parse_logsoftmax);
90
91
92
        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
93
        add_mem_op("Concat", &onnx_parser::parse_concat);
94
95
96
        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);
97
        add_mem_op("ConstantOfShape", &onnx_parser::parse_constant_of_shape);
Khalique's avatar
Khalique committed
98
        add_mem_op("Transpose", &onnx_parser::parse_transpose);
Shucai Xiao's avatar
Shucai Xiao committed
99
        add_mem_op("RNN", &onnx_parser::parse_rnn);
100
        add_mem_op("GRU", &onnx_parser::parse_gru);
Shucai Xiao's avatar
Shucai Xiao committed
101
        add_mem_op("LSTM", &onnx_parser::parse_lstm);
Khalique's avatar
Khalique committed
102
        add_mem_op("Pad", &onnx_parser::parse_pad);
Shucai Xiao's avatar
Shucai Xiao committed
103
104
        add_mem_op("ReduceSum", &onnx_parser::parse_reduce_oper<op::reduce_sum>);
        add_mem_op("ReduceMean", &onnx_parser::parse_reduce_oper<op::reduce_mean>);
105
106
107
108
109
110
111

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

    void init_actv_func()
    {
112
113
114
115
116
117
        // Support name format of all lower case or the first letter capital
        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
118
119
120
121
    }

    template <class F>
    void add_op(std::string name, F f)
Paul's avatar
Paul committed
122
123
124
125
126
127
128
129
130
    {
        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
131
132
133
134
135
136
137
    {
        ops.emplace(name, f);
    }

    template <class F>
    void add_mem_op(std::string name, F f)
    {
Paul's avatar
Paul committed
138
        add_op(name, [=](auto&&... xs) {
Paul's avatar
Paul committed
139
140
141
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }
Khalique's avatar
Khalique committed
142

143
    template <class T>
Khalique's avatar
Khalique committed
144
    void add_binary_op(std::string name, T x)
145
    {
Paul's avatar
Paul committed
146
        add_op(name, [this, x](attribute_map attributes, std::vector<instruction_ref> args) {
Scott Thornton's avatar
Scott Thornton committed
147
            if(args.size() != 2)
Paul's avatar
Paul committed
148
                MIGRAPHX_THROW("binary operators should have 2 operands");
149
            if(contains(attributes, "broadcast") and contains(attributes, "axis"))
150
151
152
153
            {
                uint64_t broadcasted = parse_value(attributes.at("broadcast")).at<uint64_t>();
                if(broadcasted != 0)
                {
154
                    uint64_t axis = parse_value(attributes.at("axis")).at<uint64_t>();
Shucai Xiao's avatar
Shucai Xiao committed
155
156
                    auto l = prog.add_instruction(op::broadcast{axis, args[0]->get_shape().lens()},
                                                  args[1]);
157
158
                    return prog.add_instruction(x, args[0], l);
                }
159
                return prog.add_instruction(x, args);
160
            }
Paul's avatar
Paul committed
161
            else
162
            {
Khalique's avatar
Khalique committed
163
                return add_broadcastable_binary_op(args[0], args[1], x);
164
165
166
167
            }
        });
    }

Shucai Xiao's avatar
Shucai Xiao committed
168
169
    std::vector<std::size_t> compute_broadcasted_lens(std::vector<std::size_t> s0,
                                                      std::vector<std::size_t> s1)
170
171
172
173
174
175
176
177
178
179
180
181
182
    {
        // 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)
Shucai Xiao's avatar
Shucai Xiao committed
183
        if(s0.size() > s1.size())
184
185
186
187
188
189
        {
            s0.swap(s1);
        }

        std::vector<std::size_t> out_lens(s1);
        auto offset = s1.size() - s0.size();
Shucai Xiao's avatar
Shucai Xiao committed
190
191
192
193
        std::transform(s0.begin(),
                       s0.end(),
                       s1.begin() + offset,
                       out_lens.begin() + offset,
194
                       [&](auto a, auto b) {
Shucai Xiao's avatar
Shucai Xiao committed
195
                           if(a != b and a != 1 and b != 1)
196
                           {
Shucai Xiao's avatar
Shucai Xiao committed
197
198
199
200
201
202
                               MIGRAPHX_THROW("COMPUTE_BROADCASTLEN: shape {" +
                                              to_string_range(s0) + "} and {" +
                                              to_string_range(s1) + "} mismatch!");
                           }
                           return std::max(a, b);
                       });
203
204
205
206

        return out_lens;
    }

Khalique's avatar
Khalique committed
207
208
209
    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
Khalique's avatar
Khalique committed
210
        if(arg0->get_shape().lens() != arg1->get_shape().lens())
Khalique's avatar
Khalique committed
211
212
        {
            // Get lengths for both arguments
Shucai Xiao's avatar
Shucai Xiao committed
213
214
            auto s0       = arg0->get_shape().lens();
            auto s1       = arg1->get_shape().lens();
215
            auto out_lens = compute_broadcasted_lens(s0, s1);
Shucai Xiao's avatar
Shucai Xiao committed
216
217
            auto l0       = prog.add_instruction(op::multibroadcast{out_lens}, arg0);
            auto l1       = prog.add_instruction(op::multibroadcast{out_lens}, arg1);
Khalique's avatar
Khalique committed
218
219
220
221
222
223
            return prog.add_instruction(x, l0, l1);
        }
        else
        {
            return prog.add_instruction(x, {arg0, arg1});
        }
224
225
    }

Paul's avatar
Paul committed
226
    template <class T>
Paul's avatar
Paul committed
227
228
    void add_generic_op(std::string name, T x)
    {
Paul's avatar
Paul committed
229
        add_op(name, [this, x](const attribute_map&, std::vector<instruction_ref> args) {
Paul's avatar
Paul committed
230
231
232
233
            return prog.add_instruction(x, args);
        });
    }

Khalique's avatar
Khalique committed
234
    template <class T>
Khalique's avatar
Khalique committed
235
    void add_variadic_op(std::string name, T x)
Khalique's avatar
Khalique committed
236
    {
Paul's avatar
Paul committed
237
        add_op(name, [this, x](const attribute_map&, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
238
            return std::accumulate(std::next(args.begin()),
Khalique's avatar
Khalique committed
239
240
241
242
243
                                   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
244
        });
Khalique's avatar
Khalique committed
245
246
    }

Khalique's avatar
Khalique committed
247
248
249
    instruction_ref parse_clip(const std::string&,
                               const attribute_map& attributes,
                               std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
250
251
252
253
254
255
256
257
258
259
260
261
262
    {
        op::clip op;
        if(contains(attributes, "max"))
        {
            op.max_val = parse_value(attributes.at("max")).at<float>();
        }
        if(contains(attributes, "min"))
        {
            op.min_val = parse_value(attributes.at("min")).at<float>();
        }
        return prog.add_instruction(op, std::move(args));
    }

Paul's avatar
Paul committed
263
    instruction_ref
Paul's avatar
Paul committed
264
    parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
265
266
    {
        auto dims = args.front()->get_shape().lens();
Scott Thornton's avatar
Scott Thornton committed
267
268
        auto r =
            prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, args.front());
269
270
        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
271
272
    }

Shucai Xiao's avatar
Shucai Xiao committed
273
274
275
    instruction_ref parse_logsoftmax(const std::string&,
                                     const attribute_map& attributes,
                                     std::vector<instruction_ref> args)
Shucai Xiao's avatar
Shucai Xiao committed
276
277
278
279
280
281
282
283
284
285
    {
        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));
    }

286
    instruction_ref parse_argmax(const std::string&,
Shucai Xiao's avatar
Shucai Xiao committed
287
288
                                 const attribute_map& attributes,
                                 std::vector<instruction_ref> args)
289
    {
290
        int64_t axis = 0;
291
292
        if(contains(attributes, "axis"))
        {
293
            axis = static_cast<int64_t>(parse_value(attributes.at("axis")).at<int>());
294
295
        }

Shucai Xiao's avatar
Shucai Xiao committed
296
        int keep_dims = 1;
Shucai Xiao's avatar
Shucai Xiao committed
297
        if(contains(attributes, "keepdims"))
Shucai Xiao's avatar
Shucai Xiao committed
298
299
300
301
        {
            keep_dims = parse_value(attributes.at("keepdims")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
302
        if(keep_dims == 0)
303
304
        {
            auto ins = prog.add_instruction(op::argmax{axis}, std::move(args));
305
            return prog.add_instruction(op::squeeze{{axis}}, ins);
306
307
308
309
310
        }
        else
        {
            return prog.add_instruction(op::argmax{axis}, std::move(args));
        }
311
312
313
    }

    instruction_ref parse_argmin(const std::string&,
Shucai Xiao's avatar
Shucai Xiao committed
314
315
                                 const attribute_map& attributes,
                                 std::vector<instruction_ref> args)
316
    {
317
        int64_t axis = 0;
318
319
        if(contains(attributes, "axis"))
        {
320
            axis = static_cast<int64_t>(parse_value(attributes.at("axis")).at<int>());
321
322
        }

Shucai Xiao's avatar
Shucai Xiao committed
323
        int keep_dims = 1;
Shucai Xiao's avatar
Shucai Xiao committed
324
        if(contains(attributes, "keepdims"))
Shucai Xiao's avatar
Shucai Xiao committed
325
326
327
328
        {
            keep_dims = parse_value(attributes.at("keepdims")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
329
        if(keep_dims == 0)
330
331
        {
            auto ins = prog.add_instruction(op::argmin{axis}, std::move(args));
332
            return prog.add_instruction(op::squeeze{{axis}}, ins);
333
334
335
336
337
        }
        else
        {
            return prog.add_instruction(op::argmin{axis}, std::move(args));
        }
338
339
    }

Paul's avatar
Paul committed
340
    instruction_ref
Paul's avatar
Paul committed
341
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
342
    {
343
        op::convolution op;
344
        auto l0 = args[0];
Paul's avatar
Paul committed
345
346
        if(contains(attributes, "pads"))
        {
Scott Thornton's avatar
Scott Thornton committed
347
            if(contains(attributes, "auto_pad"))
348
            {
Paul's avatar
Paul committed
349
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
350
            }
351
352
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
353
            if(padding.size() != 4)
354
            {
Paul's avatar
Paul committed
355
                MIGRAPHX_THROW("padding should have 4 values");
356
            }
Scott Thornton's avatar
Scott Thornton committed
357
            if(padding[0] != padding[2] || padding[1] != padding[3])
358
            {
359
360
                // 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
361
                l0      = prog.add_instruction(op::pad{padding}, l0);
362
            }
363
364
365
366
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
367
            }
Paul's avatar
Paul committed
368
        }
Paul's avatar
Paul committed
369
370
371
372
373
374
375
376
        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
377
        if(contains(attributes, "auto_pad"))
378
379
        {
            auto s = attributes["auto_pad"].s();
Scott Thornton's avatar
Scott Thornton committed
380
            if(contains(attributes, "pads") and to_upper(s) != "NOTSET")
381
            {
Paul's avatar
Paul committed
382
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
383
384
            }

wsttiger's avatar
fixes  
wsttiger committed
385
            if(s.find("SAME") != std::string::npos)
386
            {
387
                op.padding_mode = op::padding_mode_t::same;
388
389
            }
        }
Khalique's avatar
Khalique committed
390
391
392
393
        if(contains(attributes, "group"))
        {
            op.group = parse_value(attributes.at("group")).at<int>();
        }
Paul's avatar
Paul committed
394
395
396
397
        if(args.size() == 3)
        {
            uint64_t axis = 1;
            auto l1       = prog.add_instruction(op, args[0], args[1]);
Shucai Xiao's avatar
Shucai Xiao committed
398
            auto l2 = prog.add_instruction(op::broadcast{axis, l1->get_shape().lens()}, args[2]);
399
            return prog.add_instruction(op::add{}, l1, l2);
Paul's avatar
Paul committed
400
        }
401
        return prog.add_instruction(op, l0, args[1]);
Paul's avatar
Paul committed
402
    }
Paul's avatar
Paul committed
403

Paul's avatar
Paul committed
404
405
406
    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
Paul's avatar
Paul committed
407
    {
Khalique's avatar
Khalique committed
408
        op::pooling op{ends_with(name, "MaxPool") ? "max" : "average"};
409
        auto l0 = args[0];
Khalique's avatar
Khalique committed
410
        if(starts_with(name, "Global"))
411
        {
Khalique's avatar
Khalique committed
412
413
            auto lens  = args.front()->get_shape().lens();
            op.lengths = {lens[2], lens[3]};
414
        }
Paul's avatar
Paul committed
415
416
        if(contains(attributes, "pads"))
        {
417
418
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
419
            if(padding.size() != 4)
420
            {
Paul's avatar
Paul committed
421
                MIGRAPHX_THROW("padding should have 4 values");
422
            }
Scott Thornton's avatar
Scott Thornton committed
423
            if(padding[0] != padding[2] || padding[1] != padding[3])
424
            {
425
426
                // 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
427
428
                l0 = prog.add_instruction(op::pad{padding, std::numeric_limits<float>::lowest()},
                                          l0);
429
430
431
432
433
            }
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
434
            }
Paul's avatar
Paul committed
435
436
437
438
439
440
441
442
443
        }
        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
444
        if(contains(attributes, "auto_pad"))
445
446
        {
            auto s = attributes["auto_pad"].s();
447
            if(s.find("SAME_UPPER") == std::string::npos)
448
            {
449
                MIGRAPHX_THROW("auto_pad only supports SAME_UPPER for pooling");
450
            }
451
            op.padding_mode = op::padding_mode_t::same;
452
453
        }

454
        return prog.add_instruction(op, l0);
Paul's avatar
Paul committed
455
456
    }

Paul's avatar
Paul committed
457
    instruction_ref
Paul's avatar
Paul committed
458
    parse_reshape(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
459
    {
460
        op::reshape op;
Paul's avatar
Paul committed
461
462
463
464
465
466
467
        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
468
            auto s = args[1]->eval();
Paul's avatar
Paul committed
469
            if(s.empty())
Paul's avatar
Paul committed
470
                MIGRAPHX_THROW("Dynamic shape is not supported.");
Paul's avatar
Paul committed
471
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
Paul's avatar
Paul committed
472
        }
Paul's avatar
Paul committed
473
474
475
        return prog.add_instruction(op, args[0]);
    }

Paul's avatar
Paul committed
476
    instruction_ref
Paul's avatar
Paul committed
477
    parse_flatten(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
478
    {
479
        uint64_t axis = 1;
Paul's avatar
Paul committed
480
481
482
483
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
484
        return prog.add_instruction(op::flatten{axis}, args[0]);
Paul's avatar
Paul committed
485
486
    }

487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
    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
505
506
507
508
509
510
511
    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));
    }
512

513
514
515
    instruction_ref
    parse_gather(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
516
        int axis = 0;
517
518
519
520
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
521
        op::gather op{axis};
522
523
524
        return prog.add_instruction(op, std::move(args));
    }

525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
    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
545
546
547
    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
Paul's avatar
Paul committed
548
    {
Shucai Xiao's avatar
Shucai Xiao committed
549
        literal v = parse_value(attributes.at("value"));
550
        // return empty literal
Shucai Xiao's avatar
Shucai Xiao committed
551
        if(v.get_shape().elements() == 0)
552
553
554
555
        {
            return prog.add_literal(literal{});
        }

556
557
558
        auto dim_size = attributes.at("value").t().dims_size();
        // if dim_size is 0, it is a scalar
        if(dim_size == 0)
559
        {
560
            migraphx::shape scalar_shape{v.get_shape().type()};
561
562
563
            return prog.add_literal(migraphx::literal{scalar_shape, v.data()});
        }

Paul's avatar
Paul committed
564
565
        return prog.add_literal(v);
    }
Paul's avatar
Paul committed
566

Paul's avatar
Paul committed
567
    instruction_ref
Paul's avatar
Paul committed
568
    parse_gemm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
569
570
    {
        float alpha = 1.0f;
Khalique's avatar
Khalique committed
571
        float beta  = 1.0f;
Paul's avatar
Paul committed
572
573
574
575
576
577
578
579
        bool transa = false;
        bool transb = false;
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        if(contains(attributes, "beta"))
        {
580
            beta = parse_value(attributes.at("beta")).at<float>();
Paul's avatar
Paul committed
581
582
583
584
585
586
587
588
589
        }
        if(contains(attributes, "transA"))
        {
            transa = parse_value(attributes.at("transA")).at<bool>();
        }
        if(contains(attributes, "transB"))
        {
            transb = parse_value(attributes.at("transB")).at<bool>();
        }
590
591
592
593
594
595

        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));

596
597
        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
598
599
        if(args.size() == 3)
        {
600
            if(beta != 0.f && args[2]->get_shape().elements() > 0)
601
            {
Shucai Xiao's avatar
Shucai Xiao committed
602
                auto out_lens   = l1->get_shape().lens();
603
                out_lens.back() = l2->get_shape().lens().back();
Shucai Xiao's avatar
Shucai Xiao committed
604
                auto l3         = args[2];
Shucai Xiao's avatar
Shucai Xiao committed
605
606
                auto l3_lens    = l3->get_shape().lens();
                if(!std::equal(out_lens.begin(), out_lens.end(), l3_lens.begin(), l3_lens.end()))
Khalique's avatar
Khalique committed
607
                {
608
                    l3 = prog.add_instruction(op::multibroadcast{out_lens}, args[2]);
Khalique's avatar
Khalique committed
609
                }
610
                return prog.add_instruction(op::dot{alpha, beta}, l1, l2, l3);
611
            }
Paul's avatar
Paul committed
612
        }
613
614

        return prog.add_instruction(op::dot{alpha, beta}, l1, l2);
Paul's avatar
Paul committed
615
616
    }

617
    instruction_ref
Shucai Xiao's avatar
Shucai Xiao committed
618
    parse_matmul(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
619
    {
Shucai Xiao's avatar
Shucai Xiao committed
620
621
        auto l0      = args[0];
        auto l1      = args[1];
622
623
624
625
626
        auto l0_lens = l0->get_shape().lens();
        auto l1_lens = l1->get_shape().lens();

        // args[0] is a vector, prepend 1 to the shape
        bool is_a_prepended = false;
Shucai Xiao's avatar
Shucai Xiao committed
627
        if(l0_lens.size() == 1)
628
629
630
631
632
633
634
        {
            is_a_prepended = true;
            l0_lens.insert(l0_lens.begin(), 1);
            l0 = prog.add_instruction(op::unsqueeze{{0}}, args[0]);
        }

        bool is_b_appended = false;
Shucai Xiao's avatar
Shucai Xiao committed
635
        if(l1_lens.size() == 1)
636
637
638
639
640
641
642
643
        {
            is_b_appended = true;
            l1_lens.push_back(1);
            l1 = prog.add_instruction(op::unsqueeze{{1}}, args[1]);
        }

        instruction_ref bl0 = l0;
        instruction_ref bl1 = l1;
Shucai Xiao's avatar
Shucai Xiao committed
644
        if(!std::equal(l0_lens.rbegin() + 2, l0_lens.rend(), l1_lens.rbegin() + 2, l1_lens.rend()))
645
646
647
648
649
650
        {
            auto l0_it = l0_lens.begin() + l0_lens.size() - 2;
            std::vector<std::size_t> l0_broadcasted_lens(l0_lens.begin(), l0_it);
            auto l1_it = l1_lens.begin() + l1_lens.size() - 2;
            std::vector<std::size_t> l1_broadcasted_lens(l1_lens.begin(), l1_it);
            auto output_lens = compute_broadcasted_lens(l0_broadcasted_lens, l1_broadcasted_lens);
651
            l0_broadcasted_lens = output_lens;
652
            l0_broadcasted_lens.insert(l0_broadcasted_lens.end(), l0_it, l0_lens.end());
653
            l1_broadcasted_lens = output_lens;
654
            l1_broadcasted_lens.insert(l1_broadcasted_lens.end(), l1_it, l1_lens.end());
Shucai Xiao's avatar
Shucai Xiao committed
655
            if(l0_lens != l0_broadcasted_lens)
656
657
658
            {
                bl0 = prog.add_instruction(op::multibroadcast{l0_broadcasted_lens}, l0);
            }
Shucai Xiao's avatar
Shucai Xiao committed
659
            if(l1_lens != l1_broadcasted_lens)
660
661
662
663
664
            {
                bl1 = prog.add_instruction(op::multibroadcast{l1_broadcasted_lens}, l1);
            }
        }

Shucai Xiao's avatar
Shucai Xiao committed
665
        auto dot_res     = prog.add_instruction(op::dot{1.0f, 0.0f}, bl0, bl1);
666
        int64_t num_axis = static_cast<int64_t>(dot_res->get_shape().lens().size());
Shucai Xiao's avatar
Shucai Xiao committed
667
        if(is_a_prepended)
668
669
670
671
        {
            dot_res = prog.add_instruction(op::squeeze{{num_axis - 2}}, dot_res);
            --num_axis;
        }
Shucai Xiao's avatar
Shucai Xiao committed
672
        if(is_b_appended)
673
674
675
        {
            dot_res = prog.add_instruction(op::squeeze{{num_axis - 1}}, dot_res);
        }
Shucai Xiao's avatar
Shucai Xiao committed
676

677
678
679
        return dot_res;
    }

680
    instruction_ref
Paul's avatar
Paul committed
681
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
682
    {
Scott Thornton's avatar
Scott Thornton committed
683
684
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
685
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
Scott Thornton's avatar
Scott Thornton committed
686
        bool is_test                                      = false;
687
688
689
690
691
692
        if(contains(attributes, "epsilon"))
        {
            epsilon = parse_value(attributes.at("epsilon")).at<float>();
        }
        if(contains(attributes, "momentum"))
        {
693
            momentum = parse_value(attributes.at("momentum")).at<float>();
694
695
696
        }
        if(contains(attributes, "is_test"))
        {
wsttiger's avatar
wsttiger committed
697
            is_test = parse_value(attributes.at("is_test")).at<uint64_t>() > 0;
698
699
700
        }
        if(contains(attributes, "spatial"))
        {
701
            bn_mode = (parse_value(attributes.at("spatial")).at<uint64_t>() > 0)
702
703
                          ? op::batch_norm_inference::spatial
                          : op::batch_norm_inference::per_activation;
704
        }
Paul's avatar
Paul committed
705
        (void)is_test;
Paul's avatar
Paul committed
706
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
Paul's avatar
Paul committed
707
        return prog.add_instruction(op, std::move(args));
708
709
    }

710
711
712
713
    instruction_ref parse_leaky_relu(const std::string&,
                                     attribute_map attributes,
                                     std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
714
        float alpha = 0.01; // default alpha val for leaky relu
715
716
717
718
719
720
721
722
        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
723
724
    instruction_ref
    parse_elu(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
725
726
727
728
729
730
731
732
733
734
    {
        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
735
736
    instruction_ref
    parse_lrn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
737
738
    {
        float alpha = 0.0001;
Khalique's avatar
Khalique committed
739
740
741
        float beta  = 0.75;
        float bias  = 1.0;
        int size    = 1;
Khalique's avatar
Khalique committed
742
743
744
745
746
747
748
749
750
751
752
753
        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
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
    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());
        }
770
        auto input_lens = args.front()->get_shape().lens();
Khalique's avatar
Khalique committed
771

Khalique's avatar
Khalique committed
772
773
        auto scale_val = prog.add_literal(scale);
        auto bias_vals = prog.add_literal(
Paul's avatar
Paul committed
774
            migraphx::literal{migraphx::shape{migraphx::shape::float_type, {bias.size()}}, bias});
Khalique's avatar
Khalique committed
775

776
        auto scale_tensor = prog.add_instruction(migraphx::op::scalar{input_lens}, scale_val);
Paul's avatar
Paul committed
777
        auto img_scaled   = prog.add_instruction(migraphx::op::mul{}, args.front(), scale_tensor);
Shucai Xiao's avatar
Shucai Xiao committed
778
        auto bias_bcast   = prog.add_instruction(migraphx::op::broadcast{1, input_lens}, bias_vals);
Paul's avatar
Paul committed
779
        return prog.add_instruction(migraphx::op::add{}, img_scaled, bias_bcast);
Khalique's avatar
Khalique committed
780
    }
Khalique's avatar
Khalique committed
781

Khalique's avatar
Khalique committed
782
783
    instruction_ref
    parse_transpose(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
784
785
786
787
788
789
790
    {
        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
791
        return prog.add_instruction(migraphx::op::transpose{perm}, args.front());
Khalique's avatar
Khalique committed
792
793
    }

Khalique's avatar
Khalique committed
794
795
796
797
798
799
800
801
802
803
    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());
        }
804
        // check if padding is actually being done (at least one value is nonzero)
Khalique's avatar
Khalique committed
805
        if(std::all_of(pads.begin(), pads.end(), [](const int& i) { return i == 0; }))
806
807
808
        {
            return prog.add_instruction(migraphx::op::identity{}, args.front());
        }
Khalique's avatar
Khalique committed
809
810
811
812
813
814
815
816
817
818
819
820
        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());
    }
821
822
823
    // 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
824
    parse_shape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
825
826
    {
        if(args.size() != 1)
827
            MIGRAPHX_THROW("Shape: operator should have 1 operand");
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
        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
864
865
        if(contains(attributes, "extra_shape"))
        {
866
            MIGRAPHX_THROW("ConstantFill: cannot handle extra shape attribute");
867
868
        }

869
870
        if(input_as_shape == 1)
        {
Shucai Xiao's avatar
Shucai Xiao committed
871
            if(args.size() != 1)
872
            {
873
                MIGRAPHX_THROW("ConstantFill: need an input argument as output shape");
874
875
            }

Shucai Xiao's avatar
Shucai Xiao committed
876
877
            if(contains(attributes, "shape"))
            {
878
                MIGRAPHX_THROW("ConstantFill: cannot set the shape argument and pass in an input "
Shucai Xiao's avatar
Shucai Xiao committed
879
                               "at the same time");
880
881
            }

882
883
884
            migraphx::argument in = args[0]->eval();
            if(in.empty())
            {
885
                MIGRAPHX_THROW("ConstantFill: cannot handle dynamic shape as input");
886
            }
887

888
889
890
            std::vector<std::size_t> dims;
            in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
            migraphx::shape s(type, dims);
891
892
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
893
894
895
        }
        else if(input_as_shape == 0)
        {
Shucai Xiao's avatar
Shucai Xiao committed
896
897
            if(!contains(attributes, "shape"))
            {
898
                MIGRAPHX_THROW("ConstantFill: attribute output shape is needed");
899
900
901
            }

            literal ls = parse_value(attributes.at("shape"));
902
            std::vector<std::size_t> dims;
Shucai Xiao's avatar
Shucai Xiao committed
903
            ls.visit([&](auto s) { dims.assign(s.begin(), s.end()); });
904
            migraphx::shape s{type, dims};
905
906
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
907
908
909
        }
        else
        {
910
            MIGRAPHX_THROW("ConstantFill: wrong value of attribute input_as_shape");
911
912
913
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
914
915
916
    instruction_ref parse_constant_of_shape(const std::string&,
                                            attribute_map attributes,
                                            std::vector<instruction_ref> args)
917
918
    {
        literal l_val{};
Shucai Xiao's avatar
Shucai Xiao committed
919
        if(contains(attributes, "value"))
920
921
        {
            l_val = parse_value(attributes.at("value"));
Shucai Xiao's avatar
Shucai Xiao committed
922
            if(l_val.get_shape().elements() != 1)
923
924
925
926
927
928
929
930
931
932
933
            {
                MIGRAPHX_THROW("ConstantOfShape: attribute value can contain only 1 elements!");
            }
        }
        else
        {
            l_val = literal({shape::float_type, {1}, {0}}, {0.0f});
        }

        // input is empty, output is a scalar
        auto type = l_val.get_shape().type();
934

Shucai Xiao's avatar
Shucai Xiao committed
935
        if(args.empty())
936
        {
937
            MIGRAPHX_THROW("Parse ConstantOfShape : must have 1 input!");
938
939
940
        }
        else
        {
941
942
            migraphx::shape s;
            // empty input tensor, output is a scalar
Shucai Xiao's avatar
Shucai Xiao committed
943
            if(args[0]->get_shape().elements() == 0)
944
            {
945
                s = migraphx::shape{type, {1}, {0}};
946
            }
947
948
949
950
951
952
953
            else
            {
                migraphx::argument in = args[0]->eval();
                if(in.empty())
                {
                    MIGRAPHX_THROW("Parse ConstantOfShape: cannot handle dynamic shape as input");
                }
954

955
956
957
958
                std::vector<std::size_t> dims;
                in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
                s = migraphx::shape{type, dims};
            }
959
960
961
962
963
964
965
966
967
968
969
970
971

            literal l_out;
            l_val.visit([&](auto val) {
                using type = std::remove_cv_t<typename decltype(val)::value_type>;
                // l_val contains only one element
                std::vector<type> out_vec(s.elements(), *val.begin());
                l_out = literal(s, out_vec);
            });

            return prog.add_literal(l_out);
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
972
    instruction_ref
Shucai Xiao's avatar
Shucai Xiao committed
973
    parse_expand(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
974
    {
Shucai Xiao's avatar
Shucai Xiao committed
975
        auto in_lens             = args[0]->get_shape().lens();
976
        migraphx::argument arg_s = args[1]->eval();
Shucai Xiao's avatar
Shucai Xiao committed
977
        if(arg_s.empty())
978
979
980
981
982
983
        {
            MIGRAPHX_THROW("Parse Expand: cannot handle dynamic shape as input");
        }
        std::vector<std::size_t> dims;
        arg_s.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
        auto out_lens = compute_broadcasted_lens(in_lens, dims);
984

Shucai Xiao's avatar
Shucai Xiao committed
985
        return prog.add_instruction(op::multibroadcast{out_lens}, args[0]);
986
987
    }

Shucai Xiao's avatar
Shucai Xiao committed
988
    std::vector<instruction_ref>
Shucai Xiao's avatar
Shucai Xiao committed
989
990
991
    parse_rnn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        migraphx::shape input_shape = args[0]->get_shape();
992
        std::size_t hidden_size     = args[1]->get_shape().lens()[1];
Shucai Xiao's avatar
Shucai Xiao committed
993
994
995

        if(contains(attributes, "hidden_size"))
        {
Shucai Xiao's avatar
Shucai Xiao committed
996
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
997
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
998
999
1000
            {
                MIGRAPHX_THROW("RNN: hidden size mismatch in input and attribute");
            }
Shucai Xiao's avatar
Shucai Xiao committed
1001
1002
1003
1004
1005
1006
1007
1008
1009
        }

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

1010
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
1011
1012
        if(direction == "bidirectional")
        {
1013
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
1014
1015
1016
        }
        else if(direction == "reverse")
        {
1017
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
1018
1019
        }

1020
        std::vector<std::string> vec_names{"tanh"};
1021
1022
1023
1024
        if(contains(attributes, "activations"))
        {
            auto names = attributes.at("activations").strings();
            vec_names.clear();
1025
            vec_names.resize(names.size());
Shucai Xiao's avatar
Shucai Xiao committed
1026
1027
1028
            std::transform(names.begin(), names.end(), vec_names.begin(), [](auto name) {
                return to_lower(name);
            });
1029
1030
        }

1031
1032
1033
        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
1034
        if(name_it != vec_names.end())
1035
1036
1037
        {
            MIGRAPHX_THROW("RNN: activation function " + std::string(*name_it) + " not supported");
        }
1038

Shucai Xiao's avatar
Shucai Xiao committed
1039
        // bidirectional case should have two activation functions.
Shucai Xiao's avatar
Shucai Xiao committed
1040
        // one is for forward, and the other is for reverse.
Shucai Xiao's avatar
Shucai Xiao committed
1041
        // if only one actv function is provided, we use it in both
1042
        // forward and reverse direction
1043
        if(dirct == op::rnn_direction::bidirectional)
1044
        {
Shucai Xiao's avatar
Shucai Xiao committed
1045
            if(vec_names.size() == 1)
1046
1047
1048
1049
1050
            {
                vec_names.push_back(vec_names.at(0));
            }
        }

Shucai Xiao's avatar
Shucai Xiao committed
1051
1052
1053
        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];
1054
        });
Shucai Xiao's avatar
Shucai Xiao committed
1055

Shucai Xiao's avatar
Shucai Xiao committed
1056
1057
1058
1059
1060
1061
1062
        // To be added later
        float clip = 0.0;
        if(contains(attributes, "clip"))
        {
            clip = parse_value(attributes.at("clip")).at<float>();
        }

1063
1064
        // if the number of arguments is less than 6, append
        // undefined operator to have 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
1065
        if(args.size() < 6)
1066
1067
1068
1069
1070
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), (6 - args.size()), ins);
        }

Shucai Xiao's avatar
Shucai Xiao committed
1071
1072
        // 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
1073
                                                  std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
1074

1075
        // second output for the last hidden state
Shucai Xiao's avatar
Shucai Xiao committed
1076
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
1077

Shucai Xiao's avatar
Shucai Xiao committed
1078
        return {hidden_states, last_output};
Shucai Xiao's avatar
Shucai Xiao committed
1079
1080
    }

1081
    std::vector<instruction_ref>
1082
1083
1084
1085
1086
1087
1088
    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
1089
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
1090
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
1091
1092
1093
            {
                MIGRAPHX_THROW("GRU: hidden size mismatch in input and attribute");
            }
1094
1095
1096
1097
1098
1099
1100
1101
1102
        }

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

1103
        op::rnn_direction dirct = op::rnn_direction::forward;
1104
1105
        if(direction == "bidirectional")
        {
1106
            dirct = op::rnn_direction::bidirectional;
1107
1108
1109
        }
        else if(direction == "reverse")
        {
1110
            dirct = op::rnn_direction::reverse;
1111
1112
        }

1113
        std::vector<std::string> vec_names = {"sigmoid", "tanh"};
1114
1115
        if(contains(attributes, "activations"))
        {
1116
            auto names = attributes.at("activations").strings();
1117
            vec_names.clear();
Shucai Xiao's avatar
Shucai Xiao committed
1118
            vec_names.resize(names.size());
Shucai Xiao's avatar
Shucai Xiao committed
1119
1120
1121
            std::transform(names.begin(), names.end(), vec_names.begin(), [](auto name) {
                return to_lower(name);
            });
1122
1123
        }

1124
        // need 4 activation functions
1125
        if(dirct == op::rnn_direction::bidirectional)
1126
        {
Shucai Xiao's avatar
Shucai Xiao committed
1127
            // 4 activation functions are used in the bidirectional
1128
            // scenario. No spec is provided in onnx::operator. we
Shucai Xiao's avatar
Shucai Xiao committed
1129
1130
            // use the algorithm that: if 1 actv function is provided,
            // repeat 1 four times. If 2 actv functins are provided,
1131
1132
            // 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
1133
1134
1135
            // assume the 3rd one is repeated once and used by the
            // reverse direction.
            // This may need change later
1136
            if(vec_names.size() == 1)
1137
            {
1138
                vec_names.insert(vec_names.end(), 3, vec_names.at(0));
1139
            }
1140
            else if(vec_names.size() == 2)
1141
            {
1142
1143
1144
                // repeat the activation functions
                vec_names.push_back(vec_names.at(0));
                vec_names.push_back(vec_names.at(1));
1145
            }
1146
            else if(vec_names.size() == 3)
1147
            {
1148
                vec_names.push_back(vec_names.at(2));
1149
1150
            }
        }
Shucai Xiao's avatar
Shucai Xiao committed
1151
        else
1152
        {
1153
            if(vec_names.size() == 1)
1154
            {
1155
                vec_names.push_back(vec_names.at(0));
1156
1157
1158
            }
        }

1159
1160
1161
        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
1162
        if(name_it != vec_names.end())
1163
1164
1165
        {
            MIGRAPHX_THROW("GRU: activation function " + std::string(*name_it) + " not supported");
        }
1166

Shucai Xiao's avatar
Shucai Xiao committed
1167
1168
1169
        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
1170
        });
1171
1172
1173
1174
1175
1176
1177
1178

        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
1179
        if(contains(attributes, "linear_before_reset"))
1180
1181
1182
1183
        {
            linear_before_reset = parse_value(attributes.at("linear_before_reset")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
1184
        // append undefined opeator to make 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
1185
        if(args.size() < 6)
Shucai Xiao's avatar
Shucai Xiao committed
1186
1187
1188
1189
1190
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), 6 - args.size(), ins);
        }

1191
1192
        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
1193
            op::gru{hidden_size, vec_actv_funcs, dirct, clip, linear_before_reset},
Shucai Xiao's avatar
Shucai Xiao committed
1194
            std::move(args));
1195
1196

        // second output for last gru output
1197
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
1198

Shucai Xiao's avatar
Shucai Xiao committed
1199
        return {hidden_states, last_output};
1200
1201
    }

Shucai Xiao's avatar
Shucai Xiao committed
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
    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
1224
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
1225
1226
        if(direction == "bidirectional")
        {
Shucai Xiao's avatar
Shucai Xiao committed
1227
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
1228
1229
1230
        }
        else if(direction == "reverse")
        {
Shucai Xiao's avatar
Shucai Xiao committed
1231
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
1232
        }
Shucai Xiao's avatar
Shucai Xiao committed
1233
        else if(direction == "forward")
Shucai Xiao's avatar
Shucai Xiao committed
1234
        {
Shucai Xiao's avatar
Shucai Xiao committed
1235
            dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
1236
1237
1238
1239
1240
1241
        }
        else
        {
            MIGRAPHX_THROW("LSTM: incorrect direction attribute");
        }

1242
        std::vector<std::string> vec_names = {"sigmoid", "tanh", "tanh"};
Shucai Xiao's avatar
Shucai Xiao committed
1243
1244
1245
1246
1247
        if(contains(attributes, "activations"))
        {
            auto names = attributes.at("activations").strings();
            vec_names.clear();
            vec_names.resize(names.size());
Shucai Xiao's avatar
Shucai Xiao committed
1248
1249
1250
            std::transform(names.begin(), names.end(), vec_names.begin(), [](auto name) {
                return to_lower(name);
            });
Shucai Xiao's avatar
Shucai Xiao committed
1251
1252
1253
        }

        // need 6 activation functions for bidirectional directions
Shucai Xiao's avatar
Shucai Xiao committed
1254
        if(dirct == op::rnn_direction::bidirectional)
Shucai Xiao's avatar
Shucai Xiao committed
1255
1256
1257
1258
1259
1260
        {
            // 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
1261
            // if 3 actv funcs are provide, repeat all three once.
Shucai Xiao's avatar
Shucai Xiao committed
1262
1263
1264
1265
            // the same algorithm is used for 4, 5, and 6 actv funcions
            // provided. This may need change later
            switch(vec_names.size())
            {
1266
            case 1:
Shucai Xiao's avatar
Shucai Xiao committed
1267
1268
1269
1270
1271
1272
                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)};
1273
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1274
1275
1276

            case 2:
                // repeat the 2nd actv func once, then repeat all three another time
Shucai Xiao's avatar
Shucai Xiao committed
1277
1278
1279
1280
1281
1282
                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
1283
1284
1285
1286
                break;

            case 3:
                // repeat all three actv funcs once
Shucai Xiao's avatar
Shucai Xiao committed
1287
1288
1289
1290
1291
1292
                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
1293
1294
                break;

Shucai Xiao's avatar
Shucai Xiao committed
1295
1296
1297
1298
1299
1300
1301
            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)};
1302
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1303

Shucai Xiao's avatar
Shucai Xiao committed
1304
1305
1306
1307
1308
1309
1310
            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)};
1311
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1312

Shucai Xiao's avatar
Shucai Xiao committed
1313
            default: break;
Shucai Xiao's avatar
Shucai Xiao committed
1314
1315
1316
1317
1318
1319
            }
        }
        else
        {
            switch(vec_names.size())
            {
Shucai Xiao's avatar
Shucai Xiao committed
1320
            case 1: vec_names = {vec_names.at(0), vec_names.at(0), vec_names.at(0)}; break;
Shucai Xiao's avatar
Shucai Xiao committed
1321
1322
1323

            case 2:
                // repeat the 2nd actv func once, so we have 3 actv funcs
Shucai Xiao's avatar
Shucai Xiao committed
1324
                vec_names = {vec_names.at(0), vec_names.at(1), vec_names.at(1)};
Shucai Xiao's avatar
Shucai Xiao committed
1325
1326
                break;

Shucai Xiao's avatar
Shucai Xiao committed
1327
            default: break;
Shucai Xiao's avatar
Shucai Xiao committed
1328
1329
1330
            }
        }

1331
1332
1333
        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
1334
        if(name_it != vec_names.end())
1335
1336
1337
        {
            MIGRAPHX_THROW("LSTM: activation function " + std::string(*name_it) + " not supported");
        }
Shucai Xiao's avatar
Shucai Xiao committed
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359

        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
1360
            args.insert(args.end(), 8 - args.size(), ins);
Shucai Xiao's avatar
Shucai Xiao committed
1361
1362
1363
1364
        }

        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
1365
            op::lstm{hidden_size, vec_actv_funcs, dirct, clip, input_forget}, std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
1366
1367

        // second output for last lstm output
Shucai Xiao's avatar
Shucai Xiao committed
1368
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
1369
1370
1371
1372
1373
1374

        // 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};
    }
1375

Shucai Xiao's avatar
Shucai Xiao committed
1376
    template <class T>
Shucai Xiao's avatar
Shucai Xiao committed
1377
    instruction_ref parse_reduce_oper(const std::string&,
Shucai Xiao's avatar
Shucai Xiao committed
1378
1379
                                      attribute_map attributes,
                                      std::vector<instruction_ref> args)
Shucai Xiao's avatar
Shucai Xiao committed
1380
1381
1382
1383
    {
        std::size_t n_dim = args.front()->get_shape().lens().size();

        // default to reduce over all dimensions
1384
        std::vector<int64_t> axes(n_dim);
Shucai Xiao's avatar
Shucai Xiao committed
1385
1386
1387
1388
1389
        std::iota(axes.begin(), axes.end(), 0);
        if(contains(attributes, "axes"))
        {
            axes.clear();
            auto&& attr_axes = attributes["axes"].ints();
1390
            axes             = std::vector<int64_t>(attr_axes.begin(), attr_axes.end());
Shucai Xiao's avatar
Shucai Xiao committed
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
        }

        int keep_dims = 1;
        if(contains(attributes, "keepdims"))
        {
            keep_dims = parse_value(attributes.at("keepdims")).at<int>();
        }

        if(keep_dims == 1)
        {
Shucai Xiao's avatar
Shucai Xiao committed
1401
            return prog.add_instruction(T{axes}, std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
1402
1403
1404
        }
        else
        {
Shucai Xiao's avatar
Shucai Xiao committed
1405
            auto ins = prog.add_instruction(T{axes}, std::move(args));
1406
            return prog.add_instruction(op::squeeze{axes}, ins);
1407
1408
        }
    }
1409

Shucai Xiao's avatar
Shucai Xiao committed
1410
1411
    instruction_ref
    parse_cast(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
1412
    {
Shucai Xiao's avatar
Shucai Xiao committed
1413
        if(!contains(attributes, "to"))
1414
1415
1416
1417
        {
            MIGRAPHX_THROW("PARSE_CAST: missing to type attribute!");
        }

Shucai Xiao's avatar
Shucai Xiao committed
1418
        int to_type        = parse_value(attributes.at("to")).at<int>();
1419
1420
1421
        shape::type_t type = get_type(to_type);
        return prog.add_instruction(op::convert{type}, std::move(args));
    }
Shucai Xiao's avatar
Shucai Xiao committed
1422

Paul's avatar
Paul committed
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
    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
1435
            MIGRAPHX_THROW("Failed reading onnx file.");
Paul's avatar
Paul committed
1436
1437
1438
1439
1440
1441
        }
    }

    void parse_graph(const onnx::GraphProto& graph)
    {
        nodes = get_nodes(graph);
1442
1443
1444
1445
1446
        std::unordered_map<std::string, onnx::TensorProto> initializer_data;
        for(auto&& f : graph.initializer())
        {
            initializer_data[f.name()] = f;
        }
Paul's avatar
Paul committed
1447
1448
1449
        for(auto&& input : graph.input())
        {
            const std::string& name = input.name();
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
            // 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
1462
        }
Paul's avatar
Paul committed
1463
        for(auto&& output : graph.output())
Paul's avatar
Paul committed
1464
        {
Paul's avatar
Paul committed
1465
            this->parse_node(output.name());
Paul's avatar
Paul committed
1466
1467
1468
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
1469
    void parse_undefined(const std::string& name)
1470
    {
Shucai Xiao's avatar
Shucai Xiao committed
1471
        auto ins           = prog.add_instruction(op::undefined{});
1472
1473
1474
        instructions[name] = ins;
    }

Paul's avatar
Paul committed
1475
    void parse_node(const std::string& name)
Paul's avatar
Paul committed
1476
    {
Paul's avatar
Paul committed
1477
        if(name.empty())
Paul's avatar
Paul committed
1478
            MIGRAPHX_THROW("Onnx node must have a name");
Paul's avatar
Paul committed
1479
1480
1481
1482
1483
1484
1485
1486
        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
1487
1488
                    assert(name != input);
                    this->parse_node(input);
Paul's avatar
Paul committed
1489
                }
Shucai Xiao's avatar
Shucai Xiao committed
1490
                else if(input.empty())
Paul's avatar
Paul committed
1491
                {
1492
                    this->parse_undefined(input);
Paul's avatar
Paul committed
1493
                }
1494
                args.push_back(instructions.at(input));
Paul's avatar
Paul committed
1495
            }
Paul's avatar
Paul committed
1496
            std::vector<instruction_ref> result;
Paul's avatar
Paul committed
1497
1498
            if(ops.count(node.op_type()) == 0)
            {
1499
                result.push_back(prog.add_instruction(op::unknown{node.op_type()}, args));
Paul's avatar
Paul committed
1500
1501
1502
            }
            else
            {
Paul's avatar
Paul committed
1503
                result = ops[node.op_type()](get_attributes(node), args);
Paul's avatar
Paul committed
1504
            }
Paul's avatar
Paul committed
1505
            // Even no output nodes produce output in migraphx
Paul's avatar
Paul committed
1506
            if(node.output().empty() and result.size() == 1)
Paul's avatar
Paul committed
1507
1508
            {
                instructions[name] = result.front();
Paul's avatar
Paul committed
1509
1510
1511
            }
            else
            {
Paul's avatar
Paul committed
1512
1513
1514
1515
1516
1517
                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
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
            }
        }
    }

    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
1535
        std::size_t n = 0;
Paul's avatar
Paul committed
1536
1537
        for(auto&& node : graph.node())
        {
Paul's avatar
Paul committed
1538
            if(node.output().empty())
Paul's avatar
Paul committed
1539
            {
Paul's avatar
Paul committed
1540
                if(node.name().empty())
Paul's avatar
Paul committed
1541
1542
1543
1544
1545
1546
1547
1548
1549
                {
                    result["migraphx_unamed_node_" + std::to_string(n)] = node;
                    n++;
                }
                else
                {
                    result[node.name()] = node;
                }
            }
Paul's avatar
Paul committed
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
            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
1575
        case onnx::AttributeProto::FLOATS: return from_repeated(shape::float_type, attr.floats());
Paul's avatar
Paul committed
1576
1577
1578
1579
1580
        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
1581
        MIGRAPHX_THROW("Invalid attribute type");
Paul's avatar
Paul committed
1582
1583
1584
1585
1586
    }

    static literal parse_tensor(const onnx::TensorProto& t)
    {
        std::vector<std::size_t> dims(t.dims().begin(), t.dims().end());
1587
1588
        if(t.has_raw_data())
        {
wsttiger's avatar
wsttiger committed
1589
            const std::string& s = t.raw_data();
Scott Thornton's avatar
Scott Thornton committed
1590
1591
1592
            switch(t.data_type())
            {
            case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
1593
            case onnx::TensorProto::FLOAT: return create_literal(shape::float_type, dims, s.data());
Scott Thornton's avatar
Scott Thornton committed
1594
            case onnx::TensorProto::UINT8: throw std::runtime_error("");
1595
            case onnx::TensorProto::INT8: return create_literal(shape::int32_type, dims, s.data());
Khalique's avatar
Khalique committed
1596
1597
            case onnx::TensorProto::UINT16:
                return create_literal(shape::int32_type, dims, s.data());
1598
1599
1600
            case onnx::TensorProto::INT16: return create_literal(shape::int32_type, dims, s.data());
            case onnx::TensorProto::INT32: return create_literal(shape::int32_type, dims, s.data());
            case onnx::TensorProto::INT64: return create_literal(shape::int64_type, dims, s.data());
Scott Thornton's avatar
Scott Thornton committed
1601
            case onnx::TensorProto::STRING: throw std::runtime_error("");
1602
            case onnx::TensorProto::BOOL: return create_literal(shape::int32_type, dims, s.data());
Khalique's avatar
Khalique committed
1603
1604
1605
1606
            case onnx::TensorProto::FLOAT16:
                return create_literal(shape::half_type, dims, s.data());
            case onnx::TensorProto::DOUBLE:
                return create_literal(shape::double_type, dims, s.data());
Scott Thornton's avatar
Scott Thornton committed
1607
1608
1609
1610
1611
            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
1612
            MIGRAPHX_THROW("Invalid tensor type");
1613
        }
Paul's avatar
Paul committed
1614
1615
1616
1617
        switch(t.data_type())
        {
        case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
        case onnx::TensorProto::FLOAT:
Khalique's avatar
Khalique committed
1618
            return create_literal(shape::float_type, dims, t.float_data());
Paul's avatar
Paul committed
1619
1620
        case onnx::TensorProto::UINT8: throw std::runtime_error("");
        case onnx::TensorProto::INT8:
Khalique's avatar
Khalique committed
1621
            return create_literal(shape::int32_type, dims, t.int32_data());
Paul's avatar
Paul committed
1622
        case onnx::TensorProto::UINT16:
Khalique's avatar
Khalique committed
1623
            return create_literal(shape::int32_type, dims, t.int32_data());
Paul's avatar
Paul committed
1624
        case onnx::TensorProto::INT16:
Khalique's avatar
Khalique committed
1625
            return create_literal(shape::int32_type, dims, t.int32_data());
Paul's avatar
Paul committed
1626
        case onnx::TensorProto::INT32:
Khalique's avatar
Khalique committed
1627
            return create_literal(shape::int32_type, dims, t.int32_data());
Paul's avatar
Paul committed
1628
        case onnx::TensorProto::INT64:
Khalique's avatar
Khalique committed
1629
            return create_literal(shape::int64_type, dims, t.int64_data());
Paul's avatar
Paul committed
1630
1631
        case onnx::TensorProto::STRING: throw std::runtime_error("");
        case onnx::TensorProto::BOOL:
Khalique's avatar
Khalique committed
1632
            return create_literal(shape::int32_type, dims, t.int32_data());
Paul's avatar
Paul committed
1633
        case onnx::TensorProto::FLOAT16:
Khalique's avatar
Khalique committed
1634
        {
Khalique's avatar
Khalique committed
1635
            std::vector<uint16_t> data_uint16(t.int32_data().begin(), t.int32_data().end());
1636
            std::vector<half> data_half;
Khalique's avatar
Khalique committed
1637
1638
1639
            std::transform(data_uint16.begin(),
                           data_uint16.end(),
                           std::back_inserter(data_half),
1640
                           [](uint16_t raw_val) { return *reinterpret_cast<half*>(&raw_val); });
1641
            return create_literal(shape::half_type, dims, data_half);
Khalique's avatar
Khalique committed
1642
        }
Paul's avatar
Paul committed
1643
        case onnx::TensorProto::DOUBLE:
Khalique's avatar
Khalique committed
1644
            return create_literal(shape::double_type, dims, t.double_data());
Paul's avatar
Paul committed
1645
1646
1647
1648
1649
        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
1650
        MIGRAPHX_THROW("Invalid tensor type");
Paul's avatar
Paul committed
1651
1652
    }

Khalique's avatar
Khalique committed
1653
    static literal
1654
    create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, const char* data)
1655
    {
Khalique's avatar
Khalique committed
1656
        // in case of scalar constants in onnx file, use dims=1 to fill initializer data
1657
        if(dims.empty())
1658
            return literal{{shape_type}, data};
1659
1660
1661
        return literal{{shape_type, dims}, data};
    }

1662
    template <class T, MIGRAPHX_REQUIRES(not std::is_pointer<T>{})>
Khalique's avatar
Khalique committed
1663
    static literal create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, T data)
1664
1665
    {
        if(dims.empty())
1666
            return literal{{shape_type}, data.begin(), data.end()};
1667
        return literal{{shape_type, dims}, data.begin(), data.end()};
1668
1669
    }

Paul's avatar
Paul committed
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
    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
1689
        case onnx::TensorProto::FLOAT16: shape_type = shape::half_type; break;
Paul's avatar
Paul committed
1690
1691
1692
1693
1694
1695
1696
1697
1698
        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
1699
        auto&& tensor_dims = t.tensor_type().shape().dim();
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
        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
1711
1712
        return {shape_type, dims};
    }
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734

    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
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
};

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
1758
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1759
} // namespace migraphx