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

Khalique's avatar
Khalique committed
57
58
59
60
        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
61
        add_binary_op("Pow", op::pow{});
Khalique's avatar
Khalique committed
62

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

67
68
        add_mem_op("ArgMax", &onnx_parser::parse_argmax);
        add_mem_op("ArgMin", &onnx_parser::parse_argmin);
69
        add_mem_op("Cast", &onnx_parser::parse_cast);
Khalique's avatar
Khalique committed
70
        add_mem_op("Clip", &onnx_parser::parse_clip);
Khalique's avatar
Khalique committed
71
        add_mem_op("LRN", &onnx_parser::parse_lrn);
Khalique's avatar
Khalique committed
72
        add_mem_op("ImageScaler", &onnx_parser::parse_imagescaler);
73
        add_mem_op("LeakyRelu", &onnx_parser::parse_leaky_relu);
Khalique's avatar
Khalique committed
74
        add_mem_op("Elu", &onnx_parser::parse_elu);
75
        add_mem_op("Expand", &onnx_parser::parse_expand);
Paul's avatar
Paul committed
76
77
        add_mem_op("Constant", &onnx_parser::parse_constant);
        add_mem_op("Conv", &onnx_parser::parse_conv);
Paul's avatar
Paul committed
78
79
        add_mem_op("MaxPool", &onnx_parser::parse_pooling);
        add_mem_op("AveragePool", &onnx_parser::parse_pooling);
80
81
        add_mem_op("GlobalMaxPool", &onnx_parser::parse_pooling);
        add_mem_op("GlobalAveragePool", &onnx_parser::parse_pooling);
Paul's avatar
Paul committed
82
        add_mem_op("Reshape", &onnx_parser::parse_reshape);
Paul's avatar
Paul committed
83
84
        add_mem_op("Flatten", &onnx_parser::parse_flatten);
        add_mem_op("Gemm", &onnx_parser::parse_gemm);
85
        add_mem_op("MatMul", &onnx_parser::parse_matmul);
86
        add_mem_op("BatchNormalization", &onnx_parser::parse_batchnorm);
Paul's avatar
Paul committed
87
        add_mem_op("Softmax", &onnx_parser::parse_softmax);
Shucai Xiao's avatar
Shucai Xiao committed
88
        add_mem_op("LogSoftmax", &onnx_parser::parse_logsoftmax);
89
90
91
        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
92
        add_mem_op("Concat", &onnx_parser::parse_concat);
93
94
95
        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);
96
        add_mem_op("ConstantOfShape", &onnx_parser::parse_constant_of_shape);
Khalique's avatar
Khalique committed
97
        add_mem_op("Transpose", &onnx_parser::parse_transpose);
Shucai Xiao's avatar
Shucai Xiao committed
98
        add_mem_op("RNN", &onnx_parser::parse_rnn);
99
        add_mem_op("GRU", &onnx_parser::parse_gru);
Shucai Xiao's avatar
Shucai Xiao committed
100
        add_mem_op("LSTM", &onnx_parser::parse_lstm);
Khalique's avatar
Khalique committed
101
        add_mem_op("Pad", &onnx_parser::parse_pad);
102
        add_mem_op("ReduceSum", &onnx_parser::parse_reduce_sum);
103
104
105
106
107
108
109

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

    void init_actv_func()
    {
110
111
112
113
114
115
        // 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
116
117
118
119
    }

    template <class F>
    void add_op(std::string name, F f)
Paul's avatar
Paul committed
120
121
122
123
124
125
126
127
128
    {
        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
129
130
131
132
133
134
135
    {
        ops.emplace(name, f);
    }

    template <class F>
    void add_mem_op(std::string name, F f)
    {
Paul's avatar
Paul committed
136
        add_op(name, [=](auto&&... xs) {
Paul's avatar
Paul committed
137
138
139
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }
Khalique's avatar
Khalique committed
140

141
    template <class T>
Khalique's avatar
Khalique committed
142
    void add_binary_op(std::string name, T x)
143
    {
Paul's avatar
Paul committed
144
        add_op(name, [this, x](attribute_map attributes, std::vector<instruction_ref> args) {
Scott Thornton's avatar
Scott Thornton committed
145
            if(args.size() != 2)
Paul's avatar
Paul committed
146
                MIGRAPHX_THROW("binary operators should have 2 operands");
147
            if(contains(attributes, "broadcast") and contains(attributes, "axis"))
148
149
150
151
            {
                uint64_t broadcasted = parse_value(attributes.at("broadcast")).at<uint64_t>();
                if(broadcasted != 0)
                {
152
                    uint64_t axis = parse_value(attributes.at("axis")).at<uint64_t>();
Shucai Xiao's avatar
Shucai Xiao committed
153
154
                    auto l = prog.add_instruction(op::broadcast{axis, args[0]->get_shape().lens()},
                                                  args[1]);
155
156
                    return prog.add_instruction(x, args[0], l);
                }
157
                return prog.add_instruction(x, args);
158
            }
Paul's avatar
Paul committed
159
            else
160
            {
Khalique's avatar
Khalique committed
161
                return add_broadcastable_binary_op(args[0], args[1], x);
162
163
164
165
            }
        });
    }

Shucai Xiao's avatar
Shucai Xiao committed
166
167
    std::vector<std::size_t> compute_broadcasted_lens(std::vector<std::size_t> s0,
                                                      std::vector<std::size_t> s1)
168
169
170
171
172
173
174
175
176
177
178
179
180
    {
        // 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
181
        if(s0.size() > s1.size())
182
183
184
185
186
187
        {
            s0.swap(s1);
        }

        std::vector<std::size_t> out_lens(s1);
        auto offset = s1.size() - s0.size();
Shucai Xiao's avatar
Shucai Xiao committed
188
189
190
191
        std::transform(s0.begin(),
                       s0.end(),
                       s1.begin() + offset,
                       out_lens.begin() + offset,
192
                       [&](auto a, auto b) {
Shucai Xiao's avatar
Shucai Xiao committed
193
                           if(a != b and a != 1 and b != 1)
194
                           {
Shucai Xiao's avatar
Shucai Xiao committed
195
196
197
198
199
200
                               MIGRAPHX_THROW("COMPUTE_BROADCASTLEN: shape {" +
                                              to_string_range(s0) + "} and {" +
                                              to_string_range(s1) + "} mismatch!");
                           }
                           return std::max(a, b);
                       });
201
202
203
204

        return out_lens;
    }

Khalique's avatar
Khalique committed
205
206
207
    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
Khalique's avatar
Khalique committed
208
        if(arg0->get_shape().lens() != arg1->get_shape().lens())
Khalique's avatar
Khalique committed
209
210
        {
            // Get lengths for both arguments
Shucai Xiao's avatar
Shucai Xiao committed
211
212
            auto s0       = arg0->get_shape().lens();
            auto s1       = arg1->get_shape().lens();
213
            auto out_lens = compute_broadcasted_lens(s0, s1);
Shucai Xiao's avatar
Shucai Xiao committed
214
215
            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
216
217
218
219
220
221
            return prog.add_instruction(x, l0, l1);
        }
        else
        {
            return prog.add_instruction(x, {arg0, arg1});
        }
222
223
    }

Paul's avatar
Paul committed
224
    template <class T>
Paul's avatar
Paul committed
225
226
    void add_generic_op(std::string name, T x)
    {
Paul's avatar
Paul committed
227
        add_op(name, [this, x](const attribute_map&, std::vector<instruction_ref> args) {
Paul's avatar
Paul committed
228
229
230
231
            return prog.add_instruction(x, args);
        });
    }

Khalique's avatar
Khalique committed
232
    template <class T>
Khalique's avatar
Khalique committed
233
    void add_variadic_op(std::string name, T x)
Khalique's avatar
Khalique committed
234
    {
Paul's avatar
Paul committed
235
        add_op(name, [this, x](const attribute_map&, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
236
            return std::accumulate(std::next(args.begin()),
Khalique's avatar
Khalique committed
237
238
239
240
241
                                   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
242
        });
Khalique's avatar
Khalique committed
243
244
    }

Khalique's avatar
Khalique committed
245
246
247
    instruction_ref parse_clip(const std::string&,
                               const attribute_map& attributes,
                               std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
248
249
250
251
252
253
254
255
256
257
258
259
260
    {
        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
261
    instruction_ref
Paul's avatar
Paul committed
262
    parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
263
264
    {
        auto dims = args.front()->get_shape().lens();
Scott Thornton's avatar
Scott Thornton committed
265
266
        auto r =
            prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, args.front());
267
268
        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
269
270
    }

Shucai Xiao's avatar
Shucai Xiao committed
271
272
273
    instruction_ref parse_logsoftmax(const std::string&,
                                     const attribute_map& attributes,
                                     std::vector<instruction_ref> args)
Shucai Xiao's avatar
Shucai Xiao committed
274
275
276
277
278
279
280
281
282
283
    {
        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));
    }

284
    instruction_ref parse_argmax(const std::string&,
Shucai Xiao's avatar
Shucai Xiao committed
285
286
                                 const attribute_map& attributes,
                                 std::vector<instruction_ref> args)
287
    {
288
        int64_t axis = 0;
289
290
        if(contains(attributes, "axis"))
        {
291
            axis = static_cast<int64_t>(parse_value(attributes.at("axis")).at<int>());
292
293
        }

Shucai Xiao's avatar
Shucai Xiao committed
294
        int keep_dims = 1;
Shucai Xiao's avatar
Shucai Xiao committed
295
        if(contains(attributes, "keepdims"))
Shucai Xiao's avatar
Shucai Xiao committed
296
297
298
299
        {
            keep_dims = parse_value(attributes.at("keepdims")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
300
        if(keep_dims == 0)
301
302
        {
            auto ins = prog.add_instruction(op::argmax{axis}, std::move(args));
303
            return prog.add_instruction(op::squeeze{{axis}}, ins);
304
305
306
307
308
        }
        else
        {
            return prog.add_instruction(op::argmax{axis}, std::move(args));
        }
309
310
311
    }

    instruction_ref parse_argmin(const std::string&,
Shucai Xiao's avatar
Shucai Xiao committed
312
313
                                 const attribute_map& attributes,
                                 std::vector<instruction_ref> args)
314
    {
315
        int64_t axis = 0;
316
317
        if(contains(attributes, "axis"))
        {
318
            axis = static_cast<int64_t>(parse_value(attributes.at("axis")).at<int>());
319
320
        }

Shucai Xiao's avatar
Shucai Xiao committed
321
        int keep_dims = 1;
Shucai Xiao's avatar
Shucai Xiao committed
322
        if(contains(attributes, "keepdims"))
Shucai Xiao's avatar
Shucai Xiao committed
323
324
325
326
        {
            keep_dims = parse_value(attributes.at("keepdims")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
327
        if(keep_dims == 0)
328
329
        {
            auto ins = prog.add_instruction(op::argmin{axis}, std::move(args));
330
            return prog.add_instruction(op::squeeze{{axis}}, ins);
331
332
333
334
335
        }
        else
        {
            return prog.add_instruction(op::argmin{axis}, std::move(args));
        }
336
337
    }

Paul's avatar
Paul committed
338
    instruction_ref
Paul's avatar
Paul committed
339
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
340
    {
341
        op::convolution op;
342
        auto l0 = args[0];
Paul's avatar
Paul committed
343
344
        if(contains(attributes, "pads"))
        {
Scott Thornton's avatar
Scott Thornton committed
345
            if(contains(attributes, "auto_pad"))
346
            {
Paul's avatar
Paul committed
347
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
348
            }
349
350
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
351
            if(padding.size() != 4)
352
            {
Paul's avatar
Paul committed
353
                MIGRAPHX_THROW("padding should have 4 values");
354
            }
Scott Thornton's avatar
Scott Thornton committed
355
            if(padding[0] != padding[2] || padding[1] != padding[3])
356
            {
357
358
                // 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
359
                l0      = prog.add_instruction(op::pad{padding}, l0);
360
            }
361
362
363
364
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
365
            }
Paul's avatar
Paul committed
366
        }
Paul's avatar
Paul committed
367
368
369
370
371
372
373
374
        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
375
        if(contains(attributes, "auto_pad"))
376
377
        {
            auto s = attributes["auto_pad"].s();
Scott Thornton's avatar
Scott Thornton committed
378
            if(contains(attributes, "pads") and to_upper(s) != "NOTSET")
379
            {
Paul's avatar
Paul committed
380
                MIGRAPHX_THROW("auto_pad and padding cannot be specified simultaneously");
381
382
            }

wsttiger's avatar
fixes  
wsttiger committed
383
            if(s.find("SAME") != std::string::npos)
384
            {
385
                op.padding_mode = op::padding_mode_t::same;
386
387
            }
        }
Khalique's avatar
Khalique committed
388
389
390
391
        if(contains(attributes, "group"))
        {
            op.group = parse_value(attributes.at("group")).at<int>();
        }
Paul's avatar
Paul committed
392
393
394
395
        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
396
            auto l2 = prog.add_instruction(op::broadcast{axis, l1->get_shape().lens()}, args[2]);
397
            return prog.add_instruction(op::add{}, l1, l2);
Paul's avatar
Paul committed
398
        }
399
        return prog.add_instruction(op, l0, args[1]);
Paul's avatar
Paul committed
400
    }
Paul's avatar
Paul committed
401

Paul's avatar
Paul committed
402
403
404
    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
Paul's avatar
Paul committed
405
    {
Khalique's avatar
Khalique committed
406
        op::pooling op{ends_with(name, "MaxPool") ? "max" : "average"};
407
        auto l0 = args[0];
Khalique's avatar
Khalique committed
408
        if(starts_with(name, "Global"))
409
        {
Khalique's avatar
Khalique committed
410
411
            auto lens  = args.front()->get_shape().lens();
            op.lengths = {lens[2], lens[3]};
412
        }
Paul's avatar
Paul committed
413
414
        if(contains(attributes, "pads"))
        {
415
416
            std::vector<std::int64_t> padding;
            copy(attributes["pads"].ints(), std::back_inserter(padding));
Scott Thornton's avatar
Scott Thornton committed
417
            if(padding.size() != 4)
418
            {
Paul's avatar
Paul committed
419
                MIGRAPHX_THROW("padding should have 4 values");
420
            }
Scott Thornton's avatar
Scott Thornton committed
421
            if(padding[0] != padding[2] || padding[1] != padding[3])
422
            {
423
424
                // 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
425
426
                l0 = prog.add_instruction(op::pad{padding, std::numeric_limits<float>::lowest()},
                                          l0);
427
428
429
430
431
            }
            else
            {
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
432
            }
Paul's avatar
Paul committed
433
434
435
436
437
438
439
440
441
        }
        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
442
        if(contains(attributes, "auto_pad"))
443
444
        {
            auto s = attributes["auto_pad"].s();
445
            if(s.find("SAME_UPPER") == std::string::npos)
446
            {
447
                MIGRAPHX_THROW("auto_pad only supports SAME_UPPER for pooling");
448
            }
449
            op.padding_mode = op::padding_mode_t::same;
450
451
        }

452
        return prog.add_instruction(op, l0);
Paul's avatar
Paul committed
453
454
    }

Paul's avatar
Paul committed
455
    instruction_ref
Paul's avatar
Paul committed
456
    parse_reshape(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
457
    {
458
        op::reshape op;
Paul's avatar
Paul committed
459
460
461
462
463
464
465
        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
466
            auto s = args[1]->eval();
Paul's avatar
Paul committed
467
            if(s.empty())
Paul's avatar
Paul committed
468
                MIGRAPHX_THROW("Dynamic shape is not supported.");
Paul's avatar
Paul committed
469
            s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
Paul's avatar
Paul committed
470
        }
Paul's avatar
Paul committed
471
472
473
        return prog.add_instruction(op, args[0]);
    }

Paul's avatar
Paul committed
474
    instruction_ref
Paul's avatar
Paul committed
475
    parse_flatten(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
476
    {
477
        uint64_t axis = 1;
Paul's avatar
Paul committed
478
479
480
481
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
482
        return prog.add_instruction(op::flatten{axis}, args[0]);
Paul's avatar
Paul committed
483
484
    }

485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
    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
503
504
505
506
507
508
509
    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));
    }
510

511
512
513
    instruction_ref
    parse_gather(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
514
        int axis = 0;
515
516
517
518
        if(contains(attributes, "axis"))
        {
            axis = parse_value(attributes.at("axis")).at<int>();
        }
519
        op::gather op{axis};
520
521
522
        return prog.add_instruction(op, std::move(args));
    }

523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
    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
543
544
545
    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
Paul's avatar
Paul committed
546
    {
Shucai Xiao's avatar
Shucai Xiao committed
547
        literal v = parse_value(attributes.at("value"));
548
        // return empty literal
Shucai Xiao's avatar
Shucai Xiao committed
549
        if(v.get_shape().elements() == 0)
550
551
552
553
        {
            return prog.add_literal(literal{});
        }

554
555
556
        auto dim_size = attributes.at("value").t().dims_size();
        // if dim_size is 0, it is a scalar
        if(dim_size == 0)
557
        {
558
            migraphx::shape scalar_shape{v.get_shape().type()};
559
560
561
            return prog.add_literal(migraphx::literal{scalar_shape, v.data()});
        }

Paul's avatar
Paul committed
562
563
        return prog.add_literal(v);
    }
Paul's avatar
Paul committed
564

Paul's avatar
Paul committed
565
    instruction_ref
Paul's avatar
Paul committed
566
    parse_gemm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Paul's avatar
Paul committed
567
568
    {
        float alpha = 1.0f;
Khalique's avatar
Khalique committed
569
        float beta  = 1.0f;
Paul's avatar
Paul committed
570
571
572
573
574
575
576
577
        bool transa = false;
        bool transb = false;
        if(contains(attributes, "alpha"))
        {
            alpha = parse_value(attributes.at("alpha")).at<float>();
        }
        if(contains(attributes, "beta"))
        {
578
            beta = parse_value(attributes.at("beta")).at<float>();
Paul's avatar
Paul committed
579
580
581
582
583
584
585
586
587
        }
        if(contains(attributes, "transA"))
        {
            transa = parse_value(attributes.at("transA")).at<bool>();
        }
        if(contains(attributes, "transB"))
        {
            transb = parse_value(attributes.at("transB")).at<bool>();
        }
588
589
590
591
592
593

        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));

594
595
        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
596
597
        if(args.size() == 3)
        {
598
            if(beta != 0.f && args[2]->get_shape().elements() > 0)
599
            {
Shucai Xiao's avatar
Shucai Xiao committed
600
                auto out_lens   = l1->get_shape().lens();
601
                out_lens.back() = l2->get_shape().lens().back();
Shucai Xiao's avatar
Shucai Xiao committed
602
                auto l3         = args[2];
Shucai Xiao's avatar
Shucai Xiao committed
603
604
                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
605
                {
606
                    l3 = prog.add_instruction(op::multibroadcast{out_lens}, args[2]);
Khalique's avatar
Khalique committed
607
                }
608
                return prog.add_instruction(op::dot{alpha, beta}, l1, l2, l3);
609
            }
Paul's avatar
Paul committed
610
        }
611
612

        return prog.add_instruction(op::dot{alpha, beta}, l1, l2);
Paul's avatar
Paul committed
613
614
    }

615
    instruction_ref
Shucai Xiao's avatar
Shucai Xiao committed
616
    parse_matmul(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
617
    {
Shucai Xiao's avatar
Shucai Xiao committed
618
619
        auto l0      = args[0];
        auto l1      = args[1];
620
621
622
623
624
        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
625
        if(l0_lens.size() == 1)
626
627
628
629
630
631
632
        {
            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
633
        if(l1_lens.size() == 1)
634
635
636
637
638
639
640
641
        {
            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
642
        if(!std::equal(l0_lens.rbegin() + 2, l0_lens.rend(), l1_lens.rbegin() + 2, l1_lens.rend()))
643
644
645
646
647
648
        {
            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);
649
            l0_broadcasted_lens = output_lens;
650
            l0_broadcasted_lens.insert(l0_broadcasted_lens.end(), l0_it, l0_lens.end());
651
            l1_broadcasted_lens = output_lens;
652
            l1_broadcasted_lens.insert(l1_broadcasted_lens.end(), l1_it, l1_lens.end());
Shucai Xiao's avatar
Shucai Xiao committed
653
            if(l0_lens != l0_broadcasted_lens)
654
655
656
            {
                bl0 = prog.add_instruction(op::multibroadcast{l0_broadcasted_lens}, l0);
            }
Shucai Xiao's avatar
Shucai Xiao committed
657
            if(l1_lens != l1_broadcasted_lens)
658
659
660
661
662
            {
                bl1 = prog.add_instruction(op::multibroadcast{l1_broadcasted_lens}, l1);
            }
        }

Shucai Xiao's avatar
Shucai Xiao committed
663
        auto dot_res     = prog.add_instruction(op::dot{1.0f, 0.0f}, bl0, bl1);
664
        int64_t num_axis = static_cast<int64_t>(dot_res->get_shape().lens().size());
Shucai Xiao's avatar
Shucai Xiao committed
665
        if(is_a_prepended)
666
667
668
669
        {
            dot_res = prog.add_instruction(op::squeeze{{num_axis - 2}}, dot_res);
            --num_axis;
        }
Shucai Xiao's avatar
Shucai Xiao committed
670
        if(is_b_appended)
671
672
673
        {
            dot_res = prog.add_instruction(op::squeeze{{num_axis - 1}}, dot_res);
        }
Shucai Xiao's avatar
Shucai Xiao committed
674

675
676
677
        return dot_res;
    }

678
    instruction_ref
Paul's avatar
Paul committed
679
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
680
    {
Scott Thornton's avatar
Scott Thornton committed
681
682
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
683
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
Scott Thornton's avatar
Scott Thornton committed
684
        bool is_test                                      = false;
685
686
687
688
689
690
        if(contains(attributes, "epsilon"))
        {
            epsilon = parse_value(attributes.at("epsilon")).at<float>();
        }
        if(contains(attributes, "momentum"))
        {
691
            momentum = parse_value(attributes.at("momentum")).at<float>();
692
693
694
        }
        if(contains(attributes, "is_test"))
        {
wsttiger's avatar
wsttiger committed
695
            is_test = parse_value(attributes.at("is_test")).at<uint64_t>() > 0;
696
697
698
        }
        if(contains(attributes, "spatial"))
        {
699
            bn_mode = (parse_value(attributes.at("spatial")).at<uint64_t>() > 0)
700
701
                          ? op::batch_norm_inference::spatial
                          : op::batch_norm_inference::per_activation;
702
        }
Paul's avatar
Paul committed
703
        (void)is_test;
Paul's avatar
Paul committed
704
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
Paul's avatar
Paul committed
705
        return prog.add_instruction(op, std::move(args));
706
707
    }

708
709
710
711
    instruction_ref parse_leaky_relu(const std::string&,
                                     attribute_map attributes,
                                     std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
712
        float alpha = 0.01; // default alpha val for leaky relu
713
714
715
716
717
718
719
720
        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
721
722
    instruction_ref
    parse_elu(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
723
724
725
726
727
728
729
730
731
732
    {
        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
733
734
    instruction_ref
    parse_lrn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
735
736
    {
        float alpha = 0.0001;
Khalique's avatar
Khalique committed
737
738
739
        float beta  = 0.75;
        float bias  = 1.0;
        int size    = 1;
Khalique's avatar
Khalique committed
740
741
742
743
744
745
746
747
748
749
750
751
        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
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
    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());
        }
768
        auto input_lens = args.front()->get_shape().lens();
Khalique's avatar
Khalique committed
769

Khalique's avatar
Khalique committed
770
771
        auto scale_val = prog.add_literal(scale);
        auto bias_vals = prog.add_literal(
Paul's avatar
Paul committed
772
            migraphx::literal{migraphx::shape{migraphx::shape::float_type, {bias.size()}}, bias});
Khalique's avatar
Khalique committed
773

774
        auto scale_tensor = prog.add_instruction(migraphx::op::scalar{input_lens}, scale_val);
Paul's avatar
Paul committed
775
        auto img_scaled   = prog.add_instruction(migraphx::op::mul{}, args.front(), scale_tensor);
Shucai Xiao's avatar
Shucai Xiao committed
776
        auto bias_bcast   = prog.add_instruction(migraphx::op::broadcast{1, input_lens}, bias_vals);
Paul's avatar
Paul committed
777
        return prog.add_instruction(migraphx::op::add{}, img_scaled, bias_bcast);
Khalique's avatar
Khalique committed
778
    }
Khalique's avatar
Khalique committed
779

Khalique's avatar
Khalique committed
780
781
    instruction_ref
    parse_transpose(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
782
783
784
785
786
787
788
    {
        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
789
        return prog.add_instruction(migraphx::op::transpose{perm}, args.front());
Khalique's avatar
Khalique committed
790
791
    }

Khalique's avatar
Khalique committed
792
793
794
795
796
797
798
799
800
801
    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());
        }
802
        // check if padding is actually being done (at least one value is nonzero)
Khalique's avatar
Khalique committed
803
        if(std::all_of(pads.begin(), pads.end(), [](const int& i) { return i == 0; }))
804
805
806
        {
            return prog.add_instruction(migraphx::op::identity{}, args.front());
        }
Khalique's avatar
Khalique committed
807
808
809
810
811
812
813
814
815
816
817
818
        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());
    }
819
820
821
    // 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
822
    parse_shape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
823
824
    {
        if(args.size() != 1)
825
            MIGRAPHX_THROW("Shape: operator should have 1 operand");
826
827
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
        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
862
863
        if(contains(attributes, "extra_shape"))
        {
864
            MIGRAPHX_THROW("ConstantFill: cannot handle extra shape attribute");
865
866
        }

867
868
        if(input_as_shape == 1)
        {
Shucai Xiao's avatar
Shucai Xiao committed
869
            if(args.size() != 1)
870
            {
871
                MIGRAPHX_THROW("ConstantFill: need an input argument as output shape");
872
873
            }

Shucai Xiao's avatar
Shucai Xiao committed
874
875
            if(contains(attributes, "shape"))
            {
876
                MIGRAPHX_THROW("ConstantFill: cannot set the shape argument and pass in an input "
Shucai Xiao's avatar
Shucai Xiao committed
877
                               "at the same time");
878
879
            }

880
881
882
            migraphx::argument in = args[0]->eval();
            if(in.empty())
            {
883
                MIGRAPHX_THROW("ConstantFill: cannot handle dynamic shape as input");
884
            }
885

886
887
888
            std::vector<std::size_t> dims;
            in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
            migraphx::shape s(type, dims);
889
890
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
891
892
893
        }
        else if(input_as_shape == 0)
        {
Shucai Xiao's avatar
Shucai Xiao committed
894
895
            if(!contains(attributes, "shape"))
            {
896
                MIGRAPHX_THROW("ConstantFill: attribute output shape is needed");
897
898
899
            }

            literal ls = parse_value(attributes.at("shape"));
900
            std::vector<std::size_t> dims;
Shucai Xiao's avatar
Shucai Xiao committed
901
            ls.visit([&](auto s) { dims.assign(s.begin(), s.end()); });
902
            migraphx::shape s{type, dims};
903
904
            std::vector<float> values(s.elements(), value);
            return prog.add_literal(migraphx::literal(s, values));
905
906
907
        }
        else
        {
908
            MIGRAPHX_THROW("ConstantFill: wrong value of attribute input_as_shape");
909
910
911
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
912
913
914
    instruction_ref parse_constant_of_shape(const std::string&,
                                            attribute_map attributes,
                                            std::vector<instruction_ref> args)
915
916
    {
        literal l_val{};
Shucai Xiao's avatar
Shucai Xiao committed
917
        if(contains(attributes, "value"))
918
919
        {
            l_val = parse_value(attributes.at("value"));
Shucai Xiao's avatar
Shucai Xiao committed
920
            if(l_val.get_shape().elements() != 1)
921
922
923
924
925
926
927
928
929
930
931
            {
                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();
932

Shucai Xiao's avatar
Shucai Xiao committed
933
        if(args.empty())
934
        {
935
            MIGRAPHX_THROW("Parse ConstantOfShape : must have 1 input!");
936
937
938
        }
        else
        {
939
940
            migraphx::shape s;
            // empty input tensor, output is a scalar
Shucai Xiao's avatar
Shucai Xiao committed
941
            if(args[0]->get_shape().elements() == 0)
942
            {
943
                s = migraphx::shape{type, {1}, {0}};
944
            }
945
946
947
948
949
950
951
            else
            {
                migraphx::argument in = args[0]->eval();
                if(in.empty())
                {
                    MIGRAPHX_THROW("Parse ConstantOfShape: cannot handle dynamic shape as input");
                }
952

953
954
955
956
                std::vector<std::size_t> dims;
                in.visit([&](auto input) { dims.assign(input.begin(), input.end()); });
                s = migraphx::shape{type, dims};
            }
957
958
959
960
961
962
963
964
965
966
967
968
969

            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
970
    instruction_ref
Shucai Xiao's avatar
Shucai Xiao committed
971
    parse_expand(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
972
    {
Shucai Xiao's avatar
Shucai Xiao committed
973
        auto in_lens             = args[0]->get_shape().lens();
974
        migraphx::argument arg_s = args[1]->eval();
Shucai Xiao's avatar
Shucai Xiao committed
975
        if(arg_s.empty())
976
977
978
979
980
981
        {
            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);
982

Shucai Xiao's avatar
Shucai Xiao committed
983
        return prog.add_instruction(op::multibroadcast{out_lens}, args[0]);
984
985
    }

Shucai Xiao's avatar
Shucai Xiao committed
986
    std::vector<instruction_ref>
Shucai Xiao's avatar
Shucai Xiao committed
987
988
989
    parse_rnn(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        migraphx::shape input_shape = args[0]->get_shape();
990
        std::size_t hidden_size     = args[1]->get_shape().lens()[1];
Shucai Xiao's avatar
Shucai Xiao committed
991
992
993

        if(contains(attributes, "hidden_size"))
        {
Shucai Xiao's avatar
Shucai Xiao committed
994
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
995
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
996
997
998
            {
                MIGRAPHX_THROW("RNN: hidden size mismatch in input and attribute");
            }
Shucai Xiao's avatar
Shucai Xiao committed
999
1000
1001
1002
1003
1004
1005
1006
1007
        }

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

1008
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
1009
1010
        if(direction == "bidirectional")
        {
1011
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
1012
1013
1014
        }
        else if(direction == "reverse")
        {
1015
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
1016
1017
        }

1018
        std::vector<std::string> vec_names{"tanh"};
1019
1020
1021
1022
        if(contains(attributes, "activations"))
        {
            auto names = attributes.at("activations").strings();
            vec_names.clear();
1023
            vec_names.resize(names.size());
Shucai Xiao's avatar
Shucai Xiao committed
1024
1025
1026
            std::transform(names.begin(), names.end(), vec_names.begin(), [](auto name) {
                return to_lower(name);
            });
1027
1028
        }

1029
1030
1031
        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
1032
        if(name_it != vec_names.end())
1033
1034
1035
        {
            MIGRAPHX_THROW("RNN: activation function " + std::string(*name_it) + " not supported");
        }
1036

Shucai Xiao's avatar
Shucai Xiao committed
1037
        // bidirectional case should have two activation functions.
Shucai Xiao's avatar
Shucai Xiao committed
1038
        // one is for forward, and the other is for reverse.
Shucai Xiao's avatar
Shucai Xiao committed
1039
        // if only one actv function is provided, we use it in both
1040
        // forward and reverse direction
1041
        if(dirct == op::rnn_direction::bidirectional)
1042
        {
Shucai Xiao's avatar
Shucai Xiao committed
1043
            if(vec_names.size() == 1)
1044
1045
1046
1047
1048
            {
                vec_names.push_back(vec_names.at(0));
            }
        }

Shucai Xiao's avatar
Shucai Xiao committed
1049
1050
1051
        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];
1052
        });
Shucai Xiao's avatar
Shucai Xiao committed
1053

Shucai Xiao's avatar
Shucai Xiao committed
1054
1055
1056
1057
1058
1059
1060
        // To be added later
        float clip = 0.0;
        if(contains(attributes, "clip"))
        {
            clip = parse_value(attributes.at("clip")).at<float>();
        }

1061
1062
        // if the number of arguments is less than 6, append
        // undefined operator to have 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
1063
        if(args.size() < 6)
1064
1065
1066
1067
1068
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), (6 - args.size()), ins);
        }

Shucai Xiao's avatar
Shucai Xiao committed
1069
1070
        // 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
1071
                                                  std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
1072

1073
        // second output for the last hidden state
Shucai Xiao's avatar
Shucai Xiao committed
1074
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
1075

Shucai Xiao's avatar
Shucai Xiao committed
1076
        return {hidden_states, last_output};
Shucai Xiao's avatar
Shucai Xiao committed
1077
1078
    }

1079
    std::vector<instruction_ref>
1080
1081
1082
1083
1084
1085
1086
    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
1087
            std::size_t hidden_size_att = parse_value(attributes.at("hidden_size")).at<int>();
Shucai Xiao's avatar
Shucai Xiao committed
1088
            if(hidden_size != hidden_size_att)
Shucai Xiao's avatar
Shucai Xiao committed
1089
1090
1091
            {
                MIGRAPHX_THROW("GRU: hidden size mismatch in input and attribute");
            }
1092
1093
1094
1095
1096
1097
1098
1099
1100
        }

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

1101
        op::rnn_direction dirct = op::rnn_direction::forward;
1102
1103
        if(direction == "bidirectional")
        {
1104
            dirct = op::rnn_direction::bidirectional;
1105
1106
1107
        }
        else if(direction == "reverse")
        {
1108
            dirct = op::rnn_direction::reverse;
1109
1110
        }

1111
        std::vector<std::string> vec_names = {"sigmoid", "tanh"};
1112
1113
        if(contains(attributes, "activations"))
        {
1114
            auto names = attributes.at("activations").strings();
1115
            vec_names.clear();
Shucai Xiao's avatar
Shucai Xiao committed
1116
            vec_names.resize(names.size());
Shucai Xiao's avatar
Shucai Xiao committed
1117
1118
1119
            std::transform(names.begin(), names.end(), vec_names.begin(), [](auto name) {
                return to_lower(name);
            });
1120
1121
        }

1122
        // need 4 activation functions
1123
        if(dirct == op::rnn_direction::bidirectional)
1124
        {
Shucai Xiao's avatar
Shucai Xiao committed
1125
            // 4 activation functions are used in the bidirectional
1126
            // scenario. No spec is provided in onnx::operator. we
Shucai Xiao's avatar
Shucai Xiao committed
1127
1128
            // use the algorithm that: if 1 actv function is provided,
            // repeat 1 four times. If 2 actv functins are provided,
1129
1130
            // 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
1131
1132
1133
            // assume the 3rd one is repeated once and used by the
            // reverse direction.
            // This may need change later
1134
            if(vec_names.size() == 1)
1135
            {
1136
                vec_names.insert(vec_names.end(), 3, vec_names.at(0));
1137
            }
1138
            else if(vec_names.size() == 2)
1139
            {
1140
1141
1142
                // repeat the activation functions
                vec_names.push_back(vec_names.at(0));
                vec_names.push_back(vec_names.at(1));
1143
            }
1144
            else if(vec_names.size() == 3)
1145
            {
1146
                vec_names.push_back(vec_names.at(2));
1147
1148
            }
        }
Shucai Xiao's avatar
Shucai Xiao committed
1149
        else
1150
        {
1151
            if(vec_names.size() == 1)
1152
            {
1153
                vec_names.push_back(vec_names.at(0));
1154
1155
1156
            }
        }

1157
1158
1159
        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
1160
        if(name_it != vec_names.end())
1161
1162
1163
        {
            MIGRAPHX_THROW("GRU: activation function " + std::string(*name_it) + " not supported");
        }
1164

Shucai Xiao's avatar
Shucai Xiao committed
1165
1166
1167
        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
1168
        });
1169
1170
1171
1172
1173
1174
1175
1176

        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
1177
        if(contains(attributes, "linear_before_reset"))
1178
1179
1180
1181
        {
            linear_before_reset = parse_value(attributes.at("linear_before_reset")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
1182
        // append undefined opeator to make 6 arguments
Shucai Xiao's avatar
Shucai Xiao committed
1183
        if(args.size() < 6)
Shucai Xiao's avatar
Shucai Xiao committed
1184
1185
1186
1187
1188
        {
            auto ins = prog.add_instruction(op::undefined{});
            args.insert(args.end(), 6 - args.size(), ins);
        }

1189
1190
        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
1191
            op::gru{hidden_size, vec_actv_funcs, dirct, clip, linear_before_reset},
Shucai Xiao's avatar
Shucai Xiao committed
1192
            std::move(args));
1193
1194

        // second output for last gru output
1195
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
1196

Shucai Xiao's avatar
Shucai Xiao committed
1197
        return {hidden_states, last_output};
1198
1199
    }

Shucai Xiao's avatar
Shucai Xiao committed
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
    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
1222
        op::rnn_direction dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
1223
1224
        if(direction == "bidirectional")
        {
Shucai Xiao's avatar
Shucai Xiao committed
1225
            dirct = op::rnn_direction::bidirectional;
Shucai Xiao's avatar
Shucai Xiao committed
1226
1227
1228
        }
        else if(direction == "reverse")
        {
Shucai Xiao's avatar
Shucai Xiao committed
1229
            dirct = op::rnn_direction::reverse;
Shucai Xiao's avatar
Shucai Xiao committed
1230
        }
Shucai Xiao's avatar
Shucai Xiao committed
1231
        else if(direction == "forward")
Shucai Xiao's avatar
Shucai Xiao committed
1232
        {
Shucai Xiao's avatar
Shucai Xiao committed
1233
            dirct = op::rnn_direction::forward;
Shucai Xiao's avatar
Shucai Xiao committed
1234
1235
1236
1237
1238
1239
        }
        else
        {
            MIGRAPHX_THROW("LSTM: incorrect direction attribute");
        }

1240
        std::vector<std::string> vec_names = {"sigmoid", "tanh", "tanh"};
Shucai Xiao's avatar
Shucai Xiao committed
1241
1242
1243
1244
1245
        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
1246
1247
1248
            std::transform(names.begin(), names.end(), vec_names.begin(), [](auto name) {
                return to_lower(name);
            });
Shucai Xiao's avatar
Shucai Xiao committed
1249
1250
1251
        }

        // need 6 activation functions for bidirectional directions
Shucai Xiao's avatar
Shucai Xiao committed
1252
        if(dirct == op::rnn_direction::bidirectional)
Shucai Xiao's avatar
Shucai Xiao committed
1253
1254
1255
1256
1257
1258
        {
            // 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
1259
            // if 3 actv funcs are provide, repeat all three once.
Shucai Xiao's avatar
Shucai Xiao committed
1260
1261
1262
1263
            // the same algorithm is used for 4, 5, and 6 actv funcions
            // provided. This may need change later
            switch(vec_names.size())
            {
1264
            case 1:
Shucai Xiao's avatar
Shucai Xiao committed
1265
1266
1267
1268
1269
1270
                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)};
1271
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1272
1273
1274

            case 2:
                // repeat the 2nd actv func once, then repeat all three another time
Shucai Xiao's avatar
Shucai Xiao committed
1275
1276
1277
1278
1279
1280
                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
1281
1282
1283
1284
                break;

            case 3:
                // repeat all three actv funcs once
Shucai Xiao's avatar
Shucai Xiao committed
1285
1286
1287
1288
1289
1290
                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
1291
1292
                break;

Shucai Xiao's avatar
Shucai Xiao committed
1293
1294
1295
1296
1297
1298
1299
            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)};
1300
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1301

Shucai Xiao's avatar
Shucai Xiao committed
1302
1303
1304
1305
1306
1307
1308
            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)};
1309
                break;
Shucai Xiao's avatar
Shucai Xiao committed
1310

Shucai Xiao's avatar
Shucai Xiao committed
1311
            default: break;
Shucai Xiao's avatar
Shucai Xiao committed
1312
1313
1314
1315
1316
1317
            }
        }
        else
        {
            switch(vec_names.size())
            {
Shucai Xiao's avatar
Shucai Xiao committed
1318
            case 1: vec_names = {vec_names.at(0), vec_names.at(0), vec_names.at(0)}; break;
Shucai Xiao's avatar
Shucai Xiao committed
1319
1320
1321

            case 2:
                // repeat the 2nd actv func once, so we have 3 actv funcs
Shucai Xiao's avatar
Shucai Xiao committed
1322
                vec_names = {vec_names.at(0), vec_names.at(1), vec_names.at(1)};
Shucai Xiao's avatar
Shucai Xiao committed
1323
1324
                break;

Shucai Xiao's avatar
Shucai Xiao committed
1325
            default: break;
Shucai Xiao's avatar
Shucai Xiao committed
1326
1327
1328
            }
        }

1329
1330
1331
        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
1332
        if(name_it != vec_names.end())
1333
1334
1335
        {
            MIGRAPHX_THROW("LSTM: activation function " + std::string(*name_it) + " not supported");
        }
Shucai Xiao's avatar
Shucai Xiao committed
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357

        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
1358
            args.insert(args.end(), 8 - args.size(), ins);
Shucai Xiao's avatar
Shucai Xiao committed
1359
1360
1361
1362
        }

        // first output for concatenation of hidden states
        auto hidden_states = prog.add_instruction(
Shucai Xiao's avatar
Shucai Xiao committed
1363
            op::lstm{hidden_size, vec_actv_funcs, dirct, clip, input_forget}, std::move(args));
Shucai Xiao's avatar
Shucai Xiao committed
1364
1365

        // second output for last lstm output
Shucai Xiao's avatar
Shucai Xiao committed
1366
        auto last_output = prog.add_instruction(op::rnn_last_output{}, hidden_states);
Shucai Xiao's avatar
Shucai Xiao committed
1367
1368
1369
1370
1371
1372

        // 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};
    }
1373

Shucai Xiao's avatar
Shucai Xiao committed
1374
1375
1376
    instruction_ref parse_reduce_sum(const std::string&,
                                     attribute_map attributes,
                                     std::vector<instruction_ref> args)
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
    {
        std::size_t n_dim = args.front()->get_shape().lens().size();

        // default to reduce over all dimensions
        std::vector<std::size_t> axes(n_dim);
        std::iota(axes.begin(), axes.end(), 0);
        if(contains(attributes, "axes"))
        {
            axes.clear();
            auto&& attr_axes = attributes["axes"].ints();
Shucai Xiao's avatar
Shucai Xiao committed
1387
            axes             = std::vector<std::size_t>(attr_axes.begin(), attr_axes.end());
1388
1389
1390
        }

        int keep_dims = 1;
Shucai Xiao's avatar
Shucai Xiao committed
1391
        if(contains(attributes, "keepdims"))
1392
1393
1394
1395
        {
            keep_dims = parse_value(attributes.at("keepdims")).at<int>();
        }

Shucai Xiao's avatar
Shucai Xiao committed
1396
        if(keep_dims == 1)
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
        {
            return prog.add_instruction(op::reduce_sum{axes}, std::move(args));
        }
        else
        {
            auto ins = prog.add_instruction(op::reduce_sum{axes}, std::move(args));
            std::vector<int64_t> squeeze_axes{axes.begin(), axes.end()};
            return prog.add_instruction(op::squeeze{squeeze_axes}, ins);
        }
    }
1407

Shucai Xiao's avatar
Shucai Xiao committed
1408
1409
    instruction_ref
    parse_cast(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
1410
    {
Shucai Xiao's avatar
Shucai Xiao committed
1411
        if(!contains(attributes, "to"))
1412
1413
1414
1415
        {
            MIGRAPHX_THROW("PARSE_CAST: missing to type attribute!");
        }

Shucai Xiao's avatar
Shucai Xiao committed
1416
        int to_type        = parse_value(attributes.at("to")).at<int>();
1417
1418
1419
        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
1420

Paul's avatar
Paul committed
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
    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
1433
            MIGRAPHX_THROW("Failed reading onnx file.");
Paul's avatar
Paul committed
1434
1435
1436
1437
1438
1439
        }
    }

    void parse_graph(const onnx::GraphProto& graph)
    {
        nodes = get_nodes(graph);
1440
1441
1442
1443
1444
        std::unordered_map<std::string, onnx::TensorProto> initializer_data;
        for(auto&& f : graph.initializer())
        {
            initializer_data[f.name()] = f;
        }
Paul's avatar
Paul committed
1445
1446
1447
        for(auto&& input : graph.input())
        {
            const std::string& name = input.name();
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
            // 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
1460
        }
Paul's avatar
Paul committed
1461
        for(auto&& output : graph.output())
Paul's avatar
Paul committed
1462
        {
Paul's avatar
Paul committed
1463
            this->parse_node(output.name());
Paul's avatar
Paul committed
1464
1465
1466
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
1467
    void parse_undefined(const std::string& name)
1468
    {
Shucai Xiao's avatar
Shucai Xiao committed
1469
        auto ins           = prog.add_instruction(op::undefined{});
1470
1471
1472
        instructions[name] = ins;
    }

Paul's avatar
Paul committed
1473
    void parse_node(const std::string& name)
Paul's avatar
Paul committed
1474
    {
Paul's avatar
Paul committed
1475
        if(name.empty())
Paul's avatar
Paul committed
1476
            MIGRAPHX_THROW("Onnx node must have a name");
Paul's avatar
Paul committed
1477
1478
1479
1480
1481
1482
1483
1484
        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
1485
1486
                    assert(name != input);
                    this->parse_node(input);
Paul's avatar
Paul committed
1487
                }
Shucai Xiao's avatar
Shucai Xiao committed
1488
                else if(input.empty())
Paul's avatar
Paul committed
1489
                {
1490
                    this->parse_undefined(input);
Paul's avatar
Paul committed
1491
                }
1492
                args.push_back(instructions.at(input));
Paul's avatar
Paul committed
1493
            }
Paul's avatar
Paul committed
1494
            std::vector<instruction_ref> result;
Paul's avatar
Paul committed
1495
1496
            if(ops.count(node.op_type()) == 0)
            {
1497
                result.push_back(prog.add_instruction(op::unknown{node.op_type()}, args));
Paul's avatar
Paul committed
1498
1499
1500
            }
            else
            {
Paul's avatar
Paul committed
1501
                result = ops[node.op_type()](get_attributes(node), args);
Paul's avatar
Paul committed
1502
            }
Paul's avatar
Paul committed
1503
            // Even no output nodes produce output in migraphx
Paul's avatar
Paul committed
1504
            if(node.output().empty() and result.size() == 1)
Paul's avatar
Paul committed
1505
1506
            {
                instructions[name] = result.front();
Paul's avatar
Paul committed
1507
1508
1509
            }
            else
            {
Paul's avatar
Paul committed
1510
1511
1512
1513
1514
1515
                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
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
            }
        }
    }

    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
1533
        std::size_t n = 0;
Paul's avatar
Paul committed
1534
1535
        for(auto&& node : graph.node())
        {
Paul's avatar
Paul committed
1536
            if(node.output().empty())
Paul's avatar
Paul committed
1537
            {
Paul's avatar
Paul committed
1538
                if(node.name().empty())
Paul's avatar
Paul committed
1539
1540
1541
1542
1543
1544
1545
1546
1547
                {
                    result["migraphx_unamed_node_" + std::to_string(n)] = node;
                    n++;
                }
                else
                {
                    result[node.name()] = node;
                }
            }
Paul's avatar
Paul committed
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
            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
1573
        case onnx::AttributeProto::FLOATS: return from_repeated(shape::float_type, attr.floats());
Paul's avatar
Paul committed
1574
1575
1576
1577
1578
        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
1579
        MIGRAPHX_THROW("Invalid attribute type");
Paul's avatar
Paul committed
1580
1581
1582
1583
1584
    }

    static literal parse_tensor(const onnx::TensorProto& t)
    {
        std::vector<std::size_t> dims(t.dims().begin(), t.dims().end());
1585
1586
        if(t.has_raw_data())
        {
wsttiger's avatar
wsttiger committed
1587
            const std::string& s = t.raw_data();
Scott Thornton's avatar
Scott Thornton committed
1588
1589
1590
            switch(t.data_type())
            {
            case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
1591
            case onnx::TensorProto::FLOAT: return create_literal(shape::float_type, dims, s.data());
Scott Thornton's avatar
Scott Thornton committed
1592
            case onnx::TensorProto::UINT8: throw std::runtime_error("");
1593
            case onnx::TensorProto::INT8: return create_literal(shape::int32_type, dims, s.data());
Khalique's avatar
Khalique committed
1594
1595
            case onnx::TensorProto::UINT16:
                return create_literal(shape::int32_type, dims, s.data());
1596
1597
1598
            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
1599
            case onnx::TensorProto::STRING: throw std::runtime_error("");
1600
            case onnx::TensorProto::BOOL: return create_literal(shape::int32_type, dims, s.data());
Khalique's avatar
Khalique committed
1601
1602
1603
1604
            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
1605
1606
1607
1608
1609
            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
1610
            MIGRAPHX_THROW("Invalid tensor type");
1611
        }
Paul's avatar
Paul committed
1612
1613
1614
1615
        switch(t.data_type())
        {
        case onnx::TensorProto::UNDEFINED: throw std::runtime_error("");
        case onnx::TensorProto::FLOAT:
Khalique's avatar
Khalique committed
1616
            return create_literal(shape::float_type, dims, t.float_data());
Paul's avatar
Paul committed
1617
1618
        case onnx::TensorProto::UINT8: throw std::runtime_error("");
        case onnx::TensorProto::INT8:
Khalique's avatar
Khalique committed
1619
            return create_literal(shape::int32_type, dims, t.int32_data());
Paul's avatar
Paul committed
1620
        case onnx::TensorProto::UINT16:
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::INT16:
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::INT32:
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::INT64:
Khalique's avatar
Khalique committed
1627
            return create_literal(shape::int64_type, dims, t.int64_data());
Paul's avatar
Paul committed
1628
1629
        case onnx::TensorProto::STRING: throw std::runtime_error("");
        case onnx::TensorProto::BOOL:
Khalique's avatar
Khalique committed
1630
            return create_literal(shape::int32_type, dims, t.int32_data());
Paul's avatar
Paul committed
1631
        case onnx::TensorProto::FLOAT16:
Khalique's avatar
Khalique committed
1632
        {
Khalique's avatar
Khalique committed
1633
            std::vector<uint16_t> data_uint16(t.int32_data().begin(), t.int32_data().end());
1634
            std::vector<half> data_half;
Khalique's avatar
Khalique committed
1635
1636
1637
            std::transform(data_uint16.begin(),
                           data_uint16.end(),
                           std::back_inserter(data_half),
1638
                           [](uint16_t raw_val) { return *reinterpret_cast<half*>(&raw_val); });
1639
            return create_literal(shape::half_type, dims, data_half);
Khalique's avatar
Khalique committed
1640
        }
Paul's avatar
Paul committed
1641
        case onnx::TensorProto::DOUBLE:
Khalique's avatar
Khalique committed
1642
            return create_literal(shape::double_type, dims, t.double_data());
Paul's avatar
Paul committed
1643
1644
1645
1646
1647
        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
1648
        MIGRAPHX_THROW("Invalid tensor type");
Paul's avatar
Paul committed
1649
1650
    }

Khalique's avatar
Khalique committed
1651
    static literal
1652
    create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, const char* data)
1653
    {
Khalique's avatar
Khalique committed
1654
        // in case of scalar constants in onnx file, use dims=1 to fill initializer data
1655
        if(dims.empty())
1656
            return literal{{shape_type}, data};
1657
1658
1659
        return literal{{shape_type, dims}, data};
    }

1660
    template <class T, MIGRAPHX_REQUIRES(not std::is_pointer<T>{})>
Khalique's avatar
Khalique committed
1661
    static literal create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, T data)
1662
1663
    {
        if(dims.empty())
1664
            return literal{{shape_type}, data.begin(), data.end()};
1665
        return literal{{shape_type, dims}, data.begin(), data.end()};
1666
1667
    }

Paul's avatar
Paul committed
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
    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
1687
        case onnx::TensorProto::FLOAT16: shape_type = shape::half_type; break;
Paul's avatar
Paul committed
1688
1689
1690
1691
1692
1693
1694
1695
1696
        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
1697
        auto&& tensor_dims = t.tensor_type().shape().dim();
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
        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
1709
1710
        return {shape_type, dims};
    }
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732

    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
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
};

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
1756
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1757
} // namespace migraphx