tf.cpp 35.4 KB
Newer Older
Khalique's avatar
Khalique committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
#include <google/protobuf/text_format.h>
#include <google/protobuf/io/zero_copy_stream_impl.h>
#include <graph.pb.h>
#include <iostream>
#include <fstream>
#include <unordered_map>
#include <unordered_set>
#include <functional>
#include <array>
#include <utility>
#include <vector>

#include <migraphx/fallthrough.hpp>
#include <migraphx/program.hpp>
#include <migraphx/operators.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/config.hpp>
#include <migraphx/tf.hpp>

namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {

struct tf_parser
{
    using attribute_map = std::unordered_map<std::string, tensorflow::AttrValue>;
Khalique's avatar
Khalique committed
27
    using node_map      = std::unordered_map<std::string, tensorflow::NodeDef>;
Khalique's avatar
Khalique committed
28
29
    // using input_node_map = std::unordered_map<std::string, std::unordered_set<std::string>>;
    using op_func = std::function<instruction_ref(attribute_map, std::vector<instruction_ref>)>;
Khalique's avatar
Khalique committed
30

Khalique's avatar
Khalique committed
31
32
33
34
35
36
37
38
    node_map nodes;
    std::vector<tensorflow::NodeDef> input_nodes;
    std::unordered_map<std::string, instruction_ref> instructions;
    program prog = program();
    bool is_nhwc = true;

    std::unordered_map<std::string, op_func> ops;

Khalique's avatar
Khalique committed
39
    std::vector<size_t> parse_axes(const attribute_map& attributes, const std::string& s) const
40
    {
41
42
43
        auto attrs = attributes.at(s).list().i();
        std::vector<size_t> axes;
        copy(attrs.begin(), attrs.end(), std::back_inserter(axes));
Khalique's avatar
Khalique committed
44
        if(is_nhwc)
45
        {
Khalique's avatar
Khalique committed
46
47
48
            std::transform(axes.begin(), axes.end(), axes.begin(), [&](size_t axis) {
                return parse_axis(axis);
            });
49
50
51
52
        }
        return axes;
    }

Khalique's avatar
Khalique committed
53
54
55
56
57
58
    template <class T>
    std::vector<T> parse_axes(std::vector<T> axes) const
    {
        std::vector<T> new_axes;
        if(is_nhwc)
        {
Khalique's avatar
Khalique committed
59
60
61
62
            std::transform(axes.begin(),
                           axes.end(),
                           std::back_inserter(new_axes),
                           [&](size_t axis) { return parse_axis(axis); });
Khalique's avatar
Khalique committed
63
64
65
66
        }
        return new_axes;
    }

Khalique's avatar
Khalique committed
67
68
69
    // tf stores certain attributes such as strides, dilations, as a 4D input.
    // The first and last dims are equal to 1, and the relevant data is in dims 2 and 3.
    // This helper function reorders the data to store for the respective operator member variables.
70
    template <class T>
71
    void reorder_data(std::vector<T>& prev_data) const
72
73
    {
        std::vector<T> new_data(prev_data.size());
74
        for(size_t i = 0; i < new_data.size(); i++)
75
        {
Khalique's avatar
Khalique committed
76
            auto new_idx         = parse_axis(i);
77
            new_data.at(new_idx) = prev_data.at(i);
78
        }
79
80
81
82
        prev_data = new_data;
    }

    template <class T>
Khalique's avatar
Khalique committed
83
    T parse_axis(const T& dim) const
84
    {
Khalique's avatar
Khalique committed
85
        T new_dim = dim;
86
87
88
89
        if(is_nhwc)
        {
            switch(dim)
            {
Khalique's avatar
Khalique committed
90
91
92
93
94
            case 0: new_dim = 0; break;
            case 1: new_dim = 2; break;
            case 2: new_dim = 3; break;
            case 3: new_dim = 1; break;
            default: break;
95
96
            }
        }
Khalique's avatar
Khalique committed
97
        return new_dim;
98
99
    }

100
101
102
103
104
105
106
    std::vector<int64_t> get_axes(size_t num_axes) const
    {
        std::vector<int64_t> axes(num_axes);
        std::iota(axes.begin(), axes.end(), 0);
        return axes;
    }

Khalique's avatar
Khalique committed
107
108
109
110
111
    tf_parser()
    {
        add_generic_op("Identity", op::identity{});
        add_generic_op("Relu", op::relu{});

112
        add_binary_op("Add", op::add{});
Khalique's avatar
Khalique committed
113

114
        add_mem_op("AvgPool", &tf_parser::parse_pooling);
115
116
        add_mem_op("BiasAdd", &tf_parser::parse_biasadd);
        add_mem_op("ConcatV2", &tf_parser::parse_concat);
Khalique's avatar
Khalique committed
117
118
119
        add_mem_op("Const", &tf_parser::parse_constant);
        add_mem_op("Conv2D", &tf_parser::parse_conv);
        add_mem_op("FusedBatchNorm", &tf_parser::parse_batchnorm);
120
        add_mem_op("MaxPool", &tf_parser::parse_pooling);
Khalique's avatar
Khalique committed
121
122
        add_mem_op("Mean", &tf_parser::parse_mean);
        add_mem_op("Pad", &tf_parser::parse_pad);
123
124
125
        add_mem_op("Reshape", &tf_parser::parse_reshape);
        add_mem_op("Softmax", &tf_parser::parse_softmax);
        add_mem_op("Squeeze", &tf_parser::parse_squeeze);
Khalique's avatar
Khalique committed
126
127
    }

128
129
130
131
132
133
134
135
136
137
138
139
140
    template <class F>
    void add_op(std::string name, F f)
    {
        ops.emplace(name, f);
    }

    // Multi output op
    template <class F>
    void add_multi_op(std::string name, F f)
    {
        ops.emplace(name, f);
    }

Khalique's avatar
Khalique committed
141
142
143
    template <class F>
    void add_mem_op(std::string name, F f)
    {
144
        add_op(name, [=](auto&&... xs) {
Khalique's avatar
Khalique committed
145
146
147
148
149
150
151
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }

    template <class T>
    void add_binary_op(std::string name, T x)
    {
Khalique's avatar
Khalique committed
152
        add_op(name, [this, x](attribute_map attributes, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
153
154
            if(args.size() != 2)
                MIGRAPHX_THROW("binary operators should have 2 operands");
155
156
157
158
159
            auto l0 = args[1];
            if(contains(attributes, "data_format"))
            {
                if(is_nhwc)
                {
Khalique's avatar
Khalique committed
160
                    l0 = prog.add_instruction(op::transpose{{0, 3, 1, 2}}, args[1]);
161
162
163
                }
            }
            return add_broadcastable_binary_op(args[0], l0, x);
Khalique's avatar
Khalique committed
164
165
166
167
168
169
        });
    }

    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
Khalique's avatar
Khalique committed
170
        if(arg0->get_shape().lens() != arg1->get_shape().lens())
Khalique's avatar
Khalique committed
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
        {
            // 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)
            //
            // Get lengths for both arguments
186
187
            const std::vector<size_t>* s0 = &arg0->get_shape().lens();
            const std::vector<size_t>* s1 = &arg1->get_shape().lens();
Khalique's avatar
Khalique committed
188
189
190
191
192

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

193
            std::vector<size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
            auto offset = s1->size() - s0->size();
            std::transform(s0->begin(),
                           s0->end(),
                           s1->begin() + offset,
                           output_lens.begin() + offset,
                           [](auto a, auto b) { return std::max(a, b); });

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

    template <class T>
    void add_generic_op(std::string name, T x)
    {
214
        add_op(name, [this, x](attribute_map, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
215
216
217
218
219
220
221
            return prog.add_instruction(x, args);
        });
    }

    instruction_ref
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
222
223
224
        float epsilon                                     = 1e-5f;
        float momentum                                    = 0.9f;
        op::batch_norm_inference::bn_infer_mode_t bn_mode = op::batch_norm_inference::spatial;
Khalique's avatar
Khalique committed
225
226
227
228
229
230
231
232
        if(contains(attributes, "epsilon"))
        {
            epsilon = attributes.at("epsilon").f();
        }
        op::batch_norm_inference op{epsilon, momentum, bn_mode};
        return prog.add_instruction(op, std::move(args));
    }

233
    instruction_ref
Khalique's avatar
Khalique committed
234
    parse_biasadd(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
235
    {
236
        uint64_t axis = 1; // assume output of previous layer is in NCHW (broadcast on channel)
Khalique's avatar
Khalique committed
237
        auto l0       = prog.add_instruction(op::broadcast{axis, args[0]->get_shape()}, args[1]);
238
        return prog.add_instruction(op::add{}, args[0], l0);
239
240
    }

Khalique's avatar
Khalique committed
241
242
243
244
    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        // get index for axis within args
245
        size_t axis_idx = attributes.at("N").i();
Khalique's avatar
Khalique committed
246
        size_t axis     = parse_axis(args[axis_idx]->eval().at<int64_t>());
Khalique's avatar
Khalique committed
247
        op::concat op{axis};
248
        // return only first N arguments (assuming last index is the axis value)
Khalique's avatar
Khalique committed
249
        return prog.add_instruction(
250
            op, std::vector<instruction_ref>(args.begin(), args.begin() + args.size() - 1));
Khalique's avatar
Khalique committed
251
252
253
254
255
256
    }

    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
    {
Khalique's avatar
Khalique committed
257
258
        literal v       = parse_tensor(attributes.at("value").tensor());
        auto l0         = prog.add_literal(v);
259
260
261
262
263
264
        size_t num_axes = l0->get_shape().lens().size();
        if(num_axes >= 4)
        {
            std::vector<int64_t> transpose_axes = get_axes(num_axes);
            reorder_data(transpose_axes);
            l0 = prog.add_instruction(op::transpose{transpose_axes}, l0);
Khalique's avatar
Khalique committed
265
        }
266
        return l0;
Khalique's avatar
Khalique committed
267
268
269
270
271
    }

    instruction_ref
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
272
        op::convolution op;
Khalique's avatar
Khalique committed
273
274
        if(contains(attributes, "strides"))
        {
275
            std::vector<size_t> stride;
276
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
277
            reorder_data(stride);
278
279
            if(stride.size() != 4)
            {
280
                MIGRAPHX_THROW("strides should have 4 values");
281
            }
282
283
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
Khalique's avatar
Khalique committed
284
285
286
        }
        if(contains(attributes, "dilations"))
        {
287
            std::vector<size_t> dilation;
288
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
289
            reorder_data(dilation);
290
291
292
293
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
294
295
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
Khalique's avatar
Khalique committed
296
        }
Khalique's avatar
Khalique committed
297
        auto weights = args[1];
298
        // check if weights are from a constant
Khalique's avatar
Khalique committed
299
300

        if(weights->name() != "@param")
301
        {
Khalique's avatar
Khalique committed
302
303
304
305
306
307
308
309
            if(is_nhwc)
            {
                weights = prog.add_instruction(op::transpose{{1, 3, 0, 2}}, args[1]);
            }
            else
            {
                weights = prog.add_instruction(op::transpose{{3, 2, 0, 1}}, args[1]);
            }
310
        }
Khalique's avatar
Khalique committed
311

312
313
        if(contains(attributes, "padding"))
        {
Khalique's avatar
Khalique committed
314
            const std::string& pad_mode     = attributes.at("padding").s();
315
            std::vector<size_t> weight_dims = weights->get_shape().lens();
Khalique's avatar
Khalique committed
316
317
            size_t weight_h                 = weight_dims[2];
            size_t weight_w                 = weight_dims[3];
318
319
            if(pad_mode.find("SAME") != std::string::npos)
            {
320
                op.padding_mode = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
321
322
323
324
325
326
327
328
                op.padding[0] =
                    static_cast<size_t>(std::ceil(static_cast<double>(
                                            -op.stride[0] + op.dilation[0] * (weight_h - 1) + 1)) /
                                        2);
                op.padding[1] =
                    static_cast<size_t>(std::ceil(static_cast<double>(
                                            -op.stride[1] + op.dilation[1] * (weight_w - 1) + 1)) /
                                        2);
329
330
331
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
332
                op.padding_mode = op::padding_mode_t::valid;
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
            }
            else if(pad_mode.find("EXPLICIT") != std::string::npos)
            {
                std::vector<size_t> padding;
                copy(attributes.at("explicit_paddings").list().i(), std::back_inserter(padding));
                if(padding.size() != 4)
                {
                    MIGRAPHX_THROW("padding should have 4 values");
                }
                if(padding[0] != padding[2] || padding[1] != padding[3])
                {
                    MIGRAPHX_THROW("migraphx does not support asymetric padding");
                }
                op.padding[0] = padding[0];
                op.padding[1] = padding[1];
            }
        }

Khalique's avatar
Khalique committed
351
        return prog.add_instruction(op, {args[0], weights});
Khalique's avatar
Khalique committed
352
353
    }

Khalique's avatar
Khalique committed
354
355
    instruction_ref
    parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
356
    {
Khalique's avatar
Khalique committed
357
358

        auto axes      = parse_axes(args[1]->eval().get<int32_t>().to_vector());
Khalique's avatar
Khalique committed
359
        bool keep_dims = attributes.at("keep_dims").b();
Khalique's avatar
Khalique committed
360
        std::vector<int32_t> hw_axes{2, 3};
Khalique's avatar
Khalique committed
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
        if(axes == hw_axes and keep_dims)
        {
            op::pooling op{"average"};
            std::vector<size_t> input_dims{args[0]->get_shape().lens()};
            op.lengths[0] = input_dims[2];
            op.lengths[1] = input_dims[3];
            return prog.add_instruction(op, args.front());
        }
        MIGRAPHX_THROW("MIGraphX does not support mean outside of GlobalAvgPool transformation");
    }

    instruction_ref
    parse_pad(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        size_t ndims = args.front()->get_shape().lens().size();

Khalique's avatar
Khalique committed
377
378
        // in tf, the paddings are arranged as a 2d shape (ndims, 2),
        // the last dim contains the left padding and right padding respectively
Khalique's avatar
Khalique committed
379
380
        std::vector<std::pair<int32_t, int32_t>> pad_per_dim(ndims);
        auto tf_padding = args[1]->eval().get<int32_t>().to_vector();
Khalique's avatar
Khalique committed
381
        for(size_t i = 0; i < 2 * ndims; i += 2)
Khalique's avatar
Khalique committed
382
        {
Khalique's avatar
Khalique committed
383
384
            pad_per_dim[i / 2].first  = tf_padding[i];
            pad_per_dim[i / 2].second = tf_padding[i + 1];
Khalique's avatar
Khalique committed
385
386
387
388
        }
        reorder_data(pad_per_dim);

        op::pad op;
Khalique's avatar
Khalique committed
389
390
        std::vector<int64_t> pads(ndims * 2);
        for(size_t i = 0; i < ndims; i++)
Khalique's avatar
Khalique committed
391
        {
Khalique's avatar
Khalique committed
392
393
            pads[i]         = pad_per_dim[i].first;
            pads[i + ndims] = pad_per_dim[i].second;
Khalique's avatar
Khalique committed
394
395
396
397
398
        }
        op.pads = pads;
        return prog.add_instruction(op, args.front());
    }

399
400
401
402
403
    instruction_ref parse_pooling(const std::string& name,
                                  attribute_map attributes,
                                  std::vector<instruction_ref> args)
    {
        op::pooling op{starts_with(name, "Max") ? "max" : "average"};
Khalique's avatar
Khalique committed
404

405
406
407
408
409
410
411
412
413
414
415
416
417
418
        if(contains(attributes, "padding"))
        {
            const std::string& pad_mode = attributes.at("padding").s();
            if(pad_mode.find("SAME") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::same;
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::valid;
            }
        }
        if(contains(attributes, "strides"))
        {
419
            std::vector<size_t> stride;
420
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
421
            reorder_data(stride);
422
423
424
425
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
426
427
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
428
429
430
        }
        if(contains(attributes, "ksize"))
        {
431
            std::vector<size_t> ksize;
432
            copy(attributes.at("ksize").list().i(), std::back_inserter(ksize));
433
            reorder_data(ksize);
434
435
436
            if(ksize.size() != 4)
            {
                MIGRAPHX_THROW("ksize should have 4 values");
Khalique's avatar
Khalique committed
437
            }
438
439
            op.lengths[0] = ksize[2];
            op.lengths[1] = ksize[3];
440
        }
441
        return prog.add_instruction(op, args[0]);
442
    }
Khalique's avatar
Khalique committed
443

444
    instruction_ref
Khalique's avatar
Khalique committed
445
    parse_reshape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
446
447
448
449
    {
        op::reshape op;
        if(args.size() != 2)
            MIGRAPHX_THROW("reshape needs 2 arguments (input, new_shape)");
Khalique's avatar
Khalique committed
450
        auto s = args[1]->eval();
451
452
453
454
        s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
        return prog.add_instruction(op, args[0]);
    }

Khalique's avatar
Khalique committed
455
456
457
458
459
460
461
462
463
    void parse_from(std::istream& is)
    {
        tensorflow::GraphDef graph;
        if(graph.ParseFromIstream(&is))
        {
            this->parse_graph(graph);
        }
        else
        {
464
            throw std::runtime_error("Failed reading tf file");
Khalique's avatar
Khalique committed
465
466
467
        }
    }

468
469
470
471
472
473
474
475
476
477
    instruction_ref
    parse_softmax(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        auto dims = args.front()->get_shape().lens();
        auto r =
            prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1]), 1, 1}}, args.front());
        auto s = prog.add_instruction(op::softmax{}, r);
        return prog.add_instruction(op::reshape{{long(dims[0]), long(dims[1])}}, s);
    }

Khalique's avatar
Khalique committed
478
479
480
    instruction_ref parse_squeeze(const std::string&,
                                  const attribute_map& attributes,
                                  std::vector<instruction_ref> args)
481
482
    {
        op::squeeze op;
483
        auto axes = parse_axes(attributes, "squeeze_dims");
484
        copy(axes, std::back_inserter(op.axes));
485
        auto args0_dims = args[0]->get_shape().lens();
486
487
        if(op.axes.empty()) // no squeeze_dims provided, remove any dim that equals 1
        {
488
            for(size_t i = 0; i < args0_dims.size(); i++)
489
            {
490
                if(args0_dims.at(i) == 1)
491
492
493
494
                {
                    op.axes.push_back(i);
                }
            }
495
        }
496
        return prog.add_instruction(op, args[0]);
497
498
    }

Khalique's avatar
Khalique committed
499
500
501
502
503
    void parse_graph(const tensorflow::GraphDef& graph)
    {
        nodes = get_nodes(graph, input_nodes);
        for(auto&& input : input_nodes)
        {
Khalique's avatar
Khalique committed
504
            const std::string& name   = input.name();
Khalique's avatar
Khalique committed
505
            attribute_map input_attrs = get_attributes(input);
Khalique's avatar
Khalique committed
506
507
            shape::type_t shape_type  = parse_type(input_attrs.at("dtype").type());
            std::vector<size_t> dims  = parse_dims(input_attrs.at("shape").shape());
508
            if(is_nhwc and dims.size() >= 4)
509
            {
510
                reorder_data(dims);
511
            }
Khalique's avatar
Khalique committed
512
513
            shape s            = shape{shape_type, dims};
            instructions[name] = prog.add_parameter(name, s);
Khalique's avatar
Khalique committed
514
515
516
        }
        for(auto&& p : nodes)
        {
517
            this->parse_node(p.first);
Khalique's avatar
Khalique committed
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
        }
    }

    void parse_node(const std::string& name)
    {
        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)
                {
                    auto&& iname = get_name(nodes.at(input));
                    assert(name != iname);
                    this->parse_node(iname);
                    args.push_back(instructions.at(iname));
                }
                else
                {
                    args.push_back(instructions.at(input));
                }
            }
            if(ops.count(node.op()) == 0)
            {
544
                instructions[name] = prog.add_instruction(op::unknown{node.op()}, args);
Khalique's avatar
Khalique committed
545
546
547
548
549
550
551
552
553
554
555
            }
            else
            {
                instructions[name] = ops[node.op()](get_attributes(node), args);
            }
        }
    }

    static attribute_map get_attributes(const tensorflow::NodeDef& node)
    {
        attribute_map result;
Khalique's avatar
Khalique committed
556
        for(auto&& attr : node.attr())
Khalique's avatar
Khalique committed
557
558
559
560
561
562
        {
            result[attr.first] = attr.second;
        }
        return result;
    }

Khalique's avatar
Khalique committed
563
    static std::string get_name(const tensorflow::NodeDef& node) { return node.name(); }
Khalique's avatar
Khalique committed
564

Khalique's avatar
Khalique committed
565
566
    static node_map get_nodes(const tensorflow::GraphDef& graph,
                              std::vector<tensorflow::NodeDef>& input_nodes)
Khalique's avatar
Khalique committed
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
    {
        node_map result;
        for(auto&& node : graph.node())
        {
            auto node_name = get_name(node);
            // assume each node in graph has an associated name
            if(node_name.empty())
                MIGRAPHX_THROW("tf node with no name found");
            result[node_name] = node;
            if(node.op() == "Placeholder")
            {
                input_nodes.push_back(node);
            }
        }
        return result;
    }

    static shape::type_t parse_type(const tensorflow::DataType t)
    {
        shape::type_t shape_type{};
        switch(t)
        {
        case tensorflow::DataType::DT_INVALID:
            break; // throw std::runtime_error("Unsupported type UNDEFINED");
        case tensorflow::DataType::DT_FLOAT: shape_type = shape::float_type; break;
        case tensorflow::DataType::DT_DOUBLE: shape_type = shape::double_type; break;
        case tensorflow::DataType::DT_INT32: shape_type = shape::int32_type; break;
        case tensorflow::DataType::DT_UINT8:
            break; // throw std::runtime_error("Unsupported type UINT8");
        case tensorflow::DataType::DT_INT16: shape_type = shape::int16_type; break;
        case tensorflow::DataType::DT_INT8: shape_type = shape::int8_type; break;
        case tensorflow::DataType::DT_STRING:
            break; // throw std::runtime_error("Unsupported type STRING");
        case tensorflow::DataType::DT_COMPLEX64:
            break; // throw std::runtime_error("Unsupported type COMPLEX64");
        case tensorflow::DataType::DT_INT64: shape_type = shape::int64_type; break;
        case tensorflow::DataType::DT_BOOL:
            break; // throw std::runtime_error("Unsupported type BOOL");
        case tensorflow::DataType::DT_QINT8:
            break; // throw std::runtime_error("Unsupported type QINT8");
        case tensorflow::DataType::DT_QUINT8:
            break; // throw std::runtime_error("Unsupported type QUINT8");
        case tensorflow::DataType::DT_QINT32:
            break; // throw std::runtime_error("Unsupported type QINT32");
        case tensorflow::DataType::DT_BFLOAT16:
            break; // throw std::runtime_error("Unsupported type BFLOAT16");
        case tensorflow::DataType::DT_QINT16:
            break; // throw std::runtime_error("Unsupported type QINT16");
        case tensorflow::DataType::DT_QUINT16:
            break; // throw std::runtime_error("Unsupported type QUINT16");
        case tensorflow::DataType::DT_UINT16: shape_type = shape::uint16_type; break;
        case tensorflow::DataType::DT_COMPLEX128:
            break; // throw std::runtime_error("Unsupported type COMPLEX128");
        case tensorflow::DataType::DT_HALF: shape_type = shape::half_type; break;
        case tensorflow::DataType::DT_RESOURCE:
            break; // throw std::runtime_error("Unsupported type RESOURCE");
        case tensorflow::DataType::DT_VARIANT:
            break; // throw std::runtime_error("Unsupported type VARIANT");
        case tensorflow::DataType::DT_UINT32: shape_type = shape::uint32_type; break;
Khalique's avatar
Khalique committed
626
627
628
        case tensorflow::DataType::DT_UINT64:
            shape_type = shape::uint64_type;
            break;
Khalique's avatar
Khalique committed
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655

        // tf pb should not use these types
        case tensorflow::DataType::DT_FLOAT_REF: break;
        case tensorflow::DataType::DT_DOUBLE_REF: break;
        case tensorflow::DataType::DT_INT32_REF: break;
        case tensorflow::DataType::DT_UINT8_REF: break;
        case tensorflow::DataType::DT_INT16_REF: break;
        case tensorflow::DataType::DT_INT8_REF: break;
        case tensorflow::DataType::DT_STRING_REF: break;
        case tensorflow::DataType::DT_COMPLEX64_REF: break;
        case tensorflow::DataType::DT_INT64_REF: break;
        case tensorflow::DataType::DT_BOOL_REF: break;
        case tensorflow::DataType::DT_QINT8_REF: break;
        case tensorflow::DataType::DT_QUINT8_REF: break;
        case tensorflow::DataType::DT_QINT32_REF: break;
        case tensorflow::DataType::DT_BFLOAT16_REF: break;
        case tensorflow::DataType::DT_QINT16_REF: break;
        case tensorflow::DataType::DT_QUINT16_REF: break;
        case tensorflow::DataType::DT_UINT16_REF: break;
        case tensorflow::DataType::DT_COMPLEX128_REF: break;
        case tensorflow::DataType::DT_HALF_REF: break;
        case tensorflow::DataType::DT_RESOURCE_REF: break;
        case tensorflow::DataType::DT_VARIANT_REF: break;
        case tensorflow::DataType::DT_UINT32_REF: break;
        case tensorflow::DataType::DT_UINT64_REF: break;
        case tensorflow::DataType::DataType_INT_MAX_SENTINEL_DO_NOT_USE_: break;
        case tensorflow::DataType::DataType_INT_MIN_SENTINEL_DO_NOT_USE_: break;
Khalique's avatar
Khalique committed
656
657
658
659
        }
        return shape_type;
    }

Khalique's avatar
Khalique committed
660
    static literal parse_tensor(const tensorflow::TensorProto& t)
Khalique's avatar
Khalique committed
661
662
663
664
665
666
    {
        std::vector<size_t> dims = parse_dims(t.tensor_shape());
        if(dims.empty())
        {
            dims = {1};
        }
667
        size_t shape_size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<size_t>());
Khalique's avatar
Khalique committed
668
669
        if(!t.tensor_content().empty()) // has raw data
        {
Khalique's avatar
Khalique committed
670
            const std::string& s = t.tensor_content();
Khalique's avatar
Khalique committed
671
672
673
            switch(t.dtype())
            {
            case tensorflow::DataType::DT_INVALID: throw std::runtime_error("");
Khalique's avatar
Khalique committed
674
675
            case tensorflow::DataType::DT_FLOAT:
                return literal{{shape::float_type, dims}, s.data()};
Khalique's avatar
Khalique committed
676
677
            case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
            case tensorflow::DataType::DT_INT8: return literal{{shape::int32_type, dims}, s.data()};
Khalique's avatar
Khalique committed
678
679
680
681
682
683
684
685
            case tensorflow::DataType::DT_UINT16:
                return literal{{shape::int32_type, dims}, s.data()};
            case tensorflow::DataType::DT_INT16:
                return literal{{shape::int32_type, dims}, s.data()};
            case tensorflow::DataType::DT_INT32:
                return literal{{shape::int32_type, dims}, s.data()};
            case tensorflow::DataType::DT_INT64:
                return literal{{shape::int64_type, dims}, s.data()};
Khalique's avatar
Khalique committed
686
687
688
            case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
            case tensorflow::DataType::DT_BOOL: return literal{{shape::int32_type, dims}, s.data()};
            case tensorflow::DataType::DT_HALF: return literal{{shape::half_type, dims}, s.data()};
Khalique's avatar
Khalique committed
689
690
            case tensorflow::DataType::DT_DOUBLE:
                return literal{{shape::double_type, dims}, s.data()};
Khalique's avatar
Khalique committed
691
692
693
694
            case tensorflow::DataType::DT_UINT32: throw std::runtime_error("");
            case tensorflow::DataType::DT_UINT64: throw std::runtime_error("");
            case tensorflow::DataType::DT_COMPLEX64: throw std::runtime_error("");
            case tensorflow::DataType::DT_COMPLEX128: throw std::runtime_error("");
Khalique's avatar
Khalique committed
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
            case tensorflow::DataType::DT_QINT8: throw std::runtime_error("");
            case tensorflow::DataType::DT_QUINT8: throw std::runtime_error("");
            case tensorflow::DataType::DT_QINT32: throw std::runtime_error("");
            case tensorflow::DataType::DT_BFLOAT16: throw std::runtime_error("");
            case tensorflow::DataType::DT_QINT16: throw std::runtime_error("");
            case tensorflow::DataType::DT_QUINT16: throw std::runtime_error("");
            case tensorflow::DataType::DT_RESOURCE: throw std::runtime_error("");
            case tensorflow::DataType::DT_VARIANT: throw std::runtime_error("");
            case tensorflow::DataType::DT_FLOAT_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_DOUBLE_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_INT32_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_UINT8_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_INT16_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_INT8_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_STRING_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_COMPLEX64_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_INT64_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_BOOL_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_QINT8_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_QUINT8_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_QINT32_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_BFLOAT16_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_QINT16_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_QUINT16_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_UINT16_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_COMPLEX128_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_HALF_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_RESOURCE_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_VARIANT_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_UINT32_REF: throw std::runtime_error("");
            case tensorflow::DataType::DT_UINT64_REF: throw std::runtime_error("");
Khalique's avatar
Khalique committed
726
727
728
729
            case tensorflow::DataType::DataType_INT_MAX_SENTINEL_DO_NOT_USE_:
                throw std::runtime_error("");
            case tensorflow::DataType::DataType_INT_MIN_SENTINEL_DO_NOT_USE_:
                throw std::runtime_error("");
Khalique's avatar
Khalique committed
730
731
732
733
734
735
736
            }
            MIGRAPHX_THROW("Invalid tensor type");
        }
        switch(t.dtype())
        {
        case tensorflow::DataType::DT_INVALID: throw std::runtime_error("");
        case tensorflow::DataType::DT_FLOAT:
737
            return literal{{shape::float_type, dims}, get_data_vals(t.float_val(), shape_size)};
Khalique's avatar
Khalique committed
738
739
        case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT8:
740
            return literal{{shape::int32_type, dims}, get_data_vals(t.int_val(), shape_size)};
Khalique's avatar
Khalique committed
741
        case tensorflow::DataType::DT_UINT16:
742
            return literal{{shape::int32_type, dims}, get_data_vals(t.int_val(), shape_size)};
Khalique's avatar
Khalique committed
743
        case tensorflow::DataType::DT_INT16:
744
            return literal{{shape::int32_type, dims}, get_data_vals(t.int_val(), shape_size)};
Khalique's avatar
Khalique committed
745
        case tensorflow::DataType::DT_INT32:
746
            return literal{{shape::int32_type, dims}, get_data_vals(t.int_val(), shape_size)};
Khalique's avatar
Khalique committed
747
        case tensorflow::DataType::DT_INT64:
748
            return literal{{shape::int64_type, dims}, get_data_vals(t.int64_val(), shape_size)};
Khalique's avatar
Khalique committed
749
750
        case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
        case tensorflow::DataType::DT_BOOL:
751
            return literal{{shape::int32_type, dims}, get_data_vals(t.bool_val(), shape_size)};
Khalique's avatar
Khalique committed
752
        case tensorflow::DataType::DT_HALF:
Khalique's avatar
Khalique committed
753
        {
754
755
            std::vector<int> data_int32 = get_data_vals(t.half_val(), shape_size);
            std::vector<uint16_t> data_uint16(data_int32.begin(), data_int32.end());
Khalique's avatar
Khalique committed
756
757
758
759
760
761
            std::vector<half> data_half;
            std::transform(data_uint16.begin(),
                           data_uint16.end(),
                           std::back_inserter(data_half),
                           [](uint16_t raw_val) { return *reinterpret_cast<half*>(&raw_val); });
            return literal{{shape::half_type, dims}, data_half};
Khalique's avatar
Khalique committed
762
        }
Khalique's avatar
Khalique committed
763
        case tensorflow::DataType::DT_DOUBLE:
Khalique's avatar
Khalique committed
764
            return literal{{shape::double_type, dims}, get_data_vals(t.double_val(), shape_size)};
Khalique's avatar
Khalique committed
765
766
767
768
        case tensorflow::DataType::DT_UINT32: throw std::runtime_error("");
        case tensorflow::DataType::DT_UINT64: throw std::runtime_error("");
        case tensorflow::DataType::DT_COMPLEX64: throw std::runtime_error("");
        case tensorflow::DataType::DT_COMPLEX128: throw std::runtime_error("");
Khalique's avatar
Khalique committed
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
        case tensorflow::DataType::DT_QINT8: throw std::runtime_error("");
        case tensorflow::DataType::DT_QUINT8: throw std::runtime_error("");
        case tensorflow::DataType::DT_QINT32: throw std::runtime_error("");
        case tensorflow::DataType::DT_BFLOAT16: throw std::runtime_error("");
        case tensorflow::DataType::DT_QINT16: throw std::runtime_error("");
        case tensorflow::DataType::DT_QUINT16: throw std::runtime_error("");
        case tensorflow::DataType::DT_RESOURCE: throw std::runtime_error("");
        case tensorflow::DataType::DT_VARIANT: throw std::runtime_error("");
        case tensorflow::DataType::DT_FLOAT_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_DOUBLE_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT32_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_UINT8_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT16_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT8_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_STRING_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_COMPLEX64_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT64_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_BOOL_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_QINT8_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_QUINT8_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_QINT32_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_BFLOAT16_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_QINT16_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_QUINT16_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_UINT16_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_COMPLEX128_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_HALF_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_RESOURCE_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_VARIANT_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_UINT32_REF: throw std::runtime_error("");
        case tensorflow::DataType::DT_UINT64_REF: throw std::runtime_error("");
Khalique's avatar
Khalique committed
800
801
802
803
        case tensorflow::DataType::DataType_INT_MAX_SENTINEL_DO_NOT_USE_:
            throw std::runtime_error("");
        case tensorflow::DataType::DataType_INT_MIN_SENTINEL_DO_NOT_USE_:
            throw std::runtime_error("");
Khalique's avatar
Khalique committed
804
805
806
807
        }
        MIGRAPHX_THROW("Invalid tensor type");
    }

808
    template <class T>
Khalique's avatar
Khalique committed
809
    static std::vector<T> get_data_vals(const google::protobuf::RepeatedField<T>& data,
810
                                        const size_t& shape_size)
811
812
813
814
815
816
817
818
819
820
821
822
    {
        std::vector<T> data_vals(shape_size);
        // check if shape has enough data values given existing fields
        if(data.size() == 1)
        {
            std::fill(data_vals.begin(), data_vals.end(), data[0]);
        }
        else
            copy(data.begin(), data.end(), std::back_inserter(data_vals));
        return data_vals;
    }

Khalique's avatar
Khalique committed
823
824
825
826
    static std::vector<size_t> parse_dims(const tensorflow::TensorShapeProto& s)
    {
        std::vector<size_t> dims;
        auto input_dims = s.dim();
Khalique's avatar
Khalique committed
827
828
829
830
        std::transform(input_dims.begin(),
                       input_dims.end(),
                       std::back_inserter(dims),
                       [](tensorflow::TensorShapeProto_Dim dim) { return dim.size(); });
Khalique's avatar
Khalique committed
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
        return dims;
    }
};

program parse_tf(const std::string& name, bool is_nhwc)
{
    std::fstream input(name.c_str(), std::ios::in | std::ios::binary);
    tf_parser parser;
    parser.is_nhwc = is_nhwc;

#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);
}

} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx