tf.cpp 45.2 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
#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>
Khalique's avatar
Khalique committed
20
#include <migraphx/pad_calc.hpp>
Khalique's avatar
Khalique committed
21
22
23
24
25
26
27

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
28
    using node_map      = std::map<std::string, tensorflow::NodeDef>;
Khalique's avatar
Khalique committed
29
30
    // 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
31

Khalique's avatar
Khalique committed
32
33
34
35
36
37
38
39
    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
40
    std::vector<size_t> parse_axes(const attribute_map& attributes, const std::string& s) const
41
    {
42
43
44
        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
45
        if(is_nhwc)
46
        {
Khalique's avatar
Khalique committed
47
48
49
            std::transform(axes.begin(), axes.end(), axes.begin(), [&](size_t axis) {
                return parse_axis(axis);
            });
50
51
52
53
        }
        return axes;
    }

Khalique's avatar
Khalique committed
54
55
56
57
58
    template <class T>
    std::vector<T> parse_axes(std::vector<T> axes) const
    {
        if(is_nhwc)
        {
59
            std::vector<T> new_axes;
Khalique's avatar
Khalique committed
60
61
62
63
            std::transform(axes.begin(),
                           axes.end(),
                           std::back_inserter(new_axes),
                           [&](size_t axis) { return parse_axis(axis); });
64
            return new_axes;
Khalique's avatar
Khalique committed
65
        }
66
        return axes;
Khalique's avatar
Khalique committed
67
68
    }

Khalique's avatar
Khalique committed
69
70
71
    // 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.
72
    template <class T>
73
    void reorder_data(std::vector<T>& prev_data) const
74
75
    {
        std::vector<T> new_data(prev_data.size());
76
        for(size_t i = 0; i < new_data.size(); i++)
77
        {
Khalique's avatar
Khalique committed
78
            auto new_idx         = parse_axis(i);
79
            new_data.at(new_idx) = prev_data.at(i);
80
        }
81
82
83
84
        prev_data = new_data;
    }

    template <class T>
Khalique's avatar
Khalique committed
85
    T parse_axis(const T& dim) const
86
    {
Khalique's avatar
Khalique committed
87
        T new_dim = dim;
88
89
90
91
        if(is_nhwc)
        {
            switch(dim)
            {
Khalique's avatar
Khalique committed
92
93
94
95
96
            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;
97
98
            }
        }
Khalique's avatar
Khalique committed
99
        return new_dim;
100
101
    }

102
103
104
105
106
107
108
    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
109
110
111
112
    tf_parser()
    {
        add_generic_op("Identity", op::identity{});
        add_generic_op("Relu", op::relu{});
Khalique's avatar
Khalique committed
113
        add_generic_op("Relu6", op::clip{6.0, 0.0});
Khalique's avatar
Khalique committed
114
        add_generic_op("Tanh", op::tanh{});
Khalique's avatar
Khalique committed
115

116
        add_binary_op("Add", op::add{});
Khalique's avatar
Khalique committed
117
        add_binary_op("Mul", op::mul{});
Khalique's avatar
Khalique committed
118
        add_binary_op("Sub", op::sub{});
Khalique's avatar
Khalique committed
119

120
        add_mem_op("AvgPool", &tf_parser::parse_pooling);
121
122
        add_mem_op("BiasAdd", &tf_parser::parse_biasadd);
        add_mem_op("ConcatV2", &tf_parser::parse_concat);
Khalique's avatar
Khalique committed
123
124
        add_mem_op("Const", &tf_parser::parse_constant);
        add_mem_op("Conv2D", &tf_parser::parse_conv);
Khalique's avatar
Khalique committed
125
        add_mem_op("DepthwiseConv2dNative", &tf_parser::parse_depthwiseconv);
Khalique's avatar
Khalique committed
126
        add_mem_op("FusedBatchNorm", &tf_parser::parse_batchnorm);
127
        add_mem_op("MatMul", &tf_parser::parse_matmul);
128
        add_mem_op("MaxPool", &tf_parser::parse_pooling);
Khalique's avatar
Khalique committed
129
        add_mem_op("Mean", &tf_parser::parse_mean);
Khalique's avatar
Khalique committed
130
        add_mem_op("Pack", &tf_parser::parse_pack);
Khalique's avatar
Khalique committed
131
        add_mem_op("Pad", &tf_parser::parse_pad);
132
133
134
        add_mem_op("Reshape", &tf_parser::parse_reshape);
        add_mem_op("Softmax", &tf_parser::parse_softmax);
        add_mem_op("Squeeze", &tf_parser::parse_squeeze);
135
        add_mem_op("StridedSlice", &tf_parser::parse_stridedslice);
Khalique's avatar
Khalique committed
136
137
    }

138
139
140
141
142
143
144
145
146
147
148
149
150
    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
151
152
153
    template <class F>
    void add_mem_op(std::string name, F f)
    {
154
        add_op(name, [=](auto&&... xs) {
Khalique's avatar
Khalique committed
155
156
157
158
159
160
161
            return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
        });
    }

    template <class T>
    void add_binary_op(std::string name, T x)
    {
Paul's avatar
Paul committed
162
        add_op(name, [this, x](const attribute_map& attributes, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
163
164
            if(args.size() != 2)
                MIGRAPHX_THROW("binary operators should have 2 operands");
165
166
167
168
169
            auto l0 = args[1];
            if(contains(attributes, "data_format"))
            {
                if(is_nhwc)
                {
Khalique's avatar
Khalique committed
170
                    l0 = prog.add_instruction(op::transpose{{0, 3, 1, 2}}, args[1]);
171
172
173
                }
            }
            return add_broadcastable_binary_op(args[0], l0, x);
Khalique's avatar
Khalique committed
174
175
176
177
178
179
        });
    }

    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
Khalique's avatar
Khalique committed
180
        if(arg0->get_shape().lens() != arg1->get_shape().lens())
Khalique's avatar
Khalique committed
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
        {
            // 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
196
197
            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
198
199
200
201
202

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

203
            std::vector<size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
            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)
    {
Paul's avatar
Paul committed
224
        add_op(name, [this, x](const attribute_map&, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
225
226
227
228
229
230
231
            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
232
233
234
        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
235
236
237
238
239
240
241
242
        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));
    }

243
    instruction_ref
Khalique's avatar
Khalique committed
244
    parse_biasadd(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
245
    {
246
        uint64_t axis = 1; // assume output of previous layer is in NCHW (broadcast on channel)
Shucai Xiao's avatar
Shucai Xiao committed
247
        auto l0 = prog.add_instruction(op::broadcast{axis, args[0]->get_shape().lens()}, args[1]);
248
        return prog.add_instruction(op::add{}, args[0], l0);
249
250
    }

Khalique's avatar
Khalique committed
251
252
253
254
    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        // get index for axis within args
255
        size_t axis_idx = attributes.at("N").i();
Khalique's avatar
Khalique committed
256
        size_t axis     = parse_axis(args[axis_idx]->eval().at<int64_t>());
Khalique's avatar
Khalique committed
257
        op::concat op{axis};
258
        // return only first N arguments (assuming last index is the axis value)
Khalique's avatar
Khalique committed
259
        return prog.add_instruction(
260
            op, std::vector<instruction_ref>(args.begin(), args.begin() + args.size() - 1));
Khalique's avatar
Khalique committed
261
262
263
264
265
266
    }

    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
    {
Khalique's avatar
Khalique committed
267
268
        literal v       = parse_tensor(attributes.at("value").tensor());
        auto l0         = prog.add_literal(v);
269
270
271
272
273
274
        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
275
        }
276
        return l0;
Khalique's avatar
Khalique committed
277
278
279
280
281
    }

    instruction_ref
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
282
        op::convolution op;
Khalique's avatar
Khalique committed
283
284
        if(contains(attributes, "strides"))
        {
285
            std::vector<size_t> stride;
286
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
287
            reorder_data(stride);
288
289
            if(stride.size() != 4)
            {
290
                MIGRAPHX_THROW("strides should have 4 values");
291
            }
292
293
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
Khalique's avatar
Khalique committed
294
295
296
        }
        if(contains(attributes, "dilations"))
        {
297
            std::vector<size_t> dilation;
298
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
299
            reorder_data(dilation);
300
301
302
303
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
304
305
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
Khalique's avatar
Khalique committed
306
        }
Khalique's avatar
Khalique committed
307
        auto weights = args[1];
308
        // check if weights are from a constant
Khalique's avatar
Khalique committed
309
310

        if(weights->name() != "@param")
311
        {
Khalique's avatar
Khalique committed
312
313
314
315
316
317
318
319
            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]);
            }
320
        }
Khalique's avatar
Khalique committed
321

Khalique's avatar
Khalique committed
322
        auto l0 = args[0];
323
324
        if(contains(attributes, "padding"))
        {
Khalique's avatar
Khalique committed
325
            const std::string& pad_mode = attributes.at("padding").s();
326
327
            if(pad_mode.find("SAME") != std::string::npos)
            {
Khalique's avatar
Khalique committed
328
                op.padding_mode                 = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
329
330
331
                std::vector<size_t> weight_dims = weights->get_shape().lens();
                size_t weight_h                 = weight_dims[2];
                size_t weight_w                 = weight_dims[3];
Khalique's avatar
Khalique committed
332
333

                auto input_dims = l0->get_shape().lens();
Khalique's avatar
Khalique committed
334
335
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
336
337
338
339
340
341
342
343
344
345
346
                std::vector<int64_t> pads(input_dims.size());
                calculate_padding(0, pads, input_h, op.stride[0], op.dilation[0], weight_h);
                calculate_padding(1, pads, input_w, op.stride[1], op.dilation[1], weight_w);

                if(pads[0] != pads[2] || pads[1] != pads[3])
                {
                    std::vector<int64_t> padding = {0, 0, pads[0], pads[1], 0, 0, pads[2], pads[3]};
                    l0 = prog.add_instruction(migraphx::op::pad{padding}, l0);
                }
                else
                {
Khalique's avatar
Khalique committed
347
348
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
349
                }
350
351
352
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
353
                op.padding_mode = op::padding_mode_t::valid;
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
            }
            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
372
        return prog.add_instruction(op, {l0, weights});
Khalique's avatar
Khalique committed
373
374
    }

Khalique's avatar
Khalique committed
375
376
377
    instruction_ref parse_depthwiseconv(const std::string&,
                                        attribute_map attributes,
                                        std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
378
379
380
    {
        op::convolution op;
        size_t num_channels = args[0]->get_shape().lens()[1];
Khalique's avatar
Khalique committed
381
        op.group            = num_channels;
Khalique's avatar
Khalique committed
382

Khalique's avatar
Khalique committed
383
384
385
386
387
388
389
390
391
392
393
394
        if(contains(attributes, "strides"))
        {
            std::vector<size_t> stride;
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
            reorder_data(stride);
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
        }
Khalique's avatar
Khalique committed
395
396
397
398
399
400
401
402
403
404
405
406
407
        if(contains(attributes, "dilations"))
        {
            std::vector<size_t> dilation;
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
            reorder_data(dilation);
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
        }

Khalique's avatar
Khalique committed
408
409
410
411
412
413
414
415
416
417
418
        auto weights = args[1];
        // check if weights are from a constant
        if(weights->name() != "@param")
        {
            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]);
Khalique's avatar
Khalique committed
419
            }
Khalique's avatar
Khalique committed
420
        }
Khalique's avatar
Khalique committed
421

Khalique's avatar
Khalique committed
422
        auto l0 = args[0];
Khalique's avatar
Khalique committed
423
424
        if(contains(attributes, "padding"))
        {
Khalique's avatar
Khalique committed
425
426
            const std::string& pad_mode = attributes.at("padding").s();

Khalique's avatar
Khalique committed
427
428
            if(pad_mode.find("SAME") != std::string::npos)
            {
Khalique's avatar
Khalique committed
429
                op.padding_mode                 = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
430
431
432
433
434
                std::vector<size_t> weight_dims = weights->get_shape().lens();
                size_t weight_h                 = weight_dims[2];
                size_t weight_w                 = weight_dims[3];

                auto input_dims = l0->get_shape().lens();
Khalique's avatar
Khalique committed
435
436
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
437
438
439
440
441
442
443
444
445
446
447
                std::vector<int64_t> pads(input_dims.size());
                calculate_padding(0, pads, input_h, op.stride[0], op.dilation[0], weight_h);
                calculate_padding(1, pads, input_w, op.stride[1], op.dilation[1], weight_w);

                if(pads[0] != pads[2] || pads[1] != pads[3])
                {
                    std::vector<int64_t> padding = {0, 0, pads[0], pads[1], 0, 0, pads[2], pads[3]};
                    l0 = prog.add_instruction(migraphx::op::pad{padding}, l0);
                }
                else
                {
Khalique's avatar
Khalique committed
448
449
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
450
                }
Khalique's avatar
Khalique committed
451
452
453
454
455
456
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::valid;
            }
        }
Khalique's avatar
Khalique committed
457

Khalique's avatar
Khalique committed
458
459
        std::vector<int64_t> new_weights_shape;
        copy(weights->get_shape().lens(), std::back_inserter(new_weights_shape));
Khalique's avatar
Khalique committed
460
461
462
463

        // weight format is (out_channels, in_channels, h, w), but in depthwise_conv,
        // out_channels is equal to the multiplier. Adjust by inserting a reshape and
        // setting in_channels to 1
Khalique's avatar
Khalique committed
464
        int64_t multiplier   = new_weights_shape[0];
Khalique's avatar
Khalique committed
465
466
467
        int64_t out_channels = num_channels * multiplier;
        new_weights_shape[0] = out_channels;
        new_weights_shape[1] = 1;
Paul's avatar
Paul committed
468
        // Make sure weights are contiguous before doing reshape
Paul's avatar
Paul committed
469
470
        auto cweights    = prog.add_instruction(op::contiguous{}, weights);
        auto new_weights = prog.add_instruction(op::reshape{new_weights_shape}, cweights);
Khalique's avatar
Khalique committed
471

Khalique's avatar
Khalique committed
472
        return prog.add_instruction(op, {l0, new_weights});
Khalique's avatar
Khalique committed
473
474
    }

Khalique's avatar
Khalique committed
475
476
    instruction_ref
    parse_matmul(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
477
478
479
    {
        bool transa = false;
        bool transb = false;
Khalique's avatar
Khalique committed
480

481
482
483
484
485
486
487
488
489
490
491
492
        if(contains(attributes, "transpose_a"))
        {
            transa = attributes.at("transpose_a").b();
        }
        if(contains(attributes, "transpose_b"))
        {
            transb = attributes.at("transpose_a").b();
        }

        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
Khalique's avatar
Khalique committed
493
        std::iter_swap(perm.end() - 1, perm.end() - 2);
494
495
496
497
498
499
500

        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];

        return prog.add_instruction(op::dot{}, l1, l2);
    }

Khalique's avatar
Khalique committed
501
502
    instruction_ref
    parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
503
    {
Khalique's avatar
Khalique committed
504
        auto axes      = parse_axes(args[1]->eval().get<int32_t>().to_vector());
Khalique's avatar
Khalique committed
505
        bool keep_dims = attributes.at("keep_dims").b();
Khalique's avatar
Khalique committed
506
        std::vector<int32_t> hw_axes{2, 3};
Khalique's avatar
Khalique committed
507
        // check if conditions for GlobalAvgPool are met
Khalique's avatar
Khalique committed
508
509
        auto lens = args[0]->get_shape().lens();
        if(axes == hw_axes and lens.size() == 4)
Khalique's avatar
Khalique committed
510
511
        {
            op::pooling op{"average"};
Khalique's avatar
Khalique committed
512
513
            op.lengths[0] = lens[2];
            op.lengths[1] = lens[3];
Khalique's avatar
Khalique committed
514
515
516
517
518
            auto l0       = prog.add_instruction(op, args.front());
            if(keep_dims)
                return l0;
            return prog.add_instruction(
                op::squeeze{std::vector<int64_t>(hw_axes.begin(), hw_axes.end())}, l0);
Khalique's avatar
Khalique committed
519
520
521
522
        }
        MIGRAPHX_THROW("MIGraphX does not support mean outside of GlobalAvgPool transformation");
    }

Khalique's avatar
Khalique committed
523
524
525
    instruction_ref parse_pack(const std::string&,
                               const attribute_map& attributes,
                               std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
526
527
528
529
530
531
    {
        // reinterpret as unsqueeze with concat
        std::vector<instruction_ref> unsqueezed_args;
        int64_t axis = 0;
        if(contains(attributes, "axis"))
            axis = attributes.at("axis").i();
532
533
534
        size_t input_size = args.front()->get_shape().lens().size();
        if(axis > input_size)
        {
Khalique's avatar
Khalique committed
535
536
            MIGRAPHX_THROW("TF_PARSER: axis value of " + to_string(axis) +
                           " must be smaller than input size " + to_string(input_size));
537
538
539
540
541
        }
        // check if input arg needs axis to be converted to NCHW
        if(input_size >= 4)
            axis = parse_axis(axis);

Khalique's avatar
Khalique committed
542
543
544
545
546
        std::transform(
            args.begin(),
            args.end(),
            std::back_inserter(unsqueezed_args),
            [&](instruction_ref arg) { return prog.add_instruction(op::unsqueeze{{axis}}, arg); });
Khalique's avatar
Khalique committed
547
548
549
        return prog.add_instruction(op::concat{static_cast<size_t>(axis)}, unsqueezed_args);
    }

Khalique's avatar
Khalique committed
550
551
552
553
554
    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
555
556
        // 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
557
558
        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
559
        for(size_t i = 0; i < 2 * ndims; i += 2)
Khalique's avatar
Khalique committed
560
        {
Khalique's avatar
Khalique committed
561
562
            pad_per_dim[i / 2].first  = tf_padding[i];
            pad_per_dim[i / 2].second = tf_padding[i + 1];
Khalique's avatar
Khalique committed
563
564
565
566
        }
        reorder_data(pad_per_dim);

        op::pad op;
Khalique's avatar
Khalique committed
567
568
        std::vector<int64_t> pads(ndims * 2);
        for(size_t i = 0; i < ndims; i++)
Khalique's avatar
Khalique committed
569
        {
Khalique's avatar
Khalique committed
570
571
            pads[i]         = pad_per_dim[i].first;
            pads[i + ndims] = pad_per_dim[i].second;
Khalique's avatar
Khalique committed
572
573
574
575
576
        }
        op.pads = pads;
        return prog.add_instruction(op, args.front());
    }

577
578
579
580
581
    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
582

583
584
        if(contains(attributes, "strides"))
        {
585
            std::vector<size_t> stride;
586
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
587
            reorder_data(stride);
588
589
590
591
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
592
593
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
594
595
596
        }
        if(contains(attributes, "ksize"))
        {
597
            std::vector<size_t> ksize;
598
            copy(attributes.at("ksize").list().i(), std::back_inserter(ksize));
599
            reorder_data(ksize);
600
601
602
            if(ksize.size() != 4)
            {
                MIGRAPHX_THROW("ksize should have 4 values");
Khalique's avatar
Khalique committed
603
            }
604
605
            op.lengths[0] = ksize[2];
            op.lengths[1] = ksize[3];
606
        }
Khalique's avatar
Khalique committed
607
608

        auto l0 = args[0];
Khalique's avatar
Khalique committed
609
610
611
612
613
        if(contains(attributes, "padding"))
        {
            const std::string& pad_mode = attributes.at("padding").s();
            if(pad_mode.find("SAME") != std::string::npos)
            {
Khalique's avatar
Khalique committed
614
                op.padding_mode = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
615
                auto input_dims = l0->get_shape().lens();
Khalique's avatar
Khalique committed
616
617
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
618
619
620
621
622
623
624
                std::vector<int64_t> pads(input_dims.size());
                calculate_padding(0, pads, input_h, op.stride[0], 1, op.lengths[0]);
                calculate_padding(1, pads, input_w, op.stride[1], 1, op.lengths[1]);

                if(pads[0] != pads[2] || pads[1] != pads[3])
                {
                    std::vector<int64_t> padding = {0, 0, pads[0], pads[1], 0, 0, pads[2], pads[3]};
Khalique's avatar
Khalique committed
625
626
                    l0                           = prog.add_instruction(
                        migraphx::op::pad{padding, std::numeric_limits<float>::lowest()}, l0);
Khalique's avatar
Khalique committed
627
628
629
                }
                else
                {
Khalique's avatar
Khalique committed
630
631
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
632
                }
Khalique's avatar
Khalique committed
633
634
635
636
637
638
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::valid;
            }
        }
Khalique's avatar
Khalique committed
639
        return prog.add_instruction(op, l0);
640
    }
Khalique's avatar
Khalique committed
641

642
    instruction_ref
Khalique's avatar
Khalique committed
643
    parse_reshape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
644
645
646
647
    {
        op::reshape op;
        if(args.size() != 2)
            MIGRAPHX_THROW("reshape needs 2 arguments (input, new_shape)");
Khalique's avatar
Khalique committed
648
        auto s = args[1]->eval();
649
650
651
652
        s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
        return prog.add_instruction(op, args[0]);
    }

Khalique's avatar
Khalique committed
653
654
655
656
657
658
659
660
661
    void parse_from(std::istream& is)
    {
        tensorflow::GraphDef graph;
        if(graph.ParseFromIstream(&is))
        {
            this->parse_graph(graph);
        }
        else
        {
662
            throw std::runtime_error("Failed reading tf file");
Khalique's avatar
Khalique committed
663
664
665
        }
    }

666
667
668
669
670
671
672
673
674
675
    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
676
677
678
    instruction_ref parse_squeeze(const std::string&,
                                  const attribute_map& attributes,
                                  std::vector<instruction_ref> args)
679
680
    {
        op::squeeze op;
681
        auto axes = parse_axes(attributes, "squeeze_dims");
682
        copy(axes, std::back_inserter(op.axes));
683
        auto args0_dims = args[0]->get_shape().lens();
684
685
        if(op.axes.empty()) // no squeeze_dims provided, remove any dim that equals 1
        {
686
            for(size_t i = 0; i < args0_dims.size(); i++)
687
            {
688
                if(args0_dims.at(i) == 1)
689
690
691
692
                {
                    op.axes.push_back(i);
                }
            }
693
        }
694
        return prog.add_instruction(op, args[0]);
695
696
    }

Khalique's avatar
Khalique committed
697
698
699
    instruction_ref parse_stridedslice(const std::string&,
                                       const attribute_map& attributes,
                                       std::vector<instruction_ref> args)
700
701
    {
        op::slice op;
Khalique's avatar
Khalique committed
702
703
704
705
706
707
708
709
        auto starts     = args[1]->eval().get<int32_t>().to_vector();
        auto ends       = args[2]->eval().get<int32_t>().to_vector();
        size_t num_axes = args[0]->get_shape().lens().size();
        if(num_axes >= 4)
        {
            reorder_data(starts);
            reorder_data(ends);
        }
710

Khalique's avatar
Khalique committed
711
712
713
714
        op.starts = std::vector<int64_t>(starts.begin(), starts.end());
        op.ends   = std::vector<int64_t>(ends.begin(), ends.end());
        op.axes   = std::vector<int64_t>(num_axes);
        std::iota(op.axes.begin(), op.axes.end(), 0);
715
        uint32_t shrink_axis_mask = 0;
Khalique's avatar
Khalique committed
716
        uint32_t bitwise_compare  = 1;
717
718
719
        std::vector<int64_t> squeeze_axes;

        if(contains(attributes, "shrink_axis_mask"))
720
            shrink_axis_mask = static_cast<uint32_t>(attributes.at("shrink_axis_mask").i());
721

Khalique's avatar
Khalique committed
722
        for(size_t i = 0; i < num_axes; i++)
723
        {
724
            // the LSB corresponds to axis 0 when determining which axes to squeeze
Khalique's avatar
Khalique committed
725
            if(((shrink_axis_mask >> i) & bitwise_compare) == 1)
726
727
                squeeze_axes.push_back(i);
        }
Khalique's avatar
Khalique committed
728
729
730
731
732
        if(num_axes >= 4)
        {
            squeeze_axes = parse_axes(squeeze_axes);
        }

733
734
735
736
        auto l0 = prog.add_instruction(op, args[0]);
        return prog.add_instruction(op::squeeze{squeeze_axes}, l0);
    }

Khalique's avatar
Khalique committed
737
738
739
740
741
    void parse_graph(const tensorflow::GraphDef& graph)
    {
        nodes = get_nodes(graph, input_nodes);
        for(auto&& input : input_nodes)
        {
Khalique's avatar
Khalique committed
742
            const std::string& name   = input.name();
Khalique's avatar
Khalique committed
743
            attribute_map input_attrs = get_attributes(input);
Khalique's avatar
Khalique committed
744
745
            shape::type_t shape_type  = parse_type(input_attrs.at("dtype").type());
            std::vector<size_t> dims  = parse_dims(input_attrs.at("shape").shape());
746
            if(is_nhwc and dims.size() >= 4)
747
            {
748
                reorder_data(dims);
749
            }
Khalique's avatar
Khalique committed
750
751
            shape s            = shape{shape_type, dims};
            instructions[name] = prog.add_parameter(name, s);
Khalique's avatar
Khalique committed
752
753
754
        }
        for(auto&& p : nodes)
        {
755
            this->parse_node(p.first);
Khalique's avatar
Khalique committed
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
        }
    }

    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)
            {
782
                instructions[name] = prog.add_instruction(op::unknown{node.op()}, args);
Khalique's avatar
Khalique committed
783
784
785
786
787
788
789
790
791
792
793
            }
            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
794
        for(auto&& attr : node.attr())
Khalique's avatar
Khalique committed
795
796
797
798
799
800
        {
            result[attr.first] = attr.second;
        }
        return result;
    }

Khalique's avatar
Khalique committed
801
    static std::string get_name(const tensorflow::NodeDef& node) { return node.name(); }
Khalique's avatar
Khalique committed
802

Khalique's avatar
Khalique committed
803
804
    static node_map get_nodes(const tensorflow::GraphDef& graph,
                              std::vector<tensorflow::NodeDef>& input_nodes)
Khalique's avatar
Khalique committed
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
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
862
863
    {
        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
864
865
866
        case tensorflow::DataType::DT_UINT64:
            shape_type = shape::uint64_type;
            break;
Khalique's avatar
Khalique committed
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893

        // 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
894
895
896
897
        }
        return shape_type;
    }

Khalique's avatar
Khalique committed
898
    static literal parse_tensor(const tensorflow::TensorProto& t)
Khalique's avatar
Khalique committed
899
900
    {
        std::vector<size_t> dims = parse_dims(t.tensor_shape());
901
        size_t shape_size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<size_t>());
Khalique's avatar
Khalique committed
902
903
        if(!t.tensor_content().empty()) // has raw data
        {
Khalique's avatar
Khalique committed
904
            const std::string& s = t.tensor_content();
Khalique's avatar
Khalique committed
905
906
907
            switch(t.dtype())
            {
            case tensorflow::DataType::DT_INVALID: throw std::runtime_error("");
Khalique's avatar
Khalique committed
908
909
            case tensorflow::DataType::DT_FLOAT:
                return literal{{shape::float_type, dims}, s.data()};
Khalique's avatar
Khalique committed
910
            case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
911
            case tensorflow::DataType::DT_INT8: return literal{{shape::int8_type, dims}, s.data()};
Khalique's avatar
Khalique committed
912
            case tensorflow::DataType::DT_UINT16:
913
                return literal{{shape::uint16_type, dims}, s.data()};
Khalique's avatar
Khalique committed
914
            case tensorflow::DataType::DT_INT16:
915
                return literal{{shape::int16_type, dims}, s.data()};
Khalique's avatar
Khalique committed
916
917
918
919
            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
920
            case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
921
            case tensorflow::DataType::DT_BOOL: return literal{{shape::int8_type, dims}, s.data()};
Khalique's avatar
Khalique committed
922
            case tensorflow::DataType::DT_HALF: return literal{{shape::half_type, dims}, s.data()};
Khalique's avatar
Khalique committed
923
924
            case tensorflow::DataType::DT_DOUBLE:
                return literal{{shape::double_type, dims}, s.data()};
Khalique's avatar
Khalique committed
925
926
927
928
            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
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
            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
960
961
962
963
            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
964
965
966
967
968
969
970
            }
            MIGRAPHX_THROW("Invalid tensor type");
        }
        switch(t.dtype())
        {
        case tensorflow::DataType::DT_INVALID: throw std::runtime_error("");
        case tensorflow::DataType::DT_FLOAT:
Khalique's avatar
Khalique committed
971
972
            return create_literal(
                shape::float_type, dims, get_data_vals(t.float_val(), shape_size));
Khalique's avatar
Khalique committed
973
974
        case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT8:
975
            return create_literal(shape::int8_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
976
        case tensorflow::DataType::DT_UINT16:
977
            return create_literal(shape::uint16_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
978
        case tensorflow::DataType::DT_INT16:
979
            return create_literal(shape::int16_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
980
        case tensorflow::DataType::DT_INT32:
981
            return create_literal(shape::int32_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
982
        case tensorflow::DataType::DT_INT64:
Khalique's avatar
Khalique committed
983
984
            return create_literal(
                shape::int64_type, dims, get_data_vals(t.int64_val(), shape_size));
Khalique's avatar
Khalique committed
985
986
        case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
        case tensorflow::DataType::DT_BOOL:
987
            return create_literal(shape::int32_type, dims, get_data_vals(t.bool_val(), shape_size));
Khalique's avatar
Khalique committed
988
        case tensorflow::DataType::DT_HALF:
Khalique's avatar
Khalique committed
989
        {
990
991
            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
992
993
994
995
996
            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); });
997
            return create_literal(shape::half_type, dims, data_half);
Khalique's avatar
Khalique committed
998
        }
Khalique's avatar
Khalique committed
999
        case tensorflow::DataType::DT_DOUBLE:
Khalique's avatar
Khalique committed
1000
            return literal{{shape::double_type, dims}, get_data_vals(t.double_val(), shape_size)};
Khalique's avatar
Khalique committed
1001
1002
1003
1004
        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
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
        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
1036
1037
1038
1039
        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
1040
1041
1042
1043
        }
        MIGRAPHX_THROW("Invalid tensor type");
    }

1044
    template <class T>
Khalique's avatar
Khalique committed
1045
    static std::vector<T> get_data_vals(const google::protobuf::RepeatedField<T>& data,
1046
                                        const size_t& shape_size)
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
    {
        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
1059
1060
1061
1062
    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
1063
1064
1065
        std::transform(input_dims.begin(),
                       input_dims.end(),
                       std::back_inserter(dims),
Paul's avatar
Paul committed
1066
                       [](const tensorflow::TensorShapeProto_Dim& dim) { return dim.size(); });
Khalique's avatar
Khalique committed
1067
1068
        return dims;
    }
1069
1070

    template <class T>
Khalique's avatar
Khalique committed
1071
    static literal
1072
    create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, std::vector<T> data)
1073
    {
Khalique's avatar
Khalique committed
1074
        // assume if explicit value is mentioned in protobuf and dim size <= 1, treat as scalar
1075
        if(dims.empty() or (dims.size() == 1 and dims.front() == 1))
1076
            return literal{{shape_type}, data};
1077
1078
        return literal{{shape_type, dims}, data};
    }
Khalique's avatar
Khalique committed
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
};

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