tf.cpp 39 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
        add_binary_op("Mul", op::mul{});
Khalique's avatar
Khalique committed
114

115
        add_mem_op("AvgPool", &tf_parser::parse_pooling);
116
117
        add_mem_op("BiasAdd", &tf_parser::parse_biasadd);
        add_mem_op("ConcatV2", &tf_parser::parse_concat);
Khalique's avatar
Khalique committed
118
119
120
        add_mem_op("Const", &tf_parser::parse_constant);
        add_mem_op("Conv2D", &tf_parser::parse_conv);
        add_mem_op("FusedBatchNorm", &tf_parser::parse_batchnorm);
121
        add_mem_op("MatMul", &tf_parser::parse_matmul);
122
        add_mem_op("MaxPool", &tf_parser::parse_pooling);
Khalique's avatar
Khalique committed
123
        add_mem_op("Mean", &tf_parser::parse_mean);
Khalique's avatar
Khalique committed
124
        add_mem_op("Pack", &tf_parser::parse_pack);
Khalique's avatar
Khalique committed
125
        add_mem_op("Pad", &tf_parser::parse_pad);
126
127
128
        add_mem_op("Reshape", &tf_parser::parse_reshape);
        add_mem_op("Softmax", &tf_parser::parse_softmax);
        add_mem_op("Squeeze", &tf_parser::parse_squeeze);
129
        add_mem_op("StridedSlice", &tf_parser::parse_stridedslice);
Khalique's avatar
Khalique committed
130
131
    }

132
133
134
135
136
137
138
139
140
141
142
143
144
    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
145
146
147
    template <class F>
    void add_mem_op(std::string name, F f)
    {
148
        add_op(name, [=](auto&&... xs) {
Khalique's avatar
Khalique committed
149
150
151
152
153
154
155
            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
156
        add_op(name, [this, x](const attribute_map& attributes, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
157
158
            if(args.size() != 2)
                MIGRAPHX_THROW("binary operators should have 2 operands");
159
160
161
162
163
            auto l0 = args[1];
            if(contains(attributes, "data_format"))
            {
                if(is_nhwc)
                {
Khalique's avatar
Khalique committed
164
                    l0 = prog.add_instruction(op::transpose{{0, 3, 1, 2}}, args[1]);
165
166
167
                }
            }
            return add_broadcastable_binary_op(args[0], l0, x);
Khalique's avatar
Khalique committed
168
169
170
171
172
173
        });
    }

    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
Khalique's avatar
Khalique committed
174
        if(arg0->get_shape().lens() != arg1->get_shape().lens())
Khalique's avatar
Khalique committed
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
        {
            // 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
190
191
            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
192
193
194
195
196

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

197
            std::vector<size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
            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
218
        add_op(name, [this, x](const attribute_map&, std::vector<instruction_ref> args) {
Khalique's avatar
Khalique committed
219
220
221
222
223
224
225
            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
226
227
228
        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
229
230
231
232
233
234
235
236
        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));
    }

237
    instruction_ref
Khalique's avatar
Khalique committed
238
    parse_biasadd(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
239
    {
240
        uint64_t axis = 1; // assume output of previous layer is in NCHW (broadcast on channel)
Shucai Xiao's avatar
Shucai Xiao committed
241
        auto l0 = prog.add_instruction(op::broadcast{axis, args[0]->get_shape().lens()}, args[1]);
242
        return prog.add_instruction(op::add{}, args[0], l0);
243
244
    }

Khalique's avatar
Khalique committed
245
246
247
248
    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        // get index for axis within args
249
        size_t axis_idx = attributes.at("N").i();
Khalique's avatar
Khalique committed
250
        size_t axis     = parse_axis(args[axis_idx]->eval().at<int64_t>());
Khalique's avatar
Khalique committed
251
        op::concat op{axis};
252
        // return only first N arguments (assuming last index is the axis value)
Khalique's avatar
Khalique committed
253
        return prog.add_instruction(
254
            op, std::vector<instruction_ref>(args.begin(), args.begin() + args.size() - 1));
Khalique's avatar
Khalique committed
255
256
257
258
259
260
    }

    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
    {
Khalique's avatar
Khalique committed
261
262
        literal v       = parse_tensor(attributes.at("value").tensor());
        auto l0         = prog.add_literal(v);
263
264
265
266
267
268
        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
269
        }
270
        return l0;
Khalique's avatar
Khalique committed
271
272
273
274
275
276
277
278
279
280
281
    }

    instruction_ref
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::convolution op;
        if(contains(attributes, "padding"))
        {
            const std::string& pad_mode = attributes.at("padding").s();
            if(pad_mode.find("SAME") != std::string::npos)
            {
282
                op.padding_mode = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
283
            }
Khalique's avatar
Khalique committed
284
            else if(pad_mode.find("EXPLICIT") != std::string::npos)
Khalique's avatar
Khalique committed
285
            {
286
                std::vector<size_t> padding;
287
                copy(attributes.at("explicit_paddings").list().i(), std::back_inserter(padding));
Khalique's avatar
Khalique committed
288
289
290
291
292
293
294
295
296
297
298
299
300
301
                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];
            }
        }
        if(contains(attributes, "strides"))
        {
302
            std::vector<size_t> stride;
303
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
304
            reorder_data(stride);
305
306
            if(stride.size() != 4)
            {
307
                MIGRAPHX_THROW("strides should have 4 values");
308
            }
309
310
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
Khalique's avatar
Khalique committed
311
312
313
        }
        if(contains(attributes, "dilations"))
        {
314
            std::vector<size_t> dilation;
315
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
316
            reorder_data(dilation);
317
318
319
320
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
321
322
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
Khalique's avatar
Khalique committed
323
        }
Khalique's avatar
Khalique committed
324
        auto weights = args[1];
325
        // check if weights are from a constant
Khalique's avatar
Khalique committed
326
327

        if(weights->name() != "@param")
328
        {
Khalique's avatar
Khalique committed
329
330
331
332
333
334
335
336
            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]);
            }
337
        }
Khalique's avatar
Khalique committed
338

Khalique's avatar
Khalique committed
339
        return prog.add_instruction(op, {args[0], weights});
Khalique's avatar
Khalique committed
340
341
    }

Khalique's avatar
Khalique committed
342
343
    instruction_ref
    parse_matmul(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
344
345
346
    {
        bool transa = false;
        bool transb = false;
Khalique's avatar
Khalique committed
347

348
349
350
351
352
353
354
355
356
357
358
359
        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
360
        std::iter_swap(perm.end() - 1, perm.end() - 2);
361
362
363
364
365
366
367

        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
368
369
    instruction_ref
    parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
370
    {
Khalique's avatar
Khalique committed
371
372

        auto axes      = parse_axes(args[1]->eval().get<int32_t>().to_vector());
Khalique's avatar
Khalique committed
373
        bool keep_dims = attributes.at("keep_dims").b();
Khalique's avatar
Khalique committed
374
        std::vector<int32_t> hw_axes{2, 3};
Khalique's avatar
Khalique committed
375
376
        // check if conditions for GlobalAvgPool are met
        if(axes == hw_axes and args[0]->get_shape().lens().size() == 4)
Khalique's avatar
Khalique committed
377
378
379
380
381
        {
            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];
Khalique's avatar
Khalique committed
382
383
384
385
386
            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
387
388
389
390
        }
        MIGRAPHX_THROW("MIGraphX does not support mean outside of GlobalAvgPool transformation");
    }

Khalique's avatar
Khalique committed
391
392
393
    instruction_ref parse_pack(const std::string&,
                               const attribute_map& attributes,
                               std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
394
395
396
397
398
399
    {
        // reinterpret as unsqueeze with concat
        std::vector<instruction_ref> unsqueezed_args;
        int64_t axis = 0;
        if(contains(attributes, "axis"))
            axis = attributes.at("axis").i();
400
401
402
        size_t input_size = args.front()->get_shape().lens().size();
        if(axis > input_size)
        {
Khalique's avatar
Khalique committed
403
404
            MIGRAPHX_THROW("TF_PARSER: axis value of " + to_string(axis) +
                           " must be smaller than input size " + to_string(input_size));
405
406
407
408
409
        }
        // check if input arg needs axis to be converted to NCHW
        if(input_size >= 4)
            axis = parse_axis(axis);

Khalique's avatar
Khalique committed
410
411
412
413
414
        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
415
416
417
        return prog.add_instruction(op::concat{static_cast<size_t>(axis)}, unsqueezed_args);
    }

Khalique's avatar
Khalique committed
418
419
420
421
422
    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
423
424
        // 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
425
426
        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
427
        for(size_t i = 0; i < 2 * ndims; i += 2)
Khalique's avatar
Khalique committed
428
        {
Khalique's avatar
Khalique committed
429
430
            pad_per_dim[i / 2].first  = tf_padding[i];
            pad_per_dim[i / 2].second = tf_padding[i + 1];
Khalique's avatar
Khalique committed
431
432
433
434
        }
        reorder_data(pad_per_dim);

        op::pad op;
Khalique's avatar
Khalique committed
435
436
        std::vector<int64_t> pads(ndims * 2);
        for(size_t i = 0; i < ndims; i++)
Khalique's avatar
Khalique committed
437
        {
Khalique's avatar
Khalique committed
438
439
            pads[i]         = pad_per_dim[i].first;
            pads[i + ndims] = pad_per_dim[i].second;
Khalique's avatar
Khalique committed
440
441
442
443
444
        }
        op.pads = pads;
        return prog.add_instruction(op, args.front());
    }

445
446
447
448
449
    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
450

451
452
453
454
455
456
457
458
459
460
461
462
463
464
        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"))
        {
465
            std::vector<size_t> stride;
466
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
467
            reorder_data(stride);
468
469
470
471
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
472
473
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
474
475
476
        }
        if(contains(attributes, "ksize"))
        {
477
            std::vector<size_t> ksize;
478
            copy(attributes.at("ksize").list().i(), std::back_inserter(ksize));
479
            reorder_data(ksize);
480
481
482
            if(ksize.size() != 4)
            {
                MIGRAPHX_THROW("ksize should have 4 values");
Khalique's avatar
Khalique committed
483
            }
484
485
            op.lengths[0] = ksize[2];
            op.lengths[1] = ksize[3];
486
        }
487
        return prog.add_instruction(op, args[0]);
488
    }
Khalique's avatar
Khalique committed
489

490
    instruction_ref
Khalique's avatar
Khalique committed
491
    parse_reshape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
492
493
494
495
    {
        op::reshape op;
        if(args.size() != 2)
            MIGRAPHX_THROW("reshape needs 2 arguments (input, new_shape)");
Khalique's avatar
Khalique committed
496
        auto s = args[1]->eval();
497
498
499
500
        s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
        return prog.add_instruction(op, args[0]);
    }

Khalique's avatar
Khalique committed
501
502
503
504
505
506
507
508
509
    void parse_from(std::istream& is)
    {
        tensorflow::GraphDef graph;
        if(graph.ParseFromIstream(&is))
        {
            this->parse_graph(graph);
        }
        else
        {
510
            throw std::runtime_error("Failed reading tf file");
Khalique's avatar
Khalique committed
511
512
513
        }
    }

514
515
516
517
518
519
520
521
522
523
    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
524
525
526
    instruction_ref parse_squeeze(const std::string&,
                                  const attribute_map& attributes,
                                  std::vector<instruction_ref> args)
527
528
    {
        op::squeeze op;
529
        auto axes = parse_axes(attributes, "squeeze_dims");
530
        copy(axes, std::back_inserter(op.axes));
531
        auto args0_dims = args[0]->get_shape().lens();
532
533
        if(op.axes.empty()) // no squeeze_dims provided, remove any dim that equals 1
        {
534
            for(size_t i = 0; i < args0_dims.size(); i++)
535
            {
536
                if(args0_dims.at(i) == 1)
537
538
539
540
                {
                    op.axes.push_back(i);
                }
            }
541
        }
542
        return prog.add_instruction(op, args[0]);
543
544
    }

Khalique's avatar
Khalique committed
545
546
547
    instruction_ref parse_stridedslice(const std::string&,
                                       const attribute_map& attributes,
                                       std::vector<instruction_ref> args)
548
549
    {
        op::slice op;
Khalique's avatar
Khalique committed
550
551
552
553
554
555
556
557
        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);
        }
558

Khalique's avatar
Khalique committed
559
560
561
562
        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);
563
        uint32_t shrink_axis_mask = 0;
Khalique's avatar
Khalique committed
564
        uint32_t bitwise_compare  = 1;
565
566
567
        std::vector<int64_t> squeeze_axes;

        if(contains(attributes, "shrink_axis_mask"))
568
            shrink_axis_mask = static_cast<uint32_t>(attributes.at("shrink_axis_mask").i());
569

Khalique's avatar
Khalique committed
570
        for(size_t i = 0; i < num_axes; i++)
571
        {
572
            // the LSB corresponds to axis 0 when determining which axes to squeeze
Khalique's avatar
Khalique committed
573
            if(((shrink_axis_mask >> i) & bitwise_compare) == 1)
574
575
                squeeze_axes.push_back(i);
        }
Khalique's avatar
Khalique committed
576
577
578
579
580
        if(num_axes >= 4)
        {
            squeeze_axes = parse_axes(squeeze_axes);
        }

581
582
583
584
        auto l0 = prog.add_instruction(op, args[0]);
        return prog.add_instruction(op::squeeze{squeeze_axes}, l0);
    }

Khalique's avatar
Khalique committed
585
586
587
588
589
    void parse_graph(const tensorflow::GraphDef& graph)
    {
        nodes = get_nodes(graph, input_nodes);
        for(auto&& input : input_nodes)
        {
Khalique's avatar
Khalique committed
590
            const std::string& name   = input.name();
Khalique's avatar
Khalique committed
591
            attribute_map input_attrs = get_attributes(input);
Khalique's avatar
Khalique committed
592
593
            shape::type_t shape_type  = parse_type(input_attrs.at("dtype").type());
            std::vector<size_t> dims  = parse_dims(input_attrs.at("shape").shape());
594
            if(is_nhwc and dims.size() >= 4)
595
            {
596
                reorder_data(dims);
597
            }
Khalique's avatar
Khalique committed
598
599
            shape s            = shape{shape_type, dims};
            instructions[name] = prog.add_parameter(name, s);
Khalique's avatar
Khalique committed
600
601
602
        }
        for(auto&& p : nodes)
        {
603
            this->parse_node(p.first);
Khalique's avatar
Khalique committed
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
        }
    }

    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)
            {
630
                instructions[name] = prog.add_instruction(op::unknown{node.op()}, args);
Khalique's avatar
Khalique committed
631
632
633
634
635
636
637
638
639
640
641
            }
            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
642
        for(auto&& attr : node.attr())
Khalique's avatar
Khalique committed
643
644
645
646
647
648
        {
            result[attr.first] = attr.second;
        }
        return result;
    }

Khalique's avatar
Khalique committed
649
    static std::string get_name(const tensorflow::NodeDef& node) { return node.name(); }
Khalique's avatar
Khalique committed
650

Khalique's avatar
Khalique committed
651
652
    static node_map get_nodes(const tensorflow::GraphDef& graph,
                              std::vector<tensorflow::NodeDef>& input_nodes)
Khalique's avatar
Khalique committed
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
    {
        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
712
713
714
        case tensorflow::DataType::DT_UINT64:
            shape_type = shape::uint64_type;
            break;
Khalique's avatar
Khalique committed
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741

        // 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
742
743
744
745
        }
        return shape_type;
    }

Khalique's avatar
Khalique committed
746
    static literal parse_tensor(const tensorflow::TensorProto& t)
Khalique's avatar
Khalique committed
747
748
    {
        std::vector<size_t> dims = parse_dims(t.tensor_shape());
749
        size_t shape_size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<size_t>());
Khalique's avatar
Khalique committed
750
751
        if(!t.tensor_content().empty()) // has raw data
        {
Khalique's avatar
Khalique committed
752
            const std::string& s = t.tensor_content();
Khalique's avatar
Khalique committed
753
754
755
            switch(t.dtype())
            {
            case tensorflow::DataType::DT_INVALID: throw std::runtime_error("");
Khalique's avatar
Khalique committed
756
757
            case tensorflow::DataType::DT_FLOAT:
                return literal{{shape::float_type, dims}, s.data()};
Khalique's avatar
Khalique committed
758
            case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
759
            case tensorflow::DataType::DT_INT8: return literal{{shape::int8_type, dims}, s.data()};
Khalique's avatar
Khalique committed
760
            case tensorflow::DataType::DT_UINT16:
761
                return literal{{shape::uint16_type, dims}, s.data()};
Khalique's avatar
Khalique committed
762
            case tensorflow::DataType::DT_INT16:
763
                return literal{{shape::int16_type, dims}, s.data()};
Khalique's avatar
Khalique committed
764
765
766
767
            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
768
            case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
769
            case tensorflow::DataType::DT_BOOL: return literal{{shape::int8_type, dims}, s.data()};
Khalique's avatar
Khalique committed
770
            case tensorflow::DataType::DT_HALF: return literal{{shape::half_type, dims}, s.data()};
Khalique's avatar
Khalique committed
771
772
            case tensorflow::DataType::DT_DOUBLE:
                return literal{{shape::double_type, dims}, s.data()};
Khalique's avatar
Khalique committed
773
774
775
776
            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
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
            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
808
809
810
811
            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
812
813
814
815
816
817
818
            }
            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
819
820
            return create_literal(
                shape::float_type, dims, get_data_vals(t.float_val(), shape_size));
Khalique's avatar
Khalique committed
821
822
        case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT8:
823
            return create_literal(shape::int8_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
824
        case tensorflow::DataType::DT_UINT16:
825
            return create_literal(shape::uint16_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
826
        case tensorflow::DataType::DT_INT16:
827
            return create_literal(shape::int16_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
828
        case tensorflow::DataType::DT_INT32:
829
            return create_literal(shape::int32_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
830
        case tensorflow::DataType::DT_INT64:
Khalique's avatar
Khalique committed
831
832
            return create_literal(
                shape::int64_type, dims, get_data_vals(t.int64_val(), shape_size));
Khalique's avatar
Khalique committed
833
834
        case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
        case tensorflow::DataType::DT_BOOL:
835
            return create_literal(shape::int32_type, dims, get_data_vals(t.bool_val(), shape_size));
Khalique's avatar
Khalique committed
836
        case tensorflow::DataType::DT_HALF:
Khalique's avatar
Khalique committed
837
        {
838
839
            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
840
841
842
843
844
            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); });
845
            return create_literal(shape::half_type, dims, data_half);
Khalique's avatar
Khalique committed
846
        }
Khalique's avatar
Khalique committed
847
        case tensorflow::DataType::DT_DOUBLE:
Khalique's avatar
Khalique committed
848
            return literal{{shape::double_type, dims}, get_data_vals(t.double_val(), shape_size)};
Khalique's avatar
Khalique committed
849
850
851
852
        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
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
        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
884
885
886
887
        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
888
889
890
891
        }
        MIGRAPHX_THROW("Invalid tensor type");
    }

892
    template <class T>
Khalique's avatar
Khalique committed
893
    static std::vector<T> get_data_vals(const google::protobuf::RepeatedField<T>& data,
894
                                        const size_t& shape_size)
895
896
897
898
899
900
901
902
903
904
905
906
    {
        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
907
908
909
910
    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
911
912
913
        std::transform(input_dims.begin(),
                       input_dims.end(),
                       std::back_inserter(dims),
Paul's avatar
Paul committed
914
                       [](const tensorflow::TensorShapeProto_Dim& dim) { return dim.size(); });
Khalique's avatar
Khalique committed
915
916
        return dims;
    }
917
918

    template <class T>
Khalique's avatar
Khalique committed
919
    static literal
920
    create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, std::vector<T> data)
921
    {
Khalique's avatar
Khalique committed
922
        // assume if explicit value is mentioned in protobuf and dim size <= 1, treat as scalar
923
        if(dims.empty() or (dims.size() == 1 and dims.front() == 1))
924
            return literal{{shape_type}, data};
925
926
        return literal{{shape_type, dims}, data};
    }
Khalique's avatar
Khalique committed
927
928
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
};

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