tf.cpp 51.6 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>;
Paul's avatar
Paul 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;

Paul's avatar
Paul committed
40
    bool should_transpose(instruction_ref ins) const
Paul's avatar
Paul committed
41
42
43
44
45
46
    {
        return is_nhwc and ins->get_shape().lens().size() == 4;
    }

    instruction_ref to_nhwc(instruction_ref ins)
    {
Paul's avatar
Paul committed
47
        if(should_transpose(ins))
Paul's avatar
Paul committed
48
49
50
51
52
53
            return prog.add_instruction(op::transpose{{0, 2, 3, 1}}, ins);
        return ins;
    }

    instruction_ref to_nchw(instruction_ref ins)
    {
Paul's avatar
Paul committed
54
        if(should_transpose(ins))
Paul's avatar
Paul committed
55
56
57
58
59
60
            return prog.add_instruction(op::transpose{{0, 3, 1, 2}}, ins);
        return ins;
    }

    instruction_ref to_kcxy(instruction_ref ins)
    {
Paul's avatar
Paul committed
61
        if(should_transpose(ins))
Paul's avatar
Paul committed
62
63
64
65
66
67
            return prog.add_instruction(op::transpose{{3, 2, 0, 1}}, ins);
        return ins;
    }

    instruction_ref make_contiguous(instruction_ref ins)
    {
Paul's avatar
Paul committed
68
        if(ins->get_shape().standard())
Paul's avatar
Paul committed
69
70
71
72
73
74
75
76
            return ins;
        else
            return prog.add_instruction(op::contiguous{}, ins);
    }

    std::vector<instruction_ref> to_nchw(const std::vector<instruction_ref>& args)
    {
        std::vector<instruction_ref> result(args.size());
Paul's avatar
Paul committed
77
        std::transform(
Paul's avatar
Paul committed
78
            args.begin(), args.end(), result.begin(), [&](auto ins) { return this->to_nchw(ins); });
Paul's avatar
Paul committed
79
80
81
        return result;
    }

Khalique's avatar
Khalique committed
82
    std::vector<size_t>
83
    parse_axes(const attribute_map& attributes, const std::string& s, const size_t num_dims) const
84
    {
85
86
87
        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
88
        if(is_nhwc)
89
        {
Khalique's avatar
Khalique committed
90
            std::transform(axes.begin(), axes.end(), axes.begin(), [&](size_t axis) {
Khalique's avatar
Khalique committed
91
                return parse_axis(axis, num_dims);
Khalique's avatar
Khalique committed
92
            });
93
94
95
96
        }
        return axes;
    }

Khalique's avatar
Khalique committed
97
    template <class T>
98
    std::vector<T> parse_axes(std::vector<T> axes, const size_t num_dims) const
Khalique's avatar
Khalique committed
99
100
101
    {
        if(is_nhwc)
        {
102
            std::vector<T> new_axes;
Khalique's avatar
Khalique committed
103
104
105
            std::transform(axes.begin(),
                           axes.end(),
                           std::back_inserter(new_axes),
Khalique's avatar
Khalique committed
106
                           [&](size_t axis) { return parse_axis(axis, num_dims); });
107
            return new_axes;
Khalique's avatar
Khalique committed
108
        }
109
        return axes;
Khalique's avatar
Khalique committed
110
111
    }

Khalique's avatar
Khalique committed
112
113
114
    // 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.
115
    template <class T>
116
    void reorder_data(std::vector<T>& prev_data) const
117
118
    {
        std::vector<T> new_data(prev_data.size());
119
        for(size_t i = 0; i < new_data.size(); i++)
120
        {
Khalique's avatar
Khalique committed
121
            auto new_idx         = parse_axis(i, new_data.size());
122
            new_data.at(new_idx) = prev_data.at(i);
123
        }
124
125
126
127
        prev_data = new_data;
    }

    template <class T>
128
    T parse_axis(const T& dim, const size_t num_dims) const
129
    {
Khalique's avatar
Khalique committed
130
        T new_dim = dim;
Khalique's avatar
Khalique committed
131
        if(is_nhwc and num_dims >= 4)
132
133
134
        {
            switch(dim)
            {
Khalique's avatar
Khalique committed
135
136
137
138
139
            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;
140
141
            }
        }
Khalique's avatar
Khalique committed
142
        return new_dim;
143
144
    }

145
146
147
148
149
150
151
    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
152
153
154
155
    tf_parser()
    {
        add_generic_op("Identity", op::identity{});
        add_generic_op("Relu", op::relu{});
Khalique's avatar
Khalique committed
156
        add_generic_op("Relu6", op::clip{6.0, 0.0});
Khalique's avatar
Khalique committed
157
        add_generic_op("Rsqrt", op::rsqrt{});
Khalique's avatar
Khalique committed
158
        add_generic_op("Tanh", op::tanh{});
Khalique's avatar
Khalique committed
159
        add_generic_op("StopGradient", op::identity{});
Khalique's avatar
Khalique committed
160

161
        add_binary_op("Add", op::add{});
Khalique's avatar
Khalique committed
162
        add_binary_op("Mul", op::mul{});
Khalique's avatar
Khalique committed
163
        add_binary_op("Pow", op::pow{});
Khalique's avatar
Khalique committed
164
        add_binary_op("SquaredDifference", op::sqdiff{});
Khalique's avatar
Khalique committed
165
        add_binary_op("Sub", op::sub{});
Khalique's avatar
Khalique committed
166

167
        add_mem_op("AvgPool", &tf_parser::parse_pooling);
Khalique's avatar
Khalique committed
168
        add_mem_op("BatchMatMul", &tf_parser::parse_matmul, false);
169
        add_mem_op("BiasAdd", &tf_parser::parse_biasadd);
Khalique's avatar
Khalique committed
170
        add_mem_op("Cast", &tf_parser::parse_cast, false);
Paul's avatar
Paul committed
171
        add_mem_op("ConcatV2", &tf_parser::parse_concat, false);
Khalique's avatar
Khalique committed
172
        add_mem_op("Const", &tf_parser::parse_constant);
Paul's avatar
Paul committed
173
        add_mem_op("Conv2D", &tf_parser::parse_conv);
Paul's avatar
Paul committed
174
        add_mem_op("DepthwiseConv2dNative", &tf_parser::parse_depthwiseconv);
175
        add_mem_op("ExpandDims", &tf_parser::parse_expanddims, false);
Khalique's avatar
Khalique committed
176
        add_mem_op("FusedBatchNorm", &tf_parser::parse_batchnorm);
Khalique's avatar
Khalique committed
177
        add_mem_op("GatherV2", &tf_parser::parse_gather, false);
Paul's avatar
Paul committed
178
        add_mem_op("MatMul", &tf_parser::parse_matmul, false);
179
        add_mem_op("MaxPool", &tf_parser::parse_pooling);
Khalique's avatar
Khalique committed
180
        add_mem_op("Mean", &tf_parser::parse_mean);
Khalique's avatar
Khalique committed
181
        add_mem_op("OneHot", &tf_parser::parse_onehot, false);
Paul's avatar
Paul committed
182
        add_mem_op("Pack", &tf_parser::parse_pack, false);
Paul's avatar
Paul committed
183
        add_mem_op("Pad", &tf_parser::parse_pad);
Paul's avatar
Paul committed
184
        add_mem_op("Reshape", &tf_parser::parse_reshape, false);
Khalique's avatar
Khalique committed
185
        add_mem_op("Slice", &tf_parser::parse_slice, false);
186
        add_mem_op("Softmax", &tf_parser::parse_softmax);
Paul's avatar
Paul committed
187
        add_mem_op("Squeeze", &tf_parser::parse_squeeze, false);
188
        add_mem_op("StridedSlice", &tf_parser::parse_stridedslice, false);
Khalique's avatar
Khalique committed
189
        add_mem_op("Transpose", &tf_parser::parse_transpose, false);
Khalique's avatar
Khalique committed
190
191
    }

192
    template <class F>
Paul's avatar
Paul committed
193
    void add_op(std::string name, F f, bool transpose = true)
194
    {
Paul's avatar
Paul committed
195
        if(transpose)
Paul's avatar
Paul committed
196
        {
Paul's avatar
Paul committed
197
198
            ops.emplace(name,
                        op_func{[=](const attribute_map& attributes,
Paul's avatar
Paul committed
199
                                    const std::vector<instruction_ref>& args) -> instruction_ref {
Paul's avatar
Paul committed
200
201
                            return to_nhwc(f(attributes, to_nchw(args)));
                        }});
Paul's avatar
Paul committed
202
203
204
205
206
        }
        else
        {
            ops.emplace(name, f);
        }
207
208
    }

Khalique's avatar
Khalique committed
209
    template <class F>
Paul's avatar
Paul committed
210
    void add_mem_op(std::string name, F f, bool transpose = true)
Khalique's avatar
Khalique committed
211
    {
Paul's avatar
Paul committed
212
213
214
215
216
        add_op(name,
               [=](auto&&... xs) {
                   return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
               },
               transpose);
Khalique's avatar
Khalique committed
217
218
219
220
221
    }

    template <class T>
    void add_binary_op(std::string name, T x)
    {
Paul's avatar
Paul committed
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
        add_op(name,
               [this, x](const attribute_map&, std::vector<instruction_ref> args) {
                   if(args.size() != 2)
                       MIGRAPHX_THROW("binary operators should have 2 operands");
                   // TODO
                   // if(contains(attributes, "data_format"))
                   // {
                   //     if(is_nhwc)
                   //     {
                   //         l0 = prog.add_instruction(op::transpose{{0, 3, 1, 2}}, args[1]);
                   //     }
                   // }
                   return add_broadcastable_binary_op(args[0], args[1], x);
               },
               false);
Khalique's avatar
Khalique committed
237
238
239
240
241
    }

    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
Khalique's avatar
Khalique committed
242
        if(arg0->get_shape().lens() != arg1->get_shape().lens())
Khalique's avatar
Khalique committed
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
        {
            // 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
258
259
            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
260
261
262
263
264

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

265
            std::vector<size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
266
267
268
269
270
271
272
273
274
            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);
Paul's avatar
Paul committed
275
            return to_nhwc(prog.add_instruction(x, to_nchw(l0), to_nchw(l1)));
Khalique's avatar
Khalique committed
276
277
278
        }
        else
        {
Paul's avatar
Paul committed
279
            return to_nhwc(prog.add_instruction(x, {to_nchw(arg0), to_nchw(arg1)}));
Khalique's avatar
Khalique committed
280
281
282
283
        }
    }

    template <class T>
Paul's avatar
Paul committed
284
    void add_generic_op(std::string name, T x, bool transpose = true)
Khalique's avatar
Khalique committed
285
    {
Paul's avatar
Paul committed
286
287
288
289
290
        add_op(name,
               [this, x](const attribute_map&, std::vector<instruction_ref> args) {
                   return prog.add_instruction(x, args);
               },
               transpose);
Khalique's avatar
Khalique committed
291
292
293
294
295
    }

    instruction_ref
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
296
297
298
        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
299
300
301
302
303
304
305
306
        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));
    }

307
    instruction_ref
Khalique's avatar
Khalique committed
308
    parse_biasadd(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
309
    {
310
        uint64_t axis = 1; // assume output of previous layer is in NCHW (broadcast on channel)
Shucai Xiao's avatar
Shucai Xiao committed
311
        auto l0 = prog.add_instruction(op::broadcast{axis, args[0]->get_shape().lens()}, args[1]);
312
        return prog.add_instruction(op::add{}, args[0], l0);
313
314
    }

Khalique's avatar
Khalique committed
315
316
317
318
319
320
321
    instruction_ref
    parse_cast(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        shape::type_t type = parse_type(attributes.at("DstT").type());
        return prog.add_instruction(op::convert{type}, std::move(args));
    }

Khalique's avatar
Khalique committed
322
323
324
325
    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        // get index for axis within args
326
        size_t axis_idx = attributes.at("N").i();
Paul's avatar
Paul committed
327
        size_t axis     = args[axis_idx]->eval().at<int64_t>();
Khalique's avatar
Khalique committed
328
        op::concat op{axis};
329
        // return only first N arguments (assuming last index is the axis value)
Paul's avatar
Paul committed
330
331
        return prog.add_instruction(
            op, std::vector<instruction_ref>(args.begin(), args.begin() + args.size() - 1));
Khalique's avatar
Khalique committed
332
333
334
335
336
337
    }

    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
    {
Paul's avatar
Paul committed
338
        literal v = parse_tensor(attributes.at("value").tensor());
Paul's avatar
Paul committed
339
        return prog.add_literal(v);
Khalique's avatar
Khalique committed
340
341
342
343
344
345
346
347
    }

    instruction_ref
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::convolution op;
        if(contains(attributes, "strides"))
        {
348
            std::vector<size_t> stride;
349
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
350
            reorder_data(stride);
351
352
            if(stride.size() != 4)
            {
353
                MIGRAPHX_THROW("strides should have 4 values");
354
            }
355
356
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
Khalique's avatar
Khalique committed
357
358
359
        }
        if(contains(attributes, "dilations"))
        {
360
            std::vector<size_t> dilation;
361
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
362
            reorder_data(dilation);
363
364
365
366
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
367
368
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
Khalique's avatar
Khalique committed
369
        }
Khalique's avatar
Khalique committed
370

Paul's avatar
Paul committed
371
        auto weights = to_kcxy(args[1]);
Paul's avatar
Paul committed
372
        auto l0      = args[0];
Khalique's avatar
Khalique committed
373
374
375
376
377
        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
378
                op.padding_mode                 = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
379
380
381
                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
382
383

                auto input_dims = l0->get_shape().lens();
Khalique's avatar
Khalique committed
384
385
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
386
387
388
389
390
391
392
393
394
395
396
                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
397
398
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
399
                }
400
401
402
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
403
                op.padding_mode = op::padding_mode_t::valid;
Khalique's avatar
Khalique committed
404
            }
Khalique's avatar
Khalique committed
405
            else if(pad_mode.find("EXPLICIT") != std::string::npos)
Khalique's avatar
Khalique committed
406
            {
407
                std::vector<size_t> padding;
408
                copy(attributes.at("explicit_paddings").list().i(), std::back_inserter(padding));
Khalique's avatar
Khalique committed
409
410
411
412
413
414
415
416
417
418
419
420
                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];
            }
        }
Paul's avatar
Paul committed
421
        return prog.add_instruction(op, {l0, to_kcxy(args[1])});
Khalique's avatar
Khalique committed
422
423
    }

Khalique's avatar
Khalique committed
424
425
426
    instruction_ref parse_depthwiseconv(const std::string&,
                                        attribute_map attributes,
                                        std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
427
428
429
    {
        op::convolution op;
        size_t num_channels = args[0]->get_shape().lens()[1];
Khalique's avatar
Khalique committed
430
        op.group            = num_channels;
Khalique's avatar
Khalique committed
431

Khalique's avatar
Khalique committed
432
433
        if(contains(attributes, "strides"))
        {
434
            std::vector<size_t> stride;
435
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
436
            reorder_data(stride);
437
438
            if(stride.size() != 4)
            {
439
                MIGRAPHX_THROW("strides should have 4 values");
440
            }
441
442
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
Khalique's avatar
Khalique committed
443
        }
Paul's avatar
Paul committed
444
445

        auto weights = to_kcxy(args[1]);
Khalique's avatar
Khalique committed
446
447
        if(contains(attributes, "dilations"))
        {
448
            std::vector<size_t> dilation;
449
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
450
            reorder_data(dilation);
451
452
453
454
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
455
456
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
Khalique's avatar
Khalique committed
457
458
        }

Khalique's avatar
Khalique committed
459
        auto l0 = args[0];
Khalique's avatar
Khalique committed
460
461
462
        if(contains(attributes, "padding"))
        {
            const std::string& pad_mode = attributes.at("padding").s();
Khalique's avatar
Khalique committed
463

Khalique's avatar
Khalique committed
464
465
            if(pad_mode.find("SAME") != std::string::npos)
            {
Khalique's avatar
Khalique committed
466
                op.padding_mode                 = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
467
468
469
470
471
                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
472
473
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
474
475
476
477
478
479
480
481
482
483
484
                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
485
486
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
487
                }
Khalique's avatar
Khalique committed
488
            }
Khalique's avatar
Khalique committed
489
            else if(pad_mode.find("VALID") != std::string::npos)
Khalique's avatar
Khalique committed
490
            {
Khalique's avatar
Khalique committed
491
                op.padding_mode = op::padding_mode_t::valid;
Khalique's avatar
Khalique committed
492
493
            }
        }
Khalique's avatar
Khalique committed
494

Khalique's avatar
Khalique committed
495
496
        std::vector<int64_t> new_weights_shape;
        copy(weights->get_shape().lens(), std::back_inserter(new_weights_shape));
Khalique's avatar
Khalique committed
497
498
499
500

        // 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
501
        int64_t multiplier   = new_weights_shape[0];
Khalique's avatar
Khalique committed
502
503
504
        int64_t out_channels = num_channels * multiplier;
        new_weights_shape[0] = out_channels;
        new_weights_shape[1] = 1;
Paul's avatar
Paul committed
505
        // Make sure weights are contiguous before doing reshape
Paul's avatar
Paul committed
506
507
        auto new_weights =
            prog.add_instruction(op::reshape{new_weights_shape}, make_contiguous(weights));
Khalique's avatar
Khalique committed
508

Khalique's avatar
Khalique committed
509
        return prog.add_instruction(op, {l0, new_weights});
Khalique's avatar
Khalique committed
510
511
    }

Khalique's avatar
Khalique committed
512
513
    instruction_ref
    parse_expanddims(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
514
515
    {
        std::vector<size_t> input_dims = args[0]->get_shape().lens();
Khalique's avatar
Khalique committed
516
        std::vector<int64_t> new_dims(input_dims.begin(), input_dims.end());
Khalique's avatar
Khalique committed
517
        size_t num_dims = input_dims.size();
518
        int32_t dim     = args[1]->eval().at<int32_t>();
Khalique's avatar
Khalique committed
519
520

        if(dim < 0)
Khalique's avatar
Khalique committed
521
522
523
524
525
526
527
528
529
530
        {
            new_dims.insert(new_dims.begin() + (num_dims + dim + 1), 1);
        }
        else
        {
            new_dims.insert(new_dims.begin() + dim, 1);
        }
        return prog.add_instruction(op::reshape{new_dims}, args[0]);
    }

Khalique's avatar
Khalique committed
531
532
533
534
535
536
537
538
    instruction_ref
    parse_gather(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
    {
        int axis = args[2]->eval().at<int32_t>();
        op::gather op{axis};
        return prog.add_instruction(op, {args[0], args[1]});
    }

Khalique's avatar
Khalique committed
539
540
    instruction_ref
    parse_matmul(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
541
542
543
    {
        bool transa = false;
        bool transb = false;
Khalique's avatar
Khalique committed
544

545
546
547
548
549
550
551
552
553
        if(contains(attributes, "transpose_a"))
        {
            transa = attributes.at("transpose_a").b();
        }
        if(contains(attributes, "transpose_b"))
        {
            transb = attributes.at("transpose_a").b();
        }

Khalique's avatar
Khalique committed
554
555
556
557
558
559
560
561
562
        if(contains(attributes, "adj_x"))
        {
            transa = attributes.at("adj_x").b();
        }
        if(contains(attributes, "adj_y"))
        {
            transb = attributes.at("adj_y").b();
        }

563
564
565
        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
566
        std::iter_swap(perm.end() - 1, perm.end() - 2);
567
568
569
570
571
572
573

        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
574
575
    instruction_ref
    parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
576
577
    {
        bool keep_dims = attributes.at("keep_dims").b();
Paul's avatar
Paul committed
578
        std::vector<int32_t> hw_axes{2, 3};
Khalique's avatar
Khalique committed
579
        // check if conditions for GlobalAvgPool are met
Khalique's avatar
Khalique committed
580
        auto lens = args[0]->get_shape().lens();
Khalique's avatar
Khalique committed
581
582
        auto axes = parse_axes(args[1]->eval().get<int32_t>().to_vector(), lens.size());

Khalique's avatar
Khalique committed
583
        if(axes == hw_axes and lens.size() == 4)
Khalique's avatar
Khalique committed
584
585
        {
            op::pooling op{"average"};
Khalique's avatar
Khalique committed
586
587
            op.lengths[0] = lens[2];
            op.lengths[1] = lens[3];
Khalique's avatar
Khalique committed
588
589
590
591
592
            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
593
594
595
596
        }
        MIGRAPHX_THROW("MIGraphX does not support mean outside of GlobalAvgPool transformation");
    }

Khalique's avatar
Khalique committed
597
598
599
    instruction_ref
    parse_onehot(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
600
        auto indices       = args[0]->eval().get<int32_t>().to_vector();
Khalique's avatar
Khalique committed
601
        int depth          = args[1]->eval().at<int32_t>();
Khalique's avatar
Khalique committed
602
        int64_t axis       = -1;
Khalique's avatar
Khalique committed
603
        size_t num_indices = indices.size();
Khalique's avatar
Khalique committed
604
605
606
        float on_value     = args[2]->eval().at<float>();
        float off_value    = args[3]->eval().at<float>();
        if(contains(attributes, "axis"))
Khalique's avatar
Khalique committed
607
608
609
610
611
612
            axis = attributes.at("axis").i();
        if(axis == -1)
        {
            shape s{shape::float_type, {num_indices, static_cast<size_t>(depth)}};
            std::vector<float> output(num_indices * depth);
            std::fill(output.begin(), output.end(), off_value);
Khalique's avatar
Khalique committed
613
            for(size_t i = 0; i < num_indices; i++)
Khalique's avatar
Khalique committed
614
            {
Khalique's avatar
Khalique committed
615
616
                if(indices[i] >= 0 and indices[i] < depth)
                    output.at(depth * i + indices[i]) = on_value;
Khalique's avatar
Khalique committed
617
618
619
620
621
622
            }
            return prog.add_literal(s, output);
        }
        MIGRAPHX_THROW("MIGraphX does not support axis != -1");
    }

Khalique's avatar
Khalique committed
623
624
625
    instruction_ref parse_pack(const std::string&,
                               const attribute_map& attributes,
                               std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
626
627
628
629
630
631
    {
        // reinterpret as unsqueeze with concat
        std::vector<instruction_ref> unsqueezed_args;
        int64_t axis = 0;
        if(contains(attributes, "axis"))
            axis = attributes.at("axis").i();
632
633
634
        size_t input_size = args.front()->get_shape().lens().size();
        if(axis > input_size)
        {
Khalique's avatar
Khalique committed
635
636
            MIGRAPHX_THROW("TF_PARSER: axis value of " + to_string(axis) +
                           " must be smaller than input size " + to_string(input_size));
637
638
        }

Khalique's avatar
Khalique committed
639
640
641
642
643
        std::transform(
            args.begin(),
            args.end(),
            std::back_inserter(unsqueezed_args),
            [&](instruction_ref arg) { return prog.add_instruction(op::unsqueeze{{axis}}, arg); });
Paul's avatar
Paul committed
644
645
        return to_nhwc(
            prog.add_instruction(op::concat{static_cast<size_t>(axis)}, unsqueezed_args));
Khalique's avatar
Khalique committed
646
647
    }

Khalique's avatar
Khalique committed
648
649
650
651
652
    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
653
654
        // 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
655
        std::vector<std::pair<int32_t, int32_t>> pad_per_dim(ndims);
Paul's avatar
Paul committed
656
        auto tf_padding = args[1]->eval().get<int32_t>().to_vector();
Khalique's avatar
Khalique committed
657
        for(size_t i = 0; i < 2 * ndims; i += 2)
Khalique's avatar
Khalique committed
658
        {
Khalique's avatar
Khalique committed
659
660
            pad_per_dim[i / 2].first  = tf_padding[i];
            pad_per_dim[i / 2].second = tf_padding[i + 1];
Khalique's avatar
Khalique committed
661
662
663
664
        }
        reorder_data(pad_per_dim);

        op::pad op;
Khalique's avatar
Khalique committed
665
666
        std::vector<int64_t> pads(ndims * 2);
        for(size_t i = 0; i < ndims; i++)
Khalique's avatar
Khalique committed
667
        {
Khalique's avatar
Khalique committed
668
669
            pads[i]         = pad_per_dim[i].first;
            pads[i + ndims] = pad_per_dim[i].second;
Khalique's avatar
Khalique committed
670
671
        }
        op.pads = pads;
Paul's avatar
Paul committed
672
        return prog.add_instruction(op, args.front());
Khalique's avatar
Khalique committed
673
674
    }

675
676
677
678
679
    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
680

681
682
        if(contains(attributes, "strides"))
        {
683
            std::vector<size_t> stride;
684
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
685
            reorder_data(stride);
686
687
688
689
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
690
691
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
692
693
694
        }
        if(contains(attributes, "ksize"))
        {
695
            std::vector<size_t> ksize;
696
            copy(attributes.at("ksize").list().i(), std::back_inserter(ksize));
697
            reorder_data(ksize);
698
699
700
            if(ksize.size() != 4)
            {
                MIGRAPHX_THROW("ksize should have 4 values");
Khalique's avatar
Khalique committed
701
            }
702
703
            op.lengths[0] = ksize[2];
            op.lengths[1] = ksize[3];
704
        }
Khalique's avatar
Khalique committed
705
706

        auto l0 = args[0];
Khalique's avatar
Khalique committed
707
708
709
710
711
        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
712
                op.padding_mode = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
713
                auto input_dims = l0->get_shape().lens();
Khalique's avatar
Khalique committed
714
715
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
716
717
718
719
720
721
722
                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
723
724
                    l0                           = prog.add_instruction(
                        migraphx::op::pad{padding, std::numeric_limits<float>::lowest()}, l0);
Khalique's avatar
Khalique committed
725
726
727
                }
                else
                {
Khalique's avatar
Khalique committed
728
729
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
730
                }
Khalique's avatar
Khalique committed
731
732
733
734
735
736
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::valid;
            }
        }
Khalique's avatar
Khalique committed
737
        return prog.add_instruction(op, l0);
738
    }
Khalique's avatar
Khalique committed
739

740
    instruction_ref
Khalique's avatar
Khalique committed
741
    parse_reshape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
742
743
744
745
    {
        op::reshape op;
        if(args.size() != 2)
            MIGRAPHX_THROW("reshape needs 2 arguments (input, new_shape)");
Khalique's avatar
Khalique committed
746
        auto s = args[1]->eval();
747
        s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
Paul's avatar
Paul committed
748
        return prog.add_instruction(op, make_contiguous(args[0]));
749
750
    }

Khalique's avatar
Khalique committed
751
752
753
754
755
756
757
758
759
    void parse_from(std::istream& is)
    {
        tensorflow::GraphDef graph;
        if(graph.ParseFromIstream(&is))
        {
            this->parse_graph(graph);
        }
        else
        {
760
            throw std::runtime_error("Failed reading tf file");
Khalique's avatar
Khalique committed
761
762
763
        }
    }

764
765
766
767
768
769
770
771
772
773
    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
774
775
776
    instruction_ref parse_squeeze(const std::string&,
                                  const attribute_map& attributes,
                                  std::vector<instruction_ref> args)
777
778
    {
        op::squeeze op;
Khalique's avatar
Khalique committed
779
        auto input_dims = args[0]->get_shape().lens();
Khalique's avatar
Khalique committed
780
        auto axes       = attributes.at("squeeze_dims").list().i();
781
        copy(axes, std::back_inserter(op.axes));
Khalique's avatar
Khalique committed
782

783
784
        if(op.axes.empty()) // no squeeze_dims provided, remove any dim that equals 1
        {
Khalique's avatar
Khalique committed
785
            for(size_t i = 0; i < input_dims.size(); i++)
786
            {
Khalique's avatar
Khalique committed
787
                if(input_dims.at(i) == 1)
788
789
790
791
                {
                    op.axes.push_back(i);
                }
            }
792
        }
Paul's avatar
Paul committed
793
        return prog.add_instruction(op, make_contiguous(args[0]));
794
795
    }

Khalique's avatar
Khalique committed
796
797
    instruction_ref
    parse_slice(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
798
799
    {
        op::slice op;
Khalique's avatar
Khalique committed
800
        auto starts     = args[1]->eval().get<int32_t>().to_vector();
Khalique's avatar
Khalique committed
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
        auto size       = args[2]->eval().get<int32_t>().to_vector();
        auto axes       = args[0]->get_shape().lens();
        size_t num_axes = axes.size();

        op.starts = std::vector<int64_t>(starts.begin(), starts.end());
        op.ends   = std::vector<int64_t>(num_axes);
        op.axes   = std::vector<int64_t>(num_axes);
        std::iota(op.axes.begin(), op.axes.end(), 0);
        for(size_t i = 0; i < num_axes; i++)
        {
            if(size[i] == -1)
                op.ends[i] = axes[i];
            else
                op.ends[i] = starts[i] + size[i];
        }
        return prog.add_instruction(op, make_contiguous(args[0]));
    }

Khalique's avatar
Khalique committed
819
820
821
    instruction_ref parse_stridedslice(const std::string&,
                                       const attribute_map& attributes,
                                       std::vector<instruction_ref> args)
822
823
    {
        op::slice op;
Khalique's avatar
Khalique committed
824
825
826
827
        auto starts              = args[1]->eval().get<int32_t>().to_vector();
        auto ends                = args[2]->eval().get<int32_t>().to_vector();
        auto l0                  = args[0];
        size_t num_axes          = l0->get_shape().lens().size();
828
        std::vector<size_t> axes = l0->get_shape().lens();
829

Khalique's avatar
Khalique committed
830
831
832
833
        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);
Khalique's avatar
Khalique committed
834
835
        uint32_t begin_mask       = 0;
        uint32_t end_mask         = 0;
836
        uint32_t shrink_axis_mask = 0;
Khalique's avatar
Khalique committed
837
        uint32_t bitwise_compare  = 1;
Khalique's avatar
Khalique committed
838
839
        std::vector<int64_t> begin_axes;
        std::vector<int64_t> end_axes;
840
841
        std::vector<int64_t> squeeze_axes;

Khalique's avatar
Khalique committed
842
843
844
845
846
847
        if(contains(attributes, "begin_mask"))
            begin_mask = static_cast<uint32_t>(attributes.at("begin_mask").i());

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

848
        if(contains(attributes, "shrink_axis_mask"))
849
            shrink_axis_mask = static_cast<uint32_t>(attributes.at("shrink_axis_mask").i());
850

Khalique's avatar
Khalique committed
851
852
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
        for(size_t i = 0; i < num_axes; i++)
        {
            // the LSB corresponds to axis 0 when determining which axes to begin
            if(((begin_mask >> i) & bitwise_compare) == 1)
                begin_axes.push_back(1);
            else
                begin_axes.push_back(0);
        }

        for(size_t i = 0; i < num_axes; i++)
        {
            // the LSB corresponds to axis 0 when determining which axes to end
            if(((end_mask >> i) & bitwise_compare) == 1)
                end_axes.push_back(1);
            else
                end_axes.push_back(0);
        }

        for(size_t i = 0; i < num_axes; i++)
        {
            if(begin_axes.at(i) == 1)
            {
                op.starts.at(i) = 0;
            }
            if(end_axes.at(i) == 1)
            {
                op.ends.at(i) = axes.at(i);
            }
        }

881
        auto l1 = prog.add_instruction(op, l0);
Khalique's avatar
Khalique committed
882
        if(shrink_axis_mask == 0)
883
            return l1;
Khalique's avatar
Khalique committed
884

Khalique's avatar
Khalique committed
885
        for(size_t i = 0; i < num_axes; i++)
886
        {
887
            // the LSB corresponds to axis 0 when determining which axes to squeeze
Khalique's avatar
Khalique committed
888
            if(((shrink_axis_mask >> i) & bitwise_compare) == 1)
889
890
                squeeze_axes.push_back(i);
        }
Khalique's avatar
Khalique committed
891

892
        return prog.add_instruction(op::squeeze{squeeze_axes}, l1);
893
894
    }

Khalique's avatar
Khalique committed
895
896
    instruction_ref
    parse_transpose(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
897
898
899
900
901
902
903
904
    {
        auto perm = args[1]->eval().get<int32_t>().to_vector();
        op::transpose op;
        op.dims = std::vector<int64_t>(perm.begin(), perm.end());

        return prog.add_instruction(op, args.front());
    }

Khalique's avatar
Khalique committed
905
906
907
908
909
    void parse_graph(const tensorflow::GraphDef& graph)
    {
        nodes = get_nodes(graph, input_nodes);
        for(auto&& input : input_nodes)
        {
Khalique's avatar
Khalique committed
910
            const std::string& name   = input.name();
Khalique's avatar
Khalique committed
911
            attribute_map input_attrs = get_attributes(input);
Khalique's avatar
Khalique committed
912
913
            shape::type_t shape_type  = parse_type(input_attrs.at("dtype").type());
            std::vector<size_t> dims  = parse_dims(input_attrs.at("shape").shape());
914
            if(is_nhwc and dims.size() >= 4)
915
            {
916
                reorder_data(dims);
917
            }
Khalique's avatar
Khalique committed
918
            shape s            = shape{shape_type, dims};
Paul's avatar
Paul committed
919
            instructions[name] = to_nhwc(prog.add_parameter(name, s));
Khalique's avatar
Khalique committed
920
921
922
        }
        for(auto&& p : nodes)
        {
923
            this->parse_node(p.first);
Khalique's avatar
Khalique committed
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
        }
    }

    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)
            {
950
                instructions[name] = prog.add_instruction(op::unknown{node.op()}, args);
Khalique's avatar
Khalique committed
951
952
953
954
955
956
957
958
959
960
961
            }
            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
962
        for(auto&& attr : node.attr())
Khalique's avatar
Khalique committed
963
964
965
966
967
968
        {
            result[attr.first] = attr.second;
        }
        return result;
    }

Khalique's avatar
Khalique committed
969
    static std::string get_name(const tensorflow::NodeDef& node) { return node.name(); }
Khalique's avatar
Khalique committed
970

Khalique's avatar
Khalique committed
971
972
    static node_map get_nodes(const tensorflow::GraphDef& graph,
                              std::vector<tensorflow::NodeDef>& input_nodes)
Khalique's avatar
Khalique committed
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
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
    {
        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
1032
1033
1034
        case tensorflow::DataType::DT_UINT64:
            shape_type = shape::uint64_type;
            break;
Khalique's avatar
Khalique committed
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061

        // 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
1062
1063
1064
1065
        }
        return shape_type;
    }

Khalique's avatar
Khalique committed
1066
    static literal parse_tensor(const tensorflow::TensorProto& t)
Khalique's avatar
Khalique committed
1067
1068
    {
        std::vector<size_t> dims = parse_dims(t.tensor_shape());
1069
        size_t shape_size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<size_t>());
Khalique's avatar
Khalique committed
1070
1071
        if(!t.tensor_content().empty()) // has raw data
        {
Khalique's avatar
Khalique committed
1072
            const std::string& s = t.tensor_content();
Khalique's avatar
Khalique committed
1073
1074
1075
            switch(t.dtype())
            {
            case tensorflow::DataType::DT_INVALID: throw std::runtime_error("");
Khalique's avatar
Khalique committed
1076
1077
            case tensorflow::DataType::DT_FLOAT:
                return literal{{shape::float_type, dims}, s.data()};
Khalique's avatar
Khalique committed
1078
            case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
1079
            case tensorflow::DataType::DT_INT8: return literal{{shape::int8_type, dims}, s.data()};
Khalique's avatar
Khalique committed
1080
            case tensorflow::DataType::DT_UINT16:
1081
                return literal{{shape::uint16_type, dims}, s.data()};
Khalique's avatar
Khalique committed
1082
            case tensorflow::DataType::DT_INT16:
1083
                return literal{{shape::int16_type, dims}, s.data()};
Khalique's avatar
Khalique committed
1084
1085
1086
1087
            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
1088
            case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
1089
            case tensorflow::DataType::DT_BOOL: return literal{{shape::int8_type, dims}, s.data()};
Khalique's avatar
Khalique committed
1090
            case tensorflow::DataType::DT_HALF: return literal{{shape::half_type, dims}, s.data()};
Khalique's avatar
Khalique committed
1091
1092
            case tensorflow::DataType::DT_DOUBLE:
                return literal{{shape::double_type, dims}, s.data()};
Khalique's avatar
Khalique committed
1093
1094
1095
1096
            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
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
            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
1128
1129
1130
1131
            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
1132
1133
1134
1135
1136
1137
1138
            }
            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
1139
1140
            return create_literal(
                shape::float_type, dims, get_data_vals(t.float_val(), shape_size));
Khalique's avatar
Khalique committed
1141
1142
        case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT8:
1143
            return create_literal(shape::int8_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
1144
        case tensorflow::DataType::DT_UINT16:
1145
            return create_literal(shape::uint16_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
1146
        case tensorflow::DataType::DT_INT16:
1147
            return create_literal(shape::int16_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
1148
        case tensorflow::DataType::DT_INT32:
1149
            return create_literal(shape::int32_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
1150
        case tensorflow::DataType::DT_INT64:
Khalique's avatar
Khalique committed
1151
1152
            return create_literal(
                shape::int64_type, dims, get_data_vals(t.int64_val(), shape_size));
Khalique's avatar
Khalique committed
1153
1154
        case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
        case tensorflow::DataType::DT_BOOL:
1155
            return create_literal(shape::int32_type, dims, get_data_vals(t.bool_val(), shape_size));
Khalique's avatar
Khalique committed
1156
        case tensorflow::DataType::DT_HALF:
Khalique's avatar
Khalique committed
1157
        {
1158
1159
            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
1160
1161
1162
1163
1164
            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); });
1165
            return create_literal(shape::half_type, dims, data_half);
Khalique's avatar
Khalique committed
1166
        }
Khalique's avatar
Khalique committed
1167
        case tensorflow::DataType::DT_DOUBLE:
Khalique's avatar
Khalique committed
1168
            return literal{{shape::double_type, dims}, get_data_vals(t.double_val(), shape_size)};
Khalique's avatar
Khalique committed
1169
1170
1171
1172
        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
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
        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
1204
1205
1206
1207
        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
1208
1209
1210
1211
        }
        MIGRAPHX_THROW("Invalid tensor type");
    }

1212
    template <class T>
Khalique's avatar
Khalique committed
1213
    static std::vector<T> get_data_vals(const google::protobuf::RepeatedField<T>& data,
1214
                                        const size_t& shape_size)
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
    {
        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
1227
1228
1229
1230
    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
1231
1232
1233
        std::transform(input_dims.begin(),
                       input_dims.end(),
                       std::back_inserter(dims),
Paul's avatar
Paul committed
1234
                       [](const tensorflow::TensorShapeProto_Dim& dim) { return dim.size(); });
Khalique's avatar
Khalique committed
1235
1236
        return dims;
    }
1237
1238

    template <class T>
Khalique's avatar
Khalique committed
1239
    static literal
1240
    create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, std::vector<T> data)
1241
    {
Khalique's avatar
Khalique committed
1242
        // assume if explicit value is mentioned in protobuf and dim size <= 1, treat as scalar
1243
        if(dims.empty() or (dims.size() == 1 and dims.front() == 1))
1244
            return literal{{shape_type}, data};
1245
1246
        return literal{{shape_type, dims}, data};
    }
Khalique's avatar
Khalique committed
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
};

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
Paul's avatar
Paul committed
1269
    parser.to_nchw(std::prev(parser.prog.end()));
Khalique's avatar
Khalique committed
1270
1271
1272
1273
1274
    return std::move(parser.prog);
}

} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx