tf.cpp 48.1 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("SquaredDifference", op::sqdiff{});
Khalique's avatar
Khalique committed
164
        add_binary_op("Sub", op::sub{});
Khalique's avatar
Khalique committed
165

166
        add_mem_op("AvgPool", &tf_parser::parse_pooling);
167
        add_mem_op("BiasAdd", &tf_parser::parse_biasadd);
Paul's avatar
Paul committed
168
        add_mem_op("ConcatV2", &tf_parser::parse_concat, false);
Khalique's avatar
Khalique committed
169
        add_mem_op("Const", &tf_parser::parse_constant);
Paul's avatar
Paul committed
170
        add_mem_op("Conv2D", &tf_parser::parse_conv);
Paul's avatar
Paul committed
171
        add_mem_op("DepthwiseConv2dNative", &tf_parser::parse_depthwiseconv);
172
        add_mem_op("ExpandDims", &tf_parser::parse_expanddims, false);
Khalique's avatar
Khalique committed
173
        add_mem_op("FusedBatchNorm", &tf_parser::parse_batchnorm);
Paul's avatar
Paul committed
174
        add_mem_op("MatMul", &tf_parser::parse_matmul, false);
175
        add_mem_op("MaxPool", &tf_parser::parse_pooling);
Khalique's avatar
Khalique committed
176
        add_mem_op("Mean", &tf_parser::parse_mean);
Paul's avatar
Paul committed
177
        add_mem_op("Pack", &tf_parser::parse_pack, false);
Paul's avatar
Paul committed
178
        add_mem_op("Pad", &tf_parser::parse_pad);
Paul's avatar
Paul committed
179
        add_mem_op("Reshape", &tf_parser::parse_reshape, false);
Khalique's avatar
Khalique committed
180
        add_mem_op("Slice", &tf_parser::parse_slice, false);
181
        add_mem_op("Softmax", &tf_parser::parse_softmax);
Paul's avatar
Paul committed
182
        add_mem_op("Squeeze", &tf_parser::parse_squeeze, false);
183
        add_mem_op("StridedSlice", &tf_parser::parse_stridedslice);
Khalique's avatar
Khalique committed
184
        add_mem_op("Transpose", &tf_parser::parse_transpose, false);
Khalique's avatar
Khalique committed
185
186
    }

187
    template <class F>
Paul's avatar
Paul committed
188
    void add_op(std::string name, F f, bool transpose = true)
189
    {
Paul's avatar
Paul committed
190
        if(transpose)
Paul's avatar
Paul committed
191
        {
Paul's avatar
Paul committed
192
193
            ops.emplace(name,
                        op_func{[=](const attribute_map& attributes,
Paul's avatar
Paul committed
194
                                    const std::vector<instruction_ref>& args) -> instruction_ref {
Paul's avatar
Paul committed
195
196
                            return to_nhwc(f(attributes, to_nchw(args)));
                        }});
Paul's avatar
Paul committed
197
198
199
200
201
        }
        else
        {
            ops.emplace(name, f);
        }
202
203
    }

Khalique's avatar
Khalique committed
204
    template <class F>
Paul's avatar
Paul committed
205
    void add_mem_op(std::string name, F f, bool transpose = true)
Khalique's avatar
Khalique committed
206
    {
Paul's avatar
Paul committed
207
208
209
210
211
        add_op(name,
               [=](auto&&... xs) {
                   return std::mem_fn(f)(*this, name, std::forward<decltype(xs)>(xs)...);
               },
               transpose);
Khalique's avatar
Khalique committed
212
213
214
215
216
    }

    template <class T>
    void add_binary_op(std::string name, T x)
    {
Paul's avatar
Paul committed
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
        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
232
233
234
235
236
    }

    template <class T>
    instruction_ref add_broadcastable_binary_op(instruction_ref arg0, instruction_ref arg1, T x)
    {
Khalique's avatar
Khalique committed
237
        if(arg0->get_shape().lens() != arg1->get_shape().lens())
Khalique's avatar
Khalique committed
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
        {
            // 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
253
254
            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
255
256
257
258
259

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

260
            std::vector<size_t> output_lens(*s1);
Khalique's avatar
Khalique committed
261
262
263
264
265
266
267
268
269
            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
270
            return to_nhwc(prog.add_instruction(x, to_nchw(l0), to_nchw(l1)));
Khalique's avatar
Khalique committed
271
272
273
        }
        else
        {
Paul's avatar
Paul committed
274
            return to_nhwc(prog.add_instruction(x, {to_nchw(arg0), to_nchw(arg1)}));
Khalique's avatar
Khalique committed
275
276
277
278
        }
    }

    template <class T>
Paul's avatar
Paul committed
279
    void add_generic_op(std::string name, T x, bool transpose = true)
Khalique's avatar
Khalique committed
280
    {
Paul's avatar
Paul committed
281
282
283
284
285
        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
286
287
288
289
290
    }

    instruction_ref
    parse_batchnorm(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
Khalique's avatar
Khalique committed
291
292
293
        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
294
295
296
297
298
299
300
301
        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));
    }

302
    instruction_ref
Khalique's avatar
Khalique committed
303
    parse_biasadd(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
304
    {
305
        uint64_t axis = 1; // assume output of previous layer is in NCHW (broadcast on channel)
Shucai Xiao's avatar
Shucai Xiao committed
306
        auto l0 = prog.add_instruction(op::broadcast{axis, args[0]->get_shape().lens()}, args[1]);
307
        return prog.add_instruction(op::add{}, args[0], l0);
308
309
    }

Khalique's avatar
Khalique committed
310
311
312
313
    instruction_ref
    parse_concat(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        // get index for axis within args
314
        size_t axis_idx = attributes.at("N").i();
Paul's avatar
Paul committed
315
        size_t axis     = args[axis_idx]->eval().at<int64_t>();
Khalique's avatar
Khalique committed
316
        op::concat op{axis};
317
        // return only first N arguments (assuming last index is the axis value)
Paul's avatar
Paul committed
318
319
        return prog.add_instruction(
            op, std::vector<instruction_ref>(args.begin(), args.begin() + args.size() - 1));
Khalique's avatar
Khalique committed
320
321
322
323
324
325
    }

    instruction_ref parse_constant(const std::string&,
                                   attribute_map attributes,
                                   const std::vector<instruction_ref>&)
    {
Paul's avatar
Paul committed
326
        literal v = parse_tensor(attributes.at("value").tensor());
Paul's avatar
Paul committed
327
        return prog.add_literal(v);
Khalique's avatar
Khalique committed
328
329
330
331
332
333
334
335
    }

    instruction_ref
    parse_conv(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
    {
        op::convolution op;
        if(contains(attributes, "strides"))
        {
336
            std::vector<size_t> stride;
337
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
338
            reorder_data(stride);
339
340
            if(stride.size() != 4)
            {
341
                MIGRAPHX_THROW("strides should have 4 values");
342
            }
343
344
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
Khalique's avatar
Khalique committed
345
346
347
        }
        if(contains(attributes, "dilations"))
        {
348
            std::vector<size_t> dilation;
349
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
350
            reorder_data(dilation);
351
352
353
354
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
355
356
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
Khalique's avatar
Khalique committed
357
        }
Khalique's avatar
Khalique committed
358

Paul's avatar
Paul committed
359
        auto weights = to_kcxy(args[1]);
Paul's avatar
Paul committed
360
        auto l0      = args[0];
Khalique's avatar
Khalique committed
361
362
363
364
365
        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
366
                op.padding_mode                 = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
367
368
369
                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
370
371

                auto input_dims = l0->get_shape().lens();
Khalique's avatar
Khalique committed
372
373
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
374
375
376
377
378
379
380
381
382
383
384
                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
385
386
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
387
                }
388
389
390
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
391
                op.padding_mode = op::padding_mode_t::valid;
Khalique's avatar
Khalique committed
392
            }
Khalique's avatar
Khalique committed
393
            else if(pad_mode.find("EXPLICIT") != std::string::npos)
Khalique's avatar
Khalique committed
394
            {
395
                std::vector<size_t> padding;
396
                copy(attributes.at("explicit_paddings").list().i(), std::back_inserter(padding));
Khalique's avatar
Khalique committed
397
398
399
400
401
402
403
404
405
406
407
408
                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
409
        return prog.add_instruction(op, {l0, to_kcxy(args[1])});
Khalique's avatar
Khalique committed
410
411
    }

Khalique's avatar
Khalique committed
412
413
414
    instruction_ref parse_depthwiseconv(const std::string&,
                                        attribute_map attributes,
                                        std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
415
416
417
    {
        op::convolution op;
        size_t num_channels = args[0]->get_shape().lens()[1];
Khalique's avatar
Khalique committed
418
        op.group            = num_channels;
Khalique's avatar
Khalique committed
419

Khalique's avatar
Khalique committed
420
421
        if(contains(attributes, "strides"))
        {
422
            std::vector<size_t> stride;
423
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
424
            reorder_data(stride);
425
426
            if(stride.size() != 4)
            {
427
                MIGRAPHX_THROW("strides should have 4 values");
428
            }
429
430
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
Khalique's avatar
Khalique committed
431
        }
Paul's avatar
Paul committed
432
433

        auto weights = to_kcxy(args[1]);
Khalique's avatar
Khalique committed
434
435
        if(contains(attributes, "dilations"))
        {
436
            std::vector<size_t> dilation;
437
            copy(attributes.at("dilations").list().i(), std::back_inserter(dilation));
438
            reorder_data(dilation);
439
440
441
442
            if(dilation.size() != 4)
            {
                MIGRAPHX_THROW("dilation should have 4 values");
            }
443
444
            op.dilation[0] = dilation[2];
            op.dilation[1] = dilation[3];
Khalique's avatar
Khalique committed
445
446
        }

Khalique's avatar
Khalique committed
447
        auto l0 = args[0];
Khalique's avatar
Khalique committed
448
449
450
        if(contains(attributes, "padding"))
        {
            const std::string& pad_mode = attributes.at("padding").s();
Khalique's avatar
Khalique committed
451

Khalique's avatar
Khalique committed
452
453
            if(pad_mode.find("SAME") != std::string::npos)
            {
Khalique's avatar
Khalique committed
454
                op.padding_mode                 = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
455
456
457
458
459
                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
460
461
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
462
463
464
465
466
467
468
469
470
471
472
                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
473
474
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
475
                }
Khalique's avatar
Khalique committed
476
            }
Khalique's avatar
Khalique committed
477
            else if(pad_mode.find("VALID") != std::string::npos)
Khalique's avatar
Khalique committed
478
            {
Khalique's avatar
Khalique committed
479
                op.padding_mode = op::padding_mode_t::valid;
Khalique's avatar
Khalique committed
480
481
            }
        }
Khalique's avatar
Khalique committed
482

Khalique's avatar
Khalique committed
483
484
        std::vector<int64_t> new_weights_shape;
        copy(weights->get_shape().lens(), std::back_inserter(new_weights_shape));
Khalique's avatar
Khalique committed
485
486
487
488

        // 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
489
        int64_t multiplier   = new_weights_shape[0];
Khalique's avatar
Khalique committed
490
491
492
        int64_t out_channels = num_channels * multiplier;
        new_weights_shape[0] = out_channels;
        new_weights_shape[1] = 1;
Paul's avatar
Paul committed
493
        // Make sure weights are contiguous before doing reshape
Paul's avatar
Paul committed
494
495
        auto new_weights =
            prog.add_instruction(op::reshape{new_weights_shape}, make_contiguous(weights));
Khalique's avatar
Khalique committed
496

Khalique's avatar
Khalique committed
497
        return prog.add_instruction(op, {l0, new_weights});
Khalique's avatar
Khalique committed
498
499
    }

Khalique's avatar
Khalique committed
500
501
    instruction_ref
    parse_expanddims(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
502
503
    {
        std::vector<size_t> input_dims = args[0]->get_shape().lens();
Khalique's avatar
Khalique committed
504
        std::vector<int64_t> new_dims(input_dims.begin(), input_dims.end());
Khalique's avatar
Khalique committed
505
        size_t num_dims = input_dims.size();
506
        int32_t dim     = args[1]->eval().at<int32_t>();
Khalique's avatar
Khalique committed
507
508

        if(dim < 0)
Khalique's avatar
Khalique committed
509
510
511
512
513
514
515
516
517
518
        {
            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
519
520
    instruction_ref
    parse_matmul(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
521
522
523
    {
        bool transa = false;
        bool transb = false;
Khalique's avatar
Khalique committed
524

525
526
527
528
529
530
531
532
533
534
535
536
        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
537
        std::iter_swap(perm.end() - 1, perm.end() - 2);
538
539
540
541
542
543
544

        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
545
546
    instruction_ref
    parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
547
548
    {
        bool keep_dims = attributes.at("keep_dims").b();
Paul's avatar
Paul committed
549
        std::vector<int32_t> hw_axes{2, 3};
Khalique's avatar
Khalique committed
550
        // check if conditions for GlobalAvgPool are met
Khalique's avatar
Khalique committed
551
        auto lens = args[0]->get_shape().lens();
Khalique's avatar
Khalique committed
552
553
        auto axes = parse_axes(args[1]->eval().get<int32_t>().to_vector(), lens.size());

Khalique's avatar
Khalique committed
554
        if(axes == hw_axes and lens.size() == 4)
Khalique's avatar
Khalique committed
555
556
        {
            op::pooling op{"average"};
Khalique's avatar
Khalique committed
557
558
            op.lengths[0] = lens[2];
            op.lengths[1] = lens[3];
Khalique's avatar
Khalique committed
559
560
561
562
563
            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
564
565
566
567
        }
        MIGRAPHX_THROW("MIGraphX does not support mean outside of GlobalAvgPool transformation");
    }

Khalique's avatar
Khalique committed
568
569
570
    instruction_ref parse_pack(const std::string&,
                               const attribute_map& attributes,
                               std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
571
572
573
574
575
576
    {
        // reinterpret as unsqueeze with concat
        std::vector<instruction_ref> unsqueezed_args;
        int64_t axis = 0;
        if(contains(attributes, "axis"))
            axis = attributes.at("axis").i();
577
578
579
        size_t input_size = args.front()->get_shape().lens().size();
        if(axis > input_size)
        {
Khalique's avatar
Khalique committed
580
581
            MIGRAPHX_THROW("TF_PARSER: axis value of " + to_string(axis) +
                           " must be smaller than input size " + to_string(input_size));
582
583
        }

Khalique's avatar
Khalique committed
584
585
586
587
588
        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
589
590
        return to_nhwc(
            prog.add_instruction(op::concat{static_cast<size_t>(axis)}, unsqueezed_args));
Khalique's avatar
Khalique committed
591
592
    }

Khalique's avatar
Khalique committed
593
594
595
596
597
    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
598
599
        // 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
600
        std::vector<std::pair<int32_t, int32_t>> pad_per_dim(ndims);
Paul's avatar
Paul committed
601
        auto tf_padding = args[1]->eval().get<int32_t>().to_vector();
Khalique's avatar
Khalique committed
602
        for(size_t i = 0; i < 2 * ndims; i += 2)
Khalique's avatar
Khalique committed
603
        {
Khalique's avatar
Khalique committed
604
605
            pad_per_dim[i / 2].first  = tf_padding[i];
            pad_per_dim[i / 2].second = tf_padding[i + 1];
Khalique's avatar
Khalique committed
606
607
608
609
        }
        reorder_data(pad_per_dim);

        op::pad op;
Khalique's avatar
Khalique committed
610
611
        std::vector<int64_t> pads(ndims * 2);
        for(size_t i = 0; i < ndims; i++)
Khalique's avatar
Khalique committed
612
        {
Khalique's avatar
Khalique committed
613
614
            pads[i]         = pad_per_dim[i].first;
            pads[i + ndims] = pad_per_dim[i].second;
Khalique's avatar
Khalique committed
615
616
        }
        op.pads = pads;
Paul's avatar
Paul committed
617
        return prog.add_instruction(op, args.front());
Khalique's avatar
Khalique committed
618
619
    }

620
621
622
623
624
    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
625

626
627
        if(contains(attributes, "strides"))
        {
628
            std::vector<size_t> stride;
629
            copy(attributes.at("strides").list().i(), std::back_inserter(stride));
630
            reorder_data(stride);
631
632
633
634
            if(stride.size() != 4)
            {
                MIGRAPHX_THROW("strides should have 4 values");
            }
635
636
            op.stride[0] = stride[2];
            op.stride[1] = stride[3];
637
638
639
        }
        if(contains(attributes, "ksize"))
        {
640
            std::vector<size_t> ksize;
641
            copy(attributes.at("ksize").list().i(), std::back_inserter(ksize));
642
            reorder_data(ksize);
643
644
645
            if(ksize.size() != 4)
            {
                MIGRAPHX_THROW("ksize should have 4 values");
Khalique's avatar
Khalique committed
646
            }
647
648
            op.lengths[0] = ksize[2];
            op.lengths[1] = ksize[3];
649
        }
Khalique's avatar
Khalique committed
650
651

        auto l0 = args[0];
Khalique's avatar
Khalique committed
652
653
654
655
656
        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
657
                op.padding_mode = op::padding_mode_t::same;
Khalique's avatar
Khalique committed
658
                auto input_dims = l0->get_shape().lens();
Khalique's avatar
Khalique committed
659
660
                size_t input_h  = input_dims[2];
                size_t input_w  = input_dims[3];
Khalique's avatar
Khalique committed
661
662
663
664
665
666
667
                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
668
669
                    l0                           = prog.add_instruction(
                        migraphx::op::pad{padding, std::numeric_limits<float>::lowest()}, l0);
Khalique's avatar
Khalique committed
670
671
672
                }
                else
                {
Khalique's avatar
Khalique committed
673
674
                    op.padding[0] = pads[0];
                    op.padding[1] = pads[1];
Khalique's avatar
Khalique committed
675
                }
Khalique's avatar
Khalique committed
676
677
678
679
680
681
            }
            else if(pad_mode.find("VALID") != std::string::npos)
            {
                op.padding_mode = op::padding_mode_t::valid;
            }
        }
Khalique's avatar
Khalique committed
682
        return prog.add_instruction(op, l0);
683
    }
Khalique's avatar
Khalique committed
684

685
    instruction_ref
Khalique's avatar
Khalique committed
686
    parse_reshape(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
687
688
689
690
    {
        op::reshape op;
        if(args.size() != 2)
            MIGRAPHX_THROW("reshape needs 2 arguments (input, new_shape)");
Khalique's avatar
Khalique committed
691
        auto s = args[1]->eval();
692
        s.visit([&](auto v) { copy(v, std::back_inserter(op.dims)); });
Paul's avatar
Paul committed
693
        return prog.add_instruction(op, make_contiguous(args[0]));
694
695
    }

Khalique's avatar
Khalique committed
696
697
698
699
700
701
702
703
704
    void parse_from(std::istream& is)
    {
        tensorflow::GraphDef graph;
        if(graph.ParseFromIstream(&is))
        {
            this->parse_graph(graph);
        }
        else
        {
705
            throw std::runtime_error("Failed reading tf file");
Khalique's avatar
Khalique committed
706
707
708
        }
    }

709
710
711
712
713
714
715
716
717
718
    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
719
720
721
    instruction_ref parse_squeeze(const std::string&,
                                  const attribute_map& attributes,
                                  std::vector<instruction_ref> args)
722
723
    {
        op::squeeze op;
Khalique's avatar
Khalique committed
724
        auto input_dims = args[0]->get_shape().lens();
Khalique's avatar
Khalique committed
725
        auto axes       = attributes.at("squeeze_dims").list().i();
726
        copy(axes, std::back_inserter(op.axes));
Khalique's avatar
Khalique committed
727

728
729
        if(op.axes.empty()) // no squeeze_dims provided, remove any dim that equals 1
        {
Khalique's avatar
Khalique committed
730
            for(size_t i = 0; i < input_dims.size(); i++)
731
            {
Khalique's avatar
Khalique committed
732
                if(input_dims.at(i) == 1)
733
734
735
736
                {
                    op.axes.push_back(i);
                }
            }
737
        }
Paul's avatar
Paul committed
738
        return prog.add_instruction(op, make_contiguous(args[0]));
739
740
    }

Khalique's avatar
Khalique committed
741
742
    instruction_ref
    parse_slice(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
743
744
    {
        op::slice op;
Khalique's avatar
Khalique committed
745
        auto starts     = args[1]->eval().get<int32_t>().to_vector();
Khalique's avatar
Khalique committed
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
        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
764
765
766
    instruction_ref parse_stridedslice(const std::string&,
                                       const attribute_map& attributes,
                                       std::vector<instruction_ref> args)
767
768
    {
        op::slice op;
Khalique's avatar
Khalique committed
769
770
771
        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();
772

Khalique's avatar
Khalique committed
773
774
775
776
        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);
777
        uint32_t shrink_axis_mask = 0;
Khalique's avatar
Khalique committed
778
        uint32_t bitwise_compare  = 1;
779
780
781
        std::vector<int64_t> squeeze_axes;

        if(contains(attributes, "shrink_axis_mask"))
782
            shrink_axis_mask = static_cast<uint32_t>(attributes.at("shrink_axis_mask").i());
783

Khalique's avatar
Khalique committed
784
        for(size_t i = 0; i < num_axes; i++)
785
        {
786
            // the LSB corresponds to axis 0 when determining which axes to squeeze
Khalique's avatar
Khalique committed
787
            if(((shrink_axis_mask >> i) & bitwise_compare) == 1)
788
789
                squeeze_axes.push_back(i);
        }
Khalique's avatar
Khalique committed
790

Paul's avatar
Paul committed
791
792
        auto l0 = prog.add_instruction(op, make_contiguous(args[0]));
        return to_nhwc(prog.add_instruction(op::squeeze{squeeze_axes}, l0));
793
794
    }

Khalique's avatar
Khalique committed
795
796
    instruction_ref
    parse_transpose(const std::string&, const attribute_map&, std::vector<instruction_ref> args)
Khalique's avatar
Khalique committed
797
798
799
800
801
802
803
804
    {
        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
805
806
807
808
809
    void parse_graph(const tensorflow::GraphDef& graph)
    {
        nodes = get_nodes(graph, input_nodes);
        for(auto&& input : input_nodes)
        {
Khalique's avatar
Khalique committed
810
            const std::string& name   = input.name();
Khalique's avatar
Khalique committed
811
            attribute_map input_attrs = get_attributes(input);
Khalique's avatar
Khalique committed
812
813
            shape::type_t shape_type  = parse_type(input_attrs.at("dtype").type());
            std::vector<size_t> dims  = parse_dims(input_attrs.at("shape").shape());
814
            if(is_nhwc and dims.size() >= 4)
815
            {
816
                reorder_data(dims);
817
            }
Khalique's avatar
Khalique committed
818
            shape s            = shape{shape_type, dims};
Paul's avatar
Paul committed
819
            instructions[name] = to_nhwc(prog.add_parameter(name, s));
Khalique's avatar
Khalique committed
820
821
822
        }
        for(auto&& p : nodes)
        {
823
            this->parse_node(p.first);
Khalique's avatar
Khalique committed
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
        }
    }

    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)
            {
850
                instructions[name] = prog.add_instruction(op::unknown{node.op()}, args);
Khalique's avatar
Khalique committed
851
852
853
854
855
856
857
858
859
860
861
            }
            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
862
        for(auto&& attr : node.attr())
Khalique's avatar
Khalique committed
863
864
865
866
867
868
        {
            result[attr.first] = attr.second;
        }
        return result;
    }

Khalique's avatar
Khalique committed
869
    static std::string get_name(const tensorflow::NodeDef& node) { return node.name(); }
Khalique's avatar
Khalique committed
870

Khalique's avatar
Khalique committed
871
872
    static node_map get_nodes(const tensorflow::GraphDef& graph,
                              std::vector<tensorflow::NodeDef>& input_nodes)
Khalique's avatar
Khalique committed
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
    {
        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
932
933
934
        case tensorflow::DataType::DT_UINT64:
            shape_type = shape::uint64_type;
            break;
Khalique's avatar
Khalique committed
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961

        // 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
962
963
964
965
        }
        return shape_type;
    }

Khalique's avatar
Khalique committed
966
    static literal parse_tensor(const tensorflow::TensorProto& t)
Khalique's avatar
Khalique committed
967
968
    {
        std::vector<size_t> dims = parse_dims(t.tensor_shape());
969
        size_t shape_size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<size_t>());
Khalique's avatar
Khalique committed
970
971
        if(!t.tensor_content().empty()) // has raw data
        {
Khalique's avatar
Khalique committed
972
            const std::string& s = t.tensor_content();
Khalique's avatar
Khalique committed
973
974
975
            switch(t.dtype())
            {
            case tensorflow::DataType::DT_INVALID: throw std::runtime_error("");
Khalique's avatar
Khalique committed
976
977
            case tensorflow::DataType::DT_FLOAT:
                return literal{{shape::float_type, dims}, s.data()};
Khalique's avatar
Khalique committed
978
            case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
979
            case tensorflow::DataType::DT_INT8: return literal{{shape::int8_type, dims}, s.data()};
Khalique's avatar
Khalique committed
980
            case tensorflow::DataType::DT_UINT16:
981
                return literal{{shape::uint16_type, dims}, s.data()};
Khalique's avatar
Khalique committed
982
            case tensorflow::DataType::DT_INT16:
983
                return literal{{shape::int16_type, dims}, s.data()};
Khalique's avatar
Khalique committed
984
985
986
987
            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
988
            case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
989
            case tensorflow::DataType::DT_BOOL: return literal{{shape::int8_type, dims}, s.data()};
Khalique's avatar
Khalique committed
990
            case tensorflow::DataType::DT_HALF: return literal{{shape::half_type, dims}, s.data()};
Khalique's avatar
Khalique committed
991
992
            case tensorflow::DataType::DT_DOUBLE:
                return literal{{shape::double_type, dims}, s.data()};
Khalique's avatar
Khalique committed
993
994
995
996
            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
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
            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
1028
1029
1030
1031
            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
1032
1033
1034
1035
1036
1037
1038
            }
            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
1039
1040
            return create_literal(
                shape::float_type, dims, get_data_vals(t.float_val(), shape_size));
Khalique's avatar
Khalique committed
1041
1042
        case tensorflow::DataType::DT_UINT8: throw std::runtime_error("");
        case tensorflow::DataType::DT_INT8:
1043
            return create_literal(shape::int8_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
1044
        case tensorflow::DataType::DT_UINT16:
1045
            return create_literal(shape::uint16_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
1046
        case tensorflow::DataType::DT_INT16:
1047
            return create_literal(shape::int16_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
1048
        case tensorflow::DataType::DT_INT32:
1049
            return create_literal(shape::int32_type, dims, get_data_vals(t.int_val(), shape_size));
Khalique's avatar
Khalique committed
1050
        case tensorflow::DataType::DT_INT64:
Khalique's avatar
Khalique committed
1051
1052
            return create_literal(
                shape::int64_type, dims, get_data_vals(t.int64_val(), shape_size));
Khalique's avatar
Khalique committed
1053
1054
        case tensorflow::DataType::DT_STRING: throw std::runtime_error("");
        case tensorflow::DataType::DT_BOOL:
1055
            return create_literal(shape::int32_type, dims, get_data_vals(t.bool_val(), shape_size));
Khalique's avatar
Khalique committed
1056
        case tensorflow::DataType::DT_HALF:
Khalique's avatar
Khalique committed
1057
        {
1058
1059
            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
1060
1061
1062
1063
1064
            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); });
1065
            return create_literal(shape::half_type, dims, data_half);
Khalique's avatar
Khalique committed
1066
        }
Khalique's avatar
Khalique committed
1067
        case tensorflow::DataType::DT_DOUBLE:
Khalique's avatar
Khalique committed
1068
            return literal{{shape::double_type, dims}, get_data_vals(t.double_val(), shape_size)};
Khalique's avatar
Khalique committed
1069
1070
1071
1072
        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
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
        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
1104
1105
1106
1107
        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
1108
1109
1110
1111
        }
        MIGRAPHX_THROW("Invalid tensor type");
    }

1112
    template <class T>
Khalique's avatar
Khalique committed
1113
    static std::vector<T> get_data_vals(const google::protobuf::RepeatedField<T>& data,
1114
                                        const size_t& shape_size)
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
    {
        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
1127
1128
1129
1130
    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
1131
1132
1133
        std::transform(input_dims.begin(),
                       input_dims.end(),
                       std::back_inserter(dims),
Paul's avatar
Paul committed
1134
                       [](const tensorflow::TensorShapeProto_Dim& dim) { return dim.size(); });
Khalique's avatar
Khalique committed
1135
1136
        return dims;
    }
1137
1138

    template <class T>
Khalique's avatar
Khalique committed
1139
    static literal
1140
    create_literal(shape::type_t shape_type, const std::vector<size_t>& dims, std::vector<T> data)
1141
    {
Khalique's avatar
Khalique committed
1142
        // assume if explicit value is mentioned in protobuf and dim size <= 1, treat as scalar
1143
        if(dims.empty() or (dims.size() == 1 and dims.front() == 1))
1144
            return literal{{shape_type}, data};
1145
1146
        return literal{{shape_type, dims}, data};
    }
Khalique's avatar
Khalique committed
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
};

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
1169
    parser.to_nchw(std::prev(parser.prog.end()));
Khalique's avatar
Khalique committed
1170
1171
1172
1173
1174
    return std::move(parser.prog);
}

} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx