parse_resize.cpp 22.1 KB
Newer Older
1
2
3
/*
 * The MIT License (MIT)
 *
4
 * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to deal
 * in the Software without restriction, including without limitation the rights
 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in
 * all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 * THE SOFTWARE.
 */
Paul Fultz II's avatar
Paul Fultz II committed
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
#include <migraphx/onnx/op_parser.hpp>
#include <migraphx/onnx/checks.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/shape_for_each.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/make_op.hpp>

namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace onnx {

const auto& get_nearest_op(const std::string& mode)
{
    using nearest_op = std::function<std::size_t(std::size_t, double)>;
    static std::unordered_map<std::string, nearest_op> const nearest_ops = {
        {"round_prefer_floor",
         [=](std::size_t d_in, double val) {
             val = std::max(0.0, std::min(d_in - 1.0, val));
             return static_cast<std::size_t>(std::ceil((val - 0.5)));
         }},
        {"round_prefer_ceil",
         [=](std::size_t d_in, double val) {
             val = std::max(0.0, std::min(d_in - 1.0, val));
             return static_cast<std::size_t>(std::round((val)));
         }},
        {"floor",
         [=](std::size_t d_in, double val) {
             val = std::max(0.0, std::min(d_in - 1.0, val));
             return static_cast<std::size_t>(std::floor((val)));
         }},
        {"ceil", [=](std::size_t d_in, double val) {
             val = std::max(0.0, std::min(d_in - 1.0, val));
             return static_cast<std::size_t>(std::ceil((val)));
         }}};

59
    if(not contains(nearest_ops, mode))
Paul Fultz II's avatar
Paul Fultz II committed
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
    {
        MIGRAPHX_THROW("PARSE_RESIZE: nearest_mode " + mode + " not supported!");
    }

    return nearest_ops.at(mode);
}

const auto& get_original_idx_op(const std::string& mode)
{
    using original_idx_op = std::function<double(std::size_t, std::size_t, std::size_t, double)>;
    static std::unordered_map<std::string, original_idx_op> const idx_ops = {
        {"half_pixel",
         [=](std::size_t, std::size_t, std::size_t idx, double scale) {
             return (idx + 0.5) / scale - 0.5;
         }},
        {"pytorch_half_pixel",
         [=](std::size_t, std::size_t l_out, std::size_t idx, double scale) {
             return l_out > 1 ? (idx + 0.5) / scale - 0.5 : 0.0;
         }},
        {"align_corners",
         [=](std::size_t l_in, std::size_t l_out, std::size_t idx, double) {
81
             return (l_out == 1) ? 0.0 : (1.0 * idx * (l_in - 1.0) / (l_out - 1.0));
Paul Fultz II's avatar
Paul Fultz II committed
82
83
84
85
86
87
88
         }},
        {"asymmetric",
         [=](std::size_t, std::size_t, std::size_t idx, double scale) { return idx / scale; }},
        {"tf_half_pixel_for_nn", [=](std::size_t, std::size_t, std::size_t idx, double scale) {
             return (idx + 0.5) / scale;
         }}};

89
    if(not contains(idx_ops, mode))
Paul Fultz II's avatar
Paul Fultz II committed
90
91
92
93
94
95
96
    {
        MIGRAPHX_THROW("PARSE_RESIZE: coordinate_transformation_mode " + mode + " not supported!");
    }

    return idx_ops.at(mode);
}

97
98
99
static std::vector<int>
calc_neighbor_points(const std::vector<std::vector<std::vector<std::size_t>>>& vvv_ind,
                     int i_dim,
100
                     std::vector<std::vector<std::size_t>> vec_dims,
101
102
103
104
                     const shape& in_s)
{
    if(i_dim == vvv_ind.size())
    {
105
        std::vector<int> vec_ind(vec_dims.size());
106
107
108
109
110
111
        std::transform(vec_dims.begin(), vec_dims.end(), vec_ind.begin(), [&](auto idx) {
            return static_cast<int>(in_s.index(idx));
        });
        return vec_ind;
    }

112
    const auto& vv_lo = vvv_ind[i_dim][0];
113
114
115
116
117
118
119
120
121
122
123
124
125
    std::vector<std::vector<std::size_t>> vec_dims1;
    for(std::size_t start = 0; start < vec_dims.size(); start += vv_lo.size())
    {
        std::transform(vv_lo.begin(),
                       vv_lo.end(),
                       vec_dims.begin() + start,
                       std::back_inserter(vec_dims1),
                       [](auto i, auto dim) {
                           dim.push_back(i);
                           return dim;
                       });
    }

126
127
    const auto& vv_hi = vvv_ind[i_dim][1];
    for(std::size_t start = 0; start < vec_dims.size(); start += vv_hi.size())
128
129
130
131
132
133
134
135
136
137
    {
        std::transform(vv_hi.begin(),
                       vv_hi.end(),
                       vec_dims.begin() + start,
                       std::back_inserter(vec_dims1),
                       [](auto i, auto dim) {
                           dim.push_back(i);
                           return dim;
                       });
    }
138
139
    vec_dims.clear();
    return calc_neighbor_points(vvv_ind, i_dim + 1, std::move(vec_dims1), in_s);
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
}

static std::string get_coord_trans_mode(const onnx_parser::attribute_map& attr)
{
    std::string coord_trans_mode = "half_pixel";
    if(contains(attr, "coordinate_transformation_mode"))
    {
        coord_trans_mode = attr.at("coordinate_transformation_mode").s();
        // does not support transformation mode "tf_crop_and_resize"
        if(coord_trans_mode == "tf_crop_and_resize")
        {
            MIGRAPHX_THROW("PARSE_RESIZE: \"tf_crop_and_resize\" mode is not supported!");
        }
    }

    return coord_trans_mode;
}

static std::string get_mode(const onnx_parser::attribute_map& attr)
{
    std::string mode = "nearest";
    if(contains(attr, "mode"))
    {
        mode = attr.at("mode").s();
        if(mode != "nearest" and mode != "linear")
        {
            MIGRAPHX_THROW("PARSE_RESIZE: only nearest and linear modes are supported!");
        }
    }

    return mode;
}

static std::string get_nearest_mode(const onnx_parser::attribute_map& attr)
{
    std::string nearest_mode = "round_prefer_floor";
    if(contains(attr, "nearest_mode"))
    {
        nearest_mode = attr.at("nearest_mode").s();
    }

    return nearest_mode;
}

Paul Fultz II's avatar
Paul Fultz II committed
184
185
struct parse_resize : op_parser<parse_resize>
{
186
187
188
    std::vector<op_desc> operators() const { 
        return {{"Resize"}, {"Upsample"}}; 
        }
Paul Fultz II's avatar
Paul Fultz II committed
189

190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
    // A helper for one case of parse().
    // Dynamic batch:  Only args[0] can have a dynamic shape, only the 0'th
    // dimension--batch size--can be non-fixed, and the only resize mode allowed is "nearest"
    instruction_ref dynamic_nearest_parse(const std::vector<size_t>& out_lens,
                                          const std::vector<double>& vec_scale,
                                          const op_desc& opd,
                                          onnx_parser::node_info& info,
                                          const std::vector<instruction_ref>& args) const
    {
        // coord transform mode
        std::string coord_trans_mode = get_coord_trans_mode(info.attributes);
        // mode: only nearest and linear modes are supported for now
        std::string mode = get_mode(info.attributes);

        // rounding option when using "nearest"
        std::string nearest_mode = get_nearest_mode(info.attributes);

        if(mode == "nearest")
        {
            auto some_dims = args[0]->get_shape().dyn_dims();

            bool mostly_fixed =
                std::all_of(some_dims.begin() + 1,
                            some_dims.end(),
214
                            [](const shape::dynamic_dimension& dd) { return dd.is_fixed(); });
215
216

            if(not mostly_fixed)
217
                MIGRAPHX_THROW("PARSE_" + opd.onnx_name +
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
                               ": dynamic shape inputs other than batch size are not supported");

            // Get static dimension set and
            // Drop the 0'th dimension,
            auto fixed_dims = args[0]->get_shape().to_static(1).lens();
            fixed_dims.erase(fixed_dims.begin());
            // dimensions of the (scaled) output, also with the 0'th dimension dropped
            auto fixed_out_lens = out_lens;
            fixed_out_lens.erase(fixed_out_lens.begin());

            // create a shape with the scaled lens and no batch dimension
            migraphx::shape static_out_shape(args[0]->get_shape().type(), fixed_out_lens);

            //               map out_idx to in_idx
            auto idx_op     = get_original_idx_op(coord_trans_mode);
            auto nearest_op = get_nearest_op(nearest_mode);

            // For each element of static_out_shape, find the matching location of input shape.
            // The indexes we find will be an argument to the gather op.
            shape_for_each(static_out_shape, [&](const auto& out_idx_v, size_t) {
                std::vector<size_t> in_idx(out_idx_v.size());
                for(auto ii = 0; ii < fixed_dims.size(); ++ii)
                {
                    // Convert this index by scaling.
                    auto idx_val =
                        idx_op(fixed_dims[ii], fixed_out_lens[ii], out_idx_v[ii], vec_scale[ii]);
                    // round the scaled value to an int index
                    in_idx[ii] = nearest_op(fixed_dims[ii], idx_val);
                }
            });

            instruction_ref gather_ins{args[0]};
            // for each static dimension
            for(auto ii = 0; ii < fixed_dims.size(); ++ii)
            {
                std::vector<size_t> in_idx(fixed_out_lens[ii]);
                // for range of this dimension's size in output
                for(auto len : range(fixed_out_lens[ii]))
                {
                    // Convert this index by scaling.
                    auto idx_val =
                        idx_op(fixed_dims[ii], fixed_out_lens[ii], len, vec_scale[ii + 1]);

                    // round the scaled value to an index
                    in_idx[len] = nearest_op(fixed_dims[ii], idx_val);
                    // Put the value into index vector
                }
                // Create a 1D shape literal
266
                auto index_lit = info.add_literal(literal(
267
268
269
270
                    migraphx::shape(migraphx::shape::int64_type, {fixed_out_lens[ii]}), in_idx));

                // add a "gather" instruction
                gather_ins = info.add_instruction(
271
                    make_op("gather", {{"axis", 1 + ii}}), gather_ins, index_lit);
272
273
274
275
276
277
278
279
280
            }
            return gather_ins;
        }
        else
        {
            MIGRAPHX_THROW("PARSE_RESIZE: only nearest_mode supports dynamic batch size input");
        }
    }

Shucai Xiao's avatar
Shucai Xiao committed
281
    instruction_ref parse(const op_desc& opd,
282
                          const onnx_parser& parser,
Paul Fultz II's avatar
Paul Fultz II committed
283
284
285
                          onnx_parser::node_info info,
                          std::vector<instruction_ref> args) const
    {
286
287
        // coord transform mode
        std::string coord_trans_mode = get_coord_trans_mode(info.attributes);
Paul Fultz II's avatar
Paul Fultz II committed
288

289
290
        // mode: only nearest and linear modes are supported for now
        std::string mode = get_mode(info.attributes);
Paul Fultz II's avatar
Paul Fultz II committed
291
292

        // nearest mode
293
        std::string nearest_mode = get_nearest_mode(info.attributes);
Paul Fultz II's avatar
Paul Fultz II committed
294
295

        // check exclude_outside, only support 0
296
297
        if(contains(info.attributes, "exclude_outside") and
           info.attributes.at("exclude_outside").i() == 1)
Paul Fultz II's avatar
Paul Fultz II committed
298
        {
299
            MIGRAPHX_THROW("PARSE_" + opd.onnx_name + ": exclude_outside 1 is not supported!");
Paul Fultz II's avatar
Paul Fultz II committed
300
301
        }

302
        // input data shape info.  Convert static lens to dynamic to simplify referencing them later
303
        auto in_s = args[0]->get_shape().to_dynamic();
304
        if(args[0]->get_shape().dynamic() and in_s.ndim() < 2)
305
            MIGRAPHX_THROW(
306
                "PARSE_" + opd.onnx_name +
307
                ": requires 2 or more dimensions input, where first dimension is batch #");
308
        std::vector<migraphx::shape::dynamic_dimension> in_dims = in_s.dyn_dims();
Paul Fultz II's avatar
Paul Fultz II committed
309
310

        // output shape is explicitly specified
311
        std::vector<size_t> out_lens(in_s.ndim());
Paul Fultz II's avatar
Paul Fultz II committed
312
313

        // scale
314
        std::vector<double> vec_scale;
Paul Fultz II's avatar
Paul Fultz II committed
315

316
        for(const auto& arg : args)
Paul Fultz II's avatar
Paul Fultz II committed
317
        {
318
319
            if(arg != args[0] and arg->get_shape().dynamic())
            {
320
                MIGRAPHX_THROW("PARSE_" + opd.onnx_name +
321
                            ": no dynamic input shapes allowed except the first one");
322
            }
323
324
325
326
327
328
329
330
331
        }

        // Special-case handling for the deprecated Upsample-7 op, which handles scales as an 
        // attribute rather than an input.
        if(opd.onnx_name == "Upsample" and contains(info.attributes, "scales"))
        {
            // "scales" attribute is a vector of float
            literal scales = parser.parse_value(info.attributes.at("scales"));
            scales.visit([&](auto s) { vec_scale.assign(s.begin(), s.end()); });
332

333
            if(in_dims.size() != vec_scale.size())
334
            {
335
336
                MIGRAPHX_THROW("PARSE_" + opd.onnx_name +
                            ": specified scales rank does not match input shape rank");
337
            }
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
            std::transform(in_dims.begin(),
                           in_dims.end(),
                           vec_scale.begin(),
                           out_lens.begin(),
                           [&](auto idx, auto scale) {
                               // inferred output size is floor(idx.max * scale)
                               return idx.max * scale;
                           });
        }
        else
        {
            // Search the input list for either output size or scale, depending on input type.
            // The first non-empty input after the input shape is assumed to be it, since we can't read
            // Onnx input names.
            //      Todo: The input "ROI" is not currently supported
            for(const auto& arg : args)
Paul Fultz II's avatar
Paul Fultz II committed
354
            {
355
356
357
                // skip first input and any empty inputs
                auto lens = arg->get_shape().to_static(1).lens();
                if(arg->name() == "undefined" or (arg == args[0]) or lens.empty())
358
                {
359
                    continue;
360
                }
361
                auto type = arg->get_shape().type();
362

363
364
365
                // This input is inferred to mean output size if type == int64_type; otherwise
                // read it as the scales
                if(type == shape::int64_type)
366
                {
367
368
369
370
371
372
373
                    auto arg_out_s = arg->eval();
                    check_arg_empty(arg_out_s,
                                    "PARSE_" + opd.onnx_name + ": dynamic output size is not supported!");
                    out_lens.clear();
                    arg_out_s.visit([&](auto ol) { out_lens.assign(ol.begin(), ol.end()); });

                    if(out_lens.size() != in_s.ndim())
374
375
                    {
                        MIGRAPHX_THROW("PARSE_" + opd.onnx_name +
376
377
378
379
380
381
382
383
384
385
386
387
                                       ": specified output rank does not match input rank");
                    }

                    // compute the scale in each dimension
                    vec_scale.resize(in_s.ndim());

                    std::transform(in_dims.begin(),
                                   in_dims.end(),
                                   out_lens.begin(),
                                   vec_scale.begin(),
                                   [](auto iss, auto oss) { return double(1.0 * oss / iss.max); });
                    break;
388
                }
389
390
                else
                {
391
392
                    // scale input
                    auto arg_scale = arg->eval();
393
                    check_arg_empty(arg_scale,
394
395
                                    "PARSE_" + opd.onnx_name +
                                    ": dynamic input scale is not supported!");
396
397
398
399
400

                    arg_scale.visit([&](auto v) { vec_scale.assign(v.begin(), v.end()); });
                    if(in_dims.size() != vec_scale.size())
                    {
                        MIGRAPHX_THROW("PARSE_" + opd.onnx_name +
401
                                       ": specified scale rank does not match input rank");
402
                    }
403
404
405
406
407
408
409
410
411
412
                    std::transform(in_dims.begin(),
                                   in_dims.end(),
                                   vec_scale.begin(),
                                   out_lens.begin(),
                                   [&](auto idx, auto scale) {
                                       // inferred output size is floor(idx.max * scale)
                                       return idx.max * scale;
                                   });
                    break;
                }
413
            }
414
        }
Brian Pickrell's avatar
tidy  
Brian Pickrell committed
415
        if(out_lens.empty())
416
417
            MIGRAPHX_THROW("PARSE_" + opd.onnx_name +
                           ": no input was given for scale or output size");
Paul Fultz II's avatar
Paul Fultz II committed
418

419
        // Dynamic batch:  Only args[0] can have a dynamic shape, only the 0'th
420
        // dimension--batch size--can be non-fixed, and the only resize mode supported is "nearest"
421
        if(args[0]->get_shape().dynamic())
422
        {
423
            return dynamic_nearest_parse(out_lens, vec_scale, opd, info, args);
424
425
426
        }
        else
        {
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
            //
            //        Static input shape.
            //
            in_s         = args[0]->get_shape();
            auto in_lens = args[0]->get_shape().lens();

            shape out_s{in_s.type(), out_lens};
            std::size_t out_elements = out_s.elements();
            auto idx_op              = get_original_idx_op(coord_trans_mode);

            // reshape input to one-dimension
            std::vector<int64_t> rsp_lens = {static_cast<int64_t>(in_s.elements())};
            args[0]                       = info.make_contiguous(args[0]);
            auto rsp = info.add_instruction(make_op("reshape", {{"dims", rsp_lens}}), args[0]);

            if(mode == "nearest")
            {
                std::vector<int> ind(out_elements);

                // map out_idx to in_idx
                auto nearest_op = get_nearest_op(nearest_mode);
                shape_for_each(out_s, [&](const auto& out_idx_v, size_t out_idx) {
                    std::vector<size_t> in_idx(out_idx_v.size());
                    for(auto ii = 0; ii < in_lens.size(); ++ii)
                    {
                        auto idx_val =
                            idx_op(in_lens[ii], out_lens[ii], out_idx_v[ii], vec_scale[ii]);
                        in_idx[ii] = nearest_op(in_lens[ii], idx_val);
                    }
Paul Fultz II's avatar
Paul Fultz II committed
456

457
458
                    ind[out_idx] = static_cast<int64_t>(in_s.index(in_idx));
                });
459

460
461
462
463
464
465
                shape ind_s{shape::int32_type, out_lens};
                auto ins_ind = info.add_literal(literal(ind_s, ind));
                return info.add_instruction(make_op("gather", {{"axis", 0}}), rsp, ins_ind);
            }
            // linear mode
            else
Paul Fultz II's avatar
Paul Fultz II committed
466
            {
467
468
469
470
471
                auto nearest_floor = get_nearest_op("floor");
                auto nearest_ceil  = get_nearest_op("ceil");

                // get the number of dimensions
                std::size_t n_dim = out_lens.size();
Brian Pickrell's avatar
Brian Pickrell committed
472
473
474
                std::vector<std::vector<std::size_t>> vv_ind(
                    2, std::vector<std::size_t>(out_elements));
                std::vector<std::vector<std::vector<std::size_t>>> vvv_ind(n_dim, vv_ind);
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
                std::vector<std::vector<float>> delta(n_dim, std::vector<float>(out_elements));

                shape_for_each(out_s, [&](const auto& out_idx_v, size_t out_idx) {
                    for(auto ii = 0; ii < in_lens.size(); ++ii)
                    {
                        auto idx_val =
                            idx_op(in_lens[ii], out_lens[ii], out_idx_v[ii], vec_scale[ii]);
                        vvv_ind[ii][0][out_idx] = nearest_floor(in_lens[ii], idx_val);
                        vvv_ind[ii][1][out_idx] = nearest_ceil(in_lens[ii], idx_val);
                        delta[ii][out_idx]      = idx_val - vvv_ind[ii][0][out_idx];
                    }
                });

                auto ind = calc_neighbor_points(
                    vvv_ind, 0, std::vector<std::vector<std::size_t>>(out_elements), in_s);
                auto ind_lens = out_lens;
                ind_lens[0] *= (std::size_t{1} << n_dim);
                shape ind_s{shape::int32_type, ind_lens};
                auto ins_ind = info.add_literal(literal(ind_s, ind));
                auto data    = info.add_instruction(make_op("gather", {{"axis", 0}}), rsp, ins_ind);

                auto dim_lens = out_lens;
                dim_lens[0] *= (std::size_t{1} << (n_dim - 1));
                for(std::size_t i = 0; i < n_dim; ++i)
499
                {
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
                    shape dim_s{shape::float_type, dim_lens};
                    const auto& dim_delta = delta[n_dim - i - 1];
                    std::vector<float> delta_data;
                    for(std::size_t j = 0; j < dim_lens[0] / out_lens[0]; ++j)
                    {
                        delta_data.insert(delta_data.begin(), dim_delta.begin(), dim_delta.end());
                    }
                    auto ins_delta = info.add_literal(dim_s, delta_data);

                    // slice the data
                    int64_t slc_stride = dim_lens[0];
                    auto low           = info.add_instruction(
                        make_op("slice", {{"axes", {0}}, {"starts", {0}}, {"ends", {slc_stride}}}),
                        data);
                    auto hi = info.add_instruction(
                        make_op(
                            "slice",
517
                            {{"axes", {0}}, {"starts", {slc_stride}}, {"ends", {2 * slc_stride}}}),
518
519
520
521
522
523
                        data);
                    auto diff = info.add_instruction(make_op("sub"), hi, low);
                    auto ddf  = info.add_instruction(make_op("mul"), diff, ins_delta);
                    data      = info.add_instruction(make_op("add"), ddf, low);
                    dim_lens[0] /= 2;
                }
Paul Fultz II's avatar
Paul Fultz II committed
524

525
526
                return data;
            }
527
        }
Paul Fultz II's avatar
Paul Fultz II committed
528
529
530
531
    }
};

} // namespace onnx
532

Paul Fultz II's avatar
Paul Fultz II committed
533
534
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx