lowering.cpp 32.6 KB
Newer Older
Paul's avatar
Paul committed
1

Paul's avatar
Paul committed
2
3
4
5
6
7
#include <migraphx/cpu/lowering.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/dfor.hpp>
#include <migraphx/operators.hpp>
#include <migraphx/shape_for_each.hpp>
#include <migraphx/iterator_for.hpp>
Paul's avatar
Paul committed
8
#include <migraphx/par_dfor.hpp>
Paul's avatar
Paul committed
9
#include <migraphx/cpu/gemm.hpp>
Paul's avatar
Paul committed
10
#include <unordered_map>
Paul's avatar
Paul committed
11
#include <utility>
Paul's avatar
Paul committed
12

Paul's avatar
Paul committed
13
namespace migraphx {
Paul's avatar
Paul committed
14
inline namespace MIGRAPHX_INLINE_NS {
Paul's avatar
Paul committed
15
16
17
18
19
20
21
22
namespace cpu {

template <typename T>
T zero(const T&)
{
    return T(0);
}

Khalique's avatar
Khalique committed
23
24
25
26
template <class T>
typename std::conditional_t<std::is_integral<T>{}, std::make_signed<T>, std::enable_if<true, T>>::
    type
    make_signed(T x)
Khalique's avatar
Khalique committed
27
28
29
30
{
    return x;
}

31
32
33
34
//
// cpu implemenataion of batch norm for inference
//
// inputs are:
35
36
37
38
// args[0] -> input data buffer
// args[1] -> mini batch mean
// args[2] -> mini batch variance
// args[3] -> gamma
Aditya Atluri's avatar
Aditya Atluri committed
39
// args[4] -> bias
40
41
42
//
// The equation to compute batch norm for inference is:
//
Aditya Atluri's avatar
Aditya Atluri committed
43
// output[i] = bias + gamma * (input[i] + mean) / sqrt(variance + epsilon)
44
45
46
47
48
//
// the input data format should be nchw
//
struct cpu_batch_norm_inference
{
49
    op::batch_norm_inference op;
50

51
52
    std::string name() const { return "cpu::batch_norm_inference"; }

Paul's avatar
Paul committed
53
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
54

Paul's avatar
Paul committed
55
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
56
    {
57
58
        argument output{output_shape};

Aditya Atluri's avatar
Aditya Atluri committed
59
60
        double epsilon           = op.epsilon;
        auto input               = args[0];
Paul's avatar
Paul committed
61
62
63
64
        auto arg_gamma           = args[1];
        auto arg_bias            = args[2];
        auto mini_batch_mean     = args[3];
        auto mini_batch_variance = args[4];
65

66
        auto num_batch    = output_shape.lens()[0];
Aditya Atluri's avatar
Aditya Atluri committed
67
68
        auto num_channels = output_shape.lens()[1];
        auto image_height = output_shape.lens()[2];
69
        auto image_width  = output_shape.lens()[3];
Aditya Atluri's avatar
Aditya Atluri committed
70

71
        if(op.bn_mode == op::batch_norm_inference::spatial)
Scott Thornton's avatar
Scott Thornton committed
72
73
74
75
        {
            visit_all(output, input, mini_batch_mean, mini_batch_variance, arg_gamma, arg_bias)(
                [&](auto result, auto buffer, auto mean, auto variance, auto gamma, auto bias) {

Paul's avatar
Paul committed
76
                    par_dfor(num_batch, num_channels, image_height, image_width)(
Scott Thornton's avatar
Scott Thornton committed
77
                        [&](std::size_t n, std::size_t c, std::size_t h, std::size_t w) {
78
79
80
81
                            assert((variance[c] + epsilon) > 0);
                            result(n, c, h, w) = gamma[c] * (buffer(n, c, h, w) - mean[c]) /
                                                     std::sqrt(variance[c] + epsilon) +
                                                 bias[c];
Scott Thornton's avatar
Scott Thornton committed
82
83
                        });
                });
84
85
        }

86
        if(op.bn_mode == op::batch_norm_inference::per_activation)
Scott Thornton's avatar
Scott Thornton committed
87
        {
88
89
90
            visit_all(output, input, mini_batch_mean, mini_batch_mean, arg_gamma, arg_bias)(
                [&](auto result, auto buffer, auto mean, auto variance, auto gamma, auto bias) {

Paul's avatar
Paul committed
91
                    par_dfor(num_batch, num_channels, image_height, image_width)(
92
                        [&](std::size_t n, std::size_t c, std::size_t h, std::size_t w) {
Paul's avatar
Paul committed
93
                            assert((variance(c, h, w) + epsilon) > 0);
Scott Thornton's avatar
Scott Thornton committed
94
95
96
97
98
99
                            result(n, c, h, w) = gamma(c, h, w) *
                                                     (buffer(n, c, h, w) - mean(c, h, w)) /
                                                     std::sqrt(variance(c, h, w) + epsilon) +
                                                 bias(c, h, w);
                        });
                });
100
        }
101
102
103
104
105

        return output;
    }
};

Khalique's avatar
Khalique committed
106
struct cpu_lrn
Khalique's avatar
Khalique committed
107
{
Khalique's avatar
Khalique committed
108
    op::lrn op;
Khalique's avatar
Khalique committed
109

Khalique's avatar
Khalique committed
110
    std::string name() const { return "cpu::lrn"; }
Khalique's avatar
Khalique committed
111
112
113
114
115
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
    argument compute(context&, shape output_shape, std::vector<argument> args) const
    {
        argument result{output_shape};
        visit_all(result, args[0])([&](auto output, auto input) {
Khalique's avatar
Khalique committed
116
117
118
119
            int n_batch         = output_shape.lens()[0];
            int channels        = output_shape.lens()[1];
            int height          = output_shape.lens()[2];
            int width           = output_shape.lens()[3];
Khalique's avatar
Khalique committed
120
            float alphaoverarea = op.alpha / op.size;
Khalique's avatar
Khalique committed
121
            int radius          = (op.size - 1) / 2;
Khalique's avatar
Khalique committed
122

123
            par_dfor(n_batch, height, width)([&](int b, int h, int w) {
Khalique's avatar
Khalique committed
124
                float scale = 0;
Khalique's avatar
Khalique committed
125
126
                dfor(channels)([&](int c) {
                    auto start = (c - radius) < 0 ? 0 : (c - radius);
Khalique's avatar
Khalique committed
127
                    auto end   = (c + radius) > channels ? channels : (c + radius);
Khalique's avatar
Khalique committed
128
129
                    for(auto k = start; k < end; ++k)
                    {
Khalique's avatar
Khalique committed
130
                        scale += std::pow(input(b, k, h, w), 2);
Khalique's avatar
Khalique committed
131
132
133
                    }
                    scale *= alphaoverarea;
                    scale += op.bias;
Khalique's avatar
Khalique committed
134
                    scale              = std::pow(scale, -op.beta);
Khalique's avatar
Khalique committed
135
136
137
138
139
140
141
142
                    output(b, c, h, w) = input(b, c, h, w) * scale;
                });
            });
        });
        return result;
    }
};

Paul's avatar
Paul committed
143
144
struct cpu_convolution
{
145
    op::convolution op;
Paul's avatar
Paul committed
146
147

    std::string name() const { return "cpu::convolution"; }
Paul's avatar
Paul committed
148
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
Paul's avatar
Paul committed
149
150
151
152
    argument compute(context&, shape output_shape, std::vector<argument> args) const
    {
        argument result{output_shape};
        visit_all(result, args[0], args[1])([&](auto output, auto input, auto weights) {
Khalique's avatar
Khalique committed
153
            auto in   = input.get_shape().lens();
Khalique's avatar
Khalique committed
154
155
            auto in_h = in[2];
            auto in_w = in[3];
Paul's avatar
Paul committed
156

Khalique's avatar
Khalique committed
157
            auto wei   = weights.get_shape().lens();
Khalique's avatar
Khalique committed
158
159
160
161
            auto wei_n = wei[0];
            auto wei_c = wei[1];
            auto wei_h = wei[2];
            auto wei_w = wei[3];
Paul's avatar
Paul committed
162

Paul's avatar
Paul committed
163
            par_dfor(output_shape.lens()[0],
Paul's avatar
Paul committed
164
165
166
                     output_shape.lens()[1],
                     output_shape.lens()[2],
                     output_shape.lens()[3])(
Paul's avatar
Paul committed
167
                [&](std::size_t o, std::size_t w, std::size_t i, std::size_t j) {
Khalique's avatar
Khalique committed
168
169
                    const int start_x  = i * op.stride[0] - op.padding[0];
                    const int start_y  = j * op.stride[1] - op.padding[1];
Khalique's avatar
Khalique committed
170
                    const int group_id = w / (wei_n / op.group);
Paul's avatar
Paul committed
171
172
173

                    double acc = 0;
                    dfor(wei_c, wei_h, wei_w)([&](std::size_t k, std::size_t x, std::size_t y) {
Khalique's avatar
Khalique committed
174
175
                        const int in_x  = start_x + x;
                        const int in_y  = start_y + y;
Khalique's avatar
Khalique committed
176
                        const int in_ch = group_id * wei_c + k;
Paul's avatar
Paul committed
177
178
                        if(in_x >= 0 && in_x < in_h && in_y >= 0 && in_y < in_w)
                        {
Khalique's avatar
Khalique committed
179
                            acc += input(o, in_ch, in_x, in_y) * weights(w, k, x, y);
Paul's avatar
Paul committed
180
181
182
183
184
185
186
187
188
                        }
                    });
                    output(o, w, i, j) = acc;
                });
        });
        return result;
    }
};

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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
struct cpu_quant_convolution
{
    op::quant_convolution op;

    std::string name() const { return "cpu::quant_convolution"; }
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
    argument compute(context&, shape output_shape, std::vector<argument> args) const
    {
        argument result{output_shape};
        visit_all(result, args[0], args[1])([&](auto output, auto input, auto weights) {
            auto in   = input.get_shape().lens();
            auto in_h = in[2];
            auto in_w = in[3];

            auto wei   = weights.get_shape().lens();
            auto wei_n = wei[0];
            auto wei_c = wei[1];
            auto wei_h = wei[2];
            auto wei_w = wei[3];

            par_dfor(output_shape.lens()[0],
                     output_shape.lens()[1],
                     output_shape.lens()[2],
                     output_shape.lens()[3])(
                [&](std::size_t o, std::size_t w, std::size_t i, std::size_t j) {
                    const int start_x  = i * op.stride[0] - op.padding[0];
                    const int start_y  = j * op.stride[1] - op.padding[1];
                    const int group_id = w / (wei_n / op.group);

                    double acc = 0;
                    dfor(wei_c, wei_h, wei_w)([&](std::size_t k, std::size_t x, std::size_t y) {
                        const int in_x  = start_x + x;
                        const int in_y  = start_y + y;
                        const int in_ch = group_id * wei_c + k;
                        if(in_x >= 0 && in_x < in_h && in_y >= 0 && in_y < in_w)
                        {
                            acc += input(o, in_ch, in_x, in_y) * weights(w, k, x, y);
                        }
                    });
                    output(o, w, i, j) = acc;
                });
        });
        return result;
    }
};

Scott Thornton's avatar
Scott Thornton committed
235
236
struct cpu_im2col
{
237
    op::im2col op;
Scott Thornton's avatar
Scott Thornton committed
238

Scott Thornton's avatar
Scott Thornton committed
239
240
    static std::string name() { return "cpu::im2col"; }
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
Scott Thornton's avatar
Scott Thornton committed
241

wsttiger's avatar
wsttiger committed
242
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
Scott Thornton's avatar
Scott Thornton committed
243
    {
Scott Thornton's avatar
Scott Thornton committed
244
        argument result{output_shape};
Scott Thornton's avatar
Scott Thornton committed
245
        auto input_shape   = args[0].get_shape();
Scott Thornton's avatar
Scott Thornton committed
246
247
        auto weights_shape = args[1].get_shape();
        visit_all(result, args[0])([&](auto col, auto input) {
Scott Thornton's avatar
Scott Thornton committed
248
249
            const std::size_t& height   = input_shape.lens()[2];
            const std::size_t& width    = input_shape.lens()[3];
Scott Thornton's avatar
Scott Thornton committed
250
251
252
            const std::size_t& channels = weights_shape.lens()[1];
            const std::size_t& kernel_h = weights_shape.lens()[2];
            const std::size_t& kernel_w = weights_shape.lens()[3];
Scott Thornton's avatar
Scott Thornton committed
253
254
            const std::size_t& pad_h    = op.padding[0];
            const std::size_t& pad_w    = op.padding[1];
Scott Thornton's avatar
Scott Thornton committed
255
256
257
            const std::size_t& stride_h = op.stride[0];
            const std::size_t& stride_w = op.stride[1];

Paul's avatar
Paul committed
258
259
            int kdiv2_h;
            int kdiv2_w;
Scott Thornton's avatar
Scott Thornton committed
260
261
            kdiv2_h = kernel_h / 2;
            kdiv2_w = kernel_w / 2;
Scott Thornton's avatar
Scott Thornton committed
262
            // calculate output sizes
Scott Thornton's avatar
Scott Thornton committed
263
264
            const std::size_t col_height = (height - kernel_h + 2 * pad_h) / stride_h + 1;
            const std::size_t col_width  = (width - kernel_w + 2 * pad_w) / stride_w + 1;
wsttiger's avatar
wsttiger committed
265
            // account for padding for the starting position of the input pixels
Scott Thornton's avatar
Scott Thornton committed
266
            std::size_t iinput = kdiv2_h - pad_h;
wsttiger's avatar
wsttiger committed
267
            // loop over output pixels (ioutput, joutput)
Scott Thornton's avatar
Scott Thornton committed
268
269
270
271
272
273
274
275
            for(std::size_t ioutput = 0; ioutput < col_height; ioutput++, iinput += stride_h)
            {
                std::size_t jinput = kdiv2_w - pad_w;
                for(std::size_t joutput = 0; joutput < col_width; joutput++, jinput += stride_w)
                {
                    // compute linear index for output
                    std::size_t ldx = ioutput * col_width + joutput;
                    std::size_t p   = 0;
wsttiger's avatar
wsttiger committed
276
277
278
279
280
281
282
283
284
285
                    dfor(channels,
                         kernel_h,
                         kernel_w)([&](std::size_t c, std::size_t koffset, std::size_t loffset) {
                        int idx     = iinput + koffset - kdiv2_h;
                        int jdx     = jinput + loffset - kdiv2_w;
                        col(ldx, p) = ((idx >= 0) && (idx < height) && (jdx >= 0) && (jdx < width))
                                          ? input(0, c, idx, jdx)
                                          : 0;
                        p++;
                    });
Scott Thornton's avatar
Scott Thornton committed
286
287
                }
            }
Scott Thornton's avatar
Scott Thornton committed
288
        });
Scott Thornton's avatar
Scott Thornton committed
289
290
291
292
        return result;
    }
};

Paul's avatar
Paul committed
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
struct max_pool
{
    static std::string name() { return "max"; }
    static double start() { return std::numeric_limits<double>::lowest(); }

    static double apply(double x, double y)
    {
        double m = std::max(x, y);
        return (m);
    }

    static double final(double x, double) { return (x); }
};

struct avg_pool
{
    static std::string name() { return "average"; }
    static double start() { return 0.0; }

    static double apply(double x, double y) { return x + y; }

    static double final(double x, double y) { return x / y; }
};

template <class Op>
struct cpu_pooling
{
320
    op::pooling op;
Paul's avatar
Paul committed
321
322

    std::string name() const { return "cpu::pooling_" + Op::name(); }
Paul's avatar
Paul committed
323
324
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
Paul's avatar
Paul committed
325
326
327
328
329
330
331
    {
        argument result{output_shape};
        visit_all(result, args[0])([&](auto output, auto input) {
            using type = typename decltype(output)::value_type;
            auto in_h  = input.get_shape().lens()[2];
            auto in_w  = input.get_shape().lens()[3];

Paul's avatar
Paul committed
332
            par_dfor(output_shape.lens()[0],
Paul's avatar
Paul committed
333
334
335
                     output_shape.lens()[1],
                     output_shape.lens()[2],
                     output_shape.lens()[3])(
Paul's avatar
Paul committed
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
                [&](std::size_t o, std::size_t w, std::size_t i, std::size_t j) {
                    const int start_x0 = i * op.stride[0] - op.padding[0];
                    const int start_y0 = j * op.stride[1] - op.padding[1];

                    const int hend = std::min(start_x0 + op.lengths[0], in_h);
                    const int wend = std::min(start_y0 + op.lengths[1], in_w);

                    const int start_x = std::max(start_x0, 0);
                    const int start_y = std::max(start_y0, 0);

                    const int w_h       = (hend - start_x);
                    const int w_w       = (wend - start_y);
                    const int pool_size = std::max(w_h * w_w, 1);

                    double acc = Op::start();
                    dfor(w_h, w_w)([&](int x, int y) {
                        const int in_x = start_x + x;
                        const int in_y = start_y + y;
                        if(in_x >= 0 && in_x < in_h && in_y >= 0 && in_y < in_w)
                        {
                            acc = Op::apply(acc, input(o, w, in_x, in_y));
                        }
                    });
                    output(o, w, i, j) = type(Op::final(acc, pool_size));
                });
        });
        return result;
    }
};

struct cpu_contiguous
{
368
    op::contiguous op;
Paul's avatar
Paul committed
369
    std::string name() const { return "cpu::contiguous"; }
Paul's avatar
Paul committed
370
371
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
Paul's avatar
Paul committed
372
    {
Paul's avatar
Paul committed
373
        return op.compute(output_shape, std::move(args));
Paul's avatar
Paul committed
374
375
376
    }
};

Khalique's avatar
Khalique committed
377
struct cpu_pad
378
{
Khalique's avatar
Khalique committed
379
380
    op::pad op;
    std::string name() const { return "cpu::contiguous"; }
381
382
383
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
    {
Khalique's avatar
Khalique committed
384
        assert(output_shape.standard());
385
        argument result{output_shape};
Khalique's avatar
Khalique committed
386
        result.visit([&](auto output) { std::fill(output.begin(), output.end(), op.value); });
Khalique's avatar
Khalique committed
387
388

        visit_all(result, args[0])([&](auto output, auto input) {
389
            shape_for_each(input.get_shape(), [&](const auto& idx) {
Khalique's avatar
Khalique committed
390
                std::vector<std::size_t> new_idx(idx.size());
Khalique's avatar
Khalique committed
391
392
393
394
                std::transform(
                    idx.begin(), idx.end(), op.pads.begin(), new_idx.begin(), [](auto i, auto j) {
                        return i + j;
                    });
Khalique's avatar
Khalique committed
395
                output(new_idx.begin(), new_idx.end()) = input(idx.begin(), idx.end());
396
            });
Khalique's avatar
Khalique committed
397
398
        });

399
400
401
402
403
404
405
406
407
408
409
        return result;
    }
};

struct cpu_concat
{
    op::concat op;
    std::string name() const { return "cpu::concat"; }
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
    {
Paul's avatar
Paul committed
410
        return op.compute(output_shape, std::move(args));
411
412
413
    }
};

Paul's avatar
Paul committed
414
415
struct cpu_gemm
{
Shucai Xiao's avatar
Shucai Xiao committed
416
417
    op::dot op;
    std::string name() const { return "cpu::dot"; }
Shucai Xiao's avatar
Shucai Xiao committed
418
419
    shape compute_shape(const std::vector<shape>& inputs) const
    {
Shucai Xiao's avatar
Shucai Xiao committed
420
421
422
423
424
        if(inputs.size() == 3)
        {
            auto c_shape = inputs.at(2);
            check_shapes{{c_shape}}.not_broadcasted();
        }
Shucai Xiao's avatar
Shucai Xiao committed
425
        return op.compute_shape(inputs);
Shucai Xiao's avatar
Shucai Xiao committed
426
    }
Paul's avatar
Paul committed
427

Paul's avatar
Paul committed
428
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
Paul's avatar
Paul committed
429
430
    {
        argument result{output_shape};
Shucai Xiao's avatar
Shucai Xiao committed
431
        // 3 inputs, it is alpha * A * B + beta * C, then
432
        // A and B are matrices, and C is of the same shape as A * B
Shucai Xiao's avatar
Shucai Xiao committed
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
        if(args.size() == 3)
        {
            // no need to consider the value of args[2]
            if(op.beta == 0.0f)
            {
                result.visit([&](auto output) { std::fill(output.begin(), output.end(), 0); });
            }
            else
            {
                visit_all(result, args[2])([&](auto output, auto input) {
                    std::copy(input.begin(), input.end(), output.begin());
                });
            }

            migemm(result, args[0], args[1], op.alpha, op.beta);

            return result;
        }

        // 2 input arguments
        migemm(result, args[0], args[1], op.alpha, 0.0f);

Paul's avatar
Paul committed
455
456
457
458
        return result;
    }
};

459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
struct cpu_quant_gemm
{
    op::quant_dot op;
    std::string name() const { return "cpu::quant_dot"; }
    shape compute_shape(const std::vector<shape>& inputs) const
    {
        if(inputs.size() == 3)
        {
            auto c_shape = inputs.at(2);
            check_shapes{{c_shape}}.not_broadcasted();
        }
        return op.compute_shape(inputs);
    }

    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
    {
        argument result{output_shape};
        // 3 inputs, it is alpha * A * B + beta * C, then
        // A and B are matrices, and C is of the same shape to A * B

        // first, convert the args[0] and args[1] from int8_t to int32_t
        argument arg_0{{shape::int32_type, {args.at(0).get_shape().lens()}}};
        argument arg_1{{shape::int32_type, {args.at(1).get_shape().lens()}}};
        arg_0.visit([&](auto output) {
Shucai Xiao's avatar
Shucai Xiao committed
483
484
            args.at(0).visit(
                [&](auto input) { std::copy(input.begin(), input.end(), output.begin()); });
485
486
487
        });

        arg_1.visit([&](auto output) {
Shucai Xiao's avatar
Shucai Xiao committed
488
489
            args.at(1).visit(
                [&](auto input) { std::copy(input.begin(), input.end(), output.begin()); });
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
        });

        if(args.size() == 3)
        {
            // no need to consider the value of args[2]
            if(op.beta == 0)
            {
                result.visit([&](auto output) { std::fill(output.begin(), output.end(), 0); });
            }
            else
            {
                visit_all(result, args[2])([&](auto output, auto input) {
                    std::copy(input.begin(), input.end(), output.begin());
                });
            }

            migemm(result, arg_0, arg_1, op.alpha, op.beta);

            return result;
        }

        // 2 input arguments
        int8_t beta = 0;
        migemm(result, arg_0, arg_1, op.alpha, beta);

        return result;
    }
};

519
520
521
522
523
524
525
526
struct cpu_gather
{
    op::gather op;
    std::string name() const { return "cpu::gather"; }
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }

    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
    {
Shucai Xiao's avatar
Shucai Xiao committed
527
        return op.compute(output_shape, std::move(args));
528
529
530
    }
};

Paul's avatar
Paul committed
531
532
533
534
535
536
537
538
539
540
541
542
543
544
struct identity_op
{
    std::string name() const { return "cpu::identity"; }
    auto fcn() const
    {
        return [](auto x) { return x; };
    }
};

struct abs_op
{
    std::string name() const { return "cpu::abs"; }
    auto fcn() const
    {
Khalique's avatar
Khalique committed
545
        return [](auto x) { return std::abs(make_signed(x)); };
Paul's avatar
Paul committed
546
547
548
549
550
551
552
553
554
555
556
557
    }
};

struct exp_op
{
    std::string name() const { return "cpu::exp"; }
    auto fcn() const
    {
        return [](auto x) { return std::exp(x); };
    }
};

Shucai Xiao's avatar
Shucai Xiao committed
558
559
560
561
562
563
564
565
566
struct log_op
{
    std::string name() const { return "cpu::log"; }
    auto fcn() const
    {
        return [](auto x) { return std::log(x); };
    }
};

Paul's avatar
Paul committed
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
struct sin_op
{
    std::string name() const { return "cpu::sin"; }
    auto fcn() const
    {
        return [](auto x) { return std::sin(x); };
    }
};

struct cos_op
{
    std::string name() const { return "cpu::cos"; }
    auto fcn() const
    {
        return [](auto x) { return std::cos(x); };
    }
};

struct tan_op
{
    std::string name() const { return "cpu::tan"; }
    auto fcn() const
    {
        return [](auto x) { return std::tan(x); };
    }
};

struct asin_op
{
    std::string name() const { return "cpu::asin"; }
    auto fcn() const
    {
        return [](auto x) { return std::asin(x); };
    }
};

struct acos_op
{
    std::string name() const { return "cpu::acos"; }
    auto fcn() const
    {
        return [](auto x) { return std::acos(x); };
    }
};

struct atan_op
{
    std::string name() const { return "cpu::atan"; }
    auto fcn() const
    {
        return [](auto x) { return std::atan(x); };
    }
};

621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
struct sinh_op
{
    std::string name() const { return "cpu::sinh"; }
    auto fcn() const
    {
        return [](auto x) { return std::sinh(x); };
    }
};

struct cosh_op
{
    std::string name() const { return "cpu::cosh"; }
    auto fcn() const
    {
        return [](auto x) { return std::cosh(x); };
    }
};

Paul's avatar
Paul committed
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
struct tanh_op
{
    std::string name() const { return "cpu::tanh"; }
    auto fcn() const
    {
        return [](auto x) { return std::tanh(x); };
    }
};

struct sigmoid_op
{
    std::string name() const { return "cpu::sigmoid"; }
    auto fcn() const
    {
        return [](auto x) { return 1.f / (1.f + std::exp(-x)); };
    }
};

struct neg_op
{
    std::string name() const { return "cpu::neg"; }
    auto fcn() const
    {
        return [](auto x) { return -x; };
    }
};

struct relu_op
{
    std::string name() const { return "cpu::relu"; }
    auto fcn() const
    {
Paul's avatar
Paul committed
671
        return [](auto x) { return std::max(decltype(x){0}, x); };
Paul's avatar
Paul committed
672
673
674
    }
};

Khalique's avatar
Khalique committed
675
676
677
678
679
680
681
682
683
684
685
struct leaky_relu_op
{
    op::leaky_relu op;
    std::string name() const { return "cpu::leaky_relu"; }
    auto fcn() const
    {
        auto& a = op.alpha;
        return [a](auto x) { return x > 0 ? x : x * a; };
    }
};

Khalique's avatar
Khalique committed
686
687
688
689
690
691
692
693
694
695
696
struct elu_op
{
    op::elu op;
    std::string name() const { return "cpu::elu"; }
    auto fcn() const
    {
        auto& a = op.alpha;
        return [a](auto x) { return x > 0 ? x : a * std::expm1(x); };
    }
};

Paul's avatar
Paul committed
697
698
699
700
701
template <typename Op>
struct cpu_unary
{
    Op op;
    std::string name() const { return op.name(); }
Paul's avatar
Paul committed
702
703
    shape compute_shape(const std::vector<shape>& inputs) const { return inputs.front(); }
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
Paul's avatar
Paul committed
704
705
706
707
708
709
710
    {
        argument result{output_shape};
        result.visit([&](auto output) {
            args[0].visit([&](auto input) {
                std::transform(input.begin(), input.end(), output.begin(), op.fcn());
            });
        });
711

Paul's avatar
Paul committed
712
713
714
715
716
717
718
        return result;
    }
};

struct softmax2d
{
    std::string name() const { return "cpu::softmax2d"; }
Paul's avatar
Paul committed
719
720
    shape compute_shape(const std::vector<shape>& inputs) const { return inputs.front(); }
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
Paul's avatar
Paul committed
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
    {
        argument result{output_shape};
        visit_all(result, args[0])([&](auto output, auto input) {
            using value_type = typename decltype(input)::value_type;
            auto nb          = input.get_shape().lens()[0];
            auto nc          = input.get_shape().lens()[1];
            auto nh          = input.get_shape().lens()[2];
            auto nw          = input.get_shape().lens()[3];
            dfor(nb, nh, nw)([&](std::size_t b, std::size_t i, std::size_t j) {
                value_type cmax = std::numeric_limits<value_type>::lowest();
                for(int c = 0; c < nc; c++)
                {
                    cmax = std::max(cmax, input(b, c, i, j));
                }
                for(int c = 0; c < nc; c++)
                {
                    output(b, c, i, j) = std::exp(input(b, c, i, j) - cmax);
                }
                value_type sum = value_type(0);
                for(int c = 0; c < nc; c++)
                {
                    sum += output(b, c, i, j);
                }
                for(int c = 0; c < nc; c++)
                {
                    output(b, c, i, j) = output(b, c, i, j) / sum;
                }
            });
        });
        return result;
    }
};

Shucai Xiao's avatar
Shucai Xiao committed
754
755
756
757
758
759
struct cpu_logsoftmax
{
    op::logsoftmax op;
    std::string name() const { return "cpu::logsoftmax"; }
    shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }

Shucai Xiao's avatar
Shucai Xiao committed
760
    template <typename T>
Shucai Xiao's avatar
Shucai Xiao committed
761
762
    std::size_t compute_batch_index(const T& idx, shape& batch_shape, int axis) const
    {
Shucai Xiao's avatar
Shucai Xiao committed
763
        if(axis == 0)
764
765
766
767
768
769
        {
            return 0;
        }
        else
        {
            std::vector<std::size_t> batch_idx(idx.begin(), idx.begin() + axis);
Shucai Xiao's avatar
Shucai Xiao committed
770
            return batch_shape.index(batch_idx.begin(), batch_idx.end());
771
        }
Shucai Xiao's avatar
Shucai Xiao committed
772
773
774
775
776
777
    }

    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
    {
        argument result{output_shape};
        auto lens = output_shape.lens();
778
        std::vector<std::size_t> batch_lens{};
Shucai Xiao's avatar
Shucai Xiao committed
779
        if(op.axis == 0)
780
781
782
        {
            batch_lens.push_back(1);
        }
Shucai Xiao's avatar
Shucai Xiao committed
783
        else
784
785
786
        {
            batch_lens.insert(batch_lens.begin(), lens.begin(), lens.begin() + op.axis);
        }
Shucai Xiao's avatar
Shucai Xiao committed
787
788
789
        shape batch_shape{migraphx::shape::uint32_type, batch_lens};
        visit_all(result, args[0])([&](auto output, auto input) {
            using value_type = typename decltype(input)::value_type;
Shucai Xiao's avatar
Shucai Xiao committed
790
791
            std::vector<value_type> batch_max(batch_shape.elements(),
                                              std::numeric_limits<value_type>::lowest());
Shucai Xiao's avatar
Shucai Xiao committed
792
            shape_for_each(output_shape, [&](auto idx) {
793
                auto index       = this->compute_batch_index(idx, batch_shape, op.axis);
Shucai Xiao's avatar
Shucai Xiao committed
794
795
796
797
                batch_max[index] = std::max(batch_max[index], input(idx.begin(), idx.end()));
            });

            shape_for_each(output_shape, [&](auto idx) {
Shucai Xiao's avatar
Shucai Xiao committed
798
                auto index = this->compute_batch_index(idx, batch_shape, op.axis);
Shucai Xiao's avatar
Shucai Xiao committed
799
800
801
802
803
                output(idx.begin(), idx.end()) = input(idx.begin(), idx.end()) - batch_max[index];
            });

            std::vector<value_type> batch_sum(batch_shape.elements(), value_type(0));
            shape_for_each(output_shape, [&](auto idx) {
804
                auto index = this->compute_batch_index(idx, batch_shape, op.axis);
Shucai Xiao's avatar
Shucai Xiao committed
805
806
807
                batch_sum[index] += std::exp(output(idx.begin(), idx.end()));
            });

Shucai Xiao's avatar
Shucai Xiao committed
808
            for(std::size_t i = 0; i < batch_sum.size(); ++i)
Shucai Xiao's avatar
Shucai Xiao committed
809
810
811
812
813
            {
                batch_sum[i] = std::log(batch_sum[i]);
            }

            shape_for_each(output_shape, [&](auto idx) {
814
                auto index = this->compute_batch_index(idx, batch_shape, op.axis);
815
                output(idx.begin(), idx.end()) -= batch_sum[index];
Shucai Xiao's avatar
Shucai Xiao committed
816
817
818
819
820
821
822
            });
        });

        return result;
    }
};

Paul's avatar
Paul committed
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
struct add_op
{
    std::string name() const { return "add"; }
    auto fcn() const
    {
        return [](auto x, auto y) { return x + y; };
    }
};

struct sub_op
{
    std::string name() const { return "sub"; }
    auto fcn() const
    {
        return [](auto x, auto y) { return x - y; };
    }
};

struct mul_op
{
    std::string name() const { return "mul"; }
    auto fcn() const
    {
        return [](auto x, auto y) { return x * y; };
    }
};

struct div_op
{
    std::string name() const { return "div"; }
    auto fcn() const
    {
        return [](auto x, auto y) { return x / y; };
    }
};

Khalique's avatar
Khalique committed
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
struct max_op
{
    std::string name() const { return "max"; }
    auto fcn() const
    {
        return [](auto x, auto y) { return std::max(x, y); };
    }
};

struct min_op
{
    std::string name() const { return "min"; }
    auto fcn() const
    {
        return [](auto x, auto y) { return std::min(x, y); };
    }
};

Paul's avatar
Paul committed
877
878
879
880
881
template <typename Op>
struct cpu_binary
{
    Op op;
    std::string name() const { return op.name(); }
Paul's avatar
Paul committed
882
883
    shape compute_shape(const std::vector<shape>& inputs) const { return inputs.front(); }
    argument compute(context&, const shape& output_shape, std::vector<argument> args) const
Paul's avatar
Paul committed
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
    {
        argument result{output_shape};
        visit_all(result, args[0], args[1])([&](auto output, auto input1, auto input2) {
            if(input1.get_shape().packed() and input2.get_shape().packed())
            {
                std::transform(
                    input1.begin(), input1.end(), input2.begin(), output.begin(), op.fcn());
            }
            else
            {
                shape_for_each(output.get_shape(), [&](const auto& idx) {
                    output(idx.begin(), idx.end()) =
                        op.fcn()(input1(idx.begin(), idx.end()), input2(idx.begin(), idx.end()));
                });
            }
        });
        return result;
    }
};

struct cpu_apply
{
    program* prog;
    std::unordered_map<std::string, std::function<void(instruction_ref)>> apply_map{};

    template <class T>
    auto simple_op()
    {
        return [this](instruction_ref ins) { apply_simple_op<T>(ins); };
    }

    template <class T, class Op>
    auto extend_op()
    {
        return [this](instruction_ref ins) { apply_extend_op<T, Op>(ins); };
    }

    void init()
    {
Shucai Xiao's avatar
Shucai Xiao committed
923
924
        apply_map["im2col"]            = extend_op<cpu_im2col, op::im2col>();
        apply_map["convolution"]       = extend_op<cpu_convolution, op::convolution>();
925
        apply_map["quant_convolution"] = extend_op<cpu_quant_convolution, op::quant_convolution>();
Shucai Xiao's avatar
Shucai Xiao committed
926
927
        apply_map["dot"]               = extend_op<cpu_gemm, op::dot>();
        apply_map["quant_dot"]         = extend_op<cpu_quant_gemm, op::quant_dot>();
Aditya Atluri's avatar
Aditya Atluri committed
928
        apply_map["batch_norm_inference"] =
929
            extend_op<cpu_batch_norm_inference, op::batch_norm_inference>();
Khalique's avatar
Khalique committed
930
        apply_map["lrn"]        = extend_op<cpu_lrn, op::lrn>();
931
        apply_map["contiguous"] = extend_op<cpu_contiguous, op::contiguous>();
Khalique's avatar
Khalique committed
932
        apply_map["pad"]        = extend_op<cpu_pad, op::pad>();
Scott Thornton's avatar
Scott Thornton committed
933
        apply_map["concat"]     = extend_op<cpu_concat, op::concat>();
934
        apply_map["gather"]     = extend_op<cpu_gather, op::gather>();
Shucai Xiao's avatar
Shucai Xiao committed
935
        apply_map["logsoftmax"] = extend_op<cpu_logsoftmax, op::logsoftmax>();
Khalique's avatar
Khalique committed
936
        apply_map["leaky_relu"] = extend_op<cpu_unary<leaky_relu_op>, op::leaky_relu>();
Khalique's avatar
Khalique committed
937
        apply_map["elu"]        = extend_op<cpu_unary<elu_op>, op::elu>();
wsttiger's avatar
wsttiger committed
938
        apply_map["identity"]   = simple_op<cpu_unary<identity_op>>();
Khalique's avatar
Khalique committed
939
        apply_map["abs"]        = simple_op<cpu_unary<abs_op>>();
940
941
        apply_map["sinh"]       = simple_op<cpu_unary<sinh_op>>();
        apply_map["cosh"]       = simple_op<cpu_unary<cosh_op>>();
wsttiger's avatar
wsttiger committed
942
943
944
        apply_map["tanh"]       = simple_op<cpu_unary<tanh_op>>();
        apply_map["sigmoid"]    = simple_op<cpu_unary<sigmoid_op>>();
        apply_map["exp"]        = simple_op<cpu_unary<exp_op>>();
Shucai Xiao's avatar
Shucai Xiao committed
945
        apply_map["log"]        = simple_op<cpu_unary<log_op>>();
wsttiger's avatar
wsttiger committed
946
947
948
949
        apply_map["neg"]        = simple_op<cpu_unary<neg_op>>();
        apply_map["sin"]        = simple_op<cpu_unary<sin_op>>();
        apply_map["cos"]        = simple_op<cpu_unary<cos_op>>();
        apply_map["tan"]        = simple_op<cpu_unary<tan_op>>();
950
951
952
        apply_map["asin"]       = simple_op<cpu_unary<asin_op>>();
        apply_map["acos"]       = simple_op<cpu_unary<acos_op>>();
        apply_map["atan"]       = simple_op<cpu_unary<atan_op>>();
Khalique's avatar
Khalique committed
953
        apply_map["relu"]       = simple_op<cpu_unary<relu_op>>();
wsttiger's avatar
wsttiger committed
954
955
956
957
        apply_map["add"]        = simple_op<cpu_binary<add_op>>();
        apply_map["sub"]        = simple_op<cpu_binary<sub_op>>();
        apply_map["mul"]        = simple_op<cpu_binary<mul_op>>();
        apply_map["div"]        = simple_op<cpu_binary<div_op>>();
Khalique's avatar
Khalique committed
958
959
        apply_map["max"]        = simple_op<cpu_binary<max_op>>();
        apply_map["min"]        = simple_op<cpu_binary<min_op>>();
Paul's avatar
Paul committed
960
961
962
963
964
965
966
967
968

        apply_map["softmax"] = simple_op<softmax2d>();
    }

    void apply()
    {
        init();
        for(auto it : iterator_for(*prog))
        {
Khalique's avatar
Khalique committed
969
            if(it->name() == "pooling")
Paul's avatar
Paul committed
970
971
972
            {
                apply_pooling(it);
            }
Paul's avatar
Paul committed
973
            else if(apply_map.count(it->name()) > 0)
Paul's avatar
Paul committed
974
            {
Paul's avatar
Paul committed
975
                apply_map.at(it->name())(it);
Paul's avatar
Paul committed
976
977
978
979
980
981
982
            }
        }
    }

    template <class T>
    void apply_simple_op(instruction_ref ins)
    {
Paul's avatar
Paul committed
983
        prog->replace_instruction(ins, T{}, ins->inputs());
Paul's avatar
Paul committed
984
985
986
987
988
    }

    template <class T, class Op>
    void apply_extend_op(instruction_ref ins)
    {
989
        auto&& op = any_cast<Op>(ins->get_operator());
Paul's avatar
Paul committed
990
        prog->replace_instruction(ins, T{op}, ins->inputs());
Paul's avatar
Paul committed
991
992
993
994
    }

    void apply_pooling(instruction_ref ins)
    {
995
        auto&& op = any_cast<op::pooling>(ins->get_operator());
Paul's avatar
Paul committed
996
        if(op.mode == "max")
Paul's avatar
Paul committed
997
            prog->replace_instruction(ins, cpu_pooling<max_pool>{op}, ins->inputs());
Paul's avatar
Paul committed
998
        else if(op.mode == "average")
Paul's avatar
Paul committed
999
            prog->replace_instruction(ins, cpu_pooling<avg_pool>{op}, ins->inputs());
Paul's avatar
Paul committed
1000
1001
1002
    }
};

Shucai Xiao's avatar
Shucai Xiao committed
1003
void lowering::apply(program& p) const { cpu_apply{&p}.apply(); }
Paul's avatar
Paul committed
1004
1005

} // namespace cpu
Paul's avatar
Paul committed
1006
} // namespace MIGRAPHX_INLINE_NS
Paul's avatar
Paul committed
1007
} // namespace migraphx