File size: 24,588 Bytes
9375c9a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
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
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
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
184
185
186
187
188
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
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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
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
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
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
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
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
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
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
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
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
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
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
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
// Copyright (C) 2015  Davis E. King ([email protected])
// License: Boost Software License   See LICENSE.txt for the full license.
#ifndef DLIB_DNN_CuDA_H_
#define DLIB_DNN_CuDA_H_


#include "tensor.h"
#include "../geometry/rectangle.h"
#include "../dnn/misc.h"

namespace dlib
{
    namespace cuda 
    {

    // ----------------------------------------------------------------------------------------

        void set_device (
            int dev
        );

        int get_device (
        );

        int get_num_devices (
        );

        std::string get_device_name (
            int device
        );

        void set_current_device_blocking_sync(
        );

        bool can_access_peer (int device_id, int peer_device_id);
        bool can_access_peer (const tensor& device, const tensor& peer_device);

        void device_synchronize (int dev);
        void device_synchronize (const tensor& dev);


        class raii_set_device
        {
        public:
            raii_set_device() = delete;
            raii_set_device(const raii_set_device&) = delete;
            raii_set_device& operator=(const raii_set_device&) = delete;

            raii_set_device(int dev)
            {
                prev_dev = get_device();
                set_device(dev);
            }

            raii_set_device(const tensor& dev)
            {
                prev_dev = get_device();
                set_device(dev.device_id());
            }

            void operator() (int dev)
            {
                set_device(dev);
            }

            void operator() (const tensor& dev)
            {
                set_device(dev.device_id());
            }

            ~raii_set_device() noexcept(false)
            {
                set_device(prev_dev);
            }

        private:
            int prev_dev;
        };


#ifdef DLIB_USE_CUDA

        class enable_peer_access
        {
        public:

            enable_peer_access() = delete;
            enable_peer_access(const enable_peer_access&) = delete;
            enable_peer_access& operator=(const enable_peer_access&) = delete;

            enable_peer_access(
                int device_id,
                int peer_device_id
            );

            enable_peer_access(
                const tensor& device,
                const tensor& peer_device
            ) : enable_peer_access(device.device_id(), peer_device.device_id())
            {}

            ~enable_peer_access() noexcept(false);

        private:

            bool call_disable;
            int device_id;
            int peer_device_id;
        };

    // -----------------------------------------------------------------------------------

        void inverse_norms (
            resizable_tensor& invnorms,
            const tensor& data,
            const double eps
        );

        void dot_prods (
            resizable_tensor& out,
            const tensor& lhs,
            const tensor& rhs
        );

        void dot_prods (
            bool add_to,
            tensor& out,
            const tensor& lhs,
            const tensor& rhs
        );

        void scale_columns (
            tensor& out,
            const tensor& m,
            const tensor& v
        );

        void scale_rows (
            tensor& out,
            const tensor& m,
            const tensor& v
        );

        void scale_rows2 (
            float beta, 
            tensor& out,
            const tensor& m1,
            const tensor& m2,
            const tensor& v1,
            const tensor& v2
        );

        void exp (
            tensor& dest,
            const tensor& src
        );

        void log (
            tensor& dest,
            const tensor& src
        );

        void log10 (
            tensor& dest,
            const tensor& src
        );

    // ------------------------------------------------------------------------------------

        void set_tensor (
            tensor& t,
            float value
        );

        void scale_tensor (
            tensor& t,
            float value
        );

    // ------------------------------------------------------------------------------------

        void multiply (
            bool add_to,
            tensor& dest,
            const tensor& src1,
            const tensor& src2
        );

        void multiply_conv (
            bool add_to,
            tensor& dest,
            const tensor& src1,
            const tensor& src2
        );

        void multiply_zero_padded (
            bool add_to,
            tensor& dest,
            const tensor& src1,
            const tensor& src2
        );

        void scale_channels (
            bool add_to,
            tensor& dest,
            const tensor& src,
            const tensor& scales
        );

        void add (
            tensor& dest,
            const tensor& src1,
            const tensor& src2
        );

    // -----------------------------------------------------------------------------------

        void affine_transform(
            tensor& dest,
            const tensor& src,
            const float A,
            const float B
        );

        void affine_transform(
            tensor& dest,
            const tensor& src,
            const float A
        );

        void affine_transform(
            tensor& dest,
            const tensor& src1,
            const tensor& src2,
            const float A,
            const float B,
            const float C
        );

        void affine_transform(
            tensor& dest,
            const tensor& src1,
            const tensor& src2,
            const float A,
            const float B
        );

        void affine_transform(
            tensor& dest,
            const tensor& src1,
            const tensor& src2,
            const tensor& src3,
            const float A,
            const float B,
            const float C,
            const float D
        );

        void affine_transform_range(
            size_t begin,
            size_t end,
            tensor& dest,
            const tensor& src1,
            const tensor& src2,
            const tensor& src3,
            const float A,
            const float B,
            const float C
        );

        void affine_transform(
            const rectangle& rect,
            tensor& dest, 
            const tensor& src1, 
            const tensor& src2, 
            const tensor& src3, 
            float A, 
            float B,
            float C
        );

        // Note that this function isn't in the tt:: namespace because add_scaled() is
        // called by cuda::add() so we don't need a tt:: version of add_scaled().  
        void add_scaled(
            tensor& dest,
            const float scale,
            const tensor& src
        );

        void add_cv_to_all_columns(
            float beta, 
            tensor& dest, 
            float alpha, 
            const tensor& src
        );

    // -----------------------------------------------------------------------------------

        void affine_transform(
            tensor& dest,
            const tensor& src,
            const tensor& A,
            const tensor& B
        );

    // -----------------------------------------------------------------------------------

        void affine_transform_conv(
            tensor& dest,
            const tensor& src,
            const tensor& A,
            const tensor& B
        );

    // ----------------------------------------------------------------------------------------

        void compute_adam_update (
            size_t begin,
            size_t end,
            tensor& s,
            tensor& m,
            tensor& v,
            const float t,
            const float learning_rate,
            const float weight_decay,
            const float momentum1,
            const float momentum2,
            const tensor& params,
            const tensor& params_grad
        );

    // -----------------------------------------------------------------------------------

        void assign_bias_gradient (
            tensor& grad,
            const tensor& gradient_input
        );

    // -----------------------------------------------------------------------------------

        void layer_normalize (
            const double eps,
            resizable_tensor& dest,
            resizable_tensor& means,
            resizable_tensor& invstds,
            const tensor& src,
            const tensor& gamma,
            const tensor& beta
        );

        void layer_normalize_gradient (
            const double eps,
            const tensor& gradient_input,
            const tensor& means,
            const tensor& invstds,
            const tensor& src,
            const tensor& gamma,
            tensor& src_grad,
            tensor& gamma_grad,
            tensor& beta_grad
        );

    // -----------------------------------------------------------------------------------

        void threshold (
            tensor& data,
            float thresh
        );

    // ----------------------------------------------------------------------------------------

        void dot (
            const tensor& a,
            const tensor& b,
            tensor& result,
            size_t idx
        );

    // ----------------------------------------------------------------------------------------

        void prelu (
            tensor& dest,
            const tensor& src,
            const tensor& param
        );

        void prelu_gradient (
            tensor& grad,
            const tensor& src,
            const tensor& gradient_input,
            const tensor& param,
            tensor& params_grad 
        );

    // ----------------------------------------------------------------------------------------

        void leaky_relu (
            tensor& dest,
            const tensor& src,
            const float alpha
        );

        void leaky_relu_gradient (
            tensor& grad,
            const tensor& src,
            const tensor& gradient_input,
            const float alpha
        );

    // ----------------------------------------------------------------------------------------

        void mish (
            tensor& dest,
            const tensor& src
        );

        void mish_gradient (
            tensor& grad,
            const tensor& src,
            const tensor& gradient_input
        );

    // ----------------------------------------------------------------------------------------

        void gelu (
            tensor& dest,
            const tensor& src
        );

        void gelu_gradient (
            tensor& grad,
            const tensor& src,
            const tensor& gradient_input
        );

    // ----------------------------------------------------------------------------------------

        void resize_bilinear (
            tensor& dest,
            long dest_row_stride,
            long dest_channel_stride,
            const tensor& src,
            long src_row_stride,
            long src_channel_stride
        );

        void resize_bilinear_gradient (
            tensor& grad,
            long grad_row_stride,
            long grad_channel_stride,
            const tensor& gradient_input,
            long gradient_input_row_stride,
            long gradient_input_channel_stride
        );

        inline void resize_bilinear (
            tensor& dest,
            const tensor& src
        ) { resize_bilinear(dest, dest.nc(), dest.nr()*dest.nc(), src, src.nc(), src.nr()*src.nc()); }

        inline void resize_bilinear_gradient (
            tensor& grad,
            const tensor& gradient_input
        ) { resize_bilinear_gradient(grad, grad.nc(), grad.nr()*grad.nc(), gradient_input, gradient_input.nc(), gradient_input.nr()*gradient_input.nc()); }

    // ----------------------------------------------------------------------------------------

        void copy_tensor(
            bool add_to,
            tensor& dest,
            size_t dest_k_offset,
            const tensor& src,
            size_t src_k_offset,
            size_t count_k
        );


    // ----------------------------------------------------------------------------------------

        class compute_loss_binary_log_per_pixel
        {
            /*!
                The point of this class is to compute the loss computed by
                loss_binary_log_per_pixel_, but to do so with CUDA.
            !*/
        public:

            compute_loss_binary_log_per_pixel(
            )
            {
            }

            template <
                typename const_label_iterator
                >
            void operator() (
                const_label_iterator truth,
                const tensor& subnetwork_output,
                tensor& gradient,
                double& loss
            ) const
            {
                const auto image_size = subnetwork_output.nr()*subnetwork_output.nc();
                const size_t bytes_per_plane = image_size*sizeof(float);
                // Allocate a cuda buffer to store all the truth images and also one float
                // for the scalar loss output.
                buf = device_global_buffer(subnetwork_output.num_samples()*bytes_per_plane + sizeof(float));

                cuda_data_ptr<float> loss_buf = static_pointer_cast<float>(buf, 1);
                buf = buf+sizeof(float);

                // copy the truth data into a cuda buffer.
                for (long i = 0; i < subnetwork_output.num_samples(); ++i, ++truth)
                {
                    const matrix<float>& t = *truth;
                    DLIB_ASSERT(t.nr() == subnetwork_output.nr());
                    DLIB_ASSERT(t.nc() == subnetwork_output.nc());
                    memcpy(buf + i*bytes_per_plane, &t(0,0), bytes_per_plane);
                }

                auto truth_buf = static_pointer_cast<const float>(buf, subnetwork_output.num_samples()*image_size);

                do_work(loss_buf, truth_buf, subnetwork_output, gradient, loss);
            }

        private:

            static void do_work(
                cuda_data_ptr<float> loss_work_buffer,
                cuda_data_ptr<const float> truth_buffer,
                const tensor& subnetwork_output,
                tensor& gradient,
                double& loss
            );

            mutable cuda_data_void_ptr buf;
        };

    // ----------------------------------------------------------------------------------------

        class compute_loss_multiclass_log_per_pixel
        {
            /*!
                The point of this class is to compute the loss computed by
                loss_multiclass_log_per_pixel_, but to do so with CUDA.
            !*/
        public:

            compute_loss_multiclass_log_per_pixel(
            )
            {
            }

            template <
                typename const_label_iterator
                >
            void operator() (
                const_label_iterator truth,
                const tensor& subnetwork_output,
                tensor& gradient,
                double& loss
            ) const
            {
                const auto image_size = subnetwork_output.nr()*subnetwork_output.nc();
                const size_t bytes_per_plane = image_size*sizeof(uint16_t);
                // Allocate a cuda buffer to store all the truth images and also one float
                // for the scalar loss output.
                buf = device_global_buffer(subnetwork_output.num_samples()*bytes_per_plane + sizeof(float));

                cuda_data_ptr<float> loss_buf = static_pointer_cast<float>(buf, 1);
                buf = buf+sizeof(float);

                // copy the truth data into a cuda buffer.
                for (long i = 0; i < subnetwork_output.num_samples(); ++i, ++truth)
                {
                    const matrix<uint16_t>& t = *truth;
                    DLIB_ASSERT(t.nr() == subnetwork_output.nr());
                    DLIB_ASSERT(t.nc() == subnetwork_output.nc());
                    memcpy(buf + i*bytes_per_plane, &t(0,0), bytes_per_plane);
                }

                auto truth_buf = static_pointer_cast<const uint16_t>(buf, subnetwork_output.num_samples()*image_size);

                do_work(loss_buf, truth_buf, subnetwork_output, gradient, loss);
            }

        private:

            static void do_work(
                cuda_data_ptr<float> loss_work_buffer,
                cuda_data_ptr<const uint16_t> truth_buffer,
                const tensor& subnetwork_output,
                tensor& gradient,
                double& loss
            );
            
            mutable cuda_data_void_ptr buf;
        };

    // ----------------------------------------------------------------------------------------

        class compute_loss_multiclass_log_per_pixel_weighted
        {
            /*!
                The point of this class is to compute the loss computed by
                loss_multiclass_log_per_pixel_weighted_, but to do so with CUDA.
            !*/
        public:

            compute_loss_multiclass_log_per_pixel_weighted(
            )
            {
            }

            template <
                typename const_label_iterator
                >
            void operator() (
                const_label_iterator truth,
                const tensor& subnetwork_output,
                tensor& gradient,
                double& loss
            ) const
            {
                const auto image_size = subnetwork_output.nr()*subnetwork_output.nc();
                const size_t bytes_per_plane = image_size*sizeof(uint16_t);
                const size_t weight_bytes_per_plane = image_size*sizeof(float);
                matrix<uint16_t> labels(truth->nr(), truth->nc());
                matrix<float> weights(truth->nr(), truth->nc());
                // Allocate a cuda buffer to store all the truth images and also one float
                // for the scalar loss output.
                buf = device_global_buffer(subnetwork_output.num_samples()*(bytes_per_plane + weight_bytes_per_plane) + sizeof(float));

                cuda_data_ptr<float> loss_buf = static_pointer_cast<float>(buf, 1);
                buf = buf+sizeof(float);
                const auto weights_offset = subnetwork_output.num_samples() * bytes_per_plane;
                // copy the truth data into a cuda buffer.
                for (long i = 0; i < subnetwork_output.num_samples(); ++i, ++truth)
                {
                    const matrix<weighted_label<uint16_t>>& t = *truth;
                    DLIB_ASSERT(t.nr() == subnetwork_output.nr());
                    DLIB_ASSERT(t.nc() == subnetwork_output.nc());
                    for (long r = 0; r < t.nr(); ++r)
                    {
                        for (long c = 0; c < t.nc(); ++c)
                        {
                            labels(r, c) = t(r, c).label;
                            weights(r, c) = t(r, c).weight;
                        }
                    }
                    memcpy(buf + i*bytes_per_plane, &labels(0,0), bytes_per_plane);
                    memcpy(buf + weights_offset + i*weight_bytes_per_plane, &weights(0, 0), weight_bytes_per_plane);
                }

                auto truth_buf = static_pointer_cast<const uint16_t>(buf, subnetwork_output.num_samples()*image_size);
                buf = buf+weights_offset;
                auto weights_buf = static_pointer_cast<const float>(buf, subnetwork_output.num_samples()*image_size);

                do_work(loss_buf, truth_buf, weights_buf, subnetwork_output, gradient, loss);
            }

        private:

            static void do_work(
                cuda_data_ptr<float> loss_work_buffer,
                cuda_data_ptr<const uint16_t> truth_buffer,
                cuda_data_ptr<const float> weights_buffer,
                const tensor& subnetwork_output,
                tensor& gradient,
                double& loss
            );

            mutable cuda_data_void_ptr buf;
        };

    // ----------------------------------------------------------------------------------------

        class compute_loss_mean_squared_per_channel_and_pixel
        {
            /*!
                The point of this class is to compute the loss computed by
                loss_mean_squared_per_channel_and_pixel_, but to do so with CUDA.
            !*/
        public:

            compute_loss_mean_squared_per_channel_and_pixel(
            )
            {
            }

            template <
                typename const_label_iterator
                >
            void operator() (
                const_label_iterator truth,
                const tensor& subnetwork_output,
                tensor& gradient,
                double& loss
            ) const
            {
                const auto image_size = subnetwork_output.nr()*subnetwork_output.nc()*subnetwork_output.k();
                const size_t bytes_per_image = image_size*sizeof(float);
                // Allocate a cuda buffer to store all the truth images and also one float
                // for the scalar loss output.
                buf = device_global_buffer(subnetwork_output.num_samples()*bytes_per_image + sizeof(float));

                cuda_data_ptr<float> loss_buf = static_pointer_cast<float>(buf, 1);
                buf = buf+sizeof(float);

                const size_t bytes_per_plane = subnetwork_output.nr()*subnetwork_output.nc()*sizeof(float);

                // copy the truth data into a cuda buffer.
                for (long i = 0; i < subnetwork_output.num_samples(); ++i, ++truth)
                {
                    const auto& t = *truth;
                    DLIB_ASSERT(static_cast<long>(t.size()) == subnetwork_output.k());
                    for (size_t j = 0; j < t.size(); ++j) {
                        DLIB_ASSERT(t[j].nr() == subnetwork_output.nr());
                        DLIB_ASSERT(t[j].nc() == subnetwork_output.nc());
                        memcpy(buf + i*bytes_per_image + j*bytes_per_plane, &t[j](0,0), bytes_per_plane);
                    }
                }

                auto truth_buf = static_pointer_cast<const float>(buf, subnetwork_output.num_samples()*image_size);

                do_work(loss_buf, truth_buf, subnetwork_output, gradient, loss);
            }

        private:

            static void do_work(
                cuda_data_ptr<float> loss_work_buffer,
                cuda_data_ptr<const float> truth_buffer,
                const tensor& subnetwork_output,
                tensor& gradient,
                double& loss
            );

            mutable cuda_data_void_ptr buf;
        };

    // ------------------------------------------------------------------------------------
    // ------------------------------------------------------------------------------------
    // ------------------------------------------------------------------------------------
    // ------------------------------------------------------------------------------------

#else // if DLIB_USE_CUDA NOT DEFINED

        inline void set_device (
            int id
        )
        {
            DLIB_CASSERT(id == 0, "dlib::cuda::set_device(id) called with an invalid device id.");
        }

        inline int get_device (
        ){ return 0; }

        inline int get_num_devices (
        ) { return 1; }

        inline std::string get_device_name (
            int device
        ) 
        {
            DLIB_CASSERT(device == 0, "dlib::cuda::set_device(id) called with an invalid device id.");
            return "CUDA_DISABLED";
        }

        inline void set_current_device_blocking_sync(
        ) {}


        inline bool can_access_peer (int , int )
        { return false; }
        inline bool can_access_peer (const tensor& , const tensor& )
        { return false; }

        inline void device_synchronize (int ){}
        inline void device_synchronize (const tensor& ){}

        class enable_peer_access
        {
        public:
            enable_peer_access() = delete;
            enable_peer_access(const enable_peer_access&) = delete;
            enable_peer_access& operator=(const enable_peer_access&) = delete;
            enable_peer_access( int, int ){}
            enable_peer_access( const tensor&, const tensor& ) {}
        };

#endif // DLIB_USE_CUDA

    } 
}


#endif // DLIB_DNN_CuDA_H_