-
Notifications
You must be signed in to change notification settings - Fork 166
/
gpu.h
591 lines (535 loc) · 24 KB
/
gpu.h
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
// Start of gpu.h
// Generic functions that use our tiny GPU abstraction layer. The
// entire context must be defined before this header is included. In
// particular we expect the following functions to be available:
static int gpu_free_actual(struct futhark_context *ctx, gpu_mem mem);
static int gpu_alloc_actual(struct futhark_context *ctx, size_t size, gpu_mem *mem_out);
int gpu_launch_kernel(struct futhark_context* ctx,
gpu_kernel kernel, const char *name,
const int32_t grid[3],
const int32_t block[3],
unsigned int shared_mem_bytes,
int num_args,
void* args[num_args],
size_t args_sizes[num_args]);
int gpu_memcpy(struct futhark_context* ctx,
gpu_mem dst, int64_t dst_offset,
gpu_mem src, int64_t src_offset,
int64_t nbytes);
int gpu_scalar_from_device(struct futhark_context* ctx,
void *dst,
gpu_mem src, size_t offset, size_t size);
int gpu_scalar_to_device(struct futhark_context* ctx,
gpu_mem dst, size_t offset, size_t size,
void *src);
void gpu_create_kernel(struct futhark_context *ctx,
gpu_kernel* kernel,
const char* name);
// Max number of thead blocks we allow along the second or third
// dimension for transpositions.
#define MAX_TR_THREAD_BLOCKS 65535
struct builtin_kernels {
// We have a lot of ways to transpose arrays.
gpu_kernel map_transpose_1b;
gpu_kernel map_transpose_1b_low_height;
gpu_kernel map_transpose_1b_low_width;
gpu_kernel map_transpose_1b_small;
gpu_kernel map_transpose_1b_large;
gpu_kernel map_transpose_2b;
gpu_kernel map_transpose_2b_low_height;
gpu_kernel map_transpose_2b_low_width;
gpu_kernel map_transpose_2b_small;
gpu_kernel map_transpose_2b_large;
gpu_kernel map_transpose_4b;
gpu_kernel map_transpose_4b_low_height;
gpu_kernel map_transpose_4b_low_width;
gpu_kernel map_transpose_4b_small;
gpu_kernel map_transpose_4b_large;
gpu_kernel map_transpose_8b;
gpu_kernel map_transpose_8b_low_height;
gpu_kernel map_transpose_8b_low_width;
gpu_kernel map_transpose_8b_small;
gpu_kernel map_transpose_8b_large;
// And a few ways of copying.
gpu_kernel lmad_copy_1b;
gpu_kernel lmad_copy_2b;
gpu_kernel lmad_copy_4b;
gpu_kernel lmad_copy_8b;
};
struct builtin_kernels* init_builtin_kernels(struct futhark_context* ctx) {
struct builtin_kernels *kernels = malloc(sizeof(struct builtin_kernels));
gpu_create_kernel(ctx, &kernels->map_transpose_1b, "map_transpose_1b");
gpu_create_kernel(ctx, &kernels->map_transpose_1b_large, "map_transpose_1b_large");
gpu_create_kernel(ctx, &kernels->map_transpose_1b_low_height, "map_transpose_1b_low_height");
gpu_create_kernel(ctx, &kernels->map_transpose_1b_low_width, "map_transpose_1b_low_width");
gpu_create_kernel(ctx, &kernels->map_transpose_1b_small, "map_transpose_1b_small");
gpu_create_kernel(ctx, &kernels->map_transpose_2b, "map_transpose_2b");
gpu_create_kernel(ctx, &kernels->map_transpose_2b_large, "map_transpose_2b_large");
gpu_create_kernel(ctx, &kernels->map_transpose_2b_low_height, "map_transpose_2b_low_height");
gpu_create_kernel(ctx, &kernels->map_transpose_2b_low_width, "map_transpose_2b_low_width");
gpu_create_kernel(ctx, &kernels->map_transpose_2b_small, "map_transpose_2b_small");
gpu_create_kernel(ctx, &kernels->map_transpose_4b, "map_transpose_4b");
gpu_create_kernel(ctx, &kernels->map_transpose_4b_large, "map_transpose_4b_large");
gpu_create_kernel(ctx, &kernels->map_transpose_4b_low_height, "map_transpose_4b_low_height");
gpu_create_kernel(ctx, &kernels->map_transpose_4b_low_width, "map_transpose_4b_low_width");
gpu_create_kernel(ctx, &kernels->map_transpose_4b_small, "map_transpose_4b_small");
gpu_create_kernel(ctx, &kernels->map_transpose_8b, "map_transpose_8b");
gpu_create_kernel(ctx, &kernels->map_transpose_8b_large, "map_transpose_8b_large");
gpu_create_kernel(ctx, &kernels->map_transpose_8b_low_height, "map_transpose_8b_low_height");
gpu_create_kernel(ctx, &kernels->map_transpose_8b_low_width, "map_transpose_8b_low_width");
gpu_create_kernel(ctx, &kernels->map_transpose_8b_small, "map_transpose_8b_small");
gpu_create_kernel(ctx, &kernels->lmad_copy_1b, "lmad_copy_1b");
gpu_create_kernel(ctx, &kernels->lmad_copy_2b, "lmad_copy_2b");
gpu_create_kernel(ctx, &kernels->lmad_copy_4b, "lmad_copy_4b");
gpu_create_kernel(ctx, &kernels->lmad_copy_8b, "lmad_copy_8b");
return kernels;
}
void free_builtin_kernels(struct futhark_context* ctx, struct builtin_kernels* kernels) {
gpu_free_kernel(ctx, kernels->map_transpose_1b);
gpu_free_kernel(ctx, kernels->map_transpose_1b_large);
gpu_free_kernel(ctx, kernels->map_transpose_1b_low_height);
gpu_free_kernel(ctx, kernels->map_transpose_1b_low_width);
gpu_free_kernel(ctx, kernels->map_transpose_1b_small);
gpu_free_kernel(ctx, kernels->map_transpose_2b);
gpu_free_kernel(ctx, kernels->map_transpose_2b_large);
gpu_free_kernel(ctx, kernels->map_transpose_2b_low_height);
gpu_free_kernel(ctx, kernels->map_transpose_2b_low_width);
gpu_free_kernel(ctx, kernels->map_transpose_2b_small);
gpu_free_kernel(ctx, kernels->map_transpose_4b);
gpu_free_kernel(ctx, kernels->map_transpose_4b_large);
gpu_free_kernel(ctx, kernels->map_transpose_4b_low_height);
gpu_free_kernel(ctx, kernels->map_transpose_4b_low_width);
gpu_free_kernel(ctx, kernels->map_transpose_4b_small);
gpu_free_kernel(ctx, kernels->map_transpose_8b);
gpu_free_kernel(ctx, kernels->map_transpose_8b_large);
gpu_free_kernel(ctx, kernels->map_transpose_8b_low_height);
gpu_free_kernel(ctx, kernels->map_transpose_8b_low_width);
gpu_free_kernel(ctx, kernels->map_transpose_8b_small);
gpu_free_kernel(ctx, kernels->lmad_copy_1b);
gpu_free_kernel(ctx, kernels->lmad_copy_2b);
gpu_free_kernel(ctx, kernels->lmad_copy_4b);
gpu_free_kernel(ctx, kernels->lmad_copy_8b);
free(kernels);
}
static int gpu_alloc(struct futhark_context *ctx, FILE *log,
size_t min_size, const char *tag,
gpu_mem *mem_out, size_t *size_out) {
if (min_size < sizeof(int)) {
min_size = sizeof(int);
}
gpu_mem* memptr;
if (free_list_find(&ctx->gpu_free_list, min_size, tag, size_out, (fl_mem*)&memptr) == 0) {
// Successfully found a free block. Is it big enough?
if (*size_out >= min_size) {
if (ctx->cfg->debugging) {
fprintf(log, "No need to allocate: Found a block in the free list.\n");
}
*mem_out = *memptr;
free(memptr);
return FUTHARK_SUCCESS;
} else {
if (ctx->cfg->debugging) {
fprintf(log, "Found a free block, but it was too small.\n");
}
int error = gpu_free_actual(ctx, *memptr);
free(memptr);
if (error != FUTHARK_SUCCESS) {
return error;
}
}
}
*size_out = min_size;
// We have to allocate a new block from the driver. If the
// allocation does not succeed, then we might be in an out-of-memory
// situation. We now start freeing things from the free list until
// we think we have freed enough that the allocation will succeed.
// Since we don't know how far the allocation is from fitting, we
// have to check after every deallocation. This might be pretty
// expensive. Let's hope that this case is hit rarely.
if (ctx->cfg->debugging) {
fprintf(log, "Actually allocating the desired block.\n");
}
int error = gpu_alloc_actual(ctx, min_size, mem_out);
while (error == FUTHARK_OUT_OF_MEMORY) {
if (ctx->cfg->debugging) {
fprintf(log, "Out of GPU memory: releasing entry from the free list...\n");
}
gpu_mem* memptr;
if (free_list_first(&ctx->gpu_free_list, (fl_mem*)&memptr) == 0) {
gpu_mem mem = *memptr;
free(memptr);
error = gpu_free_actual(ctx, mem);
if (error != FUTHARK_SUCCESS) {
return error;
}
} else {
break;
}
error = gpu_alloc_actual(ctx, min_size, mem_out);
}
return error;
}
static int gpu_free(struct futhark_context *ctx,
gpu_mem mem, size_t size, const char *tag) {
gpu_mem* memptr = malloc(sizeof(gpu_mem));
*memptr = mem;
free_list_insert(&ctx->gpu_free_list, size, (fl_mem)memptr, tag);
return FUTHARK_SUCCESS;
}
static int gpu_free_all(struct futhark_context *ctx) {
free_list_pack(&ctx->gpu_free_list);
gpu_mem* memptr;
while (free_list_first(&ctx->gpu_free_list, (fl_mem*)&memptr) == 0) {
gpu_mem mem = *memptr;
free(memptr);
int error = gpu_free_actual(ctx, mem);
if (error != FUTHARK_SUCCESS) {
return error;
}
}
return FUTHARK_SUCCESS;
}
static int gpu_map_transpose(struct futhark_context* ctx,
gpu_kernel kernel_default,
gpu_kernel kernel_low_height,
gpu_kernel kernel_low_width,
gpu_kernel kernel_small,
gpu_kernel kernel_large,
const char *name, size_t elem_size,
gpu_mem dst, int64_t dst_offset,
gpu_mem src, int64_t src_offset,
int64_t k, int64_t n, int64_t m) {
int64_t mulx = TR_BLOCK_DIM / n;
int64_t muly = TR_BLOCK_DIM / m;
int32_t mulx32 = mulx;
int32_t muly32 = muly;
int32_t k32 = k;
int32_t n32 = n;
int32_t m32 = m;
gpu_kernel kernel = kernel_default;
int32_t grid[3];
int32_t block[3];
void* args[11];
size_t args_sizes[11] = {
sizeof(gpu_mem), sizeof(int64_t),
sizeof(gpu_mem), sizeof(int64_t),
sizeof(int32_t),
sizeof(int32_t),
sizeof(int32_t),
sizeof(int32_t),
sizeof(int32_t)
};
args[0] = &dst;
args[1] = &dst_offset;
args[2] = &src;
args[3] = &src_offset;
args[7] = &mulx;
args[8] = &muly;
if (dst_offset + k * n * m <= 2147483647L &&
src_offset + k * n * m <= 2147483647L) {
if (m <= TR_BLOCK_DIM/2 && n <= TR_BLOCK_DIM/2) {
if (ctx->logging) { fprintf(ctx->log, "Using small kernel\n"); }
kernel = kernel_small;
grid[0] = ((k * n * m) + (TR_BLOCK_DIM*TR_BLOCK_DIM) - 1) / (TR_BLOCK_DIM*TR_BLOCK_DIM);
grid[1] = 1;
grid[2] = 1;
block[0] = TR_BLOCK_DIM*TR_BLOCK_DIM;
block[1] = 1;
block[2] = 1;
} else if (m <= TR_BLOCK_DIM/2 && TR_BLOCK_DIM < n) {
if (ctx->logging) { fprintf(ctx->log, "Using low-width kernel\n"); }
kernel = kernel_low_width;
int64_t x_elems = m;
int64_t y_elems = (n + muly - 1) / muly;
grid[0] = (x_elems + TR_BLOCK_DIM - 1) / TR_BLOCK_DIM;
grid[1] = (y_elems + TR_BLOCK_DIM - 1) / TR_BLOCK_DIM;
grid[2] = k;
block[0] = TR_BLOCK_DIM;
block[1] = TR_BLOCK_DIM;
block[2] = 1;
} else if (n <= TR_BLOCK_DIM/2 && TR_BLOCK_DIM < m) {
if (ctx->logging) { fprintf(ctx->log, "Using low-height kernel\n"); }
kernel = kernel_low_height;
int64_t x_elems = (m + mulx - 1) / mulx;
int64_t y_elems = n;
grid[0] = (x_elems + TR_BLOCK_DIM - 1) / TR_BLOCK_DIM;
grid[1] = (y_elems + TR_BLOCK_DIM - 1) / TR_BLOCK_DIM;
grid[2] = k;
block[0] = TR_BLOCK_DIM;
block[1] = TR_BLOCK_DIM;
block[2] = 1;
} else {
if (ctx->logging) { fprintf(ctx->log, "Using default kernel\n"); }
kernel = kernel_default;
grid[0] = (m+TR_TILE_DIM-1)/TR_TILE_DIM;
grid[1] = (n+TR_TILE_DIM-1)/TR_TILE_DIM;
grid[2] = k;
block[0] = TR_TILE_DIM;
block[1] = TR_TILE_DIM/TR_ELEMS_PER_THREAD;
block[2] = 1;
}
args[4] = &k32;
args[5] = &m32;
args[6] = &n32;
args[7] = &mulx32;
args[8] = &muly32;
} else {
if (ctx->logging) { fprintf(ctx->log, "Using large kernel\n"); }
kernel = kernel_large;
grid[0] = (m+TR_TILE_DIM-1)/TR_TILE_DIM;
grid[1] = (n+TR_TILE_DIM-1)/TR_TILE_DIM;
grid[2] = k;
block[0] = TR_TILE_DIM;
block[1] = TR_TILE_DIM/TR_ELEMS_PER_THREAD;
block[2] = 1;
args[4] = &k;
args[5] = &m;
args[6] = &n;
args[7] = &mulx;
args[8] = &muly;
args_sizes[4] = sizeof(int64_t);
args_sizes[5] = sizeof(int64_t);
args_sizes[6] = sizeof(int64_t);
args_sizes[7] = sizeof(int64_t);
args_sizes[8] = sizeof(int64_t);
}
// Cap the number of thead blocks we launch and figure out how many
// repeats we need alongside each dimension.
int32_t repeat_1 = grid[1] / MAX_TR_THREAD_BLOCKS;
int32_t repeat_2 = grid[2] / MAX_TR_THREAD_BLOCKS;
grid[1] = repeat_1 > 0 ? MAX_TR_THREAD_BLOCKS : grid[1];
grid[2] = repeat_2 > 0 ? MAX_TR_THREAD_BLOCKS : grid[2];
args[9] = &repeat_1;
args[10] = &repeat_2;
args_sizes[9] = sizeof(repeat_1);
args_sizes[10] = sizeof(repeat_2);
if (ctx->logging) {
fprintf(ctx->log, "\n");
}
return gpu_launch_kernel(ctx, kernel, name, grid, block,
TR_TILE_DIM*(TR_TILE_DIM+1)*elem_size,
sizeof(args)/sizeof(args[0]), args, args_sizes);
}
#define GEN_MAP_TRANSPOSE_GPU2GPU(NAME, ELEM_TYPE) \
static int map_transpose_gpu2gpu_##NAME \
(struct futhark_context* ctx, \
gpu_mem dst, int64_t dst_offset, \
gpu_mem src, int64_t src_offset, \
int64_t k, int64_t m, int64_t n) \
{ \
return \
gpu_map_transpose \
(ctx, \
ctx->kernels->map_transpose_##NAME, \
ctx->kernels->map_transpose_##NAME##_low_height, \
ctx->kernels->map_transpose_##NAME##_low_width, \
ctx->kernels->map_transpose_##NAME##_small, \
ctx->kernels->map_transpose_##NAME##_large, \
"map_transpose_" #NAME, sizeof(ELEM_TYPE), \
dst, dst_offset, src, src_offset, \
k, n, m); \
}
static int gpu_lmad_copy(struct futhark_context* ctx,
gpu_kernel kernel, int r,
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r],
gpu_mem src, int64_t src_offset, int64_t src_strides[r],
int64_t shape[r]) {
if (r > 8) {
set_error(ctx, strdup("Futhark runtime limitation:\nCannot copy array of greater than rank 8.\n"));
return 1;
}
int64_t n = 1;
for (int i = 0; i < r; i++) { n *= shape[i]; }
void* args[6+(8*3)];
size_t args_sizes[6+(8*3)];
args[0] = &dst;
args_sizes[0] = sizeof(gpu_mem);
args[1] = &dst_offset;
args_sizes[1] = sizeof(dst_offset);
args[2] = &src;
args_sizes[2] = sizeof(gpu_mem);
args[3] = &src_offset;
args_sizes[3] = sizeof(src_offset);
args[4] = &n;
args_sizes[4] = sizeof(n);
args[5] = &r;
args_sizes[5] = sizeof(r);
int64_t zero = 0;
for (int i = 0; i < 8; i++) {
args_sizes[6+i*3] = sizeof(int64_t);
args_sizes[6+i*3+1] = sizeof(int64_t);
args_sizes[6+i*3+2] = sizeof(int64_t);
if (i < r) {
args[6+i*3] = &shape[i];
args[6+i*3+1] = &dst_strides[i];
args[6+i*3+2] = &src_strides[i];
} else {
args[6+i*3] = &zero;
args[6+i*3+1] = &zero;
args[6+i*3+2] = &zero;
}
}
const size_t w = 256; // XXX: hardcoded thread block size.
return gpu_launch_kernel(ctx, kernel, "copy_lmad_dev_to_dev",
(const int32_t[3]) {(n+w-1)/w,1,1},
(const int32_t[3]) {w,1,1},
0, 6+(8*3), args, args_sizes);
}
#define GEN_LMAD_COPY_ELEMENTS_GPU2GPU(NAME, ELEM_TYPE) \
static int lmad_copy_elements_gpu2gpu_##NAME \
(struct futhark_context* ctx, \
int r, \
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r], \
gpu_mem src, int64_t src_offset, int64_t src_strides[r], \
int64_t shape[r]) { \
return gpu_lmad_copy(ctx, ctx->kernels->lmad_copy_##NAME, r, \
dst, dst_offset, dst_strides, \
src, src_offset, src_strides, \
shape); \
} \
#define GEN_LMAD_COPY_GPU2GPU(NAME, ELEM_TYPE) \
static int lmad_copy_gpu2gpu_##NAME \
(struct futhark_context* ctx, \
int r, \
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r], \
gpu_mem src, int64_t src_offset, int64_t src_strides[r], \
int64_t shape[r]) { \
log_copy(ctx, "GPU to GPU", r, dst_offset, dst_strides, \
src_offset, src_strides, shape); \
int64_t size = 1; \
for (int i = 0; i < r; i++) { size *= shape[i]; } \
if (size == 0) { return FUTHARK_SUCCESS; } \
int64_t k, n, m; \
if (lmad_map_tr(&k, &n, &m, \
r, dst_strides, src_strides, shape)) { \
log_transpose(ctx, k, n, m); \
return map_transpose_gpu2gpu_##NAME \
(ctx, dst, dst_offset, src, src_offset, k, n, m); \
} else if (lmad_memcpyable(r, dst_strides, src_strides, shape)) { \
if (ctx->logging) {fprintf(ctx->log, "## Flat copy\n\n");} \
return gpu_memcpy(ctx, \
dst, dst_offset*sizeof(ELEM_TYPE), \
src, src_offset*sizeof(ELEM_TYPE), \
size * sizeof(ELEM_TYPE)); \
} else { \
if (ctx->logging) {fprintf(ctx->log, "## General copy\n\n");} \
return lmad_copy_elements_gpu2gpu_##NAME \
(ctx, r, \
dst, dst_offset, dst_strides, \
src, src_offset, src_strides, \
shape); \
} \
}
static int
lmad_copy_elements_host2gpu(struct futhark_context *ctx, size_t elem_size,
int r,
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r],
unsigned char* src, int64_t src_offset, int64_t src_strides[r],
int64_t shape[r]) {
(void)ctx; (void)elem_size; (void)r;
(void)dst; (void)dst_offset; (void)dst_strides;
(void)src; (void)src_offset; (void)src_strides;
(void)shape;
set_error(ctx, strdup("Futhark runtime limitation:\nCannot copy unstructured array from host to GPU.\n"));
return 1;
}
static int
lmad_copy_elements_gpu2host (struct futhark_context *ctx, size_t elem_size,
int r,
unsigned char* dst, int64_t dst_offset, int64_t dst_strides[r],
gpu_mem src, int64_t src_offset, int64_t src_strides[r],
int64_t shape[r]) {
(void)ctx; (void)elem_size; (void)r;
(void)dst; (void)dst_offset; (void)dst_strides;
(void)src; (void)src_offset; (void)src_strides;
(void)shape;
set_error(ctx, strdup("Futhark runtime limitation:\nCannot copy unstructured array from GPU to host.\n"));
return 1;
}
#define GEN_LMAD_COPY_ELEMENTS_HOSTGPU(NAME, ELEM_TYPE) \
static int lmad_copy_elements_gpu2gpu_##NAME \
(struct futhark_context* ctx, \
int r, \
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r], \
gpu_mem src, int64_t src_offset, int64_t src_strides[r], \
int64_t shape[r]) { \
return (ctx, ctx->kernels->lmad_copy_##NAME, r, \
dst, dst_offset, dst_strides, \
src, src_offset, src_strides, \
shape); \
} \
static int lmad_copy_host2gpu(struct futhark_context* ctx, size_t elem_size, bool sync,
int r,
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r],
unsigned char* src, int64_t src_offset, int64_t src_strides[r],
int64_t shape[r]) {
log_copy(ctx, "Host to GPU", r, dst_offset, dst_strides,
src_offset, src_strides, shape);
int64_t size = elem_size;
for (int i = 0; i < r; i++) { size *= shape[i]; }
if (size == 0) { return FUTHARK_SUCCESS; }
int64_t k, n, m;
if (lmad_memcpyable(r, dst_strides, src_strides, shape)) {
if (ctx->logging) {fprintf(ctx->log, "## Flat copy\n\n");}
return memcpy_host2gpu(ctx, sync,
dst, dst_offset*elem_size,
src, src_offset*elem_size,
size);
} else {
if (ctx->logging) {fprintf(ctx->log, "## General copy\n\n");}
int error;
error = lmad_copy_elements_host2gpu
(ctx, elem_size, r,
dst, dst_offset, dst_strides,
src, src_offset, src_strides,
shape);
if (error == 0 && sync) {
error = futhark_context_sync(ctx);
}
return error;
}
}
static int lmad_copy_gpu2host(struct futhark_context* ctx, size_t elem_size, bool sync,
int r,
unsigned char* dst, int64_t dst_offset, int64_t dst_strides[r],
gpu_mem src, int64_t src_offset, int64_t src_strides[r],
int64_t shape[r]) {
log_copy(ctx, "Host to GPU", r, dst_offset, dst_strides,
src_offset, src_strides, shape);
int64_t size = elem_size;
for (int i = 0; i < r; i++) { size *= shape[i]; }
if (size == 0) { return FUTHARK_SUCCESS; }
int64_t k, n, m;
if (lmad_memcpyable(r, dst_strides, src_strides, shape)) {
if (ctx->logging) {fprintf(ctx->log, "## Flat copy\n\n");}
return memcpy_gpu2host(ctx, sync,
dst, dst_offset*elem_size,
src, src_offset*elem_size,
size);
} else {
if (ctx->logging) {fprintf(ctx->log, "## General copy\n\n");}
int error;
error = lmad_copy_elements_gpu2host
(ctx, elem_size, r,
dst, dst_offset, dst_strides,
src, src_offset, src_strides,
shape);
if (error == 0 && sync) {
error = futhark_context_sync(ctx);
}
return error;
}
}
GEN_MAP_TRANSPOSE_GPU2GPU(1b, uint8_t)
GEN_MAP_TRANSPOSE_GPU2GPU(2b, uint16_t)
GEN_MAP_TRANSPOSE_GPU2GPU(4b, uint32_t)
GEN_MAP_TRANSPOSE_GPU2GPU(8b, uint64_t)
GEN_LMAD_COPY_ELEMENTS_GPU2GPU(1b, uint8_t)
GEN_LMAD_COPY_ELEMENTS_GPU2GPU(2b, uint16_t)
GEN_LMAD_COPY_ELEMENTS_GPU2GPU(4b, uint32_t)
GEN_LMAD_COPY_ELEMENTS_GPU2GPU(8b, uint64_t)
GEN_LMAD_COPY_GPU2GPU(1b, uint8_t)
GEN_LMAD_COPY_GPU2GPU(2b, uint16_t)
GEN_LMAD_COPY_GPU2GPU(4b, uint32_t)
GEN_LMAD_COPY_GPU2GPU(8b, uint64_t)
// End of gpu.h