| 1 |
// Start of gpu.h |
| 2 |
|
| 3 |
// Generic functions that use our tiny GPU abstraction layer. The |
| 4 |
// entire context must be defined before this header is included. In |
| 5 |
// particular we expect the following functions to be available: |
| 6 |
|
| 7 |
static int gpu_free_actual(struct futhark_context *ctx, gpu_mem mem); |
| 8 |
static int gpu_alloc_actual(struct futhark_context *ctx, size_t size, gpu_mem *mem_out); |
| 9 |
int gpu_launch_kernel(struct futhark_context* ctx, |
| 10 |
gpu_kernel kernel, |
| 11 |
const char *name, const char *provenance, |
| 12 |
const int32_t grid[3], |
| 13 |
const int32_t block[3], |
| 14 |
unsigned int shared_mem_bytes, |
| 15 |
int num_args, |
| 16 |
void* args[num_args], |
| 17 |
size_t args_sizes[num_args]); |
| 18 |
int gpu_memcpy(struct futhark_context* ctx, const char* provenance, |
| 19 |
gpu_mem dst, int64_t dst_offset, |
| 20 |
gpu_mem src, int64_t src_offset, |
| 21 |
int64_t nbytes); |
| 22 |
int gpu_scalar_from_device(struct futhark_context* ctx, const char* provenance, |
| 23 |
void *dst, |
| 24 |
gpu_mem src, size_t offset, size_t size); |
| 25 |
int gpu_scalar_to_device(struct futhark_context* ctx, const char* provenance, |
| 26 |
gpu_mem dst, size_t offset, size_t size, |
| 27 |
void *src); |
| 28 |
void gpu_create_kernel(struct futhark_context *ctx, |
| 29 |
gpu_kernel* kernel, |
| 30 |
const char* name); |
| 31 |
|
| 32 |
static void gpu_init_log(struct futhark_context *ctx) { |
| 33 |
if (ctx->cfg->logging) { |
| 34 |
fprintf(ctx->log, "Default block size: %ld\n", (long)ctx->cfg->gpu.default_block_size); |
| 35 |
fprintf(ctx->log, "Default grid size: %ld\n", (long)ctx->cfg->gpu.default_grid_size); |
| 36 |
fprintf(ctx->log, "Default tile size: %ld\n", (long)ctx->cfg->gpu.default_tile_size); |
| 37 |
fprintf(ctx->log, "Default register tile size: %ld\n", (long)ctx->cfg->gpu.default_reg_tile_size); |
| 38 |
fprintf(ctx->log, "Default cache: %ld\n", (long)ctx->cfg->gpu.default_cache); |
| 39 |
fprintf(ctx->log, "Default registers: %ld\n", (long)ctx->cfg->gpu.default_registers); |
| 40 |
fprintf(ctx->log, "Default threshold: %ld\n", (long)ctx->cfg->gpu.default_threshold); |
| 41 |
fprintf(ctx->log, "Max thread block size: %ld\n", (long)ctx->max_thread_block_size); |
| 42 |
fprintf(ctx->log, "Max grid size: %ld\n", (long)ctx->max_grid_size); |
| 43 |
fprintf(ctx->log, "Max tile size: %ld\n", (long)ctx->max_tile_size); |
| 44 |
fprintf(ctx->log, "Max threshold: %ld\n", (long)ctx->max_threshold); |
| 45 |
fprintf(ctx->log, "Max shared memory: %ld\n", (long)ctx->max_shared_memory); |
| 46 |
fprintf(ctx->log, "Max registers: %ld\n", (long)ctx->max_registers); |
| 47 |
fprintf(ctx->log, "Max cache: %ld\n", (long)ctx->max_cache); |
| 48 |
fprintf(ctx->log, "Lockstep width: %ld\n", (long)ctx->lockstep_width); |
| 49 |
} |
| 50 |
} |
| 51 |
|
| 52 |
// Generic GPU command line options. |
| 53 |
|
| 54 |
void futhark_context_config_set_default_thread_block_size(struct futhark_context_config *cfg, int size) { |
| 55 |
cfg->gpu.default_block_size = size; |
| 56 |
cfg->gpu.default_block_size_changed = 1; |
| 57 |
} |
| 58 |
|
| 59 |
void futhark_context_config_set_default_group_size(struct futhark_context_config *cfg, int size) { |
| 60 |
futhark_context_config_set_default_thread_block_size(cfg, size); |
| 61 |
} |
| 62 |
|
| 63 |
void futhark_context_config_set_default_grid_size(struct futhark_context_config *cfg, int num) { |
| 64 |
cfg->gpu.default_grid_size = num; |
| 65 |
cfg->gpu.default_grid_size_changed = 1; |
| 66 |
} |
| 67 |
|
| 68 |
void futhark_context_config_set_default_num_groups(struct futhark_context_config *cfg, int num) { |
| 69 |
futhark_context_config_set_default_grid_size(cfg, num); |
| 70 |
} |
| 71 |
|
| 72 |
void futhark_context_config_set_default_tile_size(struct futhark_context_config *cfg, int size) { |
| 73 |
cfg->gpu.default_tile_size = size; |
| 74 |
cfg->gpu.default_tile_size_changed = 1; |
| 75 |
} |
| 76 |
|
| 77 |
void futhark_context_config_set_default_reg_tile_size(struct futhark_context_config *cfg, int size) { |
| 78 |
cfg->gpu.default_reg_tile_size = size; |
| 79 |
} |
| 80 |
|
| 81 |
void futhark_context_config_set_default_cache(struct futhark_context_config *cfg, int size) { |
| 82 |
cfg->gpu.default_cache = size; |
| 83 |
} |
| 84 |
|
| 85 |
void futhark_context_config_set_default_registers(struct futhark_context_config *cfg, int size) { |
| 86 |
cfg->gpu.default_registers = size; |
| 87 |
} |
| 88 |
|
| 89 |
void futhark_context_config_set_default_threshold(struct futhark_context_config *cfg, int size) { |
| 90 |
cfg->gpu.default_threshold = size; |
| 91 |
} |
| 92 |
|
| 93 |
int futhark_context_config_set_tuning_param(struct futhark_context_config *cfg, |
| 94 |
const char *param_name, |
| 95 |
size_t new_value) { |
| 96 |
for (int i = 0; i < cfg->num_tuning_params; i++) { |
| 97 |
if (strcmp(param_name, cfg->tuning_param_names[i]) == 0) { |
| 98 |
cfg->tuning_params[i] = new_value; |
| 99 |
return 0; |
| 100 |
} |
| 101 |
} |
| 102 |
if (strcmp(param_name, "default_thread_block_size") == 0 || |
| 103 |
strcmp(param_name, "default_group_size") == 0) { |
| 104 |
cfg->gpu.default_block_size = new_value; |
| 105 |
return 0; |
| 106 |
} |
| 107 |
if (strcmp(param_name, "default_grid_size") == 0 || |
| 108 |
strcmp(param_name, "default_num_groups") == 0) { |
| 109 |
cfg->gpu.default_grid_size = new_value; |
| 110 |
return 0; |
| 111 |
} |
| 112 |
if (strcmp(param_name, "default_threshold") == 0) { |
| 113 |
cfg->gpu.default_threshold = new_value; |
| 114 |
return 0; |
| 115 |
} |
| 116 |
if (strcmp(param_name, "default_tile_size") == 0) { |
| 117 |
cfg->gpu.default_tile_size = new_value; |
| 118 |
return 0; |
| 119 |
} |
| 120 |
if (strcmp(param_name, "default_reg_tile_size") == 0) { |
| 121 |
cfg->gpu.default_reg_tile_size = new_value; |
| 122 |
return 0; |
| 123 |
} |
| 124 |
if (strcmp(param_name, "default_cache") == 0) { |
| 125 |
cfg->gpu.default_cache = new_value; |
| 126 |
return 0; |
| 127 |
} |
| 128 |
if (strcmp(param_name, "default_shared_memory") == 0) { |
| 129 |
cfg->gpu.default_shared_memory = new_value; |
| 130 |
return 0; |
| 131 |
} |
| 132 |
|
| 133 |
return 1; |
| 134 |
} |
| 135 |
|
| 136 |
// End of GPU command line optiopns. |
| 137 |
|
| 138 |
// Max number of thead blocks we allow along the second or third |
| 139 |
// dimension for transpositions. |
| 140 |
#define MAX_TR_THREAD_BLOCKS 65535 |
| 141 |
|
| 142 |
struct builtin_kernels { |
| 143 |
// We have a lot of ways to transpose arrays. |
| 144 |
gpu_kernel map_transpose_1b; |
| 145 |
gpu_kernel map_transpose_1b_low_height; |
| 146 |
gpu_kernel map_transpose_1b_low_width; |
| 147 |
gpu_kernel map_transpose_1b_small; |
| 148 |
gpu_kernel map_transpose_1b_large; |
| 149 |
gpu_kernel map_transpose_2b; |
| 150 |
gpu_kernel map_transpose_2b_low_height; |
| 151 |
gpu_kernel map_transpose_2b_low_width; |
| 152 |
gpu_kernel map_transpose_2b_small; |
| 153 |
gpu_kernel map_transpose_2b_large; |
| 154 |
gpu_kernel map_transpose_4b; |
| 155 |
gpu_kernel map_transpose_4b_low_height; |
| 156 |
gpu_kernel map_transpose_4b_low_width; |
| 157 |
gpu_kernel map_transpose_4b_small; |
| 158 |
gpu_kernel map_transpose_4b_large; |
| 159 |
gpu_kernel map_transpose_8b; |
| 160 |
gpu_kernel map_transpose_8b_low_height; |
| 161 |
gpu_kernel map_transpose_8b_low_width; |
| 162 |
gpu_kernel map_transpose_8b_small; |
| 163 |
gpu_kernel map_transpose_8b_large; |
| 164 |
|
| 165 |
// And a few ways of copying. |
| 166 |
gpu_kernel lmad_copy_1b; |
| 167 |
gpu_kernel lmad_copy_2b; |
| 168 |
gpu_kernel lmad_copy_4b; |
| 169 |
gpu_kernel lmad_copy_8b; |
| 170 |
}; |
| 171 |
|
| 172 |
struct builtin_kernels* init_builtin_kernels(struct futhark_context* ctx) { |
| 173 |
struct builtin_kernels *kernels = malloc(sizeof(struct builtin_kernels)); |
| 174 |
gpu_create_kernel(ctx, &kernels->map_transpose_1b, "map_transpose_1b"); |
| 175 |
gpu_create_kernel(ctx, &kernels->map_transpose_1b_large, "map_transpose_1b_large"); |
| 176 |
gpu_create_kernel(ctx, &kernels->map_transpose_1b_low_height, "map_transpose_1b_low_height"); |
| 177 |
gpu_create_kernel(ctx, &kernels->map_transpose_1b_low_width, "map_transpose_1b_low_width"); |
| 178 |
gpu_create_kernel(ctx, &kernels->map_transpose_1b_small, "map_transpose_1b_small"); |
| 179 |
|
| 180 |
gpu_create_kernel(ctx, &kernels->map_transpose_2b, "map_transpose_2b"); |
| 181 |
gpu_create_kernel(ctx, &kernels->map_transpose_2b_large, "map_transpose_2b_large"); |
| 182 |
gpu_create_kernel(ctx, &kernels->map_transpose_2b_low_height, "map_transpose_2b_low_height"); |
| 183 |
gpu_create_kernel(ctx, &kernels->map_transpose_2b_low_width, "map_transpose_2b_low_width"); |
| 184 |
gpu_create_kernel(ctx, &kernels->map_transpose_2b_small, "map_transpose_2b_small"); |
| 185 |
|
| 186 |
gpu_create_kernel(ctx, &kernels->map_transpose_4b, "map_transpose_4b"); |
| 187 |
gpu_create_kernel(ctx, &kernels->map_transpose_4b_large, "map_transpose_4b_large"); |
| 188 |
gpu_create_kernel(ctx, &kernels->map_transpose_4b_low_height, "map_transpose_4b_low_height"); |
| 189 |
gpu_create_kernel(ctx, &kernels->map_transpose_4b_low_width, "map_transpose_4b_low_width"); |
| 190 |
gpu_create_kernel(ctx, &kernels->map_transpose_4b_small, "map_transpose_4b_small"); |
| 191 |
|
| 192 |
gpu_create_kernel(ctx, &kernels->map_transpose_8b, "map_transpose_8b"); |
| 193 |
gpu_create_kernel(ctx, &kernels->map_transpose_8b_large, "map_transpose_8b_large"); |
| 194 |
gpu_create_kernel(ctx, &kernels->map_transpose_8b_low_height, "map_transpose_8b_low_height"); |
| 195 |
gpu_create_kernel(ctx, &kernels->map_transpose_8b_low_width, "map_transpose_8b_low_width"); |
| 196 |
gpu_create_kernel(ctx, &kernels->map_transpose_8b_small, "map_transpose_8b_small"); |
| 197 |
|
| 198 |
gpu_create_kernel(ctx, &kernels->lmad_copy_1b, "lmad_copy_1b"); |
| 199 |
gpu_create_kernel(ctx, &kernels->lmad_copy_2b, "lmad_copy_2b"); |
| 200 |
gpu_create_kernel(ctx, &kernels->lmad_copy_4b, "lmad_copy_4b"); |
| 201 |
gpu_create_kernel(ctx, &kernels->lmad_copy_8b, "lmad_copy_8b"); |
| 202 |
|
| 203 |
return kernels; |
| 204 |
} |
| 205 |
|
| 206 |
void free_builtin_kernels(struct futhark_context* ctx, struct builtin_kernels* kernels) { |
| 207 |
gpu_free_kernel(ctx, kernels->map_transpose_1b); |
| 208 |
gpu_free_kernel(ctx, kernels->map_transpose_1b_large); |
| 209 |
gpu_free_kernel(ctx, kernels->map_transpose_1b_low_height); |
| 210 |
gpu_free_kernel(ctx, kernels->map_transpose_1b_low_width); |
| 211 |
gpu_free_kernel(ctx, kernels->map_transpose_1b_small); |
| 212 |
|
| 213 |
gpu_free_kernel(ctx, kernels->map_transpose_2b); |
| 214 |
gpu_free_kernel(ctx, kernels->map_transpose_2b_large); |
| 215 |
gpu_free_kernel(ctx, kernels->map_transpose_2b_low_height); |
| 216 |
gpu_free_kernel(ctx, kernels->map_transpose_2b_low_width); |
| 217 |
gpu_free_kernel(ctx, kernels->map_transpose_2b_small); |
| 218 |
|
| 219 |
gpu_free_kernel(ctx, kernels->map_transpose_4b); |
| 220 |
gpu_free_kernel(ctx, kernels->map_transpose_4b_large); |
| 221 |
gpu_free_kernel(ctx, kernels->map_transpose_4b_low_height); |
| 222 |
gpu_free_kernel(ctx, kernels->map_transpose_4b_low_width); |
| 223 |
gpu_free_kernel(ctx, kernels->map_transpose_4b_small); |
| 224 |
|
| 225 |
gpu_free_kernel(ctx, kernels->map_transpose_8b); |
| 226 |
gpu_free_kernel(ctx, kernels->map_transpose_8b_large); |
| 227 |
gpu_free_kernel(ctx, kernels->map_transpose_8b_low_height); |
| 228 |
gpu_free_kernel(ctx, kernels->map_transpose_8b_low_width); |
| 229 |
gpu_free_kernel(ctx, kernels->map_transpose_8b_small); |
| 230 |
|
| 231 |
gpu_free_kernel(ctx, kernels->lmad_copy_1b); |
| 232 |
gpu_free_kernel(ctx, kernels->lmad_copy_2b); |
| 233 |
gpu_free_kernel(ctx, kernels->lmad_copy_4b); |
| 234 |
gpu_free_kernel(ctx, kernels->lmad_copy_8b); |
| 235 |
|
| 236 |
free(kernels); |
| 237 |
} |
| 238 |
|
| 239 |
static int gpu_alloc(struct futhark_context *ctx, FILE *log, |
| 240 |
size_t min_size, const char *tag, |
| 241 |
gpu_mem *mem_out, size_t *size_out) { |
| 242 |
if (min_size < sizeof(int)) { |
| 243 |
min_size = sizeof(int); |
| 244 |
} |
| 245 |
|
| 246 |
// Round up the allocation to be at least divisible by 4, because that is |
| 247 |
// assumed by the code generator. |
| 248 |
min_size = (min_size+3) & ~3; |
| 249 |
|
| 250 |
gpu_mem* memptr; |
| 251 |
if (free_list_find(&ctx->gpu_free_list, min_size, tag, size_out, (fl_mem*)&memptr) == 0) { |
| 252 |
// Successfully found a free block. Is it big enough? |
| 253 |
if (*size_out >= min_size) { |
| 254 |
if (ctx->cfg->debugging) { |
| 255 |
fprintf(log, "No need to allocate: Found a block in the free list.\n"); |
| 256 |
} |
| 257 |
*mem_out = *memptr; |
| 258 |
free(memptr); |
| 259 |
return FUTHARK_SUCCESS; |
| 260 |
} else { |
| 261 |
if (ctx->cfg->debugging) { |
| 262 |
fprintf(log, "Found a free block, but it was too small.\n"); |
| 263 |
} |
| 264 |
int error = gpu_free_actual(ctx, *memptr); |
| 265 |
free(memptr); |
| 266 |
if (error != FUTHARK_SUCCESS) { |
| 267 |
return error; |
| 268 |
} |
| 269 |
} |
| 270 |
} |
| 271 |
|
| 272 |
*size_out = min_size; |
| 273 |
|
| 274 |
// We have to allocate a new block from the driver. If the |
| 275 |
// allocation does not succeed, then we might be in an out-of-memory |
| 276 |
// situation. We now start freeing things from the free list until |
| 277 |
// we think we have freed enough that the allocation will succeed. |
| 278 |
// Since we don't know how far the allocation is from fitting, we |
| 279 |
// have to check after every deallocation. This might be pretty |
| 280 |
// expensive. Let's hope that this case is hit rarely. |
| 281 |
|
| 282 |
if (ctx->cfg->debugging) { |
| 283 |
fprintf(log, "Actually allocating the desired block.\n"); |
| 284 |
} |
| 285 |
|
| 286 |
int error = gpu_alloc_actual(ctx, min_size, mem_out); |
| 287 |
|
| 288 |
while (error == FUTHARK_OUT_OF_MEMORY) { |
| 289 |
if (ctx->cfg->debugging) { |
| 290 |
fprintf(log, "Out of GPU memory: releasing entry from the free list...\n"); |
| 291 |
} |
| 292 |
gpu_mem* memptr; |
| 293 |
if (free_list_first(&ctx->gpu_free_list, (fl_mem*)&memptr) == 0) { |
| 294 |
gpu_mem mem = *memptr; |
| 295 |
free(memptr); |
| 296 |
error = gpu_free_actual(ctx, mem); |
| 297 |
if (error != FUTHARK_SUCCESS) { |
| 298 |
return error; |
| 299 |
} |
| 300 |
} else { |
| 301 |
break; |
| 302 |
} |
| 303 |
error = gpu_alloc_actual(ctx, min_size, mem_out); |
| 304 |
} |
| 305 |
|
| 306 |
return error; |
| 307 |
} |
| 308 |
|
| 309 |
static int gpu_free(struct futhark_context *ctx, |
| 310 |
gpu_mem mem, size_t size, const char *tag) { |
| 311 |
gpu_mem* memptr = malloc(sizeof(gpu_mem)); |
| 312 |
*memptr = mem; |
| 313 |
free_list_insert(&ctx->gpu_free_list, size, (fl_mem)memptr, tag); |
| 314 |
return FUTHARK_SUCCESS; |
| 315 |
} |
| 316 |
|
| 317 |
static int gpu_free_all(struct futhark_context *ctx) { |
| 318 |
free_list_pack(&ctx->gpu_free_list); |
| 319 |
gpu_mem* memptr; |
| 320 |
while (free_list_first(&ctx->gpu_free_list, (fl_mem*)&memptr) == 0) { |
| 321 |
gpu_mem mem = *memptr; |
| 322 |
free(memptr); |
| 323 |
int error = gpu_free_actual(ctx, mem); |
| 324 |
if (error != FUTHARK_SUCCESS) { |
| 325 |
return error; |
| 326 |
} |
| 327 |
} |
| 328 |
|
| 329 |
return FUTHARK_SUCCESS; |
| 330 |
} |
| 331 |
|
| 332 |
static int gpu_map_transpose(struct futhark_context* ctx, |
| 333 |
const char* provenance, |
| 334 |
gpu_kernel kernel_default, |
| 335 |
gpu_kernel kernel_low_height, |
| 336 |
gpu_kernel kernel_low_width, |
| 337 |
gpu_kernel kernel_small, |
| 338 |
gpu_kernel kernel_large, |
| 339 |
const char *name, size_t elem_size, |
| 340 |
gpu_mem dst, int64_t dst_offset, |
| 341 |
gpu_mem src, int64_t src_offset, |
| 342 |
int64_t k, int64_t n, int64_t m) { |
| 343 |
int64_t mulx = TR_BLOCK_DIM / n; |
| 344 |
int64_t muly = TR_BLOCK_DIM / m; |
| 345 |
int32_t mulx32 = mulx; |
| 346 |
int32_t muly32 = muly; |
| 347 |
int32_t k32 = k; |
| 348 |
int32_t n32 = n; |
| 349 |
int32_t m32 = m; |
| 350 |
|
| 351 |
gpu_kernel kernel = kernel_default; |
| 352 |
int32_t grid[3]; |
| 353 |
int32_t block[3]; |
| 354 |
|
| 355 |
void* args[11]; |
| 356 |
size_t args_sizes[11] = { |
| 357 |
sizeof(gpu_mem), sizeof(int64_t), |
| 358 |
sizeof(gpu_mem), sizeof(int64_t), |
| 359 |
sizeof(int32_t), |
| 360 |
sizeof(int32_t), |
| 361 |
sizeof(int32_t), |
| 362 |
sizeof(int32_t), |
| 363 |
sizeof(int32_t) |
| 364 |
}; |
| 365 |
|
| 366 |
args[0] = &dst; |
| 367 |
args[1] = &dst_offset; |
| 368 |
args[2] = &src; |
| 369 |
args[3] = &src_offset; |
| 370 |
args[7] = &mulx; |
| 371 |
args[8] = &muly; |
| 372 |
|
| 373 |
if (dst_offset + k * n * m <= 2147483647L && |
| 374 |
src_offset + k * n * m <= 2147483647L) { |
| 375 |
if (m <= TR_BLOCK_DIM/2 && n <= TR_BLOCK_DIM/2) { |
| 376 |
if (ctx->logging) { fprintf(ctx->log, "Using small kernel\n"); } |
| 377 |
kernel = kernel_small; |
| 378 |
grid[0] = ((k * n * m) + (TR_BLOCK_DIM*TR_BLOCK_DIM) - 1) / (TR_BLOCK_DIM*TR_BLOCK_DIM); |
| 379 |
grid[1] = 1; |
| 380 |
grid[2] = 1; |
| 381 |
block[0] = TR_BLOCK_DIM*TR_BLOCK_DIM; |
| 382 |
block[1] = 1; |
| 383 |
block[2] = 1; |
| 384 |
} else if (m <= TR_BLOCK_DIM/2 && TR_BLOCK_DIM < n) { |
| 385 |
if (ctx->logging) { fprintf(ctx->log, "Using low-width kernel\n"); } |
| 386 |
kernel = kernel_low_width; |
| 387 |
int64_t x_elems = m; |
| 388 |
int64_t y_elems = (n + muly - 1) / muly; |
| 389 |
grid[0] = (x_elems + TR_BLOCK_DIM - 1) / TR_BLOCK_DIM; |
| 390 |
grid[1] = (y_elems + TR_BLOCK_DIM - 1) / TR_BLOCK_DIM; |
| 391 |
grid[2] = k; |
| 392 |
block[0] = TR_BLOCK_DIM; |
| 393 |
block[1] = TR_BLOCK_DIM; |
| 394 |
block[2] = 1; |
| 395 |
} else if (n <= TR_BLOCK_DIM/2 && TR_BLOCK_DIM < m) { |
| 396 |
if (ctx->logging) { fprintf(ctx->log, "Using low-height kernel\n"); } |
| 397 |
kernel = kernel_low_height; |
| 398 |
int64_t x_elems = (m + mulx - 1) / mulx; |
| 399 |
int64_t y_elems = n; |
| 400 |
grid[0] = (x_elems + TR_BLOCK_DIM - 1) / TR_BLOCK_DIM; |
| 401 |
grid[1] = (y_elems + TR_BLOCK_DIM - 1) / TR_BLOCK_DIM; |
| 402 |
grid[2] = k; |
| 403 |
block[0] = TR_BLOCK_DIM; |
| 404 |
block[1] = TR_BLOCK_DIM; |
| 405 |
block[2] = 1; |
| 406 |
} else { |
| 407 |
if (ctx->logging) { fprintf(ctx->log, "Using default kernel\n"); } |
| 408 |
kernel = kernel_default; |
| 409 |
grid[0] = (m+TR_TILE_DIM-1)/TR_TILE_DIM; |
| 410 |
grid[1] = (n+TR_TILE_DIM-1)/TR_TILE_DIM; |
| 411 |
grid[2] = k; |
| 412 |
block[0] = TR_TILE_DIM; |
| 413 |
block[1] = TR_TILE_DIM/TR_ELEMS_PER_THREAD; |
| 414 |
block[2] = 1; |
| 415 |
} |
| 416 |
args[4] = &k32; |
| 417 |
args[5] = &m32; |
| 418 |
args[6] = &n32; |
| 419 |
args[7] = &mulx32; |
| 420 |
args[8] = &muly32; |
| 421 |
} else { |
| 422 |
if (ctx->logging) { fprintf(ctx->log, "Using large kernel\n"); } |
| 423 |
kernel = kernel_large; |
| 424 |
grid[0] = (m+TR_TILE_DIM-1)/TR_TILE_DIM; |
| 425 |
grid[1] = (n+TR_TILE_DIM-1)/TR_TILE_DIM; |
| 426 |
grid[2] = k; |
| 427 |
block[0] = TR_TILE_DIM; |
| 428 |
block[1] = TR_TILE_DIM/TR_ELEMS_PER_THREAD; |
| 429 |
block[2] = 1; |
| 430 |
args[4] = &k; |
| 431 |
args[5] = &m; |
| 432 |
args[6] = &n; |
| 433 |
args[7] = &mulx; |
| 434 |
args[8] = &muly; |
| 435 |
args_sizes[4] = sizeof(int64_t); |
| 436 |
args_sizes[5] = sizeof(int64_t); |
| 437 |
args_sizes[6] = sizeof(int64_t); |
| 438 |
args_sizes[7] = sizeof(int64_t); |
| 439 |
args_sizes[8] = sizeof(int64_t); |
| 440 |
} |
| 441 |
|
| 442 |
// Cap the number of thead blocks we launch and figure out how many |
| 443 |
// repeats we need alongside each dimension. |
| 444 |
int32_t repeat_1 = grid[1] / MAX_TR_THREAD_BLOCKS; |
| 445 |
int32_t repeat_2 = grid[2] / MAX_TR_THREAD_BLOCKS; |
| 446 |
grid[1] = repeat_1 > 0 ? MAX_TR_THREAD_BLOCKS : grid[1]; |
| 447 |
grid[2] = repeat_2 > 0 ? MAX_TR_THREAD_BLOCKS : grid[2]; |
| 448 |
args[9] = &repeat_1; |
| 449 |
args[10] = &repeat_2; |
| 450 |
args_sizes[9] = sizeof(repeat_1); |
| 451 |
args_sizes[10] = sizeof(repeat_2); |
| 452 |
|
| 453 |
if (ctx->logging) { |
| 454 |
fprintf(ctx->log, "\n"); |
| 455 |
} |
| 456 |
|
| 457 |
return gpu_launch_kernel(ctx, kernel, name, provenance, grid, block, |
| 458 |
TR_TILE_DIM*(TR_TILE_DIM+1)*elem_size, |
| 459 |
sizeof(args)/sizeof(args[0]), args, args_sizes); |
| 460 |
} |
| 461 |
|
| 462 |
#define GEN_MAP_TRANSPOSE_GPU2GPU(NAME, ELEM_TYPE) \ |
| 463 |
static int map_transpose_gpu2gpu_##NAME \ |
| 464 |
(struct futhark_context* ctx, const char* provenance, \ |
| 465 |
gpu_mem dst, int64_t dst_offset, \ |
| 466 |
gpu_mem src, int64_t src_offset, \ |
| 467 |
int64_t k, int64_t m, int64_t n) \ |
| 468 |
{ \ |
| 469 |
return \ |
| 470 |
gpu_map_transpose \ |
| 471 |
(ctx, \ |
| 472 |
provenance, \ |
| 473 |
ctx->kernels->map_transpose_##NAME, \ |
| 474 |
ctx->kernels->map_transpose_##NAME##_low_height, \ |
| 475 |
ctx->kernels->map_transpose_##NAME##_low_width, \ |
| 476 |
ctx->kernels->map_transpose_##NAME##_small, \ |
| 477 |
ctx->kernels->map_transpose_##NAME##_large, \ |
| 478 |
"map_transpose_" #NAME, sizeof(ELEM_TYPE), \ |
| 479 |
dst, dst_offset, src, src_offset, \ |
| 480 |
k, n, m); \ |
| 481 |
} |
| 482 |
|
| 483 |
static int gpu_lmad_copy(struct futhark_context* ctx, const char* provenance, |
| 484 |
gpu_kernel kernel, int r, |
| 485 |
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r], |
| 486 |
gpu_mem src, int64_t src_offset, int64_t src_strides[r], |
| 487 |
int64_t shape[r]) { |
| 488 |
if (r > 8) { |
| 489 |
set_error(ctx, strdup("Futhark runtime limitation:\nCannot copy array of greater than rank 8.\n")); |
| 490 |
return 1; |
| 491 |
} |
| 492 |
|
| 493 |
int64_t n = 1; |
| 494 |
for (int i = 0; i < r; i++) { n *= shape[i]; } |
| 495 |
|
| 496 |
void* args[6+(8*3)]; |
| 497 |
size_t args_sizes[6+(8*3)]; |
| 498 |
|
| 499 |
args[0] = &dst; |
| 500 |
args_sizes[0] = sizeof(gpu_mem); |
| 501 |
args[1] = &dst_offset; |
| 502 |
args_sizes[1] = sizeof(dst_offset); |
| 503 |
args[2] = &src; |
| 504 |
args_sizes[2] = sizeof(gpu_mem); |
| 505 |
args[3] = &src_offset; |
| 506 |
args_sizes[3] = sizeof(src_offset); |
| 507 |
args[4] = &n; |
| 508 |
args_sizes[4] = sizeof(n); |
| 509 |
args[5] = &r; |
| 510 |
args_sizes[5] = sizeof(r); |
| 511 |
|
| 512 |
int64_t zero = 0; |
| 513 |
|
| 514 |
for (int i = 0; i < 8; i++) { |
| 515 |
args_sizes[6+i*3] = sizeof(int64_t); |
| 516 |
args_sizes[6+i*3+1] = sizeof(int64_t); |
| 517 |
args_sizes[6+i*3+2] = sizeof(int64_t); |
| 518 |
if (i < r) { |
| 519 |
args[6+i*3] = &shape[i]; |
| 520 |
args[6+i*3+1] = &dst_strides[i]; |
| 521 |
args[6+i*3+2] = &src_strides[i]; |
| 522 |
} else { |
| 523 |
args[6+i*3] = &zero; |
| 524 |
args[6+i*3+1] = &zero; |
| 525 |
args[6+i*3+2] = &zero; |
| 526 |
} |
| 527 |
} |
| 528 |
const size_t w = 256; // XXX: hardcoded thread block size. |
| 529 |
|
| 530 |
return gpu_launch_kernel(ctx, kernel, "copy_lmad_dev_to_dev", provenance, |
| 531 |
(const int32_t[3]) {(n+w-1)/w,1,1}, |
| 532 |
(const int32_t[3]) {w,1,1}, |
| 533 |
0, 6+(8*3), args, args_sizes); |
| 534 |
} |
| 535 |
|
| 536 |
#define GEN_LMAD_COPY_ELEMENTS_GPU2GPU(NAME, ELEM_TYPE) \ |
| 537 |
static int lmad_copy_elements_gpu2gpu_##NAME \ |
| 538 |
(struct futhark_context* ctx, const char* provenance, \ |
| 539 |
int r, \ |
| 540 |
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r], \ |
| 541 |
gpu_mem src, int64_t src_offset, int64_t src_strides[r], \ |
| 542 |
int64_t shape[r]) { \ |
| 543 |
return gpu_lmad_copy(ctx, provenance, \ |
| 544 |
ctx->kernels->lmad_copy_##NAME, \ |
| 545 |
r, \ |
| 546 |
dst, dst_offset, dst_strides, \ |
| 547 |
src, src_offset, src_strides, \ |
| 548 |
shape); \ |
| 549 |
} \ |
| 550 |
|
| 551 |
#define GEN_LMAD_COPY_GPU2GPU(NAME, ELEM_TYPE) \ |
| 552 |
static int lmad_copy_gpu2gpu_##NAME \ |
| 553 |
(struct futhark_context* ctx, const char* provenance, \ |
| 554 |
int r, \ |
| 555 |
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r], \ |
| 556 |
gpu_mem src, int64_t src_offset, int64_t src_strides[r], \ |
| 557 |
int64_t shape[r]) { \ |
| 558 |
log_copy(ctx, "GPU to GPU", provenance, r, dst_offset, dst_strides, \ |
| 559 |
src_offset, src_strides, shape); \ |
| 560 |
int64_t size = 1; \ |
| 561 |
for (int i = 0; i < r; i++) { size *= shape[i]; } \ |
| 562 |
if (size == 0) { return FUTHARK_SUCCESS; } \ |
| 563 |
int64_t k, n, m; \ |
| 564 |
if (lmad_map_tr(&k, &n, &m, \ |
| 565 |
r, dst_strides, src_strides, shape)) { \ |
| 566 |
log_transpose(ctx, k, n, m); \ |
| 567 |
return map_transpose_gpu2gpu_##NAME \ |
| 568 |
(ctx, provenance, dst, dst_offset, src, src_offset, k, n, m); \ |
| 569 |
} else if (lmad_memcpyable(r, dst_strides, src_strides, shape)) { \ |
| 570 |
if (ctx->logging) {fprintf(ctx->log, "## Flat copy\n\n");} \ |
| 571 |
return gpu_memcpy(ctx, \ |
| 572 |
provenance, \ |
| 573 |
dst, dst_offset*sizeof(ELEM_TYPE), \ |
| 574 |
src, src_offset*sizeof(ELEM_TYPE), \ |
| 575 |
size * sizeof(ELEM_TYPE)); \ |
| 576 |
} else { \ |
| 577 |
if (ctx->logging) {fprintf(ctx->log, "## General copy\n\n");} \ |
| 578 |
return lmad_copy_elements_gpu2gpu_##NAME \ |
| 579 |
(ctx, provenance, \ |
| 580 |
r, \ |
| 581 |
dst, dst_offset, dst_strides, \ |
| 582 |
src, src_offset, src_strides, \ |
| 583 |
shape); \ |
| 584 |
} \ |
| 585 |
} |
| 586 |
|
| 587 |
static int |
| 588 |
lmad_copy_elements_host2gpu(struct futhark_context *ctx, const char *provenance, |
| 589 |
size_t elem_size, |
| 590 |
int r, |
| 591 |
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r], |
| 592 |
unsigned char* src, int64_t src_offset, int64_t src_strides[r], |
| 593 |
int64_t shape[r]) { |
| 594 |
(void)ctx; (void)provenance; (void)elem_size; (void)r; |
| 595 |
(void)dst; (void)dst_offset; (void)dst_strides; |
| 596 |
(void)src; (void)src_offset; (void)src_strides; |
| 597 |
(void)shape; |
| 598 |
set_error(ctx, strdup("Futhark runtime limitation:\nCannot copy unstructured array from host to GPU.\n")); |
| 599 |
return 1; |
| 600 |
} |
| 601 |
|
| 602 |
static int |
| 603 |
lmad_copy_elements_gpu2host (struct futhark_context *ctx, const char *provenance, |
| 604 |
size_t elem_size, |
| 605 |
int r, |
| 606 |
unsigned char* dst, int64_t dst_offset, int64_t dst_strides[r], |
| 607 |
gpu_mem src, int64_t src_offset, int64_t src_strides[r], |
| 608 |
int64_t shape[r]) { |
| 609 |
(void)ctx; (void)provenance; (void)elem_size; (void)r; |
| 610 |
(void)dst; (void)dst_offset; (void)dst_strides; |
| 611 |
(void)src; (void)src_offset; (void)src_strides; |
| 612 |
(void)shape; |
| 613 |
set_error(ctx, strdup("Futhark runtime limitation:\nCannot copy unstructured array from GPU to host.\n")); |
| 614 |
return 1; |
| 615 |
} |
| 616 |
|
| 617 |
static int lmad_copy_host2gpu(struct futhark_context* ctx, const char* provenance, |
| 618 |
size_t elem_size, bool sync, |
| 619 |
int r, |
| 620 |
gpu_mem dst, int64_t dst_offset, int64_t dst_strides[r], |
| 621 |
unsigned char* src, int64_t src_offset, int64_t src_strides[r], |
| 622 |
int64_t shape[r]) { |
| 623 |
log_copy(ctx, "Host to GPU", provenance, r, dst_offset, dst_strides, |
| 624 |
src_offset, src_strides, shape); |
| 625 |
int64_t size = elem_size; |
| 626 |
for (int i = 0; i < r; i++) { size *= shape[i]; } |
| 627 |
if (size == 0) { return FUTHARK_SUCCESS; } |
| 628 |
int64_t k, n, m; |
| 629 |
if (lmad_memcpyable(r, dst_strides, src_strides, shape)) { |
| 630 |
if (ctx->logging) {fprintf(ctx->log, "## Flat copy\n\n");} |
| 631 |
return memcpy_host2gpu(ctx, provenance, |
| 632 |
sync, |
| 633 |
dst, dst_offset*elem_size, |
| 634 |
src, src_offset*elem_size, |
| 635 |
size); |
| 636 |
} else { |
| 637 |
if (ctx->logging) {fprintf(ctx->log, "## General copy\n\n");} |
| 638 |
int error; |
| 639 |
error = lmad_copy_elements_host2gpu |
| 640 |
(ctx, provenance, |
| 641 |
elem_size, r, |
| 642 |
dst, dst_offset, dst_strides, |
| 643 |
src, src_offset, src_strides, |
| 644 |
shape); |
| 645 |
if (error == 0 && sync) { |
| 646 |
error = futhark_context_sync(ctx); |
| 647 |
} |
| 648 |
return error; |
| 649 |
} |
| 650 |
} |
| 651 |
|
| 652 |
static int lmad_copy_gpu2host(struct futhark_context* ctx, const char* provenance, |
| 653 |
size_t elem_size, bool sync, |
| 654 |
int r, |
| 655 |
unsigned char* dst, int64_t dst_offset, int64_t dst_strides[r], |
| 656 |
gpu_mem src, int64_t src_offset, int64_t src_strides[r], |
| 657 |
int64_t shape[r]) { |
| 658 |
log_copy(ctx, "Host to GPU", provenance, r, dst_offset, dst_strides, |
| 659 |
src_offset, src_strides, shape); |
| 660 |
int64_t size = elem_size; |
| 661 |
for (int i = 0; i < r; i++) { size *= shape[i]; } |
| 662 |
if (size == 0) { return FUTHARK_SUCCESS; } |
| 663 |
int64_t k, n, m; |
| 664 |
if (lmad_memcpyable(r, dst_strides, src_strides, shape)) { |
| 665 |
if (ctx->logging) {fprintf(ctx->log, "## Flat copy\n\n");} |
| 666 |
return memcpy_gpu2host(ctx, provenance, |
| 667 |
sync, |
| 668 |
dst, dst_offset*elem_size, |
| 669 |
src, src_offset*elem_size, |
| 670 |
size); |
| 671 |
} else { |
| 672 |
if (ctx->logging) {fprintf(ctx->log, "## General copy\n\n");} |
| 673 |
int error; |
| 674 |
error = lmad_copy_elements_gpu2host |
| 675 |
(ctx, provenance, elem_size, r, |
| 676 |
dst, dst_offset, dst_strides, |
| 677 |
src, src_offset, src_strides, |
| 678 |
shape); |
| 679 |
if (error == 0 && sync) { |
| 680 |
error = futhark_context_sync(ctx); |
| 681 |
} |
| 682 |
return error; |
| 683 |
} |
| 684 |
} |
| 685 |
|
| 686 |
GEN_MAP_TRANSPOSE_GPU2GPU(1b, uint8_t) |
| 687 |
GEN_MAP_TRANSPOSE_GPU2GPU(2b, uint16_t) |
| 688 |
GEN_MAP_TRANSPOSE_GPU2GPU(4b, uint32_t) |
| 689 |
GEN_MAP_TRANSPOSE_GPU2GPU(8b, uint64_t) |
| 690 |
|
| 691 |
GEN_LMAD_COPY_ELEMENTS_GPU2GPU(1b, uint8_t) |
| 692 |
GEN_LMAD_COPY_ELEMENTS_GPU2GPU(2b, uint16_t) |
| 693 |
GEN_LMAD_COPY_ELEMENTS_GPU2GPU(4b, uint32_t) |
| 694 |
GEN_LMAD_COPY_ELEMENTS_GPU2GPU(8b, uint64_t) |
| 695 |
|
| 696 |
GEN_LMAD_COPY_GPU2GPU(1b, uint8_t) |
| 697 |
GEN_LMAD_COPY_GPU2GPU(2b, uint16_t) |
| 698 |
GEN_LMAD_COPY_GPU2GPU(4b, uint32_t) |
| 699 |
GEN_LMAD_COPY_GPU2GPU(8b, uint64_t) |
| 700 |
|
| 701 |
// End of gpu.h |