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
|
//------------------------------------------------------------------------------
// rmm_wrap.cpp: C-callable wrapper for an RMM memory resource
//------------------------------------------------------------------------------
// SPDX-License-Identifier: Apache-2.0
//------------------------------------------------------------------------------
// rmm_wrap.cpp contains a single global object, the RMM_Wrap_Handle that holds
// an RMM (Rapids Memory Manager) memory resource and a hash map (C++
// std:unordered_map). This allows rmm_wrap to provide 7 functions to a C
// application:
// Create/destroy an RMM resource:
// rmm_wrap_initialize: create the RMM resource
// rmm_wrap_finalize: destroy the RMM resource
// C-style malloc/calloc/realloc/free methods:
// rmm_wrap_malloc: malloc a block of memory using RMM
// rmm_wrap_calloc: calloc a block of memory using RMM
// rmm_wrap_realloc: realloc a block of allocated by this RMM wrapper
// rmm_wrap_free: free a block of memory allocated by this RMM wrapper
// PMR-based allocate/deallocate methods (C-callable):
// rmm_wrap_allocate (std::size_t *size)
// rmm_wrap_deallocate (void *p, std::size_t size)
#include "rmm_wrap.hpp"
#include <iostream>
//------------------------------------------------------------------------------
// RMM_Wrap_Handle: a global object containing the RMM context
//------------------------------------------------------------------------------
// rmm_wrap_context is a pointer to a single, global RMM_Wrap_Handle object
// that all methods in this file can access. The object and its pointer
// cannot be accessed outside this file.
typedef struct
{
RMM_MODE mode;
std::shared_ptr<rmm::mr::device_memory_resource> resource;
std::shared_ptr<std::pmr::memory_resource> host_resource;
std::shared_ptr<alloc_map> size_map ;
std::shared_ptr<cuda_stream_pool> stream_pool;
cudaStream_t main_stream;
}
RMM_Wrap_Handle ;
// rmm_wrap_context: global pointer to the single RMM_Wrap_Handle object
static RMM_Wrap_Handle *rmm_wrap_context = NULL ;
//------------------------------------------------------------------------------
// make a resource pool
//------------------------------------------------------------------------------
#if 0
inline auto make_host()
{
return std::make_shared<rmm::mr::new_delete_resource>() ;
}
inline auto make_host_pinned()
{
return std::make_shared<rmm::mr::pinned_memory_resource>() ;
}
#endif
inline auto make_cuda()
{
return std::make_shared<rmm::mr::cuda_memory_resource>() ;
}
inline auto make_managed()
{
return std::make_shared<rmm::mr::managed_memory_resource>() ;
}
#if 0
inline auto make_and_set_host_pool
(
std::size_t initial_size,
std::size_t maximum_size
)
{
auto resource = std::pmr::synchronized_pool_resource() ;
rmm::mr::set_current_device_resource( resource ) ;
return resource;
}
inline auto make_and_set_host_pinned_pool
(
std::size_t initial_size,
std::size_t maximum_size
)
{
auto resource = rmm::mr::make_owning_wrapper<pool_mr>
( make_host_pinned(), initial_size, maximum_size ) ;
rmm::mr::set_current_device_resource( resource.get()) ;
return resource;
}
#endif
// size_map is an unordered alloc_map that maps allocation address to the size
// of each allocation
inline auto make_and_set_device_pool
(
std::size_t initial_size,
std::size_t maximum_size
)
{
auto resource = rmm::mr::make_owning_wrapper<rmm::mr::pool_memory_resource>
( make_cuda(), initial_size, maximum_size ) ;
rmm::mr::set_current_device_resource( resource.get()) ;
return resource;
}
inline auto make_and_set_managed_pool
(
std::size_t initial_size,
std::size_t maximum_size
)
{
// std::cout<< " make_managed_pool called with init_size"
// <<initial_size<<" max_size "<<maximum_size<<"\n";
auto resource = rmm::mr::make_owning_wrapper<rmm::mr::pool_memory_resource>
( make_managed(), initial_size, maximum_size ) ;
rmm::mr::set_current_device_resource( resource.get()) ;
return resource;
}
inline std::shared_ptr<rmm::cuda_stream_pool> make_and_set_cuda_stream_pool
(
std::size_t num_streams
)
{
return std::make_shared<rmm::cuda_stream_pool>(num_streams);
}
inline cudaStream_t make_cuda_stream()
{
cudaStream_t user_stream;
RMM_WRAP_CHECK_CUDA(cudaStreamCreate(&user_stream));
return user_stream;
}
//------------------------------------------------------------------------------
// rmm_wrap_finalize: destroy the global rmm_wrap_context
//------------------------------------------------------------------------------
// Destroy the rmm_wrap_context. This method allows destroys the contents of
// the rmm_wrap_context: the memory resource (host or device) and the
// alloc_map.
void rmm_wrap_finalize (void)
{
if (rmm_wrap_context != NULL)
{
RMM_WRAP_CHECK_CUDA(cudaStreamDestroy(rmm_wrap_context->main_stream));
delete (rmm_wrap_context) ;
rmm_wrap_context = NULL ;
}
}
//------------------------------------------------------------------------------
// rmm_wrap_initialize: initialize the global rmm_wrap_context
//------------------------------------------------------------------------------
int rmm_wrap_initialize // returns -1 on error, 0 on success
(
RMM_MODE mode, // TODO: describe
std::size_t init_pool_size, // TODO: describe
std::size_t max_pool_size, // TODO: describe
std::size_t stream_pool_size
)
{
//--------------------------------------------------------------------------
// check inputs
//--------------------------------------------------------------------------
if (rmm_wrap_context != NULL)
{
// rmm_wrap_initialize cannot be called twice
return (-1) ;
}
if(stream_pool_size <= 0)
{
std::cout << "Stream pool size must be >=0" << std::endl;
// failed to create the alloc_map
return (-1) ;
}
// create the RMM wrap handle and save it as a global pointer.
rmm_wrap_context = new RMM_Wrap_Handle() ;
// std::cout<< " init called with mode "<<mode<<" init_size "
// <<init_pool_size<<" max_size "<<max_pool_size<<"\n";
//--------------------------------------------------------------------------
// Construct a resource that uses a coalescing best-fit pool allocator
//--------------------------------------------------------------------------
// Set CUDA stream pool
rmm_wrap_context->stream_pool = make_and_set_cuda_stream_pool(stream_pool_size);
rmm_wrap_context->main_stream = make_cuda_stream();
if (mode == rmm_wrap_host )
{
// rmm_wrap_context->host_resource =
// std::pmr::synchronized_pool_resource() ;
// // (init_pool_size, max_pool_size) ;
// rmm_wrap_context->host_resource = make_and_set_host_pool() ;
// // (init_pool_size, max_pool_size) ;
}
else if (mode == rmm_wrap_host_pinned )
{
// rmm_wrap_context->host_resource =
// std::pmr::synchronized_pool_resource() ;
// // (init_pool_size, max_pool_size) ;
}
else if (mode == rmm_wrap_device )
{
rmm_wrap_context->resource =
make_and_set_device_pool( init_pool_size, max_pool_size) ;
}
else if ( mode == rmm_wrap_managed )
{
rmm_wrap_context->resource =
make_and_set_managed_pool( init_pool_size, max_pool_size) ;
}
else
{
// invalid mode
return (-1) ;
}
// Mark down the mode for reference later
rmm_wrap_context->mode = mode;
//--------------------------------------------------------------------------
// create size map to lookup size of each allocation
//--------------------------------------------------------------------------
rmm_wrap_context->size_map = std::make_shared<alloc_map> () ;
if (rmm_wrap_context->size_map.get() == NULL)
{
// failed to create the alloc_map
return (-1) ;
}
return (0) ;
}
cudaStream_t get_next_stream_from_pool() {
return rmm_wrap_context->stream_pool->get_stream();
}
cudaStream_t get_stream_from_pool(std::size_t stream_id) {
return rmm_wrap_context->stream_pool->get_stream(stream_id);
}
cudaStream_t get_main_stream() {
return rmm_wrap_context->main_stream;
}
//------------------------------------------------------------------------------
// rmm_wrap_malloc: malloc-equivalent method using RMM
//------------------------------------------------------------------------------
// rmm_wrap_malloc is identical to the ANSI C11 malloc function, except that
// it uses RMM underneath to allocate the space.
void *rmm_wrap_malloc (std::size_t size)
{
return (rmm_wrap_allocate (&size)) ;
}
//------------------------------------------------------------------------------
// rmm_wrap_calloc: calloc-equivalent method using RMM
//------------------------------------------------------------------------------
// rmm_wrap_calloc is identical to the ANSI C11 calloc function, except that
// it uses RMM underneath to allocate the space.
void *rmm_wrap_calloc (std::size_t n, std::size_t size)
{
std::size_t s = n * size ;
void *p = rmm_wrap_allocate (&s) ;
// NOTE: this is single-threaded on the CPU. If you want a faster method,
// malloc the space and use cudaMemset for the GPU or GB_memset on the CPU.
// The GraphBLAS GB_calloc_memory method uses malloc and GB_memset.
memset (p, 0, s) ;
return (p) ;
}
//------------------------------------------------------------------------------
// rmm_wrap_realloc: realloc-equivalent method using RMM
//------------------------------------------------------------------------------
// rmm_wrap_realloc is identical to the ANSI C11 realloc function, except that
// it uses RMM underneath to allocate the space.
void *rmm_wrap_realloc (void *p, std::size_t newsize)
{
if (p == NULL)
{
// allocate a new block. This is OK.
return (rmm_wrap_allocate (&newsize)) ;
}
if (newsize == 0)
{
// free the block. This OK.
rmm_wrap_deallocate (p, 0) ;
return (NULL) ;
}
alloc_map *am = rmm_wrap_context->size_map.get() ;
std::size_t oldsize = am->at( (std::size_t)(p) ) ;
if (oldsize == 0)
{
// the block is not in the hashmap; cannot realloc it.
// This is a failure.
return (NULL) ;
}
// check for quick return
if (newsize >= oldsize/2 && newsize <= oldsize)
{
// Be lazy. If the block does not change, or is shrinking but only by a
// small amount, then leave the block as-is.
return (p) ;
}
// allocate the new space
void *pnew = rmm_wrap_allocate (&newsize) ;
if (pnew == NULL)
{
// old block is not modified. This is a failure, but the old block is
// still in the hashmap.
return (NULL) ;
}
// copy the old space into the new space
std::size_t s = (oldsize < newsize) ? oldsize : newsize ;
// FIXME: query the pointer if it's on the GPU.
memcpy (pnew, p, s) ; // NOTE: single-thread CPU, not GPU. Slow!
// free the old space
rmm_wrap_deallocate (p, oldsize) ;
// return the new space
return (pnew) ;
}
//------------------------------------------------------------------------------
// rmm_wrap_free: free a block of memory, size not needed
//------------------------------------------------------------------------------
// rmm_wrap_free is identical to the ANSI C11 free function, except that
// it uses RMM underneath to allocate the space.
void rmm_wrap_free (void *p)
{
rmm_wrap_deallocate (p, 0) ;
}
//------------------------------------------------------------------------------
// rmm_wrap_allocate: allocate a block of memory using RMM
//------------------------------------------------------------------------------
void *rmm_wrap_allocate( std::size_t *size)
{
if (rmm_wrap_context == NULL) return (NULL) ;
alloc_map *am = rmm_wrap_context->size_map.get() ;
if (am == NULL)
{
// PANIC!
// std::cout<< "Uh oh, can't allocate before initializing RMM"
// << std::endl;
return (NULL) ;
}
// ensure size is nonzero
if (*size == 0) *size = 256 ;
// round-up the allocation to a multiple of 256
std::size_t aligned = (*size) % 256 ;
if (aligned > 0)
{
*size += (256 - aligned) ;
}
// printf(" rmm_wrap_alloc %ld bytes\n",*size) ;
rmm::mr::device_memory_resource *memoryresource =
rmm::mr::get_current_device_resource() ;
void *p = memoryresource->allocate( *size ) ;
if (p == NULL)
{
// out of memory
*size = 0 ;
return (NULL) ;
}
// insert p into the hashmap
am->emplace ((std::size_t)p, (std::size_t)(*size)) ;
// return the allocated block
return (p) ;
}
//------------------------------------------------------------------------------
// rmm_wrap_deallocate: deallocate a block previously allocated by RMM
//------------------------------------------------------------------------------
void rmm_wrap_deallocate( void *p, std::size_t size)
{
if (rmm_wrap_context == NULL) return ;
// Note: there are 3 PANIC cases below. The API of rmm_wrap_deallocate
// does not allow an error condition to be returned. These PANICs could be
// logged, or they could terminate the program if debug mode enabled, etc.
// In production, all we can do is ignore the PANIC.
if (p == NULL)
{
// nothing to do; ignore a double-free
if (size > 0)
{
// PANIC! Why does a NULL pointer have a nonzero size??
}
return ;
}
// check the size given. If the input size is zero, then the
// size is unknown (say rmm_wrap_free(p)). In that case, just trust the
// hashmap. Otherwise, double-check to make sure the size is correct.
alloc_map *am = rmm_wrap_context->size_map.get() ;
size_t actual_size = 0 ;
if (am == NULL)
{
// PANIC!
// std::cout<< "Uh oh, can't deallocate before initializing RMM"
// << std::endl;
return ;
}
else
{
//actual_size = am->at( (std::size_t)(p) ) ;
auto iter = am->find( (std::size_t)(p) ) ;
if (iter != am->end() ) actual_size = iter->second;
else std::cout<< " rmm_wrap:: tried to free unallocated pointer ! " << p ;
}
if (actual_size == 0)
{
// PANIC! oops, p is not in the hashmap. Ignore it. TODO: could add
// a printf here, write to a log file, etc. if debug mode, abort, etc.
return ;
}
if (size > 0 && size != actual_size)
{
// PANIC! oops, invalid old size. Ignore the input size, and free p
// anyway. TODO: could add a printf here, write to a log file, etc.
// if debug mode, abort, etc.
}
// remove p from the hashmap
am->erase ( (std::size_t)(p) ) ;
// deallocate the block of memory
rmm::mr::device_memory_resource *memoryresource =
rmm::mr::get_current_device_resource() ;
memoryresource->deallocate( p, actual_size ) ;
}
|