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
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@author Mark Gates
*/
#include <stdlib.h>
#include <stdio.h>
#ifdef DEBUG_MEMORY
#include <map>
#include <mutex> // requires C++11
#endif
#include <cuda_runtime.h>
#include "magma_v2.h"
#include "magma_internal.h"
#include "error.h"
//#ifdef MAGMA_HAVE_CUDA
#ifdef DEBUG_MEMORY
std::mutex g_pointers_mutex; // requires C++11
std::map< void*, size_t > g_pointers_dev;
std::map< void*, size_t > g_pointers_cpu;
std::map< void*, size_t > g_pointers_pin;
#endif
/***************************************************************************//**
Allocates memory on the GPU. CUDA imposes a synchronization.
Use magma_free() to free this memory.
@param[out]
ptrPtr On output, set to the pointer that was allocated.
NULL on failure.
@param[in]
size Size in bytes to allocate. If size = 0, allocates some minimal size.
@return MAGMA_SUCCESS
@return MAGMA_ERR_DEVICE_ALLOC on failure
Type-safe versions avoid the need for a (void**) cast and explicit sizeof.
@see magma_smalloc
@see magma_dmalloc
@see magma_cmalloc
@see magma_zmalloc
@see magma_imalloc
@see magma_index_malloc
@ingroup magma_malloc
*******************************************************************************/
extern "C" magma_int_t
magma_malloc( magma_ptr* ptrPtr, size_t size )
{
// malloc and free sometimes don't work for size=0, so allocate some minimal size
if ( size == 0 )
size = sizeof(magmaDoubleComplex);
if ( cudaSuccess != cudaMalloc( ptrPtr, size )) {
return MAGMA_ERR_DEVICE_ALLOC;
}
#ifdef DEBUG_MEMORY
g_pointers_mutex.lock();
g_pointers_dev[ *ptrPtr ] = size;
g_pointers_mutex.unlock();
#endif
return MAGMA_SUCCESS;
}
/***************************************************************************//**
@fn magma_free( ptr )
Frees GPU memory previously allocated by magma_malloc().
@param[in]
ptr Pointer to free.
@return MAGMA_SUCCESS
@return MAGMA_ERR_INVALID_PTR on failure
@ingroup magma_malloc
*******************************************************************************/
extern "C" magma_int_t
magma_free_internal( magma_ptr ptr,
const char* func, const char* file, int line )
{
#ifdef DEBUG_MEMORY
g_pointers_mutex.lock();
if ( ptr != NULL && g_pointers_dev.count( ptr ) == 0 ) {
fprintf( stderr, "magma_free( %p ) that wasn't allocated with magma_malloc.\n", ptr );
}
else {
g_pointers_dev.erase( ptr );
}
g_pointers_mutex.unlock();
#endif
cudaError_t err = cudaFree( ptr );
check_xerror( err, func, file, line );
if ( err != cudaSuccess ) {
return MAGMA_ERR_INVALID_PTR;
}
return MAGMA_SUCCESS;
}
/***************************************************************************//**
Allocate size bytes on CPU.
The purpose of using this instead of malloc is to properly align arrays
for vector (SSE, AVX) instructions. The default implementation uses
posix_memalign (on Linux, MacOS, etc.) or _aligned_malloc (on Windows)
to align memory to a 64 byte boundary (typical cache line size).
Use magma_free_cpu() to free this memory.
@param[out]
ptrPtr On output, set to the pointer that was allocated.
NULL on failure.
@param[in]
size Size in bytes to allocate. If size = 0, allocates some minimal size.
@return MAGMA_SUCCESS
@return MAGMA_ERR_HOST_ALLOC on failure
Type-safe versions avoid the need for a (void**) cast and explicit sizeof.
@see magma_smalloc_cpu
@see magma_dmalloc_cpu
@see magma_cmalloc_cpu
@see magma_zmalloc_cpu
@see magma_imalloc_cpu
@see magma_index_malloc_cpu
@ingroup magma_malloc_cpu
*******************************************************************************/
extern "C" magma_int_t
magma_malloc_cpu( void** ptrPtr, size_t size )
{
// malloc and free sometimes don't work for size=0, so allocate some minimal size
if ( size == 0 )
size = sizeof(magmaDoubleComplex);
#if 1
#if defined( _WIN32 ) || defined( _WIN64 )
*ptrPtr = _aligned_malloc( size, 64 );
if ( *ptrPtr == NULL ) {
return MAGMA_ERR_HOST_ALLOC;
}
#else
int err = posix_memalign( ptrPtr, 64, size );
if ( err != 0 ) {
*ptrPtr = NULL;
return MAGMA_ERR_HOST_ALLOC;
}
#endif
#else
*ptrPtr = malloc( size );
if ( *ptrPtr == NULL ) {
return MAGMA_ERR_HOST_ALLOC;
}
#endif
#ifdef DEBUG_MEMORY
g_pointers_mutex.lock();
g_pointers_cpu[ *ptrPtr ] = size;
g_pointers_mutex.unlock();
#endif
return MAGMA_SUCCESS;
}
/***************************************************************************//**
Frees CPU memory previously allocated by magma_malloc_cpu().
The default implementation uses free(),
which works for both malloc and posix_memalign.
For Windows, _aligned_free() is used.
@param[in]
ptr Pointer to free.
@return MAGMA_SUCCESS
@return MAGMA_ERR_INVALID_PTR on failure
@ingroup magma_malloc_cpu
*******************************************************************************/
extern "C" magma_int_t
magma_free_cpu( void* ptr )
{
#ifdef DEBUG_MEMORY
g_pointers_mutex.lock();
if ( ptr != NULL && g_pointers_cpu.count( ptr ) == 0 ) {
fprintf( stderr, "magma_free_cpu( %p ) that wasn't allocated with magma_malloc_cpu.\n", ptr );
}
else {
g_pointers_cpu.erase( ptr );
}
g_pointers_mutex.unlock();
#endif
#if defined( _WIN32 ) || defined( _WIN64 )
_aligned_free( ptr );
#else
free( ptr );
#endif
return MAGMA_SUCCESS;
}
/***************************************************************************//**
Allocates memory on the CPU in pinned memory.
Use magma_free_pinned() to free this memory.
@param[out]
ptrPtr On output, set to the pointer that was allocated.
NULL on failure.
@param[in]
size Size in bytes to allocate. If size = 0, allocates some minimal size.
@return MAGMA_SUCCESS
@return MAGMA_ERR_HOST_ALLOC on failure
Type-safe versions avoid the need for a (void**) cast and explicit sizeof.
@see magma_smalloc_pinned
@see magma_dmalloc_pinned
@see magma_cmalloc_pinned
@see magma_zmalloc_pinned
@see magma_imalloc_pinned
@see magma_index_malloc_pinned
@ingroup magma_malloc_pinned
*******************************************************************************/
extern "C" magma_int_t
magma_malloc_pinned( void** ptrPtr, size_t size )
{
// malloc and free sometimes don't work for size=0, so allocate some minimal size
// (for pinned memory, the error is detected in free)
if ( size == 0 )
size = sizeof(magmaDoubleComplex);
if ( cudaSuccess != cudaHostAlloc( ptrPtr, size, cudaHostAllocPortable )) {
return MAGMA_ERR_HOST_ALLOC;
}
#ifdef DEBUG_MEMORY
g_pointers_mutex.lock();
g_pointers_pin[ *ptrPtr ] = size;
g_pointers_mutex.unlock();
#endif
return MAGMA_SUCCESS;
}
/***************************************************************************//**
@fn magma_free_pinned( ptr )
Frees CPU pinned memory previously allocated by magma_malloc_pinned().
@param[in]
ptr Pointer to free.
@return MAGMA_SUCCESS
@return MAGMA_ERR_INVALID_PTR on failure
@ingroup magma_malloc_pinned
*******************************************************************************/
extern "C" magma_int_t
magma_free_pinned_internal( void* ptr,
const char* func, const char* file, int line )
{
#ifdef DEBUG_MEMORY
g_pointers_mutex.lock();
if ( ptr != NULL && g_pointers_pin.count( ptr ) == 0 ) {
fprintf( stderr, "magma_free_pinned( %p ) that wasn't allocated with magma_malloc_pinned.\n", ptr );
}
else {
g_pointers_pin.erase( ptr );
}
g_pointers_mutex.unlock();
#endif
cudaError_t err = cudaFreeHost( ptr );
check_xerror( err, func, file, line );
if ( cudaSuccess != err ) {
return MAGMA_ERR_INVALID_PTR;
}
return MAGMA_SUCCESS;
}
/***************************************************************************//**
@fn magma_mem_info( free, total )
Sets the parameters 'free' and 'total' to the free and total memory in the
system (in bytes).
@param[in]
free Address of the result for 'free' bytes on the system
total Address of the result for 'total' bytes on the system
@return MAGMA_SUCCESS
@return MAGMA_ERR_INVALID_PTR on failure
*******************************************************************************/
extern "C" magma_int_t
magma_mem_info(size_t * freeMem, size_t * totalMem) {
cudaMemGetInfo(freeMem, totalMem);
return MAGMA_SUCCESS;
}
extern "C" magma_int_t
magma_memset(void * ptr, int value, size_t count) {
return cudaMemset(ptr, value, count);
}
extern "C" magma_int_t
magma_memset_async(void * ptr, int value, size_t count, magma_queue_t queue) {
#ifdef MAGMA_HAVE_CUDA
// return cudaMemsetAsync(ptr, value, count, queue);
return cudaMemsetAsync(ptr, value, count, queue->cuda_stream());
#elif defined(MAGMA_HAVE_HIP)
return hipMemsetAsync(ptr, value, count, queue->hip_stream());
#endif
}
//#endif // MAGMA_HAVE_CUDA
|