AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

libhsakmt: optimize big system buffer allocation

Browse Source 
To change biggest single buffer to be huge page aligned
and other optimization.

Signed-off-by: Eric Huang <jinhuieric.huang@amd.com>


[ROCm/ROCR-Runtime commit: afe7965796]
This commit is contained in:
Eric Huang
2025-05-26 15:24:09 -04:00
committed by Huang, JinHuiEric
parent 2c6f84b12c
commit 0d5e261f39
1 changed files with 10 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rocr-runtime/libhsakmt/src/fmm.c
+10 -18
View File
@@ -90,8 +90,6 @@ struct vm_object {
* size, Thunk aligns it to page size and allocates this
* aligned size on GPU
*/
uint64_t *handles; /* kfd handles array */
uint32_t handle_num; /* number of handles */
uint32_t node_id;
rbtree_node_t node;
rbtree_node_t user_node;
@@ -118,6 +116,8 @@ struct vm_object {
int mmap_fd;
off_t mmap_offset;
#endif
uint32_t handle_num; /* number of handles */
uint64_t handles[]; /* kfd handles array */
};
typedef struct vm_object vm_object_t;
@@ -331,30 +331,26 @@ static vm_area_t *vm_create_and_init_area(void *start, void *end)
return area;
}
/* One page smaller than 512GB system buffer limit,
/* One huge page smaller than 512GB system buffer limit,
* because 512GB allocation will cause TTM failure.
*/
#define BIGGEST_SINGLE_BUF_SIZE ((1ULL << 39) - PAGE_SIZE)
#define BIGGEST_SINGLE_BUF_SIZE ((1ULL << 39) - GPU_HUGE_PAGE_SIZE)
static vm_object_t *vm_create_and_init_object(void *start, uint64_t size,
uint64_t handle, HsaMemFlags mflags)
{
vm_object_t *object = (vm_object_t *) malloc(sizeof(vm_object_t));
uint64_t handle_array_size;
vm_object_t *object;
uint64_t handle_array_size = (size + BIGGEST_SINGLE_BUF_SIZE - 1) /
BIGGEST_SINGLE_BUF_SIZE;
object = (vm_object_t *) malloc(sizeof(vm_object_t) +
handle_array_size * sizeof(uint64_t));
if (object) {
object->start = start;
object->userptr = NULL;
object->userptr_size = 0;
object->size = size;
handle_array_size = (size + BIGGEST_SINGLE_BUF_SIZE - 1) /
BIGGEST_SINGLE_BUF_SIZE;
object->handles = (uint64_t *)malloc(handle_array_size *
sizeof(uint64_t));
if (!object->handles) {
free(object);
return NULL;
}
object->handles[0] = handle;
object->handle_num = 1;
object->registered_device_id_array_size = 0;
@@ -402,9 +398,6 @@ static void vm_remove_area(manageable_aperture_t *app, vm_area_t *area)
static void vm_remove_object(manageable_aperture_t *app, vm_object_t *object)
{
/* Free allocations inside the object */
if (object->handles)
free(object->handles);
if (object->registered_device_id_array)
free(object->registered_device_id_array);
@@ -1136,7 +1129,6 @@ static vm_object_t *fmm_allocate_memory_object(uint32_t gpu_id, void *mem,
struct kfd_ioctl_free_memory_of_gpu_args free_args = {0};
vm_object_t *vm_obj = NULL;
HsaMemFlags mflags;
int i;
uint64_t offset = 0, total_size, size;
if (!mem)