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

Substitute pow2bin allocator with a dlmalloc based allocator (#71)

Browse Source 
* Add dlmalloc_strat allocator strategy
 - Use mspace variant to ease encapsulation
 - Make pow2bins and dlmalloc cmake selectable
* Add unit tester for dlmalloc, rework single_heap, pow2bins unit testers
accordingly
 - add dlmalloc get_used/get_avail, and have all strats allocators also have a get_used
 - Rework memallocator unit tests: bin size is per strat, alignment is verified in singleheap
* bugfix: dlmalloc exposed that the pingpong test would write past end of
allocation with -w 32
* iostream leakage/mixed usage of cerr and fprintf(stderr

---------

Signed-off-by: Aurelien Bouteiller <aurelien.bouteiller@amd.com>
This commit is contained in:
Aurelien Bouteiller
2025-05-01 11:55:23 -04:00
committed by GitHub
parent edcd1ed57e
commit b835de6cd5
20 changed files with 6956 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/functional_tests/ping_all_tester.cpp
+5 -4
View File
@@ -47,8 +47,8 @@ __global__ void PingAllTest(int loop, int skip, long long int *start_time,
status[j] = 0;
}
if (hipThreadIdx_x == 0) {
auto blk_pe_off {hipBlockIdx_x * num_pe};
if (is_thread_zero_in_block()) {
auto blk_pe_off {wg_id * num_pe};
for (int i = 0; i < loop + skip; i++) {
if (i == skip) {
@@ -61,6 +61,7 @@ __global__ void PingAllTest(int loop, int skip, long long int *start_time,
rocshmem_int_wait_until_all(&r_buf[blk_pe_off], num_pe, status, ROCSHMEM_CMP_EQ, 1);
}
end_time[wg_id] = wall_clock64();
rocshmem_ctx_quiet(ctx);
}
rocshmem_wg_ctx_destroy(&ctx);
rocshmem_wg_finalize();
@@ -71,14 +72,14 @@ __global__ void PingAllTest(int loop, int skip, long long int *start_time,
*****************************************************************************/
PingAllTester::PingAllTester(TesterArguments args) : Tester(args) {
int num_pes {rocshmem_n_pes()};
r_buf = (int *)rocshmem_malloc(sizeof(int) * args.wg_size * num_pes);
r_buf = (int *)rocshmem_malloc(sizeof(int) * args.num_wgs * num_pes);
}
PingAllTester::~PingAllTester() { rocshmem_free(r_buf); }
void PingAllTester::resetBuffers(uint64_t size) {
int num_pes {rocshmem_n_pes()};
memset(r_buf, 0, sizeof(int) * args.wg_size * num_pes);
memset(r_buf, 0, sizeof(int) * args.num_wgs * num_pes);
}
void PingAllTester::launchKernel(dim3 gridSize, dim3 blockSize, int loop,
tests/functional_tests/ping_pong_tester.cpp
+6 -3
View File
@@ -42,7 +42,7 @@ __global__ void PingPongTest(int loop, int skip, long long int *start_time,
int pe = rocshmem_ctx_my_pe(ctx);
if (hipThreadIdx_x == 0) {
if (is_thread_zero_in_block()) {
for (int i = 0; i < loop + skip; i++) {
if (i == skip) {
@@ -60,7 +60,10 @@ __global__ void PingPongTest(int loop, int skip, long long int *start_time,
}
}
end_time[wg_id] = wall_clock64();
rocshmem_ctx_quiet(ctx);
}
rocshmem_wg_ctx_destroy(&ctx);
rocshmem_wg_finalize();
}
@@ -69,13 +72,13 @@ __global__ void PingPongTest(int loop, int skip, long long int *start_time,
* HOST TESTER CLASS METHODS
*****************************************************************************/
PingPongTester::PingPongTester(TesterArguments args) : Tester(args) {
r_buf = (int *)rocshmem_malloc(sizeof(int) * args.wg_size);
r_buf = (int *)rocshmem_malloc(sizeof(int) * args.num_wgs);
}
PingPongTester::~PingPongTester() { rocshmem_free(r_buf); }
void PingPongTester::resetBuffers(uint64_t size) {
memset(r_buf, 0, sizeof(int) * args.wg_size);
memset(r_buf, 0, sizeof(int) * args.num_wgs);
}
void PingPongTester::launchKernel(dim3 gridSize, dim3 blockSize, int loop,