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

Update bandwidth and latency calculations

Browse Source 
- Refined bandwidth and latency calculations for improved accuracy
This commit is contained in:
avinashkethineedi
2025-02-17 06:18:46 +00:00
parent 40bd8a38a0
commit c155636da4
26 changed files with 283 additions and 270 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/functional_tests/ping_all_tester.cpp
+7 -5
View File
@@ -29,9 +29,11 @@ using namespace rocshmem;
/******************************************************************************
* DEVICE TEST KERNEL
*****************************************************************************/
__global__ void PingAllTest(int loop, int skip, uint64_t *timer, int *r_buf,
__global__ void PingAllTest(int loop, int skip, long long int *start_time,
long long int *end_time, int *r_buf,
ShmemContextType ctx_type) {
__shared__ rocshmem_ctx_t ctx;
int wg_id = get_flat_grid_id();
rocshmem_wg_init();
rocshmem_wg_ctx_create(ctx_type, &ctx);
@@ -44,12 +46,11 @@ __global__ void PingAllTest(int loop, int skip, uint64_t *timer, int *r_buf,
}
if (hipThreadIdx_x == 0) {
uint64_t start;
auto blk_pe_off {hipBlockIdx_x * num_pe};
for (int i = 0; i < loop + skip; i++) {
if (i == skip) {
start = rocshmem_timer();
start_time[wg_id] = wall_clock64();
}
for (int j{0}; j < num_pe; j++) {
@@ -57,7 +58,7 @@ __global__ void PingAllTest(int loop, int skip, uint64_t *timer, int *r_buf,
}
rocshmem_int_wait_until_all(&r_buf[blk_pe_off], num_pe, status, ROCSHMEM_CMP_EQ, 1);
}
timer[hipBlockIdx_x] = rocshmem_timer() - start;
end_time[wg_id] = wall_clock64();
}
rocshmem_wg_ctx_destroy(&ctx);
rocshmem_wg_finalize();
@@ -83,7 +84,8 @@ void PingAllTester::launchKernel(dim3 gridSize, dim3 blockSize, int loop,
size_t shared_bytes = 0;
hipLaunchKernelGGL(PingAllTest, gridSize, blockSize, shared_bytes, stream,
loop, args.skip, timer, r_buf, _shmem_context);
loop, args.skip, start_time, end_time, r_buf,
_shmem_context);
num_msgs = (loop + args.skip) * gridSize.x;
num_timed_msgs = loop;