Skip to content

Bugfix - Use volatile accesses for HIP zero-copy loads/stores in gpu-copy-bw #838

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Hongtao Zhang (polarG) merged 5 commits into from
Aug 31, 2026
Merged

Bugfix - Use volatile accesses for HIP zero-copy loads/stores in gpu-copy-bw #838

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
f57ce61
Fix gpu-copy-bw correctness: use volatile SM-copy load/store on HIP
polarG Jul 26, 2026
28950ca
CI/CD - Only cap setuptools below 66 for Python 3.7
Jul 31, 2026
feb7200
Merge branch 'main' into dev/hongtaozhang/gpu-copy-bw-correctness
polarG Aug 24, 2026
b9514bf
Fix HIP volatile copy accesses
Aug 24, 2026
5d76241
Use declared HIP vector member types
Aug 24, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 12 additions & 4 deletions superbench/benchmarks/micro_benchmarks/gpu_copy_performance/gpu_copy.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -599,8 +599,12 @@ int DestroyEvent(BenchArgs *args) {
// https://github.com/ROCmSoftwarePlatform/rccl/blob/5c8380ff5b5925cae4bce00b1879a5f930226e8d/src/collectives/device/common_kernel.h#L268
inline __device__ void FetchULong2(ulong2 &v, const ulong2 *p) {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
v.x = p->x;
v.y = p->y;
// ROCm clang drops volatile when ulong2 members are accessed through a volatile ulong2 pointer,
// so use volatile scalar member pointers to prevent compiler elision or reordering.
const volatile decltype(p->x) *x_pointer = &p->x;
const volatile decltype(p->y) *y_pointer = &p->y;
v.x = *x_pointer;
v.y = *y_pointer;
#else
asm volatile("ld.volatile.global.v2.u64 {%0,%1}, [%2];" : "=l"(v.x), "=l"(v.y) : "l"(p) : "memory");
#endif
Expand All @@ -614,8 +618,12 @@ inline __device__ void FetchULong2(ulong2 &v, const ulong2 *p) {
// https://github.com/ROCmSoftwarePlatform/rccl/blob/5c8380ff5b5925cae4bce00b1879a5f930226e8d/src/collectives/device/common_kernel.h#L276
inline __device__ void StoreULong2(ulong2 *p, ulong2 &v) {
#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
p->x = v.x;
p->y = v.y;
// ROCm clang drops volatile when ulong2 members are accessed through a volatile ulong2 pointer,
// so use volatile scalar member pointers to prevent compiler elision or reordering.
volatile decltype(p->x) *x_pointer = &p->x;
volatile decltype(p->y) *y_pointer = &p->y;
*x_pointer = v.x;
*y_pointer = v.y;
#else
asm volatile("st.volatile.global.v2.u64 [%0], {%1,%2};" ::"l"(p), "l"(v.x), "l"(v.y) : "memory");
#endif
Expand Down
Loading