Fix gfx950 gating conditions to match ROCm 7.0.2 (#2003)

[ROCm/rccl commit: 8444b3c6e9]
This commit is contained in:
Nilesh M Negi
2025-10-29 23:27:04 -05:00
committed by GitHub
parent eb0b1387b7
commit 03d37f6305
2 changed files with 5 additions and 5 deletions
+4 -4
View File
@@ -270,8 +270,8 @@ private:
i4.flag2 = flag;
*((u64_gptr) dst->v) = *((u64_gptr) i4.v);
*((u64_gptr) dst->v+1) = *((u64_gptr) i4.v+1);
#if defined(__gfx950__) && ROCM_VERSION < 70200
__builtin_amdgcn_fence(__ATOMIC_RELEASE, ""); // flush cache
#if defined(__gfx950__) && ROCM_VERSION < 70002
__builtin_amdgcn_fence(__ATOMIC_RELEASE, ""); // flush cache on gfx950 if ROCr fix for hipHostMallocUncached is not available (ROCm version < 7.0.2)
#endif
#else
asm volatile("st.volatile.global.v4.u32 [%0], {%1,%2,%3,%4};" :: "l"(&dst->i4), "r"((uint32_t)val), "r"(flag), "r"((uint32_t)(val >> 32)), "r"(flag) : "memory");
@@ -346,8 +346,8 @@ private:
__builtin_nontemporal_store(u4, (uint32_t*)dst);
else
__builtin_nontemporal_store(u8, (uint64_t*)dst);
#if defined(__gfx950__) && ROCM_VERSION < 70200
__builtin_amdgcn_fence(__ATOMIC_RELEASE, ""); // flush cache
#if defined(__gfx950__) && ROCM_VERSION < 70002
__builtin_amdgcn_fence(__ATOMIC_RELEASE, ""); // flush cache on gfx950 if ROCr fix for hipHostMallocUncached is not available (ROCm version < 7.0.2)
#endif
#else
if(sizeof(U) == 1)