ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 06:49:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add validity checks for MoE FlatMM scatter and enable bf16 hardware atomic-add (#3236)

Browse Source 
* Add validity checks for MoE FlatMM scatter and enable bf16 hardware atomic

* correct clang-format

* removed unused rtol_atol variable from example code

* clang format correction

* remove unused varable max_accumulated_value from example
This commit is contained in:
msaffari-amd
2025-11-28 09:43:01 +01:00
committed by GitHub
parent 30727c48fc
commit f875ab0bbc
2 changed files with 10 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
include/ck_tile/core/arch/generic_memory_space_atomic.hpp
View File

@@ -102,6 +102,9 @@ CK_TILE_DEVICE void atomic_add(X* p_dst, const X& x);
template <>
CK_TILE_DEVICE void atomic_add<bf16x2_t>(bf16x2_t* p_dst, const bf16x2_t& x)
{
#if HAS_GLOBAL_ATOMIC_PK_ADD_BUILTIN
__builtin_amdgcn_global_atomic_fadd_v2bf16(c_style_pointer_cast<bf16x2_t*>(p_dst), x);
#else
union U32BF162_ADDR
{
uint32_t* u32_a;
@@ -128,6 +131,7 @@ CK_TILE_DEVICE void atomic_add<bf16x2_t>(bf16x2_t* p_dst, const bf16x2_t& x)
new_v = new_.u32;
cur_v.u32 = atomicCAS(dword_addr.u32_a, old_v, new_v);
} while(cur_v.u32 != old_v);
#endif
}
template <>