Blockwise Scaling for FP8 (#1932)
* F8 Blockwise Scaling * two more NumProducerThreadEvents --------- Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
co-authored by
Haicheng Wu
parent
375e284e6a
commit
ef5620dd1d
@@ -716,6 +716,23 @@ void cpasync_barrier_arrive(uint64_t const* smem_ptr) {
|
||||
#endif
|
||||
}
|
||||
|
||||
// Arrive on completion of in-flight cp.async operations issued by the calling thread (noinc)
|
||||
CUTLASS_DEVICE
|
||||
void cpasync_barrier_arrive_noinc(uint64_t const* smem_ptr) {
|
||||
#if CUDA_BARRIER_ENABLED
|
||||
uint32_t smem_addr = cute::cast_smem_ptr_to_uint(smem_ptr);
|
||||
asm volatile(
|
||||
"{\n\t"
|
||||
"cp.async.mbarrier.arrive.noinc.shared::cta.b64 [%0];\n\t"
|
||||
"}"
|
||||
:
|
||||
: "r"(smem_addr));
|
||||
cutlass::arch::synclog_emit_cpasync_barrier_arrive(__LINE__, smem_addr);
|
||||
#elif defined(__CUDA_ARCH__)
|
||||
asm volatile ("brkpt;\n" ::);
|
||||
#endif
|
||||
}
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
Reference in New Issue
Block a user