Blockwise Scaling for FP8 (#1932)

* F8 Blockwise Scaling

* two more NumProducerThreadEvents

---------

Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Manish Gupta
2025-01-09 11:22:09 -05:00
committed by GitHub
co-authored by Haicheng Wu
parent 375e284e6a
commit ef5620dd1d
21 changed files with 2337 additions and 29 deletions
+17
View File
@@ -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
}
////////////////////////////////////////////////////////////////////////////////////////////////////
////////////////////////////////////////////////////////////////////////////////////////////////////