AI/rocm-systems
2
0
Förgrening 0
Kod Ärenden Pull-förfrågningar Actions Packages Projekt Släpp Wiki Aktiviteter

Remove node-count and threshold restrictions from p2p-batching (#2077)

Bläddra i källkod 
* remove node-count and threshold restrictions from p2p-batching

* remove batching threshold usage, fix typo for using batching-enablement flag

---------

Co-authored-by: Mustafa Abduljabbar <mustafa.abduljabbar@amd.com>
This commit is contained in:
isaki001
2025-12-15 18:55:46 -06:00
committad av GitHub
förälder 5787c960fc
incheckning 7c1049d2a4
2 ändrade filer med 4 tillägg och 16 borttagningar
Download Patch File Download Diff File Expand all files Collapse all files
src/device/common.h
+2 -2
Visa fil
@@ -288,10 +288,10 @@ __device__ __forceinline__ void loadWorkBatchToShmem(
if (WARP_SIZE == 64) {
if (uint64_t(batch.offsetBitset) & (1ull<<lane)) {
int nWorksBelow = __popc(uint64_t(batch.offsetBitset) & ((1ull<<lane)-1));
int nWorksBelow = __popcll(uint64_t(batch.offsetBitset) & ((1ull<<lane)-1));
fnsOfBitset[nWorksBelow] = lane;
}
nWorks = __popc(uint64_t(batch.offsetBitset));
nWorks = __popcll(uint64_t(batch.offsetBitset));
} else {
// WARP_SIZE == 32
if (uint32_t(batch.offsetBitset) & (1u<<lane)) {