Skip to content

fix(rccl) JIRA ID : ROCM-25479 LL: cache-bypass user buffers, plain/non-temporal comm FIFO #9336

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
wenkaidu merged 3 commits into from
Jul 31, 2026
Merged

fix(rccl) JIRA ID : ROCM-25479 LL: cache-bypass user buffers, plain/non-temporal comm FIFO #9336

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
fbe5fbe
[RCCL] LL: cache-bypass user buffers, plain/non-temporal comm FIFO
wenkaidu Jul 7, 2026
9c49e83
Fix for comments
wenkaidu Jul 28, 2026
faa0b8d
Merge branch 'develop' into ll-cache-bypass-develop
wenkaidu Jul 31, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 20 additions & 10 deletions projects/rccl/src/device/prims_ll.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -154,7 +154,10 @@ class Primitives<T, RedOp, Fan, Direct, ProtoLL, P2p, isNetOffload, Metadata, Pi
: "=v"(i4.i4)
: "v"(&src->i4));
#elif RCCL_HAVE_GLOBAL_DWORDX4_BUILTINS
i4.v4u = __builtin_amdgcn_global_load_b128((v4u_gptr)src->v, RCCL_SYSTEM_SYNCSCOPE);
// Comm FIFO buffers are uncached; use non-temporal loads so the flag poll
// never observes a stale cache line (no system-scope cache-bypass needed).
*((u64_gptr)i4.v) = __builtin_nontemporal_load((u64_gptr)src->v);
Comment thread
wenkaidu marked this conversation as resolved.
Resolved
*((u64_gptr)i4.v + 1) = __builtin_nontemporal_load((u64_gptr)src->v + 1);
#else
*((u64_gptr)i4.v) = __builtin_nontemporal_load((u64_gptr)src->v);
*((u64_gptr)i4.v + 1) = __builtin_nontemporal_load((u64_gptr)src->v + 1);
Expand Down Expand Up @@ -191,10 +194,12 @@ class Primitives<T, RedOp, Fan, Direct, ProtoLL, P2p, isNetOffload, Metadata, Pi
: "=v"(line[i].i4)
: "v"(&src->i4));
#elif RCCL_HAVE_GLOBAL_DWORDX4_BUILTINS
line[i].v4u = __builtin_amdgcn_global_load_b128((v4u_gptr)src->v, RCCL_SYSTEM_SYNCSCOPE);
// Comm FIFO buffers are uncached; use non-temporal loads (no bypass).
Comment thread
wenkaidu marked this conversation as resolved.
Resolved
line[i].v[0] = __builtin_nontemporal_load((u64_gptr)src->v);
line[i].v[1] = __builtin_nontemporal_load((u64_gptr)src->v + 1);
#else
line[i].v[0] = __builtin_nontemporal_load(src->v);
line[i].v[1] = __builtin_nontemporal_load(src->v + 1);
line[i].v[0] = __builtin_nontemporal_load((u64_gptr)src->v);
line[i].v[1] = __builtin_nontemporal_load((u64_gptr)src->v + 1);
#endif
#else
asm volatile("ld.volatile.global.v4.u32 {%0,%1,%2,%3}, [%4];"
Expand All @@ -218,7 +223,9 @@ class Primitives<T, RedOp, Fan, Direct, ProtoLL, P2p, isNetOffload, Metadata, Pi
: "=v"(line[i].i4)
: "v"(&src->i4));
#elif RCCL_HAVE_GLOBAL_DWORDX4_BUILTINS
line[i].v4u = __builtin_amdgcn_global_load_b128((v4u_gptr)src->v, RCCL_SYSTEM_SYNCSCOPE);
// Comm FIFO buffers are uncached; use non-temporal loads (no bypass).
line[i].v[0] = __builtin_nontemporal_load((u64_gptr)src->v);
line[i].v[1] = __builtin_nontemporal_load((u64_gptr)src->v + 1);
#else
line[i].v[0] = __builtin_nontemporal_load((u64_gptr)src->v);
line[i].v[1] = __builtin_nontemporal_load((u64_gptr)src->v + 1);
Expand All @@ -244,8 +251,10 @@ class Primitives<T, RedOp, Fan, Direct, ProtoLL, P2p, isNetOffload, Metadata, Pi
i4.data2 = (val >> 32);
i4.flag2 = flag;
#if RCCL_HAVE_GLOBAL_DWORDX4_BUILTINS
// System scope store that bypasses the hardware caches, should generate global_store_dwordx4 instruction with sc0 and sc1 bits set to 1 on gfx942/gfx950.
__builtin_amdgcn_global_store_b128((v4u_gptr)dst->v, i4.v4u, RCCL_SYSTEM_SYNCSCOPE);
// Comm FIFO buffers are uncached; use plain stores (no system-scope cache
// bypass) for higher store throughput and lower register pressure.
*((u64_gptr)dst->v) = *((u64_gptr)i4.v);
*((u64_gptr)dst->v + 1) = *((u64_gptr)i4.v + 1);
#else
#if defined(__gfx1200__) || defined(__gfx1201__)
__hip_atomic_store((u64_gptr)dst->v, i4.v[0], __ATOMIC_RELAXED, __HIP_MEMORY_SCOPE_SYSTEM);
Expand Down Expand Up @@ -408,7 +417,8 @@ class Primitives<T, RedOp, Fan, Direct, ProtoLL, P2p, isNetOffload, Metadata, Pi

__device__ void storeData(T* dst, uint64_t val, int eltN) {
if (__all((reinterpret_cast<uintptr_t>(dst) & (sizeof(T) - 1)) == 0 && sizeof(T) * eltN == sizeof(val))) {
*((u64_gptr)dst) = val;
// Cache-bypassing (system-scope) store to the user output buffer.
store(reinterpret_cast<uint64_t*>(dst), val);
return;
}
union {
Expand All @@ -421,8 +431,8 @@ class Primitives<T, RedOp, Fan, Direct, ProtoLL, P2p, isNetOffload, Metadata, Pi
// Yes, for some template arguments this code will be unreachable. That's fine.
// coverity[dead_error_line]
if (i == 0 || i < eltN)
// store(dst+i, elt[i]);
dst[i] = elt[i];
// Cache-bypassing (system-scope) store to the user output buffer.
store(dst + i, elt[i]);
}
}

Expand Down
Loading