Skip to content

[perf] improve XQA-MLA perf #5468

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
merged 1 commit into from
Jun 26, 2025
+135 −128
Merged

[perf] improve XQA-MLA perf #5468

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
18 changes: 9 additions & 9 deletions cpp/kernels/xqa/barriers.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,7 +68,7 @@ public:
template <Scope scope = defaultScope, ArriveOrder order = ArriveOrder::RELEASE>
__device__ inline mha::conditional_t<scope == Scope::CTA, ArrivalToken, void> arrive(uint32_t update = 1)
{
ArrivalToken token;
ArrivalToken token{};
#if __CUDA_ARCH__ >= 900
if constexpr (scope == Scope::CTA)
{
Expand Down Expand Up @@ -128,9 +128,9 @@ public:

__device__ inline bool isLocal() const
{
uint32_t addrCtaRank;
uint32_t addrCtaRank{};
asm("getctarank.u64 %0, %1;\n" : "=r"(addrCtaRank) : "l"(addr()));
uint32_t ctaRank;
uint32_t ctaRank{};
asm("mov.u32 %0, %%cluster_ctarank;\n" : "=r"(ctaRank));
return addrCtaRank == ctaRank;
}
Expand All @@ -154,7 +154,7 @@ public:
#if __CUDA_ARCH__ >= 900
if constexpr (scope == Scope::CTA)
{
ArrivalToken token;
ArrivalToken token{};
asm volatile("mbarrier.arrive.expect_tx.relaxed.cta.b64 %0, [%1], %2;\n"
: "=l"(token)
: "l"(addr()), "r"(txCount)
Expand All @@ -181,7 +181,7 @@ public:
{
if constexpr (scope == Scope::CTA)
{
ArrivalToken token;
ArrivalToken token{};
switch (order)
{
case ArriveOrder::RELEASE:
Expand Down Expand Up @@ -239,7 +239,7 @@ public:
template <Scope scope = defaultScope>
__device__ inline bool test_wait(ArrivalToken&& token)
{
uint32_t ready;
uint32_t ready{};
if constexpr (scope == Scope::CGA)
{
asm volatile(
Expand Down Expand Up @@ -271,7 +271,7 @@ public:
template <Scope scope = defaultScope>
__device__ inline bool test_wait_parity(bool parity)
{
uint32_t ready;
uint32_t ready{};
if constexpr (scope == Scope::CGA)
{
asm volatile(
Expand Down Expand Up @@ -303,7 +303,7 @@ public:
template <Scope scope = defaultScope>
__device__ inline bool try_wait(ArrivalToken&& token)
{
uint32_t ready;
uint32_t ready{};
if constexpr (scope == Scope::CGA)
{
asm volatile(
Expand Down Expand Up @@ -334,7 +334,7 @@ public:
template <Scope scope = defaultScope>
__device__ inline bool try_wait_parity(bool parity)
{
uint32_t ready;
uint32_t ready{};
if constexpr (scope == Scope::CGA)
{
asm volatile(
Expand Down
4 changes: 2 additions & 2 deletions cpp/kernels/xqa/mha_components.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ template <uint32_t n>
__device__ inline QuadRegRowMaxT<n * warp_size> replicateForQuad(Warp const& warp, Vec<float, n> const& src)
{
assertWarpConverged();
QuadRegRowMaxT<n * warp_size> dst;
QuadRegRowMaxT<n * warp_size> dst{};
#pragma unroll
for (uint32_t i = 0; i < src.size; i++)
{
Expand All @@ -82,7 +82,7 @@ __device__ inline ThrdRegRowMaxT<warp_size * exactDiv(n, 4)> dedupFromQuad(Warp
assert(src[i] == __shfl_sync(~0U, src[i], laneId() / 4 * 4));
}
#endif
ThrdRegRowMaxT<warp_size * exactDiv(n, 4)> dst;
ThrdRegRowMaxT<warp_size * exactDiv(n, 4)> dst{};
uint32_t const lane = laneId();
uint32_t const idxMat = lane / 8;
uint32_t const idxRow = lane % 8;
Expand Down
2 changes: 1 addition & 1 deletion cpp/kernels/xqa/mha_sm90.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1616,7 +1616,7 @@ CUBIN_EXPORT __global__
{
if (warpElectSync())
{
tma::load1DAsync(&smem.tokens[idxBuf], &scratchMem.tokens()[idxChunk],
tma::loadLinearAsync(&smem.tokens[idxBuf], &scratchMem.tokens()[idxChunk],
sizeof(smem.tokens[idxBuf]), bar.produced);
arrive_tx(bar.produced, sizeof(smem.tokens[idxBuf]), 1);
}
Expand Down
Loading