Skip to content

add statatic dispatch for hgemm and fix copyright for code_utils.py #13

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
bin913 merged 2 commits into from
Jul 21, 2026
+301 −234
Merged

add statatic dispatch for hgemm and fix copyright for code_utils.py #13

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8e5e79b
use staic_dispatch for hgemm
bin913 Jun 29, 2026
6a8dd0a
refactor code_utils.py to fix copyright
bin913 Jul 20, 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
318 changes: 201 additions & 117 deletions src/flag_blas/runtime/backend/_nvidia/hopper/ops/gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1904,6 +1904,181 @@ def _hgemm_nn_build_kernel(
])


# ---------------------------------------------------------------------------
# Module-level condition predicates for hgemm_tn StaticDispatch
# ---------------------------------------------------------------------------
def _hgemm_tn_is_skinny_aligned(m, n, k, aligned, **_kw):
return aligned and (
(m >= 16384 and max(n, k) <= 2048)
or (n >= 16384 and max(m, k) <= 2048)
)


def _hgemm_tn_is_aligned(m, n, k, aligned, **_kw):
return aligned


def _hgemm_tn_is_default(**_kw):
return True


# ---------------------------------------------------------------------------
# Module-level condition predicates for hgemm_nt StaticDispatch
# ---------------------------------------------------------------------------
def _hgemm_nt_is_aligned(m, n, k, aligned, **_kw):
return aligned


def _hgemm_nt_is_default(**_kw):
return True


# ---------------------------------------------------------------------------
# Module-level condition predicates for hgemm_tt StaticDispatch
# ---------------------------------------------------------------------------
def _hgemm_tt_is_skinny_aligned(m, n, k, aligned, **_kw):
return aligned and m >= 16384 and max(n, k) <= 2048


def _hgemm_tt_is_aligned(m, n, k, aligned, **_kw):
return aligned


def _hgemm_tt_is_default(**_kw):
return True


# ---------------------------------------------------------------------------
# Module-level factory functions for hgemm_tn StaticDispatch
# ---------------------------------------------------------------------------
def _hgemm_tn_build_kernel3(
A, B, C, m, n, k, lda, ldb, ldc, alpha, beta, beta_is_zero,
):
return lambda: _hgemm_tn_kernel3[(
triton.cdiv(m, 128) * triton.cdiv(n, 256),
)](
TensorDescriptor(base=A, shape=[k, m], strides=[lda, 1], block_shape=[64, 128]),
TensorDescriptor(base=B, shape=[k, n], strides=[ldb, 1], block_shape=[64, 256]),
TensorDescriptor(base=C, shape=[m, n], strides=[ldc, 1], block_shape=[128, 256]),
alpha, beta, m, n, k, beta_is_zero,
BLOCK_M=128, BLOCK_N=256, BLOCK_K=64, GROUP_M=8,
num_stages=4, num_warps=8, num_ctas=1,
)


def _hgemm_tn_build_kernel2(
A, B, C, m, n, k, lda, ldb, ldc, alpha, beta, beta_is_zero,
):
grid = lambda meta: (
triton.cdiv(m, meta["BLOCK_M"]) * triton.cdiv(n, meta["BLOCK_N"]),
)
return lambda: _hgemm_tn_kernel2[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero,
)


def _hgemm_tn_build_kernel(
A, B, C, m, n, k, lda, ldb, ldc, alpha, beta, beta_is_zero,
):
grid = lambda meta: (
triton.cdiv(m, meta["BLOCK_M"]) * triton.cdiv(n, meta["BLOCK_N"]),
)
return lambda: _hgemm_tn_kernel[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero,
)


# ---------------------------------------------------------------------------
# Module-level factory functions for hgemm_nt StaticDispatch
# ---------------------------------------------------------------------------
def _hgemm_nt_build_kernel2(
A, B, C, m, n, k, lda, ldb, ldc, alpha, beta, beta_is_zero,
):
grid = lambda meta: (
triton.cdiv(m, meta["BLOCK_M"]) * triton.cdiv(n, meta["BLOCK_N"]),
)
return lambda: _hgemm_nt_kernel2[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero,
)


def _hgemm_nt_build_kernel(
A, B, C, m, n, k, lda, ldb, ldc, alpha, beta, beta_is_zero,
):
grid = lambda meta: (
triton.cdiv(m, meta["BLOCK_M"]) * triton.cdiv(n, meta["BLOCK_N"]),
)
return lambda: _hgemm_nt_kernel[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero,
)


# ---------------------------------------------------------------------------
# Module-level factory functions for hgemm_tt StaticDispatch
# ---------------------------------------------------------------------------
def _hgemm_tt_build_kernel3(
A, B, C, m, n, k, lda, ldb, ldc, alpha, beta, beta_is_zero,
):
return lambda: _hgemm_tt_kernel3[(
triton.cdiv(m, 128) * triton.cdiv(n, 256),
)](
TensorDescriptor(base=A, shape=[k, m], strides=[lda, 1], block_shape=[64, 128]),
TensorDescriptor(base=B, shape=[n, k], strides=[ldb, 1], block_shape=[256, 64]),
TensorDescriptor(base=C, shape=[m, n], strides=[ldc, 1], block_shape=[128, 256]),
alpha, beta, m, n, k, beta_is_zero,
BLOCK_M=128, BLOCK_N=256, BLOCK_K=64, GROUP_M=8,
num_stages=4, num_warps=8, num_ctas=1,
)


def _hgemm_tt_build_kernel2(
A, B, C, m, n, k, lda, ldb, ldc, alpha, beta, beta_is_zero,
):
grid = lambda meta: (
triton.cdiv(m, meta["BLOCK_M"]) * triton.cdiv(n, meta["BLOCK_N"]),
)
return lambda: _hgemm_tt_kernel2[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero,
)


def _hgemm_tt_build_kernel(
A, B, C, m, n, k, lda, ldb, ldc, alpha, beta, beta_is_zero,
):
grid = lambda meta: (
triton.cdiv(m, meta["BLOCK_M"]) * triton.cdiv(n, meta["BLOCK_N"]),
)
return lambda: _hgemm_tt_kernel[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero,
)


_HGEMM_TN_DISPATCH = StaticDispatch([
# skinny + aligned → kernel3 (TensorDescriptor, hardcoded config)
(_hgemm_tn_is_skinny_aligned, _hgemm_tn_build_kernel3),
# aligned → kernel2 (block_ptr)
(_hgemm_tn_is_aligned, _hgemm_tn_build_kernel2),
# default → kernel (original)
(_hgemm_tn_is_default, _hgemm_tn_build_kernel),
])

_HGEMM_NT_DISPATCH = StaticDispatch([
# aligned → kernel2 (block_ptr)
(_hgemm_nt_is_aligned, _hgemm_nt_build_kernel2),
# default → kernel (original)
(_hgemm_nt_is_default, _hgemm_nt_build_kernel),
])

_HGEMM_TT_DISPATCH = StaticDispatch([
# skinny + aligned → kernel3 (TensorDescriptor, hardcoded config)
(_hgemm_tt_is_skinny_aligned, _hgemm_tt_build_kernel3),
# aligned → kernel2 (block_ptr)
(_hgemm_tt_is_aligned, _hgemm_tt_build_kernel2),
# default → kernel (original)
(_hgemm_tt_is_default, _hgemm_tt_build_kernel),
])


def hgemm(
transa: int,
transb: int,
Expand Down Expand Up @@ -1975,126 +2150,35 @@ def hgemm(
)
runner()
elif transa == CUBLAS_OP_T and transb == CUBLAS_OP_N:
is_skinny = (m >= 16384 and max(n, k) <= 2048) or (
n >= 16384 and max(m, k) <= 2048
runner = _HGEMM_TN_DISPATCH.lookup_and_build(
m, n, k, aligned,
context=dict(
A=A, B=B, C=C, m=m, n=n, k=k,
lda=lda, ldb=ldb, ldc=ldc,
alpha=alpha, beta=beta, beta_is_zero=beta_is_zero,
),
)
if aligned and is_skinny:
BLOCK_M = 128
BLOCK_N = 256
BLOCK_K = 64
GROUP_M = 8
NUM_STAGES = 4
NUM_WARPS = 8
NUM_CTAS = 1
desc_a = TensorDescriptor(
base=A,
shape=[k, m],
strides=[lda, 1],
block_shape=[BLOCK_K, BLOCK_M],
)
desc_b = TensorDescriptor(
base=B,
shape=[k, n],
strides=[ldb, 1],
block_shape=[BLOCK_K, BLOCK_N],
)
desc_c = TensorDescriptor(
base=C,
shape=[m, n],
strides=[ldc, 1],
block_shape=[BLOCK_M, BLOCK_N],
)
grid = (triton.cdiv(m, BLOCK_M) * triton.cdiv(n, BLOCK_N),)
_hgemm_tn_kernel3[grid](
desc_a,
desc_b,
desc_c,
alpha,
beta,
m,
n,
k,
beta_is_zero,
BLOCK_M=BLOCK_M,
BLOCK_N=BLOCK_N,
BLOCK_K=BLOCK_K,
GROUP_M=GROUP_M,
num_stages=NUM_STAGES,
num_warps=NUM_WARPS,
num_ctas=NUM_CTAS,
)
elif aligned:
_hgemm_tn_kernel2[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero
)
else:
_hgemm_tn_kernel[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero
)
runner()
elif transa == CUBLAS_OP_N and transb == CUBLAS_OP_T:
if aligned:
_hgemm_nt_kernel2[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero
)
else:
_hgemm_nt_kernel[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero
)
runner = _HGEMM_NT_DISPATCH.lookup_and_build(
m, n, k, aligned,
context=dict(
A=A, B=B, C=C, m=m, n=n, k=k,
lda=lda, ldb=ldb, ldc=ldc,
alpha=alpha, beta=beta, beta_is_zero=beta_is_zero,
),
)
runner()
else:
is_skinny = m >= 16384 and max(n, k) <= 2048
if is_skinny and aligned:
BLOCK_M = 128
BLOCK_N = 256
BLOCK_K = 64
GROUP_M = 8
NUM_STAGES = 4
NUM_WARPS = 8
NUM_CTAS = 1
desc_a = TensorDescriptor(
base=A,
shape=[k, m],
strides=[lda, 1],
block_shape=[BLOCK_K, BLOCK_M],
)
desc_b = TensorDescriptor(
base=B,
shape=[n, k],
strides=[ldb, 1],
block_shape=[BLOCK_N, BLOCK_K],
)
desc_c = TensorDescriptor(
base=C,
shape=[m, n],
strides=[ldc, 1],
block_shape=[BLOCK_M, BLOCK_N],
)
grid = (triton.cdiv(m, BLOCK_M) * triton.cdiv(n, BLOCK_N),)
_hgemm_tt_kernel3[grid](
desc_a,
desc_b,
desc_c,
alpha,
beta,
m,
n,
k,
beta == 0.0,
BLOCK_M=BLOCK_M,
BLOCK_N=BLOCK_N,
BLOCK_K=BLOCK_K,
GROUP_M=GROUP_M,
num_stages=NUM_STAGES,
num_warps=NUM_WARPS,
num_ctas=NUM_CTAS,
)
elif aligned:
_hgemm_tt_kernel2[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero
)
else:
_hgemm_tt_kernel[grid](
A, B, C, alpha, beta, m, n, k, lda, ldb, ldc, beta_is_zero
)
runner = _HGEMM_TT_DISPATCH.lookup_and_build(
m, n, k, aligned,
context=dict(
A=A, B=B, C=C, m=m, n=n, k=k,
lda=lda, ldb=ldb, ldc=ldc,
alpha=alpha, beta=beta, beta_is_zero=beta_is_zero,
),
)
runner()


@libentry()
Expand Down
Loading
Loading