Skip to content

[TritonCPU] Prepare L1 matmul and elementwise kernels for benchmarking #78

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
sandlbn merged 4 commits into from
Jun 1, 2026
Merged

[TritonCPU] Prepare L1 matmul and elementwise kernels for benchmarking #78

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
86a937a
Prep L1 matmuls for benchmarking
jopperm May 31, 2026
ab9dceb
Prep L1 elementwise kernels for benchmarking
jopperm May 31, 2026
851c4f9
Fix dtype
jopperm May 31, 2026
432d23c
Add mem_bytes tags, more blocksizes for eltwise, don't use persistent…
jopperm Jun 1, 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
24 changes: 16 additions & 8 deletions backends/triton/cpu/KernelBench/level1/10_3D_tensor_matrix_multiplication.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,15 +10,15 @@
import triton.language as tl


def _configs():
return [
@triton.autotune(
configs=[
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": 4},
),
]


@triton.autotune(configs=_configs(), key=["M", "N", "K"])
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": gs},
)
for gs in [1, 2, 4, 8]
],
key=["M", "N", "K"],
)
@triton.jit
def _matmul_kernel(
a_ptr,
Expand Down Expand Up @@ -101,6 +101,13 @@ def forward(self, A, B):
c_flat = torch.empty((total_m, l), device=a.device, dtype=torch.bfloat16)

def grid(META):
assert (
m % META["BLOCK_M"] == 0
and l % META["BLOCK_N"] == 0
and k % META["BLOCK_K"] == 0
), (
"M, L, and K must be divisible by BLOCK_M, BLOCK_N, and BLOCK_K respectively"
)
return (
triton.cdiv(total_m, META["BLOCK_M"]) * triton.cdiv(l, META["BLOCK_N"]),
)
Expand All @@ -118,6 +125,7 @@ def grid(META):
b.stride(1),
c_flat.stride(0),
c_flat.stride(1),
assume_in_bounds=True,
)

return c_flat.reshape(batch, m, l)
Expand Down
28 changes: 16 additions & 12 deletions backends/triton/cpu/KernelBench/level1/11_4D_tensor_matrix_multiplication.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,16 +29,13 @@ def swizzle_tile(
return pid_m, pid_n


def get_autotune_configs():
return [
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": 4},
),
]


@triton.autotune(
configs=get_autotune_configs(),
configs=[
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": gs},
)
for gs in [1, 2, 4, 8]
],
key=["M", "N", "K"],
)
@triton.jit
Expand Down Expand Up @@ -111,9 +108,15 @@ def forward(self, A, B):

C_2d = torch.empty((M, N), device=A.device, dtype=torch.bfloat16)

grid = lambda META: (
triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),
)
def grid(META):
assert (
M % META["BLOCK_M"] == 0
and N % META["BLOCK_N"] == 0
and K % META["BLOCK_K"] == 0
), (
"M, N, and K must be divisible by BLOCK_M, BLOCK_N, and BLOCK_K respectively"
)
return (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),)

_gemm_kernel[grid](
A_flat,
Expand All @@ -128,6 +131,7 @@ def forward(self, A, B):
B_fp16.stride(1),
C_2d.stride(0),
C_2d.stride(1),
assume_in_bounds=True,
)

result = C_2d.view(b_dim, i_dim, j_dim, k_dim)
Expand Down
3 changes: 3 additions & 0 deletions backends/triton/cpu/KernelBench/level1/12_Matmul_with_diagonal_matrices_.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,9 @@
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32},
),
triton.Config(
{"BLOCK_M": 64, "BLOCK_N": 64},
),
],
key=["N", "M"],
)
Expand Down
30 changes: 19 additions & 11 deletions backends/triton/cpu/KernelBench/level1/13_Matmul_for_symmetric_matrices.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,15 +29,15 @@ def swizzle_tile(
return pid_m, pid_n


def _configs():
return [
@triton.autotune(
configs=[
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": 4},
),
]


@triton.autotune(configs=_configs(), key=["M", "N", "K"])
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": gs},
)
for gs in [1, 2, 4, 8]
],
key=["M", "N", "K"],
)
@triton.jit
def _matmul_kernel(
a_ptr,
Expand Down Expand Up @@ -111,9 +111,16 @@ def forward(self, A, B):
N = B.shape[1]
C = torch.empty((M, N), device=A.device, dtype=torch.bfloat16)

grid = lambda META: (
triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),
)
def grid(META):
assert (
M % META["BLOCK_M"] == 0
and N % META["BLOCK_N"] == 0
and K % META["BLOCK_K"] == 0
), (
"M, N, and K must be divisible by BLOCK_M, BLOCK_N, and BLOCK_K respectively"
)
return (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),)

_matmul_kernel[grid](
A,
B,
Expand All @@ -130,6 +137,7 @@ def forward(self, A, B):
DIVISIBLE_M=(M % 256 == 0),
DIVISIBLE_N=(N % 128 == 0),
DIVISIBLE_K=(K % 32 == 0),
assume_in_bounds=True,
)
return C

Expand Down
18 changes: 13 additions & 5 deletions backends/triton/cpu/KernelBench/level1/14_Matmul_for_upper_triangular_matrices.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,8 +34,9 @@ def swizzle_tile(
@triton.autotune(
configs=[
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": 4},
),
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": gs},
)
for gs in [1, 2, 4, 8]
],
key=["M", "N", "K"],
)
Expand Down Expand Up @@ -117,9 +118,15 @@ def forward(self, A, B):
N = B.shape[1]
C = torch.zeros((M, N), device=A.device, dtype=A.dtype)

grid = lambda META: (
triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),
)
def grid(META):
assert (
M % META["BLOCK_M"] == 0
and N % META["BLOCK_N"] == 0
and K % META["BLOCK_K"] == 0
), (
"M, N, and K must be divisible by BLOCK_M, BLOCK_N, and BLOCK_K respectively"
)
return (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),)

_triu_matmul_kernel[grid](
A,
Expand All @@ -134,6 +141,7 @@ def forward(self, A, B):
B.stride(1),
C.stride(0),
C.stride(1),
assume_in_bounds=True,
)
return C

Expand Down
19 changes: 13 additions & 6 deletions backends/triton/cpu/KernelBench/level1/15_Matmul_for_lower_triangular_matrices.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,10 +32,11 @@ def swizzle_tile(
@triton.autotune(
configs=[
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": 4},
),
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": gs},
)
for gs in [1, 2, 4, 8]
],
key=["M"],
key=["M", "N", "K"],
)
@triton.jit
def tril_matmul_kernel(
Expand Down Expand Up @@ -111,9 +112,14 @@ def forward(self, A, B):
M = A.shape[0]
C = torch.zeros(M, M, device=A.device, dtype=A.dtype)

grid = lambda META: (
triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(M, META["BLOCK_N"]),
)
def grid(META):
assert (
M % META["BLOCK_M"] == 0
and M % META["BLOCK_N"] == 0
and M % META["BLOCK_K"] == 0
), "M must be divisible by BLOCK_M, BLOCK_N, and BLOCK_K respectively"
return (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(M, META["BLOCK_N"]),)

tril_matmul_kernel[grid](
A,
B,
Expand All @@ -125,6 +131,7 @@ def forward(self, A, B):
B.stride(1),
C.stride(0),
C.stride(1),
assume_in_bounds=True,
)
return C

Expand Down
23 changes: 14 additions & 9 deletions backends/triton/cpu/KernelBench/level1/16_Matmul_with_transposed_A.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,16 +10,13 @@
import triton.language as tl


def get_autotune_configs():
return [
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": 4},
),
]


@triton.autotune(
configs=get_autotune_configs(),
configs=[
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": gs},
)
for gs in [1, 2, 4, 8]
],
key=["M", "N", "K"],
)
@triton.jit
Expand Down Expand Up @@ -96,6 +93,13 @@ def forward(self, A: torch.Tensor, B: torch.Tensor) -> torch.Tensor:
C = torch.empty((M, N), device=A.device, dtype=A.dtype)

def grid(META):
assert (
M % META["BLOCK_M"] == 0
and N % META["BLOCK_N"] == 0
and K % META["BLOCK_K"] == 0
), (
"M, N, and K must be divisible by BLOCK_M, BLOCK_N, and BLOCK_K respectively"
)
return (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),)

_matmul_at_kernel[grid](
Expand All @@ -111,6 +115,7 @@ def grid(META):
B.stride(1),
C.stride(0),
C.stride(1),
assume_in_bounds=True,
)
return C

Expand Down
23 changes: 14 additions & 9 deletions backends/triton/cpu/KernelBench/level1/17_Matmul_with_transposed_B.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,16 +10,13 @@
import triton.language as tl


def get_autotune_configs():
return [
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": 4},
),
]


@triton.autotune(
configs=get_autotune_configs(),
configs=[
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": gs},
)
for gs in [1, 2, 4, 8]
],
key=["M", "N", "K"],
)
@triton.jit
Expand Down Expand Up @@ -100,6 +97,13 @@ def forward(self, A: torch.Tensor, B: torch.Tensor) -> torch.Tensor:
C = torch.empty((M, N), device=A.device, dtype=A.dtype)

def grid(META):
assert (
M % META["BLOCK_M"] == 0
and N % META["BLOCK_N"] == 0
and K % META["BLOCK_K"] == 0
), (
"M, N, and K must be divisible by BLOCK_M, BLOCK_N, and BLOCK_K respectively"
)
return (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),)

_matmul_bt_kernel[grid](
Expand All @@ -115,6 +119,7 @@ def grid(META):
B.stride(1),
C.stride(0),
C.stride(1),
assume_in_bounds=True,
)

return C
Expand Down
24 changes: 16 additions & 8 deletions backends/triton/cpu/KernelBench/level1/18_Matmul_with_transposed_both.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,15 +10,15 @@
import triton.language as tl


def _configs():
return [
@triton.autotune(
configs=[
triton.Config(
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": 4},
),
]


@triton.autotune(configs=_configs(), key=["M", "N", "K"])
{"BLOCK_M": 32, "BLOCK_N": 32, "BLOCK_K": 32, "GROUP_SIZE_M": gs},
)
for gs in [1, 2, 4, 8]
],
key=["M", "N", "K"],
)
@triton.jit
def _matmul_tt_kernel(
A_ptr,
Expand Down Expand Up @@ -96,6 +96,13 @@ def forward(self, A: torch.Tensor, B: torch.Tensor) -> torch.Tensor:
C = torch.empty((M, N), device=A.device, dtype=A.dtype)

def grid(META):
assert (
M % META["BLOCK_M"] == 0
and N % META["BLOCK_N"] == 0
and K % META["BLOCK_K"] == 0
), (
"M, N, and K must be divisible by BLOCK_M, BLOCK_N, and BLOCK_K respectively"
)
return (triton.cdiv(M, META["BLOCK_M"]) * triton.cdiv(N, META["BLOCK_N"]),)

_matmul_tt_kernel[grid](
Expand All @@ -111,6 +118,7 @@ def grid(META):
B.stride(1),
C.stride(0),
C.stride(1),
assume_in_bounds=True,
)

return C
Expand Down
Loading
Loading