Skip to content
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

Jump to bottom

Peel off the last iteration of matmul loop #617

Draft
wants to merge 2 commits into
base: benchmark_gemm
Choose a base branch
from
Draft

Peel off the last iteration of matmul loop #617

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fdf34f3
Peel off the last iteration and remove masked load inside the loop
zhanglx13 Jul 25, 2024
8705aac
Workaround the bug in stream-pipeline
zhanglx13 Jul 27, 2024
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
22 changes: 15 additions & 7 deletions scripts/amd/gemm/matmul_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,16 +41,24 @@ def matmul_kernel(
bias = tl.load(bias_ptrs, mask=offs_am < M, other=0.0)
acc_dtype = tl.float32 if a_ptr.type.element_ty != tl.int8 else tl.int32
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=acc_dtype)
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K * SPLIT_K)):
if EVEN_K:
a = tl.load(a_ptrs)
b = tl.load(b_ptrs)
else:
a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)

max_k = tl.cdiv(K, BLOCK_SIZE_K * SPLIT_K)

for k in range(0, max_k-1):
a = tl.load(tl.multiple_of(a_ptrs, (1, 16)))
b = tl.load(tl.multiple_of(b_ptrs, (16, 1)))
accumulator += tl.dot(a, b)
a_ptrs += BLOCK_SIZE_K * SPLIT_K * stride_ak
b_ptrs += BLOCK_SIZE_K * SPLIT_K * stride_bk

k = max_k - 1
offs_k = k * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
a_ptrsX = a_ptr + offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak
b_ptrsX = b_ptr + offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn
a = tl.load(a_ptrsX, mask=offs_k[None, :] < K, other=0.0)
b = tl.load(b_ptrsX, mask=offs_k[:, None] < K, other=0.0)
accumulator += tl.dot(a, b)

c = accumulator.to(c_ptr.type.element_ty)
if BIAS:
c += bias[:, None]
Expand Down
Loading