gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[TUTORIALS] adjust heuristics for dwdb kernel (#565)

Browse Source
This commit is contained in:
Natalia Gimelshein
2022-06-29 17:00:22 -07:00
committed by GitHub
parent 1895ceaa2d
commit 1bbb2430d9
1 changed files with 26 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
python/tutorials/05-layer-norm.py
View File

@@ -128,8 +128,10 @@ def _layer_norm_bwd_dwdb(
cols = pid * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
dw = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
db = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for i in range(0, M, BLOCK_SIZE_M):
rows = i + tl.arange(0, BLOCK_SIZE_M)
UNROLL: tl.constexpr = 4
for i in range(0, M, BLOCK_SIZE_M * UNROLL):
for j in range(UNROLL):
rows = i + j * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
mask = (rows[:, None] < M) & (cols[None, :] < N)
offs = rows[:, None] * N + cols[None, :]
a = tl.load(A + offs, mask=mask, other=0.).to(tl.float32)
@@ -211,7 +213,15 @@ class LayerNorm(torch.autograd.Function):
BLOCK_SIZE_N=ctx.BLOCK_SIZE,
num_warps=ctx.num_warps,
)
# accumulate partial sums in separate kernel
if N > 10240:
BLOCK_SIZE_N = 128
BLOCK_SIZE_M = 32
num_warps = 4
else:
# maximize occupancy for small N
BLOCK_SIZE_N = 16
BLOCK_SIZE_M = 16
num_warps = 8
grid = lambda meta: [triton.cdiv(N, meta["BLOCK_SIZE_N"])]
_layer_norm_bwd_dwdb[grid](
a, dout,
@@ -220,17 +230,11 @@ class LayerNorm(torch.autograd.Function):
dbias,
M,
N,
BLOCK_SIZE_M=32,
BLOCK_SIZE_N=128,
BLOCK_SIZE_M=BLOCK_SIZE_M,
BLOCK_SIZE_N=BLOCK_SIZE_N,
num_warps=num_warps
)
return (da, None, dweight, dbias, None, None,
None, None, None, None,
None,
None, None, None,
None,
None, None, None,
None, None, None,
None, None, None)
return (da, None, dweight, dbias, None)
def layer_norm(a, normalized_shape, weight, bias, eps):