Skip to content

Workaround for matmul kernel crash with i8xf32 operands. #12

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.

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
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
12 changes: 10 additions & 2 deletions lib/Dialect/TritonGPU/Transforms/AccelerateMatmul.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -310,8 +310,16 @@ class BlockedToMMA : public mlir::OpRewritePattern<DotOp> {
dotOp.getMaxNumImpreciseAcc(), false);
} else {
// convert operands
int minBitwidth =
std::min(computeOrigBitWidth(a), computeOrigBitWidth(b));
int bitwidthA = computeOrigBitWidth(a);
int bitwidthB = computeOrigBitWidth(b);
int minBitwidth = std::min(bitwidthA, bitwidthB);
int maxBitwidth = std::max(bitwidthA, bitwidthB);
if (minBitwidth == 8 && maxBitwidth == 32) {
// workaround for i8xf32 matmul, issue #2853
// f32 x kWidth=4 is not supported in triton nvidiagpu to nvvm lowering
// use kWidth = 2
minBitwidth = 16;
}
Type minType = rewriter.getIntegerType(minBitwidth);
// convert A operand
auto newAEncoding = DotOperandEncodingAttr::get(
Expand Down
78 changes: 78 additions & 0 deletions python/test/unit/language/test_core.py
Original file line number Diff line number Diff line change
Expand Up @@ -5263,6 +5263,84 @@ def test_dot_max_num_imprecise_acc(M, N, K, BLOCK_M, BLOCK_N, BLOCK_K, in_type_s
assert h.asm["ptx"].count("add.f32") == (BLOCK_M * BLOCK_N) // (32 * num_warps) * (BLOCK_K // low_precision_acc)


# -----------------------------
# test i8x*fp32 fp32*xi8 dot
# -----------------------------


@triton.jit
def mixed_matmul_kernel(lhs_ptr, # (M, K)
rhs_ptr, # (K, N)
out_ptr, # (M, N)
# shape information (strides)
M, N, K,
# block information
block_m: tl.constexpr, block_n: tl.constexpr, block_k: tl.constexpr, lhs_mixed: tl.constexpr,
dtype: tl.constexpr):

start_m = tl.program_id(0) # start (axis m)
start_n = tl.program_id(1) # start (axis n)

acc = tl.zeros([block_m, block_n], dtype=tl.float32)

for start_k in range(0, K, block_k):
lhs_tile = (start_m * block_m + tl.arange(0, block_m))[:, None] * K + (start_k + tl.arange(0, block_k))[None, :]

rhs_tile = (start_k + tl.arange(0, block_k))[:, None] * N + (start_n * block_n + tl.arange(0, block_n))[None, :]

lhs_mask = ((start_m * block_m + tl.arange(0, block_m)) < M)[:, None] * (
(start_k + tl.arange(0, block_k)) < K)[None, :]
rhs_mask = ((start_n * block_n + tl.arange(0, block_n)) < N)[None, :] * (
(start_k + tl.arange(0, block_k)) < K)[:, None]

lhs = tl.load(lhs_ptr + lhs_tile, mask=lhs_mask, other=0.0)
rhs = tl.load(rhs_ptr + rhs_tile, mask=rhs_mask, other=0.0)

if lhs_mixed:
lhs = lhs.to(dtype)
else:
rhs = rhs.to(dtype)
acc += tl.dot(lhs, rhs)

out_tile = ((start_m * block_m + tl.arange(0, block_m))[:, None] * N + start_n * block_n +
tl.arange(0, block_n)[None, :])

mask = ((start_m * block_m + tl.arange(0, block_m)) < M)[:, None] * (
(start_n * block_n + tl.arange(0, block_n)) < N)[None, :]

tl.store(out_ptr + out_tile, acc, mask=mask)


@pytest.mark.interpreter
@pytest.mark.parametrize("M, N, K", [(128, 64, 32)])
@pytest.mark.parametrize("BLOCK_M, BLOCK_N, BLOCK_K", [(16, 16, 16)])
@pytest.mark.parametrize("is_fp32", [True, False])
@pytest.mark.parametrize("lhs_mixed", [True, False])
def test_mixed_dot(M, N, K, BLOCK_M, BLOCK_N, BLOCK_K, is_fp32, lhs_mixed, device):
if is_cuda():
cc = torch.cuda.get_device_capability()
if cc[0] < 8:
pytest.skip("Need at least sm80")

torch_dtype = torch.float32 if is_fp32 else torch.float16
tl_dtype = tl.float32 if is_fp32 else tl.float16

if lhs_mixed:
lhs = torch.randint(0, 127, (M, K), dtype=torch.int8, device=device)
rhs = torch.randn((K, N), dtype=torch.float16, device=device).to(dtype=torch_dtype)
else:
lhs = torch.randn((M, K), dtype=torch.float16, device=device).to(dtype=torch_dtype)
rhs = torch.randint(0, 127, (K, N), dtype=torch.int8, device=device)
out = torch.empty((M, N), dtype=torch.float32, device=device)

mixed_matmul_kernel[(triton.cdiv(M, BLOCK_M), triton.cdiv(N,
BLOCK_N))](lhs, rhs, out, M=M, N=N, K=K, block_m=BLOCK_M,
block_n=BLOCK_N, block_k=BLOCK_K,
lhs_mixed=lhs_mixed, dtype=tl_dtype)
ref = torch.mm(lhs.to(dtype=torch.float32), rhs.to(dtype=torch.float32))
torch.testing.assert_close(ref, out, rtol=1e-3, atol=1e-3)


# -----------------------
# test enable_fp_fusion
# -----------------------
Expand Down