Skip to content

Add INT8 GEMM support to the GEMM operator #94

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

Open
albiol2004 wants to merge 3 commits into
base: devel
Choose a base branch
from
Open

Add INT8 GEMM support to the GEMM operator #94

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
24e6204
Add INT8 GEMM support to the GEMM operator
albiol2004 Apr 9, 2026
d3280eb
Fix xclbin filename collision between bf16 and int8 GEMM variants
albiol2004 Apr 9, 2026
a2cb969
Move dtype map to shared np_dtype_map in test_utils
albiol2004 Apr 9, 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
11 changes: 11 additions & 0 deletions iron/common/test_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,17 @@
"i32": torch.int32,
}

# Numpy equivalent of torch_dtype_map, for use in AIERuntimeArgSpec and other
# numpy-based interfaces. Derived from torch_dtype_map to stay in sync.
np_dtype_map = {
k: (
np.dtype(bfloat16)
if v == torch.bfloat16
else torch.tensor([], dtype=v).numpy().dtype
)
for k, v in torch_dtype_map.items()
}

# TODO: Consider upstreaming generic buffer utilities to mlir-aie once operator abstractions stabilize.


Expand Down
9 changes: 7 additions & 2 deletions iron/operators/gemm/design.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,16 +26,19 @@
microkernel_mac_dim_map = {
"npu1": {
"bf16": (4, 8, 4),
"i8": (8, 8, 8),
},
"npu1": {
"bf16": (4, 8, 4),
"i8": (8, 8, 8),
},
"npu2": {
"bf16": {
# emulate_bf16_mmul_with_bfp16
True: (8, 8, 8),
False: (4, 8, 8),
},
"i8": (8, 8, 8),
},
}

Expand Down Expand Up @@ -68,11 +71,13 @@ def main():
default=None,
help="Name of the archive file for the AIE kernels",
)
argparser.add_argument("--dtype_in", type=str, choices=["bf16"], default="bf16")
argparser.add_argument(
"--dtype_in", type=str, choices=["bf16", "i8"], default="bf16"
)
argparser.add_argument(
"--dtype_out",
type=str,
choices=["bf16", "f32"],
choices=["bf16", "f32", "i8", "i16", "i32"],
default="bf16",
)
argparser.add_argument("--trace_size", type=int, default=0)
Expand Down
46 changes: 35 additions & 11 deletions iron/operators/gemm/op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
DesignGenerator,
)
from iron.common.device_utils import get_kernel_dir
from iron.common.test_utils import np_dtype_map
import aie.utils as aie_utils


Expand Down Expand Up @@ -61,7 +62,9 @@ def __post_init__(self):
if self.N % min_N != 0:
raise ValueError(f"N ({self.N}) must be a multiple of {min_N}")

if self.emulate_bf16_mmul_with_bfp16:
if self.dtype_in == "i8":
min_tile_m, min_tile_k, min_tile_n = 16, 8, 16
elif self.emulate_bf16_mmul_with_bfp16:
min_tile_m, min_tile_k, min_tile_n = 8, 8, 8
else:
min_tile_m, min_tile_k, min_tile_n = 4, 8, 8
Expand All @@ -74,9 +77,21 @@ def __post_init__(self):

MLIROperator.__init__(self, context=self.context)

@property
def name(self) -> str:
"""Include dtype in operator name when not the default bf16, to avoid
xclbin filename collisions between bf16 and int8 GEMM variants with
identical dimensions."""
base = super().name
if self.dtype_in != "bf16":
base += f"_{self.dtype_in}_{self.dtype_out}"
Copy link
Copy Markdown
Collaborator

@hunhoffe hunhoffe Apr 9, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it might make sense to always change it to include dtype in/out... thoughts @andrej ?

return base

@property
def _kernel_flags_suffix(self):
"""Suffix encoding compile-time flags that affect the kernel binary."""
if self.dtype_in == "i8":
return f"_{self.dtype_in}_{self.dtype_out}"
return f"_{int(self.prio_accuracy)}_{int(self.emulate_bf16_mmul_with_bfp16)}_{int(self.round_conv_even)}"

def get_mlir_artifact(self):
Expand Down Expand Up @@ -117,14 +132,17 @@ def get_kernel_artifacts(self):
f"-DDIM_K={self.tile_k}",
f"-DDIM_N={self.tile_n}",
]
if self.prio_accuracy:
kernel_flags.append("-Dbf16_f32_ONLY")
if self.dtype_in == "i8":
kernel_flags.append(f"-D{self.dtype_in}_{self.dtype_out}_ONLY")
else:
kernel_flags.append("-Dbf16_bf16_ONLY")
if self.round_conv_even:
kernel_flags.append("-DROUND_CONV_EVEN")
if self.emulate_bf16_mmul_with_bfp16:
kernel_flags.append("-DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16")
if self.prio_accuracy:
kernel_flags.append("-Dbf16_f32_ONLY")
else:
kernel_flags.append("-Dbf16_bf16_ONLY")
if self.round_conv_even:
kernel_flags.append("-DROUND_CONV_EVEN")
if self.emulate_bf16_mmul_with_bfp16:
kernel_flags.append("-DAIE_API_EMULATE_BFLOAT16_MMUL_WITH_BFP16")
if self.b_col_maj:
kernel_flags.append("-DB_COL_MAJ")
if self.c_col_maj:
Expand All @@ -150,13 +168,19 @@ def get_kernel_artifacts(self):
]

def get_arg_spec(self):
dtype_in_np = np_dtype_map[self.dtype_in]
dtype_out_np = np_dtype_map[self.dtype_out]
return [
AIERuntimeArgSpec("in", (self.M, self.K)), # input A
AIERuntimeArgSpec("in", (self.M, self.K), dtype=dtype_in_np), # input A
AIERuntimeArgSpec(
"in", (self.K, self.N) if not self.b_col_maj else (self.N, self.K)
"in",
(self.K, self.N) if not self.b_col_maj else (self.N, self.K),
dtype=dtype_in_np,
), # input B (weights)
AIERuntimeArgSpec(
"out", (self.M, self.N) if not self.c_col_maj else (self.N, self.M)
"out",
(self.M, self.N) if not self.c_col_maj else (self.N, self.M),
dtype=dtype_out_np,
), # output C
]

Expand Down
19 changes: 16 additions & 3 deletions iron/operators/gemm/reference.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@ def generate_golden_reference(
K: int,
N: int,
dtype="bf16",
dtype_out=None,
seed=42,
b_col_maj=False,
c_col_maj=False,
Expand All @@ -18,9 +19,21 @@ def generate_golden_reference(
torch.manual_seed(seed)
val_range = 4
dtype_torch = torch_dtype_map[dtype]
input_a = torch.randn(M, K, dtype=dtype_torch) * val_range
input_b_full = torch.rand(K, N, dtype=dtype_torch) * val_range
output_full = torch.matmul(input_a, input_b_full)
if dtype in ("i8",):
input_a = torch.randint(-val_range, val_range + 1, (M, K), dtype=dtype_torch)
input_b_full = torch.randint(
-val_range, val_range + 1, (K, N), dtype=dtype_torch
)
output_full = torch.matmul(
input_a.to(torch.int32), input_b_full.to(torch.int32)
)
if dtype_out is not None and dtype_out != "i32":
dtype_out_torch = torch_dtype_map[dtype_out]
output_full = output_full.to(dtype_out_torch)
else:
input_a = torch.randn(M, K, dtype=dtype_torch) * val_range
input_b_full = torch.rand(K, N, dtype=dtype_torch) * val_range
output_full = torch.matmul(input_a, input_b_full)
if False:
# The following inputs are useful for debugging;
# the A matrix becomes a matrix where each element encodes its row and column index,
Expand Down
79 changes: 50 additions & 29 deletions iron/operators/gemm/test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,34 +21,41 @@ def get_params():
max_aie_columns = dev.cols
device_type = dev.resolve().name
# fmt: off
# M, K, N, num_aie_columns, b_col_maj, c_col_maj, m, k, n, trace_size, partition_N
# M, K, N, num_aie_columns, b_col_maj, c_col_maj, m, k, n, trace_size, partition_N, dtype_in, dtype_out
regular_params = [
(2048, 2048, 2048, 1, False, False, 64, 64, 64, 0, 1),
(2048, 2048, 2048, 2, True, False, 64, 64, 64, 0, 1),
(2048, 2048, 2048, 8, True, True, 64, 64, 64, 0, 1),
( 384, 1536, 1792, 4, True, False, 32, 48, 64, 0, 1),
(1792, 896, 1152, 8, False, True, 64, 32, 48, 0, 1),
( 896, 1792, 640, 8, False, True, 32, 64, 80, 0, 1),
( 192, 384, 64, 4, False, False, 48, 96, 16, 0, 1),
( 192, 384, 64, 4, True, True, 48, 96, 16, 0, 1),
( 64, 512, 256, 4, True, False, 16, 64, 64, 0, 4),
(2048, 2048, 2048, 1, False, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 2048, 2048, 2, True, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 2048, 2048, 8, True, True, 64, 64, 64, 0, 1, "bf16", "bf16"),
( 384, 1536, 1792, 4, True, False, 32, 48, 64, 0, 1, "bf16", "bf16"),
(1792, 896, 1152, 8, False, True, 64, 32, 48, 0, 1, "bf16", "bf16"),
( 896, 1792, 640, 8, False, True, 32, 64, 80, 0, 1, "bf16", "bf16"),
( 192, 384, 64, 4, False, False, 48, 96, 16, 0, 1, "bf16", "bf16"),
( 192, 384, 64, 4, True, True, 48, 96, 16, 0, 1, "bf16", "bf16"),
( 64, 512, 256, 4, True, False, 16, 64, 64, 0, 4, "bf16", "bf16"),
]
int8_params = [
(2048, 2048, 2048, 4, False, False, 64, 64, 64, 0, 1, "i8", "i32"),
( 512, 512, 512, 4, False, False, 32, 32, 32, 0, 1, "i8", "i32"),
(2048, 2048, 2048, 8, False, False, 64, 64, 64, 0, 1, "i8", "i32"),
(1024, 1024, 1024, 8, False, False, 64, 64, 64, 0, 1, "i8", "i8"),
(1024, 1024, 1024, 8, True, False, 64, 64, 64, 0, 1, "i8", "i16"),
]
extensive_params = [
(2048, 2048, 2048, 8, False, False, 32, 32, 128, 0, 1),
(2048, 2048, 8192, 2, False, False, 64, 64, 64, 0, 1),
(2048, 8192, 2048, 2, False, False, 64, 64, 64, 0, 1),
(2048, 64, 2048, 2, False, False, 64, 64, 64, 0, 1),
(2048, 64, 8192, 2, False, False, 64, 64, 64, 0, 1),
(2048, 2048, 2048, 8, True, False, 128, 32, 32, 0, 1),
(2048, 2048, 8192, 2, True, False, 64, 64, 64, 0, 1),
(2048, 8192, 2048, 2, True, False, 64, 64, 64, 0, 1),
(2048, 64, 2048, 2, True, False, 64, 64, 64, 0, 1),
(2048, 64, 8192, 2, True, False, 64, 64, 64, 0, 1),
(2048, 2048, 2048, 2, False, True, 8, 16, 32, 0, 1),
(2048, 2048, 8192, 2, False, True, 64, 64, 64, 0, 1),
(2048, 8192, 2048, 2, False, True, 64, 64, 64, 0, 1),
(2048, 64, 2048, 2, False, True, 64, 64, 64, 0, 1),
(2048, 64, 8192, 2, False, True, 64, 64, 64, 0, 1),
(2048, 2048, 2048, 8, False, False, 32, 32, 128, 0, 1, "bf16", "bf16"),
(2048, 2048, 8192, 2, False, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 8192, 2048, 2, False, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 64, 2048, 2, False, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 64, 8192, 2, False, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 2048, 2048, 8, True, False, 128, 32, 32, 0, 1, "bf16", "bf16"),
(2048, 2048, 8192, 2, True, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 8192, 2048, 2, True, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 64, 2048, 2, True, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 64, 8192, 2, True, False, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 2048, 2048, 2, False, True, 8, 16, 32, 0, 1, "bf16", "bf16"),
(2048, 2048, 8192, 2, False, True, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 8192, 2048, 2, False, True, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 64, 2048, 2, False, True, 64, 64, 64, 0, 1, "bf16", "bf16"),
(2048, 64, 8192, 2, False, True, 64, 64, 64, 0, 1, "bf16", "bf16"),
]
# fmt: on

Expand All @@ -69,6 +76,8 @@ def add_params(param_list, is_extensive):
n,
trace_size,
partition_N,
dtype_in,
dtype_out,
) = p

# Skip tests that require more columns than available on the device
Expand All @@ -84,6 +93,7 @@ def add_params(param_list, is_extensive):
params.append(pytest.param(*p, marks=marks))

add_params(regular_params, is_extensive=False)
add_params(int8_params, is_extensive=False)
add_params(extensive_params, is_extensive=True)

return params
Expand All @@ -95,7 +105,7 @@ def add_params(param_list, is_extensive):
Throughput=r"Throughput: (?P<value>[\d\.e\+-]+) GFLOP/s",
)
@pytest.mark.parametrize(
"M,K,N,num_aie_columns,b_col_maj,c_col_maj,m,k,n,trace_size,partition_N",
"M,K,N,num_aie_columns,b_col_maj,c_col_maj,m,k,n,trace_size,partition_N,dtype_in,dtype_out",
get_params(),
)
def test_gemm(
Expand All @@ -110,6 +120,8 @@ def test_gemm(
n,
trace_size,
partition_N,
dtype_in,
dtype_out,
aie_context,
):
total_N = N * partition_N
Expand All @@ -118,6 +130,8 @@ def test_gemm(
M=M,
K=K,
N=total_N,
dtype=dtype_in,
dtype_out=dtype_out,
b_col_maj=b_col_maj,
c_col_maj=c_col_maj,
)
Expand All @@ -130,13 +144,20 @@ def test_gemm(
tile_k=k,
tile_n=n,
num_aie_columns=num_aie_columns,
prio_accuracy=True,
prio_accuracy=dtype_in == "bf16",
emulate_bf16_mmul_with_bfp16=False,
b_col_maj=b_col_maj,
c_col_maj=c_col_maj,
dtype_in=dtype_in,
dtype_out=dtype_out,
context=aie_context,
)

if dtype_in == "i8":
rel_tol, abs_tol = 1e-10, 1e-10
else:
rel_tol, abs_tol = 0.005, 0.005

if partition_N == 1:
input_buffers = {
"A": golden_ref["input"].flatten(),
Expand All @@ -146,7 +167,7 @@ def test_gemm(
"C": golden_ref["output"][0].flatten(),
}
errors, latency_us, bandwidth_gbps = run_test(
operator, input_buffers, output_buffers, rel_tol=0.005, abs_tol=0.005
operator, input_buffers, output_buffers, rel_tol=rel_tol, abs_tol=abs_tol
)
else:
compilable = operator.compile()
Expand Down Expand Up @@ -200,7 +221,7 @@ def test_gemm(
# Compare concatenated output to full reference
C_expected = golden_ref["output"][0]
buf_errors = verify_buffer(
C_concat, "C", C_expected, rel_tol=0.005, abs_tol=0.005
C_concat, "C", C_expected, rel_tol=rel_tol, abs_tol=abs_tol
)
errors = {"C": buf_errors} if buf_errors else {}

Expand Down
Loading