biondizzle/DeepGEMM
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make various updates and fixes (#198)

Browse Source
This commit is contained in:
Ray Wang
2025-09-25 16:19:07 +08:00
committed by GitHub
parent 79f48ee15a
commit 3f71de7aa9
45 changed files with 3281 additions and 1060 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
tests/test_bf16.py
View File

@@ -7,6 +7,7 @@ from deep_gemm.testing import (
calc_diff, count_bytes
)
from generators import (
get_arch_major,
enumerate_normal, enumerate_m_grouped_contiguous, enumerate_m_grouped_masked, generate_normal,
generate_m_grouped_contiguous, generate_m_grouped_masked
)
@@ -14,14 +15,18 @@ from generators import (
def test_gemm() -> None:
print('Testing GEMM:')
for _, m, n, k, major_a, major_b, accumulate, out_dtype in enumerate_normal(use_bf16=True):
for kernel_type, m, n, k, major_a, major_b, accumulate, out_dtype in enumerate_normal(torch.bfloat16):
# TODO: support accumulation for SM90 BF16 GEMM
if get_arch_major() == 9 and accumulate:
continue
major_opt = 'N' if major_a.is_k_major() else 'T'
major_opt += 'T' if major_b.is_k_major() else 'N'
out_opt = 'FP32' if out_dtype == torch.float else 'BF16'
acc_opt = f'acc={int(accumulate)}'
for test_alias in (False, True):
a, b, c, d, ref_d = generate_normal(m, n, k, major_a, major_b, accumulate, out_dtype, use_bf16=True)
a, b, c, d, ref_d = generate_normal(m, n, k, major_a, major_b, accumulate, out_dtype, kernel_type, use_bf16=True)
func_name = f'bf16_gemm_{major_opt.lower() if test_alias else "nt"}'
if test_alias:
a = a if major_a.is_k_major() else a.T
@@ -31,28 +36,22 @@ def test_gemm() -> None:
diff = calc_diff(d, ref_d)
assert diff < 0.0001, (f'{m=}, {n=}, {k=}, {major_opt=}, {accumulate=}, {out_dtype=}, '
f'{diff:.5f}, alias={test_alias}')
a, b, c, d, ref_d = generate_normal(m, n, k, major_a, major_b, accumulate, out_dtype, use_bf16=True)
a, b, c, d, ref_d = generate_normal(m, n, k, major_a, major_b, accumulate, out_dtype, kernel_type, use_bf16=True)
cublas_t = 0
t = bench_kineto(lambda: deep_gemm.bf16_gemm_nt(a, b, d, c=c), 'bf16_gemm', suppress_kineto_output=True)
if accumulate == 0 and out_dtype == torch.bfloat16:
# noinspection PyBroadException
try:
cublas_t = bench_kineto(lambda: a @ b.T, 'nvjet', suppress_kineto_output=True)
except Exception:
pass
print(f' > Perf (m={m:5}, n={n:5}, k={k:5}, layout={major_opt}, {out_opt}, {acc_opt}): '
f'{t * 1e6:4.0f} us | '
cublas_t, split_k_t = bench_kineto(lambda: deep_gemm.cublaslt_gemm_nt(a, b, d, c=c), ('nvjet', 'reduce'), suppress_kineto_output=True)
print(f' > Perf (m={m:6}, n={n:6}, k={k:6}, layout={major_opt}, {out_opt}, {acc_opt}): '
f'{t * 1e6:5.0f} us | '
f'{2 * m * n * k / t / 1e12:4.0f} TFLOPS | '
f'{(count_bytes(a, b, d) + count_bytes(c) * int(accumulate)) / 1e9 / t:4.0f} GB/s | '
f'{cublas_t / t:.2f}x cuBLAS')
f'{(cublas_t + split_k_t) / t:.2f}x cuBLAS')
print()
def test_m_grouped_gemm_contiguous() -> None:
print('Testing m-grouped contiguous GEMM:')
for _, num_groups, expected_m_per_group, n, k, major_a, major_b in enumerate_m_grouped_contiguous(use_bf16=True):
for _, num_groups, expected_m_per_group, n, k, major_a, major_b in enumerate_m_grouped_contiguous(torch.bfloat16):
major_opt = 'N' if major_a.is_k_major() else 'T'
major_opt += 'T' if major_b.is_k_major() else 'N'
@@ -85,7 +84,7 @@ def test_m_grouped_gemm_masked() -> None:
print('Testing m-grouped masked GEMM:')
# TODO: when the actual `m` is greater than `expected_m_per_group`, efficiency may significantly decrease.
for _, num_groups, max_m, expected_m_per_group, n, k in enumerate_m_grouped_masked():
for _, num_groups, max_m, expected_m_per_group, n, k in enumerate_m_grouped_masked(torch.bfloat16):
# Test correctness
for i in range(10):
a, b, masked_m, d, ref_d = generate_m_grouped_masked(num_groups, max_m, expected_m_per_group, n, k, use_bf16=True)
@@ -111,6 +110,27 @@ def test_m_grouped_gemm_masked() -> None:
print()
def test_cublaslt_gemm() -> None:
print('Testing cuBLASLt GEMM:')
for kernel_type, m, n, k, major_a, major_b, accumulate, out_dtype in enumerate_normal(dtype=torch.bfloat16):
major_opt = 'N' if major_a.is_k_major() else 'T'
major_opt += 'T' if major_b.is_k_major() else 'N'
out_opt = 'FP32' if out_dtype == torch.float else 'BF16'
acc_opt = f'acc={int(accumulate)}'
a, b, c, d, ref_d = generate_normal(m, n, k, major_a, major_b, accumulate, out_dtype, kernel_type, use_bf16=True)
deep_gemm.cublaslt_gemm_nt(a, b, d, c=c)
diff = calc_diff(d, ref_d)
assert diff < 5e-7, f'{diff=}, ({m=}, {n=}, {k=}, {major_opt=}, {accumulate=}, {out_dtype=})'
t = bench_kineto(lambda: deep_gemm.cublaslt_gemm_nt(a, b, d, c=c), 'nvjet', suppress_kineto_output=True,)
print(f' > Perf (m={m:6}, n={n:6}, k={k:6}, layout={major_opt}, {out_opt}, {acc_opt}): '
f'{t * 1e6:5.0f} us | '
f'{2 * m * n * k / t / 1e12:4.0f} TFLOPS | '
f'{(count_bytes(a, b, d) + count_bytes(c) * int(accumulate)) / 1e9 / t:4.0f} GB/s')
print()
if __name__ == '__main__':
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
@@ -121,5 +141,9 @@ if __name__ == '__main__':
print(f' > {deep_gemm.__path__}\n')
test_gemm()
test_m_grouped_gemm_contiguous()
test_m_grouped_gemm_masked()
# TODO: support SM100
if get_arch_major() == 9:
test_m_grouped_gemm_contiguous()
test_m_grouped_gemm_masked()
test_cublaslt_gemm()