mirror of
https://github.com/tile-ai/tilelang.git
synced 2026-10-22 06:50:25 +08:00
Lei Wang
7bde63d526
* Add kernel caching mechanism to TileLang - Implement a new `cached` function in `tilelang/cache/__init__.py` to cache and reuse compiled kernels - Expose the `cached` function in the main `tilelang/__init__.py` - Add a test case for cached matrix multiplication in `testing/python/cache/test_tilelang_cache_matmul.py` - Provide a `clear_cache()` function to reset the kernel cache when needed * Refactor kernel caching test and implementation - Simplify the `cached` function in `tilelang/cache/__init__.py` - Update test script `test_tilelang_cache_matmul.py` to use `tilelang.testing.main()` - Remove unnecessary whitespace and improve code formatting * Update import for `cached` function in MHA examples - Modify import statement in `example_mha_bwd.py` and `test_tilelang_kernel_mha_bwd.py` - Change import from `tilelang.profiler import cached` to `tilelang import cached` - Align with recent refactoring of kernel caching mechanism * Refactor `cached` function signature in kernel caching - Update function signature to use keyword-only arguments for `target` and `target_host` - Improve parameter order and readability of the `cached` decorator - Maintain existing functionality while enhancing function definition
125 lines
2.8 KiB
Python
125 lines
2.8 KiB
Python
# Copyright (c) Microsoft Corporation.
|
|
# Licensed under the MIT License.
|
|
|
|
from tilelang import tvm as tvm
|
|
import tilelang.testing
|
|
from tilelang import cached
|
|
|
|
|
|
def matmul(
|
|
M,
|
|
N,
|
|
K,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
accum_dtype,
|
|
num_stages,
|
|
threads,
|
|
):
|
|
A_shape = (K, M) if trans_A else (M, K)
|
|
B_shape = (N, K) if trans_B else (K, N)
|
|
A_shared_shape = (block_K, block_M) if trans_A else (block_M, block_K)
|
|
B_shared_shape = (block_N, block_K) if trans_B else (block_K, block_N)
|
|
|
|
import tilelang.language as T
|
|
|
|
@T.prim_func
|
|
def main(
|
|
A: T.Buffer(A_shape, in_dtype),
|
|
B: T.Buffer(B_shape, in_dtype),
|
|
C: T.Buffer((M, N), out_dtype),
|
|
):
|
|
with T.Kernel(T.ceildiv(N, block_N), T.ceildiv(M, block_M), threads=threads) as (bx, by):
|
|
A_shared = T.alloc_shared(A_shared_shape, in_dtype)
|
|
B_shared = T.alloc_shared(B_shared_shape, in_dtype)
|
|
C_local = T.alloc_fragment((block_M, block_N), accum_dtype)
|
|
T.clear(C_local)
|
|
for k in T.Pipelined(T.ceildiv(K, block_K), num_stages=num_stages):
|
|
if trans_A:
|
|
T.copy(A[k * block_K, by * block_M], A_shared)
|
|
else:
|
|
T.copy(A[by * block_M, k * block_K], A_shared)
|
|
if trans_B:
|
|
T.copy(B[bx * block_N, k * block_K], B_shared)
|
|
else:
|
|
T.copy(B[k * block_K, bx * block_N], B_shared)
|
|
T.gemm(A_shared, B_shared, C_local, trans_A, trans_B)
|
|
T.copy(C_local, C[by * block_M, bx * block_N])
|
|
|
|
return main
|
|
|
|
|
|
def run_cache_matmul(
|
|
M,
|
|
N,
|
|
K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
dtypeAccum,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
num_stages=3,
|
|
num_threads=128,
|
|
):
|
|
program = matmul(
|
|
M,
|
|
N,
|
|
K,
|
|
block_M,
|
|
block_N,
|
|
block_K,
|
|
trans_A,
|
|
trans_B,
|
|
in_dtype,
|
|
out_dtype,
|
|
dtypeAccum,
|
|
num_stages,
|
|
num_threads,
|
|
)
|
|
|
|
kernel = cached(program, [2])
|
|
|
|
profiler = kernel.get_profiler()
|
|
|
|
def ref_program(A, B):
|
|
import torch
|
|
|
|
if trans_A:
|
|
A = A.T
|
|
if trans_B:
|
|
B = B.T
|
|
C = torch.matmul(A.to(torch.float), B.to(torch.float))
|
|
C = C.to(torch.__getattribute__(out_dtype))
|
|
return C
|
|
|
|
profiler.assert_allclose(ref_program, atol=1e-2, rtol=1e-2)
|
|
|
|
|
|
def test_cache_matmul_f16f16f16_nn():
|
|
run_cache_matmul(
|
|
512,
|
|
1024,
|
|
768,
|
|
False,
|
|
False,
|
|
"float16",
|
|
"float16",
|
|
"float16",
|
|
128,
|
|
256,
|
|
32,
|
|
2,
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
tilelang.testing.main()
|