Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
test_tilelang_cache_matmul.py 2.78 KB
Newer Older
Lei Wang's avatar
[Feat] Introduce new caching mechanism for compiled kernels (#176)  
Lei Wang committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124 
# 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()