MCPcopy Create free account
hub / github.com/tile-ai/tilelang / main

Function main

testing/python/kernel/test_tilelang_kernel_gemv_simt.py:51–110  ·  view source on GitHub ↗
(
        A: T.Tensor(A_shape, in_dtype),
        B: T.Tensor(B_shape, in_dtype),
        Bias: T.Tensor(Bias_shape, out_dtype),
        C: T.Tensor(C_shape, out_dtype),
    )

Source from the content-addressed store, hash-verified

49
50 @T.prim_func
51 def main(
52 A: T.Tensor(A_shape, in_dtype),
53 B: T.Tensor(B_shape, in_dtype),
54 Bias: T.Tensor(Bias_shape, out_dtype),
55 C: T.Tensor(C_shape, out_dtype),
56 ):
57 with T.Kernel(T.ceildiv(N, n_partition), M, threads=(reduce_thread, n_partition)) as (
58 bx,
59 by,
60 ):
61 A_local = T.alloc_local((micro_size_k,), in_dtype)
62 B_local = T.alloc_local((micro_size_k,), in_dtype)
63 accum_res = T.alloc_local((1,), accum_dtype)
64 reduced_accum_res = T.alloc_local((1,), accum_dtype)
65
66 kr = T.get_thread_binding(0)
67 ni = T.get_thread_binding(1)
68
69 T.clear(accum_res)
70 for ko in T.serial(T.ceildiv(K, block_K)):
71 for v in T.vectorized(micro_size_k):
72 A_local[v] = A[by, ko * block_K + kr * micro_size_k + v]
73
74 for v in T.vectorized(micro_size_k):
75 B_local[v] = B[
76 bx * n_partition + ni,
77 ko * block_K + kr * micro_size_k + v,
78 ]
79
80 if use_dp4a:
81 for ki in T.serial(micro_size_k // dp4a_size):
82 T.dp4a(
83 A_local[ki * dp4a_size],
84 B_local[ki * dp4a_size],
85 accum_res[0],
86 )
87 else:
88 for ki in T.serial(micro_size_k):
89 accum_res[0] += A_local[ki].astype(accum_dtype) * B_local[ki].astype(accum_dtype)
90
91 with T.attr(
92 T.comm_reducer(lambda x, y: x + y, [T.cast(0, accum_dtype)]),
93 "reduce_scope",
94 T.reinterpret(T.uint64(0), dtype="handle"),
95 ):
96 T.evaluate(
97 T.tvm_thread_allreduce(
98 T.uint32(1),
99 accum_res[0],
100 True,
101 reduced_accum_res[0],
102 kr,
103 dtype="handle",
104 )
105 )
106 if kr == 0:
107 if with_bias:
108 C[by, bx * n_partition + ni] = reduced_accum_res[0] + Bias[bx * n_partition + ni]

Callers

nothing calls this directly

Calls 1

get_thread_bindingMethod · 0.45

Tested by

no test coverage detected

Used in the wild real call sites across dependent graphs

searching dependent graphs…