MCPcopy Create free account
hub / github.com/apache/tvm / wmma_load_impl

Function wmma_load_impl

python/tvm/s_tir/tensor_intrin/cuda.py:859–897  ·  view source on GitHub ↗
(a: T.handle, c: T.handle)

Source from the content-addressed store, hash-verified

857
858 @T.prim_func(s_tir=True)
859 def wmma_load_impl(a: T.handle, c: T.handle) -> None:
860 s1 = T.int32()
861 s0 = T.int32()
862 d1 = T.int32()
863 d0 = T.int32()
864 A = T.match_buffer(
865 a,
866 (frag_m, frag_n),
867 dtype,
868 align=64,
869 offset_factor=offset_factor,
870 scope=shared_scope,
871 strides=[s1, s0],
872 )
873 C = T.match_buffer(
874 c,
875 (frag_m, frag_n),
876 dtype,
877 align=64,
878 offset_factor=offset_factor,
879 scope=wmma_fragment_scope,
880 strides=[d1, d0],
881 )
882 with T.sblock("root"):
883 T.reads(A[0:frag_m, 0:frag_n])
884 T.writes(C[0:frag_m, 0:frag_n])
885 T.evaluate(
886 T.tvm_load_matrix_sync(
887 C.data,
888 m_dim,
889 n_dim,
890 k_dim,
891 get_wmma_fragment_index(C, d1, frag_m, frag_n),
892 A.access_ptr("r"),
893 s1,
894 layout,
895 dtype="handle",
896 )
897 )
898
899 return wmma_load_desc, wmma_load_impl
900

Callers

nothing calls this directly

Calls 3

get_wmma_fragment_indexFunction · 0.85
access_ptrMethod · 0.80
evaluateMethod · 0.45

Tested by

no test coverage detected

Used in the wild real call sites across dependent graphs

searching dependent graphs…