Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
sddmm.cu 4.18 KB
Newer Older
Zihao Ye's avatar
[kernel] New SpMM & SDDMM kernel on CPU and CUDA (#1644)  
Zihao Ye 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 
/*!
 *  Copyright (c) 2020 by Contributors
 * \file array/cuda/sddmm.cu
 * \brief SDDMM C APIs and definitions.
 */
#include <dgl/array.h>
#include "./sddmm.cuh"
#include "./functor.cuh"

namespace dgl {
namespace aten {

#define SWITCH_OP(op, Op, ...)                                      \
  do {                                                              \
    if ((op) == "add") {                                            \
      typedef cuda::binary::Add<DType> Op;                          \
      { __VA_ARGS__ }                                               \
    } else if ((op) == "sub") {                                     \
      typedef cuda::binary::Sub<DType> Op;                          \
      { __VA_ARGS__ }                                               \
    } else if ((op) == "mul") {                                     \
      typedef cuda::binary::Mul<DType> Op;                          \
      { __VA_ARGS__ }                                               \
    } else if ((op) == "div") {                                     \
      typedef cuda::binary::Div<DType> Op;                          \
      { __VA_ARGS__ }                                               \
    } else if ((op) == "copy_u") {                                  \
      typedef cuda::binary::CopyU<DType> Op;                        \
      { __VA_ARGS__ }                                               \
    } else if ((op) == "copy_e") {                                  \
      typedef cuda::binary::CopyE<DType> Op;                        \
      { __VA_ARGS__ }                                               \
    } else if ((op) == "dot") {                                     \
      typedef cuda::binary::Dot<DType> Op;                          \
      { __VA_ARGS__ }                                               \
    } else {                                                        \
      LOG(FATAL) << "Unsupported SpMM/SDDMM binary operator: " << op;     \
    }                                                               \
  } while (0)


/*!
 * \brief CUDA implementation of g-SDDMM on Csr format.
 */
template <int XPU, typename IdType, typename DType>
void SDDMMCsr(const std::string& op,
              const BcastOff& bcast,
              const CSRMatrix& csr,
              NDArray ufeat,
              NDArray vfeat,
              NDArray out) {
  SWITCH_OP(op, Op, {
    cuda::SDDMMCsr<IdType, DType, Op>(bcast, csr, ufeat, vfeat, out);
  });
}

/*!
 * \brief CUDA implementation of g-SDDMM on Coo format.
 */
template <int XPU, typename IdType, typename DType>
void SDDMMCoo(const std::string& op,
              const BcastOff& bcast,
              const COOMatrix& coo,
              NDArray ufeat,
              NDArray vfeat,
              NDArray out) {
  SWITCH_OP(op, Op, {
    cuda::SDDMMCoo<IdType, DType, Op>(bcast, coo, ufeat, vfeat, out);
  });
}

template void SDDMMCsr<kDLGPU, int32_t, float>(
    const std::string& op, const BcastOff& bcast, const CSRMatrix& csr,
    NDArray ufeat, NDArray vfeat, NDArray out);
template void SDDMMCsr<kDLGPU, int64_t, float>(
    const std::string& op, const BcastOff& bcast, const CSRMatrix& csr,
    NDArray ufeat, NDArray vfeat, NDArray out);
template void SDDMMCsr<kDLGPU, int32_t, double>(
    const std::string& op, const BcastOff& bcast, const CSRMatrix& csr,
    NDArray ufeat, NDArray vfeat, NDArray out);
template void SDDMMCsr<kDLGPU, int64_t, double>(
    const std::string& op, const BcastOff& bcast, const CSRMatrix& csr,
    NDArray ufeat, NDArray vfeat, NDArray out);

template void SDDMMCoo<kDLGPU, int32_t, float>(
    const std::string& op, const BcastOff& bcast, const COOMatrix& coo,
    NDArray ufeat, NDArray vfeat, NDArray out);
template void SDDMMCoo<kDLGPU, int64_t, float>(
    const std::string& op, const BcastOff& bcast, const COOMatrix& coo,
    NDArray ufeat, NDArray vfeat, NDArray out);
template void SDDMMCoo<kDLGPU, int32_t, double>(
    const std::string& op, const BcastOff& bcast, const COOMatrix& coo,
    NDArray ufeat, NDArray vfeat, NDArray out);
template void SDDMMCoo<kDLGPU, int64_t, double>(
    const std::string& op, const BcastOff& bcast, const COOMatrix& coo,
    NDArray ufeat, NDArray vfeat, NDArray out);

}  // namespace aten
}  // namespace dgl