Skip to content

Commit 0834bc7

Browse files
author
Chao Liu
committed
compiler parameter use stream
1 parent f2ac783 commit 0834bc7

File tree

1 file changed

+90
-85
lines changed

1 file changed

+90
-85
lines changed

‎host/solver/include/conv_igemm_fwd_v6r1_dlops_nchw_kcyx_nkhw.hpp

Lines changed: 90 additions & 85 deletions
Original file line numberDiff line numberDiff line change
@@ -2,6 +2,7 @@
22
#define CONV_IGEMM_FWD_V6R1_DLOPS_NCHW_KCYX_NKHW_HPP
33

44
#include <numeric>
5+
#include <sstream>
56

67
namespace ck {
78
namespace driver {
@@ -10,93 +11,97 @@ struct CompileParameterConvIgemmFwdV6r1DlopsNchwKcyxNkhw
1011
{
1112
auto GetCompileParameterString() const
1213
{
14+
auto param = std::stringstream();
15+
1316
// clang-format off
14-
return
15-
" -DCK_PARAM_ABDataTypeEnum=" +
16-
std::to_string(ABDataTypeEnum) +
17-
" -DCK_PARAM_AccDataTypeEnum=" +
18-
std::to_string(AccDataTypeEnum) +
19-
" -DCK_PARAM_CDataTypeEnum=" +
20-
std::to_string(CDataTypeEnum) +
21-
" -DCK_PARAM_BlockSize=" +
22-
std::to_string(BlockSize) +
23-
" -DCK_PARAM_GN0=" +
24-
std::to_string(GN0) +
25-
" -DCK_PARAM_GK1=" +
26-
std::to_string(GK1) +
27-
" -DCK_PARAM_GM1PerBlockGM11=" +
28-
std::to_string(GM1PerBlockGM11) +
29-
" -DCK_PARAM_GN1PerBlockGN11=" +
30-
std::to_string(GN1PerBlockGN11) +
31-
" -DCK_PARAM_GK0PerBlock=" +
32-
std::to_string(GK0PerBlock) +
33-
" -DCK_PARAM_BM1PerThreadBM11=" +
34-
std::to_string(BM1PerThreadBM11) +
35-
" -DCK_PARAM_BN1PerThreadBN11=" +
36-
std::to_string(BN1PerThreadBN11) +
37-
" -DCK_PARAM_BK0PerThread=" +
38-
std::to_string(BK0PerThread) +
39-
" -DCK_PARAM_BM10BN10ThreadClusterBM10Xs=" +
40-
std::to_string(BM10BN10ThreadClusterBM10Xs[0]) + "," +
41-
std::to_string(BM10BN10ThreadClusterBM10Xs[1]) +
42-
" -DCK_PARAM_BM10BN10ThreadClusterBN10Xs=" +
43-
std::to_string(BM10BN10ThreadClusterBN10Xs[0]) + "," +
44-
std::to_string(BM10BN10ThreadClusterBN10Xs[1]) +
45-
" -DCK_PARAM_ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1=" +
46-
std::to_string(ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[0]) + "," +
47-
std::to_string(ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[1]) + "," +
48-
std::to_string(ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[2]) + "," +
49-
std::to_string(ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[3]) + "," +
50-
std::to_string(ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[4]) +
51-
" -DCK_PARAM_ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1=" +
52-
std::to_string(ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[0]) + "," +
53-
std::to_string(ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[1]) + "," +
54-
std::to_string(ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[2]) + "," +
55-
std::to_string(ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[3]) + "," +
56-
std::to_string(ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[4]) +
57-
" -DCK_PARAM_ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1=" +
58-
std::to_string(ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[0]) + "," +
59-
std::to_string(ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[1]) + "," +
60-
std::to_string(ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[2]) + "," +
61-
std::to_string(ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[3]) + "," +
62-
std::to_string(ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[4]) +
63-
" -DCK_PARAM_ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1=" +
64-
std::to_string(ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[0]) + "," +
65-
std::to_string(ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[1]) + "," +
66-
std::to_string(ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[2]) + "," +
67-
std::to_string(ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[3]) + "," +
68-
std::to_string(ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[4]) +
69-
" -DCK_PARAM_BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1=" +
70-
std::to_string(BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[0]) + "," +
71-
std::to_string(BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[1]) + "," +
72-
std::to_string(BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[2]) + "," +
73-
std::to_string(BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[3]) + "," +
74-
std::to_string(BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[4]) +
75-
" -DCK_PARAM_BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1=" +
76-
std::to_string(BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[0]) + "," +
77-
std::to_string(BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[1]) + "," +
78-
std::to_string(BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[2]) + "," +
79-
std::to_string(BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[3]) + "," +
80-
std::to_string(BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[4]) +
81-
" -DCK_PARAM_BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1=" +
82-
std::to_string(BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[0]) + "," +
83-
std::to_string(BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[1]) + "," +
84-
std::to_string(BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[2]) + "," +
85-
std::to_string(BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[3]) + "," +
86-
std::to_string(BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[4]) +
87-
" -DCK_PARAM_BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1=" +
88-
std::to_string(BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[0]) + "," +
89-
std::to_string(BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[1]) + "," +
90-
std::to_string(BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[2]) + "," +
91-
std::to_string(BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[3]) + "," +
92-
std::to_string(BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[4]) +
93-
" -DCK_PARAM_CThreadTransferDstScalarPerVector=" +
94-
std::to_string(CThreadTransferDstScalarPerVector) +
95-
" -DCK_PARAM_HasMainKBlockLoop=" +
96-
std::to_string(static_cast<int>(HasMainKBlockLoop)) +
97-
" -DCK_PARAM_HasDoubleTailKBlockLoop=" +
98-
std::to_string(static_cast<int>(HasDoubleTailKBlockLoop));
17+
param <<
18+
" -DCK_PARAM_ABDataTypeEnum=" <<
19+
ABDataTypeEnum <<
20+
" -DCK_PARAM_AccDataTypeEnum=" <<
21+
AccDataTypeEnum <<
22+
" -DCK_PARAM_CDataTypeEnum=" <<
23+
CDataTypeEnum <<
24+
" -DCK_PARAM_BlockSize=" <<
25+
BlockSize <<
26+
" -DCK_PARAM_GN0=" <<
27+
GN0 <<
28+
" -DCK_PARAM_GK1=" <<
29+
GK1 <<
30+
" -DCK_PARAM_GM1PerBlockGM11="
31+
<< GM1PerBlockGM11 <<
32+
" -DCK_PARAM_GN1PerBlockGN11=" <<
33+
GN1PerBlockGN11 <<
34+
" -DCK_PARAM_GK0PerBlock=" <<
35+
GK0PerBlock <<
36+
" -DCK_PARAM_BM1PerThreadBM11=" <<
37+
BM1PerThreadBM11 <<
38+
" -DCK_PARAM_BN1PerThreadBN11=" <<
39+
BN1PerThreadBN11 <<
40+
" -DCK_PARAM_BK0PerThread=" <<
41+
BK0PerThread <<
42+
" -DCK_PARAM_BM10BN10ThreadClusterBM10Xs=" <<
43+
BM10BN10ThreadClusterBM10Xs[0] << "," <<
44+
BM10BN10ThreadClusterBM10Xs[1] <<
45+
" -DCK_PARAM_BM10BN10ThreadClusterBN10Xs=" <<
46+
BM10BN10ThreadClusterBN10Xs[0] << "," <<
47+
BM10BN10ThreadClusterBN10Xs[1] <<
48+
" -DCK_PARAM_ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1=" <<
49+
ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[0] << "," <<
50+
ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[1] << "," <<
51+
ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[2] << "," <<
52+
ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[3] << "," <<
53+
ABlockTransferThreadSliceLengths_GK0_GM0_GM10_GM11_GK1[4] <<
54+
" -DCK_PARAM_ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1=" <<
55+
ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[0] << "," <<
56+
ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[1] << "," <<
57+
ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[2] << "," <<
58+
ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[3] << "," <<
59+
ABlockTransferThreadClusterLengths_GK0_GM0_GM10_GM11_GK1[4] <<
60+
" -DCK_PARAM_ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1=" <<
61+
ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[0] << "," <<
62+
ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[1] << "," <<
63+
ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[2] << "," <<
64+
ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[3] << "," <<
65+
ABlockTransferSrcVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[4] <<
66+
" -DCK_PARAM_ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1=" <<
67+
ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[0] << "," <<
68+
ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[1] << "," <<
69+
ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[2] << "," <<
70+
ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[3] << "," <<
71+
ABlockTransferDstVectorTensorLengths_GK0_GM0_GM10_GM11_GK1[4] <<
72+
" -DCK_PARAM_BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1=" <<
73+
BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[0] << "," <<
74+
BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[1] << "," <<
75+
BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[2] << "," <<
76+
BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[3] << "," <<
77+
BBlockTransferThreadSliceLengths_GK0_GN0_GN10_GN11_GK1[4] <<
78+
" -DCK_PARAM_BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1=" <<
79+
BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[0] << "," <<
80+
BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[1] << "," <<
81+
BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[2] << "," <<
82+
BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[3] << "," <<
83+
BBlockTransferThreadClusterLengths_GK0_GN0_GN10_GN11_GK1[4] <<
84+
" -DCK_PARAM_BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1=" <<
85+
BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[0] << "," <<
86+
BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[1] << "," <<
87+
BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[2] << "," <<
88+
BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[3] << "," <<
89+
BBlockTransferSrcVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[4] <<
90+
" -DCK_PARAM_BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1=" <<
91+
BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[0] << "," <<
92+
BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[1] << "," <<
93+
BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[2] << "," <<
94+
BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[3] << "," <<
95+
BBlockTransferDstVectorTensorLengths_GK0_GN0_GN10_GN11_GK1[4] <<
96+
" -DCK_PARAM_CThreadTransferDstScalarPerVector=" <<
97+
CThreadTransferDstScalarPerVector <<
98+
" -DCK_PARAM_HasMainKBlockLoop=" <<
99+
static_cast<int>(HasMainKBlockLoop) <<
100+
" -DCK_PARAM_HasDoubleTailKBlockLoop=" <<
101+
static_cast<int>(HasDoubleTailKBlockLoop);
99102
// clang-format on
103+
104+
return param.str();
100105
}
101106

102107
ck::DataTypeEnum_t ABDataTypeEnum = ck::DataTypeEnum_t::Unknown;

0 commit comments

Comments
 (0)