Skip to content

Commit 4067d14

Browse files
author
Timmy
committed
fix linux build
1 parent c4e7964 commit 4067d14

File tree

4 files changed

+64
-64
lines changed

4 files changed

+64
-64
lines changed

src/library/blas/trtri/triple_dgemm_update_128_16_R.cpp

Lines changed: 16 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -98,10 +98,10 @@ int PagesPerNB = NB / (blk * 2); \n
9898

9999
do {\n
100100
double a[4]; \n
101-
a[0] = READA; incA += lda; \n
102-
a[1] = READA; incA += lda; \n
103-
a[2] = READA; incA += lda; \n
104-
a[3] = READA; incA += lda; \n
101+
a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
102+
a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
103+
a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
104+
a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
105105

106106
bs[inx][iny] = B[0 * ldb]; \n
107107
bs[inx][iny + 4] = B[4 * ldb]; \n
@@ -122,20 +122,20 @@ int PagesPerNB = NB / (blk * 2); \n
122122
//__syncthreads();
123123
barrier(CLK_LOCAL_MEM_FENCE); \n
124124

125-
daxpy(a[0], &bs[0][0], c); a[0] = READA; incA += lda; \n
126-
daxpy(a[1], &bs[1][0], c); a[1] = READA; incA += lda; \n
127-
daxpy(a[2], &bs[2][0], c); a[2] = READA; incA += lda; \n
128-
daxpy(a[3], &bs[3][0], c); a[3] = READA; incA += lda; \n
125+
daxpy(a[0], &bs[0][0], c); a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
126+
daxpy(a[1], &bs[1][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
127+
daxpy(a[2], &bs[2][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
128+
daxpy(a[3], &bs[3][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
129129

130-
daxpy(a[0], &bs[4][0], c); a[0] = READA; incA += lda; \n
131-
daxpy(a[1], &bs[5][0], c); a[1] = READA; incA += lda; \n
132-
daxpy(a[2], &bs[6][0], c); a[2] = READA; incA += lda; \n
133-
daxpy(a[3], &bs[7][0], c); a[3] = READA; incA += lda; \n
130+
daxpy(a[0], &bs[4][0], c); a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
131+
daxpy(a[1], &bs[5][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
132+
daxpy(a[2], &bs[6][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
133+
daxpy(a[3], &bs[7][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
134134

135-
daxpy(a[0], &bs[8][0], c); a[0] = READA; incA += lda; \n
136-
daxpy(a[1], &bs[9][0], c); a[1] = READA; incA += lda; \n
137-
daxpy(a[2], &bs[10][0], c); a[2] = READA; incA += lda; \n
138-
daxpy(a[3], &bs[11][0], c); a[3] = READA; incA += lda; \n
135+
daxpy(a[0], &bs[8][0], c); a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
136+
daxpy(a[1], &bs[9][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
137+
daxpy(a[2], &bs[10][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
138+
daxpy(a[3], &bs[11][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
139139

140140
daxpy(a[0], &bs[12][0], c); \n
141141
daxpy(a[1], &bs[13][0], c); \n

src/library/blas/trtri/triple_dgemm_update_128_32_PART1_R.cpp

Lines changed: 16 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -96,10 +96,10 @@ int PagesPerNB = NB / (blk * 2); \n
9696

9797
do {\n
9898
double a[4]; \n
99-
a[0] = READA; incA += lda; \n
100-
a[1] = READA; incA += lda; \n
101-
a[2] = READA; incA += lda; \n
102-
a[3] = READA; incA += lda; \n
99+
a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
100+
a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
101+
a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
102+
a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
103103

104104
bs[inx][iny] = B[0 * ldb]; \n
105105
bs[inx][iny + 4] = B[4 * ldb]; \n
@@ -112,20 +112,20 @@ int PagesPerNB = NB / (blk * 2); \n
112112
//__syncthreads();
113113
barrier(CLK_LOCAL_MEM_FENCE); \n
114114

115-
daxpy(a[0], &bs[0][0], c); a[0] = READA; incA += lda; \n
116-
daxpy(a[1], &bs[1][0], c); a[1] = READA; incA += lda; \n
117-
daxpy(a[2], &bs[2][0], c); a[2] = READA; incA += lda; \n
118-
daxpy(a[3], &bs[3][0], c); a[3] = READA; incA += lda; \n
115+
daxpy(a[0], &bs[0][0], c); a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
116+
daxpy(a[1], &bs[1][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
117+
daxpy(a[2], &bs[2][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
118+
daxpy(a[3], &bs[3][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
119119
\n
120-
daxpy(a[0], &bs[4][0], c); a[0] = READA; incA += lda; \n
121-
daxpy(a[1], &bs[5][0], c); a[1] = READA; incA += lda; \n
122-
daxpy(a[2], &bs[6][0], c); a[2] = READA; incA += lda; \n
123-
daxpy(a[3], &bs[7][0], c); a[3] = READA; incA += lda; \n
120+
daxpy(a[0], &bs[4][0], c); a[0] = ( (incA < maxA ) ? Ain[incA] : 0 ) ; incA += lda; \n
121+
daxpy(a[1], &bs[5][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
122+
daxpy(a[2], &bs[6][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
123+
daxpy(a[3], &bs[7][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
124124
\n
125-
daxpy(a[0], &bs[8][0], c); a[0] = READA; incA += lda; \n
126-
daxpy(a[1], &bs[9][0], c); a[1] = READA; incA += lda; \n
127-
daxpy(a[2], &bs[10][0], c); a[2] = READA; incA += lda; \n
128-
daxpy(a[3], &bs[11][0], c); a[3] = READA; incA += lda; \n
125+
daxpy(a[0], &bs[8][0], c); a[0] = ( (incA < maxA ) ? Ain[incA] : 0 ) ; incA += lda; \n
126+
daxpy(a[1], &bs[9][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
127+
daxpy(a[2], &bs[10][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
128+
daxpy(a[3], &bs[11][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
129129

130130
daxpy(a[0], &bs[12][0], c);\n
131131
daxpy(a[1], &bs[13][0], c);\n

src/library/blas/trtri/triple_dgemm_update_128_64_PART1_R.cpp

Lines changed: 16 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -96,10 +96,10 @@ int PagesPerNB = NB / (blk * 2); \n
9696

9797
do {\n
9898
double a[4]; \n
99-
a[0] = READA; incA += lda; \n
100-
a[1] = READA; incA += lda; \n
101-
a[2] = READA; incA += lda; \n
102-
a[3] = READA; incA += lda; \n
99+
a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
100+
a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
101+
a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
102+
a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
103103

104104
bs[inx][iny] = B[0 * ldb]; \n
105105
bs[inx][iny + 4] = B[4 * ldb]; \n
@@ -108,20 +108,20 @@ int PagesPerNB = NB / (blk * 2); \n
108108
//__syncthreads();
109109
barrier(CLK_LOCAL_MEM_FENCE); \n
110110

111-
daxpy(a[0], &bs[0][0], c); a[0] = READA; incA += lda;\n
112-
daxpy(a[1], &bs[1][0], c); a[1] = READA; incA += lda;\n
113-
daxpy(a[2], &bs[2][0], c); a[2] = READA; incA += lda;\n
114-
daxpy(a[3], &bs[3][0], c); a[3] = READA; incA += lda;\n
111+
daxpy(a[0], &bs[0][0], c); a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
112+
daxpy(a[1], &bs[1][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
113+
daxpy(a[2], &bs[2][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
114+
daxpy(a[3], &bs[3][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
115115
\n
116-
daxpy(a[0], &bs[4][0], c); a[0] = READA; incA += lda;\n
117-
daxpy(a[1], &bs[5][0], c); a[1] = READA; incA += lda;\n
118-
daxpy(a[2], &bs[6][0], c); a[2] = READA; incA += lda;\n
119-
daxpy(a[3], &bs[7][0], c); a[3] = READA; incA += lda;\n
116+
daxpy(a[0], &bs[4][0], c); a[0] = ( (incA < maxA ) ? Ain[incA] : 0 ); incA += lda;\n
117+
daxpy(a[1], &bs[5][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
118+
daxpy(a[2], &bs[6][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
119+
daxpy(a[3], &bs[7][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
120120
\n
121-
daxpy(a[0], &bs[8][0], c); a[0] = READA; incA += lda;\n
122-
daxpy(a[1], &bs[9][0], c); a[1] = READA; incA += lda;\n
123-
daxpy(a[2], &bs[10][0], c); a[2] = READA; incA += lda; \n
124-
daxpy(a[3], &bs[11][0], c); a[3] = READA; incA += lda; \n
121+
daxpy(a[0], &bs[8][0], c); a[0] = ( (incA < maxA ) ? Ain[incA] : 0 ); incA += lda;\n
122+
daxpy(a[1], &bs[9][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
123+
daxpy(a[2], &bs[10][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
124+
daxpy(a[3], &bs[11][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
125125

126126
daxpy(a[0], &bs[12][0], c);\n
127127
daxpy(a[1], &bs[13][0], c);\n

src/library/blas/trtri/triple_dgemm_update_128_ABOVE64_PART1_R.cpp

Lines changed: 16 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -96,10 +96,10 @@ int PagesPerNB = NB / (blk * 2); \n
9696

9797
do {\n
9898
double a[4]; \n
99-
a[0] = READA; incA += lda; \n
100-
a[1] = READA; incA += lda; \n
101-
a[2] = READA; incA += lda; \n
102-
a[3] = READA; incA += lda; \n
99+
a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
100+
a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
101+
a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
102+
a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
103103

104104
bs[inx][iny] = B[0 * ldb]; \n
105105
bs[inx][iny + 4] = B[4 * ldb]; \n
@@ -108,20 +108,20 @@ int PagesPerNB = NB / (blk * 2); \n
108108
//__syncthreads();
109109
barrier(CLK_LOCAL_MEM_FENCE); \n
110110

111-
daxpy(a[0], &bs[0][0], c); a[0] = READA; incA += lda; \n
112-
daxpy(a[1], &bs[1][0], c); a[1] = READA; incA += lda; \n
113-
daxpy(a[2], &bs[2][0], c); a[2] = READA; incA += lda; \n
114-
daxpy(a[3], &bs[3][0], c); a[3] = READA; incA += lda; \n
111+
daxpy(a[0], &bs[0][0], c); a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
112+
daxpy(a[1], &bs[1][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
113+
daxpy(a[2], &bs[2][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
114+
daxpy(a[3], &bs[3][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
115115

116-
daxpy(a[0], &bs[4][0], c); a[0] = READA; incA += lda; \n
117-
daxpy(a[1], &bs[5][0], c); a[1] = READA; incA += lda; \n
118-
daxpy(a[2], &bs[6][0], c); a[2] = READA; incA += lda; \n
119-
daxpy(a[3], &bs[7][0], c); a[3] = READA; incA += lda; \n
116+
daxpy(a[0], &bs[4][0], c); a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
117+
daxpy(a[1], &bs[5][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
118+
daxpy(a[2], &bs[6][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
119+
daxpy(a[3], &bs[7][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
120120

121-
daxpy(a[0], &bs[8][0], c); a[0] = READA; incA += lda; \n
122-
daxpy(a[1], &bs[9][0], c); a[1] = READA; incA += lda; \n
123-
daxpy(a[2], &bs[10][0], c); a[2] = READA; incA += lda; \n
124-
daxpy(a[3], &bs[11][0], c); a[3] = READA; incA += lda; \n
121+
daxpy(a[0], &bs[8][0], c); a[0] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
122+
daxpy(a[1], &bs[9][0], c); a[1] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
123+
daxpy(a[2], &bs[10][0], c); a[2] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
124+
daxpy(a[3], &bs[11][0], c); a[3] = ((incA < maxA) ? Ain[incA] : 0); incA += lda; \n
125125

126126
daxpy(a[0], &bs[12][0], c); \n
127127
daxpy(a[1], &bs[13][0], c); \n

0 commit comments

Comments
 (0)