Minor fixes
This commit is contained in:
parent
e3a80e6aa8
commit
a8cc6b365d
|
@ -801,32 +801,20 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
|
||||||
|
|
||||||
i = (m & 1);
|
i = (m & 1);
|
||||||
if (i > 0) {
|
if (i > 0) {
|
||||||
if (X > posY) {
|
|
||||||
/* a01 += 2;
|
|
||||||
a02 += 2; */
|
|
||||||
b += 4;
|
|
||||||
} else
|
|
||||||
if (X < posY) {
|
|
||||||
b[ 0] = *(a01 + 0);
|
|
||||||
b[ 1] = *(a01 + 1);
|
|
||||||
b[ 2] = *(a01 + 2);
|
|
||||||
b[ 3] = *(a01 + 3);
|
|
||||||
|
|
||||||
/* a01 += lda;
|
|
||||||
a02 += lda; */
|
|
||||||
b += 4;
|
|
||||||
} else {
|
|
||||||
#ifdef UNIT
|
#ifdef UNIT
|
||||||
b[ 0] = ONE;
|
if (X < posY) {
|
||||||
b[ 1] = ZERO;
|
|
||||||
#else
|
|
||||||
b[ 0] = *(a01 + 0);
|
|
||||||
b[ 1] = *(a01 + 1);
|
|
||||||
#endif
|
#endif
|
||||||
b[ 2] = *(a01 + 2);
|
b[ 0] = *(a01 + 0);
|
||||||
b[ 3] = *(a01 + 3);
|
b[ 1] = *(a01 + 1);
|
||||||
b += 4;
|
#ifdef UNIT
|
||||||
}
|
} else {
|
||||||
|
b[ 0] = ONE;
|
||||||
|
b[ 1] = ZERO;
|
||||||
|
}
|
||||||
|
#endif
|
||||||
|
b[ 2] = *(a01 + 2);
|
||||||
|
b[ 3] = *(a01 + 3);
|
||||||
|
b += 4;
|
||||||
}
|
}
|
||||||
posY += 2;
|
posY += 2;
|
||||||
}
|
}
|
||||||
|
|
|
@ -301,7 +301,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
|
||||||
dtrmm_kernel_4x8( temp, &alpha , ptrba, ptrbb, C0, C1, C2, C3, C4, C5, C6, C7);
|
dtrmm_kernel_4x8( temp, &alpha , ptrba, ptrbb, C0, C1, C2, C3, C4, C5, C6, C7);
|
||||||
|
|
||||||
ptrba = ptrba + temp * 4;
|
ptrba = ptrba + temp * 4;
|
||||||
ptrbb = ptrbb + temp * 8;
|
// ptrbb = ptrbb + temp * 8;
|
||||||
|
|
||||||
/*
|
/*
|
||||||
for (k=0; k<temp; k++)
|
for (k=0; k<temp; k++)
|
||||||
|
@ -446,7 +446,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
|
||||||
temp - 8; // number of values in B
|
temp - 8; // number of values in B
|
||||||
|
|
||||||
ptrba += temp*4; // number of values in A
|
ptrba += temp*4; // number of values in A
|
||||||
ptrbb += temp*8; // number of values in B
|
// ptrbb += temp*8; // number of values in B
|
||||||
}
|
}
|
||||||
#ifdef LEFT
|
#ifdef LEFT
|
||||||
off += 4; // number of values in A
|
off += 4; // number of values in A
|
||||||
|
@ -709,14 +709,14 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
|
||||||
off += 1; // number of values in A
|
off += 1; // number of values in A
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
C0 = C0+1;
|
/* C0 = C0+1;
|
||||||
C1 = C1+1;
|
C1 = C1+1;
|
||||||
C2 = C2+1;
|
C2 = C2+1;
|
||||||
C3 = C3+1;
|
C3 = C3+1;
|
||||||
C4 = C4+1;
|
C4 = C4+1;
|
||||||
C5 = C5+1;
|
C5 = C5+1;
|
||||||
C6 = C6+1;
|
C6 = C6+1;
|
||||||
C7 = C7+1;
|
C7 = C7+1; */
|
||||||
|
|
||||||
}
|
}
|
||||||
|
|
||||||
|
@ -862,7 +862,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
|
||||||
temp - 4; // number of values in B */
|
temp - 4; // number of values in B */
|
||||||
|
|
||||||
ptrba += temp*4; // number of values in A
|
ptrba += temp*4; // number of values in A
|
||||||
ptrbb += temp*4; // number of values in B
|
// ptrbb += temp*4; // number of values in B
|
||||||
}
|
}
|
||||||
#ifdef LEFT
|
#ifdef LEFT
|
||||||
off += 4; // number of values in A
|
off += 4; // number of values in A
|
||||||
|
@ -1049,10 +1049,10 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
|
||||||
off += 1; // number of values in A
|
off += 1; // number of values in A
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
C0 = C0+1;
|
/* C0 = C0+1;
|
||||||
C1 = C1+1;
|
C1 = C1+1;
|
||||||
C2 = C2+1;
|
C2 = C2+1;
|
||||||
C3 = C3+1;
|
C3 = C3+1; */
|
||||||
|
|
||||||
}
|
}
|
||||||
|
|
||||||
|
@ -1311,8 +1311,8 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
|
||||||
off += 1; // number of values in A
|
off += 1; // number of values in A
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
C0 = C0+1;
|
/* C0 = C0+1;
|
||||||
C1 = C1+1;
|
C1 = C1+1; */
|
||||||
|
|
||||||
}
|
}
|
||||||
|
|
||||||
|
@ -1532,7 +1532,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
|
||||||
off += 1; // number of values in A
|
off += 1; // number of values in A
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
C0 = C0+1;
|
// C0 = C0+1;
|
||||||
|
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
|
@ -54,7 +54,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
|
||||||
BLASLONG n, info;
|
BLASLONG n, info;
|
||||||
BLASLONG bk, i, blocking, start_i;
|
BLASLONG bk, i, blocking, start_i;
|
||||||
int mode;
|
int mode;
|
||||||
BLASLONG lda, range_N[2];
|
BLASLONG lda;//, range_N[2];
|
||||||
blas_arg_t newarg;
|
blas_arg_t newarg;
|
||||||
FLOAT *a;
|
FLOAT *a;
|
||||||
FLOAT alpha[2] = { ONE, ZERO};
|
FLOAT alpha[2] = { ONE, ZERO};
|
||||||
|
@ -100,8 +100,8 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
|
||||||
bk = n - i;
|
bk = n - i;
|
||||||
if (bk > blocking) bk = blocking;
|
if (bk > blocking) bk = blocking;
|
||||||
|
|
||||||
range_N[0] = i;
|
/* range_N[0] = i;
|
||||||
range_N[1] = i + bk;
|
range_N[1] = i + bk; */
|
||||||
|
|
||||||
newarg.lda = lda;
|
newarg.lda = lda;
|
||||||
newarg.ldb = lda;
|
newarg.ldb = lda;
|
||||||
|
|
|
@ -54,7 +54,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
|
||||||
BLASLONG n, info;
|
BLASLONG n, info;
|
||||||
BLASLONG bk, i, blocking;
|
BLASLONG bk, i, blocking;
|
||||||
int mode;
|
int mode;
|
||||||
BLASLONG lda, range_N[2];
|
BLASLONG lda; // , range_N[2];
|
||||||
blas_arg_t newarg;
|
blas_arg_t newarg;
|
||||||
FLOAT *a;
|
FLOAT *a;
|
||||||
FLOAT alpha[2] = { ONE, ZERO};
|
FLOAT alpha[2] = { ONE, ZERO};
|
||||||
|
@ -96,8 +96,8 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
|
||||||
bk = n - i;
|
bk = n - i;
|
||||||
if (bk > blocking) bk = blocking;
|
if (bk > blocking) bk = blocking;
|
||||||
|
|
||||||
range_N[0] = i;
|
/* range_N[0] = i;
|
||||||
range_N[1] = i + bk;
|
range_N[1] = i + bk; */
|
||||||
|
|
||||||
newarg.lda = lda;
|
newarg.lda = lda;
|
||||||
newarg.ldb = lda;
|
newarg.ldb = lda;
|
||||||
|
|
Loading…
Reference in New Issue