Merge pull request #986 from ksraste/develop

SGEMM, DGEMM, CGEMM, ZGEMM functions data prefetch
This commit is contained in:
Zhang Xianyi 2016-10-18 12:38:52 +08:00 committed by GitHub
commit e54b6ddaa0
4 changed files with 873 additions and 250 deletions

View File

@ -1092,18 +1092,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#if defined(TRMMKERNEL)
BLASLONG off;
#endif
FLOAT *pc0, *pc1, *pc2, *pc3;
FLOAT *pa0, *pb0;
FLOAT *pc0, *pc1, *pc2, *pc3, *pa0, *pb0;
FLOAT res0, res1, res2, res3, res4, res5, res6, res7;
FLOAT res8, res9, res10, res11, res12, res13, res14, res15;
FLOAT a0_r, a1_r;
FLOAT a0_i, a1_i;
FLOAT a0_r, a1_r, a0_i, a1_i, b0_i, b1_i, b2_i, b3_i;
FLOAT b0_r, b1_r, b2_r, b3_r;
FLOAT b0_i, b1_i, b2_i, b3_i;
v4f32 src_a0, src_a1, src_a2, src_a3, src_b0, src_b1;
v4f32 src_a0r, src_a0i, src_a1r, src_a1i, src_br, src_bi;
v4f32 dst0, dst1, dst2, dst3;
v4f32 alpha_r, alpha_i;
v4f32 dst0, dst1, dst2, dst3, alpha_r, alpha_i;
v4f32 res0_r, res0_i, res1_r, res1_i, res2_r, res2_i, res3_r, res3_i;
v4f32 res4_r, res4_i, res5_r, res5_i, res6_r, res6_i, res7_r, res7_i;
v4f32 dst0_r, dst0_i, dst1_r, dst1_i;
@ -1122,12 +1118,12 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
pc2 = pc1 + 2 * ldc;
pc3 = pc2 + 2 * ldc;
pa0 = A;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
pa0 = A;
for (i = (m >> 3); i--;)
{
#if defined(TRMMKERNEL)
@ -1150,6 +1146,17 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
temp = k;
#endif
#ifdef ENABLE_PREFETCH
__asm__ __volatile__(
"pref 0, 64(%[pa0]) \n\t"
"pref 0, 96(%[pa0]) \n\t"
"pref 0, 32(%[pb0]) \n\t"
:
: [pa0] "r" (pa0), [pb0] "r" (pb0)
);
#endif
#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
CGEMM_KERNEL_8X4_MSA(, -, , +, +);
#endif
@ -1165,6 +1172,17 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
for (l = (temp - 1); l--;)
{
#ifdef ENABLE_PREFETCH
__asm__ __volatile__(
"pref 0, 64(%[pa0]) \n\t"
"pref 0, 96(%[pa0]) \n\t"
"pref 0, 32(%[pb0]) \n\t"
:
: [pa0] "r" (pa0), [pb0] "r" (pb0)
);
#endif
#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
CGEMM_KERNEL_8X4_MSA(+, -, +, +,);
#endif
@ -1340,6 +1358,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
CGEMM_SCALE_2X4
#endif
pc0 += 4;
pc1 += 4;
pc2 += 4;
pc3 += 4;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1357,11 +1379,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 2; // number of values in A
#endif
#endif
pc0 += 4;
pc1 += 4;
pc2 += 4;
pc3 += 4;
}
if (m & 1)
@ -1426,6 +1443,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
CGEMM_SCALE_1X4
#endif
pc0 += 2;
pc1 += 2;
pc2 += 2;
pc3 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1443,21 +1464,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 1; // number of values in A
#endif
#endif
pc0 += 2;
pc1 += 2;
pc2 += 2;
pc3 += 2;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 4; // number of values in A
#endif
l = k << 3;
B = B + l;
i = ldc << 3;
C = C + i;
B += (k << 3);
C += (ldc << 3);
}
if (n & 2)
@ -1465,12 +1479,12 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
pc0 = C;
pc1 = pc0 + 2 * ldc;
pa0 = A;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
pa0 = A;
for (i = (m >> 3); i--;)
{
#if defined(TRMMKERNEL)
@ -1691,6 +1705,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
CGEMM_SCALE_2X2
#endif
pc0 += 4;
pc1 += 4;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1708,9 +1724,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 2; // number of values in A
#endif
#endif
pc0 += 4;
pc1 += 4;
}
if (m & 1)
@ -1775,6 +1788,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
CGEMM_SCALE_1X2
#endif
pc0 += 2;
pc1 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1792,30 +1807,26 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 1; // number of values in A
#endif
#endif
pc0 += 2;
pc1 += 2;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 2; // number of values in A
#endif
l = k << 2;
B = B + l;
i = ldc << 2;
C = C + i;
B += (k << 2);
C += (ldc << 2);
}
if (n & 1)
{
pc0 = C;
pa0 = A;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
pa0 = A;
for (i = (m >> 3); i--;)
{
#if defined(TRMMKERNEL)
@ -2036,6 +2047,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
CGEMM_SCALE_2X1
#endif
pc0 += 4;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -2053,8 +2065,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 2; // number of values in A
#endif
#endif
pc0 += 4;
}
if (m & 1)
@ -2119,6 +2129,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
CGEMM_SCALE_1X1
#endif
pc0 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -2136,18 +2147,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 1; // number of values in A
#endif
#endif
pc0 += 2;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 1; // number of values in A
#endif
l = k << 1;
B = B + l;
i = ldc << 1;
C = C + i;
B += (k << 1);
C += (ldc << 1);
}
return 0;

File diff suppressed because it is too large Load Diff

View File

@ -91,6 +91,15 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pb0 = B;
temp = k;
#endif
#ifdef ENABLE_PREFETCH
__asm__ __volatile__(
"pref 0, 32(%[pa0]) \n\t"
"pref 0, 32(%[pb0]) \n\t"
:
: [pa0] "r" (pa0), [pb0] "r" (pb0)
);
#endif
LD_SP2_INC(pa0, 4, src_a0, src_a1);
LD_SP2_INC(pb0, 4, src_b0, src_b1);
@ -129,6 +138,18 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
for (l = ((temp - 1) >> 1); l--;)
{
#ifdef ENABLE_PREFETCH
__asm__ __volatile__(
"pref 0, 64(%[pa0]) \n\t"
"pref 0, 96(%[pa0]) \n\t"
"pref 0, 64(%[pb0]) \n\t"
"pref 0, 96(%[pb0]) \n\t"
:
: [pa0] "r" (pa0), [pb0] "r" (pb0)
);
#endif
LD_SP2_INC(pa0, 4, src_a0, src_a1);
LD_SP2_INC(pb0, 4, src_b0, src_b1);
@ -500,6 +521,15 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
ST_SP(dst2, pc6);
ST_SP(dst3, pc7);
pc0 += 4;
pc1 += 4;
pc2 += 4;
pc3 += 4;
pc4 += 4;
pc5 += 4;
pc6 += 4;
pc7 += 4;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = k - off;
@ -516,15 +546,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 4; // number of values in A
#endif
#endif
pc0 += 4;
pc1 += 4;
pc2 += 4;
pc3 += 4;
pc4 += 4;
pc5 += 4;
pc6 += 4;
pc7 += 4;
}
if (m & 2)
@ -763,6 +784,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc6[1] += tmp13;
pc7[1] += tmp15;
#endif
pc0 += 2;
pc1 += 2;
pc2 += 2;
pc3 += 2;
pc4 += 2;
pc5 += 2;
pc6 += 2;
pc7 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -780,15 +809,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 2; // number of values in A
#endif
#endif
pc0 += 2;
pc1 += 2;
pc2 += 2;
pc3 += 2;
pc4 += 2;
pc5 += 2;
pc6 += 2;
pc7 += 2;
}
if (m & 1)
@ -959,6 +979,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc6[0] += tmp6;
pc7[0] += tmp7;
#endif
pc0 += 1;
pc1 += 1;
pc2 += 1;
pc3 += 1;
pc4 += 1;
pc5 += 1;
pc6 += 1;
pc7 += 1;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -976,24 +1004,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 1; // number of values in A
#endif
#endif
pc0 += 1;
pc1 += 1;
pc2 += 1;
pc3 += 1;
pc4 += 1;
pc5 += 1;
pc6 += 1;
pc7 += 1;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 8; // number of values in A
#endif
l = (k << 3);
B = B + l;
i = (ldc << 3);
C = C + i;
B += (k << 3);
C += (ldc << 3);
}
if (n & 4)
@ -1003,12 +1021,12 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc2 = pc1 + ldc;
pc3 = pc2 + ldc;
pa0 = A;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
pa0 = A;
for (i = (m >> 3); i--;)
{
#if defined(TRMMKERNEL)
@ -1145,7 +1163,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
dst6 += res6 * v_alpha;
dst7 += res7 * v_alpha;
#endif
ST_SP2_INC(dst0, dst1, pc0, 4);
ST_SP2_INC(dst2, dst3, pc1, 4);
ST_SP2_INC(dst4, dst5, pc2, 4);
@ -1268,6 +1285,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pa0 += 4;
pb0 += 4;
}
#if defined(TRMMKERNEL)
dst0 = res0 * v_alpha;
dst1 = res1 * v_alpha;
@ -1289,6 +1307,11 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
ST_SP(dst2, pc2);
ST_SP(dst3, pc3);
pc0 += 4;
pc1 += 4;
pc2 += 4;
pc3 += 4;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = k - off;
@ -1305,10 +1328,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 4; // number of values in A
#endif
#endif
pc0 += 4;
pc1 += 4;
pc2 += 4;
pc3 += 4;
}
if (m & 2)
@ -1459,6 +1478,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc2[1] += tmp5;
pc3[1] += tmp7;
#endif
pc0 += 2;
pc1 += 2;
pc2 += 2;
pc3 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1476,11 +1499,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 2; // number of values in A
#endif
#endif
pc0 += 2;
pc1 += 2;
pc2 += 2;
pc3 += 2;
}
if (m & 1)
@ -1591,6 +1609,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc2[0] += tmp2;
pc3[0] += tmp3;
#endif
pc0 += 1;
pc1 += 1;
pc2 += 1;
pc3 += 1;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1608,20 +1630,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 1; // number of values in A
#endif
#endif
pc0 += 1;
pc1 += 1;
pc2 += 1;
pc3 += 1;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 4; // number of values in A
#endif
l = (k << 2);
B = B + l;
i = (ldc << 2);
C = C + i;
B += (k << 2);
C += (ldc << 2);
}
if (n & 2)
@ -1629,12 +1645,12 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc0 = C;
pc1 = pc0 + ldc;
pa0 = A;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
pa0 = A;
for (i = (m >> 3); i--;)
{
#if defined(TRMMKERNEL)
@ -1847,6 +1863,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
ST_SP(dst0, pc0);
ST_SP(dst1, pc1);
pc0 += 4;
pc1 += 4;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = k - off;
@ -1863,8 +1882,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 4; // number of values in A
#endif
#endif
pc0 += 4;
pc1 += 4;
}
if (m & 2)
@ -1967,6 +1984,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc0[1] += tmp1;
pc1[1] += tmp3;
#endif
pc0 += 2;
pc1 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1984,9 +2003,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 2; // number of values in A
#endif
#endif
pc0 += 2;
pc1 += 2;
}
if (m & 1)
@ -2067,6 +2083,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc0[0] += tmp0;
pc1[0] += tmp1;
#endif
pc0 += 1;
pc1 += 1;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -2084,28 +2102,26 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 1; // number of values in A
#endif
#endif
pc0 += 1;
pc1 += 1;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 2; // number of values in A
#endif
l = (k << 1);
B = B + l;
i = (ldc << 1);
C = C + i;
B += (k << 1);
C += (ldc << 1);
}
if (n & 1)
{
pc0 = C;
pa0 = A;
#if defined(TRMMKERNEL) && defined(LEFT)
off = offset;
#endif
pa0 = A;
for (i = (m >> 3); i--;)
{
#if defined(TRMMKERNEL)
@ -2272,6 +2288,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
#endif
ST_SP(dst0, pc0);
pc0 += 4;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = k - off;
@ -2288,7 +2306,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 4; // number of values in A
#endif
#endif
pc0 += 4;
}
if (m & 2)
@ -2359,6 +2376,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pb0 += 1;
}
tmp0 = alpha * tmp0;
tmp1 = alpha * tmp1;
#if defined(TRMMKERNEL)
pc0[0] = tmp0;
pc0[1] = tmp1;
@ -2366,6 +2386,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
pc0[0] += tmp0;
pc0[1] += tmp1;
#endif
pc0 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -2383,8 +2404,6 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
off += 2; // number of values in A
#endif
#endif
pc0 += 2;
}
if (m & 1)
@ -2448,34 +2467,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, FLOAT *A, FLOAT *B,
#else
pc0[0] += alpha * tmp0;
#endif
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
temp = k - off;
#ifdef LEFT
temp -= 1; // number of values in A
#else
temp -= 1; // number of values in B
#endif
pa0 += temp * 1;
pb0 += temp * 1;
#endif
#ifdef LEFT
off += 1; // number of values in A
#endif
#endif
pc0 += 1;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 1; // number of values in A
#endif
l = (k << 0);
B = B + l;
i = (ldc << 0);
C = C + i;
}
return 0;

View File

@ -851,6 +851,18 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
temp = k;
#endif
#ifdef ENABLE_PREFETCH
__asm__ __volatile__(
"pref 0, 64(%[pa0]) \n\t"
"pref 0, 96(%[pa0]) \n\t"
"pref 0, 64(%[pb0]) \n\t"
"pref 0, 96(%[pb0]) \n\t"
:
: [pa0] "r" (pa0), [pb0] "r" (pb0)
);
#endif
#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
ZGEMM_KERNEL_4X4_MSA(, -, , +, +);
#endif
@ -866,6 +878,18 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
for (l = (temp - 1); l--;)
{
#ifdef ENABLE_PREFETCH
__asm__ __volatile__(
"pref 0, 64(%[pa0]) \n\t"
"pref 0, 96(%[pa0]) \n\t"
"pref 0, 64(%[pb0]) \n\t"
"pref 0, 96(%[pb0]) \n\t"
:
: [pa0] "r" (pa0), [pb0] "r" (pb0)
);
#endif
#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
ZGEMM_KERNEL_4X4_MSA(+, -, +, +,);
#endif
@ -1039,6 +1063,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
ZGEMM_SCALE_1X4_MSA
#endif
pc0 += 2;
pc1 += 2;
pc2 += 2;
pc3 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1056,21 +1084,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 1; // number of values in A
#endif
#endif
pc0 += 2;
pc1 += 2;
pc2 += 2;
pc3 += 2;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 4; // number of values in A
#endif
l = k << 3;
B = B + l;
i = ldc << 3;
C = C + i;
B += (k << 3);
C += (ldc << 3);
}
if (n & 2)
@ -1294,6 +1315,8 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
ZGEMM_SCALE_1X2_MSA
#endif
pc0 += 2;
pc1 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1311,19 +1334,14 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 1; // number of values in A
#endif
#endif
pc0 += 2;
pc1 += 2;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 2; // number of values in A
#endif
l = k << 2;
B = B + l;
i = ldc << 2;
C = C + i;
B += (k << 2);
C += (ldc << 2);
}
if (n & 1)
@ -1555,6 +1573,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
#else
ZGEMM_SCALE_1X1
#endif
pc0 += 2;
#if defined(TRMMKERNEL)
#if (defined(LEFT) && defined(TRANSA)) || (!defined(LEFT) && !defined(TRANSA))
@ -1572,18 +1591,15 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alphar, FLOAT alphai,
off += 1; // number of values in A
#endif
#endif
pc0 += 2;
}
#if defined(TRMMKERNEL) && !defined(LEFT)
off += 1; // number of values in A
#endif
l = k << 1;
B = B + l;
i = ldc << 1;
C = C + i;
B += (k << 1);
C += (ldc << 1);
}
return 0;
}