Hugo MeilandandCursor
560caa56e6
TARGET=CORTEXA72 previously reused the A57 8x4 path. Add a dedicated 6x8 NEON ukernel, contiguous MR=6 panel packers (stock gemm_*copy_6 is 4+2), and UNROLL_M=6-aware TRSM kernels so HPL/dtrsm does not corrupt the heap. DTRMM falls back to generic 2x2 until a matching kernel exists. Co-authored-by: Cursor <cursoragent@cursor.com>
64 lines
1.5 KiB
C
64 lines
1.5 KiB
C
/* True MR=6 transpose pack (ITCOPY) for OpenBLAS DGEMM 6xN.
|
|
*
|
|
* Call signature matches gemm_tcopy_*: (m, n, a, lda, b) where for NN
|
|
* ICOPY uses ITCOPY(m=Kc, n=Mc). Output is Goto-style micropanels:
|
|
* for each panel of 6 rows: Kc contiguous packs of 6 doubles.
|
|
* Remainders after full panels: 4, then 2, then 1 (kernel cascade).
|
|
*
|
|
* Stock generic/gemm_tcopy_6.c is a clone of tcopy_4 (4+2) and does NOT
|
|
* match a contiguous 6-wide ukernel — do not use it for 6x8.
|
|
*/
|
|
#include "common.h"
|
|
|
|
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
|
|
{
|
|
BLASLONG j, js;
|
|
FLOAT *b_ptr = b;
|
|
BLASLONG n6 = n / 6;
|
|
BLASLONG nr = n - n6 * 6;
|
|
|
|
for (js = 0; js < n6; js++) {
|
|
FLOAT *a_row = a + js * 6;
|
|
for (j = 0; j < m; j++) {
|
|
FLOAT *ap = a_row + j * lda;
|
|
b_ptr[0] = ap[0];
|
|
b_ptr[1] = ap[1];
|
|
b_ptr[2] = ap[2];
|
|
b_ptr[3] = ap[3];
|
|
b_ptr[4] = ap[4];
|
|
b_ptr[5] = ap[5];
|
|
b_ptr += 6;
|
|
}
|
|
}
|
|
|
|
FLOAT *a_row = a + n6 * 6;
|
|
if (nr >= 4) {
|
|
for (j = 0; j < m; j++) {
|
|
FLOAT *ap = a_row + j * lda;
|
|
b_ptr[0] = ap[0];
|
|
b_ptr[1] = ap[1];
|
|
b_ptr[2] = ap[2];
|
|
b_ptr[3] = ap[3];
|
|
b_ptr += 4;
|
|
}
|
|
a_row += 4;
|
|
nr -= 4;
|
|
}
|
|
if (nr >= 2) {
|
|
for (j = 0; j < m; j++) {
|
|
FLOAT *ap = a_row + j * lda;
|
|
b_ptr[0] = ap[0];
|
|
b_ptr[1] = ap[1];
|
|
b_ptr += 2;
|
|
}
|
|
a_row += 2;
|
|
nr -= 2;
|
|
}
|
|
if (nr >= 1) {
|
|
for (j = 0; j < m; j++) {
|
|
*b_ptr++ = a_row[j * lda];
|
|
}
|
|
}
|
|
return 0;
|
|
}
|