sbgemm: spr: oncopy: use tile load/store instead

This commit is contained in:
Wangyang Guo 2021-09-16 20:08:42 -07:00
parent f2485352a6
commit a52456b168
1 changed files with 59 additions and 121 deletions

View File

@ -28,18 +28,45 @@
#include <immintrin.h>
#include "common.h"
#define COPY_32(N) _mm512_storeu_si512(boffset + 32 * N, _mm512_loadu_si512(aoffset##N + i))
#define MASK_COPY_32(N) _mm512_mask_storeu_epi16(boffset + tail_m * N, mmask, _mm512_maskz_loadu_epi16(mmask, aoffset##N + i))
#define COPY_ODD_TAIL(N) *(boffset + N) = *(aoffset##N + i);
typedef struct {
char palette_id;
char start_row;
char dummy0[14]; // bytes 2-15 reserved, must be zero
short tile_colsb[8];
char dummy1[16]; // bytes 32-47 reserved, must be zero
char tile_rows[8];
char dummy2[16]; // bytes 56-63 reserved, must be zero
} tilecfg;
#define T_16x32 0
#define T_16xm 1
#define T_nx32 2
#define T_nxm 3
#define TCONF(cfg, m, n) \
memset(&cfg, 0, sizeof(tilecfg)); \
cfg.palette_id = 1; \
cfg.tile_rows[T_16x32] = 16; \
cfg.tile_colsb[T_16x32] = 64; \
if (m) { \
cfg.tile_rows[T_16xm] = 16; \
cfg.tile_colsb[T_16xm] = m * 2; \
} \
if (n) { \
cfg.tile_rows[T_nx32] = n; \
cfg.tile_colsb[T_nx32] = 64; \
} \
if (m && n) { \
cfg.tile_rows[T_nxm] = n; \
cfg.tile_colsb[T_nxm] = m * 2; \
} \
_tile_loadconfig(&cfg);
int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
BLASLONG i, j;
IFLOAT *aoffset, *boffset;
IFLOAT *aoffset0, *aoffset1, *aoffset2, *aoffset3;
IFLOAT *aoffset4, *aoffset5, *aoffset6, *aoffset7;
IFLOAT *aoffset8, *aoffset9, *aoffset10, *aoffset11;
IFLOAT *aoffset12, *aoffset13, *aoffset14, *aoffset15;
IFLOAT *aoffset0;
aoffset = a;
boffset = b;
@ -48,141 +75,52 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b) {
BLASLONG m32 = m & ~31;
BLASLONG m2 = m & ~1;
BLASLONG tail_m = m2 - m32;
BLASLONG tail_n = n - n16;
tilecfg cfg;
TCONF(cfg, tail_m, tail_n);
for (j = 0; j < n16; j += 16) {
IFLOAT *boffset0 = boffset;
aoffset0 = aoffset;
aoffset1 = aoffset0 + lda;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
for (i = 0; i < m32; i += 32) {
COPY_32(0); COPY_32(1); COPY_32(2); COPY_32(3);
boffset += 32 * 16;
}
aoffset4 = aoffset3 + lda;
aoffset5 = aoffset4 + lda;
aoffset6 = aoffset5 + lda;
aoffset7 = aoffset6 + lda;
boffset = boffset0;
for (i = 0; i < m32; i += 32) {
COPY_32(4); COPY_32(5); COPY_32(6); COPY_32(7);
boffset += 32 * 16;
}
aoffset8 = aoffset7 + lda;
aoffset9 = aoffset8 + lda;
aoffset10 = aoffset9 + lda;
aoffset11 = aoffset10 + lda;
boffset = boffset0;
for (i = 0; i < m32; i += 32) {
COPY_32(8); COPY_32(9); COPY_32(10); COPY_32(11);
boffset += 32 * 16;
}
aoffset12 = aoffset11 + lda;
aoffset13 = aoffset12 + lda;
aoffset14 = aoffset13 + lda;
aoffset15 = aoffset14 + lda;
boffset = boffset0;
for (i = 0; i < m32; i += 32) {
COPY_32(12); COPY_32(13); COPY_32(14); COPY_32(15);
_tile_loadd(T_16x32, aoffset0, lda * 2);
_tile_stored(T_16x32, boffset, 32 * 2);
aoffset0 += 32;
boffset += 32 * 16;
}
if (i < m2) {
int tail_m = m2 - i;
__mmask32 mmask = (1UL << tail_m) - 1;
MASK_COPY_32(0); MASK_COPY_32(1); MASK_COPY_32(2); MASK_COPY_32(3);
MASK_COPY_32(4); MASK_COPY_32(5); MASK_COPY_32(6); MASK_COPY_32(7);
MASK_COPY_32(8); MASK_COPY_32(9); MASK_COPY_32(10); MASK_COPY_32(11);
MASK_COPY_32(12); MASK_COPY_32(13); MASK_COPY_32(14); MASK_COPY_32(15);
i = m2;
_tile_loadd(T_16xm, aoffset0, lda * 2);
_tile_stored(T_16xm, boffset, tail_m * 2);
aoffset0 += tail_m;
boffset += tail_m * 16;
i = m2;
}
if (i < m) {
/* the tail odd k should put alone */
COPY_ODD_TAIL(0); COPY_ODD_TAIL(1); COPY_ODD_TAIL(2); COPY_ODD_TAIL(3);
COPY_ODD_TAIL(4); COPY_ODD_TAIL(5); COPY_ODD_TAIL(6); COPY_ODD_TAIL(7);
COPY_ODD_TAIL(8); COPY_ODD_TAIL(9); COPY_ODD_TAIL(10); COPY_ODD_TAIL(11);
COPY_ODD_TAIL(12); COPY_ODD_TAIL(13); COPY_ODD_TAIL(14); COPY_ODD_TAIL(15);
for (int ii = 0; ii < 16; ii++) {
*(boffset + ii) = *(aoffset0 + lda * ii);
}
boffset += 16;
}
aoffset += 16 * lda;
}
if (j < n) {
int remain_n = n - j;
aoffset0 = aoffset;
aoffset1 = aoffset0 + lda;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset5 = aoffset4 + lda;
aoffset6 = aoffset5 + lda;
aoffset7 = aoffset6 + lda;
aoffset8 = aoffset7 + lda;
aoffset9 = aoffset8 + lda;
aoffset10 = aoffset9 + lda;
aoffset11 = aoffset10 + lda;
aoffset12 = aoffset11 + lda;
aoffset13 = aoffset12 + lda;
aoffset14 = aoffset13 + lda;
aoffset15 = aoffset14 + lda;
for (i = 0; i < m32; i += 32) {
switch(remain_n) {
case 15: COPY_32(14);
case 14: COPY_32(13);
case 13: COPY_32(12);
case 12: COPY_32(11);
case 11: COPY_32(10);
case 10: COPY_32(9);
case 9: COPY_32(8);
case 8: COPY_32(7);
case 7: COPY_32(6);
case 6: COPY_32(5);
case 5: COPY_32(4);
case 4: COPY_32(3);
case 3: COPY_32(2);
case 2: COPY_32(1);
case 1: COPY_32(0);
}
boffset += 32 * remain_n;
_tile_loadd(T_nx32, aoffset0, lda * 2);
_tile_stored(T_nx32, boffset, 32 * 2);
aoffset0 += 32;
boffset += 32 * tail_n;
}
if (i < m2) {
int tail_m = m2 - i;
__mmask32 mmask = (1UL << tail_m) - 1;
switch(remain_n) {
case 15: MASK_COPY_32(14);
case 14: MASK_COPY_32(13);
case 13: MASK_COPY_32(12);
case 12: MASK_COPY_32(11);
case 11: MASK_COPY_32(10);
case 10: MASK_COPY_32(9);
case 9: MASK_COPY_32(8);
case 8: MASK_COPY_32(7);
case 7: MASK_COPY_32(6);
case 6: MASK_COPY_32(5);
case 5: MASK_COPY_32(4);
case 4: MASK_COPY_32(3);
case 3: MASK_COPY_32(2);
case 2: MASK_COPY_32(1);
case 1: MASK_COPY_32(0);
}
i = m2;
boffset += tail_m * remain_n;
_tile_loadd(T_nxm, aoffset0, lda * 2);
_tile_stored(T_nxm, boffset, tail_m * 2);
aoffset0 += tail_m;
boffset += tail_m * tail_n;
}
if (i < m) {
switch(remain_n) {
case 15: COPY_ODD_TAIL(14);
case 14: COPY_ODD_TAIL(13);
case 13: COPY_ODD_TAIL(12);
case 12: COPY_ODD_TAIL(11);
case 11: COPY_ODD_TAIL(10);
case 10: COPY_ODD_TAIL(9);
case 9: COPY_ODD_TAIL(8);
case 8: COPY_ODD_TAIL(7);
case 7: COPY_ODD_TAIL(6);
case 6: COPY_ODD_TAIL(5);
case 5: COPY_ODD_TAIL(4);
case 4: COPY_ODD_TAIL(3);
case 3: COPY_ODD_TAIL(2);
case 2: COPY_ODD_TAIL(1);
case 1: COPY_ODD_TAIL(0);
for (int ii = 0; ii < tail_n; ii++) {
*(boffset + ii) = *(aoffset0 + lda * ii);
}
}
}