Skip to content

Commit beccb83

Browse files
guowangyguowangy
committed
sbgemm: cooperlake: add n24 kernel for tcopy_4
1 parent 5fcacad commit beccb83

1 file changed

Lines changed: 99 additions & 2 deletions

File tree

‎kernel/x86_64/sbgemm_tcopy_4_cooperlake.c‎

Lines changed: 99 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -29,20 +29,117 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
2929
#include <immintrin.h>
3030
#include "common.h"
3131

32+
#define STORE_VEC(Bx, By, vec) \
33+
if (By == 0) asm("vmovdqu16 %0, (%1)": : "v"(vec), "r"(boffset##Bx)); \
34+
else asm("vmovdqu16 %0, (%1, %2, %c3)": : "v"(vec), "r"(boffset##Bx), "r"(blk_size), "n"(By * 2));
35+
3236
int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
3337
BLASLONG i, j;
3438

3539
IFLOAT *boffset0, *boffset1;
3640

3741
boffset0 = b;
3842

43+
BLASLONG n24 = n - (n % 24);
3944
BLASLONG n8 = n & ~7;
45+
BLASLONG m8 = m & ~7;
4046
BLASLONG m4 = m & ~3;
4147
BLASLONG m2 = m & ~1;
4248

43-
for (j = 0; j < n8; j += 8) {
49+
int permute_table[] = {
50+
0x0, 0x1, 0x2, 0x3, 0x10, 0x11, 0x12, 0x13, 0x8, 0x9, 0xa, 0xb, 0x18, 0x19, 0x1a, 0x1b,
51+
0x4, 0x5, 0x6, 0x7, 0x14, 0x15, 0x16, 0x17, 0xc, 0xd, 0xe, 0xf, 0x1c, 0x1d, 0x1e, 0x1f,
52+
0x0, 0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x10, 0x11, 0x12, 0x13, 0x14, 0x15, 0x16, 0x17,
53+
0x8, 0x9, 0xa, 0xb, 0xc, 0xd, 0xe, 0xf, 0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f,
54+
};
55+
56+
j = 0;
57+
if (n > 23) {
58+
/* n = 24 is the max width in current blocking setting */
59+
__m512i idx_lo_128 = _mm512_loadu_si512(permute_table);
60+
__m512i idx_hi_128 = _mm512_loadu_si512(permute_table + 16);
61+
__m512i idx_lo_256 = _mm512_loadu_si512(permute_table + 32);
62+
__m512i idx_hi_256 = _mm512_loadu_si512(permute_table + 48);
63+
__mmask32 mask24 = (1UL << 24) - 1;
64+
BLASLONG blk_size = m * 4;
65+
BLASLONG stride = blk_size * 3;
66+
67+
for (; j < n24; j += 24) {
68+
boffset1 = boffset0 + stride;
69+
for (i = 0; i < m8; i += 8) {
70+
__m512i r0, r1, r2, r3, r4, r5, r6, r7;
71+
__m512i t0, t1, t2, t3, t4, t5, t6, t7;
72+
r0 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 0)*lda + j]);
73+
r1 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 1)*lda + j]);
74+
r2 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 2)*lda + j]);
75+
r3 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 3)*lda + j]);
76+
r4 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 4)*lda + j]);
77+
r5 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 5)*lda + j]);
78+
r6 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 6)*lda + j]);
79+
r7 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 7)*lda + j]);
80+
81+
t0 = _mm512_unpacklo_epi16(r0, r1);
82+
t1 = _mm512_unpackhi_epi16(r0, r1);
83+
t2 = _mm512_unpacklo_epi16(r2, r3);
84+
t3 = _mm512_unpackhi_epi16(r2, r3);
85+
t4 = _mm512_unpacklo_epi16(r4, r5);
86+
t5 = _mm512_unpackhi_epi16(r4, r5);
87+
t6 = _mm512_unpacklo_epi16(r6, r7);
88+
t7 = _mm512_unpackhi_epi16(r6, r7);
89+
90+
r0 = _mm512_permutex2var_epi32(t0, idx_lo_128, t2);
91+
r1 = _mm512_permutex2var_epi32(t1, idx_lo_128, t3);
92+
r2 = _mm512_permutex2var_epi32(t4, idx_lo_128, t6);
93+
r3 = _mm512_permutex2var_epi32(t5, idx_lo_128, t7);
94+
r4 = _mm512_permutex2var_epi32(t0, idx_hi_128, t2);
95+
r5 = _mm512_permutex2var_epi32(t1, idx_hi_128, t3);
96+
r6 = _mm512_permutex2var_epi32(t4, idx_hi_128, t6);
97+
r7 = _mm512_permutex2var_epi32(t5, idx_hi_128, t7);
98+
99+
t0 = _mm512_permutex2var_epi32(r0, idx_lo_256, r2);
100+
t1 = _mm512_permutex2var_epi32(r1, idx_lo_256, r3);
101+
t2 = _mm512_permutex2var_epi32(r4, idx_lo_256, r6);
102+
t3 = _mm512_permutex2var_epi32(r5, idx_lo_256, r7);
103+
t4 = _mm512_permutex2var_epi32(r0, idx_hi_256, r2);
104+
t5 = _mm512_permutex2var_epi32(r1, idx_hi_256, r3);
105+
106+
STORE_VEC(0, 0, t0); STORE_VEC(0, 1, t1); STORE_VEC(0, 2, t2);
107+
STORE_VEC(1, 0, t3); STORE_VEC(1, 1, t4); STORE_VEC(1, 2, t5);
108+
boffset0 += 32;
109+
boffset1 += 32;
110+
}
111+
for (; i < m2; i += 2) {
112+
__m512i r0, r1, t0, t1;
113+
r0 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 0)*lda + j]);
114+
r1 = _mm512_maskz_loadu_epi16(mask24, &a[(i + 1)*lda + j]);
115+
t0 = _mm512_unpacklo_epi16(r0, r1);
116+
t1 = _mm512_unpackhi_epi16(r0, r1);
117+
STORE_VEC(0, 0, _mm512_extracti32x4_epi32(t0, 0));
118+
STORE_VEC(0, 1, _mm512_extracti32x4_epi32(t1, 0));
119+
STORE_VEC(0, 2, _mm512_extracti32x4_epi32(t0, 1));
120+
STORE_VEC(1, 0, _mm512_extracti32x4_epi32(t1, 1));
121+
STORE_VEC(1, 1, _mm512_extracti32x4_epi32(t0, 2));
122+
STORE_VEC(1, 2, _mm512_extracti32x4_epi32(t1, 2));
123+
boffset0 += 8;
124+
boffset1 += 8;
125+
}
126+
for (; i < m; i++) {
127+
*(uint64_t *)(boffset0 + blk_size * 0) = *(uint64_t *)&a[i * lda + j + 0];
128+
*(uint64_t *)(boffset0 + blk_size * 1) = *(uint64_t *)&a[i * lda + j + 4];
129+
*(uint64_t *)(boffset0 + blk_size * 2) = *(uint64_t *)&a[i * lda + j + 8];
130+
*(uint64_t *)(boffset1 + blk_size * 0) = *(uint64_t *)&a[i * lda + j + 12];
131+
*(uint64_t *)(boffset1 + blk_size * 1) = *(uint64_t *)&a[i * lda + j + 16];
132+
*(uint64_t *)(boffset1 + blk_size * 2) = *(uint64_t *)&a[i * lda + j + 20];
133+
boffset0 += 4;
134+
boffset1 += 4;
135+
}
136+
boffset0 += stride * 2;
137+
}
138+
}
139+
140+
for (; j < n8; j += 8) {
44141
boffset1 = boffset0 + m * 4;
45-
for (i = 0; i < m4; i +=4) {
142+
for (i = 0; i < m4; i += 4) {
46143
__m128i a0 = _mm_loadu_si128((void *)&a[(i + 0)*lda + j]);
47144
__m128i a1 = _mm_loadu_si128((void *)&a[(i + 1)*lda + j]);
48145
__m128i a2 = _mm_loadu_si128((void *)&a[(i + 2)*lda + j]);

0 commit comments

Comments
 (0)