OpenBLAS/kernel/x86_64/sbgemm_kernel_16x16_spr_tmpl.c

530 lines
14 KiB
C

/***************************************************************************
* Copyright (c) 2021, The OpenBLAS Project
* All rights reserved.
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are
* met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in
* the documentation and/or other materials provided with the
* distribution.
* 3. Neither the name of the OpenBLAS project nor the names of
* its contributors may be used to endorse or promote products
* derived from this software without specific prior written permission.
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
* AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
* LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
* SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
* CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
* USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
* *****************************************************************************/
#include <immintrin.h>
#include <string.h>
#include "common.h"
#ifndef SBGEMM_KERNEL_SPR
#define SBGEMM_KERNEL_SPR
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;
/* tile0/tile1 -- A (m x 2k)
* tile2/tile3 -- B (2k x n)
* tile4-7 -- C (m x n)
*/
#define TCONF(cfg, m, n, k2) \
memset(&cfg, 0, sizeof(tilecfg)); \
cfg.palette_id = 1; \
cfg.tile_rows[0] = m; \
cfg.tile_rows[1] = m; \
cfg.tile_rows[2] = k2>>1; \
cfg.tile_rows[3] = k2>>1; \
cfg.tile_rows[4] = m; \
cfg.tile_rows[5] = m; \
cfg.tile_rows[6] = m; \
cfg.tile_rows[7] = m; \
cfg.tile_colsb[0] = k2<<1; \
cfg.tile_colsb[1] = k2<<1; \
cfg.tile_colsb[2] = n * 4; \
cfg.tile_colsb[3] = n * 4; \
cfg.tile_colsb[4] = n * 4; \
cfg.tile_colsb[5] = n * 4; \
cfg.tile_colsb[6] = n * 4; \
cfg.tile_colsb[7] = n * 4; \
_tile_loadconfig(&cfg);
/* CONFIG for handling k2 and odd tail at the same time
* tile0 -- A (m x 2k)
* tile1 -- A (m x 1)
* tile2 -- B (2k x n)
* tile3 -- B (1 x n)
* tile4 -- C (m x n)
*/
#define TCONF_TAIL(cfg, m, n, k2) \
memset(&cfg, 0, sizeof(tilecfg)); \
cfg.palette_id = 1; \
cfg.tile_rows[0] = m; \
cfg.tile_rows[1] = m; \
cfg.tile_rows[2] = k2>>1; \
cfg.tile_rows[3] = 1; \
cfg.tile_rows[4] = m; \
cfg.tile_colsb[0] = k2<<1; \
cfg.tile_colsb[1] = 4; \
cfg.tile_colsb[2] = n * 4; \
cfg.tile_colsb[3] = n * 4; \
cfg.tile_colsb[4] = n * 4; \
_tile_loadconfig(&cfg);
#define T_A0 0
#define T_A1 1
#define T_B0 2
#define T_B1 3
#define T_C00 4
#define T_C01 5
#define T_C10 6
#define T_C11 7
#define LOAD_A(M, N) _tile_loadd(T_A##M, ptr_a##M, lda * 2)
#define LOAD_A_TAIL(M, N) {\
__m256i ymm = _mm256_loadu_epi16(ptr_a##M); \
__m512i zmm = _mm512_cvtepu16_epi32(ymm); \
_mm512_storeu_epi16(tail_a + 16 * M, zmm); \
_tile_loadd(T_A##M, tail_a + 16 * M, 2 * 2); \
}
#define MASK_LOAD_A_TAIL(M, N) {\
__m256i ymm = _mm256_maskz_loadu_epi16(amask, ptr_a##M); \
__m512i zmm = _mm512_cvtepu16_epi32(ymm); \
_mm512_storeu_epi16(tail_a + 16 * M, zmm); \
_tile_loadd(T_A##M, tail_a + 16 * M, 2 * 2); \
}
#define LOAD_B(M, N) _tile_loadd(T_B##N, ptr_b##N, ldb * 2)
#define LOAD_B_TAIL(M, N) {\
__m256i ymm = _mm256_loadu_epi16(ptr_b##N); \
__m512i zmm = _mm512_cvtepu16_epi32(ymm); \
_mm512_storeu_epi16(tail_b + 16 * N, zmm); \
_tile_loadd(T_B##N, tail_b + 16 * N, 2 * 2); \
}
#define MASK_LOAD_B_TAIL(M, N) {\
__m256i ymm = _mm256_maskz_loadu_epi16(bmask, ptr_b##N); \
__m512i zmm = _mm512_cvtepu16_epi32(ymm); \
_mm512_storeu_epi16(tail_b + 16 * N, zmm); \
_tile_loadd(T_B##N, tail_b + 16 * N, 2 * 2); \
}
#define MATMUL(M, N) _tile_dpbf16ps(T_C##M##N, T_A##M, T_B##N)
#define MATMUL_TAIL(M, N) _tile_dpbf16ps(T_C00, T_A##M, T_B##N)
#define STORE_C(M, N) _tile_stored(T_C##M##N, ptr_c##M##N, ldc * 4)
#define LOAD_C_F(M, N) _tile_loadd(T_C##M##N, ptr_c##M##N, ldc * 4)
#endif // end of SBGEMM_KERNEL_SPR
#ifdef ALPHA_ONE
#undef LOAD_C
#define LOAD_C(M, N) _tile_loadd(T_C##M##N, ptr_c##M##N, ldc * 4)
#else
#undef LOAD_C
#define LOAD_C(M, N) _tile_zero(T_C##M##N)
#define ALPHA_STORE(N) \
__m512 zmm_d##N = _mm512_loadu_ps(dst##N + noffset); \
__m512 zmm_s##N = _mm512_loadu_ps(src##N + noffset); \
zmm_d##N = _mm512_fmadd_ps(alpha_512, zmm_s##N, zmm_d##N); \
_mm512_storeu_ps(dst##N + noffset, zmm_d##N);
#define MASK_APLPHA_STORE(N) \
__m512 zmm_d##N = _mm512_maskz_loadu_ps(mask, dst##N + noffset); \
__m512 zmm_s##N = _mm512_maskz_loadu_ps(mask, src##N + noffset); \
zmm_d##N = _mm512_fmadd_ps(alpha_512, zmm_s##N, zmm_d##N); \
_mm512_mask_storeu_ps(dst##N + noffset, mask, zmm_d##N);
#endif // end of ALPHA_ONE
#ifdef ALPHA_ONE
int sbgemm_kernel_spr_alpha_one(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT * A, IFLOAT * B, FLOAT * C, BLASLONG ldc)
#else
int sbgemm_kernel_spr_alpha(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT alpha, IFLOAT * A, IFLOAT * B, FLOAT * C, BLASLONG ldc)
#endif
{
/* Row Major matrix for AMX requirement */
IFLOAT *ptr_a = A, *ptr_b = B;
IFLOAT *ptr_b0, *ptr_b1;
IFLOAT *ptr_a0, *ptr_a1;
FLOAT *ptr_c = C;
FLOAT *ptr_c00, *ptr_c01, *ptr_c10, *ptr_c11;
BLASLONG lda, ldb;
BLASLONG m_count = m;
BLASLONG n_count, k_count;
#ifndef ALPHA_ONE
// make sure each row is 64 bytes aligned
BLASLONG cn = (n & 31) ? (n & ~31) + 32 : n;
FLOAT *raw_tmp_c;
if (k < 32) {
// only need to zero buff in this situation
raw_tmp_c = (FLOAT *)calloc(1, sizeof(FLOAT) * m * cn + 64);
} else {
raw_tmp_c = (FLOAT *)malloc(sizeof(FLOAT) * m * cn + 64);
}
// align buf to 64 byte boundary
FLOAT *tmp_c = (FLOAT *)(((uintptr_t) raw_tmp_c + 63) & ~(uintptr_t)63);
ptr_c = tmp_c;
BLASLONG ldc_o = ldc;
ldc = cn;
#endif
IFLOAT tail_a[32 * 2] __attribute__ ((aligned (64)));
IFLOAT tail_b[32 * 2] __attribute__ ((aligned (64)));
tilecfg cfg;
if (k > 31) {
for (; m_count > 31; m_count -= 32) {
ptr_b = B;
ptr_c00 = ptr_c;
ptr_c01 = ptr_c00 + 16;
ptr_c10 = ptr_c + 16 * ldc;
ptr_c11 = ptr_c10 + 16;
ptr_c += 32 * ldc;
n_count = n;
TCONF(cfg, 16, 16, 32);
for (; n_count > 31; n_count -= 32) {
ptr_a0 = ptr_a;
ptr_a1 = ptr_a + 16 * k;
ptr_b0 = ptr_b;
ptr_b1 = ptr_b + 16 * k;
ptr_b += 32 * k;
lda = 32;
ldb = 32;
LOAD_C(0, 0); LOAD_C(0, 1);
LOAD_C(1, 0); LOAD_C(1, 1);
k_count = k;
for (; k_count > 31; k_count -= 32) {
LOAD_A(0, x); LOAD_A(1, x);
ptr_a0 += 16 * 32;
ptr_a1 += 16 * 32;
LOAD_B(x, 0); LOAD_B(x, 1);
ptr_b0 += 16 * 32;
ptr_b1 += 16 * 32;
MATMUL(0, 0); MATMUL(0, 1);
MATMUL(1, 0); MATMUL(1, 1);
}
STORE_C(0, 0); STORE_C(0, 1);
STORE_C(1, 0); STORE_C(1, 1);
ptr_c00 += 32;
ptr_c01 += 32;
ptr_c10 += 32;
ptr_c11 += 32;
}
for (; n_count > 0; n_count -= 16) {
int tail_n = (n_count > 16) ? 16: n_count;
ptr_a0 = ptr_a;
ptr_a1 = ptr_a + 16 * k;
ptr_b0 = ptr_b;
ptr_b += tail_n * k;
lda = 32;
ldb = 2 * tail_n;
TCONF(cfg, 16, tail_n, 32);
LOAD_C(0, 0);
LOAD_C(1, 0);
k_count = k;
for (; k_count > 31; k_count -= 32) {
LOAD_A(0, x); LOAD_A(1, x);
ptr_a0 += 16 * 32;
ptr_a1 += 16 * 32;
LOAD_B(x, 0);
ptr_b0 += tail_n * 32;
MATMUL(0, 0);
MATMUL(1, 0);
}
STORE_C(0, 0);
STORE_C(1, 0);
ptr_c00 += tail_n;
ptr_c10 += tail_n;
}
ptr_a += 32 * k;
}
for (; m_count > 0; m_count -= 16) {
// process at most 16 m at a time
int tail_m = (m_count > 16) ? 16: m_count;
ptr_b = B;
ptr_c00 = ptr_c;
ptr_c01 = ptr_c00 + 16;
ptr_c += tail_m * ldc;
n_count = n;
TCONF(cfg, tail_m, 16, 32);
for (; n_count > 31; n_count -= 32) {
ptr_a0 = ptr_a;
ptr_b0 = ptr_b;
ptr_b1 = ptr_b + 16 * k;
ptr_b += 32 * k;
lda = 32;
ldb = 32;
LOAD_C(0, 0); LOAD_C(0, 1);
k_count = k;
for (; k_count > 31; k_count -= 32) {
LOAD_A(0, x);
ptr_a0 += tail_m * 32;
LOAD_B(x, 0); LOAD_B(x, 1);
ptr_b0 += 16 * 32;
ptr_b1 += 16 * 32;
MATMUL(0, 0); MATMUL(0, 1);
}
STORE_C(0, 0); STORE_C(0, 1);
ptr_c00 += 32;
ptr_c01 += 32;
}
for (; n_count > 0; n_count -= 16) {
int tail_n = (n_count > 16) ? 16: n_count;
ptr_a0 = ptr_a;
ptr_b0 = ptr_b;
ptr_b += tail_n * k;
lda = 32;
ldb = 2 * tail_n;
TCONF(cfg, tail_m, tail_n, 32);
LOAD_C(0, 0);
k_count = k;
for (; k_count > 31; k_count -= 32) {
LOAD_A(0, x);
ptr_a0 += tail_m * 32;
LOAD_B(x, 0);
ptr_b0 += tail_n * 32;
MATMUL(0, 0);
}
STORE_C(0, 0);
ptr_c00 += tail_n;
}
ptr_a += tail_m * k;
}
}
// process for k < 32
BLASLONG k32 = k & ~31;
BLASLONG k2 = k & ~1;
if (k32 != k) {
int remain_k2 = k2 - k32;
m_count = m;
ptr_a = A;
#ifndef ALPHA_ONE
ptr_c = tmp_c;
#else
ptr_c = C;
#endif
if (remain_k2 > 0 && k2 != k) { // k%32 = 2x + 1 (x != 0)
for (; m_count > 0; m_count -= 16) {
int tail_m = (m_count > 16) ? 16: m_count;
__mmask16 amask = (1UL << tail_m) - 1;
ptr_a0 = ptr_a + tail_m * k32;
ptr_a1 = ptr_a + tail_m * k2;
ptr_a += tail_m * k;
ptr_b = B;
ptr_c00 = ptr_c;
ptr_c += tail_m * ldc;
n_count = n;
lda = remain_k2;
ldb = 32;
if (n_count > 15) {
TCONF_TAIL(cfg, tail_m, 16, remain_k2);
LOAD_A(0, x); MASK_LOAD_A_TAIL(1, x);
for (; n_count > 15; n_count -= 16) {
ptr_b0 = ptr_b + 16 * k32;
ptr_b1 = ptr_b + 16 * k2;
LOAD_C_F(0, 0);
LOAD_B(x, 0); LOAD_B_TAIL(x, 1);
MATMUL(0, 0); MATMUL_TAIL(1, 1);
STORE_C(0, 0);
ptr_b += 16 * k;
ptr_c00 += 16;
}
}
if (n_count > 0) {
int tail_n = (n_count > 16) ? 16: n_count;
__mmask16 bmask = (1UL << tail_n) - 1;
ptr_b0 = ptr_b + tail_n * k32;
ptr_b1 = ptr_b + tail_n * k2;
ldb = 2 * tail_n;
TCONF_TAIL(cfg, tail_m, tail_n, remain_k2);
LOAD_C_F(0, 0);
LOAD_A(0, x); MASK_LOAD_A_TAIL(1, x);
LOAD_B(x, 0); MASK_LOAD_B_TAIL(x, 1);
MATMUL(0, 0); MATMUL_TAIL(1, 1);
STORE_C(0, 0);
}
}
} else if (remain_k2 > 0) { // k%32 = 2x
for (; m_count > 0; m_count -= 16) {
int tail_m = (m_count > 16) ? 16: m_count;
ptr_a0 = ptr_a + tail_m * k32;
ptr_a += tail_m * k;
ptr_b = B;
ptr_c00 = ptr_c;
ptr_c += tail_m * ldc;
n_count = n;
lda = remain_k2;
ldb = 32;
if (n_count > 15) {
TCONF(cfg, tail_m, 16, remain_k2);
LOAD_A(0, x);
for (; n_count > 15; n_count -= 16) {
ptr_b0 = ptr_b + 16 * k32;
LOAD_C_F(0, 0);
LOAD_B(x, 0);
MATMUL(0, 0);
STORE_C(0, 0);
ptr_b += 16 * k;
ptr_c00 += 16;
}
}
if (n_count > 0) {
int tail_n = (n_count > 16) ? 16: n_count;
ptr_b0 = ptr_b + tail_n * k32;
ldb = 2 * tail_n;
TCONF(cfg, tail_m, tail_n, remain_k2);
LOAD_C_F(0, 0);
LOAD_A(0, x);
LOAD_B(x, 0);
MATMUL(0, 0);
STORE_C(0, 0);
}
}
} else { // k%32 = 1
for (; m_count > 0; m_count -= 16) {
int tail_m = (m_count > 16) ? 16: m_count;
__mmask16 amask = (1UL << tail_m) - 1;
ptr_a0 = ptr_a + tail_m * k2;
ptr_a += tail_m * k;
ptr_b = B;
ptr_c00 = ptr_c;
ptr_c += tail_m * ldc;
n_count = n;
if (n_count > 15) {
TCONF(cfg, tail_m, 16, 2);
MASK_LOAD_A_TAIL(0, x);
for (; n_count > 15; n_count -= 16) {
ptr_b0 = ptr_b + 16 * k2;
LOAD_C_F(0, 0);
LOAD_B_TAIL(x, 0);
MATMUL(0, 0);
STORE_C(0, 0);
ptr_b += 16 * k;
ptr_c00 += 16;
}
}
if (n_count > 0) {
int tail_n = (n_count > 16) ? 16: n_count;
__mmask16 bmask = (1UL << tail_n) - 1;
ptr_b0 = ptr_b + tail_n * k2;
TCONF(cfg, tail_m, tail_n, 2);
LOAD_C_F(0, 0);
MASK_LOAD_A_TAIL(0, x);
MASK_LOAD_B_TAIL(x, 0);
MATMUL(0, 0);
STORE_C(0, 0);
}
}
}
}
#ifndef ALPHA_ONE
__m512 alpha_512 = _mm512_broadcastss_ps(_mm_load_ss(&alpha));
BLASLONG n16 = n & ~15;
BLASLONG noffset;
FLOAT *src0, *src1, *src2, *src3;
FLOAT *dst0, *dst1, *dst2, *dst3;
FLOAT *src = tmp_c;
FLOAT *dst = C;
m_count = m;
for (; m_count > 3; m_count -= 4) {
src0 = src;
src1 = src0 + ldc;
src2 = src1 + ldc;
src3 = src2 + ldc;
src += 4 * ldc;
dst0 = dst;
dst1 = dst0 + ldc_o;
dst2 = dst1 + ldc_o;
dst3 = dst2 + ldc_o;
dst += 4 * ldc_o;
noffset = 0;
for (; noffset < n16; noffset += 16) {
ALPHA_STORE(0);
ALPHA_STORE(1);
ALPHA_STORE(2);
ALPHA_STORE(3);
}
if (noffset < n) {
__mmask16 mask = (1UL << (n - noffset)) - 1;
MASK_APLPHA_STORE(0);
MASK_APLPHA_STORE(1);
MASK_APLPHA_STORE(2);
MASK_APLPHA_STORE(3);
}
}
for (; m_count > 1; m_count -= 2) {
src0 = src;
src1 = src0 + ldc;
src += 2 * ldc;
dst0 = dst;
dst1 = dst0 + ldc_o;
dst += 2 * ldc_o;
noffset = 0;
for (; noffset < n16; noffset += 16) {
ALPHA_STORE(0);
ALPHA_STORE(1);
}
if (noffset < n) {
__mmask16 mask = (1UL << (n - noffset)) - 1;
MASK_APLPHA_STORE(0);
MASK_APLPHA_STORE(1);
}
}
for (; m_count > 0; m_count -= 1) {
src0 = src;
dst0 = dst;
noffset = 0;
for (; noffset < n16; noffset += 16) {
ALPHA_STORE(0);
}
if (noffset < n) {
__mmask16 mask = (1UL << (n - noffset)) - 1;
MASK_APLPHA_STORE(0);
}
}
free(raw_tmp_c);
#endif
return 0;
}