floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update avx512 zgemm3m kernel

This commit is contained in:
wjc404 2019-12-02 16:01:35 +08:00 committed by GitHub
parent 87773b9be8
commit 7c52e0a567
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
kernel/x86_64/zgemm3m_kernel_8x4_skylakex.c
View File

@ -1,5 +1,5 @@
/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store */ /* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store */
/* r12 = k << 5(const), r13 = k(const), r14 = b_head_pos(const), r15 = %1 + 3r12 */ /* r11 = m(const), r12 = k << 5(const), r13 = k(const), r14 = b_head_pos(const), r15 = %1 + 3r12 */
#include "common.h" #include "common.h"
#include <stdint.h> #include <stdint.h>
@ -153,6 +153,7 @@
#define COMPUTE_m4_n20 COMPUTE_L_m4(12,33633) COMPUTE_R_m4(8,33933) #define COMPUTE_m4_n20 COMPUTE_L_m4(12,33633) COMPUTE_R_m4(8,33933)
#define COMPUTE_m4_n24 COMPUTE_L_m4(12,33533) COMPUTE_R_m4(12,33933) #define COMPUTE_m4_n24 COMPUTE_L_m4(12,33533) COMPUTE_R_m4(12,33933)
#define COMPUTE_m4(ndim) COMPUTE_m4_n##ndim #define COMPUTE_m4(ndim) COMPUTE_m4_n##ndim
/* m = 2 *//* vmm0 for alpha, vmm1-vmm3 for temporary use, vmm4-vmm15 for accumulators */ /* m = 2 *//* vmm0 for alpha, vmm1-vmm3 for temporary use, vmm4-vmm15 for accumulators */
#define KERNEL_k1m2n1 \ #define KERNEL_k1m2n1 \
"vmovupd (%0),%%xmm1; addq $16,%0;"\ "vmovupd (%0),%%xmm1; addq $16,%0;"\
@ -219,6 +220,7 @@
"decq %4; jmp "#ndim"002022b;"\ "decq %4; jmp "#ndim"002022b;"\
#ndim"002023:\n\t"\ #ndim"002023:\n\t"\
SAVE_m2(ndim) SAVE_m2(ndim)
/* m = 1 *//* vmm0 for alpha, vmm1-vmm3 and vmm10-vmm15 for temporary use, vmm4-vmm9 for accumulators */ /* m = 1 *//* vmm0 for alpha, vmm1-vmm3 and vmm10-vmm15 for temporary use, vmm4-vmm9 for accumulators */
#define KERNEL_k1m1n1 \ #define KERNEL_k1m1n1 \
"vmovsd (%0),%%xmm1; addq $8,%0;"\ "vmovsd (%0),%%xmm1; addq $8,%0;"\
@ -278,6 +280,7 @@
"decq %4; jmp "#ndim"001011b;"\ "decq %4; jmp "#ndim"001011b;"\
#ndim"001012:\n\t"\ #ndim"001012:\n\t"\
SAVE_m1(ndim) SAVE_m1(ndim)
#define COMPUTE(ndim) {\ #define COMPUTE(ndim) {\
next_b = b_pointer + ndim * K;\ next_b = b_pointer + ndim * K;\
__asm__ __volatile__(\ __asm__ __volatile__(\