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

Optimise cswap for POWER9 

Use lxvd2x instruction instead of lxvw4x.
lxvd2x performs far better on the new POWER architecture than lxvw4x.
This commit is contained in:
Matt Brown 2017-06-14 16:36:10 +10:00
parent 6f4eca5ea4
commit 4f09030fdc
1 changed files with 64 additions and 64 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

128
kernel/power/cswap_microk_power8.c
View File

@ -42,91 +42,91 @@ static void cswap_kernel_32 (long n, float *x, float *y)
".p2align 5 \n" ".p2align 5 \n"
"1: \n\t" "1: \n\t"
"lxvw4x 32, 0, %4 \n\t" "lxvd2x 32, 0, %4 \n\t"
"lxvw4x 33, %5, %4 \n\t" "lxvd2x 33, %5, %4 \n\t"
"lxvw4x 34, %6, %4 \n\t" "lxvd2x 34, %6, %4 \n\t"
"lxvw4x 35, %7, %4 \n\t" "lxvd2x 35, %7, %4 \n\t"
"lxvw4x 36, %8, %4 \n\t" "lxvd2x 36, %8, %4 \n\t"
"lxvw4x 37, %9, %4 \n\t" "lxvd2x 37, %9, %4 \n\t"
"lxvw4x 38, %10, %4 \n\t" "lxvd2x 38, %10, %4 \n\t"
"lxvw4x 39, %11, %4 \n\t" "lxvd2x 39, %11, %4 \n\t"
"addi %4, %4, 128 \n\t" "addi %4, %4, 128 \n\t"
"lxvw4x 40, 0, %4 \n\t" "lxvd2x 40, 0, %4 \n\t"
"lxvw4x 41, %5, %4 \n\t" "lxvd2x 41, %5, %4 \n\t"
"lxvw4x 42, %6, %4 \n\t" "lxvd2x 42, %6, %4 \n\t"
"lxvw4x 43, %7, %4 \n\t" "lxvd2x 43, %7, %4 \n\t"
"lxvw4x 44, %8, %4 \n\t" "lxvd2x 44, %8, %4 \n\t"
"lxvw4x 45, %9, %4 \n\t" "lxvd2x 45, %9, %4 \n\t"
"lxvw4x 46, %10, %4 \n\t" "lxvd2x 46, %10, %4 \n\t"
"lxvw4x 47, %11, %4 \n\t" "lxvd2x 47, %11, %4 \n\t"
"addi %4, %4, -128 \n\t" "addi %4, %4, -128 \n\t"
"lxvw4x 48, 0, %3 \n\t" "lxvd2x 48, 0, %3 \n\t"
"lxvw4x 49, %5, %3 \n\t" "lxvd2x 49, %5, %3 \n\t"
"lxvw4x 50, %6, %3 \n\t" "lxvd2x 50, %6, %3 \n\t"
"lxvw4x 51, %7, %3 \n\t" "lxvd2x 51, %7, %3 \n\t"
"lxvw4x 0, %8, %3 \n\t" "lxvd2x 0, %8, %3 \n\t"
"lxvw4x 1, %9, %3 \n\t" "lxvd2x 1, %9, %3 \n\t"
"lxvw4x 2, %10, %3 \n\t" "lxvd2x 2, %10, %3 \n\t"
"lxvw4x 3, %11, %3 \n\t" "lxvd2x 3, %11, %3 \n\t"
"addi %3, %3, 128 \n\t" "addi %3, %3, 128 \n\t"
"lxvw4x 4, 0, %3 \n\t" "lxvd2x 4, 0, %3 \n\t"
"lxvw4x 5, %5, %3 \n\t" "lxvd2x 5, %5, %3 \n\t"
"lxvw4x 6, %6, %3 \n\t" "lxvd2x 6, %6, %3 \n\t"
"lxvw4x 7, %7, %3 \n\t" "lxvd2x 7, %7, %3 \n\t"
"lxvw4x 8, %8, %3 \n\t" "lxvd2x 8, %8, %3 \n\t"
"lxvw4x 9, %9, %3 \n\t" "lxvd2x 9, %9, %3 \n\t"
"lxvw4x 10, %10, %3 \n\t" "lxvd2x 10, %10, %3 \n\t"
"lxvw4x 11, %11, %3 \n\t" "lxvd2x 11, %11, %3 \n\t"
"addi %3, %3, -128 \n\t" "addi %3, %3, -128 \n\t"
"stxvw4x 32, 0, %3 \n\t" "stxvd2x 32, 0, %3 \n\t"
"stxvw4x 33, %5, %3 \n\t" "stxvd2x 33, %5, %3 \n\t"
"stxvw4x 34, %6, %3 \n\t" "stxvd2x 34, %6, %3 \n\t"
"stxvw4x 35, %7, %3 \n\t" "stxvd2x 35, %7, %3 \n\t"
"stxvw4x 36, %8, %3 \n\t" "stxvd2x 36, %8, %3 \n\t"
"stxvw4x 37, %9, %3 \n\t" "stxvd2x 37, %9, %3 \n\t"
"stxvw4x 38, %10, %3 \n\t" "stxvd2x 38, %10, %3 \n\t"
"stxvw4x 39, %11, %3 \n\t" "stxvd2x 39, %11, %3 \n\t"
"addi %3, %3, 128 \n\t" "addi %3, %3, 128 \n\t"
"stxvw4x 40, 0, %3 \n\t" "stxvd2x 40, 0, %3 \n\t"
"stxvw4x 41, %5, %3 \n\t" "stxvd2x 41, %5, %3 \n\t"
"stxvw4x 42, %6, %3 \n\t" "stxvd2x 42, %6, %3 \n\t"
"stxvw4x 43, %7, %3 \n\t" "stxvd2x 43, %7, %3 \n\t"
"stxvw4x 44, %8, %3 \n\t" "stxvd2x 44, %8, %3 \n\t"
"stxvw4x 45, %9, %3 \n\t" "stxvd2x 45, %9, %3 \n\t"
"stxvw4x 46, %10, %3 \n\t" "stxvd2x 46, %10, %3 \n\t"
"stxvw4x 47, %11, %3 \n\t" "stxvd2x 47, %11, %3 \n\t"
"addi %3, %3, 128 \n\t" "addi %3, %3, 128 \n\t"
"stxvw4x 48, 0, %4 \n\t" "stxvd2x 48, 0, %4 \n\t"
"stxvw4x 49, %5, %4 \n\t" "stxvd2x 49, %5, %4 \n\t"
"stxvw4x 50, %6, %4 \n\t" "stxvd2x 50, %6, %4 \n\t"
"stxvw4x 51, %7, %4 \n\t" "stxvd2x 51, %7, %4 \n\t"
"stxvw4x 0, %8, %4 \n\t" "stxvd2x 0, %8, %4 \n\t"
"stxvw4x 1, %9, %4 \n\t" "stxvd2x 1, %9, %4 \n\t"
"stxvw4x 2, %10, %4 \n\t" "stxvd2x 2, %10, %4 \n\t"
"stxvw4x 3, %11, %4 \n\t" "stxvd2x 3, %11, %4 \n\t"
"addi %4, %4, 128 \n\t" "addi %4, %4, 128 \n\t"
"stxvw4x 4, 0, %4 \n\t" "stxvd2x 4, 0, %4 \n\t"
"stxvw4x 5, %5, %4 \n\t" "stxvd2x 5, %5, %4 \n\t"
"stxvw4x 6, %6, %4 \n\t" "stxvd2x 6, %6, %4 \n\t"
"stxvw4x 7, %7, %4 \n\t" "stxvd2x 7, %7, %4 \n\t"
"stxvw4x 8, %8, %4 \n\t" "stxvd2x 8, %8, %4 \n\t"
"stxvw4x 9, %9, %4 \n\t" "stxvd2x 9, %9, %4 \n\t"
"stxvw4x 10, %10, %4 \n\t" "stxvd2x 10, %10, %4 \n\t"
"stxvw4x 11, %11, %4 \n\t" "stxvd2x 11, %11, %4 \n\t"
"addi %4, %4, 128 \n\t" "addi %4, %4, 128 \n\t"