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

Optimise sdot 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 16:43:31 +10:00
parent 32c7fe6bff
commit e0034de22d
1 changed files with 32 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

64
kernel/power/sdot_microk_power8.c
View File

@ -57,22 +57,22 @@ static float sdot_kernel_16 (long n, float *x, float *y)
"xxlxor 38, 38, 38 \n\t" "xxlxor 38, 38, 38 \n\t"
"xxlxor 39, 39, 39 \n\t" "xxlxor 39, 39, 39 \n\t"
"lxvw4x 40, 0, %2 \n\t" "lxvd2x 40, 0, %2 \n\t"
"lxvw4x 48, 0, %3 \n\t" "lxvd2x 48, 0, %3 \n\t"
"lxvw4x 41, %10, %2 \n\t" "lxvd2x 41, %10, %2 \n\t"
"lxvw4x 49, %10, %3 \n\t" "lxvd2x 49, %10, %3 \n\t"
"lxvw4x 42, %11, %2 \n\t" "lxvd2x 42, %11, %2 \n\t"
"lxvw4x 50, %11, %3 \n\t" "lxvd2x 50, %11, %3 \n\t"
"lxvw4x 43, %12, %2 \n\t" "lxvd2x 43, %12, %2 \n\t"
"lxvw4x 51, %12, %3 \n\t" "lxvd2x 51, %12, %3 \n\t"
"lxvw4x 44, %13, %2 \n\t" "lxvd2x 44, %13, %2 \n\t"
"lxvw4x %x4, %13, %3 \n\t" "lxvd2x %x4, %13, %3 \n\t"
"lxvw4x 45, %14, %2 \n\t" "lxvd2x 45, %14, %2 \n\t"
"lxvw4x %x5, %14, %3 \n\t" "lxvd2x %x5, %14, %3 \n\t"
"lxvw4x 46, %15, %2 \n\t" "lxvd2x 46, %15, %2 \n\t"
"lxvw4x %x6, %15, %3 \n\t" "lxvd2x %x6, %15, %3 \n\t"
"lxvw4x 47, %16, %2 \n\t" "lxvd2x 47, %16, %2 \n\t"
"lxvw4x %x7, %16, %3 \n\t" "lxvd2x %x7, %16, %3 \n\t"
"addi %2, %2, 128 \n\t" "addi %2, %2, 128 \n\t"
"addi %3, %3, 128 \n\t" "addi %3, %3, 128 \n\t"
@ -84,29 +84,29 @@ static float sdot_kernel_16 (long n, float *x, float *y)
"1: \n\t" "1: \n\t"
"xvmaddasp 32, 40, 48 \n\t" "xvmaddasp 32, 40, 48 \n\t"
"lxvw4x 40, 0, %2 \n\t" "lxvd2x 40, 0, %2 \n\t"
"lxvw4x 48, 0, %3 \n\t" "lxvd2x 48, 0, %3 \n\t"
"xvmaddasp 33, 41, 49 \n\t" "xvmaddasp 33, 41, 49 \n\t"
"lxvw4x 41, %10, %2 \n\t" "lxvd2x 41, %10, %2 \n\t"
"lxvw4x 49, %10, %3 \n\t" "lxvd2x 49, %10, %3 \n\t"
"xvmaddasp 34, 42, 50 \n\t" "xvmaddasp 34, 42, 50 \n\t"
"lxvw4x 42, %11, %2 \n\t" "lxvd2x 42, %11, %2 \n\t"
"lxvw4x 50, %11, %3 \n\t" "lxvd2x 50, %11, %3 \n\t"
"xvmaddasp 35, 43, 51 \n\t" "xvmaddasp 35, 43, 51 \n\t"
"lxvw4x 43, %12, %2 \n\t" "lxvd2x 43, %12, %2 \n\t"
"lxvw4x 51, %12, %3 \n\t" "lxvd2x 51, %12, %3 \n\t"
"xvmaddasp 36, 44, %x4 \n\t" "xvmaddasp 36, 44, %x4 \n\t"
"lxvw4x 44, %13, %2 \n\t" "lxvd2x 44, %13, %2 \n\t"
"lxvw4x %x4, %13, %3 \n\t" "lxvd2x %x4, %13, %3 \n\t"
"xvmaddasp 37, 45, %x5 \n\t" "xvmaddasp 37, 45, %x5 \n\t"
"lxvw4x 45, %14, %2 \n\t" "lxvd2x 45, %14, %2 \n\t"
"lxvw4x %x5, %14, %3 \n\t" "lxvd2x %x5, %14, %3 \n\t"
"xvmaddasp 38, 46, %x6 \n\t" "xvmaddasp 38, 46, %x6 \n\t"
"lxvw4x 46, %15, %2 \n\t" "lxvd2x 46, %15, %2 \n\t"
"lxvw4x %x6, %15, %3 \n\t" "lxvd2x %x6, %15, %3 \n\t"
"xvmaddasp 39, 47, %x7 \n\t" "xvmaddasp 39, 47, %x7 \n\t"
"lxvw4x 47, %16, %2 \n\t" "lxvd2x 47, %16, %2 \n\t"
"lxvw4x %x7, %16, %3 \n\t" "lxvd2x %x7, %16, %3 \n\t"
"addi %2, %2, 128 \n\t" "addi %2, %2, 128 \n\t"
"addi %3, %3, 128 \n\t" "addi %3, %3, 128 \n\t"