OmniNX/Horizon-OC
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

TinyMemBenchNX multi-threads support, but max_threads = 2 ?

Browse Source
This commit is contained in:
KazushiM
2021-10-25 22:20:36 +08:00
parent 5797bc0af7
commit e947b47598
6 changed files with 239 additions and 138 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
Source/TinyMemBenchNX/source/aarch64-asm.s
View File

@@ -23,7 +23,7 @@
#ifdef __aarch64__
.cpu cortex-a53+fp+simd
.cpu cortex-a57+fp+simd
.text
.align 2
@@ -39,6 +39,18 @@
SIZE .req x2
.endm
asm_function aligned_block_read_ldp_x_aarch64
0:
ldp x3, x4, [DST, #(0 * 16)]
ldp x5, x6, [DST, #(1 * 16)]
ldp x7, x8, [DST, #(2 * 16)]
ldp x9, x10, [DST, #(3 * 16)]
add DST, DST, #64
subs SIZE, SIZE, #64
bgt 0b
ret
.endfunc
asm_function aligned_block_copy_ldpstp_x_aarch64
0:
ldp x3, x4, [SRC, #(0 * 16)]
@@ -56,6 +68,16 @@ asm_function aligned_block_copy_ldpstp_x_aarch64
ret
.endfunc
asm_function aligned_block_read_ldp_q_aarch64
0:
ldp q0, q1, [DST, #(0 * 32)]
ldp q2, q3, [DST, #(1 * 32)]
add DST, DST, #64
subs SIZE, SIZE, #64
bgt 0b
ret
.endfunc
asm_function aligned_block_copy_ldpstp_q_aarch64
0:
ldp q0, q1, [SRC, #(0 * 32)]