sravan/system76-edk2
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

BaseMemoryLibSse2: Take advantage of write combining buffers

Browse Source 
The current SSE2 implementation of the ZeroMem(), SetMem(),
SetMem16(), SetMem32 and SetMem64 functions is writing 16 bytes per 16
bytes. It hurts the performances so bad that this is even slower than
a simple 'rep stos' (4% slower) in regular DRAM.

To take full advantages of the 'movntdq' instruction it is better to
"queue" a total of 64 bytes in the write combining buffers.  This
patch implement such a change.  Below is a table where I measured
(with 'rdtsc') the time to write an entire 100MB RAM buffer. These
functions operate almost two times faster.

| Function | Arch | Untouched | 64 bytes | Result |
|----------+------+-----------+----------+--------|
| ZeroMem  | Ia32 |  17765947 |  9136062 | 1.945x |
| ZeroMem  | X64  |  17525170 |  9233391 | 1.898x |
| SetMem   | Ia32 |  17522291 |  9137272 | 1.918x |
| SetMem   | X64  |  17949261 |  9176978 | 1.956x |
| SetMem16 | Ia32 |  18219673 |  9372062 | 1.944x |
| SetMem16 | X64  |  17523331 |  9275184 | 1.889x |
| SetMem32 | Ia32 |  18495036 |  9273053 | 1.994x |
| SetMem32 | X64  |  17368864 |  9285885 | 1.870x |
| SetMem64 | Ia32 |  18564473 |  9241362 | 2.009x |
| SetMem64 | X64  |  17506951 |  9280148 | 1.886x |

Signed-off-by: Jeremy Compostella <jeremy.compostella@intel.com>
Reviewed-by: Liming Gao <gaoliming@byosoft.com.cn>
This commit is contained in:
Compostella, Jeremy
2020-10-10 04:42:34 +08:00
committed by mergify[bot]
parent 19c87b7d44
commit d25fd8710d
10 changed files with 85 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
MdePkg/Library/BaseMemoryLibSse2/Ia32/SetMem64.nasm
View File

@@ -38,17 +38,29 @@ ASM_PFX(InternalMemSetMem64):
add edx, 8
dec ecx
.0:
shr ecx, 1
push ebx
mov ebx, ecx
and ebx, 7
shr ecx, 3
jz @SetQwords
movlhps xmm0, xmm0
.1:
movntdq [edx], xmm0
lea edx, [edx + 16]
movntdq [edx + 16], xmm0
movntdq [edx + 32], xmm0
movntdq [edx + 48], xmm0
lea edx, [edx + 64]
loop .1
mfence
@SetQwords:
jnc .2
test ebx, ebx
jz .3
mov ecx, ebx
.2
movq qword [edx], xmm0
.2:
lea edx, [edx + 8]
loop .2
.3:
pop ebx
ret