AVX512 support?

[missdeer]

The AVX2 edition is super fast, is it possible to have an AVX512 edition?

[zufuliu]

AVX512 is uncommon in desktop CPU (especially Intel CPUs after 12 Gen).

[missdeer]

https://github.com/zufuliu/notepad4/pull/1006

[zufuliu]

Build configuration for AVX512 is added by commit 01a16c5677325c098bf0a837a0738b98b9878ef9. it's tentative, will be replaced with AVX10.X (e.g. AVX10.2) in the further when the later become popular in desktop CPUs.

[zufuliu]

Some code (line endings detection, brace matching, etc.) has been port to AVX512 (guarded by NP2_USE_AVX512), the remaining (UTF-8 validation and image processing) will be added later. I don't have access to computer that supports AVX512, so not sure whether these AVX512 code is faster than their AVX2 counterpart.

[missdeer]

I have a Windows 11 box supports AVX512, I would like help to test the new feature if you can provide efficient test cases.

[zufuliu]

Just test big files (hundred MBs with word wrap off and LF / CR+LF line endings) on your system.

1. enable console output, uncomment line above AttachConsole() inside wWinMain() in Notepad4.cpp: https://github.com/zufuliu/notepad4/blob/07874e0c9c6624ad15ae14ac664415aa4d6998ee/src/Notepad4.cpp#L490

2. EOL detection, uncomment code at end of EditLoadFile(): https://github.com/zufuliu/notepad4/blob/07874e0c9c6624ad15ae14ac664415aa4d6998ee/src/Edit.cpp#L1173-L1180

3. brace matching, uncomment code for case Message::BraceMatch: https://github.com/zufuliu/notepad4/blob/07874e0c9c6624ad15ae14ac664415aa4d6998ee/scintilla/src/Editor.cxx#L8219-L8227

4. insert string, uncomment code inside CellBuffer::InsertString() or CellBuffer::BasicInsertString(): https://github.com/zufuliu/notepad4/blob/07874e0c9c6624ad15ae14ac664415aa4d6998ee/scintilla/src/CellBuffer.cxx#L446-L449 https://github.com/zufuliu/notepad4/blob/07874e0c9c6624ad15ae14ac664415aa4d6998ee/scintilla/src/CellBuffer.cxx#L827 https://github.com/zufuliu/notepad4/blob/07874e0c9c6624ad15ae14ac664415aa4d6998ee/scintilla/src/CellBuffer.cxx#L1213-L1215 for BasicInsertString(), AVX512 is currently disabled due to stack (compiled with cl /utf-8 /W4 /c /EHsc /std:c++20 /O2 /GS- /GR- /Gv /FAcs /DNDEBUG /DUNICODE /DNOMINMAX /I../include /I../lexlib /arch:AVX512 CellBuffer.cxx) is 200 bytes larger than AVX2: https://github.com/zufuliu/notepad4/blob/07874e0c9c6624ad15ae14ac664415aa4d6998ee/scintilla/src/CellBuffer.cxx#L844

[zufuliu]

Here is the script I used to make big files:

import json

path = r'D:\Program Files\Microsoft Visual Studio\Packages\_Instances\953ef33b\catalog.json'
with open(path, encoding='utf-8') as fd:
	doc = fd.read()
catalog = json.loads(doc)

with open('catalog.json', 'w', encoding='utf-8') as fd:
	fd.write(doc)

pretty = json.dumps(catalog, ensure_ascii=False, indent='\t')
pretty = '\n' + pretty + '\n'
with open('pretty.json', 'w', encoding='utf-8') as fd:
	fd.write(pretty)

with open('catalog2.json', 'w', encoding='utf-8') as fd:
	fd.write(doc)
	fd.write(pretty)

big = {}
for i in range(30):
	big[i] = catalog
pretty = json.dumps(big, ensure_ascii=False, indent='\t')
pretty = '\n[\n(\n<\n' + pretty + '\n>\n)\n]\n'
with open('big.json', 'w', encoding='utf-8') as fd:
	fd.write(pretty)

pretty = pretty.encode('gbk', 'backslashreplace')
with open('big-gbk.json', 'wb') as fd:
	fd.write(pretty)

[missdeer]

Got it. I will try to run these test cases recently.

[missdeer]

simple quick test result:

>notepad4-avx2.exe big.json

C:\Users\HP\Downloads\notepad4-test>
H:\Projects\notepad2\src\Notepad4.cpp:495 wWinMain
Notepad4 EOL time: 37.976700
CR+LF: 17837319, LF: 0, CR: 0
BasicInsertString avx2=1, cache=256, perLine=1023, duration=58.045100
InsertString duration=145.876400
BraceMatch 2 / -1, 16606350 / 677007975, 28.856600
BraceMatch 5 / 23182016, 23181522 / 677007975, 0.955200
BraceMatch 8 / 677007966, 36735725 / 677007975, 16.262400
BraceMatch 11 / 677007963, 38494926 / 677007975, 37.358900
BraceMatch 20 / 22566941, 44255015 / 677007975, 2.368700
BraceMatch 674 / 2436, 65692067 / 677007975, 0.000600
BraceMatch 680 / 814, 78084145 / 677007975, 0.000500
BraceMatch 2454 / 22543705, 154572928 / 677007975, 3.739700
BraceMatch 2460 / 4534, 156856227 / 677007975, 0.001000
BraceMatch 22543724 / 22543933, 170396159 / 677007975, 0.000400
BraceMatch 22543952 / 22566937, 173210459 / 677007975, 0.006800
BraceMatch 22543970 / 22544144, 175255517 / 677007975, 0.000600

======================================================================
notepad4-avx2.exe big-gbk.json

H:\Projects\notepad2\src\Notepad4.cpp:495 wWinMain
Notepad4 EOL time: 26.452200
CR+LF: 0, LF: 17837319, CR: 0
BasicInsertString avx2=1, cache=256, perLine=1023, duration=50.733500
InsertString duration=130.058000
BraceMatch 1 / -1, 37170555 / 663802026, 37.276200
BraceMatch 3 / 52713878, 52298785 / 663802026, 2.221700
BraceMatch 5 / 663802020, 59846479 / 663802026, 19.668100
BraceMatch 7 / 663802018, 80585195 / 663802026, 42.052400
BraceMatch 5 / 663802020, 85119126 / 663802026, 31.121200
BraceMatch 1 / -1, 95948811 / 663802026, 39.473900
BraceMatch 15 / 22126739, 115442847 / 663802026, 4.701800
BraceMatch 633 / 91, 165515441 / 663802026, 0.000600
BraceMatch 649 / 2364, 176147389 / 663802026, 0.000700
BraceMatch 1032 / 1164, 207105042 / 663802026, 0.000700
BraceMatch 22103559 / 22103764, 246998259 / 663802026, 0.000500
BraceMatch 2381 / 22103541, 251947793 / 663802026, 4.444000

======================================================================
notepad4-avx512.exe big.json

H:\Projects\notepad2\src\Notepad4.cpp:495 wWinMain
Notepad4 EOL time: 40.889300
CR+LF: 17837319, LF: 0, CR: 0
BasicInsertString avx2=1, cache=256, perLine=1023, duration=61.784000
InsertString duration=161.582300
BraceMatch 2 / -1, 16432709 / 677007975, 31.505500
BraceMatch 5 / 24935957, 24794339 / 677007975, 2.519400
BraceMatch 8 / 677007966, 36954773 / 677007975, 14.782100
BraceMatch 11 / 677007963, 38701202 / 677007975, 41.382900
BraceMatch 20 / 22566941, 45498942 / 677007975, 2.217700
BraceMatch 674 / 2436, 83889867 / 677007975, 0.000800
BraceMatch 8 / 677007966, 677007975 / 677007975, 14.773800
BraceMatch 11 / 677007963, 677007975 / 677007975, 69.885200
BraceMatch 20 / 22566941, 677007975 / 677007975, 2.742900
BraceMatch 2454 / 22543705, 677007975 / 677007975, 2.213800
BraceMatch 2460 / 4534, 677007975 / 677007975, 0.001000
BraceMatch 22543952 / 22566937, 677007975 / 677007975, 0.003200
BraceMatch 22543970 / 22544144, 677007975 / 677007975, 0.000500

======================================================================
notepad4-avx512.exe big-gbk.json

H:\Projects\notepad2\src\Notepad4.cpp:495 wWinMain
Notepad4 EOL time: 20.167400
CR+LF: 0, LF: 17837319, CR: 0
BasicInsertString avx2=1, cache=256, perLine=1023, duration=49.942100
InsertString duration=141.289800
BraceMatch 1 / -1, 16133616 / 663802026, 34.732700
BraceMatch 3 / 25321651, 24560536 / 663802026, 1.218400
BraceMatch 5 / 663802020, 36304327 / 663802026, 15.959800
BraceMatch 7 / 663802018, 38896663 / 663802026, 35.508500
BraceMatch 15 / 22126739, 46926987 / 663802026, 2.065700
BraceMatch 649 / 2364, 89402873 / 663802026, 0.000800
BraceMatch 2381 / 22103541, 109116522 / 663802026, 2.249700
BraceMatch 2381 / 22103541, 109116522 / 663802026, 3.724500
BraceMatch 2386 / 4395, 114974590 / 663802026, 0.001100
BraceMatch 22103559 / 22103764, 121131402 / 663802026, 0.000500
BraceMatch 22103782 / 22126736, 125064168 / 663802026, 0.002100
BraceMatch 3 / 141220810, 140834221 / 663802026, 12.278700
BraceMatch 1 / -1, 141898139 / 663802026, 50.197100

CPU info:

C:\Users\HP>wmic cpu list full


AddressWidth=64
Architecture=9
Availability=3
Caption=AMD64 Family 25 Model 117 Stepping 2
ConfigManagerErrorCode=
ConfigManagerUserConfig=
CpuStatus=1
CreationClassName=Win32_Processor
CurrentClockSpeed=2646
CurrentVoltage=12
DataWidth=64
Description=AMD64 Family 25 Model 117 Stepping 2
DeviceID=CPU0
ErrorCleared=
ErrorDescription=
ExtClock=100
Family=107
InstallDate=
L2CacheSize=8192
L2CacheSpeed=
LastErrorCode=
Level=25
LoadPercentage=10
Manufacturer=AuthenticAMD
MaxClockSpeed=3301
Name=AMD Ryzen 7 8840HS w/ Radeon 780M Graphics
OtherFamilyDescription=
PNPDeviceID=
PowerManagementCapabilities=
PowerManagementSupported=FALSE
ProcessorId=178BFBFF00A70F52
ProcessorType=3
Revision=29954
Role=CPU
SocketDesignation=FP8
Status=OK
StatusInfo=3
Stepping=2
SystemCreationClassName=Win32_ComputerSystem
SystemName=DESKTOP-CRIAF8P
UniqueId=
UpgradeMethod=6
Version=Model 5, Stepping 2
VoltageCaps=

Test with Notepad4 25.07 r5738 (08d0b0f0) with MSVC 19.44.35214.0

[missdeer]

Test with Notepad4 25.07 r5738 (08d0b0f) with MSVC 19.44.35214.0

Seems that the about dialog is not rebuilt as expected, the commit should be c774a0ba45cfc5759f8d881b7ef84a9e2641e0dd :

commit c774a0ba45cfc5759f8d881b7ef84a9e2641e0dd (HEAD -> main, origin/main, origin/HEAD)
Author: zufuliu <[email protected]>
Date:   Fri Aug 8 21:32:53 2025 +0800

    Fix wrong pixel count for `RGBAImage::BGRAFromRGBA()`.

[zufuliu]

simple quick test result:

Thanks, so it looks AVX512 is not obviously faster than AVX2.

Seems that the about dialog is not rebuilt

It will be rebuilt after Dialogs.cpp or it's includes changed, CI has a step to update reversion: https://github.com/zufuliu/notepad4/blob/8e4af459957c4918334aad081b462633b80fd937/.github/workflows/main.yml#L16-L18

[missdeer]

Thanks, so it looks AVX512 is not obviously faster than AVX2.

It seems that my computer is a laptop, and the AVX-512 instruction set may cause the CPU to throttle, so its performance boost might not be as high as expected.

[zufuliu]

and the AVX-512 instruction set may cause the CPU to throttle, so its performance boost might not be as high as expected.

I don't have more on this, maybe it will faster on server CPU (only has P cores?).

I abandoned AVX512 version UTF-8 validation:

1. The code doesn't compile with just -march=x86-64-v4 (Clang) due to the use of _mm512_permutex2var_epi8.
2. The code from https://github.com/lemire/fastvalidate-utf-8 is complex, the code from https://github.com/zwegner/faster-utf8-validator doesn't pass the tests (see https://github.com/zwegner/faster-utf8-validator/issues/6).

[missdeer]

2. The code from https://github.com/lemire/fastvalidate-utf-8 is complex, the code from https://github.com/zwegner/faster-utf8-validator doesn't pass the tests (see [AVX512 version form wip branch doesn't pass the tests zwegner/faster-utf8-validator#6](https://github.com/zwegner/faster-utf8-validator/issues/6)).

How about this library: https://github.com/simdutf/

[zufuliu]

How about this library: https://github.com/simdutf/

The C++ template code (also https://github.com/simdutf/is_utf8) is hard to isolate only the AVX515 parts.

[zufuliu]

Close this for now, AVX512 UTF-8 validation could be implemented later.