[B! simd] yassã®ãƒ–ãƒƒã‚¯ãƒžãƒ¼ã‚¯

yass id:yass

simdã«é–¢ã™ã‚‹yassã®ãƒ–ãƒƒã‚¯ãƒžãƒ¼ã‚¯ (25)

${{author_name}}$

{{{comment_expanded}}}

{{label}}

{{#is_bookmark}}ãƒªã‚¹ãƒˆ{{/is_bookmark}}{{^is_bookmark}}ãƒªãƒ³ã‚¯{{/is_bookmark}}

${{author_name}}$
{{author_name}}{{created}}
{{ #comment }}{{ comment }}{{ /comment }}
- {{ label }}

${{author_name}}$

{{{comment_expanded}}}

{{label}}

{{#is_bookmark}}ãƒªã‚¹ãƒˆ{{/is_bookmark}}{{^is_bookmark}}ãƒªãƒ³ã‚¯{{/is_bookmark}}

HotSpot JavaVM ã® SIMD æœ€é©åŒ–ã‚’åŠ¹ã‹ã›ã‚‹æ–¹æ³• - Qiita
HotSpot JavaVM ã®ãƒ™ã‚¯ãƒˆãƒ«åŒ–å¤‰æ› æœ€è¿‘ã® HotSpot JavaVM ã¯ã‚¹ã‚«ãƒ©ãƒ¼æ¼”ç®—ã®ç¹°ã‚Šè¿”ã—å‡¦ç†ã‚’ãƒ™ã‚¯ãƒˆãƒ«åŒ–ã— SIMD å‘½ä»¤ã«å¤‰æ›ã™ã‚‹æœ€é©åŒ–ã‚’è¡Œã£ã¦ã„ã¾ã™ (SIMD ã¨ã¯ä½•ãžã‚„ã¨ã„ã†è©±ã¯å¾ŒåŠå‚ç…§)ã€‚å®Ÿéš›ã«æœ€é©åŒ–ãŒåŠ¹ãã‚³ãƒ¼ãƒ‰ã§è©¦ã—ã¦ã¿ãŸã¨ã“ã‚ 1.5ï½ž2.8 å€ç¨‹åº¦ã®é€Ÿåº¦å‘ä¸ŠãŒè¦‹ã‚‰ã‚ŒãŸã®ã§ã€å¤§é‡ã®æ¼”ç®—å‡¦ç†ã‚’è¡Œã† (GPGPU ã«é ¼ã‚‰ãªã„) Java ãƒ©ã‚¤ãƒ–ãƒ©ãƒªã§ã†ã¾ãä½¿ã†ã“ã¨ãŒå‡ºæ¥ã‚Œã°æœ‰åŠ¹ãªæœ€é©åŒ–æ‰‹æ®µã«ãªã‚‹ã‹ã‚‚ã—ã‚Œã¾ã›ã‚“ã€‚ HotSpot ã® SIMD æœ€é©åŒ–ã¯ Superword-Level Parallelism (SLP) ã«åŸºã¥ã„ã¦ã„ã¾ã™ (ä»¥é™ã“ã®æœ€é©åŒ–ã¯ SLP ã¨å‘¼ã³ã¾ã™)ã€‚å…ƒã€…ã“ã®è«–æ–‡ã¯æ™‚ä»£ã‚’åæ˜ ã—ã¦ SIMD æœªå¯¾å¿œã®ç”»åƒãƒ»éŸ³å£°å‡¦ç†ã‚’ã‚³ãƒ³ãƒ‘ã‚¤ãƒ©ã‚„ãƒ©ãƒ³ã‚¿ã‚¤ãƒ ã®ãƒ¬ã‚¤ãƒ¤ãƒ¼ã§ SIMD ã‚’åˆ©ç”¨ã™ã‚‹å‘½ä»¤ã«å¤‰æ›ã™ã‚‹ã“ã¨ã‚’ç›®çš„ã¨ã—ã¦ã„ã¾ã—ãŸãŒã€ã“ã‚Œã¯ SIMD å‘½ä»¤
yass 2017/09/23
java

SIMD
ãƒªãƒ³ã‚¯
http://prestodb.rocks/code/simd/
yass 2017/06/28
java

SIMD
ãƒªãƒ³ã‚¯
Vectorization in HotSpot JVM
yass 2017/04/11
JVM

SIMD
ãƒªãƒ³ã‚¯
lzturbo
LzTurbo is a high performance lz77 based compressor featuring the fastest compression/decompression. Serveral times faster than the fastest compressor (see LzTurbo benchmarks) Download: LzTurbo v1.2 for windows and linux - Parallel and linear scala ble compression/decompression on multiprocessor/multicore systems. - Overlapping i/o with compression/decompression. - Parallel compression/decompressio
yass 2016/06/11
compression

LZ77

SIMD
ãƒªãƒ³ã‚¯
OpenCLã‚„ã‚‹å‰ã«SIMDä½¿ã„åˆ‡ã‚Œã£ã¦ã„ã†å¹»æƒ³ - aokomoriuta's blog
ãŠè©«ã³ ã‚¢ãƒ©ã‚¤ãƒ³å¿˜ã‚Œã¦ã¾ã—ãŸã”ã‚ã‚“ãªã•ã„ã€ã§ã‚‚æ™‚é–“ã‚ã‚“ã¾ã‚Šå¤‰ã‚ã£ã¦ãªã„ã‹ã‚‰è¨±ã—ã¦ãƒ»ãƒ»ãƒ» https://t.co/JNtq2U2kMqâ€” é’åå®ˆæŒ (@aokomoriuta) April 29, 2015 ã§ã¯æœ¬ç·¨ã©ã†ãžâ†“ æœ¬ç·¨ è‹¥å¹²è©±é¡Œã«ä¹—ã‚Šé…ã‚ŒãŸæ„Ÿã‚ã‚Šã¾ã™ãŒã€‚ d.hatena.ne.jp ã‘ã©ã€SSEã‚‚çŸ¥ã‚‰ããƒ¼ã€SIMDã‚‚çŸ¥ã‚‰ããƒ¼ã€ãªã‚“ã‹ä¿ºãŒæ›¸ã„ãŸã‚¢ãƒ«ã‚´ãƒªã‚ºãƒ é…ã„ã‘ã©ã¨ã‚Šã‚ãˆãšOpenCLã¨ã‹ã§é«˜é€ŸåŒ–ã—ã‚ˆã£ã‹ãªãƒ¼ã¨ã‹ãã€ç”˜ã£ãŸã‚Œã¦ã‚“ã˜ã‚ƒããˆã‚ˆã€‚CPUã‚’ã‚‚ã£ã¨ä½¿ã„ãã‚Œã‚ˆã€‚ãŠå‰ã®ã‚¢ã‚ªã„ã‚³ãƒ¼ãƒ‰ã®ã›ã„ã§CPUãŒæ³£ã„ã¦ã‚‹ã‚ˆã€‚ã£ã¦ã„ã†è©±ã§ã™ã‚ˆã€‚ GPGPUãªã‚“ã¦ã®ã¯SIMDã‚’ä½¿ã„åˆ‡ã£ãŸå¾Œã®è©±ã§ã—ã‚‡ã€‚ GPGPUã™ã‚‹ã®ã«GPUã®ãƒ‘ãƒ¯ãƒ¼ã¨ãƒ¡ãƒ¢ãƒªãŒè¶³ã‚Šã¾ã›ã‚“ã¨ã‹è¨€ã†å‰ã«ã¾ãšSIMDã‹ã‚‰ã ã‚ã€‚ ã¨ã‹è¨€ã‚ã‚Œã¦ãŸã®ã§ã€æ¤œè¨¼ã™ã‚‹ã“ã¨ã«ã—ãŸï¼ˆã‚„ã£ã¤ã‘ï¼‰ã€‚ ç’°å¢ƒ OS: Windows 7 Profession
yass 2015/04/29
AVX

opencl

GPGPU

simd
ãƒªãƒ³ã‚¯
ãƒ—ãƒã‚°ãƒ©ãƒ ã‚’é«˜é€ŸåŒ–ã™ã‚‹è©±
9. 9 æœ€é©åŒ–ã«ã¤ã„ã¦ ã€Œç´°ã‹ã„åŠ¹çŽ‡ã®ã“ã¨ã¯å¿˜ã‚Œã¦ã€æ™‚é–“ã® 97% ã«ã¤ã„ã¦è€ƒãˆ ã‚ˆã†ã€‚æ™‚æœŸå°šæ—©ãªæœ€é©åŒ–ã¯è«¸æ‚ªã®æ ¹æºã ã€‚ãã‚Œã§ã‚‚æ®‹ã‚Š 3% ã«ã¤ã„ã¦ã‚‚æ©Ÿä¼šã‚’é€ƒã™ã¹ãã§ã¯ãªã„ã€ - Donald E. Knuth ã€Œãƒ—ãƒã‚°ãƒ©ãƒ æœ€é©åŒ–ã®ç¬¬ä¸€æ³•å‰‡ : æœ€é©åŒ–ã™ã‚‹ãªã€‚ ãƒ—ãƒã‚°ãƒ©ãƒ æœ€é©åŒ–ã®ç¬¬äºŒæ³•å‰‡ ( ä¸Šç´šè€…é™å®š ): ã¾ã ã™ã‚‹ãªã€‚ ã€ - Michael A. Jackson 11. 11 æœ€é©åŒ–ã®å¯¾è±¡ ä¸»ã« Intel ã® Haswell ãƒžã‚¤ã‚¯ãƒã‚¢ãƒ¼ã‚ãƒ†ã‚¯ãƒãƒ£ä»¥é™ã‚’å¯¾è±¡ å¤šãã®ãƒ†ã‚¯ãƒ‹ãƒƒã‚¯ã¯ä»–ã®ãƒ—ãƒã‚»ãƒƒã‚µã«ã‚‚å¿œç”¨ã§ãã¾ã™ ãƒ™ãƒ¼ã‚¹ ãƒžã‚¤ã‚¯ãƒã‚¢ãƒ¼ã‚ãƒ†ã‚¯ãƒãƒ£ ãƒ—ãƒã‚»ã‚¹ãƒ«ãƒ¼ãƒ« ç™»å ´å¹´ Nehalem Nehalem 45nm 2008 ã€ƒ Westmere 32nm 2010 Sandy Bridge Sandy Bridge 32nm 2011 ã€ƒ Ivy Bridge 22nm 2012 Hasw
yass 2015/03/16
CPU

simd

cache

bit
ãƒªãƒ³ã‚¯
SIMD auto-vectorization in hotspot
Krystal Mok rednaxelafx at gmail.com Mon Jan 16 22:36:20 PST 2012 Previous message: SIMD auto-vectorization in hotspot Next message: SIMD auto-vectorization in hotspot Messages sorted by: [ date ] [ thread ] [ subject ] [ author ] Hi, Apparently, the source code to read is opto/superword.[hpp|cpp] [1][2]. The flag to trace it is -XX:+TraceSuperWord. There's a paper mentioned in [1], which is the o
yass 2014/12/21
" -XX:+UseSuperWord "

java

simd
ãƒªãƒ³ã‚¯
Takayuki Kaneko on Twitter: "Intelã•ã‚“æ›°ãã€‚E5 v3ã®avx2å‘½ä»¤ã‚’ã‚µãƒãƒ¼ãƒˆã—ãŸJava8ã¯3å€é€Ÿã„ã‚‰ã—ã„ã€‚ #cwt2014"
yass 2014/11/07
java

java8

simd

avx

xeon
ãƒªãƒ³ã‚¯
Vitesse Data | Home page
More Speed Deepgreen runs TPCH up to 5X faster than Greenplum and comes with many OLAP specific optimizations. These performance improvements means the cluster can do more without costly expansion. More Connected Connect effortlessly to cloud storage and data sources such as HDFS, S3, Oracle, Geode and Elastic Search. Storing the fact tables on HDFS/S3 makes for lean DW and eases ETL chores. Fresh
yass 2014/10/25
postgresql

llvm

simd
ãƒªãƒ³ã‚¯
AVX2æ™‚ä»£ã®æ£è¦è¡¨ç¾ãƒžãƒƒãƒãƒ³ã‚° ã€œåŠç¾¤ã§ãã‚“ãã‚“ï¼ã€œ
x86/x64æœ€é©åŒ–å‹‰å¼·ä¼šã§ã®ç™ºè¡¨è³‡æ–™ã§ã™ http://atnd.org/events/25823 UST â†’ http://www.ustream.tv/recorded/21484937 2013å¹´åº¦ã®Haswellã‹ã‚‰å…¥ã‚‹AVX2ã‚’ä½¿ã£ãŸæ£è¦è¡¨ç¾ãƒžãƒƒãƒãƒ³ã‚°ã¨ã„ã†ãƒžãƒ‹ã‚¢ãƒƒã‚¯ãªãƒã‚¿ :-)Read less
yass 2014/10/17
AVX

RegExp

simd
ãƒªãƒ³ã‚¯
Quick Post: Should you ever use Linked-Lists? | Future Chips
This question has been bugging me for the last few days. Why would anyone use a linked-list instead of arrays? I argue that linked lists have become irrelevant in the context of a modern computer system. I have asked around a few colleagues and I include their counter arguments and my rebuttal for your reference. Please point out if I am missing something. Lists have several disadvantages that did
yass 2014/02/16
data structure

linked list

simd

prefetch

memory

TBL
ãƒªãƒ³ã‚¯
ãƒžãƒ«ãƒã‚³ã‚¢CPUä¸Šã®ä¸¦åˆ—åŒ–æ‰‹æ³•ã€ãã®ä¸¦åˆ—æ€§èƒ½ã¨å•é¡Œç‚¹
GPU ã‚’ä½¿ç”¨ã—ãŸä¸¦åˆ—è¨ˆç®—ç’°å¢ƒãŒä¸€èˆ¬çš„ã«ãªã‚Šã¤ã¤ã‚ã‚Šã¾ã™ã€‚ãƒ¦ãƒ¼ã‚¶ã‚µã‚¤ãƒ‰ç«‹ã£ã¦ã¿ã‚Œã°ã€ã€Œä¸¦åˆ—åŒ–é©ç”¨æ–¹æ³•ã®ç¨®é¡žã¨ãã®åŠ¹æžœã€ã‚’ã‚‚ã†ä¸€åº¦æ•´ç†ã—ã¦è€ƒãˆã€è‡ªåˆ†ã«åˆã£ãŸæ‰‹æ³•ã‚’é¸æŠžã—ãªã‘ã‚Œã°ã„ã‘ãªã„ã¨æ€ã£ã¦ã„ã‚‹æ–¹ã‚‚å¤šã„ã“ã¨ã§ã—ã‚‡ã†ã€‚ä¸¦åˆ—è¨ˆç®—ã¨è¨€ã£ã¦ã‚‚ã€ãƒ—ãƒã‚»ãƒƒã‚µãƒ»ã‚³ã‚¢è‡ªä½“ãŒè‡ªå‹•çš„ã«SIMDï¼ˆãƒ™ã‚¯ãƒˆãƒ«å‡¦ç†ï¼‰å®Ÿè¡Œã™ã‚‹ã‚‚ã®ã‹ã‚‰ã€ãƒžãƒ«ãƒã‚³ã‚¢ã‚’åˆ©ç”¨ã—ãŸè‡ªå‹•ä¸¦åˆ—ã€OpenMPã«ã‚ˆã‚‹ã‚¹ãƒ¬ãƒƒãƒ‰ãƒ¬ãƒ™ãƒ«ã®ä¸¦åˆ—å®Ÿè¡Œã€ã•ã‚‰ã«ãƒ—ãƒã‚°ãƒ©ãƒ ã‚’æ”¹é€ ã—ã¦ãƒžãƒ«ãƒãƒ—ãƒã‚»ã‚¹ã«ã‚ˆã‚‹ MPI ä¸¦åˆ—å®Ÿè¡Œã¨æ§˜ã€…ãªæ‰‹æ³•ãŒç”¨æ„ã•ã‚Œã¦ã„ã¾ã™ã€‚ã•ã‚‰ã«ã€ä»Šã€GPUã‚’åˆ©ç”¨ã—ãŸ many cores ä¸¦åˆ—ã‚’åˆ©ç”¨ã—ãŸä¸¦åˆ—å®Ÿè¡Œã‚‚åŠ ã‚ã‚Šã€ãƒ¦ãƒ¼ã‚¶ã«ã¨ã£ã¦ã¯ã€è‡ªåˆ†ã§ã‚‚æ‰‹è»½ã«ã§ãã‚‹æ–¹æ³•ã®åˆ¤æ–ã‚„ãã®æ€§èƒ½åŠ¹æžœã«é–¢ã™ã‚‹ä¸€ã¤ã®åˆ¤æ–åŸºæº–ã‚’æŒã£ã¦ãŠãã“ã¨ãŒå¿…è¦ã§ã—ã‚‡ã†ã€‚ã¾ãŸã€ç°¡å˜ãªä¸¦åˆ—å®Ÿè¡Œã®åŽŸç†ç‰ã®çŸ¥è˜ãŒã‚ã‚‹ã ã‘ã§ã‚‚ã€ä»Šå¾Œã® many cores ä¸¦åˆ—æ™‚ä»£ã«å‘ã‘ã¦ã€åˆç†çš„ãªä¸¦åˆ—åŒ–ã®ãŸã‚ã®æŒ‡
yass 2014/01/13
"ã“ã®ã‚·ã‚¹ãƒ†ãƒ ã ã‘ã§è¨€ã†ã¨ã€ 2ã‚¹ãƒ¬ãƒƒãƒ‰ã‚‚ 4 ã‚¹ãƒ¬ãƒƒãƒ‰ã®ä¸¦åˆ—ãƒ¡ãƒ¢ãƒªãƒãƒ³ãƒ‰å¹…ãŒ 16GB/sec ç¨‹åº¦ã¨åŒã˜ã§ã‚ã‚Šã€å¸¯åŸŸãŒé£½å’Œã—ã¦ã„ã‚‹ã“ã¨ãŒåˆ†ã‹ã‚Šã¾ã™ã€‚ã“ã†ã—ãŸçŠ¶æ³ãŒç¾åœ¨ã® x64 ç³»ã®ãƒ—ãƒã‚»ãƒƒã‚µã®ãƒ¡ãƒ¢ãƒªãƒ»ã‚¢ãƒ¼ã‚ãƒ†ã‚¯ãƒãƒ£ã®é™ç•Œ"

cpu

concurrency

memory

cache

bandwidth

simd
ãƒªãƒ³ã‚¯
æˆ–ã‚‹ãƒ—ãƒã‚°ãƒ©ãƒžã®ä¸€ç”Ÿ Â» AVX-512 ãŒã‚¢ãƒŠã‚¦ãƒ³ã‚¹ã•ã‚Œã¦ã„ãŸ
x86/x64 æœ€é©åŒ–å‹‰å¼·ä¼š #6 ã§èžã„ã¦ãã¾ã—ãŸãŒã€AVX-512 ãªã‚‹ã‚‚ã®ãŒã‚¢ãƒŠã‚¦ãƒ³ã‚¹ã•ã‚Œã¦ã¾ã—ãŸã€‚åå‰ã‹ã‚‰äºˆæƒ³ã•ã‚Œã‚‹é€šã‚Šã€SIMD ãƒ¬ã‚¸ã‚¹ã‚¿ãŒ 512bit ã«ãªã‚Šã¾ã™ã€‚SIMD ãƒ¬ã‚¸ã‚¹ã‚¿ãŒ 512bit ã¨ã„ã†ã¨ Xeon Phi ã¯æœ€åˆã‹ã‚‰ãã†ãªã£ã¦ã„ã‚‹ã®ã§ã™ãŒã€ã“ã‚ŒãŒæ™®é€šã® Xeon ã«é™ã‚Šã¦ãã‚‹ã‚¤ãƒ¡ãƒ¼ã‚¸ã§ã™ã€‚ AVX ãŒã€Œå˜ãªã‚‹ legacy SSE ã® 256bit ç‰ˆã€ã§ã¯ãªã„ã®ã¨åŒæ§˜ã€AVX-512 ã‚‚ã€Œå˜ãªã‚‹ AVX ã® 512bit ç‰ˆã€ã§ã¯ã‚ã‚Šã¾ã›ã‚“ã€‚ SIMD ãƒ¬ã‚¸ã‚¹ã‚¿ãŒ zmm ã¨ã„ã†åå‰ã§ 512bit ã«ãªã‚‹ã€‚ä¸‹ä½ 256bit ã¯ ymm ã¨ã—ã¦ã‚¢ã‚¯ã‚»ã‚¹ã§ãã‚‹ï¼ˆxmm ã«å¯¾ã™ã‚‹ ymm ã¨åŒã˜ï¼‰ 64bit ãƒ¢ãƒ¼ãƒ‰ã®å ´åˆã€SIMD ãƒ¬ã‚¸ã‚¹ã‚¿ãŒ 32æœ¬ä½¿ãˆã‚‹ã€‚ã“ã‚Œã¯ã€zmm ã ã‘ã§ãªã ymm ã‚„ xmm ã§ã‚‚åŒæ§˜ã§ã‚ã‚‹ã€‚32bit ãƒ¢ãƒ¼ãƒ‰ã§ã¯8æœ¬ã®ã¾ã¾
yass 2014/01/09
" SIMD ãƒ¬ã‚¸ã‚¹ã‚¿ãŒ 512bit ã«ãªã‚Šã¾ã™ã€‚SIMD ãƒ¬ã‚¸ã‚¹ã‚¿ãŒ 512bit ã¨ã„ã†ã¨ Xeon Phi ã¯æœ€åˆã‹ã‚‰ãã†ãªã£ã¦ã„ã‚‹ã®ã§ã™ãŒã€ã“ã‚ŒãŒæ™®é€šã® Xeon ã«é™ã‚Šã¦ãã‚‹ã‚¤ãƒ¡ãƒ¼ã‚¸ã§ã™ã€‚"

SIMD

AVX

intel

CPU

xeon

Xeon Phi
ãƒªãƒ³ã‚¯
Javaâ„¢ SE Development Kit 7 Update 40 Bug Fixes
yass 2013/09/24
"6340864 / hotspot / Implement vectorization optimizations in hotspot-server"

java

jvm

SIMD

SSE

vectorization
ãƒªãƒ³ã‚¯
ORC File and Vectorization - Hadoop Summit 2013
yass 2013/09/24
ORCFile

hive

hadoop

java

SIMD

vectorization
ãƒªãƒ³ã‚¯
https://bugs.java.com/?bug_id=6340864
yass 2013/09/23
" Almost all C-compilers available are able to vectorize code, so that simple code in loops for example can be processed in parallel by using the SIMD registers of modern CPUs like SSE1/2 or AltiVec. "

java

SIMD

SSE

vectorization
ãƒªãƒ³ã‚¯
çµ„ã¿è¾¼ã¿é–¢æ•°ï¼ˆintrinsicï¼‰ã«ã‚ˆã‚‹SIMDå…¥é–€
6. ç›®æ¨™ç›®æ¨™ OpenCVã«ã‚ˆã‚Šã‚‚é«˜é€Ÿãªãƒ—ãƒã‚°ãƒ©ãƒ ã‚’ç°¡OpenCVã«ã‚ˆã‚Šã‚‚é«˜é€Ÿãªãƒ—ãƒã‚°ãƒ©ãƒ ã‚’ç°¡ å˜ã«ã¤ãã‚Œã‚‹æ›¸ãæ–¹ã‚’è¦šãˆã‚ˆã†å˜ ãã‚Œã‚‹æ›¸ãæ–¹ã‚’è¦šãˆã‚ˆã† å‡ºæ¥ã‚‹ã‚ˆã†ã«ãªã‚‹ã“ã¨ ãƒ»é«˜é€Ÿã‚³ãƒ”ãƒ¼ src.copyTo(dest);é«˜é€Ÿã‚³ãƒ” src.copyTo(dest); ãƒ»é«˜é€ŸãªåŠ æ¸›ä¹—ç®— dest =Â src1Â +Â src2; ãƒ»é«˜é€Ÿåž‹å¤‰æ› src.convert(dest,CV_16U); ãƒ»é«˜é€ŸBGR2RGB cvtColor(src,dest,CV_BGR2RGB); ãƒ»é«˜é€ŸBGR2Gray cvtColor(src,dest,CV_BGR2Gray); â€»å‡ºæ¥ãªã„ã“ã¨ æ•´æ•°æ¼”ç®—ã ã‘ æµ®å‹•å°æ•°ç‚¹ã®å‘½ä»¤ã¯ä½¿ã‚ãªã„ æ¯”è¼ƒæ¼”ç®—ã¯ä½¿æ•´æ•°æ¼”ç®—ã ã‘ï¼Œæµ®å‹•å°æ•°ç‚¹ã®å‘½ä»¤ã¯ä½¿ã‚ãªã„ï¼Œæ¯”è¼ƒæ¼”ç®—ã¯ä½¿ ã‚ãªã„ï¼Œæ°´å¹³æ¼”ç®—ã¯ä½¿ã‚ãªã„ï¼ŒAVXã¯å«ã¾ãªã„(SSE4.1ã¾ã§) 7. ç›®çš„ç›®çš„ ã‚’ä½¿ãˆã° åŒ–ã¯é›£ã—ããªã„ã‚ˆ
yass 2013/08/14
CPU

simd

OpenCV

SSE

AVX

intrinsic

intel

cpu
ãƒªãƒ³ã‚¯
SSEã¨AVXã§é«˜æ¬¡å…ƒãƒ™ã‚¯ãƒˆãƒ«ã®å†…ç©è¨ˆç®—ã‚’é«˜é€ŸåŒ–ã—ã¦ã¿ãŸ | ã•ã‹ãªå‰ç·š
ä¸–ç•Œæœ€é€Ÿã®ãŠéšã¨è¨€ãˆã°ã‚«ã‚¸ã‚é¡žã§ï¼Œæ³³ãé€Ÿåº¦ã¯æ™‚é€Ÿ100km/hã‚’è¶…ãˆã‚‹ã¨ã‹ï¼Ž55ãƒŽãƒƒãƒˆç¨‹ã«ãªã‚‹ã®ã§ã“ã‚Œã¯Mk-48éšé›·ã«ã‚‚åŒ¹æ•µã™ã‚‹ã»ã©ã§ã™ï¼Ž ä¸€æ–¹ã¡ã‚‡ã£ã¨ãƒãƒ¼ãƒˆãªé«˜é€ŸãŠéšã¨ã—ã¦ã¯ï¼ŒãŠé¦´æŸ“ã¿ãƒˆãƒ“ã‚¦ã‚ªãŒé£›è¡Œä¸ã«æœ€å¤§70km/hã»ã©ã«é”ã™ã‚‹ã¨ã‹ï¼Ž ä»Šæ—¥ã¯ãã‚“ãªè‹¥å¹²ãƒãƒ¼ãƒˆãªé«˜é€ŸåŒ–ã®ãŠè©±ï¼ˆï¼Ÿï¼‰ã¨ã„ã†ã“ã¨ã§ï¼ŒSSEçµ„ã¿è¾¼ã¿å‘½ä»¤ã«ã¤ã„ã¦ï¼Ž SSEã‚„AVXã¨ã„ãˆã°ãŠé¦´æŸ“ã¿SIMDå‘½ä»¤ã§ï¼Œãã‚Œã‚’ãƒ—ãƒã‚°ãƒ©ãƒ ã‹ã‚‰æ§‹é€ ä½“ã¨é–¢æ•°ã®å½¢å¼ã§é«˜ç§»æ®–ã«è¨˜è¿°ã™ã‚‹æ–¹æ³•ãŒSIMDçµ„ã¿è¾¼ã¿é–¢æ•°(SIMD Intrinsic)ãªã‚ã‘ã§ã™ãŒï¼Œã“ã‚Œã‚’ä½¿ã£ã¦ã”ãå…¸åž‹çš„ãªãƒ™ã‚¯ãƒˆãƒ«ã®å†…ç©è¨ˆç®—ã‚’é«˜é€ŸåŒ–ã—ã¦ã¿ã¾ã—ãŸï¼Ž ãƒ™ã‚¯ãƒˆãƒ«ã®å†…ç©ã®é«˜é€ŸåŒ–ã¨è¨€ãˆã°æ˜Ÿã®æ•°ã»ã©ã‚‚ã•ã‚Œã¦ã‚‹è©±ãªã‚ã‘ã§ï¼Œã„ã¾ã•ã‚‰éšã®æƒ…å ±ãªã‚“ã‹å½¹ã«ç«‹ã¤æ°—ã¯å…¨ãã—ãªã„ã®ã§ã™ãŒï¼Œç´”ç²‹ã«è‡ªåˆ†ã§ã‚„ã‚‰ãªã„ã¨ã‚ã‹ã‚“ãªã„ï¼žï¼œã€€ã¨ã„ã†ã“ã¨ã§ï¼Œ ã¨ã«ã‹ãæ›¸ã„ã¦ã¿ã‚ˆã† åŠ¹æžœã®ç¨‹ã¯ã„ã‹ã«ï¼Ÿ ã‚’èª¿ã¹ã¦ã¿ãŸãã¦ï¼Œã‚„
yass 2013/08/13
" SSEã‚„AVXã¨ã„ãˆã°ãŠé¦´æŸ“ã¿SIMDå‘½ä»¤ / ã“ã‚Œã‚’ä½¿ã£ã¦ã”ãå…¸åž‹çš„ãªãƒ™ã‚¯ãƒˆãƒ«ã®å†…ç©è¨ˆç®—ã‚’é«˜é€ŸåŒ– / æœ€å¤§ã§10å€ã®ã‚¹ãƒ«ãƒ¼ãƒ—ãƒƒãƒˆã¨ã„ã†ã®ã¯é©šã / 8192æ¬¡å…ƒï¼Œãƒ‡ãƒ¼ã‚¿é‡ã«ã—ã¦8192*sizeof(float)*2=64KBã«é”ã™ã‚‹ã¨ï¼Œé€Ÿåº¦å·®ãŒãªããªã£ã¦ã¾ã™ã"

sse

simd

AVX

intel

vector

Intrinsic

c

cpu
ãƒªãƒ³ã‚¯
https://bugs.java.com/?bug_id=7121648
yass 2013/08/13
java

java7

intel

cpu

AVX

SIMD
ãƒªãƒ³ã‚¯
ãªãªã‚45Â°ã‹ã‚‰è¦‹ãŸJavaOne
2. è‡ªå·±ç´¹ä»‹ ã™ãˆãªãŒ ã‚„ã™ã¾ã• ï‚— æœ«æ°¸ ææ£ï¼ˆ@YaSuenagï¼‰ ï‚— NTT OSSã‚»ãƒ³ã‚¿å‹¤å‹™ â—¦ Javaã®éšœå®³è§£æžã€Q&Aå¯¾å¿œ ï‚– åˆè¨€è‘‰ã¯ã€ŒJavaã„ã„åã ã‹ã‚‰ï¼ã€ ï‚– hs_errãƒã‚°ã‚„ã‚³ã‚¢ã®è§£æž ï‚– Javaã¨libcã€Linuxã‚«ãƒ¼ãƒãƒ«ã®é–“ã‚’è¡Œã£ãŸã‚Šæ¥ãŸã‚Šâ€¦ â—¦ ãŸã¾ï½žã«OpenJDKã®ãƒ‘ãƒƒãƒæ›¸ã„ã¦ã¾ã™ ï‚— æ ¹ã£ã‹ã‚‰ã®ã‚µãƒ³ãƒ‡ãƒ¼ãƒ—ãƒã‚°ãƒ©ãƒžãƒ¼ â—¦ Object Pascal (Delphi) å¤§å¥½ã â—¦ Cã¨Javaã‚‚å¥½ã â—¦ ã‚¢ã‚»ãƒ³ãƒ–ãƒ©ã‚‚çµæ§‹å¥½ã ï‚— JavaOneæ´ï¼š2å¹´ï¼ˆ2010 & 2011ï¼‰ 3. JavaOneã§æ„Ÿã˜ãŸã“ã¨ ï‚— æ±‚ã‚ã‚‰ã‚Œã‚‹ã‚¹ã‚ãƒ«ã®2æ¥µåŒ– (EE vs SE) â—¦ EE ï‚– ã¨ã‚Šã‚ãˆãšã€Œã‚¯ãƒ©ã‚¦ãƒ‰ã€ ï‚– ã¨ã«ã‹ãã€Œãƒãƒ¼ã‚¿ãƒ“ãƒªãƒ†ã‚£ã€ï¼ ï‚– EoD å¯¾ â—¦ SE æ¥µ ï‚– ã¨ã‚Šã‚ãˆãšã€Œä¸¦åˆ—åŒ–ã€ ï‚– ãƒžãƒ«ãƒãƒªãƒ³ã‚¬ãƒ«ï¼ŸãªVMã®å®Ÿç¾ ï‚– VMã®ä½œã‚Šã‚„å‹•ãã‚’æ„è˜
yass 2013/08/13
java

Performance

memory

cpu

hugepage

String

SSE

AVX

simd
ãƒªãƒ³ã‚¯
1 2 æ¬¡ã®ãƒšãƒ¼ã‚¸