Adds an AVX2 option for the parser. #10
Conversation
This gives significant performance boost on Haswell CPUs (up to 2X). The change requires the entire parse_headers function to change - instead looking for the next token, it now creates a bitmap of all tokens. The code was not tested much, so it might contain some bugs, I mostly wrote it to check the performance potential of AVX2 for parsing.
|
Wow! This is very interesting. I am not sure if I can maintain this, but it's definitely a wonderful work. Please wait while I look into it. |
|
any updates on this? |
|
Sorry, I had thought that I had left my comment on this issue. The code in PR looks great. It shows the true potential of what could be done by using SIMD instructions, and my kudos go to @vkrasnov for the work. OTOH I am afraid I would not be capable of maintaining the code. The approach used in the PR is a big leap from string operations that the non-SIMD version of picohttpparser and the SIMD-optimized version based on PCMPxSTRx instruction use. And personally I dot have any practical issue in the level of performance achieved by the current implementation of picohttpparser. Considering all the facts, I am inclined to not merging this PR, but keep it open so that others interested could notice the work. |
|
@vkrasnov could volunteer if there was demand for it. But on my fork I see zero stars :) |
|
Has this been merged? |
|
We do not have a plan to merge this. HTTP/1 parsing performance is becoming less issue, as high volume traffic moves to HTTP/2 and /3. We think that the cost of maintaining the added complexity is not worth the performance benefit. Of course, people are free to use a fork or create and maintain ones own. |
|
Yeah, it has been open for too long, no reason to keep it now. |
This gives significant performance boost on Haswell CPUs (compared to PCMPESTRI)
1.78X for bench.c up to 2X for some other on i5-4278U, with gcc 4.9.2 -O3 -mavx2 -mbmi2.
The change requires the entire parse_headers function to change -
instead looking for the next token, it now creates a bitmap of all tokens.
When compiled without -mavx2 it uses the old parse_headers.
The code was not tested much, so it might contain some bugs, I mostly
wrote it to check the performance potential of AVX2 for parsing.