diff options
| author | flywind <xzsflywind@gmail.com> | 2021-04-29 19:39:36 +0800 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2021-04-29 13:39:36 +0200 |
| commit | 5edddd68d04648553bc02d52cd53d3a1a14102f3 (patch) | |
| tree | 0a50c544369fcca5e063de23df6d0ad390d0a371 | |
| parent | 87229e272ecd4012b750215d35e914580cd8475c (diff) | |
| download | Nim-5edddd68d04648553bc02d52cd53d3a1a14102f3.tar.gz | |
improve cache line size (#17885)
* improve cache line size - 64bit system tend to use cache line of 64 bytes - add align https://trishagee.com/2011/07/22/dissecting_the_disruptor_why_its_so_fast_part_two__magic_cache_line_padding Though I'm not sure, what do you think? @timotheecour * Update lib/pure/concurrency/threadpool.nim
| -rw-r--r-- | lib/pure/concurrency/threadpool.nim | 9 |
1 files changed, 3 insertions, 6 deletions
diff --git a/lib/pure/concurrency/threadpool.nim b/lib/pure/concurrency/threadpool.nim index 9beb39522..f8a5f9ba8 100644 --- a/lib/pure/concurrency/threadpool.nim +++ b/lib/pure/concurrency/threadpool.nim @@ -52,17 +52,14 @@ proc signal(cv: var Semaphore) = release(cv.L) signal(cv.c) -const CacheLineSize = 32 # true for most archs +const CacheLineSize = 64 # true for most archs type Barrier {.compilerproc.} = object entered: int cv: Semaphore # Semaphore takes 3 words at least - when sizeof(int) < 8: - cacheAlign: array[CacheLineSize-4*sizeof(int), byte] - left: int - cacheAlign2: array[CacheLineSize-sizeof(int), byte] - interest: bool # whether the master is interested in the "all done" event + left {.align(CacheLineSize).}: int + interest {.align(CacheLineSize).} : bool # whether the master is interested in the "all done" event proc barrierEnter(b: ptr Barrier) {.compilerproc, inline.} = # due to the signaling between threads, it is ensured we are the only |