On Sat, Jan 20, 2018 at 5:18 PM, wm4 <nfx...@googlemail.com> wrote: > On Sat, 20 Jan 2018 12:52:46 +0700 > Muhammad Faiz <mfc...@gmail.com> wrote: > >> On Sat, Jan 20, 2018 at 11:49 AM, James Almer <jamr...@gmail.com> wrote: >> > On 1/20/2018 1:29 AM, Muhammad Faiz wrote: >> >> Help avoiding malloc-free cycles when allocating-freeing common >> >> structures. >> >> >> >> Signed-off-by: Muhammad Faiz <mfc...@gmail.com> >> >> --- >> >> libavutil/staticpool.h | 117 >> >> +++++++++++++++++++++++++++++++++++++++++++++++++ >> >> 1 file changed, 117 insertions(+) >> >> create mode 100644 libavutil/staticpool.h >> >> >> >> diff --git a/libavutil/staticpool.h b/libavutil/staticpool.h >> >> new file mode 100644 >> >> index 0000000000..9c9b2784bc >> >> --- /dev/null >> >> +++ b/libavutil/staticpool.h >> >> @@ -0,0 +1,117 @@ >> >> +/* >> >> + * This file is part of FFmpeg. >> >> + * >> >> + * FFmpeg is free software; you can redistribute it and/or >> >> + * modify it under the terms of the GNU Lesser General Public >> >> + * License as published by the Free Software Foundation; either >> >> + * version 2.1 of the License, or (at your option) any later version. >> >> + * >> >> + * FFmpeg is distributed in the hope that it will be useful, >> >> + * but WITHOUT ANY WARRANTY; without even the implied warranty of >> >> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU >> >> + * Lesser General Public License for more details. >> >> + * >> >> + * You should have received a copy of the GNU Lesser General Public >> >> + * License along with FFmpeg; if not, write to the Free Software >> >> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA >> >> 02110-1301 USA >> >> + */ >> >> + >> >> +#ifndef AVUTIL_STATICPOOL_H >> >> +#define AVUTIL_STATICPOOL_H >> >> + >> >> +#include <stdatomic.h> >> >> +#include "avassert.h" >> >> +#include "mem.h" >> >> + >> >> +/** >> >> + * FF_STATICPOOL allocate memory without av_malloc if possible >> >> + * @param size must be 2^n between 64 and 4096 >> >> + */ >> >> +#define FF_STATICPOOL_DECLARE(type, size) >> >> \ >> >> +typedef struct type##_StaticPoolWrapper { >> >> \ >> >> + type buf; >> >> \ >> >> + unsigned index; >> >> \ >> >> + atomic_uint next; >> >> \ >> >> +} type##_StaticPoolWrapper; >> >> \ >> >> + >> >> \ >> >> +static atomic_uint type##_staticpool_next; >> >> \ >> >> +static atomic_uint type##_staticpool_last; >> >> \ >> >> +static type##_StaticPoolWrapper type##_staticpool_table[size]; >> >> \ >> >> + >> >> \ >> >> +static type *type##_staticpool_malloc(void) >> >> \ >> >> +{ >> >> \ >> >> + unsigned val, index, serial, new_val; >> >> \ >> >> + >> >> \ >> >> + av_assert0((size) >= 64 && (size) <= 4096 && !((size) & ((size) - >> >> 1))); \ >> >> + >> >> \ >> >> + /* use serial, avoid spinlock */ >> >> \ >> >> + /* acquire, so we don't get stalled table[index].next */ >> >> \ >> >> + val = atomic_load_explicit(&type##_staticpool_next, >> >> memory_order_acquire); \ >> >> + do { >> >> \ >> >> + index = val & ((size) - 1); >> >> \ >> >> + serial = val & ~((size) - 1); >> >> \ >> >> + new_val = >> >> atomic_load_explicit(&type##_staticpool_table[index].next, >> >> memory_order_relaxed) | (serial + (size)); \ >> >> + } while >> >> (!atomic_compare_exchange_strong_explicit(&type##_staticpool_next, &val, >> >> new_val, \ >> > >> > The wrappers for atomic_compare_exchange_* in the compat folder are not >> > really working and fixing them is supposedly not trivial, so this will >> > only work with GCC and Clang but not with for example MSVC or SunCC. >> >> What's the problem? I only see that stdatomic compat make typedef >> every atomic type to intptr_t, so atomic_*64_t won't work if >> sizeof(intptr_t) == 32. >> Here I use atomic_uint, so I guess it will work. >> >> Note that if atomic_compare_exchange_* is broken then atomic_fetch_* >> will also be broken because atomic_fetch_* call >> atomic_compare_exchange_* on suncc compat. >> >> > >> > Can you implement this using mutexes instead, or otherwise avoid using >> > atomic_compare_exchange_*? You can even use static mutex initialization >> > now for all targets, and not just native pthreads. >> >> Using mutex makes implementation slower. >> Using atomic_exchange requires spin lock. > > Uncontended mutexes are spinlocks.
Of course. But this code isn't spinlock, so probably the performance won't suffer so much when contended. _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel