Luca Barbato <[email protected]> writes: > On 24/11/11 12:03, Robert Nagy wrote: >> Just scrolling through the prores code. And found a possible minor >> optimization. Don't have a setup for compiling right now so I can't measure >> the perf diff. >> >> In static inline int decode_vlc_codeword(GetBitContext *gb, uint8_t >> codebook) >> >> Replace: >> >>> log = 31 - av_log2(buf) /* count prefix bits (zeroes) */ >> >> with something like: >> >>> #ifdef X86 >>> #define NLZ(x) // use BSR instruction >>> #else >>> #define NLZ(x) 31 - av_log2_c(x) >>> #endif >>> >>> log = NLZ(buf); /* count prefix bits (zeroes) */ > > If the compiler doesn't do that for us, might be useful adding a > mathop for it.
av_log2() already uses __builtin_clz() if available. -- Måns Rullgård [email protected] _______________________________________________ libav-devel mailing list [email protected] https://lists.libav.org/mailman/listinfo/libav-devel
