Hi, sorry for the late reply, I missed this updated patch On 2014-05-17 12:21:10 +0800, [email protected] wrote: > From: James Yu <[email protected]> > > Using NEON intrinsics to support ARMv7 and ARMv8 simultaneously. > Rewrite ARMv7 assembly to intinsics version. > > Toolchain: > armv7: gcc-linaro-arm-linux-gnueabihf-4.8-2014.01_linux > armv8: gcc-linaro-aarch64-linux-gnu-4.8-2013.10_linux > > Add libavcodec/arm/mpegvideo_neon.c. > > Result of "avconv -benchmark -i sample -an -f null -" on ARMv7 > Original NEON asm : bench: utime=307.030s maxrss=27988kB > NEON intrinsics : bench: utime=305.480s maxrss=28004kB
I can't reproduce this. On a cortex-a9 the asm is slightly faster: asm: 821 UNITS in dct_unquantize_h263_intra_neon, 4193974 runs, 330 skips 919 UNITS in dct_unquantize_h263_inter_neon, 8387930 runs, 678 skips intrinsics: 842 UNITS in dct_unquantize_h263_intra_intrin, 4193926 runs, 378 skips 1035 UNITS in dct_unquantize_h263_inter_intrin, 8387778 runs, 830 skips 10 UNITS are one cpu cycle. That's after removing the av_assert0 from your patch. It's slightly faster overall too. Updated patch fully integrated follows. Janne _______________________________________________ libav-devel mailing list [email protected] https://lists.libav.org/mailman/listinfo/libav-devel
