On Tue, Mar 29, 2011 at 5:40 PM, Barry Song <21cn...@gmail.com> wrote: > 2011/3/24 Michael Hope <michael.h...@linaro.org>: >> Hi Barry. The short answer is 'it depends' :) >> >> GCC is more than a compiler and includes other things such as a >> runtime library (libgcc) and hooks into the libc for features like >> thread local storage. These are built for the architecture and >> floating point unit options you pass to GCC's configure, so if you >> want one toolchain that runs everywhere then you need to configure it >> for the lowest common denominator (normally a ARMv5T in ARM mode with >> no FPU). A similar argument applies to GLIBC. >> >> The next question is, does this matter for your application? What >> workload will your product run and will it be meaningfully affected by >> this lowest common denominator build? >> * If your application uses a lot of floating point, then the lack of >> FPU support in GLIBC matters > > Completely right. I got a linaro 4.5 by: > > sudo add-apt-repository ppa:linaro-maintainers/toolchain > sudo apt-get update > sudo apt-get install gcc-4.5-arm-linux-gnueabi > > Then I did WHETSTONE benchmark on a low-frequency cortex A9 FPGA with vfpv3. > > Result showed linaro by apt-get is using generic glibc in fact. its > gave only 3.3 WHETSTONE MIPS. Then i compiled a glibc with vfpv3 > support to replace the glibc by apt-get, the new toolchain gave 16.7 > WHETSTONE MIPS, 400% improvement in WHETSTONE.
Whetstone is a floating point benchmark, so I'd expect this type of improvement. > >> * If your product has limited memory, then the smaller code size of >> Thumb-2 is worthwhile >> * If you need to squeeze out another 5 % in performance, then using >> ARMv7 instead of ARMv5 will help > > Completely right. I did a edn benchmark from > http://www.mrtc.mdh.se/projects/wcet/wcet_bench/edn/edn.c. EDN can > show memory and fixed-point performance. My testing result shows edn > has a 5-10% performance improvement by linaro 4.5 with armv7_a option. Nice :) This matches what I've seen. > Then my conclusion is that the best choice for cortex A9 with vfpv3 is > 1. using linaro gcc 4.4/4.5 with ARMV7 optimization > 2. using glibc compiled with vfpv3 supporting > > A strange phenomenon i found was nbench would enter dead loop in > ASSIGNMENT case. It could finish ASSIGNMENT case forever. nbench can > be downloaded from > http://www.tux.org/~mayer/linux/nbench-byte-2.2.3.tar.gz. I haven't > time to figure out whether it is a linaro gcc bug. Then i just ignore > this case by writing NBENCH.CONF like the following: > CUSTOMRUN=T > DONUMSORT=T > DOLU=T > DOSTRINGSORT=T > DOBITFIELD=T > DOEMF=T > DOFOUR=T > DOASSIGN=F > ASSIGNMINSECONDS=1 > DOHUFF=T > DOIDEA=T > > If you have time, may you help to check the reason? I couldn't reproduce this, sorry. I built nbench using gcc-linaro-4.5-2011.03-0 using both -O3 -mtune=cortex-a9 -mfpu=vfpv3-d16 and -O3 -mtune=cortex-a9 -mfpu=neon and in both cases ASSIGNMENT completes. -- Michael _______________________________________________ linaro-toolchain mailing list linaro-toolchain@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-toolchain
