in the file blis.h, change the line #elif defined(__GNUC__) to #elif 0 vector extension is useful for those cpu such as powerpc that have simd unit but lack blis microkernel. For arm family, only arm64 neon is a decent simd unit. armv7 neon does not support double precision float. vpf is not even a simd unit.
On Jul 14, 2017 1:54 AM, "Thomas Costigliola" <fo...@iocane.net> wrote: Undefining __GNUC__ seemed to work. On 07/13/2017 01:51 PM, Thomas Costigliola wrote: > Thanks Bill, but how do I disable it in blis.h? > > On 07/12/2017 06:05 PM, Bill wrote: > >> gcc vector extension seems available for some years >> https://gcc.gnu.org/onlinedocs/gcc-3.3.4/gcc/Vector-Extensions.html >> >> It also complied successfully for raspberry using linaro gcc toolchain. >> Perhaps kindle toolchain is special. If you cannot fix it then you may >> disable it in blis.h and let it fall through to c ref version. Performance >> will be more or less the same (no improvement) if there is no decent >> hardware simd support. >> >> Sent from my iPhone >> >> On 13 Jul, 2017, at 3:29 AM, Thomas Costigliola <fo...@iocane.net> wrote: >> >> I am using an old version of GCC built to target the kindle (arm >>> architecture). >>> >>> gcc version 4.5.4 20110505 (prerelease) (Linaro GCC 4.5-2011.05-0) >>> >>> The new BLIS routines are causing compilation failures as follows: >>> >>> jsrc/blis/gemm_vec-ref.c: In function 'bli_zgemm_128_3x2': >>> jsrc/blis/gemm_vec-ref.c:582:11: error: subscripted value is neither >>> array nor pointer >>> jsrc/blis/gemm_vec-ref.c:582:11: error: subscripted value is neither >>> array nor pointer >>> jsrc/blis/gemm_vec-ref.c:583:11: error: subscripted value is neither >>> array nor pointer >>> >>> >>> It seems like a GCC version, behavior issue. Is there either 1) a quick >>> syntax fix that will get it to compile or 2) a way to disable the BLIS >>> routines? Or are they required in the latest J engine? >>> >>> Thanks, >>> -Thomas >>> >>> ---------------------------------------------------------------------- >>> For information about J forums see http://www.jsoftware.com/forums.htm >>> >> ---------------------------------------------------------------------- >> For information about J forums see http://www.jsoftware.com/forums.htm >> >> > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm > ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
