Re: [Discuss-gnuradio] Problem with volk and fftw memory alignment

[Tom Rondeau]

On Fri, May 29, 2015 at 4:33 PM, <address@hidden> wrote:

I've been trying out the (bleeding edge) corr_est() code and the test_corr_est.grc sometimes segvs. Not repeatable, won't crash under gdb :-) From a core file, it's crashing loading the first element of aVector in volk_32fc_x2_multiply_32fc_a_avx
instruction is vmovaps (%eax),%ymm1    where eax is 0x8b8b590.

My CPU supports the 256bit AVX instructions, and I believe such data needs to be 256bit=32 byte aligned. EAX here isn't. Looking at the backtrace, it's crashing in fft_filter_ccc::filter, where the d_fwdfft is generated from fft::fft_complex

In turn this creates d_inbuf by calling fftwf_malloc. I believe this only guarantees 16 byte alignment. (I'm on fftw3.3.4, on a 32 bit Linux 3.18.1, CPU Intel i5-3470)

As we move to use volk and the fast SIMD, should the code in fft.cc (and perhaps other places) move to using a volk memory allocator to get the right alignment, rather than fftwf_malloc?

I'm on gentoo, so I could fix the source to fftwf_malloc to return 32 byte aligned, but that's not a general solution.

I'm not sure how I should report this issue, please pass on my report if there's a better place to discuss such things
--
 G8SQH
<address@hidden>

Yes, this is absolutely a problem with using the fftw_malloc functions. Unless told, FFTW doesn't build AVX support, which means that it can return vectors that are not AVX-aligned, and the code is calling the _a kernel and so requires it to be appropriately aligned.

Can you change and fftw_malloc'd arrays used in volk calls to volk_malloc and a) make sure that works and b) submit a patch with where you're running into this? I thought we had fixed these already (also, what version are you running?).

Tom