Convolution produces garbage result with SSE2?

[ienapliss]

In my project, I simply copy the files and define MUFFT_HAVE_X86 and MUFFT_HAVE_SSE inside fft_internals.h.
Then I compile my program with gcc (mingw on windows) with -msse2 for enabling sse2.

If I disable SIMD in the flags, the output results is OK.
mufft_create_plan_conv(N, MUFFT_FLAG_CPU_NO_SSE | MUFFT_FLAG_CPU_NO_AVX | MUFFT_FLAG_CPU_NO_SSE3, MUFFT_CONV_METHOD_FLAG_MONO_MONO);

It's not a problem of alignment (all my data is aligned even to 64 bytes). I tested the pffft library with the same data and simd works just fine.