Memory alignment issues (DSP-80)

[Barabas5532]

There are some places in this library where the user must align memory passed to functions correctly. These are not documented anywhere I can find.

E.G. The dsps_fft4r_fc32_ansi_ function casts the float *data parameter to the following struct:

typedef union fc32_u
{
    struct
    {
        float re;
        float im;
    };
    uint64_t data;
}fc32_t;

The uint64_t member of this union requires 8 bytes alignment. Actually, this member is unused, so it would be best to remove it, and make the default 4 byte alignment work. Why doesn't this library use the C99 complex type feature?

I've also noticed that 16 byte alignment was added to a lot of the example apps with the support for ESP32S3 chip in 108493c. Is this alignment only required for implementations using assembly code for that chip?

esp-dsp/modules/common/include/dsp_tests.h

Line 34 in 3175e82

#define memalign(align_, size_) malloc(size_)

The tests disable alignment on esp-idf version <4.3.0 by defining memalign as malloc and ignoring the alignment parameter. Would this not cause issues when loading memory from an unaligned address?

[igrr]

The tests disable alignment on esp-idf version <4.3.0 by defining memalign as malloc and ignoring the alignment parameter.

I think the reason is that older versions of IDF didn't have memalign implemented. And the 16 byte alignment is necessary for esp32-s3, which is supported only since IDF 4.4.

Good point about documenting the alignment requirements, we'll update the docs!