I use 128 bit __m128 for vector optimization in my convolution.
To save memory, and processing.
Vector optimization in convolution, you make me curious. For which part in particular?
And with “my convolution” you mean you implemented convolution?
The direct approach, or in the frequency domain?
I’ve had a few people ask if my software is M1 optimized, so…
I am only using 128 bit for complex multiplication on the M1 for FFT and am looking for other methods to utilize vector calculation in my code.
Maybe you made a typo: the complex multiplication is not used for the FFT… But used after the forward FFT and before the inverse FFT.
What about using vDSP for the complex multiplication? If you also need to support Windows, you can use IPP for it. I bet that their implementation will be the fastest you can get (for free!).
So when you say vector optimization, you mean you are using 4 byte reals in a 128 bit register?
Cheers