bfly load/stores

[fbarchard]

The bfly functions do multiple loads and stores of Fout e.g. C_ADDTO(*Fout,scratch[3]);

Code size and performance is improved by loading Fout at the top of the loop and storing it at the bottom, using registers for the operations. On ARMv7 the bfly4 function main loop

Was 170 instructions 46 loads 36 stores

Now 140 instructions 35 loads 19 stores

This is bfly4 with the change:

static void kf_bfly4(
        kiss_fft_cpx * Fout,
        const size_t fstride,
        const kiss_fft_cfg st,
        const size_t m
        )
{
    kiss_fft_cpx *tw1,*tw2,*tw3;
    kiss_fft_cpx scratch[6];
    kiss_fft_cpx Fout0, Fout1, Fout2, Fout3;
    size_t k=m;
    const size_t m2=2*m;
    const size_t m3=3*m;


    tw3 = tw2 = tw1 = st->twiddles;

    do {
        Fout0 = Fout[0];
        Fout1 = Fout[m];
        Fout2 = Fout[m2];
        Fout3 = Fout[m3];
        C_FIXDIV(Fout0,4); C_FIXDIV(Fout1,4); C_FIXDIV(Fout2,4); C_FIXDIV(Fout3,4);

        C_MUL(scratch[0],Fout1 , *tw1 );
        C_MUL(scratch[1],Fout2 , *tw2 );
        C_MUL(scratch[2],Fout3 , *tw3 );

        C_SUB( scratch[5] , Fout0, scratch[1] );
        C_ADDTO(Fout0, scratch[1]);
        C_ADD( scratch[3] , scratch[0] , scratch[2] );
        C_SUB( scratch[4] , scratch[0] , scratch[2] );
        C_SUB( Fout2, Fout0, scratch[3] );
        tw1 += fstride;
        tw2 += fstride*2;
        tw3 += fstride*3;
        C_ADDTO( Fout0 , scratch[3] );

        if(st->inverse) {
            Fout1.r = scratch[5].r - scratch[4].i;
            Fout1.i = scratch[5].i + scratch[4].r;
            Fout3.r = scratch[5].r + scratch[4].i;
            Fout3.i = scratch[5].i - scratch[4].r;
        }else{
            Fout1.r = scratch[5].r + scratch[4].i;
            Fout1.i = scratch[5].i - scratch[4].r;
            Fout3.r = scratch[5].r - scratch[4].i;
            Fout3.i = scratch[5].i + scratch[4].r;
        }
        Fout[0]  = Fout0;
        Fout[m]  = Fout1;
        Fout[m2] = Fout2;
        Fout[m3] = Fout3;
        ++Fout;
    }while(--k);
}

[fbarchard]

The ADDTO macro seems like it would work ok on intel, but the above code is smaller/faster on x86 as well. So I'd suggest removing that macro and using C_ADD, simplifying the implementation and improving code generated.

[VioletGiraffe]

Hi, thank you for your work, fast FFT on ARMv7a is exactly what I need. Can I assume your have tested this change in terms of correctness?

[fbarchard]

yes, this is a small change with no impact on quality. I only did bfly4... it should be done for the bfly2 and others. And then we can remove C_ADDTO and use C_ADD