FFT performance on the 5s 64-bit mode vs. 32-bit mode

perf.c

/*
100,000 iterations of a 1024 point real->imag FFT, computing magnitude & phase and then its inverse again
Run twice, once with double precision and then again using the single precision float vDSP

5s in 32 bit mode
2013-09-29 13:48:47.009 perf[3513:60b] doFFT double precision took 157.924796 seconds
2013-09-29 13:49:13.707 perf[3513:60b] doFFT single precision took 26.697247 seconds

5s in 64 bit mode
2013-09-29 13:51:47.402 perf[3524:60b] doFFT double took 57.861582 seconds (2.7x speedup over 32-bit mode)
2013-09-29 13:52:06.063 perf[3524:60b] doFFT single took 18.660364 seconds (1.4x speedup over 32-bit mode)

The code being iterated (N=1024):
*/

void dofftD() {
    vDSP_ctozD((DSPDoubleComplex*)xD, 2, &tempSplitComplexD, 1, N/2);
    vDSP_fft_zripD(fftSetupD, &tempSplitComplexD, 1, LOG_N, kFFTDirection_Forward);
    
    vDSP_zvabsD(&tempSplitComplexD, 1, magD, 1, N/2);
    vDSP_zvphasD(&tempSplitComplexD, 1, phaseD, 1, N/2);
    
    tempSplitComplexD.realp = magD;
    tempSplitComplexD.imagp = phaseD;
    
    vDSP_ztocD(&tempSplitComplexD, 1, tempComplexD, 2, N/2);
    vDSP_rectD((double*)tempComplexD, 2, (double*)tempComplexD, 2, N/2);
    vDSP_ctozD(tempComplexD, 2, &tempSplitComplexD, 1, N/2);
    
    vDSP_fft_zripD(fftSetupD, &tempSplitComplexD, 1, LOG_N, kFFTDirection_Inverse);
    vDSP_ztocD(&tempSplitComplexD, 1, (DSPDoubleComplex*)yD, 2, N/2);

    double scale = 0.5/N;
    vDSP_vsmulD(yD, 1, &scale, yD, 1, N);
}