Simple fix that should save 2x in memory during calculation. As a simple check for complex64's, I computed the wavelet transform for a signal, n drawn from N(0, 1) and another signal x = 100*unit_step + n for both 128 and 64 precision. In both cases, the absolute difference between 128 and 64 was < 1e-6, so I think we should be fine with 64.
Simple fix that should save 2x in memory during calculation. As a simple check for complex64's, I computed the wavelet transform for a signal, n drawn from N(0, 1) and another signal x = 100*unit_step + n for both 128 and 64 precision. In both cases, the absolute difference between 128 and 64 was < 1e-6, so I think we should be fine with 64.