Errors - Githubissues

I have some results from tests of both precisions. Two measurements of errors has been carried out. Both of them are the comparison of every single element in the reuslt of the GPU implementation [gpu], and the original python implementation [cpu].

absolute error = abs(cpu_i - gpu_i) relative error = abs( (cpu_i - gpu_i)/cpu_i )

_Following results are produced by unit_test2.py_

def perband_emulators ( emulators, band_pass)

Notice: the errors of float32 are comparisons between CUDA double and Python single precision, since python functions, e.g. scipy.spatial.distance do not yield float32 output, regardless its inputs data type.

mu float64 absolute_val 0.21 error_absolute 8.1e-13 [average] 3.5e-12 [max] error_relative 8.3e-12 [average] 3.8e-10 [max]

mu float32 absolute_val 0.21 error_absolute 0.00032 [average] 0.0014 [max] error_relative 0.0031 [average] 0.15 [max]

var float64 absolute_val 5.5e-07 error_absolute 1e-10 [average] 6.2e-10 [max] error_relative 0.00055 [average] 0.006 [max]

var float32 absolute_val 5.5e-07 error_absolute 0.1 [average] 0.55 [max] error_relative 5.4e+05 [average] 5.9e+06 [max]

deriv float64 absolute_val 0.088 error_absolute 1.3e-13 [average] 3.8e-12 [max] error_relative 5.6e-12 [average] 4.6e-09 [max]

deriv float32 absolute_val 0.088 error_absolute 4.1e-05 [average] 0.00094 [max] error_relative 0.0014 [average] 1 [max]

UCL / gp_emulator

Errors #45