stuartarchibald
commented
4 years ago Thanks for the report. SVML should make your code faster, but only where vectorized math functions are applicable.
In the code sample provided, this line:
ratio = a[i] / b[i]would appear to Numba as doing a division with a complex scalar as the numerator and denominator. As Numba doesn't differentiate between NumPy and Python scalars, this would end up as a division operation based on the python semantics, such that a branch would be added into the loop to handle raising a ZeroDivisionError in the case of b[i] being zero, this branch would prevent the loop vectorizing as well as possible. Further, not specifying 'reassoc' in the flags means the loop cannot vectorize as the reduction on error has to be executed in order. Net result of these two issues results in no SVML instructions being selected.
"Fixing" the code to look like:
import numpy as np
import numba as nb
@nb.njit(fastmath=True)
# ^--- set to True, selectively applying parts of fastmath flags rarely goes well in practice.
# See https://github.com/numba/numba/pull/3847#issuecomment-472822965
def foo(a, b):
assert a.size == b.size
error = 0.0
for i in range(a.size):
ratio = np.divide(a[i], b[i]) # <--- explicit numpy divide
magnitude_error = np.log10(np.abs(ratio))
phase_error = ((np.angle(ratio) + np.pi) % (2.0 * np.pi)) - np.pi
error += ((magnitude_error**2.0 + phase_error**2.0) / 2.0)
return error
a = np.arange(1, 1000) + 1j * np.arange(1, 1000)
b = np.arange(1, 1000) + 1j * np.arange(1, 1000)
# Compile function
foo(a, b)
print(foo.inspect_asm(foo.signatures[0]))should print out assembly that contains things like:
movabsq $__svml_hypot2, %rax
callq *%rax
movabsq $__svml_log102, %rax
callq *%raxwhere the __svml_ prefix indicates use of an SVML intrinsic.
Hope this helps?
cmartinezdem
davo417
AdityaSoni19031997
Hello,
I am running Numba 0.48 (last version available in PyPi) on Windows 10 (64 bits) with Python 3.8.2 and I am not able to get better performance after installing Intel's icc_rt library. It seems that SVML libraries are detected by Numba, as it can be checked in
numba -soutput:However, the following code takes the same amount of time to run with or without SVML libraries:
I tried to install Numba and icc_rt with conda and I get the same result. Shouldn't perform better with SVML libraries?
Thank you so much for your support.