@Cuda.jit is there cache=True?

Numba Support: How do I do ...?
1

Is it possible to cache cuda jit compiled code? I tried to pass cache=True in @cuda.jit but can’t find any compiled code in the cache dir.

Thanks in advance!

Best,
Kris

1 Like
2

Are you using a Python script or Jupyter notebook?

3

Thanks for your reply. For both, in a jupyter notebook and in a python script, there have been no entries generated in the corresponding cache directories. There are, however, entries for compiled CPU code.

This is a simple test code which generates code for subtract but not for sub_cuda:

from numba import cuda
import numba as nb
import numpy as np

@nb.jit("int64[:](int64[:],int64[:])",cache=True)
def subtract(a,b):
    return a - b

@cuda.jit("void(int64[:],int64[:],int64[:])",cache=True)
def sub_cuda(a,b,c):
    start = cuda.grid(1)
    stride = cuda.gridsize(1)
    for i in range(start, a.shape[0], stride):
        c[i] = a[i] - b[i]

a = np.ones(3,dtype=np.int64)
b = np.ones(3,dtype=np.int64)
c = np.ndarray(3,dtype=np.int64)
sub_cuda[64,128](a,b,c)
d = subtract(a,b)
print(c,d)
4

Unfortunately there is no support for caching in the CUDA target at present. (Apologies for brevity, on mobile)

3 Likes