Floating-point bit twiddling in CUDA

Feature request

I need to use view individual floating-point numbers as integers and backwards. For example:

np.array(np.array(23.65).view('i8') | 1).view('d').item()

should give 23.650000000000002 in CUDA device functions.

I also need to use frexp and ldexp in CUDA device functions.

[gmarkall]

I also need to use frexp and ldexp in CUDA device functions.

@zhihaoy You can use the code in https://numba.discourse.group/t/cuda-building-extensions-to-call-libdevice-functions-cube-root-example/140 as a starting point for making a Numba extension to call __nv_ldexp and __nv_frexp as a workaround for now - if you modify the names of the functions (and the Python implementation of cbrt so that it is ldexp / frexp) then that should give you ldexp and frexp functions you can call from CUDA kernels.

[gmarkall]

Now that @overload is supported on the CUDA target, I have made a start on reusing the implementation in the core of Numba for this: https://github.com/gmarkall/numba/tree/cuda-scalar-view - the change itself is only:

diff --git a/numba/cpython/numbers.py b/numba/cpython/numbers.py
index d01152359..9021e3ee4 100644
--- a/numba/cpython/numbers.py
+++ b/numba/cpython/numbers.py
@@ -1359,5 +1359,5 @@ def scalar_view(scalar, viewty):
         return impl
 
 
-overload_method(types.Float, 'view')(scalar_view)
-overload_method(types.Integer, 'view')(scalar_view)
+overload_method(types.Float, 'view', target='generic')(scalar_view)
+overload_method(types.Integer, 'view', target='generic')(scalar_view)

however there are some test failures for this functionality on the CUDA target - I have not yet looked into whether there is a problem in the implementation related to CUDA, or if the tests need some adjustment.