Conversation
numba/dppl/tests/dppl/test_ufuncs.py
Outdated
|
|
||
| input_types = (input_type,) * ufunc.nin | ||
| output_types = (output_type,) * ufunc.nout | ||
| cfunc = dppl.kernel(pyfunc) |
There was a problem hiding this comment.
We are supposed to test the numpy compatibility for numba.njit(parallel:{'offload':True}) as well
| continue | ||
|
|
||
| output_type = self._determine_output_type( | ||
| input_type, int_output_type, float_output_type) |
There was a problem hiding this comment.
The issue with testing some functions (e.g. sin, cos) for integer is the incompatibility of the dtype of the output buffer we pass and the actual output. For integer input we need a float type output for function like cos. We need to make changes in _determine_output_type() to accomodate this.
numba/dppl/tests/dppl/test_ufuncs.py
Outdated
| enable_pyobj_flags.set("no_rewrites") | ||
|
|
||
| no_pyobj_flags = Flags() | ||
| no_pyobj_flags.set("no_rewrites") |
There was a problem hiding this comment.
This flag might prevent rewrite of numpy calls to array expressions, which in turn will prevent us from generating parfor nodes. The parfor nodes trigger the gpu code generation.
|
@reazulhoque I fixed all according your comments, could you please take a look? |
| input_types = (input_type,) * ufunc.nin | ||
| output_types = (output_type,) * ufunc.nout | ||
| cfunc_kernel = dppl.kernel(pyfunc) | ||
| cfunc = njit(parallel={'offload':True})(pyfunc) |
There was a problem hiding this comment.
Unfortunately numba is not rewriting this ufuncs as array expression because the output argument is being provided. (https://github.com/IntelPython/numba/blob/pydppl/numba/np/ufunc/array_exprs.py#L88)
We need to find a solution to this as without the ufuncs being rewritten as array expression we will not generate a parfor node and in turn will not create a GPU kernel for it.
2 participants
No description provided.