| 161 | ids=context_choice_ids) |
| 162 | @pytest.mark.parametrize("dtype", dtypes, ids=dtypes) |
| 163 | def test_numba_memalloc(c, dtype): |
| 164 | ctx, nb_ctx = context_choices[c] |
| 165 | dtype = np.dtype(dtype) |
| 166 | # Allocate memory using numba context |
| 167 | # Warning: this will not be reflected in pyarrow context manager |
| 168 | # (e.g bytes_allocated does not change) |
| 169 | size = 10 |
| 170 | mem = nb_ctx.memalloc(size * dtype.itemsize) |
| 171 | darr = DeviceNDArray((size,), (dtype.itemsize,), dtype, gpu_data=mem) |
| 172 | darr[:5] = 99 |
| 173 | darr[5:] = 88 |
| 174 | np.testing.assert_equal(darr.copy_to_host()[:5], 99) |
| 175 | np.testing.assert_equal(darr.copy_to_host()[5:], 88) |
| 176 | |
| 177 | # wrap numba allocated memory with CudaBuffer |
| 178 | cbuf = cuda.CudaBuffer.from_numba(mem) |
| 179 | arr2 = np.frombuffer(cbuf.copy_to_host(), dtype=dtype) |
| 180 | np.testing.assert_equal(arr2, darr.copy_to_host()) |
| 181 | |
| 182 | |
| 183 | @pytest.mark.parametrize("c", range(len(context_choice_ids)), |