Enable OpenCL tests on gh actions. (#317)

* Enable opencl tests and errors with memoization.

This used to work with old versions of pytools but caused many failures
with recent versions.  Fix issue where the CUDA tests would fail. This was 
because the memoization was ignoring the kwargs.  Also fix an issue 
when using start_idx on CUDA due to type inconsistencies.

.github/workflows/tests.yml

@@ -25,8 +25,8 @@ jobs:
       - name: Install dependencies on Linux/MacOS
         run: |
           conda info
-          #conda install -c conda-forge pocl=1.5 pyopencl
-          #python -c 'import pyopencl as cl'
+          conda install -c conda-forge pocl pyopencl
+          python -c 'import pyopencl as cl'
         if: ${{ runner.os != 'Windows' }}
       - name: Install dependencies
         run: |

pysph/base/gpu_helper_kernels.py

@@ -4,12 +4,16 @@
 from pytools import memoize
 
 
-@memoize(key=lambda *args: tuple(args))
+def key_func(*args, **kw):
+    return args, frozenset(kw.items())
+
+
+@memoize(key=key_func)
 def get_elwise(f, backend):
     return Elementwise(f, backend=backend)
 
 
-@memoize(key=lambda *args: tuple(args))
+@memoize(key=key_func)
 def get_scan(inp_f, out_f, dtype, backend):
     return Scan(input=inp_f, output=out_f, dtype=dtype,
                 backend=backend)

pysph/sph/acceleration_eval_gpu_helper.py

@@ -339,7 +339,7 @@ def _call_kernel(self, info, extra_args):
             n = dest.get_number_of_particles(info.get('real', True))
         else:
             n = self._get_index(dest, stop_idx)
-        n_iter = n - start_idx
+        n_iter = int(n - start_idx)
 
         # args is actually [queue, None, None, actual_meaningful_args]
         # we do not need the first 3 args on CUDA.