Replace numba.generated_jit with njit (pygae#430)

Done by avoid the numpy.result_type (not supported by numba). Instead
do the array multiplication at-once, then use the result dtype.
This requires some numpy array-indexing tricks.

clifford/__init__.py

@@ -172,19 +172,14 @@ def _get_mult_function(mt: sparse.COO):
     k_list, l_list, m_list = mt.coords
     mult_table_vals = mt.data
 
-    @_numba_utils.generated_jit(nopython=True)
+    @_numba_utils.njit
     def mv_mult(value, other_value):
-        # this casting will be done at jit-time
-        ret_dtype = _get_mult_function_result_type(value, other_value, mult_table_vals.dtype)
-        mult_table_vals_t = mult_table_vals.astype(ret_dtype)
-
-        def mult_inner(value, other_value):
-            output = np.zeros(dims, dtype=ret_dtype)
-            for k, l, m, val in zip(k_list, l_list, m_list, mult_table_vals_t):
-                output[l] += value[k] * val * other_value[m]
-            return output
-
-        return mult_inner
+        res = value[k_list] * mult_table_vals * other_value[m_list]
+        output = np.zeros(dims, dtype=res.dtype)
+        # Can not use "np.add.at(output, l_list, res)", as ufunc.at is not supported by numba
+        for l, val in zip(l_list, res):
+            output[l] += val
+        return output
 
     return mv_mult
 
@@ -203,24 +198,16 @@ def _get_mult_function_runtime_sparse(mt: sparse.COO):
     k_list, l_list, m_list = mt.coords
     mult_table_vals = mt.data
 
-    @_numba_utils.generated_jit(nopython=True)
+    @_numba_utils.njit
     def mv_mult(value, other_value):
-        # this casting will be done at jit-time
-        ret_dtype = _get_mult_function_result_type(value, other_value, mult_table_vals.dtype)
-        mult_table_vals_t = mult_table_vals.astype(ret_dtype)
-
-        def mult_inner(value, other_value):
-            output = np.zeros(dims, dtype=ret_dtype)
-            for ind, k in enumerate(k_list):
-                v_val = value[k]
-                if v_val != 0.0:
-                    m = m_list[ind]
-                    ov_val = other_value[m]
-                    if ov_val != 0.0:
-                        l = l_list[ind]
-                        output[l] += v_val * mult_table_vals_t[ind] * ov_val
-            return output
-        return mult_inner
+        # Use mask where both operands are non-zero, to avoid zero-multiplications
+        nz_mask = (value != 0.0)[k_list] & (other_value != 0.0)[m_list]
+        res = value[k_list[nz_mask]] * mult_table_vals[nz_mask] * other_value[m_list[nz_mask]]
+        output = np.zeros(dims, dtype=res.dtype)
+        # Can not use "np.add.at(output, l_list[nz_mask], res)", as ufunc.at is not supported by numba
+        for l, val in zip(l_list[nz_mask], res):
+            output[l] += val
+        return output
 
     return mv_mult