Add a Numba implementation for Generator.dirichlet

Author: brandonwillard

aesara/link/numba/dispatch/random.py

@@ -6,7 +6,9 @@
 import numba.np.unsafe.ndarray as numba_ndarray
 import numpy as np
 from numba import types
-from numba.extending import overload
+from numba.extending import overload, overload_method, register_jitable
+from numba.np.random.distributions import random_beta, random_standard_gamma
+from numba.np.random.generator_methods import check_size, check_types, is_nonelike
 
 import aesara.tensor.random.basic as aer
 from aesara.graph.basic import Apply
@@ -296,3 +298,78 @@ def dirichlet_rv(rng, size, dtype, alphas):
             return (rng, rng.dirichlet(alphas, size))
 
     return dirichlet_rv
+
+
+@register_jitable
+def random_dirichlet(bitgen, alpha, size):
+    """
+    This implementation is straight from ``numpy/random/_generator.pyx``.
+    """
+
+    k = len(alpha)
+    alpha_arr = np.asarray(alpha, dtype=np.float64)
+
+    if np.any(np.less_equal(alpha_arr, 0)):
+        raise ValueError("alpha <= 0")
+
+    shape = size + (k,)
+
+    diric = np.zeros(shape, np.float64)
+
+    i = 0
+    totsize = diric.size
+
+    if (k > 0) and (alpha_arr.max() < 0.1):
+        alpha_csum_arr = np.empty_like(alpha_arr)
+        csum = 0.0
+        for j in range(k - 1, -1, -1):
+            csum += alpha_arr[j]
+            alpha_csum_arr[j] = csum
+
+        while i < totsize:
+            acc = 1.0
+            for j in range(k - 1):
+                v = random_beta(bitgen, alpha_arr[j], alpha_csum_arr[j + 1])
+                diric[i + j] = acc * v
+                acc *= 1.0 - v
+            diric[i + k - 1] = acc
+            i = i + k
+
+    else:
+        while i < totsize:
+            acc = 0.0
+            for j in range(k):
+                diric[i + j] = random_standard_gamma(bitgen, alpha_arr[j])
+                acc = acc + diric[i + j]
+            invacc = 1.0 / acc
+            for j in range(k):
+                diric[i + j] = diric[i + j] * invacc
+            i = i + k
+
+    return diric
+
+
+@overload_method(types.NumPyRandomGeneratorType, "dirichlet")
+def NumPyRandomGeneratorType_dirichlet(inst, alphas, size=None):
+    check_types(alphas, [types.Array, types.List], "alphas")
+
+    if isinstance(size, types.Omitted):
+        size = size.value
+
+    if is_nonelike(size):
+
+        def impl(inst, alphas, size=None):
+            return random_dirichlet(inst.bit_generator, alphas, ())
+
+    elif isinstance(size, (int, types.Integer)):
+
+        def impl(inst, alphas, size=None):
+            return random_dirichlet(inst.bit_generator, alphas, (size,))
+
+    else:
+        check_size(size)
+
+        def impl(inst, alphas, size=None):
+            return random_dirichlet(inst.bit_generator, alphas, size)
+
+    return impl

tests/link/numba/test_random.py

@@ -520,7 +520,6 @@ def test_CategoricalRV(dist_args, size, cm):
         )
 
 
-@pytest.mark.skip(reason="Not yet supported in Numba via `Generator`s")
 @pytest.mark.parametrize(
     "a, size, cm",
     [