fix: address review feedback for PR #184

Author: ReviewLoop

benchmarks/benchmarks/bench_continuous.py

@@ -40,11 +40,14 @@
     "normal": ("normal", (1.0, 2.0)),
     "standard_exponential": ("standard_exponential", ()),
     "exponential": ("exponential", (2.0,)),
-    # MKL's GNORM gamma method switches algorithm at shape 1 and at 0.6
+    # MKL's GNORM gamma method may select different internal algorithms
+    # around shape 1 and 0.6; these shapes exercise those regimes
     "standard_gamma": ("standard_gamma", (3.0,)),
     "standard_gamma_mid_shape": ("standard_gamma", (0.8,)),
     "standard_gamma_small_shape": ("standard_gamma", (0.5,)),
-    # MKL's CJA beta method uses Cheng, Johnk or Atkinson depending on shapes
+    # MKL's CJA beta method may select different internal algorithms (Cheng,
+    # Johnk or Atkinson) depending on shapes; these shapes exercise those
+    # regimes
     "beta": ("beta", (2.0, 5.0)),
     "beta_small_shape": ("beta", (0.5, 0.5)),
     "beta_mixed_shape": ("beta", (0.5, 2.0)),

benchmarks/benchmarks/bench_discrete.py

@@ -34,16 +34,16 @@
 
 # name -> (MKLRandomState method, positional arguments before ``size``)
 _DISCRETE = {
-    # MKL's BTPE binomial method uses acceptance/rejection only when
-    # n * min(p, 1 - p) >= 30
+    # MKL's BTPE binomial method may select a different internal algorithm
+    # when n * min(p, 1 - p) >= 30; these parameters exercise both regimes
     "binomial": ("binomial", (10, 0.5)),
     "binomial_large_n": ("binomial", (1000, 0.3)),
     "negative_binomial": ("negative_binomial", (5, 0.5)),
     # default method (POISNORM); per-method timings are in bench_methods.py
     "poisson": ("poisson", (10.0,)),
     "geometric": ("geometric", (0.3,)),
-    # MKL's H2PE hypergeometric method uses acceptance/rejection for a large
-    # mode
+    # MKL's H2PE hypergeometric method may select a different internal
+    # algorithm for a large mode; these parameters exercise both regimes
     "hypergeometric": ("hypergeometric", (10, 20, 5)),
     "hypergeometric_large_mode": ("hypergeometric", (1000, 2000, 500)),
     "zipf": ("zipf", (2.0,)),

benchmarks/benchmarks/bench_methods.py

@@ -37,7 +37,8 @@
 # lognormal accepts no BoxMuller2
 _LOGNORMAL_METHODS = ["ICDF", "BoxMuller"]
 _POISSON_METHODS = ["POISNORM", "PTPE"]
-# PTPE uses table lookup below lam = 27 and acceptance/rejection above
+# MKL's PTPE poisson method may select a different internal algorithm
+# depending on lam (around lam = 27); these values exercise both regimes
 _POISSON_LAMS = [10.0, 100.0]