[test] Use numpy's tolerance for float16 (#491)

Some CI runs were spuriously failing due to overly stringent tolerance
for fp16 tests (#490). Use numpy's epsilon for 16b floats instead of the
default relative tolerance for np.testing.assert_allclose().

This PR passes `CUDATestCase.FLOAT16_RTOL` to relevant calls to
`np.testing.assert_allclose`.
Alternatively, we could add `assertAllClose` to `CUDATestCase` that
infers the appropriate relative tolerance from the data types passed in.
Please let me know if we'd like to go that route instead. Thanks in
advance!

<!--

Thank you for contributing to numba-cuda :)

Here are some guidelines to help the review process go smoothly.

1. Please write a description in this text box of the changes that are
being
   made.

2. Please ensure that you have written units tests for the changes
made/features
   added.

3. If you are closing an issue please use one of the automatic closing
words as
noted here:
https://help.github.com/articles/closing-issues-using-keywords/

4. If your pull request is not ready for review but you want to make use
of the
continuous integration testing facilities please label it with `[WIP]`.

5. If your pull request is ready to be reviewed without requiring
additional
work on top of it, then remove the `[WIP]` label (if present) and
replace
it with `[REVIEW]`. If assistance is required to complete the
functionality,
for example when the C/C++ code of a feature is complete but Python
bindings
are still required, then add the label `[HELP-REQ]` so that others can
triage
and assist. The additional changes then can be implemented on top of the
same PR. If the assistance is done by members of the rapidsAI team, then
no
additional actions are required by the creator of the original PR for
this,
otherwise the original author of the PR needs to give permission to the
person(s) assisting to commit to their personal fork of the project. If
that
doesn't happen then a new PR based on the code of the original PR can be
opened by the person assisting, which then will be the PR that will be
   merged.

6. Once all work has been done and review has taken place please do not
add
features or make changes out of the scope of those requested by the
reviewer
(doing this just add delays as already reviewed code ends up having to
be
re-reviewed/it is hard to tell what is new etc!). Further, please do not
rebase your branch on main/force push/rewrite history, doing any of
these
   causes the context of any comments made by reviewers to be lost. If
   conflicts occur against main they should be resolved by merging main
   into the branch used for making the pull request.

Many thanks in advance for your cooperation!

-->

Author: ashermancinelli

numba_cuda/numba/cuda/testing.py

@@ -17,6 +17,7 @@
 from typing import Iterable, Union
 from io import StringIO
 import unittest
+import numpy as np
 
 if PYVERSION >= (3, 10):
     from filecheck.matcher import Matcher
@@ -44,6 +45,8 @@ class CUDATestCase(TestCase):
     matches FileCheck checks, and is not specific to CUDADispatcher.
     """
 
+    FLOAT16_RTOL = np.finfo(np.float16).eps
+
     def setUp(self):
         self._low_occupancy_warnings = config.CUDA_LOW_OCCUPANCY_WARNINGS
         self._warn_on_implicit_copy = config.CUDA_WARN_ON_IMPLICIT_COPY

numba_cuda/numba/cuda/tests/cudapy/test_intrinsics.py

@@ -629,7 +629,7 @@ def test_hadd(self):
         arg1 = np.array([3.0], dtype=np.float16)
         arg2 = np.array([4.0], dtype=np.float16)
         compiled[1, 1](ary, arg1, arg2)
-        np.testing.assert_allclose(ary[0], arg1 + arg2)
+        np.testing.assert_allclose(ary[0], arg1 + arg2, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_hadd_scalar(self):
@@ -639,7 +639,7 @@ def test_hadd_scalar(self):
         arg2 = np.float16(3.0)
         compiled[1, 1](ary, arg1, arg2)
         ref = arg1 + arg2
-        np.testing.assert_allclose(ary[0], ref)
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
 
     @skip_on_cudasim("Compilation unsupported in the simulator")
     @skip_if_nvjitlink_missing("Numbast generated bindings")
@@ -657,7 +657,9 @@ def test_hfma(self):
         arg2 = np.array([3.0], dtype=np.float16)
         arg3 = np.array([4.0], dtype=np.float16)
         compiled[1, 1](ary, arg1, arg2, arg3)
-        np.testing.assert_allclose(ary[0], arg1 * arg2 + arg3)
+        np.testing.assert_allclose(
+            ary[0], arg1 * arg2 + arg3, rtol=self.FLOAT16_RTOL
+        )
 
     @skip_unless_cc_53
     def test_hfma_scalar(self):
@@ -668,7 +670,7 @@ def test_hfma_scalar(self):
         arg3 = np.float16(4.0)
         compiled[1, 1](ary, arg1, arg2, arg3)
         ref = arg1 * arg2 + arg3
-        np.testing.assert_allclose(ary[0], ref)
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
 
     @skip_on_cudasim("Compilation unsupported in the simulator")
     @skip_if_nvjitlink_missing("Numbast generated bindings")
@@ -687,7 +689,7 @@ def test_hsub(self):
         arg1 = np.array([3.0], dtype=np.float16)
         arg2 = np.array([4.0], dtype=np.float16)
         compiled[1, 1](ary, arg1, arg2)
-        np.testing.assert_allclose(ary[0], arg1 - arg2)
+        np.testing.assert_allclose(ary[0], arg1 - arg2, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_hsub_scalar(self):
@@ -697,7 +699,7 @@ def test_hsub_scalar(self):
         arg2 = np.float16(1.57)
         compiled[1, 1](ary, arg1, arg2)
         ref = arg1 - arg2
-        np.testing.assert_allclose(ary[0], ref)
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
 
     @skip_on_cudasim("Compilation unsupported in the simulator")
     @skip_if_nvjitlink_missing("Numbast generated bindings")
@@ -714,7 +716,7 @@ def test_hmul(self):
         arg1 = np.array([3.0], dtype=np.float16)
         arg2 = np.array([4.0], dtype=np.float16)
         compiled[1, 1](ary, arg1, arg2)
-        np.testing.assert_allclose(ary[0], arg1 * arg2)
+        np.testing.assert_allclose(ary[0], arg1 * arg2, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_hmul_scalar(self):
@@ -724,7 +726,7 @@ def test_hmul_scalar(self):
         arg2 = np.float16(1.57)
         compiled[1, 1](ary, arg1, arg2)
         ref = arg1 * arg2
-        np.testing.assert_allclose(ary[0], ref)
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
 
     @skip_on_cudasim("Compilation unsupported in the simulator")
     @skip_if_nvjitlink_missing("Numbast generated bindings")
@@ -743,7 +745,7 @@ def test_hdiv_scalar(self):
 
         compiled[1, 1](ary, arg1, arg2)
         ref = arg1 / arg2
-        np.testing.assert_allclose(ary[0], ref)
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_hdiv(self):
@@ -754,15 +756,15 @@ def test_hdiv(self):
 
         compiled.forall(ary.size)(ary, arry1, arry2)
         ref = arry1 / arry2
-        np.testing.assert_allclose(ary, ref)
+        np.testing.assert_allclose(ary, ref, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_hneg(self):
         compiled = cuda.jit("void(f2[:], f2[:])")(simple_hneg)
         ary = np.zeros(1, dtype=np.float16)
         arg1 = np.array([3.0], dtype=np.float16)
         compiled[1, 1](ary, arg1)
-        np.testing.assert_allclose(ary[0], -arg1)
+        np.testing.assert_allclose(ary[0], -arg1, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_hneg_scalar(self):
@@ -771,7 +773,7 @@ def test_hneg_scalar(self):
         arg1 = np.float16(3.1415926)
         compiled[1, 1](ary, arg1)
         ref = -arg1
-        np.testing.assert_allclose(ary[0], ref)
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
 
     @skip_on_cudasim("Compilation unsupported in the simulator")
     @skip_if_nvjitlink_missing("Numbast generated bindings")
@@ -787,7 +789,7 @@ def test_habs(self):
         ary = np.zeros(1, dtype=np.float16)
         arg1 = np.array([-3.0], dtype=np.float16)
         compiled[1, 1](ary, arg1)
-        np.testing.assert_allclose(ary[0], abs(arg1))
+        np.testing.assert_allclose(ary[0], abs(arg1), rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_habs_scalar(self):
@@ -796,7 +798,7 @@ def test_habs_scalar(self):
         arg1 = np.float16(-3.1415926)
         compiled[1, 1](ary, arg1)
         ref = abs(arg1)
-        np.testing.assert_allclose(ary[0], ref)
+        np.testing.assert_allclose(ary[0], ref, rtol=self.FLOAT16_RTOL)
 
     @skip_on_cudasim("Compilation unsupported in the simulator")
     @skip_if_nvjitlink_missing("Numbast generated bindings")
@@ -849,15 +851,15 @@ def test_fp16_intrinsics_common(self):
                 kernel = cuda.jit("void(f2[:], f2[:])")(kernel)
                 kernel[1, N](r, x)
                 expected = fn(x, dtype=np.float16)
-                np.testing.assert_allclose(r, expected)
+                np.testing.assert_allclose(r, expected, rtol=self.FLOAT16_RTOL)
 
         x2 = np.random.randint(1, 10, size=N).astype(np.float16)
         for kernel, fn in zip(exp_kernels, expected_exp_functions):
             with self.subTest(fn=fn):
                 kernel = cuda.jit("void(f2[:], f2[:])")(kernel)
                 kernel[1, N](r, x2)
                 expected = fn(x2, dtype=np.float16)
-                np.testing.assert_allclose(r, expected)
+                np.testing.assert_allclose(r, expected, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_hexp10(self):
@@ -876,7 +878,7 @@ def hexp10_vectors(r, x):
 
         # Run the kernel
         hexp10_vectors[1, N](r, x)
-        np.testing.assert_allclose(r, 10**x)
+        np.testing.assert_allclose(r, 10**x, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_fp16_comparison(self):
@@ -948,10 +950,10 @@ def test_hmax(self):
         arg1 = np.float16(3.0)
         arg2 = np.float16(4.0)
         compiled[1, 1](ary, arg1, arg2)
-        np.testing.assert_allclose(ary[0], arg2)
+        np.testing.assert_allclose(ary[0], arg2, rtol=self.FLOAT16_RTOL)
         arg1 = np.float16(5.0)
         compiled[1, 1](ary, arg1, arg2)
-        np.testing.assert_allclose(ary[0], arg1)
+        np.testing.assert_allclose(ary[0], arg1, rtol=self.FLOAT16_RTOL)
 
     @skip_unless_cc_53
     def test_hmin(self):
@@ -960,10 +962,10 @@ def test_hmin(self):
         arg1 = np.float16(3.0)
         arg2 = np.float16(4.0)
         compiled[1, 1](ary, arg1, arg2)
-        np.testing.assert_allclose(ary[0], arg1)
+        np.testing.assert_allclose(ary[0], arg1, rtol=self.FLOAT16_RTOL)
         arg1 = np.float16(5.0)
         compiled[1, 1](ary, arg1, arg2)
-        np.testing.assert_allclose(ary[0], arg2)
+        np.testing.assert_allclose(ary[0], arg2, rtol=self.FLOAT16_RTOL)
 
     def test_cbrt_f32(self):
         compiled = cuda.jit("void(float32[:], float32)")(simple_cbrt)