ulpwise 0.3.1__tar.gz → 0.3.2__tar.gz

{ulpwise-0.3.1 → ulpwise-0.3.2}/CHANGELOG.md

@@ -1,5 +1,28 @@
 # Changelog
 
+## 0.3.2 (2026-10-06)
+
+- Two regression corpus cases for pytorch #199850: `torch.erf` in bfloat16 and float16 on CPU returns 0 at and
+  below 1.8e-7 and loses relative accuracy below 1e-3 (13404 bfloat16 ulps, 5 float16 ulps at worst), found with
+  the half precision rows of `ulpwise survey --functions erf --backends torch --dtypes f16,bf16`.
+- Two regression corpus cases for pytorch #199867: `torch.special.logit` in float16 and bfloat16 on CPU rounds `1 - x`
+  and `x / (1 - x)` to the input dtype before the log, so logit(0.499756) in float16 is -0.000488 for an exact -0.000977
+  (512 ulp, 64 bfloat16 ulp at worst), found with `ulpwise survey --functions logit --backends torch --dtypes f16,bf16`.
+- `studies/accuracy-survey-2026-10-half`: the second accuracy survey, torch 2.14.0+cpu in float16 and bfloat16 for the
+  45 functions with half precision CPU kernels. 33 float16 and 34 bfloat16 rows of 45 are correctly rounded; `erf`
+  (pytorch #199850) and `logit` (pytorch #199867) lose their digits before the rounding step, `polygamma(2, x)` in
+  float16 is off at every half integer in `(-1024, -256)` because the Hurwitz zeta sum accumulates in float for the
+  reduced types, and the activation tails and the `rsqrt` and `i0e` vector versus scalar disagreements of September
+  show again.
+- `ulpwise survey`: the `polygamma_1` and `polygamma_2` references at a non positive integer are now the signed
+  infinity of the pole, `+inf` for an odd order (the limit from both sides) and `-inf` for an even one (the sign of
+  `(-1) ** (n + 1) * n! * zeta(n + 1, x)`, which is what scipy, torch and jax return), instead of a domain error
+  that expected `nan`. The f16 and bf16 grids reach the integers from 2048 and 256 up, so every negative point
+  past there was a pole and counted as a nonfinite mismatch: on torch 2.14 the two functions had 354 of 3546
+  float16 points and 492 of 4016 bfloat16 points each, `polygamma_2` has 0 now in every dtype and `polygamma_1`
+  keeps 491 bfloat16, 4 float32 and 2 float16 points where torch returns a large finite value instead of `inf`
+  at a negative integer (pytorch #198663).
+
 ## 0.3.1 (2026-10-05)
 
 - `ulpwise survey` takes `f16` and `bf16` in `--dtypes` (the default stays `f32,f64`): torch and jax get half

{ulpwise-0.3.1 → ulpwise-0.3.2}/Cargo.lock

@@ -120,7 +120,7 @@ checksum = "adb6935a6f5c20170eeceb1a3835a49e12e19d792f6dd344ccc76a985ca5a6ca"
 
 [[package]]
 name = "ulpwise"
-version = "0.3.1"
+version = "0.3.2"
 dependencies = [
  "pyo3",
 ]

{ulpwise-0.3.1 → ulpwise-0.3.2}/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "ulpwise"
-version = "0.3.1"
+version = "0.3.2"
 edition = "2021"
 rust-version = "1.86"
 license = "MIT OR Apache-2.0"

{ulpwise-0.3.1 → ulpwise-0.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ulpwise
-Version: 0.3.1
+Version: 0.3.2
 Classifier: Development Status :: 3 - Alpha
 Classifier: Framework :: Pytest
 Classifier: Intended Audience :: Developers
@@ -226,6 +226,8 @@ PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 31 present, 0 fi
 | torchvision #9676 `clamp_bounding_boxes` collapses slightly tilted rotated boxes to a point | geometry | open |
 | pytorch #198663 `polygamma(1, x)` float64 keeps 9 digits (series stops at `1/42`), float32 loses all for large negative `x` | truncation, rounding | open |
 | pytorch #198664 `erfcx` off by `x*x/2` ulps for negative `x` (`exp` at the rounded square) | rounding | open |
+| pytorch #199850 `torch.erf` in bfloat16 and float16 on CPU returns 0 at and below 1.8e-7 and loses relative accuracy below 1e-3 (13404 bfloat16 ulps, 5 float16 ulps) | cancellation | open |
+| pytorch #199867 `torch.special.logit` in float16 and bfloat16 on CPU rounds `1 - x` and `x / (1 - x)` to the input dtype before the log, logit(0.499756) is -0.000488 for -0.000977 (512 float16 ulps, 64 bfloat16 ulps) | rounding | open |
 
 Cases marked `open` have an issue with the complete patch attached and no merged fix yet. Every case
 is an expected failure until the installed release contains the fix (`fixed_in_release` in
@@ -341,6 +343,21 @@ logged and skipped.
 - scipy's float64 `lgamma` does not handle the zeros at 1 and 2, and its Bessel functions lose the
   phase at large `x`.
 
+[`studies/accuracy-survey-2026-10-half`](https://github.com/Nicholas022400701/ulpwise/blob/main/studies/accuracy-survey-2026-10-half/README.md)
+is the second run, the 45 functions with torch float16 and bfloat16 CPU kernels measured in the ulps of
+those dtypes on the same build. 33 of the 45 float16 rows and 34 of the bfloat16 rows are correctly
+rounded over the whole grid; the exceptions:
+
+- `erf` goes through a formula with 1.5e-7 absolute error, 254 bfloat16 ulps and `-0` below
+  `|x| = 1.8e-7` (pytorch #199850).
+- `logit` rounds `1 - x` and `x / (1 - x)` to the input dtype before the log, 512 float16 ulps at
+  `x = 0.499756` while the CUDA kernel computes in float (pytorch #199867).
+- `polygamma(2, x)` in float16 is off at every half integer in `(-1024, -256)`, 1,540 ulps at
+  `x = -1023.5`, because the Hurwitz zeta sum accumulates in float for the reduced types and in
+  double for float32.
+- the `gelu`, `gelu_tanh`, `silu` and `sigmoid` tails and the `rsqrt` and `i0e` vector versus scalar
+  disagreements are the float32 findings of September seen through coarser ulps.
+
 ## How the exact oracle works
 
 For `sqrt(x)` the candidate `r` and the midpoint `m` between `r` and its neighbour are written as

{ulpwise-0.3.1 → ulpwise-0.3.2}/README.md

@@ -197,6 +197,8 @@ PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 31 present, 0 fi
 | torchvision #9676 `clamp_bounding_boxes` collapses slightly tilted rotated boxes to a point | geometry | open |
 | pytorch #198663 `polygamma(1, x)` float64 keeps 9 digits (series stops at `1/42`), float32 loses all for large negative `x` | truncation, rounding | open |
 | pytorch #198664 `erfcx` off by `x*x/2` ulps for negative `x` (`exp` at the rounded square) | rounding | open |
+| pytorch #199850 `torch.erf` in bfloat16 and float16 on CPU returns 0 at and below 1.8e-7 and loses relative accuracy below 1e-3 (13404 bfloat16 ulps, 5 float16 ulps) | cancellation | open |
+| pytorch #199867 `torch.special.logit` in float16 and bfloat16 on CPU rounds `1 - x` and `x / (1 - x)` to the input dtype before the log, logit(0.499756) is -0.000488 for -0.000977 (512 float16 ulps, 64 bfloat16 ulps) | rounding | open |
 
 Cases marked `open` have an issue with the complete patch attached and no merged fix yet. Every case
 is an expected failure until the installed release contains the fix (`fixed_in_release` in
@@ -312,6 +314,21 @@ logged and skipped.
 - scipy's float64 `lgamma` does not handle the zeros at 1 and 2, and its Bessel functions lose the
   phase at large `x`.
 
+[`studies/accuracy-survey-2026-10-half`](https://github.com/Nicholas022400701/ulpwise/blob/main/studies/accuracy-survey-2026-10-half/README.md)
+is the second run, the 45 functions with torch float16 and bfloat16 CPU kernels measured in the ulps of
+those dtypes on the same build. 33 of the 45 float16 rows and 34 of the bfloat16 rows are correctly
+rounded over the whole grid; the exceptions:
+
+- `erf` goes through a formula with 1.5e-7 absolute error, 254 bfloat16 ulps and `-0` below
+  `|x| = 1.8e-7` (pytorch #199850).
+- `logit` rounds `1 - x` and `x / (1 - x)` to the input dtype before the log, 512 float16 ulps at
+  `x = 0.499756` while the CUDA kernel computes in float (pytorch #199867).
+- `polygamma(2, x)` in float16 is off at every half integer in `(-1024, -256)`, 1,540 ulps at
+  `x = -1023.5`, because the Hurwitz zeta sum accumulates in float for the reduced types and in
+  double for float32.
+- the `gelu`, `gelu_tanh`, `silu` and `sigmoid` tails and the `rsqrt` and `i0e` vector versus scalar
+  disagreements are the float32 findings of September seen through coarser ulps.
+
 ## How the exact oracle works
 
 For `sqrt(x)` the candidate `r` and the midpoint `m` between `r` and its neighbour are written as

{ulpwise-0.3.1 → ulpwise-0.3.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "ulpwise"
-version = "0.3.1"
+version = "0.3.2"
 description = "Numerical conformance testing for ML code: float edge values, exact rounding oracles, knife-edge finders and a pytest plugin"
 readme = "README.md"
 license = { text = "MIT OR Apache-2.0" }

{ulpwise-0.3.1 → ulpwise-0.3.2}/python/ulpwise/__init__.py

@@ -30,7 +30,7 @@ from ._core import special as _special_core
 from ._core import ulp_distance as _ulp_distance_core
 from ._core import ulp_distances as _ulp_distances_core
 
-__version__ = "0.3.1"
+__version__ = "0.3.2"
 
 __all__ = [
     "UNARY_OPS",

{ulpwise-0.3.1 → ulpwise-0.3.2}/python/ulpwise/corpus/cases.json

@@ -1019,6 +1019,90 @@
     "rtol": 0.0
    },
    "before_fix": "[1] on ultralytics 8.4.163: nc = 1 and no labels, \"ignoring corrupt image/label: non-normalized or out of bounds coordinates [1.15 1.15 ...]\""
+  },
+  {
+   "id": "pytorch-199850-erf-bfloat16-small-arguments",
+   "repo": "pytorch/pytorch",
+   "pr": null,
+   "issue": 199850,
+   "merged_at": null,
+   "kind": "cancellation",
+   "title": "torch.erf in bfloat16 on CPU returns 0 for |x| <= 1.8e-7 and loses relative accuracy below 1e-3",
+   "symptom": "bfloat16 and float16 erf go through Vectorized<float>::erf(), the Abramowitz and Stegun 7.1.26 formula 1 - r * t * exp(-x * x), whose absolute error is 1.5e-7; below |x| = 1e-3 that error is a visible fraction of erf(x) ~ 1.128 x and below 1.5e-7 the 1 - 1 rounds to exactly 0. Of the 16128 bfloat16 inputs in (0, 0.5], 14102 are more than 1 ulp off, 13379 return 0, the worst is 13404 ulp. The MKL float32 and float64 paths are correct; float32 without MKL and under torch.compile takes the same code (issue #198183). The patch adds a Taylor series branch for |x| < 0.5 (Draft PR #199851).",
+   "requires": [
+    "torch"
+   ],
+   "fixed_in_release": null,
+   "repro": "import torch\nxs = torch.tensor([1e-7, 5e-7, 1e-6, 1e-5], dtype=torch.bfloat16)\nactual = torch.erf(xs).tolist()\n# erf at the bfloat16 inputs in float64, rounded to bfloat16\nexpected = [1.1315569281578064e-07, 5.62518835067749e-07, 1.125037670135498e-06, 1.1324882507324219e-05]\n",
+   "check": {
+    "type": "max_ulp",
+    "dtype": "bf16",
+    "max_ulp": 1
+   },
+   "before_fix": "[0.0, 4.172325134277344e-07, 1.1920928955078125e-06, 1.1563301086425781e-05] on torch 2.14.0 (13299, 55, 9 and 4 ulp)"
+  },
+  {
+   "id": "pytorch-199850-erf-float16-small-arguments",
+   "repo": "pytorch/pytorch",
+   "pr": null,
+   "issue": 199850,
+   "merged_at": null,
+   "kind": "cancellation",
+   "title": "torch.erf in float16 on CPU returns 0 for |x| <= 1.8e-7 and is up to 5 ulp off below 1e-3",
+   "symptom": "Same Vectorized<float>::erf() path as the bfloat16 case; float16 has three more bits, so the 1.5e-7 absolute error shows as 1240 of the 14336 inputs in (0, 0.5] more than 1 ulp off, at most 5 ulp, and the three inputs at or below 1.8e-7 return 0.",
+   "requires": [
+    "torch"
+   ],
+   "fixed_in_release": null,
+   "repro": "import torch\nxs = torch.tensor([1e-7, 5e-7, 1e-6, 1e-5], dtype=torch.float16)\nactual = torch.erf(xs).tolist()\n# erf at the float16 inputs in float64, rounded to float16\nexpected = [1.1920928955078125e-07, 5.364418029785156e-07, 1.1324882507324219e-06, 1.1324882507324219e-05]\n",
+   "check": {
+    "type": "max_ulp",
+    "dtype": "f16",
+    "max_ulp": 1
+   },
+   "before_fix": "[0.0, 4.172325134277344e-07, 1.1920928955078125e-06, 1.1563301086425781e-05] on torch 2.14.0 (2, 2, 1 and 4 ulp)"
+  },
+  {
+   "id": "pytorch-199867-logit-bfloat16-intermediate-rounding",
+   "repo": "pytorch/pytorch",
+   "pr": null,
+   "issue": 199867,
+   "merged_at": null,
+   "kind": "rounding",
+   "title": "torch.special.logit in bfloat16 on CPU rounds 1 - x and x / (1 - x) to bfloat16 before the log, 64 ulp at x = 0.498047",
+   "symptom": "The CPU kernel computes 1 - x and x / (1 - x) in the input dtype and only the log in float, so near x = 0.5, where logit is small, the two roundings take most of the digits: logit(0.499756) in float16 is -0.000488 for an exact -0.000977 (512 ulp), 582 of the 15358 float16 values in (0, 1) are more than 1 ulp off and 59 of the 16254 bfloat16 values, 64 ulp at worst. The CUDA kernel computes in float. The patch computes in float and rounds once, as sigmoid_kernel does for the reduced types.",
+   "requires": [
+    "torch"
+   ],
+   "fixed_in_release": null,
+   "repro": "import torch\nxs = torch.tensor([0.498046875, 0.490234375, 0.470703125, 0.51953125], dtype=torch.bfloat16)\nactual = torch.special.logit(xs).tolist()\n# logit at the bfloat16 inputs in float64, rounded to bfloat16\nexpected = [-0.0078125, -0.0390625, -0.1171875, 0.078125]\n",
+   "check": {
+    "type": "max_ulp",
+    "dtype": "bf16",
+    "max_ulp": 1
+   },
+   "before_fix": "[-0.00390625, -0.035888671875, -0.1201171875, 0.0751953125] on torch 2.14.0 (64, 13, 5.7 and 6.1 ulp)"
+  },
+  {
+   "id": "pytorch-199867-logit-float16-intermediate-rounding",
+   "repo": "pytorch/pytorch",
+   "pr": null,
+   "issue": 199867,
+   "merged_at": null,
+   "kind": "rounding",
+   "title": "torch.special.logit in float16 on CPU rounds 1 - x and x / (1 - x) to float16 before the log, 512 ulp at x = 0.499756",
+   "symptom": "The CPU kernel computes 1 - x and x / (1 - x) in the input dtype and only the log in float, so near x = 0.5, where logit is small, the two roundings take most of the digits: logit(0.499756) in float16 is -0.000488 for an exact -0.000977 (512 ulp), 582 of the 15358 float16 values in (0, 1) are more than 1 ulp off and 59 of the 16254 bfloat16 values, 64 ulp at worst. The CUDA kernel computes in float. The patch computes in float and rounds once, as sigmoid_kernel does for the reduced types.",
+   "requires": [
+    "torch"
+   ],
+   "fixed_in_release": null,
+   "repro": "import torch\nxs = torch.tensor([0.499755859375, 0.497314453125, 0.489990234375, 0.52001953125], dtype=torch.float16)\nactual = torch.special.logit(xs).tolist()\n# logit at the float16 inputs in float64, rounded to float16\nexpected = [-0.0009765625, -0.0107421875, -0.0400390625, 0.08013916015625]\n",
+   "check": {
+    "type": "max_ulp",
+    "dtype": "f16",
+    "max_ulp": 1
+   },
+   "before_fix": "[-0.00048828125, -0.01129150390625, -0.039337158203125, 0.0797119140625] on torch 2.14.0 (512, 72, 23 and 6.7 ulp)"
   }
  ]
-}
+}

{ulpwise-0.3.1 → ulpwise-0.3.2}/python/ulpwise/survey.py

@@ -230,7 +230,9 @@ def _ref_polygamma(n):
     def f(x):
         mp = _mp()
         if x <= 0 and x == mp.floor(x):
-            return None
+            # the pole: +inf from both sides for an odd n; for an even n the two sides disagree and the reference
+            # is the sign of (-1) ** (n + 1) * n! * zeta(n + 1, x), which is what scipy, torch and jax return
+            return mp.mpf("inf") if n % 2 else mp.mpf("-inf")
         return mp.polygamma(n, x)
 
     return f

ulpwise-0.3.2/studies/accuracy-survey-2026-10-half/README.md

@@ -0,0 +1,133 @@
+# Accuracy survey, October 2026: torch in float16 and bfloat16
+
+The 45 functions of the September survey for which torch has a float16 and a bfloat16 CPU kernel,
+measured in the ulps of those dtypes, on the same machine and the same torch build. The question
+this time: a half precision kernel in torch is supposed to compute in float and round once, so the
+rows should all read 0.5. Which ones do not, and why.
+
+| | |
+|---|---|
+| command | `ulpwise survey --backends torch --dtypes f16,bf16 --out studies/accuracy-survey-2026-10-half` on ulpwise `main` at 0b12fb8 (0.3.1 plus the `polygamma` pole references below) |
+| backends | torch 2.14.0+cpu (CPU capability AVX512, MKL available) |
+| reference | mpmath at 200 bits, evaluated at the float16 or bfloat16 input the kernel actually saw |
+| inputs | 600 log spaced points over each function's domain, clipped to the finite range of the dtype, plus the named edge values of the dtype, 480 to 662 inputs per row, subnormals included |
+| platform | Linux x86_64, Python 3.13, mpmath 1.3.0 |
+| files | `results.csv` (one row per function and dtype with every metric and the worst input), `results.md` (the pivot table) |
+
+The 16 functions without half precision kernels in torch (`bessel_*`, `modified_bessel_*`,
+`scaled_modified_bessel_*`, `spherical_bessel_j0`, `airy_ai`, `erfcx`, `log_ndtr`, `ndtri`, `zeta`)
+are logged and skipped, so there are 90 rows.
+
+## How to read a number
+
+The same metric as in September: `|got - exact| / spacing(dtype, exact)`, correct rounding is 0.5,
+and an exact result of zero or in the subnormal range produces enormous numbers that are not
+precision bugs. Three things are specific to the half precision grids:
+
+1. **bfloat16 has 8 significant bits and float16 11.** Every bfloat16 value with `|x| >= 256` is
+   an integer, every float16 value with `|x| >= 2048` is one. A log spaced grid over a domain that
+   reaches 1e5 therefore spends a quarter of its points on integers, which is why `sinc` in
+   bfloat16 reads `8.4e32` (the exact value at an integer is 0) and why `polygamma_1` in bfloat16
+   has 75 non finite mismatches (every negative point past 256 is a pole).
+2. **The float16 range is narrow.** Normal numbers stop at 6.1e-5, the smallest subnormal is 6e-8
+   and the largest value is 65504. A tail value such as `gelu(-5.57) = -7.1e-8` is a subnormal with
+   one significant bit; its ulp is 6e-8 and torch's `-0` there counts as 1.19 ulps, not as the
+   millions the same formula gives in float32.
+3. **The two tolerance columns** use `torch.testing`'s defaults for the dtype, `rtol = 1e-3,
+   atol = 1e-5` for float16 and `rtol = 1.6e-2, atol = 1e-5` for bfloat16, and the per op override
+   from `op_db`. One ulp of float16 is between 4.9e-4 and 9.8e-4 relative, so the default float16
+   tolerance is about one ulp and the bfloat16 one about four.
+
+The `polygamma` reference changed since September: at a non positive integer it is now the signed
+infinity of the pole (`+inf` for `polygamma_1`, `-inf` for `polygamma_2`, the sign torch, scipy and
+jax return) instead of a domain error that expected `nan`, so a kernel that returns the infinity at
+a pole is no longer counted as a mismatch. The September `polygamma_1` float32 row counts 7 non
+finite mismatches with the old reference and would count 4 with this one (the changelog entry on
+the `polygamma` references has the per dtype numbers); the measured errors do not depend on the
+reference at the poles, and the September files are left as they were measured.
+
+## Findings
+
+### Two kernels lose their digits before the rounding step
+
+In both cases the half precision path is not "compute in float and round once", and the errors
+are far outside what the dtype can represent.
+
+- **`erf` in bfloat16 and float16** (pytorch #199850, patch and Draft PR #199851). 254 ulps in
+  bfloat16 at `x = -1.3e-8` (torch returns `-0`), p99 241, 34.6 percent of the grid beyond 1 ulp;
+  4.55 ulps in float16, 24.5 percent beyond 1 ulp. The reduced types go through
+  `Vectorized<float>::erf()`, the Abramowitz and Stegun 7.1.26 formula `1 - r * t * exp(-x * x)`,
+  whose absolute error of 1.5e-7 is invisible for a float32 result of order 1 but is the whole
+  result when `erf(x) ~ 1.128 x` is small: below `|x| = 1.8e-7` the `1 - 1` rounds to exactly 0.
+  Over every bfloat16 value in `(0, 0.5]`, 14,102 of 16,128 inputs are more than 1 ulp off and
+  13,379 return 0; float32 and float64 are correct on this build because they take the MKL path.
+- **`logit` in float16 and bfloat16** (pytorch #199867, patch on a branch). 72 ulps in float16 at
+  `x = 0.49731`, 29 of the 575 inputs fail the default float16 tolerance; 64 ulps in bfloat16, 10
+  of 480. The CPU kernel computes `1 - x` and `x / (1 - x)` in the input dtype and only the `log`
+  in float, so near `x = 0.5`, where the result is small, the two roundings take most of its
+  digits: `logit(0.499756)` in float16 is `-0.000488` for an exact `-0.000977`, 512 ulps, and 582 of
+  the 15,358 float16 values in `(0, 1)` are beyond 1 ulp (59 of 16,254 in bfloat16). Computing in
+  float and rounding once is correctly rounded at every one of those inputs, and that is what the
+  CUDA kernel already does. The `OpInfo` reference test carries
+  `precisionOverride({torch.bfloat16: 5e-1, torch.float16: 5e-1})`, an absolute tolerance of 0.5
+  on values below 0.01, so the `op fail` column is 0 for both rows.
+
+### `polygamma(2, x)` in float16 at the negative half integers
+
+14 ulps on the grid at `x = -578.5` (got `-2.146e-6`, exact `-2.983e-6`, a 28 percent error). Off
+the grid it is worse: at every one of the 768 float16 half integers in `(-1024, -256)` the result is
+more than 1 ulp off, 1,540 ulps at `x = -1023.5` (got `-2.38e-7`, exact `-9.54e-7`), and so are 203
+of the 255 half integers in `(-256, -1)`. float32 is within 1e-4 float16 ulps at the same inputs.
+These are the points where the function is small: at a half integer the reflection term of
+`polygamma(2, x)` vanishes and the value is `polygamma(2, 1 - x) ~ -1 / x^2`, while the Hurwitz
+zeta sum behind it adds terms of order 1 that cancel. torch's `calc_polygamma<Half>` accumulates
+that sum in `float` (`acc_type<Half, false>`) and `calc_polygamma<float>` in `double`, which is why
+float32 keeps its digits and float16 keeps an absolute error of about 1e-6. bfloat16 does not show
+it only because bfloat16 has no half integers past 256 and the values below 256 are of order 1. Not
+reported upstream yet (the fix would compute the reduced types through the float path, as
+`calc_digamma` does). On CUDA `acc_type<float, true>` is `float`, so by the same reading of the
+source the float32 kernel there has the same problem; not measured.
+
+### The formula tails, as in September
+
+- **`gelu` and `gelu_tanh` in bfloat16** return `-0` from about `x = -5.5` down, where
+  `1 + erf(x / sqrt 2)` and `1 + tanh(t)` are exactly 0 in float: 242 and 250 ulps at the worst grid
+  point, 2.1 and 1.6 percent of the grid beyond 10 ulps. bfloat16 has the float32 exponent range, so
+  `gelu(-7.875) = -1.3e-14` is an ordinary normal number there. In float16 the same inputs are
+  below the subnormal range and the row reads 1.19 ulps (see above).
+- **`silu(-92.5)` and `sigmoid(-92.5)` in bfloat16** return `-0` and 0 because `exp(92.5)` overflows
+  float32: 67.7 and 0.73 ulps, exactly the float32 behaviour of September measured in coarser ulps.
+
+### The vectorized kernel and the scalar tail disagree
+
+`rsqrt` (128 of 571 float16 inputs, 184 of 615 bfloat16) and `i0e` (303 of 588, 413 of 620) return
+different bits for the same input depending on whether it landed in the vector body or in the
+scalar tail of the tensor; `gelu` in float16 differs at 3 inputs. Both paths are within 1.5 ulps
+(`rsqrt` 0.95 and 1.08, `i0e` 1.46 and 0.65), so this is determinism and not precision: the result
+of an element depends on the tensor size and alignment.
+
+### Small things
+
+- `entr` reads 1.19 ulps in float16 (`x = 3088`) and 1.2 in bfloat16: `-x * log(x)` with the log
+  rounded to the dtype before the product.
+- `polygamma_1` in bfloat16 returns a large finite number instead of `inf` at the 75 negative
+  integers of its grid (`4.3e9` at `-300`), the float32 reflection `pi^2 / sin(pi x)^2` with a
+  rounded argument. float16 returns `inf` at the same points because `4.3e9` overflows float16.
+- The single non finite mismatch of `reciprocal` and of `digamma` in each dtype is the edge value
+  `-0.0`, where torch returns an infinity and the reference has no value.
+
+### Everything else rounds correctly
+
+33 of the 45 float16 rows and 34 of the 45 bfloat16 rows have a maximum of 0.5 ulps over the whole
+grid, 39 and 36 stay at or below 1 ulp. `exp`, `log`, `sqrt`, the trigonometric and hyperbolic
+functions and their inverses, `erfc`, `erfinv`, `lgamma`, `digamma`, `i0`, `i1`, `i1e`, `elu`,
+`selu`, `mish`, `softplus` and `logsigmoid` are all correctly rounded in both dtypes: they convert
+to float, compute, and convert back once.
+
+## What the survey does not measure
+
+One machine, CPU only, one build. 600 points per domain, so the all values scans quoted above for
+`erf`, `logit` and `polygamma_2` were separate runs over every value of the dtype in the stated
+interval, not part of `results.csv`. CUDA kernels are not measured. The half precision grids are
+coarse at the top of their range (see the integer note above), which is also why the bfloat16
+`polygamma_2` row cannot show the half integer problem.

ulpwise-0.3.2/studies/accuracy-survey-2026-10-half/results.csv

@@ -0,0 +1,91 @@
+function,backend,dtype,n,max_ulp,p99_ulp,median_ulp,frac_gt1,frac_gt10,max_abs_err,nonfinite_mismatch,worst_x,worst_got,worst_ref,vec_scalar_mismatch,default_tol_fail,op_tol_fail,op_rtol,op_atol,seconds
+exp,torch,f16,577,0.49652399858703933,0.4877496877181343,0.0708642683168237,0.0,0.0,5.618725877364341,0,-0.0012197494506835938,0.9990234375,0.9987809941413149,0,0,0,0.001,1e-05,0.02
+exp,torch,bf16,620,0.4974970990796693,0.47520794301700675,4.470348358154297e-08,0.0,0.0,1.5855092962538616e+33,0,-0.021728515625,0.9765625,0.97850584804328,0,0,0,0.016,1e-05,0.03
+exp2,torch,f16,577,0.4962890576058865,0.49205187864370914,0.06404980535758294,0.0,0.0,5.911688801792479,0,0.0006990432739257812,1.0,1.0004846572828183,0,0,0,0.001,1e-05,0.02
+exp2,torch,bf16,617,0.4889077724274732,0.48072804765681126,8.876224730507275e-09,0.0,0.0,1.003981677138867e+32,0,-0.01422119140625,0.98828125,0.9901910459860448,0,0,0,0.016,1e-05,0.03
+expm1,torch,f16,577,0.49730118760972325,0.49034771682641204,0.04686579643374022,0.0,0.0,4.618725877364341,0,0.039947509765625,0.040740966796875,0.040756143224719536,0,0,0,0.001,1e-05,0.03
+expm1,torch,bf16,620,0.4853231448379511,0.4672336336794082,1.4028955774656958e-08,0.0,0.0,1.5855092962538616e+33,0,-0.431640625,-0.349609375,-0.35055727176726165,0,0,0,0.016,1e-05,0.04
+log,torch,f16,571,0.4999053175932843,0.49424945169839446,0.25053770135421477,0.0,0.0,0.005092666561314019,0,3.12109375,1.1376953125,1.138183501286712,0,0,0,0.001,1e-05,0.03
+log,torch,bf16,615,0.49900858890093847,0.49542647977390203,0.25221120721920237,0.0,0.0,0.24950429445047462,0,2.1122164116337979e-38,-87.0,-86.75049570554953,0,0,0,0.016,0.05,0.03
+log2,torch,f16,571,0.49982244041977186,0.4957887013732117,0.24917759192167044,0.0,0.0,0.007716188891969011,0,27184.0,14.734375,14.73047013718422,0,0,0,0.001,1e-05,0.03
+log2,torch,bf16,615,0.49984227461230096,0.49631140505807564,0.24992113730615048,0.0,0.0,0.4594316186372964,0,4.4994390158148434e-18,-57.5,-57.62496056865307,0,0,0,0.016,0.1,0.03
+log10,torch,f16,571,0.49963245913465204,0.4903260309973072,0.2422109430332955,0.0,0.0,0.0019428033816639712,0,0.0159454345703125,-1.7978515625,-1.7973636401766264,0,0,0,0.001,1e-05,0.03
+log10,torch,bf16,615,0.4998269941082355,0.4948880258445741,0.24490373120930514,0.0,0.0,0.12395643537827539,0,3.651923115891872e+28,28.625,28.56252162573647,0,0,0,0.016,0.05,0.03
+log1p,torch,f16,571,0.4995147834324868,0.49338139163683464,0.17714651543802104,0.0,0.0,0.0038929874031676803,0,0.07012939453125,0.06781005859375,0.06777957063089401,0,0,0,0.001,1e-05,0.03
+log1p,torch,bf16,615,0.4983324042823015,0.48845749265665506,0.04134085372358378,0.0,0.0,0.24828420141527374,0,3.37944351430359e+22,51.75,51.87458310107058,0,0,0,0.016,0.1,0.03
+sqrt,torch,f16,571,0.4999389499380398,0.49007737105917587,0.24356979992962996,0.0,0.0,0.06249236874225517,0,0.99951171875,0.99951171875,0.9997558295653994,0,0,0,0.001,1e-05,0.02
+sqrt,torch,bf16,615,0.499510762740992,0.49902723367208435,0.2612480950072464,0.0,0.0,3.599354375916544e+16,0,1.1990042378387333e-35,3.4558944247975454e-18,3.462664057974341e-18,0,0,0,0.016,0.07,0.03
+rsqrt,torch,f16,571,0.948287752861986,0.4995526197759731,0.2426170559979162,0.0,0.0,0.8267026007069944,0,29824.0,0.005786895751953125,0.005790513182650993,128,0,0,0.001,0.05,0.03
+rsqrt,torch,bf16,615,1.0782885165447518,0.4993185592317671,0.2564640308922517,0.0016260162601626016,0.0,5.716504393996698e+16,0,1.010213276796536e+38,9.994988777600744e-20,9.949321482592776e-20,184,0,0,0.016,1e-05,0.03
+reciprocal,torch,f16,588,0.49975574010747437,0.49818181818181817,0.21811460258780038,0.0,0.0,11.026506024099945,1,0.007808685302734375,128.125,128.06253053248656,0,0,0,0.001,1e-05,0.02
+reciprocal,torch,bf16,619,0.4980392156862745,0.49751396316102214,0.23943661971830985,0.0,0.0,5.596749455936647e+35,1,-33423360.0,-3.003515303134918e-08,-2.991919424019608e-08,0,0,0,0.016,1e-05,0.02
+sin,torch,f16,588,0.4996588404833843,0.4945322035448394,0.08753395421070723,0.0,0.0,0.0002419961642482482,0,-0.418212890625,-0.40625,-0.4061280129783976,0,0,0,0.001,1e-05,0.03
+sin,torch,bf16,620,0.49902617601409,0.48721123707134933,0.024317169873750665,0.0,0.0,0.001949321000055071,0,-4544.0,-0.953125,-0.9511756789999449,0,0,0,0.016,0.01,0.03
+cos,torch,f16,588,0.49845340422660517,0.4956050200685431,0.09112808529822407,0.0,0.0,0.00024338545128255173,0,-37568.0,0.6708984375,0.6711418229512826,0,0,0,0.001,1e-05,0.02
+cos,torch,bf16,620,0.4982183001878292,0.47264074304242,0.052667672036868424,0.0,0.0,0.00194616523510871,0,-15.0625,-0.796875,-0.7988211652351087,0,0,0,0.016,0.01,0.03
+tan,torch,f16,588,0.49881488045603833,0.4914911839161815,0.08113939790216393,0.0,0.0,0.017764182039798015,0,-0.028411865234375,-0.0284271240234375,-0.028419512712395387,0,0,0,0.001,1e-05,0.03
+tan,torch,bf16,620,0.49997112617668704,0.4875677336521684,0.04762034554969279,0.0,0.0,0.11281124397999065,0,-248.0,0.1884765625,0.18798830944709308,0,0,0,0.016,1e-05,0.03
+asin,torch,f16,648,0.4941364115688654,0.4890526633905064,0.009329841781712579,0.0,0.0,0.0004825550894227426,0,-0.99951171875,-1.5390625,-1.5395450550894227,0,0,0,0.001,1e-05,0.04
+asin,torch,bf16,662,0.4903648339847289,0.47865063368268224,1.127157265195468e-06,0.0,0.0,0.0037394580756460183,0,-0.1435546875,-0.14453125,-0.1440523780918118,0,0,0,0.016,0.01,0.03
+acos,torch,f16,648,0.4998616549945954,0.49918054207310525,0.3402882157076168,0.0,0.0,0.0009668216433849075,0,-4.589557647705078e-06,1.5712890625,1.5708009163525443,0,0,0,0.001,0.01,0.03
+acos,torch,bf16,662,0.49946004853325654,0.48679677796885956,0.06253255191473825,0.0,0.0,0.007723588354033417,0,0.703125,0.7890625,0.791013515814583,0,0,0,0.016,0.1,0.03
+atan,torch,f16,588,0.4992192365768905,0.49369285920882655,0.10639318712486037,0.0,0.0,0.0004838466629295546,0,-0.028411865234375,-0.0283966064453125,-0.02840422392633937,0,0,0,0.001,1e-05,0.02
+atan,torch,bf16,620,0.4965529153521348,0.47865676775623917,0.030024177423035394,0.0,0.0,0.0038793196511885952,0,-8.75,-1.453125,-1.4570043196511886,0,0,0,0.016,0.01,0.03
+sinh,torch,f16,577,0.49732908474709986,0.4746697418863143,1.5251472540330284e-05,0.0,0.0,15.33758697706071,0,-3.3203125,-13.8203125,-13.816427116525412,0,0,0,0.001,0.01,0.02
+sinh,torch,bf16,620,0.4788486391066387,0.43272709337056176,3.102249621785879e-18,0.0,0.0,7.927546481269308e+32,0,-35.25,-1020346790576128.0,-1018240791989558.1,0,0,0,0.016,1e-05,0.03
+cosh,torch,f16,577,0.49670313581569203,0.4866177109694043,3.419863079152353e-05,0.0,0.0,15.337576687445107,0,-0.09375,1.0048828125,1.00439775084393,0,0,0,0.001,1e-05,0.02
+cosh,torch,bf16,618,0.49961859403523684,0.4890192345539579,6.899470525787083e-06,0.0,0.0,2.593927829559106e+34,0,-4.65625,52.75,52.62509535149119,0,0,0,0.016,1e-05,0.03
+tanh,torch,f16,565,0.4968461339304618,0.4728154092331018,0.00037039433914303285,0.0,0.0,0.0002260244619315399,0,-0.05157470703125,-0.051544189453125,-0.051529026912416676,0,0,0,0.001,1e-05,0.02
+tanh,torch,bf16,619,0.49394065721331387,0.4158991837313057,1.1847957170688265e-17,0.0,0.0,0.0019294556922394568,0,-1.8046875,-0.94921875,-0.9472892943077605,0,0,0,0.016,0.01,0.03
+asinh,torch,f16,588,0.49726901812367386,0.485833174223409,0.06690126703671712,0.0,0.0,0.003734970759392553,0,-0.045166015625,-0.045135498046875,-0.04515067349298473,0,0,0,0.001,1e-05,0.03
+asinh,torch,bf16,622,0.49695720494730633,0.49049542780111755,0.13255825285142203,0.0,0.0,0.2484786024736536,0,-1.5966312839994595e+35,-82.0,-81.75152139752635,0,0,0,0.016,0.05,0.03
+acosh,torch,f16,494,0.4988606428465956,0.4929749887821485,0.24133997218624037,0.0,0.0,0.003897348772239795,0,9024.0,9.8046875,9.80079015122776,0,0,0,0.001,1e-05,0.02
+acosh,torch,bf16,595,0.4987632823816887,0.4923519073210921,0.2587197539399728,0.0,0.0,0.24685276677669776,0,50855936.0,18.5,18.437654589702287,0,0,0,0.016,0.05,0.03
+atanh,torch,f16,648,0.49854056428787036,0.4888009692464037,0.013543825964255069,0.0,0.0,0.0013952518564153138,0,-0.98583984375,-2.47265625,-2.4716825379603753,0,0,0,0.001,1e-05,0.03
+atanh,torch,bf16,662,0.49854328708069856,0.4774971482177477,2.2543145421625954e-06,0.0,0.0,0.0068152048981477975,0,-0.431640625,-0.4609375,-0.4619112173575795,0,0,0,0.016,0.01,0.03
+erf,torch,f16,556,4.554008308720993,3.9606023515251345,0.36747878381461374,0.2446043165467626,0.0,0.0002343458794934472,0,-3.4570693969726562e-06,-4.172325134277344e-06,-3.900885086731851e-06,0,0,0,0.001,0.01,0.03
+erf,torch,bf16,619,253.8853125964903,241.47314175843889,0.4011357235244519,0.345718901453958,0.2455573505654281,0.001888891120270908,0,-1.3096723705530167e-08,-0.0,-1.4778070186526183e-08,0,0,0,0.016,0.01,0.03
+erfc,torch,f16,563,0.49941349865129464,0.49188257526419277,0.0921491285628032,0.0,0.0,0.0004877084947767507,0,-0.33154296875,1.3603515625,1.3608392709947768,0,0,0,0.001,0.01,0.03
+erfc,torch,bf16,619,0.4990151959322128,0.48589553314533057,0.010950456983912267,0.0,0.0,0.003898556218220328,0,-0.09375,1.109375,1.1054764437817797,0,0,0,0.016,0.01,0.04
+erfinv,torch,f16,648,0.499711282699754,0.4963725486861071,0.24345604528932485,0.0,0.0,0.0009112087227549637,0,-0.984375,-1.708984375,-1.7094723742995115,0,0,0,0.001,0.01,0.27
+erfinv,torch,bf16,662,0.49853190518721646,0.49689033519167136,0.2345729315275233,0.0,0.0,0.006422985394390679,0,-0.0419921875,-0.037353515625,-0.037231803734085155,0,0,0,0.016,0.01,0.23
+lgamma,torch,f16,588,0.49855249039133065,0.49142581178497385,0.19514052288979372,0.0,0.0,15.447817746615328,0,-0.55224609375,1.2763671875,1.2758803198336022,0,0,0,0.001,0.7,0.04
+lgamma,torch,bf16,616,0.4991600213797292,0.4921353382997607,0.2136705084395268,0.0,0.0,1799.9055113499053,0,-0.00049591064453125,7.625,7.609401249331883,0,0,0,0.016,1e-05,0.04
+digamma,torch,f16,588,0.49909345681262235,0.49319664738344154,0.14984071040633115,0.0,0.0,11.603681000764482,1,0.0003020763397216797,-3310.0,-3310.9981869136254,0,0,0,0.001,0.5,0.06
+digamma,torch,bf16,620,0.4980392156862745,0.4926691600549886,0.1755873052660539,0.0,0.0,5.233181085766195e+33,1,1.0799670077492329e-19,-9.295429630892704e+18,-9.259542123273814e+18,0,0,0,0.016,1e-05,0.06
+sigmoid,torch,f16,569,0.4995072238814441,0.49611778689062863,0.07485496826868755,0.0,0.0,0.00024390001166085806,0,7.21875,0.99951171875,0.9992678187383391,0,0,0,0.001,0.01,0.03
+sigmoid,torch,bf16,619,0.7324025492357978,0.4965613536741928,0.015380859300971217,0.0,0.0,0.0019531002499254546,0,-92.5,0.0,6.726055149644708e-41,0,0,0,0.016,0.01,0.03
+i0,torch,f16,567,0.4980641717544879,0.4810795082375317,0.00021057227777656481,0.0,0.0,7.726867658057017,0,-0.33251953125,1.0283203125,1.027833921707271,0,0,0,0.001,1e-05,0.06
+i0,torch,bf16,619,0.4966370950157683,0.4818175297966319,2.772361190232269e-05,0.0,0.0,7.244326886666297e+29,0,-0.70703125,1.1328125,1.1289325226951894,0,0,0,0.016,1e-05,0.06
+i0e,torch,f16,588,1.4580990190361116,0.7490650014782869,0.17596994104951222,0.003401360544217687,0.0,0.0002436430109009624,0,54400.0,0.0017118453979492188,0.0017104548463639868,303,0,0,0.001,0.3,0.11
+i0e,torch,bf16,620,0.6481236314694578,0.4971856378121076,0.1216757309092941,0.0,0.0,0.0019515878602806014,0,724992.0,0.0004673004150390625,0.0004685366127614392,413,0,0,0.016,0.3,0.16
+i1,torch,f16,567,0.5000001700996843,0.4999999999999796,0.04735283585266004,0.0,0.0,5.632916459231637,0,-5.453824996948242e-05,-2.7239322662353516e-05,-2.726912499487994e-05,0,0,0,0.001,1e-05,0.06
+i1,torch,bf16,619,0.5,0.4887155677084712,2.099691547751623e-05,0.0,0.0,4.653187276137042e+30,0,-9.183549615799121e-41,-0.0,-4.591774807899561e-41,0,0,0,0.016,1e-05,0.07
+i1e,torch,f16,588,0.4999999701976787,0.4999986326697982,0.23838889927140242,0.0,0.0,6.1009957499358336e-05,0,-5.960464477539063e-08,-0.0,-2.980232061133854e-08,0,0,0,0.001,1e-05,0.1
+i1e,torch,bf16,620,0.5,0.482013289876223,0.11544587685223048,0.0,0.0,0.00047008773643925084,0,-9.183549615799121e-41,-0.0,-4.591774807899561e-41,0,0,0,0.016,1e-05,0.11
+logit,torch,f16,575,71.98646013347366,7.128095961116031,0.285489468988363,0.07130434782608695,0.008695652173913044,0.004545946250987143,0,0.497314453125,-0.01129150390625,-0.01074229080098363,0,29,0,0.001,0.5,0.03
+logit,torch,bf16,480,64.0006510476272,5.753001711505866,0.2887366319272604,0.0625,0.004166666666666667,0.24954241196000737,0,0.498046875,-0.00390625,-0.007812539736793652,0,10,0,0.016,0.5,0.02
+sinc,torch,f16,588,1.0,1.0,0.059106073512140145,0.0,0.0,0.00024115754367781683,0,-37568.0,5.960464477539063e-08,0.0,0,0,0,0.001,1e-05,0.03
+sinc,torch,bf16,616,8.366493961506314e+32,6.591783121186793e+32,0.05288922090080902,0.24675324675324675,0.24675324675324675,0.0018124671372343748,0,-1352.0,7.683411240577698e-08,0.0,0,0,0,0.016,1e-05,0.03
+entr,torch,f16,571,1.1911701490793096,0.941464544697664,0.231875189432804,0.0035026269702276708,0.0,27.97501637348614,0,3088.0,-24832.0,-24812.941277614733,0,0,0,0.001,0.1,0.02
+entr,torch,bf16,608,1.2041861399076887,1.0059213428202278,0.30769754676849176,0.013157894736842105,0.0,2528.1952394642867,0,6.716849298982197e-15,2.2026824808563106e-13,2.1919871594286306e-13,0,0,0,0.016,0.1,0.03
+polygamma_1,torch,f16,588,0.5313563672898688,0.4812044346716291,0.03925370341215999,0.0,0.0,12.837055171883549,0,-27.140625,53.9375,53.95410488647781,0,0,0,0.001,1e-05,0.28
+polygamma_1,torch,bf16,616,0.49804144352119323,0.491512863022444,0.21772804902178775,0.0,0.0,3.9146045550699396e+33,75,2.015625,0.640625,0.6386795256112453,0,0,0,0.1,10.0,0.26
+polygamma_2,torch,f16,588,14.045201776369824,7.387401080132452,0.0,0.013605442176870748,0.006802721088435374,13.40193968929816,0,-578.5,-2.1457672119140625e-06,-2.982926474593271e-06,0,0,0,0.001,1e-05,0.3
+polygamma_2,torch,bf16,616,1.2061074283109008,0.4984591141373135,0.2007174268790859,0.0016233766233766235,0.0,4.2978092811015023e+21,0,-110.5,-8.058547973632812e-05,-8.116059657493157e-05,0,0,0,0.1,10.0,0.28
+gelu,torch,f16,565,1.1881916773348211,0.499999854473665,0.20044231985055688,0.005309734513274336,0.0,0.0009315074694753456,0,-5.5703125,-0.0,-7.082174285261757e-08,3,0,0,0.001,1e-05,0.05
+gelu,torch,bf16,618,241.6946820417233,155.95439972661566,0.011518837421012803,0.024271844660194174,0.021035598705501618,0.00770458022777154,0,-7.875,-0.0,-1.3416750046608877e-14,0,0,0,0.016,1e-05,0.06
+gelu_tanh,torch,f16,565,0.9071118194365897,0.4999999933419324,0.19785984246498592,0.0,0.0,0.0008413279348808977,0,-5.20703125,-0.0,-5.406807776907621e-08,0,0,0,0.001,1e-05,0.04
+gelu_tanh,torch,bf16,618,250.04118844059732,163.29967685968398,0.010957276982555794,0.01779935275080906,0.016181229773462782,0.007754819576731631,0,-9.125,-0.0,-1.23279823909867e-29,0,0,0,0.016,1e-05,0.05
+silu,torch,f16,569,0.5000000149011612,0.49999938428401947,0.1983509299786978,0.0,0.0,0.0020256671500202827,0,5.960464477539063e-08,0.0,2.980232327587373e-08,0,0,0,0.001,1e-05,0.03
+silu,torch,bf16,619,67.7472358043113,0.49195979713031013,0.024109840208281073,0.0016155088852988692,0.0016155088852988692,0.014749840050201968,0,-92.5,-0.0,-6.221601013421355e-39,0,0,0,0.016,1e-05,0.03
+mish,torch,f16,569,0.4999120478296914,0.48618690812981913,0.2010713378968312,0.0,0.0,0.0016890522381167727,0,0.4521484375,0.333251953125,0.33337400196480216,0,0,0,0.001,0.01,0.04
+mish,torch,bf16,619,0.49449324044528936,0.4857935599377371,0.20385758177384955,0.0,0.0,0.007381023396877406,0,0.00112152099609375,0.000675201416015625,0.0006733150740034283,0,0,0,0.016,1e-05,0.04
+softplus,torch,f16,577,0.49851656125264876,0.4927780320059944,0.2876126008884409,0.0,0.0,0.001748125916124721,0,1.056640625,1.35546875,1.3549819174206517,0,0,0,0.01,0.01,0.03
+softplus,torch,bf16,618,0.4996284353988572,0.49560855289998085,0.42088462033541374,0.0,0.0,0.013653602756594552,0,0.00042724609375,0.6953125,0.6933608264242233,0,0,0,0.016,0.01,0.04
+logsigmoid,torch,f16,577,0.49851656125264876,0.4939507931537791,0.2876126008884409,0.0,0.0,0.001748125916124721,0,-1.056640625,-1.35546875,-1.3549819174206517,0,0,0,0.001,0.01,0.03
+logsigmoid,torch,bf16,618,0.4996284353988572,0.49560855289998085,0.42088462033541374,0.0,0.0,0.013653602756594552,0,-0.00042724609375,-0.6953125,-0.6933608264242233,0,0,0,0.016,0.005,0.04
+elu,torch,f16,569,0.4993940605685672,0.4719907161479319,0.0,0.0,0.0,0.00024384475613703493,0,-7.21875,-0.9990234375,-0.999267282256137,0,0,0,0.001,1e-05,0.03
+elu,torch,bf16,619,0.49681778446169805,0.4637805853750317,0.0,0.0,0.0,0.0019304541362277217,0,-0.056884765625,-0.05517578125,-0.0552970746544096,0,0,0,0.016,1e-05,0.03
+selu,torch,f16,569,0.4992712076423498,0.49346596296490464,0.26541042774774526,0.0,0.0,0.02847055252087216,0,-0.1016845703125,-0.169921875,-0.16998282119233915,0,0,0,0.001,1e-05,0.04
+selu,torch,bf16,619,0.4982157977990872,0.4950991039854694,0.19653120297821333,0.0,0.0,0.22652257062782155,0,-0.09765625,-0.1630859375,-0.16357247636503816,0,0,0,0.016,1e-05,0.04

ulpwise-0.3.2/studies/accuracy-survey-2026-10-half/results.md

@@ -0,0 +1,105 @@
+# Accuracy survey
+
+Max error in ulps of the dtype against mpmath at 200 bits, about 60 decimal digits (p99 in brackets), on log spaced inputs over each function's domain plus the named edge values of the dtype. The reference is evaluated at the rounded input, so the numbers measure the implementation and not the conditioning of the function. An error is |got - exact| divided by the spacing of the dtype at the exact value, which is why an exact result of zero or a result in the subnormal range gives enormous numbers: read those rows together with the notes in README.md next to this file.
+
+Versions: ulpwise 0.3.1, numpy 2.2.6, torch 2.14.0+cpu, scipy 1.18.1, mpmath 1.3.0, torch cpu capability AVX512
+
+torch columns: `vec!=scalar` is the number of inputs where the vectorized kernel and the scalar tail disagree; `default fail` and `op fail` count inputs that fail torch's reference test tolerance for the dtype, with the default tolerance and with the op's own override respectively. A row with `default fail > 0` and `op fail = 0` is an error that torch's test suite tolerates because of the override.
+
+| function | dtype | torch max (p99) | torch vec!=scalar | torch default fail | torch op fail | op (rtol, atol) |
+|---|---|---|---|---|---|---|
+| exp | bf16 | 0.5 (0.48) | 0 | 0 | 0 | (0.016, 1e-05) |
+| exp | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| exp2 | bf16 | 0.49 (0.48) | 0 | 0 | 0 | (0.016, 1e-05) |
+| exp2 | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| expm1 | bf16 | 0.49 (0.47) | 0 | 0 | 0 | (0.016, 1e-05) |
+| expm1 | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| log | bf16 | 0.5 (0.5) | 0 | 0 | 0 | (0.016, 0.05) |
+| log | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| log2 | bf16 | 0.5 (0.5) | 0 | 0 | 0 | (0.016, 0.1) |
+| log2 | f16 | 0.5 (0.5) | 0 | 0 | 0 | (0.001, 1e-05) |
+| log10 | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 0.05) |
+| log10 | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| log1p | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 0.1) |
+| log1p | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| sqrt | bf16 | 0.5 (0.5) | 0 | 0 | 0 | (0.016, 0.07) |
+| sqrt | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| rsqrt | bf16 | 1.1 (0.5) | 184 | 0 | 0 | (0.016, 1e-05) |
+| rsqrt | f16 | 0.95 (0.5) | 128 | 0 | 0 | (0.001, 0.05) |
+| reciprocal | bf16 | 0.5 (0.5) +1 nonfinite | 0 | 0 | 0 | (0.016, 1e-05) |
+| reciprocal | f16 | 0.5 (0.5) +1 nonfinite | 0 | 0 | 0 | (0.001, 1e-05) |
+| sin | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 0.01) |
+| sin | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| cos | bf16 | 0.5 (0.47) | 0 | 0 | 0 | (0.016, 0.01) |
+| cos | f16 | 0.5 (0.5) | 0 | 0 | 0 | (0.001, 1e-05) |
+| tan | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 1e-05) |
+| tan | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| asin | bf16 | 0.49 (0.48) | 0 | 0 | 0 | (0.016, 0.01) |
+| asin | f16 | 0.49 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| acos | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 0.1) |
+| acos | f16 | 0.5 (0.5) | 0 | 0 | 0 | (0.001, 0.01) |
+| atan | bf16 | 0.5 (0.48) | 0 | 0 | 0 | (0.016, 0.01) |
+| atan | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| sinh | bf16 | 0.48 (0.43) | 0 | 0 | 0 | (0.016, 1e-05) |
+| sinh | f16 | 0.5 (0.47) | 0 | 0 | 0 | (0.001, 0.01) |
+| cosh | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 1e-05) |
+| cosh | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| tanh | bf16 | 0.49 (0.42) | 0 | 0 | 0 | (0.016, 0.01) |
+| tanh | f16 | 0.5 (0.47) | 0 | 0 | 0 | (0.001, 1e-05) |
+| asinh | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 0.05) |
+| asinh | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| acosh | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 0.05) |
+| acosh | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| atanh | bf16 | 0.5 (0.48) | 0 | 0 | 0 | (0.016, 0.01) |
+| atanh | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+| erf | bf16 | 254 (241) | 0 | 0 | 0 | (0.016, 0.01) |
+| erf | f16 | 4.6 (4) | 0 | 0 | 0 | (0.001, 0.01) |
+| erfc | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 0.01) |
+| erfc | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 0.01) |
+| erfinv | bf16 | 0.5 (0.5) | 0 | 0 | 0 | (0.016, 0.01) |
+| erfinv | f16 | 0.5 (0.5) | 0 | 0 | 0 | (0.001, 0.01) |
+| lgamma | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 1e-05) |
+| lgamma | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 0.7) |
+| digamma | bf16 | 0.5 (0.49) +1 nonfinite | 0 | 0 | 0 | (0.016, 1e-05) |
+| digamma | f16 | 0.5 (0.49) +1 nonfinite | 0 | 0 | 0 | (0.001, 0.5) |
+| sigmoid | bf16 | 0.73 (0.5) | 0 | 0 | 0 | (0.016, 0.01) |
+| sigmoid | f16 | 0.5 (0.5) | 0 | 0 | 0 | (0.001, 0.01) |
+| i0 | bf16 | 0.5 (0.48) | 0 | 0 | 0 | (0.016, 1e-05) |
+| i0 | f16 | 0.5 (0.48) | 0 | 0 | 0 | (0.001, 1e-05) |
+| i0e | bf16 | 0.65 (0.5) | 413 | 0 | 0 | (0.016, 0.3) |
+| i0e | f16 | 1.5 (0.75) | 303 | 0 | 0 | (0.001, 0.3) |
+| i1 | bf16 | 0.5 (0.49) | 0 | 0 | 0 | (0.016, 1e-05) |
+| i1 | f16 | 0.5 (0.5) | 0 | 0 | 0 | (0.001, 1e-05) |
+| i1e | bf16 | 0.5 (0.48) | 0 | 0 | 0 | (0.016, 1e-05) |
+| i1e | f16 | 0.5 (0.5) | 0 | 0 | 0 | (0.001, 1e-05) |
+| logit | bf16 | 64 (5.8) | 0 | 10 | 0 | (0.016, 0.5) |
+| logit | f16 | 72 (7.1) | 0 | 29 | 0 | (0.001, 0.5) |
+| sinc | bf16 | 8.4e+32 (6.6e+32) | 0 | 0 | 0 | (0.016, 1e-05) |
+| sinc | f16 | 1 (1) | 0 | 0 | 0 | (0.001, 1e-05) |
+| entr | bf16 | 1.2 (1) | 0 | 0 | 0 | (0.016, 0.1) |
+| entr | f16 | 1.2 (0.94) | 0 | 0 | 0 | (0.001, 0.1) |
+| polygamma_1 | bf16 | 0.5 (0.49) +75 nonfinite | 0 | 0 | 0 | (0.1, 10) |
+| polygamma_1 | f16 | 0.53 (0.48) | 0 | 0 | 0 | (0.001, 1e-05) |
+| polygamma_2 | bf16 | 1.2 (0.5) | 0 | 0 | 0 | (0.1, 10) |
+| polygamma_2 | f16 | 14 (7.4) | 0 | 0 | 0 | (0.001, 1e-05) |
+| gelu | bf16 | 242 (156) | 0 | 0 | 0 | (0.016, 1e-05) |
+| gelu | f16 | 1.2 (0.5) | 3 | 0 | 0 | (0.001, 1e-05) |
+| gelu_tanh | bf16 | 250 (163) | 0 | 0 | 0 | (0.016, 1e-05) |
+| gelu_tanh | f16 | 0.91 (0.5) | 0 | 0 | 0 | (0.001, 1e-05) |
+| silu | bf16 | 68 (0.49) | 0 | 0 | 0 | (0.016, 1e-05) |
+| silu | f16 | 0.5 (0.5) | 0 | 0 | 0 | (0.001, 1e-05) |
+| mish | bf16 | 0.49 (0.49) | 0 | 0 | 0 | (0.016, 1e-05) |
+| mish | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 0.01) |
+| softplus | bf16 | 0.5 (0.5) | 0 | 0 | 0 | (0.016, 0.01) |
+| softplus | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.01, 0.01) |
+| logsigmoid | bf16 | 0.5 (0.5) | 0 | 0 | 0 | (0.016, 0.005) |
+| logsigmoid | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 0.01) |
+| elu | bf16 | 0.5 (0.46) | 0 | 0 | 0 | (0.016, 1e-05) |
+| elu | f16 | 0.5 (0.47) | 0 | 0 | 0 | (0.001, 1e-05) |
+| selu | bf16 | 0.5 (0.5) | 0 | 0 | 0 | (0.016, 1e-05) |
+| selu | f16 | 0.5 (0.49) | 0 | 0 | 0 | (0.001, 1e-05) |
+
+## Errors tolerated by a torch override
+
+- `logit` f16: 29 of 575 inputs fail the default tolerance, none fail the op's (rtol 0.001, atol 0.5); worst x = 0.497314453125, got -0.01129150390625, exact -0.01074229080098363 (72 ulp).
+- `logit` bf16: 10 of 480 inputs fail the default tolerance, none fail the op's (rtol 0.016, atol 0.5); worst x = 0.498046875, got -0.00390625, exact -0.007812539736793652 (64 ulp).

{ulpwise-0.3.1 → ulpwise-0.3.2}/tests/test_survey.py

@@ -217,6 +217,9 @@ def test_every_reference_agrees_with_the_libraries_it_measures(entry):
         ("lgamma", -2.0, math.inf),
         ("lgamma", -2.5, "torch"),
         ("digamma", -2.0, None),  # torch returns nan at the poles
+        ("polygamma_1", -2.0, math.inf),  # the pole of an odd order is +inf from both sides
+        ("polygamma_1", 0.0, math.inf),
+        ("polygamma_2", -2.0, -math.inf),  # an even order takes the sign of its formula, as scipy, torch and jax do
         ("erfinv", 1.0, math.inf),
         ("erfinv", -1.0, -math.inf),
         ("ndtri", 0.0, -math.inf),