llvm: Add option to select 32bit floating point representation #2400
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Fixes parallel execution when using fp32 type. Signed-off-by: Jan Vesely <[email protected]>
Signed-off-by: Jan Vesely <[email protected]>
Predator-prey runs both fp32 and fp64 variant irrespective of the --fp-precision setting. Signed-off-by: Jan Vesely <[email protected]>
Add tests. Conversion from double -> half is not accurate, because of an issue with a call to runtime library [0]. [0] numba/llvmlite#834 Signed-off-by: Jan Vesely <[email protected]>
llvmlite does not handle special FP string values for other than fp64 precision [0] [0] numba/llvmlite#833 Signed-off-by: Jan Vesely <[email protected]>
Make sure we always have fp64 variant available. Add type specific tests. Signed-off-by: Jan Vesely <[email protected]>
Signed-off-by: Jan Vesely <[email protected]>
…ding fp64 Signed-off-by: Jan Vesely <[email protected]>
Bump vector length to 1500 to span more than one 4kB page. Signed-off-by: Jan Vesely <[email protected]>
Add const dimension variant to every tested op. Convert arguments to the right fp type. Signed-off-by: Jan Vesely <[email protected]>
This is tested in test_builtins_matrix.py const dimensions tests. Signed-off-by: Jan Vesely <[email protected]>
… again Signed-off-by: Jan Vesely <[email protected]>
Signed-off-by: Jan Vesely <[email protected]>
This should switch to using pytest.mark if it gets more widespread Signed-off-by: Jan Vesely <[email protected]>
The original formula would return NaN if the input was large enough so that exp(2x) == Inf and would need an extra condition check. The new formula handles large inputs without extra checks. Add tests with extreme value to builtins tests. Signed-off-by: Jan Vesely <[email protected]>
Signed-off-by: Jan Vesely <[email protected]>
…le in fp32 There are no special input values in the tests might as well make them representable in 32b floating point. Fixes fp32 test failures in SUB operation tests. Signed-off-by: Jan Vesely <[email protected]>
Signed-off-by: Jan Vesely <[email protected]>
The original formula would return NaN if the input was large enough so that exp(2x) == Inf and would need an extra condition check. The new formula handles large inputs without extra checks. Add tests with extreme values to builtins tests. Signed-off-by: Jan Vesely <[email protected]>
The original formula would return NaN if the input was large enough so that exp(2x) == Inf and would need an extra condition check. The new formula handles large inputs without extra checks. Add tests with extreme value to builtins tests. Use stricter tolerance for fp64 tests. Signed-off-by: Jan Vesely <[email protected]>
Move expected values out of param list. Integrate test names into param list. Re-enable small drift rate test, override expected results on windows and mac. Enable small drift rate test in compiled mode, override expected results. Add more small drift rate tests. Signed-off-by: Jan Vesely <[email protected]>
The results differ from fp64 because of rounding and use of Philox PRNG. Signed-off-by: Jan Vesely <[email protected]>
Signed-off-by: Jan Vesely <[email protected]>
Signed-off-by: Jan Vesely <[email protected]>
Signed-off-by: Jan Vesely <[email protected]>
…ision Signed-off-by: Jan Vesely <[email protected]>
These are now high/low enough to work with fp32 Philox random sequence. Make sure initializer values are different from the test values. Signed-off-by: Jan Vesely <[email protected]>
Drop the hack using inf input values to select PRNG. Signed-off-by: Jan Vesely <[email protected]>
|
This PR causes the following changes to the html docs (ubuntu-latest-3.7-x64): See CI logs for the full diff. |
1 similar comment
|
This PR causes the following changes to the html docs (ubuntu-latest-3.7-x64): See CI logs for the full diff. |
PTX results for fp32 DDA function with neg input differ due to operation accuracy/rounding. Signed-off-by: Jan Vesely <[email protected]>
|
This PR causes the following changes to the html docs (ubuntu-latest-3.7-x64): See CI logs for the full diff. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Add
--fp-precisionpytest cmdline option, accepted values are 'fp64' and 'fp32', 'fp64' is the default.Consolidate tests of compiler builtin helpers.
Refactor implementation of hyperbolic functions (tanh, coth, csch) to handle corner cases that result in the infinite value of exponential.
Adjust tests and add fp32 results for cases where the lower precision and rounding give different results.