POWER: Fix issues in zscal to address lapack failures #4712
Conversation
This patch fixes following lapack failures with clang compiler on POWER. zed.out: ZVX: 18 out of 5190 tests failed to pass the threshold zgd.out: ZGV drivers: 25 out of 1092 tests failed to pass the threshold zgd.out: ZGV drivers: 6 out of 1092 tests failed to pass the threshold
CodSpeed Performance ReportMerging #4712 will not alter performanceComparing Summary
|
| @@ -38,6 +38,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |||
|
|
|||
| #pragma GCC optimize "O1" | |||
|
|
|||
| #if defined(__clang__) | |||
There was a problem hiding this comment.
Is your intention to disable FMA for clang for both SINGLE and DOUBLE versions of this algorithm?
|
At some point it will probably make more sense to replace the Reference-LAPACK testsuite, or at least relax its criteria. I only wish I knew how to define that point, and creating a replacement would be a huge task... |
|
How often are changes/fixes to LAPACK pulled in from? |
|
If it is not a documentation-only fix (and not a huge change either, like having both 32 and 64bit integer versions of all functions in the library), usually within days of the PR getting merged (or sometimes as soon as a plausible PR gets posted). |
|
Just wondering if I had to manually merge LAPACK changes. Thanks for the clarification. |
|
I realize we already have the same fix/workaround on x86_64, so merging this after all. Can still discuss adding another |
This patch fixes following lapack failures with clang compiler on POWER.
zed.out: ZVX: 18 out of 5190 tests failed to pass the threshold
zgd.out: ZGV drivers: 25 out of 1092 tests failed to pass the threshold
zgd.out: ZGV drivers: 6 out of 1092 tests failed to pass the threshold