Hotfix: Fix offset calculation to prevent overflow if offset is really large #1298
Conversation
amcamd
commented
Mar 17, 2023
amcamd
commented
- Contains:
- Torre's fix to add offset to array pointers in call to RocblasContractionProblem
- Andrew's trsm changes to avoid int32_t overflow in trsm offset calculations
- Alex's tensile_tag.txt for changes in Tensile
* works around tensile side offset size limitations
amcamd
requested review from
TorreZuk,
mahmoodw,
daineAMD,
bragadeesh,
NaveenElumalaiAMD,
rkamd,
yoichiyoshida and
babakpst
as code owners
|
@amcamd not sure how you have clang-format failure if you cherry-picked |
|
I had to resolve a merge conflict, from ROCm 5.5 to now trtri_gemm_block has changed to rocblas_trtri_gemm_block. I needed to reverse the change. I guess that clang-format did not run on the commit after resolving the merge conflict. I will fix it. |
clients/gtest/trsm_gtest.yaml
Outdated
| - { M: 46345, N: 4, lda: 46345, ldb: 46345 } | ||
| # - { M: 47000, N: 4, lda: 47000, ldb: 47000 } # calls rocblas_internal_gemm_template with batch_count=367, stride_a=6016128 | ||
|
|
||
| - &size_t_right_matrix_size_range | ||
| - { M: 4, N: 46345, lda: 46345, ldb: 4 } |
There was a problem hiding this comment.
These tests timed out for me on gfx90a, but passed when I increased the timeout threshold. I'm testing all other configs as well just to be safe.
There was a problem hiding this comment.
@daineAMD Naveen's speedup for the initialization of the triangle matrix in https://github.com/ROCmSoftwarePlatform/rocBLAS-internal/pull/1551 . This PR is in the develop branch but it is not in release/rocm-rel-5.5. Tests will run faster on the develop branch than on release/rocm-rel-5.5.
* split out blas1_gtest to build faster
