64-bit offset support in hipBLASLt#7585

Open

shbae wants to merge 25 commits into

users/sbae/64bit_offset_support

Commits on Jun 25, 2026

minor fix of the type of an internal function name.
shbae
committed
[hipblaslt] implement host-side code for the 64bit offset support in hipblaslt.
shbae
committed
[hipblaslt] pass batch offset values as kernel arguments.
shbae
committed
[hipblaslt] add offset parameters into the kernel signature.
shbae
committed
[hipblaslt] update kernel generation to use offsets in general batch mode.
shbae
committed
[hipblaslt] add waitcnt instruction after SLoadB64 before updating the offset.
shbae
committed
[hipblaslt] implement tests for 64-bit offset support
shbae
committed
[hipblaslt] update tensilelite code-gen part for 64bit offset support to add arguments appropriately and to use only two extra SGPRs
shbae
committed
[hipblaslt] fix bugs in computeStoreSrd() and remove unnecessary change.
shbae
committed
[hipblaslt] update 64-bit offset support for BetaOnly and Conversion.
shbae
committed
[hipblaslt] fix bugs in the changes of KernelWriterConversion.py
shbae
committed
[hipblaslt] remove unnecessary if-condition and minor update.
shbae
committed
[hipblaslt] fix CI test failures.
shbae
committed
remove temporary debug implementation.
shbae
committed
[hipblaslt] fix the bug related to CI failures.
shbae
committed
[hipblaslt] place the batch offset argument at the tail of the kernarg buffer
shbae
committed
[hipblaslt] modify the batch offset value as in elements.
shbae
committed
remove temporary debug test and update matmul_batch_offset_large test.
shbae
committed
[hipblaslt] modify matmul_batch_offset_large test with actual 64-bit offset inputs.
shbae
committed
limit the tested gpu_arch for matmul_batch_offset_large tests.
shbae
committed
[hipblaslt] remove checking the positive offset value
shbae
committed
[hipblaslt] add matmul_batch_offset_all_solutions test.
shbae
committed
[hipblaslt] don't add batchOffset arguments to the Custom Kernels, and excludes custom kernels solutions for General batched GEMM.
shbae
committed
[hipblaslt] add tests for negative batch offset values.
shbae
committed
[hipblaslt] check validity with non-zero offsets with POINTER_ARRAY mode only.
shbae
committed