unified-runtime icon indicating copy to clipboard operation
unified-runtime copied to clipboard

Results 324 unified-runtime issues
Sort by recently updated
recently updated
newest added

- Run fuzztests in CI on Level Zero instead of the Null Adapter - Fixes [kernel not triggering](https://github.com/PatKamin/unified-runtime/actions/runs/8235356302/job/22519248157#step:8:237) in fuzztests: `Failed to open a kernel source file: /home/runner/work/unified-runtime/unified-runtime/build/test/conformance/device_binaries/bar/sycl_spir641.spv` - Fixes...

ci/cd

There is no point in running a test if it's allowed to segfault (unless maybe we're testing an operating system kernel). If a test can optionally segfault, it's worse than...

conformance

This fixes the implementation of UR_PROGRAM_INFO_DEVICES and UR_PROGRAM_INFO_NUM_DEVICES when using multi-devices programs.

level-zero

This is now reflected in the spec and in the CTS tests. Also implement a few trivial fixes for adapters failing the now enforced non-optional queries.

conformance
specification
level-zero
cuda
hip
opencl
native-cpu

This PR is built on top of [[OPENCL] Add UR handles to OPENCL adapter](https://github.com/oneapi-src/unified-runtime/pull/1176). [Intel/llvm testing](https://github.com/intel/llvm/pull/12662)

opencl

The `UR_DEVICE_INFO_PCI_ADDRESS` query previously returned a string of the form `"DDDD:BB:SS.F"` whereas Level Zero and OpenCL (via the `cl_khr_pci_bus_info` extension) return structs with each field separated out. Additionally, CUDA and...

loader
specification
level-zero
cuda
hip
opencl

- Adds a new entry point `urEventGetSyncPointProfilingInfoExp`. This function queries the profiling information of a sync-point. It takes the handle of the event returned from the graph submission and the...

loader
specification
level-zero
cuda
hip
opencl
native-cpu

have to iron out a few more things, but the initial concept is this: for regular commandlist, we want to give the events device scope thus avoiding L3 flush. To...

level-zero

This patch is adding a microbenchmark for the enqueue operation (of an empty kernel). This is in preparation for upcoming L0 optimization patches so that we can track performance changes...

ci/cd

- `UR_DEVICE_INFO_SUB_GROUP_SIZES_INTEL`: The spec says it's an array of uint32_t, but CUDA, HIP, OpenCL, and LevelZero adapters were returning arrays of size_t. As a result, we had extra dummy zeroes...

conformance
level-zero
cuda
hip
opencl