Merge adapters branch into main #1158

This change ensures that USM allocation APIs don't return `UR_RESULT_SUCCESS` when an error occurs within `USMAllocationMakeResident`. Signed-off-by: Michael Aziz <michael.aziz@intel.com>

Input and match files were printed only on one error. Now, it will be printed for both places (where this makes sense). This is an extension to commit: be85a49

…ch start

[UR][Tests] Update match scripts

The E2E test assumes that the device supports priorities (sm_35 or newer).

[UR][HIP] Move HIP adapter sourcer to Unified Runtime

On Windows we may build L0 using ICX , this is to revert changes to check WIN32 instead of MSVC, to fix build failures with ICX on Windows.

[UR][L0][CMake] Disable warning for ICX on Windows

[L0] fix windows cmake scripts

We allow reusing existing pre-built L0 Libs and Headers. However, using LEVEL_ZERO_LIBRARY and LEVEL_ZERO_INCLUDE_DIR directly may leak the path of shared folder. This PR introduces two L0_LIBRARY and L0_INCLUDE_DIR , and copy L0 loader/headers locally to the build to avoid leaking their location on a shared drive.

[UR][L0] Copy prebuilt L0 to avoid leaking shared folder path

Adapter implementations are being moved into the repo we should mirror the code ownership of those implementations. This patch adds a CODEOWNERS file which matches the pre adapter move ownership to be used as a baseline for future changes for each individual adapter.

The new priority parameter hadn't been reflected here so the ownership property was being used as priority and default to `true` which caused crashes.

For kernel fusion support for hip, we need to finalize the kernels using comgr. The patch finalizes tagged binaries during buildProgram before handing it over to the hip runtime. Signed-off-by: Victor Lomuller <victor@codeplay.com>

- Second return has no effect - Checking the same flag twice has no effect

Add CODEOWNERS file

[UR][L0] Propagate errors from `USMAllocationMakeResident`

…elzero-adapter [UR][L0] Fix minor code duplication

The HIP team needs some fixing up currently so use the CUDA team for now.

Add code owners for the CUDA and HIP adapters

…-zero [UR][L0] Add adapter source to codeowners file

Update HIP adapter codeowners to use specific team

The L0 events used to implement UR sync-points in a command-buffer aren't reset after the first execution. Leaving them in a signalled state before subsequent submissions of the UR command-buffer. This patch resets the L0 events used as UR command-buffer sync-points to the non-signalled state, by appending event reset commands to the end of the user defined command-list. There is also a reset added for the UR command-buffer wait event that is signalled when the wait list dependencies of `urCommandBufferEnqueueExp` are met. Additionally this patch removes the host scope flag from being set on L0 events created by the UR command-buffer, as we never signal/reset them from host, only device.

Rather than having a command-list wide barrier and separate signal command, have a single barrier command which signals the UR command-buffer signal event and waits on the events that correspond to sync-points. Co-authored-by: Pablo Reble <pablo.reble@intel.com>

Co-authored-by: Martin Morrison-Grant <martin.morrisongrant@codeplay.com> Co-authored-by: Petr Vesely <petr.vesely@codeplay.com> Co-authored-by: Callum Fare <callum@codeplay.com> Co-authored-by: aarongreig <aarongreig01@gmail.com>

- Update Cmake to use local adapter source files - Update license headers - Add virtual memory entrypoints to interface_loader - Add .clang-format - Update CODEOWNERS file for OpenCL adapter

[HIP] Re add extended deleters

…#969) - Upstream CI is unable to use find_package() for OpenCL. This commit allows it to set the target manually and skip the call to find_package(). - This commit also changes UR to use the OpenCL::OpenCL target. It provides a simpler alternative to import the opencl library without having to use add_library().

Set [unified-runtime-command-buffer-write](https://github.com/orgs/oneapi-src/teams/unified-runtime-command-buffer-write] as the codeowners of the command-buffer experimental feature. Including: * Adapter implementations * Specification document * YML definition

Also merge urQueueCreate InvalidValueProperties test into InvalidQueueProperties test.

…rror_checking Revert "[UR][L0] Propagate errors from `USMAllocationMakeResident`"

…eowners Set command-buffer experimental feature codeowners

- Implement initial subset of command buffer functions for CUDA - Report command buffer extension as supported from device query - Some additional methods and refactoring to reuse existing adapter code

[CUDA] Fix queue creation with native handle

[UR][HIP] Enable kernel finalization using comgr

Add `unified-runtime-bindless-images-write` team as code owners for anything related to images.

…owners CODEOWNERS for Bindless Images

[Command-buffer][L0] Reset sync-point events

Print all CTS outputs (from gtest) in CI.

This change ensures that USM allocation APIs don't return `UR_RESULT_SUCCESS` when an error occurs within `USMAllocationMakeResident`. Signed-off-by: Michael Aziz <michael.aziz@intel.com>

…iant [HIP][CUDA] Change unions in ur_mem_handle_t_ to stdvariant

…me-main-changes-into-adapters

…-changes-into-adapters Bump adapters branch

This is required after oneapi-src#961

…some-main-changes-into-adapters Revert "Bump adapters branch"

Signed-off-by: Michael Aziz <michael.aziz@intel.com>

Normally OpenCL limits fill type operations to a max pattern size of 128, this patch includes a workaround to extend that.

[CUDA][EXP] CUDA adapter support for command buffers

Also ignore flags in no-op urEnqueueUSMPrefetch hint.

Signed-off-by: Michael Aziz <michael.aziz@intel.com>

Instead of creating an event, appending, and synchronizing it on it to wait for all commands in an immediate command list, emit a zeCommandListHostSynchronize instead. Signed-off-by: Jaime Arteaga <jaime.a.arteaga.molina@intel.com>

Signed-off-by: Michael Aziz <michael.aziz@intel.com>

Signed-off-by: Jaime Arteaga <jaime.a.arteaga.molina@intel.com>

This patch moves the `test-adapter-cuda` and `test-adapter-hip` test suites out of the `test/conformance` directory. The goal of this move is to ensure that all adapters will be tested with the same set and number of conformance tests, making conformance pass rates directly comparable between adapters. Up to now the CUDA and HIP adapters have included additional unit tests of internals which skews pass rates.

…SpecializationConstants [OpenCL] Add version check for urProgramSetSpecializationConstants

The code introduced in 74f42f8 uses the signature of cuGraphInstantiate from CUDA 12.x. In CUDA 11.x, this function has different parameters.

[CUDA][HIP] Move adapter specific testing out of CTS

We rather want to fail CMake configuration than silently move on, to find out about missing example, after the build. Codegen example is only build "on demand", when UR_BUILD_EXAMPLE_CODEGEN is set.

automatically generate linker version script for adapter libraries

…calMemSizeQuery [UR] [CUDA] Changed the output of querying localMemSize

Also return RESULT_SUCCESS for no-op UR_KERNEL_EXEC_INFO_CACHE_CONFIG hint.

…anch

Also includes a few other GetInfo related fixes: * Add missing device info queries * Add mapping of CL command type to UR command type * Correct mapping of UR_QUEUE_INFO_FLAGS * Add mapping of cl_command_queue_properties to ur_queue_flags_t * Add mapping of cl_unified_shared_memory_type_intel to ur_usm_type_t * Add UNSUPPORTED_ENUMERATION path to KernelGeGroupInfo tests. And a fix related to one of the fixed queries: * Populate pfnReadHostPipe and pfnWriteHostPipe ddi table entries.

Fail codegen example configuration if no dependcies

When initializing conda few extra commands are added to user's profile. These commands make it impossible to log in as the test user to the machine, since conda may be missing. We can safely remove the whole '.profile' after the job.

…leak [OpenCL] Fix memory leak

The CTS is already dependant on the DPC++ compiler to generate programs inputs for the program, kernel, and enqueue test suites specified via the `UR_DPCXX` CMake variable. If the DPC++ compiler is not installed on the system the executables it outputs will likely fail to find the SYCL runtime library when executed, breaking the generation of CTS program inputs. The patch introduces the `UR_SYCL_LIBRARY_DIR` CMake variable enabling the user to specify the path to the SYCL runtime library for use when generating CTS program inputs.

…e-sycl-library-dir [CTS] Add UR_SYCL_LIBRARY_DIR CMake variable

…changes-into-adapters

This allows us to return UR_ERROR_INVALID_SIZE when we should. Extra checks are only performed on a non-success error code. Also adds a missing bounds check to urMemBufferPartition

[UR][L0] Add support for urAdapterGetLastError in L0

[UR][L0] Add support for zeCommandListHostSynchronize

[UR][L0] Propagate OOM errors from `USMAllocationMakeResident`

[SYCL][UR][L0] Distinguish min/max calls from macros

This PR refactors the current SYCL Native CPU Implementation as a Unified Runtime adapter. No new functionalities are added. Contents of this PR: - Native CPU implementation divided in multiple source files at `sycl/plugins/unified_runtime/ur/adapters/native_cpu` - Native CPU PI plugin is now implemented through the `pi2ur.hpp` interface - _pi_* classes have now an equivalent ur_* class - `UR_PLATFORM_BACKEND_NATIVE_CPU` value added to `ur2piPlatformInfoValue`

Bump the Unified Runtime commit, and make adapter changes needed for the newly added adapter handles (see oneapi-src#715 for details) This fixes #10066 by providing an implementation of `piPluginGetLastError` in pi2ur.

This PR fixes some errors while building the Native CPU plugin: Updated urPlatformGet signature Added missing function urAdapterGetLastError

…invocation (#10495) This PR adds support to multiple SYCL targets alongside `native_cpu` in the same compiler invocation (e.g. `clang++ -fsycl -fsycl-targets=native_cpu,spir64 input.cpp`). In order to implement this we had to make changes to multiple components, here is a quick overview: * Driver: changes in the Driver allow to correctly parse all the targets passed to `-fsycl-targets` (before we were just looking for `native_cpu`, ignoring the others). The Driver now also calls `sycl-post-link` and `clang-offload-wrapper`, performing a compilation flow more similar to the one used for other targets. * Sema: since the kernel name needs to be the same for all the SYCL targets, the change to the kernel name in Sema has been removed, and replaced with an LLVM Pass that gets run when lowering the device module (`llvm/lib/SYCLLowerIR/RenameKernelSYCLNativeCPU.cpp`). * Runtime: The definition for `_pi_program` in the Native CPU Plug-In now supports multiple kernels in one program, and the `__SYCL_PI_DEVICE_BINARY_TARGET_NATIVE_CPU` binary type has been added in order to identify kernels compiled for Native CPU. * clang-offload-wrapper: for Native CPU, the offload-wrapper doesn't bundle the device code in the host module, but instead produces an array containing function declarations that are resolved by the linker, see `sycl/doc/design/SYCLNativeCPU.md` for more information. --------- Co-authored-by: Uwe Dolinsky <uwe@codeplay.com> Co-authored-by: Alexey Sachkov <alexey.sachkov@intel.com> Co-authored-by: Steffen Larsen <steffen.larsen@intel.com>

This PR adds some initial support for memory copy operations to the NativeCPU device. It also fixes global offsets and the query for queue properties in the NativeCPU PI/UR adapter.

Removes some unneeded UR_ASSERTS Added initial GlobalTimeStamp implementation for `sycl-ls`.

Processed sycl, libclc and libdevice directories.

Fix the license headers at the top of each source file in the unified runtime directory. --------- Co-authored-by: Alexey Bader <alexey.bader@intel.com>

Adding more UR adapter functionality to pass more tests related to buffers and subbuffers, and to enable `sycl-ls --verbose`.

…and ABI Check (#10883) This PR adds the remaining (not fully implemented) UR interface functions to the NativeCPU/UR adapter. The NativeCPU PI plugin now also exports all PI interface functions and an ABI check was added.

@call

This PR allows linking to libclc when compiling for SYCL Native CPU. Currently only the `x86_64-unknown-linux-gnu` target triple is supported, additional target triples (and possibly a more versatile way of setting them) will come with follow up PRs. Some useful information for reviewing: * We start using an `AddrSpaceMap` (set in `TargetInfo.cpp`) because the mangled names emitted by the device compiler need to match with the names provided by `libclc`. The AddressSpaceMap is taken from the `PTX` Target. * Changes in `Driver` are needed to find and link to `libclc`. * `libclc/ptx-nvidiacl/libspirv/atomic/loadstore_helpers.ll` has been split into 4 modules, one for each memory ordering constraint. Copies of these modules have been added in `generic` (because some functions in `generic/libspirv/atomic` needed them), and the module split allows to specialize the file for targets that may not support some orderings. Currently only a couple of function for `acquire` and `seq_cst` have been implemented for `generic`, but the others will be implemented in a follow up PR. * We've added a target in `libclc` for `x86_64-unknown-linux`. This has been done because some math builtins in `generic` have been defined as ``` typedef char vec __attribute__((ext_vector_type(8))); __attribute__((overloadable)) vec __clc_native_popcount(vec x) __asm("llvm.ctpop" ".v16i" "8"); vec call(vec x) { return __clc_native_popcount(x); } ``` While this approach conveniently allows to call directly LLVM intrinsics, it does seem to play well with the ABI for `x86_64-unknown-linux`, since it leads to this IR: ``` define dso_local double @call(double noundef %x.coerce) #0 { entry: %0 = bitcast double %x.coerce to <8 x i8> %1 = bitcast <8 x i8> %0 to double %call = call double @llvm.ctpop.v8i8(double noundef %1) oneapi-src#8 %2 = bitcast double %call to <8 x i8> %3 = bitcast <8 x i8> %2 to double ret double %3 } ``` Which is invalid because `lvm.ctpop.v8i8` expect a vector of `i8` and not a `double`, leading to failing asserts in the compiler that prevented from building `libclc`. As a temporary work around we have added empty files that override the files in `generic` when building for `x86_64-unknown-linux`, allowing to complete the build, even though the corresponding builtins will be missing from the library. We are working on a proper solution for this. --------- Co-authored-by: Uwe Dolinsky <uwe@codeplay.com>

This PR adds support to `local_accessors` by implementing `urKernelSetArgLocal`.

Fetch the latest revision of unified runtime: Notable changes * Some command buffer entry-points have been renamed, also 2 additions of membufferFill and USMFill * UrInit/urTearDown have been removed - replaced with loader only versions.

This change adds a new aspect for esimd, `ext_intel_esimd`, and annotates the two fundamental esimd classes, `simd_obj_impl` and `simd_view_impl` with the `uses_aspect` attribute. `simd_obj_impl` is the base class of `simd` and `simd_mask` which are the fundamental user-facing classes. `simd_view_impl` is the base of only `simd_view`. `simd_obj_impl` is not a base of `simd_view_impl`, but every `simd_view_impl` requires a reference to a `simd` or `simd_mask` at construction time, so I am not sure if we truly need to annotate `simd_view_impl`, but I added it to be safe. It also adds a new PI device info query, `PI_EXT_INTEL_DEVICE_INFO_ESIMD_SUPPORT` that is used to query at runtime if a device supports ESIMD. For UR-based plugins, we map that to `UR_DEVICE_INFO_ESIMD_SUPPOR`. The implementation simply returns false for cuda, hip and native_cpu. For l0 and opencl, we check that the device is an intel gpu by querying the device type is gpu and the vendor id is `0x8086`. For ESIMD emulator we simply return true. I would appreciate careful review on the plugin changes in particular, as I am not an expert. This change also updates the esimd spec to document the new aspect. In a future change, I plan to use the new aspect to remove the requirement for the `SYCL_ESIMD_FUNCTION` and `SYCL_ESIMD_KERNEL` macros that set function attributes, but I am not doing that as part of this PR as it requires more investigation. --------- Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>

Reverts intel/llvm#11155. I'm almost sure this is what broke our CI for the past few days.

This PR removes dependencies to the sycl headers/library in the Native CPU UR adapter. `sycl/include/sycl/detail/native_cpu.hpp` has been moved to `sycl/plugins/unified_runtime/ur/adapters/native_cpu/nativecpu_state.hpp`, and the definitions of work item builtins has been moved from that header to the compiler, which now emits them in the `PrepareSYCLNativeCPUPass`.

…Branch [OpenCL] Combined CTS fixes

…bined

…xes-combined

…me-main-changes-into-adapters-second-patch

…cond-patch

Use a new environment variable, UR_L0_LEAKS_DEBUG, to check for leaks in the UR L0 adapter, instead of relying on a specific value being set in UR_L0_DEBUG. Signed-off-by: Jaime Arteaga <jaime.a.arteaga.molina@intel.com>

…patch

…reate`

- The UR spec was recently changed to guarantee that hint entryponts never return errors. This commit changes the CUDA adapter to be conformant with this change. - This commit also changes the type of PointerRangeSize which was causing a stack corruption.

CUDA does not make a distinction between binaryTypes (it treats PTX and binaries using the same entrypoints). However, for UR, by definition: * urProgramCompile should set the binary type to UR_PROGRAM_BINARY_TYPE_COMPILED_OBJECT * urProgramBuild / urProgramLink should set it to UR_PROGRAM_BINARY_TYPE_EXECUTABLE. * urProgramCreateWithBinary should set the binary type to UR_PROGRAM_BINARY_TYPE_COMPILED_OBJECT

…andBufferExp, PhysicalMem, USMExp, VirtualMem.

Also fix a similar issue coverity had with CreateProgramWithBinary which was passing array pointers straight through to the cl entry point by putting these params in local const arrays.

…PUAdapterToUR [NATIVECPU] Move Native CPU adapter to UR.

…cond-patch

…kflow [CI] Clean up after conda init in cmake.yml

…xes_cuda_hip [CUDA][HIP] Combined CTS Fixes

…o merge-some-main-changes-into-adapters-third-patch

piProgramBuild receives a list of devices, while urProgramBuild does not. This produces a series of issues when a UR program needs to be created for a specific device. So define a new API, called urProgramBuildExp to pass this list. Authored-by: jaime.a.arteaga.molina@intel.com

Expand upon the introduction of `urProgramBuildExp` and include `urProgramCompileExp` and `urProgramLinkExp` which include a device-list in place of a context. These more closely align with the PI/OpenCL analogues but only to introduce device-lists, not all extant arguments from those entry-points. This patch also moves the `urProgramBuildExp` definition into an experimental feature file and introduces a brief document containing motivation.

…ompile piProgramBuild receives a list of devices, while urProgramBuild does not. This produces a series of issues when a UR program needs to be created for a specific device. So define a new API, called urProgramBuildExp to pass this list. Signed-off-by: Spruit, Neil R <neil.r.spruit@intel.com> Co-authored-by: Jaime Arteaga <jaime.a.arteaga.molina@intel.com>

Add the `urProgramBuildExp`, `urProgramCompileExp`, and `urProgramLinkExp` to the loader `ur_program_exp_dditable_t`. Also add `"ur_exp_multi_device_compile"` to the list of extensions supported by the L0 adapter, enables the SYCL RT to query support.

Align the `urProgramLinkExp` spec and implementation argument orders to fix Windows link error.

…nition to match spec Signed-off-by: Spruit, Neil R <neil.r.spruit@intel.com>

…fferCreate_leak [OpenCL] Fix memory leak and coverity issue with struct-to-array casts

[UR][L0] Add support for passing device list to urProgramBuild/Link/Compile

Ensure that all adapters have the correct signatures for the multi-device compile experimental feature entry points and that they entry points exist even when returning `UR_RESULT_ERROR_UNSUPPORTED_FEATURE`.

…1025c040110c332

…ice-compile [CUDA][HIP][OpenCL][NATIVECPU] Fix multi-device compile

[UR][L0] Add UR_L0_LEAKS_DEBUG key

…dBuffers [OpenCL] Add Command Buffer extension to OpenCL adapter.

[HIP] Revert add prefetch for USM hip allocations a6b8fa66b537753415d24076f…

[CMake] Bump version to v0.9.0

[UR][CUDA] Fix compatibility with CUDA 11.x

[NATIVECPU] Implement correct reference counting for Native CPU adapter

…-warnings [NATIVECPU] Fix build errors when -Werror=on

…-of-regex-in-opencl [UR][OPENCL] Eliminate usage of regex in opencl

Introduces the follow CMake options: * `UR_OPENCL_INCLUDE_DIR` - directory containing the OpenCL Headers * `UR_OPENCL_ICD_LOADER_LIBRARY` - path of the OpenCL ICD Loader library In the event that `UR_OPENCL_INCLUDE_DIR` is not specified, clone [KhronosGroup/OpenCL-Headers](https://github.com/KhronosGroup/OpenCL-Headers.git). In the event that `UR_OPENCL_ICD_LOADER_LIBRARY` is not specified, first inspect the system with `find_package(OpenCL 3.0)` and use that if found. Otherwise, clone [KhronosGroup/OpenCL-ICD-Loader](https://github.com/KhronosGroup/OpenCL-ICD-Loader.git).

[CI] Add support for OpenCL CI on CPU

…ird-patch

[HIP] Ifdef entry point introduced in HIP 5.6

…patch

…-l0-clang-cl [GHA] Disable Windows L0 clang-cl job

…ters.

…ird-patch

fixes oneapi-src#1002

Add system info in CI workflows run on self-hosted runners

…-changes-into-adapters-third-patch Merge main into adapters branch

[HIP] Adapter PR batch

…pters CMake Flag for building all adapters

…er-refcount-fix [UR] Fix adapter refcount tracking

…erGet [UR] Fix urAdapterGet ignores NumEntries

Separating the device code compilation from the rest of the setup lets a test query device or platform properties to decide whether it should skip without potentially hitting a compilation error because the kernel being built as part of the setup uses an unsupported feature.

Update the `scripts/ctest_parser.py` to work with the adapter branch changes.

…rnelFixtures Allow fixtures derived from KernelTest to defer compilation.

This is accomplished with the various size queries for buffers, images and USM allocations. Since not all adapters have these queries implemented the bounds checking isn't entirely comprehensive on all platforms just yet.

…rser [CTS] Update ctest_parser.py script

[HIP] Hip adapter multi dev ctx

…ckingValidation Add bounds checking for enqueue operations to the validation layer.

Commits on Oct 3, 2023

[L0] fix windows cmake scripts

pbalcer committed Oct 3, 2023

Configuration menu

View commit details

Copy full SHA for ec915f8

Browse repository at this point

Copy the full SHA

ec915f8 View commit details

Browse the repository at this point in the history

Merge adapters branch into main #1158

Merge adapters branch into main #1158

Commits on Sep 28, 2023

Commits on Sep 29, 2023

Commits on Oct 2, 2023

Commits on Oct 3, 2023

Commits on Oct 4, 2023

Commits on Oct 5, 2023

Commits on Oct 9, 2023

Commits on Oct 10, 2023

Commits on Oct 11, 2023

Commits on Oct 12, 2023

Commits on Oct 13, 2023

Commits on Oct 16, 2023

Commits on Oct 17, 2023

Commits on Oct 18, 2023

Commits on Oct 19, 2023

Commits on Oct 20, 2023

Commits on Oct 23, 2023

Commits on Oct 24, 2023

Commits on Oct 25, 2023

Commits on Oct 26, 2023

Commits on Oct 27, 2023

Commits on Oct 30, 2023

Commits on Oct 31, 2023

Commits on Nov 1, 2023

Commits on Nov 2, 2023

Commits on Nov 3, 2023

Commits on Nov 6, 2023

Commits on Nov 7, 2023

Commits on Nov 8, 2023

Commits on Nov 9, 2023

Commits on Nov 10, 2023

Commits on Nov 12, 2023

Commits on Nov 13, 2023

Commits on Nov 14, 2023

Commits on Nov 15, 2023

Commits on Nov 16, 2023

Commits on Nov 17, 2023

Commits on Nov 20, 2023

Commits on Nov 21, 2023

Commits on Nov 22, 2023

Commits on Nov 23, 2023

Commits on Nov 24, 2023

Commits on Nov 27, 2023

Commits on Nov 28, 2023

Commits on Nov 29, 2023

Commits on Nov 30, 2023

Commits on Dec 1, 2023

Commits on Dec 4, 2023

Commits on Dec 5, 2023