[SYCL][Graph][Doc] Add SYCL-Graph usage guide and example doc #379

Bensuo · 2024-07-29T17:06:11Z

Move examples from spec to reduce bloat
Adds some usage guidelines for common scenarios
Remove reductions from examples since they are not supported

sycl/doc/SYCLGraphUsageGuide.md

EwanC

LGTM as a starting point for this document we can improve on over time.

sycl/doc/SYCLGraphUsageGuide.md

julianmi

LGTM

…ntel#14513) (intel#14899) Reverts intel#14513 This reverts commit 56b1410. Build in postcommit broken https://github.com/intel/llvm/actions/runs/10196353773

The problem was this was a runtime `if` and not `if constexpr`. This function was copy-pasted from the ESIMD headers, and the headers one was already fixed to be `if constexpr`, so just use that. Closes: intel#14877 Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>

) Layout is slightly different on systems like RHEL8 with older libstdc++. Adjust regexp accordingly.

…14883) Returning a std::string from device.cpp to device.hpp will cross the ABI boundary. So, we need to neutralize it via detail::string.

In the same vein as intel#14273, this PR prevents exceptions from leaking in additional destructors caught by Coverity. I'd like to draw attention to `device_impl.cpp` however: A comment was left suggesting that exceptions in the `device_impl` destructor be added to the asynchronous exceptions list. Given that devices are usually destroyed during shutdown, adding exceptions to the exceptions list doesn't seem to make sense, as there would be nothing to handle the exceptions anyway. However, if this understanding is incorrect, and I should still add exceptions to an asynchronous exceptions list, please let me know. Thanks!

For intel#14763

…TOR (intel#14865) replaced XFAIL approach with "not" usage in RUN lines Signed-off-by: Tikhomirova, Kseniya <kseniya.tikhomirova@intel.com>

This patch introduces `map_external_linear_memory` to enable mapping interop memory to linear USM, returning a `void *`. The PI function `piextMemMapExternalLinearMemory` has been added to enable this in the PI layer. A Vulkan test case has been added to test this functionality. --------- Co-authored-by: Duncan Brawley <duncan.brawley@codeplay.com>

Some files mistakenly had the wrong permissions. This removes all execute permissions from asciidoc, .cpp or .hpp files.

Add option "-fsycl-allow-device-dependencies" to enable support for dynamic linking. Also: 1. No functions are importable without "-fsycl-allow-device-dependencies" 2. Deal with SYCL_EXTERNAL header decls that have lost SYCL_EXTERNAL attribute in LLVM IR 3. SPIRV/SYCL/ESIMD builtins cannot be an imported function Tested in three E2E test cases. Minor change: Change SYCL-EXTERNAL to SYCL_EXTERNAL in testcase comment. --------- Signed-off-by: Lu, John <john.lu@intel.com> Co-authored-by: Marcos Maronas <marcos.maronas@intel.com>

…intel#14872) - fix for intel#14444 Signed-off-by: Neil R. Spruit <neil.r.spruit@intel.com>

Signed-off-by: Neil R. Spruit <neil.r.spruit@intel.com> Co-authored-by: Peter Žužek <peterzuzek@gmail.com>

`map.insert` doesn't insert values if the set already contains them. This can happen when UR/PI happens to reuse the same program pointer that it used for a previous program. -- This was causing some tests in the PI 2 UR conversion to randomly fail, including at least intel#14765 . Fixes intel#14819.

…ntel#14936) This would fix the post-commit fails for [14873](intel#14873)

This to align with what we had before PI removal. We need at least the /EHsc flag to be able to build UR in some Windows systems.

…l#14889)

Fix debug build issue [-Werror,-Wmicrosoft-include] llvm\include\sycl\handler.hpp:52:

intel#14819 has been fixed, so turning the tests back on.

See intel#14932 These tests were added 4 years ago with XFAIL and have never been enabled.

…#14875) Implementation design explaining those changes in a bigger picture can be found in intel#10540 Key things implemented here: - device code split to outline virtual functions into separate device images - emission of new properties for virtual functions - generation of `calls-indirectly` LLVM IR attribute for kernels that construct objects with virtual functions, but don't do calls - device image manipulations to cleanup or preserve virtual functions depending on a device image Even though those pieces are technically independent from each other, it is hard to split them apart into separate PRs, because they all have to be either present or absent for existing E2E tests for virtual functions to work.

…ion (intel#14827) This option allows to save generated SPIRV files in the specified directory.

Consists of two parts: 1. Fix the setting of `-Werror` on Windows in general for E2E tests * Just appending to `SYCL_E2E_CLANG_CXX_FLAGS` adds an extra semicolon that causes the command line to fail 3. Address remaining `-Werror` failures in Bindless Images E2E tests * Add return from function, even though there's an `assert(false)`

Those tests currently fail on AMD devices. To enable them, [sycl_ext_oneapi_free_function_kernels.asciidoc](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/proposed/sycl_ext_oneapi_free_function_kernels.asciidoc) needs to be investigated how to make it work for HIP backend.

This patch refactors the way we lower SYCL attributes and properties to NVVM annotations, through function attributes and metadata. It unifies the flow better with the SPIR-V paths at the same time. Previously we had: 1. Clang handling function attributes in two places: 1. CodeGenFunction did generic lowering of attributes to function metadata 2. NVPTXTargetCodeGenInfo did its own additional lowering of attributes to NVVM annotations 2. Kernel properties being handled in clang. NVPTXTargetCodeGenInfo lowered kernel properties, which had already been converted to function attributes, to NVVM annotations. 3. Kernel properties for HIP/CUDA *not* being turned into function metadata. Because function metadata is how the ComputeModuleRuntimeInfo library creates runtime info, this meant that target backends couldn't act on properties, as they were lost during lowering. Fundamentally, clang is not the correct place for handling SYCL kernel properties as it unnecessarily touches the front-end for what is really library code. With this patch, we now have: 1. Clang handles function attributes in one place, lowering to function metadata 2. SYCLLowerIR lowers kernel properties to the same function metadata: now working for HIP and CUDA kernels too. Because function metadata is consistently present for kernel properties, they behave as expected on HIP/CUDA targets now. 3. A new pass converts function metadata to NVVM annotations, replacing the code we've taken out of clang.

Some system headers included are not needed anymore.

For reference_argument.cpp, the test `RUN` line was wrong, and I added the real error message to the `CHECK` string. For `simd32_platform_error.cpp`, all Intel GPUs we test on report as support SIMD32 (for the ones that don't in hardware, it is probably emulated but it definitely works), so they don't throw this error so we can't test it, so just remove the test. Closes: intel#14643 Closes: intel#14644 Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>

The test was disabled on Linux but the error message shows it failed on Windows: ``` # | ...\llvm\install\bin\clang-offload-wrapper: error: '...\AppData\Local\Temp\lit-tmp-36f_juo7\ctor_load_usm_fp_extra-167ca4_esimd_14.sym': Operation did not complete successfully because the file contains a virus or potentially unwanted software. ``` `AppData` is a Windows folder and the linked PRs show it failing in the Windows job (example [here](https://github.com/intel/llvm/actions/runs/9994419232/job/27630212012)) The test isn't failing on Windows where the issue was reported to occur, so just remove the wrong `UNSUPPORTED` line. Closes: intel#14650 Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>

We were using the `x` prefix to the `RUN` lit command to comment out a `RUN` check. However, LIT still detects this as a `RUN` command, which illustrated in a CI fail of this test recently https://github.com/intel/llvm/actions/runs/10198164930/job/28215254220 Fixed by changing `xRUN` to `RUNx` which I've verified locally is not detected as a `RUN` command by LIT, and can be used to comment the `RUN` command out. Relates to intel#14473

This patch renames the directory (`sycl-fusion` to `sycl-jit`) as well as the involved library names (following the same principle: `fusion` to `jit`). To keep the user facing names consistent, buildbot switches were updated. And finally the codeowners file had to reflect the directory change.

Included PRs: * oneapi-src/unified-runtime#1917 * oneapi-src/unified-runtime#1906 * oneapi-src/unified-runtime#1896 * oneapi-src/unified-runtime#1810 * oneapi-src/unified-runtime#1812 * oneapi-src/unified-runtime#1929

@AlexeySachkov

4f86ab replaced `insert` with `[]` in order to fix a pointer re-use issue, however that was not the correct fix. Instead, a multimap was incorrectly converted to a regular map during the PI->UR conversion. This change reverts the rest of 4f86ab (besides the test, which is still valid), and converts NativePrograms back to a multimap. --- Thanks to @AlexeySachkov for finding the real issue in intel#14873 (comment)

Fix wrong link

e.g. ``` $ ninja check-sycl-dumps [...] # .---command stderr------------ # | 'FileCheck': command not found # `----------------------------- # error: command failed with exit status: 127 ```

- Move examples from spec to reduce bloat - Adds some usage guidelines for common scenarios - Remove reductions from examples since they are not supported

Bensuo · 2024-08-06T09:46:10Z

Upstream PR: intel#14965

Bensuo requested review from EwanC, fabiomestre, reble and julianmi July 29, 2024 17:06

reble reviewed Jul 29, 2024

View reviewed changes