Tags: chengjunlu/llvm
Tags
[SYCL][Doc] Document global objects handling policy (intel#2636)
[Driver][SYCL] Fix processing for partial link step with static archi… …ves (intel#2661) When encountering specific linker specific arguments (-rpath) we do not pass them along to the partial link step. Improve the processing for these options for cases where -Xlinker <arg1> -Xlinker <arg2> is used.
[NFC][SYCL] Use visitors to print kernel name type (intel#2660) This is the part of integration header generator refactoring to make it use clang visitors and simplify the code.
[SYCL] Change in buffer management for integrated GPUs. (intel#2631) This change special-cases GPU devices with integrated memory. Buffers for these devices is allocated in host shared memory, improving performance of buffer data transfers. Signed-off-by: rdeodhar <rajiv.deodhar@intel.com>
[SYCL] Reinitialize moved containers (intel#2648) Signed-off-by: Sergey Kanaev <sergey.kanaev@intel.com>
[SYCL][ESIMD] Pass -vc-codegen instead of -cmc option to backend comp… …iler (intel#2642)
[SYCL] Fixed bug regarding device caching (intel#2566) Created a new function, populateDeviceCacheIfNeeded, so that cached devices could be shared across both piDevicesGet and piextDeviceCreateWithNativeHandle. This new function will check/fill and return cached devices. Renamed getOrCreatePlatforms to getPlatformCache, and refactored the function so that it limits level zero driver calls and removes redundant code. Called getOrMakePlatformImpl() from make_platform() so that the PlatformImpl cache is accessible to platforms that are created using a native handle. Called getOrMakeDeviceImpl() from make_device() so that the DeviceImpl cache is accessible to devices that are created using a native handle. Also, added an E2E test for these changes.
[SYCL] Don't block execution when flushing a stream (intel#2581) Currently stream flush is a blocking operation. This is incorrect, there are cases when kernels must be executed in parallel, for example, when 2 kernels are connected with FPGA pipe. This commit implements stream flush as a non-blocking operation by means of host_tasks.
PreviousNext