53fc17578f
- Use std::uint64_t instead of size_t to avoid implementation-defined type. - Rename to bst_idx_t, to account for other types of indexing. - Small cleanup to the base header.
SYCL-based Algorithm for Tree Construction
This plugin adds support of SYCL programming model for prediction algorithms to XGBoost.
Usage
Specify the 'device' parameter as described in the table below to offload model training and inference on SYCL device.
Algorithms
| device | Description |
|---|---|
| sycl | use default sycl device |
| sycl:gpu | use default sycl gpu |
| sycl:cpu | use default sycl cpu |
| sycl:gpu:N | use sycl gpu number N |
| sycl:cpu:N | use sycl cpu number N |
Python example:
param['device'] = 'sycl:gpu:0'
Note: 'sycl:cpu' devices have full functional support but can't provide good enough performance. We recommend use 'sycl:cpu' devices only for test purposes. Note: if device is specified to be 'sycl', device type will be automatically chosen. In case the system has both sycl GPU and sycl CPU, GPU will on use.
Dependencies
To build and use the plugin, install Intel® oneAPI DPC++/C++ Compiler. See also Intel® oneAPI Programming Guide.
Build
From the xgboost directory, run:
$ mkdir build
$ cd build
$ cmake .. -DPLUGIN_SYCL=ON
$ make -j