brintos

brintos / llvm-project-archived public Read only

0
0
Text · 1.8 KiB · 0e059f1 Raw
59 lines · plain
1.. title:: clang-tidy - altera-single-work-item-barrier2 3altera-single-work-item-barrier4===============================5 6Finds OpenCL kernel functions that call a barrier function but do not call7an ID function (``get_local_id``, ``get_local_id``, ``get_group_id``, or8``get_local_linear_id``).9 10These kernels may be viable single work-item kernels, but will be forced to11execute as NDRange kernels if using a newer version of the Altera Offline12Compiler (>= v17.01).13 14If using an older version of the Altera Offline Compiler, these kernel15functions will be treated as single work-item kernels, which could be16inefficient or lead to errors if NDRange semantics were intended.17 18Based on the `Altera SDK for OpenCL: Best Practices Guide19<https://www.altera.com/en_US/pdfs/literature/hb/opencl-sdk/aocl_optimization_guide.pdf>`_.20 21Examples:22 23.. code-block:: c++24 25  // error: function calls barrier but does not call an ID function.26  void __kernel barrier_no_id(__global int * foo, int size) {27    for (int i = 0; i < 100; i++) {28      foo[i] += 5;29    }30    barrier(CLK_GLOBAL_MEM_FENCE);31  }32 33  // ok: function calls barrier and an ID function.34  void __kernel barrier_with_id(__global int * foo, int size) {35    for (int i = 0; i < 100; i++) {36      int tid = get_global_id(0);37      foo[tid] += 5;38    }39    barrier(CLK_GLOBAL_MEM_FENCE);40  }41 42  // ok with AOC Version 17.01: the reqd_work_group_size turns this into43  // an NDRange.44  __attribute__((reqd_work_group_size(2,2,2)))45  void __kernel barrier_with_id(__global int * foo, int size) {46    for (int i = 0; i < 100; i++) {47      foo[tid] += 5;48    }49    barrier(CLK_GLOBAL_MEM_FENCE);50  }51 52Options53-------54 55.. option:: AOCVersion56 57   Defines the version of the Altera Offline Compiler. Defaults to ``1600``58   (corresponding to version 16.00).59