DPCT1086

Intel® DPC++ Compatibility Tool Developer Guide and Reference

Download PDF

ID 768918

Date 6/24/2024

Version

Public

A newer version of this document is available. Customers should click here to go to the newest version.

Visible to Intel only — GUID: GUID-16B4E3A5-D89B-4115-BB0F-81732B257F50

View Details

DPCT1086

Message

__activemask() is migrated to 0xffffffff. You may need to adjust the code.

Detailed Help

There is currently no functional equivalent of __activemask() in SYCL*. If there is flow control in your code that will make the thread inactive, you need to rewrite the thread logic.

For example, this original CUDA* code:

  __device__ inline int SHFL_SYNC(unsigned mask, int val, unsigned offset,
                                  unsigned w = warpSize) {
    return __shfl_down_sync(mask, val, offset, w);
  }

  __global__ void kernel(int *array) {
    unsigned int tid = threadIdx.x;
    if (tid >= 8)
      return;
    unsigned mask = __activemask();
    array[tid] = SHFL_SYNC(mask, array[tid], 4);
  }

results in the following migrated SYCL code:

  inline int SHFL_SYNC(unsigned mask, int val, unsigned offset,
                       const sycl::nd_item<3> &item_ct1, unsigned w = 0) {
    /*
    DPCT1023:0: The SYCL sub-group does not support mask options for
    dpct::shift_sub_group_left. You can specify
    "--use-experimental-features=masked-sub-group-operation" to use the
    experimental helper function to migrate __shfl_down_sync.
    */
    if (!w) w = item_ct1.get_sub_group().get_local_range().get(0);
    // This call will wait for all work-items to arrive which will never happen since only work-items with tid < 8 will encounter this call.
    return dpct::shift_sub_group_left(item_ct1.get_sub_group(), val, offset, w);
  }

  void kernel(int *array, const sycl::nd_item<3> &item_ct1) {
    unsigned int tid = item_ct1.get_local_id(2);
    if (tid >= 8)
      return;
    /*
    DPCT1086:1: __activemask() is migrated to 0xffffffff. You may need to adjust
    the code.
    */
    unsigned mask = 0xffffffff;
    array[tid] = SHFL_SYNC(mask, array[tid], 4, item_ct1);
  }

which is rewritten to:

  // remove mask parameter, as it is not used
  inline int SHFL_SYNC(int val, unsigned offset,
                       const sycl::nd_item<3> &item_ct1, unsigned w = 0) {
    if (!w) w = item_ct1.get_sub_group().get_local_range().get(0);
    unsigned int tid = item_ct1.get_local_id(2);
    // Use a temporary variable to save the result of sycl::shift_group_left() to make sure all work-items can encounter this call.
    int v_tmp = sycl::shift_group_left(item_ct1.get_sub_group(), val, offset);
    return (tid < 8) ? v_tmp : val;
  }

  void kernel(int *array, const sycl::nd_item<3> &item_ct1) {
    unsigned int tid = item_ct1.get_local_id(2);
    // remove mask parameter, as it is not used
    array[tid] = SHFL_SYNC(array[tid], 4, item_ct1);
  }

Suggestions to Fix

Check if 0xffffffff can be used instead of __activemask(). If it cannot be used, redesign the thread logic.

Select Your Language

Using Intel.com Search

Quick Links

Recent Searches

Advanced Search

Only search in

Intel® DPC++ Compatibility Tool Developer Guide and Reference

DPCT1086

Message

Detailed Help

Suggestions to Fix