libgomp/doc/amd-radeon-gcn.rst

   1 ..
   2   Copyright 1988-2022 Free Software Foundation, Inc.
   3   This is part of the GCC manual.
   4   For copying conditions, see the copyright.rst file.
   5
   6 .. _amd-radeon:
   7
   8 AMD Radeon (GCN)
   9 ****************
  10
  11 On the hardware side, there is the hierarchy (fine to coarse):
  12
  13 * work item (thread)
  14
  15 * wavefront
  16
  17 * work group
  18
  19 * compute unite (CU)
  20
  21 All OpenMP and OpenACC levels are used, i.e.
  22
  23 * OpenMP's simd and OpenACC's vector map to work items (thread)
  24
  25 * OpenMP's threads ('parallel') and OpenACC's workers map
  26         to wavefronts
  27
  28 * OpenMP's teams and OpenACC's gang use a threadpool with the
  29         size of the number of teams or gangs, respectively.
  30
  31 The used sizes are
  32
  33 * Number of teams is the specified ``num_teams`` (OpenMP) or
  34         ``num_gangs`` (OpenACC) or otherwise the number of CU
  35
  36 * Number of wavefronts is 4 for gfx900 and 16 otherwise;
  37         ``num_threads`` (OpenMP) and ``num_workers`` (OpenACC)
  38         overrides this if smaller.
  39
  40 * The wavefront has 102 scalars and 64 vectors
  41
  42 * Number of workitems is always 64
  43
  44 * The hardware permits maximally 40 workgroups/CU and
  45         16 wavefronts/workgroup up to a limit of 40 wavefronts in total per CU.
  46
  47 * 80 scalars registers and 24 vector registers in non-kernel functions
  48         (the chosen procedure-calling API).
  49
  50 * For the kernel itself: as many as register pressure demands (number of
  51         teams and number of threads, scaled down if registers are exhausted)
  52
  53 The implementation remark:
  54
  55 * I/O within OpenMP target regions and OpenACC parallel/kernels is supported
  56         using the C library ``printf`` functions and the Fortran
  57         ``print`` / ``write`` statements.