Skip to content

Compiling for AMD with Open64

Hardware♯

Compilation Flags♯

  • consult AMD Open64 Documentation[2][3]
  • -march=bdver2[4][5]
    • specific flags for various CPU features may be used, instead, e.g. "-mfma4"
    • other possible flags:
      • -mso -- optimize for multicore scalability
      • -apo -- automatic parallelization

Math Libraries♯

  • Use the AMD Core Math Library:[6]
    • select an appropriate acml module -- for Proteus, use acml/open64/fma4
      • acml/open64/64 -- base ACML library, 32-bit integers
      • acml/open64/fma4 -- ACML with FMA4 (Fused Multiply Accumulate v. 4), 32-bit integers
      • acml/open64/mp/64 -- ACML with OpenMP[7]
      • acml/open64/mp/fma4 -- ACML with OpenMP + FMA4
      • acml/open64-int64/64 -- base ACML library, 64-bit integers
      • acml/open64-int64/fma4 -- ACML with FMA4, 64-bit integers
      • acml/open64-int64/mp/64 -- ACML with OpenMP, 64-bit integers
      • acml/open64-int64/mp/fma4 -- ACML with OpenMP + FMA4

MPI Implementations♯

See the Message Passing Interface article for implementations and versions available on Proteus.

MVAPICH2♯

Refer to official MVAPICH2 website for documentation.

Module:

proteus-mvapich2/open64

Environment variables:

MPI_HOME=/mnt/HA/opt/mvapich2/open64/64/1.9-mlnx-ofed MPI_LDFLAGS=-L/mnt/HA/opt/mvapich2/open64/64/1.9-mlnx-ofed/lib MPI_LIBDIR=/mnt/HA/opt/mvapich2/open64/64/1.9-mlnx-ofed/lib MPI_LIBS=-lmpich -lmpl -libmad -libumad -libverbs -ldl -lrt -lnuma -lrdmacm -lm -lpthread MPI_RUN=/mnt/HA/opt/mvapich2/open64/64/1.9-mlnx-ofed/bin/mpirun_rsh

Example makefile settings:

CPPFLAGS=-I$(MPI_HOME)/include LDFLAGS=$(MPI_LDFLAGS) $(MPI_LIBS)

OpenMPI♯

Module:

proteus-openmpi/open64

Guidance on Building Specific Applications♯

AMD has produced a series of guides on building some widely-used scientific software: Apps & Libraries Built by x86 Open64.

References♯

[1] New-Bulldozer-and-Piledriver-Instructions.pdf

[2] AMD x86 Open64 Compiler Suite web site

[3] CompilerOptQuickRef-63004300.pdf

[4] AMD Developer Central: Developer Guides & Manuals

[5] Phoronix - FX-8350 Piledriver Tuning On AMD's Open64 Compiler

[6] AMD Core Math Library website

[7] OpenMP® official website, 32-bit integers (NB OpenMP is an API for shared memory multiprocessing. It is distinct from OpenMPI, which is an implementation of the MPI-2 standard.)