Section DBCSR

• Configuration options for the DBCSR library.

• Section path: CP2K_INPUT / GLOBAL / DBCSR

• This section cannot be repeated and can be optional.

• Subsections

  • none

• Keywords

  • COMBINED_TYPES
  • COMM_THREAD_LOAD
  • DETAILED_TIMING
  • KERNEL_TIMING
  • MM_DRIVER
  • MM_STACK_SIZE
  • MM_WORKSHARE
  • N_SIZE_MNK_STACKS
  • N_STACK_BUFFERS
  • N_STACK_MEMORY_REGIONS
  • SUBCOMMUNICATORS
  • USE_COMM_THREAD
  • USE_CUDA_HOST_ALLOCATION
  • USE_MPI_ALLOCATION

• Keyword descriptions

  COMBINED_TYPES
  	COMBINED_TYPES T
  	Combine matrix index and data into a single MPI derived data type for communication (avoid using).
  	This optional keyword cannot be repeated and it expects precisely one logical.
  	Default value: .FALSE.

  COMM_THREAD_LOAD
  	COMM_THREAD_LOAD 50
  	If a communications thread is used, specify how much multiplication workload (%) the thread should perform in addition to communication tasks
  	This optional keyword cannot be repeated and it expects precisely one integer.
  	Default value: 100

  DETAILED_TIMING
  	DETAILED_TIMING T
  	Output multiplication timing (development)
  	This optional keyword cannot be repeated and it expects precisely one logical.
  	Default value: .FALSE.

  KERNEL_TIMING
  	KERNEL_TIMING T
  	Time multiplication kernel (development)
  	This optional keyword cannot be repeated and it expects precisely one logical.
  	Default value: .FALSE.

  MM_DRIVER
  	MM_DRIVER BLAS
  	Select which routines to use for matrix block multiplications.
  	This optional keyword cannot be repeated and it expects precisely one keyword.
  	Default value: SMM
  	List of valid keywords: BLAS BLAS CUDA CUDA (not available) MATMUL Fortran MATMUL PLASMA PLASMA (not available) SMM libsmm: a library for small matrix multiplies (only BLAS wrapper available in this binary)

  MM_STACK_SIZE
  	MM_STACK_SIZE 1000
  	Size of multiplication parameter stack.
  	This optional keyword cannot be repeated and it expects precisely one integer.
  	Default value: 1000

  MM_WORKSHARE
  	MM_WORKSHARE T
  	Share MM work (e.g., multiple threads, CPU/GPU, ...).
  	This optional keyword cannot be repeated and it expects precisely one logical.
  	Default value: .FALSE.

  N_SIZE_MNK_STACKS
  	N_SIZE_MNK_STACKS 2
  	Number of stacks to use for distinct atomic sizes (e.g., 2 for a system of mostly waters).
  	This optional keyword cannot be repeated and it expects precisely one integer.
  	Default value: 0

  N_STACK_BUFFERS
  	N_STACK_BUFFERS 2
  	Number of stack buffers to use (e.g., 2 when using GPUs)
  	This optional keyword cannot be repeated and it expects precisely one integer.
  	Default value: 1

  N_STACK_MEMORY_REGIONS
  	N_STACK_MEMORY_REGIONS 2
  	Number of memory regions to use for stacks (e.g., number of GPU cards per MPI process)
  	This optional keyword cannot be repeated and it expects precisely one integer.
  	Default value: 1

  SUBCOMMUNICATORS
  	SUBCOMMUNICATORS T
  	Use MPI subcommunicators for transfers that are limited to process grid rows and columns.
  	This optional keyword cannot be repeated and it expects precisely one logical.
  	Default value: .FALSE.

  USE_COMM_THREAD
  	USE_COMM_THREAD T
  	During multiplication, use a thread to periodically poll MPI to progress outstanding message completions. This is beneficial on systems without a DMA-capable network adapter e.g. Cray XE6.
  	This optional keyword cannot be repeated and it expects precisely one logical.
  	Default value: .TRUE.

  USE_CUDA_HOST_ALLOCATION
  	USE_CUDA_ALLOCATION F
  	Allow use of CUDA-allocated host-pinned memory (avoid using). (not available)
  	This optional keyword cannot be repeated and it expects precisely one logical.
  	Default value: .FALSE.

  USE_MPI_ALLOCATION
  	USE_MPI_ALLOCATION T
  	Allow use of MPI-allocated memory for potentially faster network communication.
  	This optional keyword cannot be repeated and it expects precisely one logical.
  	Default value: .TRUE.