Functions that perform parallel reductions on the GPU. More...

Functions
cudaReal	sum (DeviceArray< cudaReal > const &in)
	Compute sum of array elements (GPU kernel wrapper).

cudaReal	max (DeviceArray< cudaReal > const &in)
	Get maximum of array elements (GPU kernel wrapper).

cudaReal	maxAbs (DeviceArray< cudaReal > const &in)
	Get maximum absolute magnitude of array elements (GPU kernel wrapper).

cudaReal	min (DeviceArray< cudaReal > const &in)
	Get minimum of array elements (GPU kernel wrapper).

cudaReal	minAbs (DeviceArray< cudaReal > const &in)
	Get minimum absolute magnitude of array elements (GPU kernel wrapper).

cudaReal	innerProduct (DeviceArray< cudaReal > const &a, DeviceArray< cudaReal > const &b)
	Compute inner product of two real arrays (GPU kernel wrapper).

Detailed Description

Functions that perform parallel reductions on the GPU.

A reduction is any operation that involves reducing all of the elements of an array (or a set of multiple arrays) down to a single result. Examples include taking the sum or finding the maximum of all array elements. Highly efficient algorithms have been developed to perform such operations in parallel on a GPU, and those algorithms are implemented here.

A kernel wrapper is provided for each reduction operation, which takes a DeviceArray as an input and returns a single value. The wrappers are called on the host, and they call CUDA kernels internally to perform the reductions in parallel.

The CUDA kernels that perform the parallel reductions are not intended to be called directly. The wrappers and CUDA kernels are implemented together in a way that optimizes the speed of the overall operation. To do this, several important aspects of the algorithm are performed by the wrapper, detailed further below.

First, the CUDA kernels do not perform the entire reduction. Rather, each block of threads in the GPU is reduced to a single value, and the kernel returns an array containing one value for each thread block. The kernel wrapper then performs the remaining reduction by either calling the kernel again or performing the remaining reduction on the host, depending on the size of the array output by the first kernel. Once the array is smaller than 1e5 elements, the remaining reduction is performed on the CPU. (This value is chosen because the GPU does not significantly speed up reductions on arrays smaller than 1e5 elements.)

Second, the CUDA kernels are designed to use a number of threads that is only half the size of the input array (rounded up to the nearest multiple of the thread block size). The kernel wrapper properly determines the GPU configuration before calling the kernel.

Function Documentation

◆ sum()

cudaReal Pscf::Prdc::Cuda::Reduce::sum ( DeviceArray< cudaReal > const & in )

Compute sum of array elements (GPU kernel wrapper).

Parameters

in	input array

Definition at line 480 of file Reduce.cu.

References Pscf::DeviceArray< Data >::allocate(), Pscf::DeviceArray< Data >::capacity(), Pscf::DeviceArray< Data >::cArray(), Pscf::DeviceArray< Data >::deallocate(), Util::Log::file(), Pscf::DeviceArray< Data >::isAllocated(), Pscf::ThreadArray::setThreadsLogical(), Pscf::ThreadArray::setThreadsPerBlock(), sum(), UTIL_CHECK, and Pscf::ThreadArray::warpSize().

Referenced by Pscf::Rpg::FourthOrderParameter< D >::compute(), Pscf::Rpg::Solvent< D >::compute(), Pscf::Rpg::Simulator< D >::computeHamiltonian(), Pscf::Rpg::Block< D >::computeStress(), innerProduct(), Pscf::Rpg::ForceBiasMove< D >::move(), Pscf::Rpg::Mask< D >::rGridAverage(), and sum().

◆ max()

cudaReal Pscf::Prdc::Cuda::Reduce::max ( DeviceArray< cudaReal > const & in )

Get maximum of array elements (GPU kernel wrapper).

Parameters

in	input array

Definition at line 569 of file Reduce.cu.

References Pscf::DeviceArray< Data >::allocate(), Pscf::DeviceArray< Data >::capacity(), Pscf::DeviceArray< Data >::cArray(), Pscf::DeviceArray< Data >::deallocate(), Util::Log::file(), Pscf::DeviceArray< Data >::isAllocated(), max(), Pscf::ThreadArray::setThreadsLogical(), Pscf::ThreadArray::setThreadsPerBlock(), UTIL_CHECK, and Pscf::ThreadArray::warpSize().

Referenced by max(), and maxAbs().

◆ maxAbs()

cudaReal Pscf::Prdc::Cuda::Reduce::maxAbs ( DeviceArray< cudaReal > const & in )

Get maximum absolute magnitude of array elements (GPU kernel wrapper).

Parameters

in	input array

Definition at line 648 of file Reduce.cu.

References Pscf::DeviceArray< Data >::allocate(), Pscf::DeviceArray< Data >::capacity(), Pscf::DeviceArray< Data >::cArray(), Pscf::DeviceArray< Data >::deallocate(), Util::Log::file(), Pscf::DeviceArray< Data >::isAllocated(), max(), Pscf::ThreadArray::setThreadsLogical(), Pscf::ThreadArray::setThreadsPerBlock(), UTIL_CHECK, and Pscf::ThreadArray::warpSize().

Referenced by Pscf::Rpg::MaxOrderParameter< D >::compute().

◆ min()

cudaReal Pscf::Prdc::Cuda::Reduce::min ( DeviceArray< cudaReal > const & in )

Get minimum of array elements (GPU kernel wrapper).

Parameters

in	input array

Definition at line 730 of file Reduce.cu.

References Pscf::DeviceArray< Data >::allocate(), Pscf::DeviceArray< Data >::capacity(), Pscf::DeviceArray< Data >::cArray(), Pscf::DeviceArray< Data >::deallocate(), Util::Log::file(), Pscf::DeviceArray< Data >::isAllocated(), min(), Pscf::ThreadArray::setThreadsLogical(), Pscf::ThreadArray::setThreadsPerBlock(), UTIL_CHECK, and Pscf::ThreadArray::warpSize().

Referenced by min(), and minAbs().

◆ minAbs()

cudaReal Pscf::Prdc::Cuda::Reduce::minAbs ( DeviceArray< cudaReal > const & in )

Get minimum absolute magnitude of array elements (GPU kernel wrapper).

Parameters

in	input array

Definition at line 809 of file Reduce.cu.

References Pscf::DeviceArray< Data >::allocate(), Pscf::DeviceArray< Data >::capacity(), Pscf::DeviceArray< Data >::cArray(), Pscf::DeviceArray< Data >::deallocate(), Util::Log::file(), Pscf::DeviceArray< Data >::isAllocated(), min(), Pscf::ThreadArray::setThreadsLogical(), Pscf::ThreadArray::setThreadsPerBlock(), UTIL_CHECK, and Pscf::ThreadArray::warpSize().

◆ innerProduct()

cudaReal Pscf::Prdc::Cuda::Reduce::innerProduct	(	DeviceArray< cudaReal > const &	a,
		DeviceArray< cudaReal > const &	b )

Compute inner product of two real arrays (GPU kernel wrapper).

Parameters

a	first input array
b	second input array

Definition at line 891 of file Reduce.cu.

References Pscf::DeviceArray< Data >::allocate(), Pscf::DeviceArray< Data >::capacity(), Pscf::DeviceArray< Data >::cArray(), Pscf::DeviceArray< Data >::deallocate(), Util::Log::file(), Pscf::DeviceArray< Data >::isAllocated(), Pscf::ThreadArray::setThreadsLogical(), Pscf::ThreadArray::setThreadsPerBlock(), sum(), UTIL_CHECK, and Pscf::ThreadArray::warpSize().

Referenced by Pscf::Rpg::System< D >::computeFreeEnergy(), Pscf::Rpg::Simulator< D >::computeHamiltonian(), Pscf::Rpg::Propagator< D >::computeQ(), Pscf::Rpg::EinsteinCrystalPerturbation< D >::hamiltonian(), Pscf::Rpg::ExtGenFilm< D >::stressTerm(), and Pscf::Rpg::MaskGenFilm< D >::stressTerm().