CUDA Fortran for Scientists and Engineers

Index

A

allocatable variable attribute, see Attributes

Assigning devices to MPI ranks, see Message Passing Interface

Asynchronous data transfers, 52–60

Hyper-Q, 59

pinned memory, 52

profiling, 60

streams, 52

Atomics

atomicCAS, 173

reduction with atomic locks, 173

attributes(), 7

Attributes

variable attributes

allocatable, 10

with pinned memory, 45

with multi-GPU code, 116

constant, 83

device, 7

pinned, 46

shared, 86

texture, 76

value, 8

function attributes

device, 28

global, 7–8

B

Bandwidth

device

effective, 41

theoretical peak, 39–41

throughput vs. bandwidth, 42

between host and device, 44

asynchronous transfers, 52–60

batching transfers, 49

cudaMemcpy(), 50

pinned memory, 45

BandwidthTest code, 49

Bank conflicts, see Shared memory

bind(), see ...