Module sync 

Synchronization and atomic operations for GPU programming.

Structs

AddF

Atomic operation kind for Atomic<'a, T>::atomic_addf

AddI

Atomic operation kind for Atomic<'a, T>::atomic_addi

AndI

Atomic operation kind for Atomic<'a, T>::atomic_andi

Assign

Atomic operation kind for Atomic<'a, T>::atomic_assign

Atomic

To prevent the user from mixing use of Atomic operation and chunk-based access, we wrap the reference to the data to be modified in an Atomic struct. ensuring that the user cannot access the data without using atomic operations. If user wants to repurpose the data for non-atomic access, they need to drop the Atomic struct first and need a sync across all blocks. For now, sync across all blocks is not supported, we should reject such code in analysis. TODO: Avoid repurposing data for atomic or chunk-based access.

MaxNumF

Atomic operation kind for Atomic<'a, T>::atomic_maxnumf

MaxS

Atomic operation kind for Atomic<'a, T>::atomic_maxs

MaxU

Atomic operation kind for Atomic<'a, T>::atomic_maxu

MaximumF

Atomic operation kind for Atomic<'a, T>::atomic_maximumf

MinNumF

Atomic operation kind for Atomic<'a, T>::atomic_minnumf

MinS

Atomic operation kind for Atomic<'a, T>::atomic_mins

MinU

Atomic operation kind for Atomic<'a, T>::atomic_minu

MinimumF

Atomic operation kind for Atomic<'a, T>::atomic_minimumf

MulF

Atomic operation kind for Atomic<'a, T>::atomic_mulf

MulI

Atomic operation kind for Atomic<'a, T>::atomic_muli

OrI

Atomic operation kind for Atomic<'a, T>::atomic_ori

SharedAtomic

Functions

sync_threads

Synchronization within a thread block. Disallow divergent control flow to ensure all threads in a block can reach the sync point to avoid deadlock.