AtomicCas

Applicability

Function Usage

Performs an atomic comparison on the specified global memory address. If the current value at that address equals value1, it is replaced with value2. Otherwise, the value at the global memory address remains unchanged.

Prototype

1
2

template <typename T>
__aicore__ inline T AtomicCas(__gm__ T *address, T value1, T value2)

Parameters

Returns

Data before the atomic operation is performed on the global memory address.

Constraints

The atomic operation involves scalar computation. If the Scalar Unit and DMA unit (MTE2/MTE3) have data dependencies during global memory read and write, you need to insert synchronization as needed.

Example

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

// Pass the address of the global data to initialize dstGlobal and dstLocal.
dstGlobal.SetGlobalBuffer(reinterpret_cast<__gm__ T *>(dstGm), dataSize);

LocalTensor<T> dstLocal = inQueueX.AllocTensor<T>();
uint32_t value1 = 1;
uint32_t value2 = 2;
uint32_t a = AscendC::AtomicCas(reinterpret_cast<__gm__ uint32_t *>(dstGm), value1, value2);
// The movement operation can be performed only after the atomic operation is complete. There are data dependencies.
event_t eventIdSToMte2 = static_cast<event_t>(GetTPipePtr()->AllocEventID<HardEvent::S_MTE2>());
// Manually insert the MTE2 wait for Scalar Unit synchronization.
SetFlag<HardEvent::S_MTE2>(eventIdSToMte2);
WaitFlag<HardEvent::S_MTE2>(eventIdSToMte2);
DataCopy(dstLocal, dstGlobal, dataSize);
// ...

Assume that the preceding function is executed on three cores, and core 1, core 2, and core 3 are scheduled in sequence. The result example is as follows:

Original global memory data dst: [1,1,1,1,1,...,1] 
Core 1:
Global memory data dst after atomic computation: [2,1,1,1,1,...,1] 
Return value a: 1
Core 2:
Global memory data dst after atomic computation: [2,1,1,1,1,...,1] 
Return value a: 2
Core 3:
Global memory data dst after atomic computation: [2,1,1,1,1,...,1] 
Return value a: 2