SetFixPipeConfig

Applicability

Product	Supported
Atlas 350 Accelerator Card	√
Atlas A3 training product/Atlas A3 inference product	√
Atlas A2 training product/Atlas A2 inference product	√
Atlas 200I/500 A2 inference product	√
Atlas inference product AI Core	x
Atlas inference product Vector Core	x
Atlas training product	x

Function Usage

Sets the tensor quantization parameters in the real-time quantization during DataCopy (CO1 -> GM or CO1 -> A1).

Prototype

template <typename T>
__aicore__ inline void SetFixPipeConfig(const LocalTensor<T>& reluPre, const LocalTensor<T>& quantPre, bool isUnitFlag = false)
template <typename T, bool setRelu = false>
__aicore__ inline void SetFixPipeConfig(const LocalTensor<T>& preData, bool isUnitFlag = false)

Parameters

**Table 1** Template parameters
Parameter	Description
T	Operand data type.
setRelu	For one tensor, when setRelu is set to true, reluPre needs to be set. Otherwise, quantPre needs to be set. setRelu can only be set to false.

**Table 2** Parameters
Parameter	Input/Output	Description
reluPre	Input	Source operand—the tensor involved in the ReLU operation. Its type is LocalTensor, and the supported TPosition is C2PIPE2GM. reluPre is reserved for future use. You can pass an empty LocalTensor.
quantPre	Input	Source operand—the quantization tensor involved in the quantization operation. Its type is LocalTensor, and the supported TPosition is C2PIPE2GM.
isUnitFlag	Input	Whether to enable UnitFlag. The default value is false. false: Disabled true: Enabled
preData	Input	Used to set a tensor. A switch can be used to control whether the tensor is a ReLU tensor or quantization tensor. The supported TPosition is C2PIPE2GM. Only the quantization tensor is supported.

Restrictions

quantPre and reluPre must be tensors on the Fixpipe Buffer.

Returns

None

Examples

For details about the complete example, see Complete Example.

__aicore__inline void SetFPC(const LocalTensor <int32_t>& reluPreTensor, const LocalTensor <int32_t>& quantPreTensor)
{
 
    AscendC::LocalTensor<uint64_t> workA1 = inQueueDeqA1.AllocTensor<uint64_t>();
    uint16_t deqSize = 128; // Size of the dequantization tensor
    AscendC::DataCopy(workA1, deqGlobal, deqSize); // deqGlobal is the GM address of quantization coefficients.
    AscendC::LocalTensor<uint64_t> deqFB = inQueueDeqFB.AllocTensor<uint64_t>(); // Dequantization tensor address on the Fix.
    uint16_t fbufBurstLen = deqSize / 128;  // l1->fix, burst_len unit is 128Bytes
    AscendC::DataCopyParams dataCopyParams(1, fbufBurstLen, 0, 0);
    AscendC::DataCopy(deqFB, workA1, dataCopyParams); // Move the data to C2PIPE2GM through DataCopy.
    AscendC::SetFixPipeConfig(deqFB); // Set the quantization tensor.
    AscendC::PipeBarrier<PIPE_FIX>();
}

Parent topic: Data Movement