BlockReduceMax

Applicability

Product	Supported
Atlas A3 training products / Atlas A3 inference products	√
Atlas A2 training products / Atlas A2 inference products	√
Atlas 200I/500 A2 inference products	√
Atlas inference product 's AI Core	√
Atlas inference product 's Vector Core	x
Atlas training products	√

Function

Computes the maximum of all elements in each data block. For details about reduction instructions, see How to Use Reduction Compute APIs.

Prototype

Bitwise mask mode

        
             template <typename T, bool isSetMask = true>
__aicore__ inline void BlockReduceMax(const LocalTensor<T>& dst, const LocalTensor<T>& src,const int32_t repeatTime, const uint64_t mask[], const int32_t dstRepStride, const int32_t srcBlkStride, const int32_t srcRepStride)

Contiguous mask mode

        
             template <typename T, bool isSetMask = true>
__aicore__ inline void BlockReduceMax(const LocalTensor<T>& dst, const LocalTensor<T>& src,const int32_t repeatTime, const int32_t mask, const int32_t dstRepStride, const int32_t srcBlkStride, const int32_t srcRepStride)

Parameters

**Table 1** Template parameters
Parameter	Description
T	Operand data type. For the Atlas A3 training products / Atlas A3 inference products , the supported data types are half and float. For the Atlas A2 training products / Atlas A2 inference products , the supported data types are half and float. For Atlas 200I/500 A2 inference products , the supported data types are half and float. For the Atlas inference product 's AI Core, the supported data types are half and float. For the Atlas training products , the supported data type is half.
isSetMask	Indicates whether to set mask inside the API. true: sets mask inside the API. false: sets mask outside the API. Developers need to use the SetVectorMask API to set the mask value. In this mode, the mask value in the input parameter of this API must be set to the placeholder MASK_PLACEHOLDER.

**Table 2** Parameters
Parameter	Input/Output	Description
dst	Output	Destination operand. The type is LocalTensor, and the supported TPosition is VECIN, VECCALC, or VECOUT. The start address of the LocalTensor must be 16-byte aligned (for data of the half type) or 32-byte aligned (for data of the float type).
src	Input	Source operand. The type is LocalTensor, and the supported TPosition is VECIN, VECCALC, or VECOUT. The start address of the LocalTensor must be 32-byte aligned.
repeatTime	Input	Number of iteration repeats. The value range is [0, 255]. For details about this parameter, see High-dimensional Sharding APIs.
mask/mask[]	Input	mask is used to control the elements that participate in computation in each iteration. Bitwise mode: controls which elements are involved in computation bit by bit. A bit value of 1 means the corresponding element participates in computation, while 0 means it does not. The mask value is an array. The array length and the value range of the array elements are related to the operand data type. When the operand is 16-bit, the array length is 2, mask[0] and mask[1] ∈ [0, 2⁶⁴ -1] and cannot be 0 at the same time. When the operand is 32-bit, the array length is 1 and mask[0] ∈ (0, 2⁶⁴ – 1]. When the operand is 64-bit, the array length is 1 and mask[0] ∈ (0, 2³² – 1]. For example, if mask = [0, 8] and 8 = 0b1000, only the fourth element participates in computation. Contiguous mode: indicates the number of contiguous elements that participate in computation. The value range is related to the operand data type. The maximum number of elements that can be processed in each repeat varies according to the data type. When the operand is 16-bit, mask ∈ [1, 128]. When the operand is 32-bit, mask ∈ [1, 64]. When the operand is 64-bit, mask ∈ [1, 32].
dstRepStride	Input	Address stride between adjacent iterations of the destination operand. The unit is the length of one reduced repeat. After each repeat (eight data blocks) is reduced, eight elements are obtained. Therefore, when the input type is half, the unit of RepStride is 16 bytes. When the input type is float, the unit of RepStride is 32 bytes. Note that this parameter cannot be set to 0 for the Atlas training products .
srcBlkStride	Input	Address stride of data blocks in a single iteration. For details, see dataBlockStride.
srcRepStride	Input	Address stride between adjacent iterations of the source operand, that is, the number of data blocks skipped of the source operand in each iteration. For details, see repeatStride.

Returns

None

Restrictions

For details about the operand address alignment requirements, see General Address Alignment Restrictions.

To save memory space, you can define a tensor shared by the source and destination operands (by address overlapping). Note that the computed destination operand data cannot overwrite the source operands that are not involved in the computation. Exercise caution when defining the tensor.
For the Atlas 200I/500 A2 inference products , if mask/mask[] is configured and no element in a data block is involved in the computation, the maximum value of all elements in the data block is filled with -inf and returned. For example, in the float scenario, if mask is set to 32, that is, only the first four data blocks are computed, -inf is returned for the maximum value in the last four data blocks. In the half scenario, -65504 is returned.
Proper use of reduction instructions in different scenarios can improve performance. For details, see Selecting Low-Latency Instructions to Optimize Reduction Operation Performance. For details about examples, see ReduceCustom.

Example

This example shows only part of the code used in the computation process (Compute). To run the sample code, copy the code snippet and replace parts of code of the Compute function in Template Samples.

Example of BlockReduceMax – high-dimensional tensor sharding computation (contiguous mask mode)

        
             int32_t mask = 256/sizeof(half);
int repeat = 1;
// repeat = 1, 128 elements one repeat, 128 elements total
// srcBlkStride = 1, no gap between blocks in one repeat
// dstRepStride = 1, srcRepStride = 8, no gap between repeats
AscendC::BlockReduceMax<half>(dstLocal, srcLocal, repeat, mask, 1, 1, 8);

Example of BlockReduceMax – high-dimensional tensor sharding computation (bitwise mask mode)

        
             uint64_t mask[2] = { UINT64_MAX, UINT64_MAX };
int repeat = 1;
// repeat = 1, 128 elements one repeat, 128 elements total
// srcBlkStride = 1, no gap between blocks in one repeat
// dstRepStride = 1, srcRepStride = 8, no gap between repeats
AscendC::BlockReduceMax<half>(dstLocal, srcLocal, repeat, mask, 1, 1, 8);

Result example:

Input (src_gm):
[-8.781, 4.688, -0.09607, -5.445, 4.957, -4.832, 9.555, 8.391, 
 6.273, -2.412, 7.969, 3.9, -0.4238, 2.988, -6.855, -1.335, 
 ...
 9.68, -6.672, -6.488, -7.398, 8.562, 3.508, 3.135, -5.512, 
 -7.883, -8.594, -5.895, -8.938, -7.676, -7.867, -9.188, -5.715]  

Output (dst_gm):
[9.555, ..., 9.68, 0, ... 0]

Template Sample

      
       
         
         
           #include "kernel_operator.h"
class KernelReduce {
public:
    __aicore__ inline KernelReduce() {}
    __aicore__ inline void Init(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
    {
        srcGlobal.SetGlobalBuffer((__gm__ half*)src);
        dstGlobal.SetGlobalBuffer((__gm__ half*)dstGm);
        pipe.InitBuffer(inQueueSrc, 1, srcDataSize * sizeof(half));
        pipe.InitBuffer(outQueueDst, 1, dstDataSize * sizeof(half));
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        Compute();
        CopyOut();
    }
private:
    __aicore__ inline void CopyIn()
    {
        AscendC::LocalTensor<half> srcLocal = inQueueSrc.AllocTensor<half>();
        AscendC::DataCopy(srcLocal, srcGlobal, srcDataSize);
        inQueueSrc.EnQue(srcLocal);
    }
    __aicore__ inline void Compute()
    {
        AscendC::LocalTensor<half> srcLocal = inQueueSrc.DeQue<half>();
        AscendC::LocalTensor<half> dstLocal = outQueueDst.AllocTensor<half>();
        half zero(0);
        AscendC::Duplicate(dstLocal, zero, dstDataSize);
        // Command execution part (replace with the preceding code)
        outQueueDst.EnQue<half>(dstLocal);
        inQueueSrc.FreeTensor(srcLocal);
    }
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<half> dstLocal = outQueueDst.DeQue<half>();
        AscendC::DataCopy(dstGlobal, dstLocal, dstDataSize);
        outQueueDst.FreeTensor(dstLocal);
    }
private:
    AscendC::TPipe pipe;
    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueueSrc;
    AscendC::TQue<AscendC::TPosition::VECOUT, 1> outQueueDst;
    AscendC::GlobalTensor<half> srcGlobal, dstGlobal;
    int srcDataSize = 128;
    int dstDataSize = 64;
};
extern "C" __global__ __aicore__ void reduce_simple_kernel(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
{
    KernelReduce op;
    op.Init(src, dstGm);
    op.Process();
}

          

        

      
     

Parent topic: Reduction Computation