BlockReduceSum

Applicability

Product	Supported/Unsupported
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	√

Functions

Sums all elements in each data block. Source operands are added in binary tree mode. For details about reduction instructions, see How to Use Reduction Compute APIs.

To sum up 128 data elements of the half type, sum up 8 data blocks, with each data block containing 16 data elements of the half type. In each data block, data is added in binary tree mode. The following figure shows the schematic diagram of BlockReduceSum.

Figure 1 BlockReduceSum schematic diagram

When being greater than 65504, the computation result is truncated to 65504. For example, the source operand is [60000, 60000, –30000, 100], 60000 + 60000 > 65504, meaning that the result overflows. In this case, the maximum value 65504 will be used as the result. Similarly, –30000 + 100 = –29900, 65504 – 29900 = 35604. The following figure shows the computation.

Figure 2 Computation in the overflow scenario

Prototype

Mask bit-by-bit mode

        
             template <typename T, bool isSetMask = true>
__aicore__ inline void BlockReduceSum(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 BlockReduceSum(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** Parameters in the template
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 the 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	The mask parameter is used to control the elements involved in computation in each iteration. Bitwise mode: controls the elements that participate in computation by bit. If a bit is set to 1, the corresponding element participates in the computation. If a bit is set to 0, the corresponding element is masked in the computation. The mask is in array form. The array length and the value range of the array elements are related to the data type of the operand. When the operand is 16-bit, the array length is 2. In this case, mask[0] and mask[1] must be in the range of [0, 2⁶⁴ – 1] and cannot be 0 at the same time. When the operand is 32-bit, the array length is 1. In this case, mask[0] must be in the range of (0, 2⁶⁴ – 1]. When the operand is 64-bit, the array length is 1. In this case, mask[0] must be in the range of (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 stride is measured in the length of the result after one repeatTime reduction. 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 Product , if mask/mask[] is configured and no element in a data block is involved in the computation, the sum of all elements in the data block is filled with 0 to return. For example, in the float scenario, if mask is set to 32, that is, only the first four data blocks are computed, 0 is returned for the sum of the last four data blocks.

Examples

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 BlockReduceSum – 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::BlockReduceSum<half>(dstLocal, srcLocal, repeat, mask, 1, 1, 8);

Example of BlockReduceSum – 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::BlockReduceSum<half>(dstLocal, srcLocal, repeat, mask, 1, 1, 8);

Result example:

Input (src_gm):
[-7.289, 4.48, -5.898, -6.199, 1.422, -6.168, -3.178, -1.198, 
 7.789, 6.754, -5.191, -0.6797, 2.883, 2.08, 8.664, -8.539,
 ...,
 -7.625, 2.529, 7.855, -2.012, -6.52, -6.652, -8.422, -9.914,
 -4.355, 1.849, 5.406, 1.483, -6.074, -1.897, 8.625, 1.969]  
Output (dst_gm):
[-10.27, ..., -23.77, 0, ..., 0]

Proper use of the reduction instruction in different scenarios can improve performance. For details about the introduction, see Using the Reduction Instruction Properly in Different Scenarios. For details about examples, see ReduceCustom.

Parent topic: Reduction computation