文档

CastDeq

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

x

Atlas inference product 's AI Core

√

Atlas inference product 's Vector Core

x

Atlas training products

x

Functions

Quantizes the input and converts the precision. The conversion formula varies according to the data type.

  • When the input type is int16_t, the int16_t input is quantized and the precision is converted to obtain the int8_t/uint8_t data. Before using this API, you need to call SetDeqScale to set quantization parameters such as scale, offset, and signMode.

    The template parameter isVecDeq determines whether to select the vector quantization mode.

    • Quantizes the input and converts the precision based on scale, offset, and signMode set by SetDeqScale when isVecDeq is set to false. The formula is as follows.

    • Quantizes the input in loop mode and converts the precision based on the 16 groups of quantization parameters scale0–scale15, offset0–offset15, and signMode0–signMode15 set on a 128-byte UB segment by SetDeqScale when isVecDeq is set to true. The formula is as follows.

  • When the input type is int32_t, the int32_t input is quantized and the precision is converted to obtain the half data. Before using this API, you need to call SetDeqScale to set the scale parameter.

    .

Prototype

  • Computation of the first n pieces of data of a tensor 
    1
2
    		 
    template <typename T, typename U, bool isVecDeq = true, bool halfBlock = true>
__aicore__ inline void CastDeq(const LocalTensor<T>& dst, const LocalTensor<U>& src, const uint32_t count)
  • High-dimensional tensor sharding computation
    • Bitwise mask mode 
      1
2
      		 
      template <typename T, typename U, bool isSetMask = true, bool isVecDeq = true, bool halfBlock = true>
__aicore__ inline void CastDeq(const LocalTensor<T>& dst, const LocalTensor<U>& src, const uint64_t mask[], uint8_t repeatTime, const UnaryRepeatParams& repeatParams)
    • Contiguous mask mode 
      1
2
      		 
      template <typename T, typename U, bool isSetMask = true, bool isVecDeq = true, bool halfBlock = true>
__aicore__ inline void CastDeq(const LocalTensor<T>& dst, const LocalTensor<U>& src, const int32_t mask, uint8_t repeatTime, const UnaryRepeatParams& repeatParams)

Parameters

Table 1 Parameters in the template

Parameter

Description

T

Data type of the output Tensor.

Atlas A3 training products / Atlas A3 inference products : The supported data types are int8_t, uint8_t, and half.

Atlas A2 training products / Atlas A2 inference products : The supported data types are int8_t, uint8_t, and half.

For the Atlas inference product 's AI Core, the supported data types are int8_t and uint8_t.

This parameter is used together with the input parameter signMode of the SetDeqScale API. When signMode is set to true, the output data type is int8_t. When signMode is set to false, the output data type is uint8_t.

U

Data type of the input Tensor.

For the Atlas A3 training products / Atlas A3 inference products , the supported data types are int16_t and int32_t.

For the Atlas A2 training products / Atlas A2 inference products , the supported data types are int16_t and int32_t.

For the Atlas inference product 's AI Core, the supported data type is int16_t.

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.

isVecDeq

Controls whether to select the vector quantization mode. This parameter is used together with SetDeqScale(const LocalTensor<T>& src). When a tensor is passed through SetDeqScale, isVecDeq must be true.

halfBlock

When the int16_t input is quantized and converted to the int8_t or uint8_t data type, the halfBlock parameter is used to indicate whether the output element is stored in the upper or lower block. If true, the result is stored in the lower half block, and if false, the result is stored in the upper half block, as shown in Figure 1.
Table 2 API 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 32-byte aligned.

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.

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].

When the number of bits of the source operand is different from that of the destination operand, the data type with more bytes is used for the computation. For example, if the source operand is of the int16_t type and the destination operand is of the int8_t type, the int16_t type is used for mask calculation.

repeatTime

Input

Number of iteration repeats. The Vector Unit reads 256 bytes of contiguous data for computation each time. To read the complete data for processing, the unit needs to read the input data in multiple repeats. repeatTime indicates the number of repeats.

For details about this parameter, see High-dimensional Sharding APIs.

repeatParams

Input

Parameters that control the operand address strides. They are of the UnaryRepeatParams type, and contain such parameters as those that specify the address stride of the operand for the same data block between adjacent iterations and address stride of the operand between different data blocks in a single iteration.

For details about the address stride parameters between adjacent iterations, see repeatStride. For details about the address stride parameters of DataBlock in the same iteration, see dataBlockStride.

count

Input

Number of elements involved in the computation.
Figure 1 Description of halfBlock

Returns

None

Restrictions

Examples

To run the sample code, copy the code snippet and replace some code of the Compute function in Template Sample.

  • Example of high-dimensional sharding computation API - contiguous mask mode 
    1
2
3
4
5
    		 
    int32_t mask = 256 / sizeof(int16_t);
// repeatTime = 2, 128 elements one repeat, 256 elements total
// dstBlkStride, srcBlkStride = 1, no gap between blocks in one repeat
// dstRepStride, srcRepStride = 8, no gap between repeats
AscendC::CastDeq<uint8_t, int16_t, true, true, true>(dstLocal, srcLocal, mask, 2, { 1, 1, 8, 8 });
  • Example of high-dimensional sharding computation API - bitwise mask mode 
    1
2
3
4
5
    		 
    uint64_t mask[2] = { UINT64_MAX, UINT64_MAX };
// repeatTime = 2, 128 elements one repeat, 256 elements total
// dstBlkStride, srcBlkStride = 1, no gap between blocks in one repeat
// dstRepStride, srcRepStride = 8, no gap between repeats
AscendC::CastDeq<uint8_t, int16_t, true, true, true>(dstLocal, srcLocal, mask, 2, { 1, 1, 8, 8 });
  • Example of first n pieces of tensor data computation API 
    1
    		 
    AscendC::CastDeq<uint8_t, int16_t, true, true>(dstLocal, srcLocal, 256);

Result example:

Input data (srcLocal):
[20 53 26 12 36  6 20 93 66 30 56 99 59 92  7 37 22 47 98 10 85 29 14 46
 17 34 45 17 25 45 82 17 66 94 68 23 67  8 89  8 92  6 10 80 87 20  9 81
 70 62 11 58 38 83 32 14 38 47 41 63 94 26 96 89 88 35 86 55 60 82 15 65
 92 67 83 23 63 25 85 93 50 91 75 60 80 10 55 20 71 14 67 23 31 63  7 93
 69 45 61 23 43 86 11 81 81 36 76 58 53 25 23 51 59 78 82 10 39 40 24 50
 68 49 79 40  4 53 22 38 45 17 29 54  9 66 98 47 12 47 47 20 98  0 59 77
  1 21 39 70 66 20 68  8 77 77 54  0  3 33 37 37 48 60 83 88 27 70 31 49
 75 21 59  3 99 84 92 84 14 44 26 56 72 56 37 52 39 11  2 59 59 65 71 64
 10 65 62 48 42 79 69 69 27 99  8 38 36 77 34 34 60 50 52 50 41 31 95 68
 27 16 42 64 19 47  0 10 36 36 33 62 98 64 32 81 49 53 27 70 35  9 63  7
 10 89  3 39 94 23 89 16 23 60 71 42 46 58 65 90]
Output (dstLocal):
[ 0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 20 53 26 12 36  6 20 93
 66 30 56 99 59 92  7 37  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 22 47 98 10 85 29 14 46 17 34 45 17 25 45 82 17  0  0  0  0  0  0  0  0
  0  0  0  0  0  0  0  0 66 94 68 23 67  8 89  8 92  6 10 80 87 20  9 81
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 70 62 11 58 38 83 32 14
 38 47 41 63 94 26 96 89  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 88 35 86 55 60 82 15 65 92 67 83 23 63 25 85 93  0  0  0  0  0  0  0  0
  0  0  0  0  0  0  0  0 50 91 75 60 80 10 55 20 71 14 67 23 31 63  7 93
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 69 45 61 23 43 86 11 81
 81 36 76 58 53 25 23 51  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 59 78 82 10 39 40 24 50 68 49 79 40  4 53 22 38  0  0  0  0  0  0  0  0
  0  0  0  0  0  0  0  0 45 17 29 54  9 66 98 47 12 47 47 20 98  0 59 77
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  1 21 39 70 66 20 68  8
 77 77 54  0  3 33 37 37  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 48 60 83 88 27 70 31 49 75 21 59  3 99 84 92 84  0  0  0  0  0  0  0  0
  0  0  0  0  0  0  0  0 14 44 26 56 72 56 37 52 39 11  2 59 59 65 71 64
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10 65 62 48 42 79 69 69
 27 99  8 38 36 77 34 34  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0
 60 50 52 50 41 31 95 68 27 16 42 64 19 47  0 10  0  0  0  0  0  0  0  0
  0  0  0  0  0  0  0  0 36 36 33 62 98 64 32 81 49 53 27 70 35  9 63  7
  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0  0 10 89  3 39 94 23 89 16
 23 60 71 42 46 58 65 90]

Template Sample

This section provides a template sample to help you quickly run reference instruction samples. 
  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
 
#include "kernel_operator.h"
template <typename srcType, typename dstType>
class KernelCastDeq {
public:
    __aicore__ inline KernelCastDeq() {}
    __aicore__ inline void Init(GM_ADDR src_gm, GM_ADDR dst_gm, uint32_t inputSize, bool halfBlock, bool isVecDeq)
    {
        srcSize = inputSize;
        dstSize = inputSize * 2;
        this->halfBlock = halfBlock;
        this->isVecDeq = isVecDeq;
        src_global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType*>(src_gm), srcSize);
        dst_global.SetGlobalBuffer(reinterpret_cast<__gm__ dstType*>(dst_gm), dstSize);
        pipe.InitBuffer(inQueueX, 1, srcSize * sizeof(srcType));
        pipe.InitBuffer(outQueue, 1, dstSize * sizeof(dstType));
        pipe.InitBuffer(tmpQueue, 1, 128);
    }
    __aicore__ inline void Process()
    {
        CopyIn();
        Compute();
        CopyOut();
    }
private:
    __aicore__ inline void CopyIn()
    {
        AscendC::LocalTensor<srcType> srcLocal = inQueueX.AllocTensor<srcType>();
        AscendC::DataCopy(srcLocal, src_global, srcSize);
        inQueueX.EnQue(srcLocal);
    }
    __aicore__ inline void Compute()
    {
        AscendC::LocalTensor<dstType> dstLocal = outQueue.AllocTensor<dstType>();
        AscendC::LocalTensor<uint64_t> tmpBuffer = tmpQueue.AllocTensor<uint64_t>();
        AscendC::Duplicate(tmpBuffer.ReinterpretCast<int32_t>(), static_cast<int32_t>(0), 32);
        AscendC::PipeBarrier<PIPE_V>();
        AscendC::Duplicate<int32_t>(dstLocal.template ReinterpretCast<int32_t>(), static_cast<int32_t>(0), dstSize / sizeof(int32_t));
        AscendC::PipeBarrier<PIPE_ALL>();
        bool signMode = false;
        if constexpr (AscendC::Std::is_same<dstType, int8_t>::value) {
            signMode = true;
        }
        AscendC::LocalTensor<srcType> srcLocal = inQueueX.DeQue<srcType>();
        if (halfBlock) {
            if (isVecDeq) {
                float vdeqScale[16] = { 1.0 };
                int16_t vdeqOffset[16] = { 0 };
                bool vdeqSignMode[16] = { signMode };
                AscendC::VdeqInfo vdeqInfo(vdeqScale, vdeqOffset, vdeqSignMode);
                AscendC::SetDeqScale(tmpBuffer, vdeqInfo);
                AscendC::CastDeq<dstType, srcType, true, true>(dstLocal, srcLocal, srcSize);
            } else {
                float scale = 1.0;
                int16_t offset = 0;
                AscendC::SetDeqScale(scale, offset, signMode);
                AscendC::CastDeq<dstType, srcType, false, true>(dstLocal, srcLocal, srcSize);
            }
        } else {
            if (isVecDeq) {
                float vdeqScale[16] = { 1.0 };
                int16_t vdeqOffset[16] = { 0 };
                bool vdeqSignMode[16] = { signMode };
                AscendC::VdeqInfo vdeqInfo(vdeqScale, vdeqOffset, vdeqSignMode);
                AscendC::SetDeqScale(tmpBuffer, vdeqInfo);
                AscendC::CastDeq<dstType, srcType, true, false>(dstLocal, srcLocal, srcSize);
            } else {
                float scale = 1.0;
                int16_t offset = 0;
                AscendC::SetDeqScale(scale, offset, signMode);
                AscendC::CastDeq<dstType, srcType, false, false>(dstLocal, srcLocal, srcSize);
            }
        }
        outQueue.EnQue<dstType>(dstLocal);
        tmpQueue.FreeTensor(tmpBuffer);
        inQueueX.FreeTensor(srcLocal);
    }
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<dstType> dstLocal = outQueue.DeQue<dstType>();
        AscendC::DataCopy(dst_global, dstLocal, dstSize);
        outQueue.FreeTensor(dstLocal);
    }
private:
    AscendC::GlobalTensor<srcType> src_global;
    AscendC::GlobalTensor<dstType> dst_global;
    AscendC::TPipe pipe;
    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueueX;
    AscendC::TQue<AscendC::TPosition::VECIN, 1> tmpQueue;
    AscendC::TQue<AscendC::TPosition::VECOUT, 1> outQueue;
    bool halfBlock = false;
    bool isVecDeq = false;
    uint32_t srcSize = 0;
    uint32_t dstSize = 0;
};
template <typename srcType, typename dstType>
__aicore__ void kernel_cast_deqscale_operator(GM_ADDR src_gm, GM_ADDR dst_gm, uint32_t dataSize, bool halfBlock, bool isVecDeq)
{
    KernelCastDeq<srcType, dstType> op;
    op.Init(src_gm, dst_gm, dataSize, halfBlock, isVecDeq);
    op.Process();
}
extern "C" __global__ __aicore__ void kernel_cast_deqscale_operator_256_int16_t_uint8_t_true_true(GM_ADDR src_gm, GM_ADDR dst_gm)
{
    kernel_cast_deqscale_operator<int16_t, uint8_t>(src_gm, dst_gm, 256, true, true);
}
Parent topic: Complex computation
我们使用cookie来确保您的高速浏览体验。继续浏览本站，即表示您同意我们使用cookie。查看详情