文档

WholeReduceMax

Function Usage

Computes the maximum value of all data and its index in each repeat. The returned index value is the internal index of each repeat. For details about reduction instructions, see Reduction Instructions.

Prototype

  • Bitwise mask mode 
    1
2
    		 
    template <typename T, bool isSetMask = true>
__aicore__ inline void WholeReduceMax(const LocalTensor<T>& dstLocal, const LocalTensor<T>& srcLocal, const uint64_t mask[], const int32_t repeatTimes, const int32_t dstRepStride, const int32_t srcBlkStride, const int32_t srcRepStride, ReduceOrder order = ReduceOrder::ORDER_VALUE_INDEX)
  • Contiguous mask mode 
    1
2
    		 
    template <typename T, bool isSetMask = true>
__aicore__ inline void WholeReduceMax(const LocalTensor<T>& dstLocal, const LocalTensor<T>& srcLocal, const int32_t mask, const int32_t repeatTimes, const int32_t dstRepStride, const int32_t srcBlkStride, const int32_t srcRepStride, ReduceOrder order = ReduceOrder::ORDER_VALUE_INDEX)

Parameters

Table 1 Parameters in the template

Parameter

Description

T

Operand data type.

For the Atlas Training Series Product , 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 MASK_PLACEHOLDER.
Table 2 Parameters

Parameter

Input/Output

Description

dstLocal

Output

Destination operand.

The type is LocalTensor, and the supported TPosition is VECIN, VECCALC, or VECOUT.

The start address of the LocalTensor must be 4-byte aligned (for data of the half type) or 8-byte aligned (for data of the float type).

srcLocal

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.

The source operand must have the same data type as the destination operand.

mask

Input

mask is used to control the elements that participate in computation in each iteration.

  • 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 iteration 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].
  • 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 parameter type is a uint64_t array whose length is 2.

    For example, if mask = [0, 8] and 8 = 0b1000, only the fourth element participates in computation.

    The parameter value range is related to the operand data type. The maximum number of elements that can be processed in each iteration varies according to the data type. When the operand is 16-bit, mask[0] and mask[1] ∈ [0, 264 -1] and cannot be 0 at the same time. When the operand is 32-bit, mask[1] is 0 and mask[0] ∈ (0, 264 – 1]. When the operand is 64-bit, mask[1] is 0 and mask[0] ∈ (0, 232 – 1].

repeatTimes

Input

Number of repeats (iterations). The value range is [0, 255].

For details about this parameter, see Common Parameters.

dstRepStride

Input

Address stride between adjacent iterations of the destination operand. The unit is the length after reduction of a repeat.

When the index and minimum value are returned, the unit is twice the length of the data type of dstLocal. When dstLocal is of the half type, the unit is 4 bytes.

When only the minimum value is returned, the unit is the length of the data type of dstLocal.

When only the index is returned, the unit is the length of the data type of uint32_t.

For the Atlas Training Series Product , this parameter cannot be set to 0.

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 in each iteration of the source operand. For details, see repeatStride.

order

Input

Specifies the relative position between the index and value in dstLocal and the return behavior. The parameter is of the ReduceOrder type. The default value is ORDER_VALUE_INDEX. The values are as follows:

  • ORDER_VALUE_INDEX: indicates that the value is in the lower half. The storage sequence of returns is [value, index].
  • ORDER_INDEX_VALUE: indicates that the index is in the lower half. The storage sequence of returns is [index, value].
  • ORDER_ONLY_VALUE: indicates that only the maximum value is returned. The storage sequence of returns is [value].
  • ORDER_ONLY_INDEX: indicates that only the index of the maximum value is returned. The storage sequence of returns is [value].

For the Atlas Training Series Product , ORDER_VALUE_INDEX is supported.

Returns

None

Availability

Atlas Training Series Product

Constraints

  • For details about the address alignment requirements of srcLocal and dstLocal, see General Restrictions.
  • To save address space, you can define a tensor for the source operand and destination operand to use at the same time (that is, address overlapping). The restrictions are as follows:
    • For a single repeat (repeatTimes = 1), the source operand must completely overlap the destination operand.
    • For multiple repeats (repeatTimes > 1), if there is a dependency between the source operand and the destination operand, that is, the destination operand of the Nth iteration is the source operand of the (N + 1)th iteration, address overlapping is not allowed.
  • The dstLocal result storage sequence is determined by order. The default values are the maximum value and its index. In the returned result, the indexes are stored based on the data type of dstLocal. For example, if dstLocal uses the half type, the indexes are stored based on the half type. The reinterpret_cast method is used to convert the indexes to the corresponding integer type when they are read. If the input type is half, reinterpret_cast<uint16_t*> is required. If the input type is float, reinterpret_cast<uint32_t*> is required. In the complete example, the first two computation results are [9.980e-01 5.364e-06], use the reinterpret_cast method to convert 5.364e-06 and obtain the index value 90.
  • Using proper reduction instructions 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.

Example

  • Example of high-dimensional tensor sharding computation (contiguous mask mode) 
    1
2
3
4
5
6
    		 
    // Both dstLocal and srcLocal are of the half type. For srcLocal, the computation data is of size 512 and is continuously arranged. Its compute result is also continuously arranged. It uses the high-dimensional tensor sharding computation API. mask is set to 128, indicating that all elements are involved in the computation.
// Based on the preceding information, repeatTimes is 4, dstRepStride is 1, srcBlkStride is 1, and srcRepStride is 8.
// To obtain the maximum value and index that are stored in the format of [value, index], you can use the default order. The following is an example:
AscendC::WholeReduceMax<half>(dstLocal, srcLocal, 128, 4, 1, 1, 8);
// To obtain the maximum value and index that are stored in the format of [index, value], you can use the following example API:
AscendC::WholeReduceMax<half>(dstLocal, srcLocal, 128, 4, 1, 1, 8, AscendC::ReduceOrder::ORDER_INDEX_VALUE);
  • Example of high-dimensional tensor sharding computation (bitwise mask mode) 
    1
2
3
4
5
    		 
    // Both dstLocal and srcLocal are of the half type. For srcLocal, the computation data is of size 512 and is continuously arranged. Its compute result is also continuously arranged. It uses the high-dimensional tensor sharding computation API. mask is set to 128, indicating that all elements are involved in the computation.
// Based on the preceding information, repeatTimes is 4, dstRepStride is 1, srcBlkStride is 1, and srcRepStride is 8.
// To obtain the maximum value and index that are stored in the format of [value, index], you can use the default order. The following is an example:
uint64_t mask[2] = { 0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF };
AscendC::WholeReduceMax<half>(dstLocal, srcLocal, mask, 4, 1, 1, 8);
  • Complete example: 
     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
    		 
    #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);
        repeat = srcDataSize / mask;
        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>();
        AscendC::WholeReduceMax<half> (dstLocal, srcLocal, mask, repeat, 1, 1, 8); // Use the default order. ReduceOrder::ORDER_VALUE_INDEX
        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::QuePosition::VECIN, 1> inQueueSrc;
    AscendC::TQue<AscendC::QuePosition::VECOUT, 1> outQueueDst;
    AscendC::GlobalTensor<half> srcGlobal, dstGlobal;
    int srcDataSize = 1024;
    int dstDataSize = 16;
    int mask = 128;
    int repeat = 0;
};
extern "C" __global__ __aicore__ void reduce_kernel(__gm__ uint8_t* src, __gm__ uint8_t* dstGm)
{
    KernelReduce op;
    op.Init(src, dstGm);
    op.Process();
}

    The following is an example:

    Input (src_gm):
[0.00787  0.8516   0.01558  0.152    0.887    0.2532   0.2272   0.1295
 0.7207   0.628    0.5522   0.991    0.3164   0.961    0.526    0.5513
 0.03973  0.3293   0.809    0.562    0.915    0.56     0.3464   0.3438
 0.6094   0.1201   0.8384   0.848    0.004436 0.4263   0.01917  0.753
 0.9126   0.2307   0.1066   0.644    0.8657   0.7085   0.7915   0.1707
 0.3806   0.957    0.0483   0.858    0.10675  0.21     0.03345  0.55
 0.3757   0.3281   0.927    0.09406  0.6445   0.985    0.405    0.09393
 0.773    0.7227   0.03714  0.595    0.889    0.0948   0.4202   0.2747
 0.5894   0.3022   0.894    0.675    0.6016   0.938    0.585    0.5244
 0.8643   0.888    0.794    0.636    0.976    0.148    0.7427   0.1742
 0.32     0.0649   0.2954   0.2018   0.833    0.0976   0.4048   0.2861
 0.8765   0.722    0.998    0.03041  0.005512 0.9087   0.9873   0.1436
 0.4812   0.1901   0.78     0.6934   0.2317   0.3782   0.8613   0.808
 0.06885  0.3584   0.5684   0.541    0.5415   0.3096   0.5957   0.9043
 0.7964   0.501    0.4324   0.7544   0.687    0.8447   0.526    0.548
 0.926    0.9106   0.1616   0.183    0.6704   0.642    0.4783   0.1797
 0.2078   0.59     0.4866   0.4683   0.649    0.7266   0.4976   0.8364
 0.6245   0.07385  0.0786   0.586    0.7827   0.3298   0.9497   0.1617
 0.4375   0.3572   0.2896   0.6465   0.1156   0.4905   0.2617   0.8267
 0.2054   0.1415   0.2993   0.8374   0.754    0.942    0.6416   0.1222
 0.1465   0.3335   0.3577   0.6484   0.614    0.5825   0.6807   0.9297
 0.694    0.759    0.908    0.9126   0.4731   0.963    0.3271   0.724
 0.4077   0.335    0.672    0.4219   0.1818   0.843    0.2708   0.0816
 0.457    0.3481   0.67     0.6895   0.6924   0.191    0.2013   0.2484
 0.8833   0.9146   0.4102   0.1063   0.6685   0.804    0.6606   0.2491
 0.34     0.3281   0.823    0.603    0.521    0.6797   0.401    0.5
 0.03683  0.04758  0.507    0.667    0.9014   0.263    0.2477   0.0179
 0.8735   0.007023 0.545    0.758    0.3508   0.6333   0.9375   0.5903
 0.2732   0.0847   0.489    0.196    0.5557   0.403    0.9204   0.3655
 0.5083   0.7515   0.3347   0.6914   0.2185   0.2458   0.5537   0.3457
 0.4878   0.869    0.908    0.0877   0.295    0.9      0.9307   0.05545
 0.4639   0.4001   0.8433   0.4883   0.916    0.7026   0.5063   0.05164
 0.936    0.844    0.2086   0.625    0.0197   0.4312   0.3677   0.983
 0.625    0.004665 0.2479   0.3093   0.9214   0.003672 0.7915   0.921
 0.331    0.01127  0.703    0.6416   0.4053   0.53     0.9688   0.10297
 0.5547   0.07367  0.2305   0.02821  0.8115   0.4202   0.0561   0.0917
 0.04828  0.536    0.0905   0.328    0.8413   0.3696   0.982    0.3733
 0.436    0.753    0.1937   0.8706   0.991    0.273    0.763    0.418
 0.4446   0.513    0.6724   0.1179   0.921    0.756    0.7144   0.6196
 0.9634   0.562    0.3088   0.864    0.709    0.6797   0.2114   0.534
 0.5225   0.1852   0.038    0.5454   0.8823   0.849    0.608    0.7734
 0.7446   0.7236   0.1903   0.1031   0.497    0.57     0.172    0.1907
 0.6333   0.641    0.681    0.2323   0.1007   0.4094   0.3655   0.4248
 0.08044  0.1483   0.08716  0.354    0.128    0.3933   0.775    0.215
 0.728    0.909    0.4204   0.618    0.2517   0.9106   0.3647   0.5977
 0.3445   0.315    0.488    0.99     0.9443   0.6196   0.9287   0.088
 0.9946   0.796    0.7515   0.1912   0.4312   0.7974   0.735    0.01536
 0.7456   0.643    0.484    0.218    0.9272   0.1703   0.1885   0.1982
 0.754    0.902    0.848    0.05832  0.4138   0.6885   0.3853   0.3499
 0.639    0.5786   0.6353   0.5664   0.02621  0.56     0.532    0.08246
 0.733    0.1334   0.0728   0.7817   0.5273   0.126    0.179    0.7334
 0.1565   0.457    0.4807   0.6987   0.5845   0.6206   0.902    0.9277
 0.501    0.6763   0.3418   0.7925   0.07556  0.0929   0.9014   0.3145
 0.04907  0.7188   0.958    0.7275   0.1963   0.1742   0.785    0.518
 0.61     0.1112   0.481    0.10583  0.198    0.181    0.3271   0.2773
 0.2391   0.5625   0.621    0.173    0.05936  0.5654   0.838    0.865
 0.01523  0.6724   0.546    0.737    0.778    0.8613   0.7085   0.8213
 0.08826  0.818    0.4866   0.159    0.4143   0.1007   0.7773   0.487
 0.5225   0.8984   0.4907   0.525    0.4075   0.2632   0.2292   0.134
 0.4622   0.65     0.294    0.607    0.2725   0.2603   0.9326   0.787
 0.9478   0.941    0.3066   0.2944   0.3928   0.73     0.1797   0.2157
 0.609    0.4216   0.8984   0.8477   0.863    0.2478   0.993    0.6274
 0.724    0.03668  0.0991   0.5825   0.662    0.6904   0.7017   0.2379
 0.514    0.1646   0.3245   0.03072  0.3232   0.907    0.9966   0.6396
 0.2969   0.02539  0.66     0.764    0.7803   0.515    0.04074  0.2258
 0.08887  0.1782   0.875    0.1517   0.2351   0.3848   0.5933   0.6875
 0.1969   0.1283   0.06232  0.4348   0.168    0.6904   0.5464   0.12036
 0.885    0.007717 0.5967   0.2856   0.628    0.62     0.854    0.4297
 0.733    0.2274   0.9736   0.01622  0.456    0.4763   0.9707   0.874
 0.8794   0.511    0.1628   0.03458  0.506    0.1464   0.3674   0.1532
 0.786    0.3809   0.406    0.015434 0.901    0.951    0.3018   0.3584
 0.5337   0.4983   0.85     0.833    0.7324   0.492    0.39     0.09845
 0.8965   0.862    0.4033   0.181    0.2203   0.3738   0.2761   0.9653
 0.3577   0.289    0.3167   0.91     0.2688   0.3972   0.585    0.2178
 0.307    0.4966   0.513    0.5225   0.786    0.1888   0.9287   0.5093
 0.1193   0.3987   0.799    0.9995   0.611    0.9897   0.7515   0.4478
 0.3232   0.2426   0.3323   0.7134   0.77     0.7275   0.02043  0.3132
 0.3555   0.03122  0.8623   0.4705   0.6357   0.3157   0.5063   0.1711
 0.885    0.7554   0.815    0.0213   0.4346   0.049    0.905    0.525
 0.921    0.02411  0.771    0.7227   0.1786   0.278    0.03387  0.7744
 0.05875  0.8955   0.8374   0.715    0.3765   0.02075  0.675    0.9883
 0.63     0.7017   0.299    0.92     0.1644   0.3977   0.487    0.818
 0.636    0.3452   0.6406   0.783    0.3728   0.1619   0.7725   0.4673
 0.297    0.9375   0.083    0.0914   0.6704   0.08923  0.332    0.0973
 0.507    0.201    0.1658   0.2358   0.8706   0.6846   0.6396   0.289
 0.831    0.669    0.4683   0.2568   0.219    0.616    0.978    0.1564
 0.925    0.4265   0.6055   0.7246   0.235    0.5376   0.03668  0.2441
 0.7935   0.383    0.2996   0.3523   0.2544   0.6006   0.8896   0.757
 0.7134   0.3196   0.3657   0.249    0.2429   0.921    0.877    0.728
 0.8853   0.1635   0.546    0.9243   0.676    0.4749   0.3928   0.4187
 0.612    0.3953   0.2372   0.4092   0.1523   0.1599   0.03108  0.1602
 0.2474   0.3572   0.0643   0.9434   0.52     0.8574   0.959    0.7593
 0.2318   0.5444   0.2222   0.3884   0.8066   0.4573   0.664    0.335
 0.02025  0.1519   0.01386  0.989    0.852    0.695    0.01289  0.3433
 0.2148   0.9404   0.6753   0.704    0.11163  0.675    0.5264   0.1514
 0.5273   0.9785   0.2769   0.4846   0.2747   0.558    0.742    0.681
 0.835    0.9546   0.941    0.588    0.785    0.2095   0.07294  0.4343
 0.086    0.5825   0.513    0.6313   0.04236  0.4072   0.558    0.681
 0.4805   0.492    0.625    0.7744   0.002626 0.662    0.9043   0.4766
 0.6597   0.6934   0.3394   0.05453  0.9146   0.2222   0.7925   0.605
 0.812    0.671    0.4329   0.2118   0.363    0.1444   0.0955   0.692
 0.675    0.3      0.6846   0.535    0.9834   0.929    0.3582   0.964
 0.3835   0.1466   0.801    0.954    0.2554   0.01357  0.6636   0.8325
 0.6494   0.817    0.2268   0.00904  0.0487   0.08716  0.6753   0.3833
 0.663    0.396    0.6685   0.983    0.0728   0.694    0.02364  0.137
 0.1727   0.231    0.7896   0.8057   0.478    0.883    0.1785   0.5938
 0.11456  0.6997   0.1945   0.02365  0.7236   0.8623   0.2178   0.1295
 0.3867   0.7188   0.11475  0.6      0.419    0.2673   0.4404   0.0107
 0.4304   0.1364   0.3708   0.1158   0.1714   0.3123   0.3403   0.7163
 0.079    0.6245   0.719    0.558    0.4526   0.09924  0.512    0.2452
 0.519    0.999    0.7207   0.5605   0.7217   0.653    0.1164   0.789
 0.4724   0.2727   0.10315  0.9644   0.7573   0.06464  0.858    0.7847
 0.958    0.618    0.9536   0.46     0.9766   0.4263   0.4363   0.4434
 0.95     0.3032   0.4338   0.809    0.1642   0.0561   0.2668   0.1853
 0.356    0.934    0.968    0.327    0.913    0.434    0.6616   0.00502
 0.05066  0.5327   0.276    0.5176   0.0674   0.6143   0.8345   0.2976
 0.315    0.6646   0.527    0.791    0.0299   0.4558   0.8354   0.3115
 0.3735   0.3582   0.742    0.2637   0.8877   0.7603   0.4568   0.2045
 0.4746   0.392    0.65     0.391    0.972    0.6973   0.2297   0.568
 0.49     0.1895   0.547    0.79     0.747    0.5205   0.313    0.3809
 0.7817   0.32     0.1012   0.339    0.716    0.8955   0.8564   0.126
 0.6597   0.228    0.1194   0.4775   0.173    0.0265   0.7456   0.859
 0.4841   0.595    0.4553   0.1351   0.2246   0.3564   0.1832   0.8535
 0.703    0.2423   0.04187  0.145    0.997    0.1919   0.571    0.8555
 0.1578   0.2688   0.405    0.3909   0.1428   0.863    0.7295   0.3267
 0.1294   0.5986   0.677    0.7065   0.8853   0.923    0.9385   0.935
 0.1747   0.32     0.2292   0.2676   0.1161   0.4666   0.3826   0.2588
 0.1863   0.7993   0.3984   0.2961   0.2952   0.3247   0.923    0.05746 ]
Output (dst_gm):
[9.980e-01 5.364e-06 9.629e-01 2.682e-06 9.946e-01 6.676e-06 9.966e-01
 7.510e-06 9.995e-01 5.424e-06 9.888e-01 6.378e-06 9.990e-01 6.735e-06
 9.971e-01 5.484e-06]
Parent topic: Reduction Instructions