Conv2D (Deprecated)

Atlas 350 Accelerator Card

x

Atlas A3 training product/Atlas A3 inference product

x

Atlas A2 training product/Atlas A2 inference product

x

Atlas 200I/500 A2 inference product

x

Atlas inference product AI Core

√

Atlas inference product Vector Core

x

Atlas training product

√

dst

Output

Destination operand.

For the Atlas training product, the supported TPosition is CO1 or CO2.

For the Atlas inference product AI Core, the supported TPosition is CO1 or CO2.

In the results, the valid tensor format is [Cout/16, Ho, Wo, 16], with a total size of Cout × Ho × Wo. The values of Ho and Wo can be calculated as follows:

Ho = floor((H + pad_top + pad_bottom – dilation_h × (Kh – 1) – 1)/stride_h + 1)

Wo = floor((W + pad_left + pad_right – dilation_w × (Kw – 1) – 1)/stride_w + 1)

The hardware requires Ho × Wo to be a multiple of 16. When defining the dst tensor, shape should be rounded up to the multiple of 16. The actual shape size should be Cout × round_howo:

round_howo = ceil(Ho × Wo/16) × 16

featureMap

Input

Input tensor. TPosition of the tensor is A1.

Shape of feature_map, in the format [C1, H, W, C0].

C1 * C0 indicates the number of input channels.

• If feature_map is of type half, C0 is 16.
• If feature_map is of type int8_t, C0 is 32.
• Value range of C1: [1, 4]. Value range of input channel: [16, 32, 64, 128].

H indicates the height. Value range: [1, 40].

W indicates the width. Value range: [1, 40].

weight

Input

Convolution kernel (weight) tensor. TPosition of the tensor is B1.

Shape of weight, in the format [C1, Kh, Kw, Cout, C0].

C1 × C0 indicates the number of input channels. The requirements for C0 are as follows:

• If feature_map is of type half, C0 is 16.
• If feature_map is of type int8_t, C0 is 32.
• Value range of C1: [1, 4].
• kernel_shape has the same number of input channels as fm_shape.

Cout indicates the number of convolution kernels. The value range is [16, 32, 64, 128], which must be a multiple of 16.

Kh indicates the height of the convolution kernel. Value range: [1, 5].

Kw indicates the width of the convolution kernel. Value range is [1, 5].

conv2dParams

Input

Status parameters such as the input matrix shape. The type is Conv2dParams. The specific definition of the structure is as follows:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

struct Conv2dParams {
    uint32_t imgShape[CONV2D_IMG_SIZE];       // [H, W]
    uint32_t kernelShapeIn[CONV2D_KERNEL_SIZE]; // [Kh, Kw]
    uint32_t stride[CONV2D_STRIDE];          // [stride_h, stride_w]
    uint32_t cin;                            // cin = C0 * C1;
    uint32_t cout;
    uint32_t padList[CONV2D_PAD];       // [pad_left, pad_right, pad_top, pad_bottom]
    uint32_t dilation[CONV2D_DILATION]; // [dilation_h, dilation_w]
    uint32_t initY;
    uint32_t partialSum;
};

tilling

Input

Fractal control parameter. The type is Conv2dTilling. The specific definition of the structure is as follows:

 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

struct Conv2dTilling {
    const uint32_t blockSize = 16; // # M block size is always 16
    LoopMode loopMode = LoopMode::MODE_NM;

    uint32_t c0Size = 32;
    uint32_t dTypeSize = 1;

    uint32_t strideH = 0;
    uint32_t strideW = 0;
    uint32_t dilationH = 0;
    uint32_t dilationW = 0;
    uint32_t hi = 0;
    uint32_t wi = 0;
    uint32_t ho = 0;
    uint32_t wo = 0;

    uint32_t height = 0;
    uint32_t width = 0;

    uint32_t howo = 0;

    uint32_t mNum = 0;
    uint32_t nNum = 0;
    uint32_t kNum = 0;

    uint32_t mBlockNum = 0;
    uint32_t kBlockNum = 0;
    uint32_t nBlockNum = 0;

    uint32_t roundM = 0;
    uint32_t roundN = 0;
    uint32_t roundK = 0;

    uint32_t mTileBlock = 0;
    uint32_t nTileBlock = 0;
    uint32_t kTileBlock = 0;

    uint32_t mIterNum = 0;
    uint32_t nIterNum = 0;
    uint32_t kIterNum = 0;

    uint32_t mTileNums = 0;

    bool mHasTail = false;
    bool nHasTail = false;
    bool kHasTail = false;

    uint32_t kTailBlock = 0;
    uint32_t mTailBlock = 0;
    uint32_t nTailBlock = 0;

    uint32_t mTailNums = 0;
};

imgShape

vector<int>

kernelShape

vector<int>

Shape of weight, in the format [Kh, Kw].

• Kh indicates the height. Value range: [1, 5].
• Kw indicates the width. Value range: [1, 5].

stride

vector<int>

cin

int

Fractal layout parameter. Cin = C1 * C0. Cin indicates the number of input channels. The value range of C1 is [1, 4].

• If feature_map is of type float, C0 = 8. The value range of the input channel is [8, 16, 24, 32].
• If feature_map is of type half, C0 is 16. The value range of the input channel is [16, 32, 48, 64].
• If feature_map is of type int8_t, C0 is 32. The value range of channel is [32, 64, 96, 128].

cout

int

Cout indicates the number of convolution kernels. The value range is [16, 32, 64, 128], which must be a multiple of 16.

padList

vector<int>

dilation

vector<int>

The width of the dilated convolution kernel is dilation_w × (Kw – 1) + 1, and the height of the dilated convolution kernel is dilation_h × (Kh – 1) + 1.

initY

uint32_t

Indicates whether dst needs to be initialized.

• 0: No bias is used. L0C must be initialized. The dst initial matrix retains the previous results, and the new Conv2D computation results will be accumulated on top of the previous Conv2D results.
• 1: No bias is used. L0C does not need to be initialized. The data in the dst initial matrix is meaningless, and the new computation results will directly overwrite the data in dst.

partialSum

uint32_t

blockSize

uint32_t

Number of elements stored in a dimension. The value is fixed at 16.

loopMode

LoopMode

Traversal mode. The structure is defined as follows:

1
2
3
4
5
6

enum class LoopMode {
    MODE_NM = 0,
    MODE_MN = 1,
    MODE_KM = 2,
    MODE_KN = 3
};

c0Size

uint32_t

Length of a block. The value can be 16 or 32.

dtypeSize

uint32_t

Length of the input data, in bytes. The value range is [1, 2].

strideH

uint32_t

Height of the convolution stride. The value range is [1, 4].

strideW

uint32_t

Width of the convolution stride. The value range is [1, 4].

dilationH

uint32_t

Height of the convolution dilation factor. The value range is [1, 4].

dilationW

uint32_t

Width of the convolution dilation factor. The value range is [1, 4].

hi

uint32_t

Height of the feature_map shape. The value range is [1, 40].

wi

uint32_t

Width of the feature_map shape. The value range is [1, 40].

ho

uint32_t

Height of the feature_map shape. The value range is [1, 40].

wo

uint32_t

Width of the feature_map shape. The value range is [1, 40].

height

uint32_t

Height of the weight shape. The value range is [1, 5].

width

uint32_t

Width of the weight shape. The value range is [1, 5].

howo

uint32_t

Size of the feature_map shape (ho * wo)

mNum

uint32_t

Equivalent data length of the M axis. The value range is [1, 4096].

nNum

uint32_t

Equivalent data length of the N axis. The value range is [1, 4096].

kNum

uint32_t

Equivalent data length of the K axis. The value range is [1, 4096].

roundM

uint32_t

Equivalent data length of the M axis. The value is rounded up to an integer multiple of blockSize. The value range is [1, 4096].

roundN

uint32_t

Equivalent data length of the N axis. The value is rounded up to an integer multiple of blockSize. The value range is [1, 4096].

roundK

uint32_t

Equivalent data length of the K axis. The value is rounded up to an integer multiple of c0Size. The value range is [1, 4096].

mBlockNum

uint32_t

Number of blocks on the M axis. mBlockNum = mNum/blockSize. The value range is [1, 4096].

nBlockNum

uint32_t

Number of blocks on the N axis. nBlockNum = nNum/blockSize. The value range is [1, 4096].

kBlockNum

uint32_t

Number of blocks on the K axis. kBlockNum = kNum/blockSize. The value range is [1, 4096].

mIterNum

uint32_t

Number of dimensions traversed on the M axis. The value range is [1, 4096].

nIterNum

uint32_t

Number of dimensions traversed on the N axis. The value range is [1, 4096].

kIterNum

uint32_t

Number of dimensions traversed on the K axis. The value range is [1, 4096].

mTileBlock

uint32_t

Number of split blocks on the M axis. The value range is [1, 4096].

nTileBlock

uint32_t

Number of split blocks on the N axis. The value range is [1, 4096].

kTileBlock

uint32_t

Number of split blocks on the K axis. The value range is [1, 4096].

kTailBlock

uint32_t

Number of tail blocks on the K axis. The value range is [1, 4096].

mTailBlock

uint32_t

Number of tail blocks on the M axis. The value range is [1, 4096].

nTailBlock

uint32_t

Number of tail blocks on the N axis. The value range is [1, 4096].

kHasTail

bool

Indicates whether a tail block exists on the K axis.

mHasTail

bool

Indicates whether a tail block exists on the M axis.

nHasTail

bool

Indicates whether a tail block exists on the N axis.

mTileNums

uint32_t

Length of split blocks on the M axis. The value range is [1, 4096].

mTailNums

uint32_t

Length of tail blocks on the M axis. The value range is [1, 4096].

int8_t

int8_t

int32_t

half

half

float

half

half

half