Conv2D

dstLocal

Output

Destination operand.

The Atlas Training Series Product supports QuePosition values CO1 and CO2.

Has format [Cout/16, Ho, Wo, 16], and size Cout * Ho * Wo, where 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. The QuePosition of the tensor is A1.

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

C1 * C0 equals the input channel count.

• 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. The value range is [1, 40].

W indicates the width. The value range is [1, 40].

weight

Input

Convolution kernel (weight) tensor. The QuePosition of the tensor is B1.

Shape of weight, in the format [C1, Kh, Kw, Cout, 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].
• Has the same number of input channels as fm_shape.

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

Kh indicates the height of the filter. The value range is [1, 5].

Kw indicates the width of the filter. The 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[kConv2dImgSize];       // [H, W]
    uint32_t kernelShape[kConv2dkernelSize]; // [Kh, Kw]
    uint32_t stride[kConv2dStride];          // [stride_h, stride_w]
    uint32_t cin;                            // cin = C0 * C1;
    uint32_t cout;
    uint32_t padList[kConv2dPad];       // [pad_left, pad_right, pad_top, pad_bottom]
    uint32_t dilation[kConv2dDilation]; // [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. The value range is [1, 5].
• Kw indicates the width. The value range is [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 filters. The value range is [16, 32, 64, 128], which must be a multiple of 16.

padList

vector<int>

dilation

vector<int>

The width and height of the dilated convolution kernel is calculated as follows: dilation_w * (Kw – 1) + 1; dilation_h * (Kh – 1) + 1

initY

uint32_t

Indicates whether dstLocal needs to be initialized.

• 0: Bias is not used. L0C needs to be initialized. The dstLocal initial matrix stores the previous conv2d result and will be added up with the new conv2d result.
• 1: Bias is not used. L0C does not need to be initialized. The dstLocal initial matrix will be overwritten by the compute result.

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