Transpose

Atlas 350 Accelerator Card

√

Atlas A3 training product/Atlas A3 inference product

√

Atlas A2 training product/Atlas A2 inference product

√

Atlas 200I/500 A2 inference product

√

Atlas inference product AI Core

√

Atlas inference product Vector Core

x

Atlas training product

√

T

Operand data type.

• Standard transpose:

  For the Atlas 350 Accelerator Card, the supported data types are uint16_t, int16_t, and half.

  For the Atlas inference product AI Core, the supported data types are uint16_t, int16_t, and half.

  For the Atlas inference product AI Core, the supported data types are uint16_t, int16_t, and half.

  For the Atlas 200I/500 A2 inference product, the supported data types are uint16_t, int16_t, and half.

  For the Atlas inference product AI Core, the supported data types are uint16_t, int16_t, and half.

  For the Atlas training product, the supported data types are uint16_t, int16_t, and half.

• Enhanced transpose:
  • When transposeType is TRANSPOSE_ND2ND_B16:

    For the Atlas 350 Accelerator Card, the supported data types are uint16_t, int16_t, and half.

    For the Ascend 610's AI Core, the supported data type is uint16_t.

    For the Atlas A2 training product/Atlas A2 inference product, the supported data type is uint16_t.

    For the Atlas 200I/500 A2 inference product, the supported data type is uint16_t.

    For the Atlas inference product AI Core, the supported data type is uint16_t.

  • When transposeType is TRANSPOSE_NCHW2NHWC or TRANSPOSE_NHWC2NCHW:

    For the Atlas 350 Accelerator Card, the supported data types are int8_t, uint8_t, fp4x2_e2m1_t, fp4x2_e1m2_t, hifloat8_t, fp8_e5m2_t, fp8_e4m3fn_t, fp8_e8m0_t, int4x2_t, int16_t, uint16_t, half, bfloat16_t, int32_t, uint32_t, float, and complex32.

    For the Atlas A3 training product/Atlas A3 inference product, the supported data types are int8_t, uint8_t, int16_t, uint16_t, half, int32_t, uint32_t, and float.

    For the Atlas A2 training product/Atlas A2 inference product, the supported data types are int8_t, uint8_t, int16_t, uint16_t, half, int32_t, uint32_t, and float.

    For the Atlas inference product AI Core, the supported data types are int8_t, uint8_t, int16_t, uint16_t, half, int32_t, uint32_t, and float.

dst

Output

Destination operand.

The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT.

The start address of LocalTensor must be 32-byte aligned.

src

Input

Source operand.

The type is LocalTensor, and TPosition can be VECIN, VECCALC, or VECOUT.

The start address of LocalTensor must be 32-byte aligned.

Its data type must be the same as that of dst.

sharedTmpBuffer

Input

Shared temporary buffer. For details about the size of sharedTmpBuffer, see Table 4.

transposeParams

Input

Data structure for controlling Transpose. The structure contains the input shape information and the transposeType parameter. For details about the definition of this data structure, see Table 3.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16

struct TransposeParamsExt {
    __aicore__ TransposeParamsExt() {}
    __aicore__ TransposeParamsExt(const uint16_t nSizeIn, const uint16_t cSizeIn, const uint16_t hSizeIn,
        const uint16_t wSizeIn, const TransposeType transposeTypeIn)
        : nSize(nSizeIn),
          cSize(cSizeIn),
          hSize(hSizeIn),
          wSize(wSizeIn),
          transposeType(transposeTypeIn)
    {}
    uint16_t nSize = 0;
    uint16_t cSize = 0;
    uint16_t hSize = 0;
    uint16_t wSize = 0;
    TransposeType transposeType = TransposeType::TRANSPOSE_ND2ND_B16;
};

nSize

Length of the n axis. Defaults to 0.

• In the transpose of the 2D matrix data block, this parameter does not need to be passed in. The input value is invalid.
• In the transpose between [N,C,H,W] and [N,H,W,C], the value range is [0, 65535].

cSize

Length of the c axis. Defaults to 0.

• In the transpose of the 2D matrix data block, this parameter does not need to be passed in. The input value is invalid.

• In the transpose between [N,C,H,W] and [N,H,W,C], the value range is [0, 4095].

hSize

Length of the h axis. Defaults to 0.

• In the transpose of the 2D matrix data block, the value is fixed at 16.

• In the transpose between [N,C,H,W] and [N,H,W,C], the value range of hSize × wSize is [0, 4095]. The product of hSize × wSize × sizeof(T) must be 32-byte aligned.

wSize

Length of the w axis. Defaults to 0.

• In the transpose of the 2D matrix data block, the value is fixed at 16.

• In the transpose between [N,C,H,W] and [N,H,W,C], the value range of hSize × wSize is [0, 4095]. The product of hSize × wSize × sizeof(T) must be 32-byte aligned.

transposeType

Data layout and reshape type. The type is the TransposeType enumeration type. The default value is TRANSPOSE_ND2ND_B16.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15

enum class TransposeType : uint8_t {
    TRANSPOSE_TYPE_NONE,           // No transpose operation is performed on the API.
    TRANSPOSE_NZ2ND_0213,          // Not supported currently.
    TRANSPOSE_NZ2NZ_0213,          // Not supported currently.
    TRANSPOSE_NZ2NZ_012_WITH_N,    // Not supported currently.
    TRANSPOSE_NZ2ND_012_WITH_N,    // Not supported currently.
    TRANSPOSE_NZ2ND_012_WITHOUT_N, // Not supported currently.
    TRANSPOSE_NZ2NZ_012_WITHOUT_N, // Not supported currently.
    TRANSPOSE_ND2ND_ONLY,          // Not supported currently.
    TRANSPOSE_ND_UB_GM,            // Not supported currently.
    TRANSPOSE_GRAD_ND_UB_GM,       // Not supported currently.
    TRANSPOSE_ND2ND_B16,           // Transpose the [16,16] two-dimensional matrix.
    TRANSPOSE_NCHW2NHWC,           // [N,C,H,W]->[N,H,W,C]
    TRANSPOSE_NHWC2NCHW            // [N,H,W,C]->[N,C,H,W]
};

TRANSPOSE_ND2ND_B16

No temporary buffer is required.

TRANSPOSE_NCHW2NHWC

For the following models:

• Atlas inference product AI Core

No temporary buffer is required.

For the following models:

• Atlas A2 training product/Atlas A2 inference product
• Atlas A3 training product/Atlas A3 inference product
• Atlas 350 Accelerator Card

The size of the temporary buffer is computed according to the following rule (pseudocode):

1
2
3

auto h0 = 16; // When the bit width of the data type is 8, h0 = 32. In other cases, h0 = 16.
auto w0 = 32 / sizeof(type);  // type indicates the data type.
auto tmpBufferSize = (cSize + 2)  * h0 * w0 * sizeof(type);

TRANSPOSE_NHWC2NCHW

For the following models:

• Atlas inference product AI Core

No temporary buffer is required.

For the following models:

• Atlas A2 training product/Atlas A2 inference product
• Atlas A3 training product/Atlas A3 inference product
• Atlas 350 Accelerator Card

The size of the temporary buffer is computed according to the following rule (pseudocode):

1
2
3

auto h0 = 16; // When the bit width of the data type is 8, h0 = 32. In other cases, h0 = 16.
auto w0 = 32 / sizeof(type);  // type indicates the data type.
auto tmpBufferSize = (cSize  * 2 + 1)  * h0 * w0 * sizeof(type);