High-Level API Porting Guide

Ascend C high-level APIs are basically compatible with the 351x and 220x architectures, and some APIs are extended. Currently, the 351x architecture does not support high-level APIs for convolution computation.

Matmul High-level APIs

Supported data types are changed.

**Table 1** Data type compatibility
Matrix A	Matrix B	Bias Matrix	Matrix C	Description
int4b_t	int4b_t	int32_t	int32_t, half	Not supported in the 351x architecture.
fp8_e4m3fn_t, fp8_e5m2_t	fp8_e4m3fn_t, fp8_e5m2_t	float, half, bfloat16_t	fp8_e4m3fn_t, half, bfloat16_t, float	Added in the 351x architecture.
hifloat8_t	hifloat8_t	float, half, bfloat16_t	hifloat8_t, half, bfloat16_t, float	Added in the 351x architecture.
float	float	bfloat16_t	float, half, bfloat16_t	Added in the 351x architecture.
bfloat16_t	bfloat16_t	bfloat16_t	float, half, bfloat16_t	Added in the 351x architecture.
half	half	bfloat16_t	float, half, bfloat16_t	Added in the 351x architecture.
int8_t	int8_t	int32_t	bfloat16_t	Added in the 351x architecture.

The 4:2 sparsity feature is not supported. For details about the compatibility solution, see the 4:2 structured sparsity.

Other High-Level APIs

**Table 2** Mathematical computation
Ascend C High-Level API	Compatibility Description
Tanh, Asin, Sin, Acos, Cos, Log, Atan, Fmod	Compatible with the 220x architecture. Extended support for algorithm configuration. The algorithm used by the API can be configured through template parameters, providing selection of high-precision and high-performance algorithms.
Sinh, Cosh, Tan, Trunc, Frac, Erf, Erfc, Atanh, Asinh, Acosh, Floor, Ceil, Round, Axpy, Exp, Lgamma, Digamma, Xor, Cumsum	Compatible with the 220x architecture.
Power	Compatible with the 220x architecture. Extended support for the uint8_t, int8_t, uint16_t, int16_t, uint32_t, and bfloat16_t data types.
Sign	Compatible with the 220x architecture. Extended support for the int64_t data type.

**Table 3** Activation functions
Ascend C High-Level API	Compatibility Description
SoftMax, SimpleSoftMax, SoftmaxFlash, SoftmaxGrad, SoftmaxFlashV2, SoftmaxFlashV3, SoftmaxGradFront, AdjustSoftMaxRes, LogSoftMax, FasterGelu, FasterGeluV2, Gelu, SwiGLU, Silu, Swish, GeGLU, ReFlu, Sigmoid	Compatible with the 220x architecture.

**Table 4** Data normalization
Ascend C High-Level API	Compatibility Description
LayerNormGrad, LayerNormGradBeta, RmsNorm, BatchNorm, DeepNorm, GroupNorm	Compatible with the 220x architecture.
Normalize, WelfordUpdate	Compatible with the 220x architecture. Extended support for the bfloat16_t data type.
LayerNorm	Compatible with the 220x architecture. Extended support for variance computation.
WelfordFinalize	Compatible with the 220x architecture. Extended support for algorithm configuration. The template parameter is used to specify whether to use the correction coefficient when computing the variance.

**Table 5** Quantization
Ascend C High-Level API	Compatibility Description
AscendQuant	Compatible with the 220x architecture. Extended support for PRE_TOKEN quantization and PRE_GROUP quantization. Extended support for quantization from the half, bfloat16_t, and float types to the fp8_e5m2_t, fp8_e4m3fn_t, hifloat8_t, and int8_t types. Extended support for quantization from the half and bfloat16_t types to the fp4x2_e1m2_t and fp4x2_e2m1_t types.
AscendDequant	Compatible with the 220x architecture. Extended support for PRE_TOKEN quantization and PRE_GROUP quantization. Extended support for dequantization from the int32_t type to the half, bfloat16_t, and float types. Extended support for dequantization from the float type to the half, bfloat16_t, and float types.
AscendAntiQuant	Compatible with the 220x architecture. Extended support for PRE_TOKEN quantization and PRE_GROUP quantization. Extended support for fake-quantization from the int8_t, hifloat8_t, fp8_e5m2_t, and fp8_e4m3fn_t types to the half, bfloat16_t, float, and half types. Extended support for fake-quantization from the fp4x2_e1m2_t and fp4x2_e2m1_t types to the half and bfloat16_t types.

**Table 6** Reduce operations
Ascend C High-Level API	Compatibility Description
Sum, Mean, ReduceXorSum, ReduceMean, ReduceAny, ReduceAll, ReduceProd	Compatible with the 220x architecture.
ReduceSum	Compatible with the 220x architecture. Extended support for the int32_t, uint32_t, int64_t, and uint64_t data types.
ReduceMax, ReduceMin	Compatible with the 220x architecture. Extended support for the int8_t, uint8_t, int16_t, uint16_t, bfloat16_t, int32_t, uint32_t, int64_t, and uint64_t data types.

**Table 7** Sorting operations
Ascend C High-Level API	Compatibility Description
Concat, Extract, GetSortOffset, GetSortLen, MrgSort	Compatible with the 220x architecture.
TopK	Compatible with the 220x architecture. Extended support for the uint8_t, int8_t, uint16_t, int16_t, uint32_t, int32_t, bfloat16_t, uint64_t, and int64_t data types when the RADIX_SELECT algorithm is used.
Sort	Compatible with the 220x architecture. Extended support for algorithm configuration. The template parameter is used to specify the sorting algorithm and the ascending or descending order.

**Table 8** Index computation
Ascend C High-Level API	Compatibility Description
Arange	Compatible with the 220x architecture. Extended support for the int64_t data type.

**Table 9** Data filtering
Ascend C High-Level API	Compatibility Description
Select	Compatible with the 220x architecture.
DropOut	Compatible with the 220x architecture. Extended support for the bfloat16_t data type.

**Table 10** Tensor transformation
Ascend C High-Level API	Compatibility Description
Transpose	Compatible with the 220x architecture. Supported data layout conversion scenarios. 2D tensor transpose or transpose of the last two dimensions of a 3D tensor Interchange of the first and second dimensions in a 3D tensor Interchange of the first and third dimensions in a 3D tensor 2D ND2NZ transpose using interleaving instructions
TransData, Pad, UnPad	Compatible with the 220x architecture.
BroadCast	Compatible with the 220x architecture. Extended support for dynamic shapes. Extended support for the int16_t, uint16_t, bfloat16_t, int32_t, and uint32_t data types.
Fill	Compatible with the 220x architecture. Extended support for the uint8_t, int8_t, bfloat16_t, uint64_t, and int64_t data types.

**Table 11** HCCL
Ascend C High-Level API	Compatibility Description
HCCL template parameters	Supported the HCCL_SERVER_TYPE_CCU server type.
InitV2, SetCcTilingV2, AllReduce, AllGather, ReduceScatter, AlltoAll, AlltoAllV, Commit, Wait, Finalize	Compatible with the 220x architecture.
BatchWrite, iterate, Query, InterHcclGroupSync, GetWindowsInAddr, GetWindowsOutAddr, GetRankId, GetRankDim, QueueBarrier, GetQueueNum	Not supported in the 351x architecture.
SetReduceType, AlltoAllvWrite	Added in the 351x architecture.

Parent topic: 220x-to-351x Porting Guide