Init

Applicability

Product	API with Tiling Parameter Passed as a Stack Address	API with Tiling Parameter Passed as a Global Memory Address
Atlas A3 training products/Atlas A3 inference products	√	√
Atlas A2 training products/Atlas A2 inference products	√	√
Atlas 200I/500 A2 inference products	√	x
Atlas inference product's AI Core	√	x
Atlas inference product's Vector Core	x	x
Atlas training products	x	x

Function

Initializes tiling data in Matmul objects. Resources are divided based on the tiling parameters. For details about the tiling parameters, see Matmul Tiling APIs.

You can use REGIST_MATMUL_OBJ to initialize a single Matmul object without passing the tiling parameters. Later, you can use the Init API to pass the tiling parameters to adjust the tiling data of the Matmul object. For example, in scenarios where tiling is variable, you can call Init to reset the tiling parameters for multiple times.

If no tiling change is required, you are advised to use REGIST_MATMUL_OBJ to pass tiling parameters for initialization.

Prototype

Tiling parameter passed as a stack address:

__aicore__ inline void Init(const TCubeTiling* __restrict cubeTiling, TPipe* tpipe = nullptr)

Tiling parameter passed as a global memory address:

__aicore__ inline void Init(const __gm__ TCubeTiling* gmCubeTiling, TPipe* tpipe = nullptr)

Parameters

**Table 1** Parameters of the API with the tiling parameter passed as a stack address
Parameter	Input/Output	Description
cubeTiling	Input	Matmul tiling parameter. For details about the definition of the TCubeTiling structure, see Table 1. The tiling parameter can be obtained through GetTiling on the host side and passed to the kernel side for use. On the kernel side, GET_TILING_DATA is called to move the tiling parameter to the stack space of AI Core. This API receives the tiling parameter by passing the stack address of the TCubeTiling structure.
tpipe	Input	Tpipe object.

**Table 2** Parameters of the API with the tiling parameter passed as a global memory address
Parameter	Input/Output	Description
gmCubeTiling	Input	Matmul tiling parameter. It points to a memory address in GM, where the data type is the TCubeTiling structure. For details about the definition of the TCubeTiling struct, see Table 1. The tiling parameter can be obtained through GetTiling on the host side and passed to the kernel side for use. On the kernel side, GET_TILING_DATA_PTR_WITH_STRUCT is called to obtain the pointer to the tiling parameter in GM. This API receives the tiling parameter by passing the GM address of the TCubeTiling structure.
tpipe	Input	Tpipe object.

Returns

None

Restrictions

API with the tiling parameter passed as a stack address:
None
API with the tiling parameter passed as a global memory address:
- This API is supported in only the partial constant quantization scenario of Matmul tiling parameters.
- CPU domain debugging is not supported.

Example

Tiling parameter passed as a stack address:

GET_TILING_DATA(tilingData, tiling);
// ...
REGIST_MATMUL_OBJ(&pipe, GetSysWorkSpacePtr(), mm);
mm.Init(&(tiling.cubeTilingData));

Tiling parameter passed as a global memory address:

CUBE-ONLY mode

#define ASCENDC_CUBE_ONLY

GET_TILING_DATA_PTR_WITH_STRUCT(MatmulCustomTilingData, tilingDataPtr, tiling);
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIC_ONLY);
// ...
REGIST_MATMUL_OBJ(&pipe, GetSysWorkSpacePtr(), mm);
mm.Init(&(tilingDataPtr->cubeTilingData));

MIX mode

GET_TILING_DATA_PTR_WITH_STRUCT(MatmulCustomTilingData, tilingDataPtr, tiling);
KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_MIX_AIC_1_2);
// ...
// In MIX mode, only the REGIST_MATMUL_OBJ interface is called to pass the GM address of the tiling parameter. The Init interface does not need to be called.
REGIST_MATMUL_OBJ(&pipe, GetSysWorkSpacePtr(), mm, &(tilingDataPtr->cubeTilingData));

Parent topic: Matmul Kernel APIs