[Priority] Medium

[Description] In Ascend C operator projects, the workspace parameter passed when writing the kernel function has been directly assigned the value Workspace. Therefore, you do not need to use SetSysWorkspace and GetUserWorkspace to set and obtain the workspace. After these redundant judgments are reduced, the compiler can further optimize unused workspace variables without using this parameter.

[Negative Example]

The workspace parameter in the fast_gelu function is equivalent to user workspace and is not empty. However, the system still checks whether the workspace is empty and sets SetSysWorkspace and GetUserWorkspace to obtain the user workspace.

template <uint64_t schMode, uint64_t dType>
__global__ __aicore__ void fast_gelu(GM_ADDR x, GM_ADDR y, GM_ADDR workspace, GM_ADDR tiling)
{
    // This is a negative example for redundant judgment.
    if (workspace == nullptr) {
        return;
    }
    SetSysWorkspace(workspace);
    GM_ADDR userWS = GetUserWorkspace(workspace);
    if (userWS == nullptr) {
        return;
    }
    REGISTER_TILING_DEFAULT(EleBaseTilingDataV2);
    GET_TILING_DATA_WITH_STRUCT(EleBaseTilingDataV2, tilingData, tiling);
    KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIV_ONLY);
    TPipe pipe;
    if constexpr (dType == static_cast<uint64_t>(TPL_FP16)) {
        ElementwiseSch<schMode, FastGeluDag::FastGeluNeedCast<half>::OpDag> sch(&tilingData, &pipe);
        sch.Init(x, y);
        sch.Process();
    } else if constexpr (dType == static_cast<uint64_t>(TPL_BF16)) {
        ElementwiseSch<schMode, FastGeluDag::FastGeluNeedCast<bfloat16_t>::OpDag> sch(&tilingData, &pipe);
        sch.Init(x, y);
        sch.Process();
    } else if constexpr (dType == static_cast<uint64_t>(TPL_FP32)) {
        ElementwiseSch<schMode, FastGeluDag::FastGeluNoCast<float>::OpDag> sch(&tilingData, &pipe);
        sch.Init(x, y);
        sch.Process();
    }
}

[Positive Example]

The null pointer check for the workspace parameter is deleted from the fast_gelu function. In addition, you do not need to set SetSysWorkspace or use GetUserWorkspace to obtain the workspace.

template <uint64_t schMode, uint64_t dType>
__global__ __aicore__ void fast_gelu(GM_ADDR x, GM_ADDR y, GM_ADDR workspace, GM_ADDR tiling)
{
    REGISTER_TILING_DEFAULT(EleBaseTilingDataV2);
    GET_TILING_DATA_WITH_STRUCT(EleBaseTilingDataV2, tilingData, tiling);
    KERNEL_TASK_TYPE_DEFAULT(KERNEL_TYPE_AIV_ONLY);
    TPipe pipe;
    if constexpr (dType == static_cast<uint64_t>(TPL_FP16)) {
        ElementwiseSch<schMode, FastGeluDag::FastGeluNeedCast<half>::OpDag> sch(&tilingData, &pipe);
        sch.Init(x, y);
        sch.Process();
    } else if constexpr (dType == static_cast<uint64_t>(TPL_BF16)) {
        ElementwiseSch<schMode, FastGeluDag::FastGeluNeedCast<bfloat16_t>::OpDag> sch(&tilingData, &pipe);
        sch.Init(x, y);
        sch.Process();
    } else if constexpr (dType == static_cast<uint64_t>(TPL_FP32)) {
        ElementwiseSch<schMode, FastGeluDag::FastGeluNoCast<float>::OpDag> sch(&tilingData, &pipe);
        sch.Init(x, y);
        sch.Process();
    }
}

Deleting Redundant Workspace Operations from the Kernel Function