tf.raw_ops.ScatterSub(
    ref, indices, updates, use_locking=False, name=None
)

// 最后一维是对其的情况处理
uint32_t firstTiling = lastDim;
if (lastDim * sizeOfDataType % BLOCK_SIZE == 0) {
    // bufferSize是初始化的Buffer数量
    uint32_t bufferSize = 6;
    
    // Int8类型比其他类型多了两块
    if (tilingKey == 4) {
        bufferSize = 10;
    }

    // Ub大小如果大于Buffer区域大小则直接按照lastDim分块
    // 否则，则按照32B依次减小，知道满足Ub大小
    if (ub_size > bufferSize * lastDim * sizeOfDataType) {
        firstTiling = lastDim;
    } else {
        while(ub_size < bufferSize * firstTiling * sizeOfDataType) {
            firstTiling -= 32 / sizeOfDataType;
        }
    }
    tilingKey += 10;
}

// half转float
int16_t varInt16 = *(int16_t*)&var;
int32_t fltInt32    =  ((varInt16 & 0x8000) << 16);
fltInt32        |= ((varInt16 & 0x7fff) << 13) + 0x38000000;
float varFp32 = *(float*)&fltInt32;

int16_t updateInt16 = *(int16_t*)&update;
fltInt32    =  ((updateInt16 & 0x8000) << 16);
fltInt32        |= ((updateInt16 & 0x7fff) << 13) + 0x38000000;
float updateFp32 = *(float*)&fltInt32;
float res = varFp32 - updateFp32;

// float转half
int16_t fltInt16;
fltInt32 = *(int32_t*)&res;
fltInt16    =  ((fltInt32 & 0x7fffffff) >> 13) - (0x38000000 >> 13);
fltInt16    |= ((fltInt32 & 0x80000000) >> 16);
half resHalf = *(half*)&fltInt16;