If I could use SIMD process ProtoBuf RepeatedField data directly,or is there a way to parse RepeatedField to vector efficiently?

Question

I use ProtoBuf RepeatedField<uint32_t> store my 64 bit uint vector.Now I want to get the sum or product of two RepeatedField,and I want to use SIMD to accelerate the speed.

If I could access the RepeatedField's memory,just like a vector,to finish it? is there a efficient way to copy the data from RepeatedField to vector.Then I will use SIMD at vectors.

My code report a heap-buffer-overflow at first iteration of _mm256_loadu_si256.

Parameter papply(Parameter& a, Parameter& b) {
    Parameter ret;

    std::vector<uint32_t> tmp(8);

    for(int i = 0;i < a.datas_size();i += 8) {

        __m256i ma = _mm256_loadu_si256((__m256i*) a.mutable_datas()->Mutable(i));
        __m256i mb = _mm256_loadu_si256((__m256i*) b.mutable_datas()->Mutable(i));

        _mm256_add_epi32(ma,mb);
        _mm256_storeu_si256((__m256i*) &tmp[0], ma);
        ret.mutable_datas()->Add(tmp.begin(),tmp.end());
    }
}

If I could access the RepeatedField's memory,just like a vector,to finish it? is there a efficient way to copy the data from RepeatedField to vector.Then I will use SIMD at vectors.

If I could use SIMD process ProtoBuf RepeatedField data directly,or is there a way to parse RepeatedField to vector efficiently?

0 Answers0