Question

我编写了CUDA代码。

unsigned long mask_buffer;
int s;
off_t p,

for(p=0;p!=5000;p++)
{
    for(s=start;s!=end;s++)
    {
        ref_off = *(((unsigned int*)(idx_base)) + p);

        if((int)(first_indexes[s-start_sequence] % 8 - ref_off % 8) < 0)
        {
            int shamt2 = (ref_off % 8 - first_indexes[s-start_sequence] % 8);
            mask_buffer = *((unsigned long *)(msk_base + (ref_off - first_indexes[s-start_sequence])/8)) >> shamt2;

            if( ( (*(unsigned long *)(seqmaskc + 16 * (s-start_sequence))) ^ mask_buffer ) << shamt2) 
                continue;
        }

        else if((int)(first_indexes[s-start_sequence] % 8 - ref_off % 8) == 0)
        {
            mask_buffer = *((unsigned long *)(msk_base + (ref_off)/8));

            if( (*(unsigned long *)(seqmaskc + 16 * (s-start_sequence)) ^ mask_buffer))
                continue;
        }

        else
        {
            int shamt2 = 8 - (first_indexes[s-start_sequence] % 8 - ref_off % 8);
            mask_buffer = *((unsigned long *)(msk_base + (ref_off/8- first_indexes[s-start_sequence]/8) - 1)) >> shamt2;

            if( ( (*(unsigned long *)(seqmaskc + 16 * (s-start_sequence))) ^ mask_buffer ) << shamt2) 
                continue;
        }

        int shamt = (ref_off % 4 - first_indexes[s-start_sequence] % 4) * 2;

        memcpy(reference_blk, ref_base + ref_off / 4 - first_indexes[s-start_sequence] / 4, sequence_bytes);

        for (rp = last_rp ; rp != (unsigned long *) reference_blk ; rp--) 
        {
            unsigned long tmp = ((*rp) & ((1 << shamt) - 1)) << (8 * sizeof(unsigned long) - shamt);
            *rp = (*rp >> shamt) | shifted_in;
            shifted_in = tmp;
        }

        *rp = (*rp >> shamt) | shifted_in;

        if (sequence_length & 0x3)
            reference_blk[sequence_length >> 2] &= (1 << ((sequence_length & 0x3) << 1)) - 1;

        for ( i = sequence_length >> 2 ; i & (SEQUENCE_ALIGN - 1) ; i++ )
            reference_blk[i] = 0;

        //-- instead of memcmp --//
        int v = 0;
        char *p1 = (char *)sequence;
        char *p2 = (char *)reference_blk;
        int tmp_asd = sequence_bytes;

        while(tmp_asd!=0)
        {
            v = *(p1++) - *(p2++);

            if(v!=0)
                break;

            tmp_asd--;
        }

        if(v == 0)
        {
            mat_count[s - (int)start_sequence]++;      /* Maintain count */
            mat_position[s - (int)start_sequence] = ref_off-first_indexes[s-start_sequence]; /* Record latest position */
        }

    }
}

这个for循环是我的代码的主要功能。但问题是变量“p”从未超过5或6。我的计算机上有GT530，我的CUDA驱动程序版本和运行时版本也是4.0。这段代码有什么问题???

Answer 1

假设这确实是一个内核，在你的新代码编辑之前你的原始帖子声称它是，你不能在CUDA内核中调用标准的C库函数，如memcpy（你这样做大约一半 - 通过代码）。您只能在内核中调用__device__个函数。因此，除非这是在CUDA中重新实现memcpy，否则在内核中调用该函数是行不通的......

此外，如果您要尝试编写需要检查组中其他线程编写结果的CUDA内核，例如您在内核的内存比较部分进行的操作，则需要在代码中放置一些同步点，以确保在开始检查这些线程的结果之前所有线程都已到达某个点。如果您正在尝试检查另一个块可能写入主内存的内容，那么还有同步原语，以确保任何先前线程块的结果已写入主内存。

CUDA没有按预期工作？

1 个答案: