Question

I am trying to solve this problem myself but I can’t.
So I want to get yours advice.

I am writing kernel code like this. VGA is GTX 580.

xxxx <<< blockNum, threadNum, SharedSize >>> (... threadNum ...)
(note. SharedSize is set 2*threadNum)

__global__ void xxxx(..., int threadNum, ...)
{
    extern __shared__ int shared[];
    int* sub_arr = &shared[0];
    int* sub_numCounting = &shared[threadNum];
    ...
}

My program creates about 1085 blocks and 1024 threads per block (I am trying to handle a huge array).

So the size of shared memory per block is 8192 (1024*2*4) bytes, right?

I figured out that I can use maximum 49152 bytes in shared memory per block on GTX 580 by using cudaDeviceProp.

And I know that the GTX 580 has 16 processors, thread block can be implemented on processor.

But my program returns an error (8192 bytes < 49152 bytes).

I used printf in the kernel to see whether it operates well or not but several blocks do not operate: Even though I create 1085 blocks, actually only 50~100 blocks operate.

And I want to know whether blocks which operate on the same processor share the same shared memory address or not. If not, do they allocate other memory for shared memory?

I can’t quite understand what maximum size of shared memory per block means.

Give me advice.

Answer 1

Yes, blocks on the same multiprocessor shared the same amount of shared memory, which is 48KB per multiprocessor for your GPU card (compute capability 2.0). So if you have N blocks on the same multiprocessor, the maximum size of shared memory per block is (48/N) KB.