cudaMemcpyToSymbol

Accelerated Computing CUDA CUDA Programming and Performance
1

I’m trying to use

CUDA_SAFE_CALL( cudaMemcpyToSymbol(d_A, h_A, mem_size_A) );

to load data from host to constant memory. But the timing I get is much much larger than when I use

CUDA_SAFE_CALL(cudaMemcpy(d_A, h_A, mem_size_A,cudaMemcpyHostToDevice));

to load the same data from host to device.

Am I doing anything wrong here Or is it supposed to be this way?

Thanks.

2

If you manage to get the memory address if the constant you can just use the second. But last time I tried this wasn’t possible when using the Cuda runtime, only when using the API directly External Image

3

My main question is: Is writing to constant memory slower than writing to global memory?

4

What do you mean with slower?

It can be that the memCpyToSymbol is the first cuda method and the card need to be initialized this can cause a large slowdown. I’m using cudaMemcpyToSymbol like this:

CUDA_SAFE_CALL(cudaMemcpyToSymbol("const_d_a", h_a, sizeof(h_a)));

Don’t pin me down on it doing this out of my head.

I know accessing Global memory is a little slower than Constant.

But I don’t know about initializing.

5

It should be the same speed…

6

Constant memory size is limited to 64 kilobytes so I don’t see how slower memcpy can affect overall performance. And I don’t see why copying to const space should be slower than copying to global space: const is just global with some caching.