Compute capability 2.1 - no. of banks

[thecoder3]

Hey guys.
My CUDA compute capability is 2.1 and i read that this version has a 32 bank shared memory and 32 warp size but i cant find anywhere whether it is executing simultaneously in half warps or warps.

My second question is: suppose i have 1 block in my code with 32 x 32 threads and i declare a shared memory in kernel as:

shared int As[32][32].

Now, in what order are the threads going to execute the kernel since i want to achieve bank conflicts so are they going to do: (imagine 32 at the same time)

First: ty=0, tx: 0 -31.
Second: ty =1, tx: 32 - 63

etc

or maybe in reverse order or at all at random. I want to know this since i want to avoid bank conflicts with shared memory and if i do:

As[tx][ty] - and they execute as shown first above then they all access same bank!

Thanks for any help!!!

[seibert]

My second question is: suppose i have 1 block in my code with 32 x 32 threads and i declare a shared memory in kernel as:

shared int As[32][32].

Now, in what order are the threads going to execute the kernel since i want to achieve bank conflicts so are they going to do: (imagine 32 at the same time)

First: ty=0, tx: 0 -31.

Second: ty =1, tx: 32 - 63

The CUDA Programming Guide Section 2.2 gives the mapping from 2D or 3D thread indexing to thread ID. For the 2D case, it is threadIdx.x + blockDim.x * threadIdx.y. Threads are then grouped into warps by consecutive thread ID.

[thecoder3]

Ok, so if i write out on paper the thread IDs from 2D to 1D and divide it into 32 chunks (warps), they will be executed in such order simultaneously. Is that correct?