Whats the max value for blockIdx.x for a <<<16,32>>> call config?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello ,

suppose we are using <<<16,32>>> call config for a kernal function, then is the below table correct?

     [u] variable[/u]       -------------------                                             [u]max value[/u]

        blockIdx.x        ------------------------                                                       15

        blockDim.x        -----------------------                                                       32

        threadIdx.x         ------------------                                                   31

Because, Im using above config and the kernal is executing 16*32 times in Emulation modes.

but in Debug and Release modes, it is executing 2 blocks ( some times 5 blocks and some times 8 blocks ).

Why is it so?

2

How do you measure the number of blocks that execute?

3

yes thats right. (but blockDim isn’t really a max value)

how do you know that?

or better how do you count them?

its hard to say why some of your thread blocks aren’t working.

I am pretty sure that they are executed, but perhaps they do nothing.

i can only guess what is happening

4

foo<<<16,32>>>( …, int* count );

global void foo( …, int* count )

{

…

*count = blcokIdx.x;

…

}

Now, print count variable.

5

All blocks are run in parallel, so you have a race condition. Count could randomly be any value from 0 to blockDim.x-1.

6

this does not work.

because it depends on which thread was the last to write to this variable.

so its random which value count has.

better try something like this:

if (threadIdx.x==0) *count++;

this means only thread with number 0 of each block increases the count variable.

so at the end the result should be the number of blocks.

and of course it would be better to use an atomic add

7

This still has a race condition between blocks and may not give you the correct answer.

Not just better, atomicAdd is the only way to do this correctly. :)

8

Can you give me some examples how to use “atomicAdd” please?

9

have a look at the programming guide (appendix C).

an example:

int count;

atomicAdd(&count,1);

this is the same like count++, but it is done as an atomic operation.

or a more general example:

x = x + y;

int x;

int y;

atomicAdd(&x, y);