Archives Discussions

[spectral]

Hi,

I have an array of float4 values and I would like to use atomic_add with.

The problem is that it is illegal to get the address of .x .y .z of the float4 and so it is impossible to use atomic_add.

Is there a solution ?

Thanks

[rick_weber]

You can't even atomically add scalar floating point numbers together. It would simple be too expensive to build this feature from a hardware perspective.

[himanshu_gautam]

I suggest to create copy of float4 in 4 floats and do atomic_add.

Also check out the atomics_counter extension, if it is useful in your case.

[Edit] notzd, Thanks for the correction.

[notzed]

Originally posted by: himanshu.gautam I suggest to create copy of float4 in 4 floats and do atomic_add.

Also check out the atomics_counter extension, if it is useful in your case.

How exactly do you atomic add with 4 floats?

And hou could an atomic_counter extension possibly be any help to working with floating point addition?

[rick_weber]

Atomic operations only apply to integers. You'll have to use a reduction to add floating point numbers in parallel.