Still unclear on 16-bit float atomic operations for consumer Pascal GPUs

Accelerated Computing CUDA CUDA Programming and Performance
#1

Is there any indication if there will be 16-bit atomic operations (preferably an atomicAdd()) on either the ‘half’ type as a float point value or as a 16-bit integer) ? This would be for shared or global memory (hopefully both but I will be happy with either possibility).

Made my own 16-bit unsigned int atomicAdd() hack for shared memory which I am currently using for a real-time image reconstruction, but it is not as efficient as a 32-bit atomic operation.

Even if it is not hardware supported I would guess that a NVIDIA version of a 16-bit atomicAdd() would be better than my ‘rolled-my-own’ version.

Heard rumors that this may be available in some inefficient form for the GTX line, but cannot find any documentation.

Maybe the final release of CUDA 8?

#2

These have existed since sm_52:

RED.E.ADD.F16x2.FTZ.RN
ATOM.E.ADD.F16x2.FTZ.RN

I don’t think the shared version ATOMS.ADD.F16x2.FTZ.RN exists. I have no idea why nvidia has chosen not to expose these in cuda or ptx as of yet. I use them from time to time in my sass programming.

Maybe submit a bug report. Or perhaps cuda-8 final will finally add support.