[Feature Request]: atomicAdd() to support half2

[ZJLi2013]

Suggestion Description

hi, hip team,

here is cuda version,

void atomic_add_gmem_h2(half2* addr, half2 in) {
	atomicAdd(addr, in);
}

looks there's non hip alternative yet, if built with hipcc, it gives:

/opt/rocm-6.1.3/include/hip/amd_detail/amd_hip_atomic.h:216:5: note: candidate function not viable: no known conversion from 'half2 *' (aka '__half2 *') to 'int *' for 1st argument
  216 | int atomicAdd(int* address, int val) {
      |     ^         ~~~~~~~~~~~~
/opt/rocm-6.1.3/include/hip/amd_detail/amd_hip_atomic.h:228:14: note: candidate function not viable: no known conversion from 'half2 *' (aka '__half2 *') to 'unsigned int *' for 1st argument
  228 | unsigned int atomicAdd(unsigned int* address, unsigned int val) {
      |              ^         ~~~~~~~~~~~~~~~~~~~~~
/opt/rocm-6.1.3/include/hip/amd_detail/amd_hip_atomic.h:240:15: note: candidate function not viable: no known conversion from 'half2 *' (aka '__half2 *') to 'unsigned long *' for 1st argument
  240 | unsigned long atomicAdd(unsigned long* address, unsigned long val) {
      |               ^         ~~~~~~~~~~~~~~~~~~~~~~
/opt/rocm-6.1.3/include/hip/amd_detail/amd_hip_atomic.h:252:20: note: candidate function not viable: no known conversion from 'half2 *' (aka '__half2 *') to 'unsigned long long *' for 1st argument
  252 | unsigned long long atomicAdd(unsigned long long* address, unsigned long long val) {
      |                    ^         ~~~~~~~~~~~~~~~~~~~~~~~~~~~
/opt/rocm-6.1.3/include/hip/amd_detail/amd_hip_atomic.h:264:7: note: candidate function not viable: no known conversion from 'half2 *' (aka '__half2 *') to 'float *' for 1st argument
  264 | float atomicAdd(float* address, float val) {
      |       ^         ~~~~~~~~~~~~~~
/opt/rocm-6.1.3/include/hip/amd_detail/amd_hip_atomic.h:290:8: note: candidate function not viable: no known conversion from 'half2 *' (aka '__half2 *') to 'double *' for 1st argument
  290 | double atomicAdd(double* address, double val) {

Operating System

Ubuntu 22.04

GPU

mi300

ROCm Component

6.1.3 + rocblas + rocwmma

[cjatin]

For half, we have unsafeAtomicAdd instead of atomicAdd.

https://github.com/ROCm/clr/blob/aa6d07518fdb211c49fd617ee9f69408f1acddfd/hipamd/include/hip/amd_detail/amd_hip_fp16.h#L1511

[ZJLi2013]

For half, we have unsafeAtomicAdd instead of atomicAdd.

https://github.com/ROCm/clr/blob/aa6d07518fdb211c49fd617ee9f69408f1acddfd/hipamd/include/hip/amd_detail/amd_hip_fp16.h#L1511

is there any risk concern for unsafeAtomicAdd, just wonder in which way it's unsafe

[b-sumner]

Its unsafe because it causes the fast HW instruction to be generated, but those instructions don't work if they act on memory that is not cached, e.g. across a PCIe bus. The developer needs to assert that they are willing to take that risk.

[jinz2014]

Does ROCm 6.2 support it ?

/opt/rocm-6.2.0/lib/llvm/bin/../../../include/hip/amd_detail/amd_hip_fp16.h does not contain the function.

[jinz2014]

Do you think it is better to have two types of atomic add functions than a single function in CUDA ?

[tcgu-amd]

Hi @jinz2014, the function is recently added part of this. You can see it on staging but it hasn't made to 6.2.2 release yet, but please keep an eye out for it since it should be out soon. Meanwhile if you are interested, you can try to compile and install clr staging to see how it works.

Do you think it is better to have two types of atomic add functions than a single function in CUDA ?

In our case we wanted to ensure developers assert that they are using a unsafe function. However, I see your point in having a uniform interface to access these functions. We can will bring it up internally for discussion for sure. Thanks!

[tcgu-amd]

Hi @jinz2014, to follow up on the discussion, in our opinion, it is best to have two versions of the atomic functions because atomic* and unsafeAtomic* APIs behave differently, hence the distinction to avoid confusion. We think this aspect outweigh the syntax discrepancy with CUDA, and should be explicitly noted by developers that use these APIs. Thanks!

[jinz2014]

Hi @tcgu-amd Thanks for your answers.

[tcgu-amd]

@jinz2014 No problem! Is there anything else we can help you with? If not I will close this issue for now. Thanks!

[jinz2014]

I will try the atomicAdd() for half data types. I am not sure if @ZJLi2013 has more questions.

[ZJLi2013]

hi, @tcgu-amd , I tried rocm/torch images with rocm6.2.3, rocm6.2.4, neither has unsafeAtomicAdd yet.

anther thing may need you clarify, looks in /opt/rocm/include, there is no clr headers, only found: /opt/rocm/include/./hip/amd_detail/amd_hip_bf16.h , supposing clr headers has merged to hip header dir?

will wait on a official release then

Thanks again

[tcgu-amd]

anther thing may need you clarify, looks in /opt/rocm/include, there is no clr headers, only found: /opt/rocm/include/./hip/amd_detail/amd_hip_bf16.h , supposing clr headers has merged to hip header dir?

Yes, that's correct. There is no clr directory under /opt/rocm, since the repository mainly hosts runtimes for hip and OpenCL. The amd_detail headers from clr are indeed under /opt/rocm/include/hip/amd_detail.

will wait on a official release then

Sounds good.

Thanks!

[tcgu-amd]

This issue will be closed since there is no further actionable item/activity. Please feel free to re-open for follow ups and further inquires requiring the release status. Thanks!

[jinz2014]

@tcgu-amd
Could you close the issue after the feature is available in the release ?

[tcgu-amd]

@jinz2014 sure! Just a note that we would prefer closing issues with no further actionable items to help better track which ones still require active attention. Since the feature requested specifically for this issue is already in staging, we can keep it open a little longer. However, in general, feature requests that are on internal roadmaps will be closed, but we do encourage users to poll for progress by continue to ask follow up questions. Hope this makes sense. Thanks!