Any update on 5700 Xt support?

[devnarekm]

Just curious if you have any plans on supporting the 5700 Xt card in the near future. As of now I haven't seen it in the supported GPUs list. Would be nice to also have it included!

[kentrussell]

There are different levels of support in ROCm, since it's an entire stack, which makes this a difficult question to answer.

1-Kernel support. This usually comes pretty quickly from upstream, and is already there in its infancy. The basic code is there, and there is work going on to thoroughly test the kernel functionality 2-Thunk support. This is usually pretty quick too, since it's mostly just adding the chip information, and then adding any weird quirks that the HW has (shader engine distribution, SDMA queues, etc) 3-Runtime support. This usually takes longer as there is a lot of testing to go on 4-Official ROCm support - Once we have 1-3 supported, we need to run the full gambit of tests and applications against the stack. This requires a lot of time, and a lot of bug-fixes. Once this is done, that's when we add it to the Supported GPUs list.

So there may be some partial support for it right now, or it might work almost perfectly. It's all a bit of a crapshoot until we update the documentation, as we haven't tested everything to fulfill our exit criteria for supporting a GPU. You can always try it out, and post PRs to help to support them in the interim, though. We always like help from the community!

[Moading]

Hi, in my experience GPUs listed as "supported" have no guarantee to work properly. I have opend a few issues here and the speed at which theses issues are fixed is extremely disappointing. The most annoying thing is missing OpenCL 2.0 support for hardware that is still beeing sold (gfx803).

I have zero confidence that the latest hardware will have OpenCL 2.0 support in ROCm, therefore I am not buying new hardware. There would be more business for AMD if the drivers delivered better OpenCL support.

[EwoutH]

@kentrussell Could you create a new issue for us to track Navi integration? You could create a task list to mark the process:

• [ ] 1-Kernel support
• [ ] 2-Thunk support
• [ ] 3-Runtime support
• [ ] 4-Official ROCm support

- [ ] 1-Kernel support
- [ ] 2-Thunk support
- [ ] 3-Runtime support
- [ ] 4-Official ROCm support

[onfoot]

I'm new here, and don't know anything about anything, so it's perfectly fine that I chip in. ;) I'm interested in ROCm since I'm currently buying the 5700 XT for myself and wanted, aside from gaming, to play around with PyTorch and Tensorflow.

Seems like kernel support has been merged in on Sep 23rd, Thunk has support for Navi 10 since July, so we're getting there.

[EwoutH]

@aak-amd @zhang2amd @Rmalavally @kentrussell Any updates about Navi/RDNA support?

[mritunjaymusale]

Bump. Still wondering if the support is there or not?

998

[jemzipx]

It is kind of sad that AMD has forgotten Navi users. I have found a temporary solution to use my RX 5700XT for deep learning with rather astonishing results. I thought I should share it here. Using Linux kernel 5.6rc and AMDGPU PRO driver I was able to set up OpenCL 2.0 on Manjaro Linux. Then I installed PlaidML which supports opencl devices. After that, I set Keras backend to PlaidML and just used Keras. I ran some benchmarks and the results are just amazing. It outperforms my 12C/24T Ryzen 3900 CPU by a massive margin. While it takes more than 12 minutes to train MobileNet on Ryzen 3900X, it takes less than a minute on Radeon 5700XT. Here are the Inference latency and Time/FPS comparison in mobilenet benchmark:

RX 5700 XT GPU:
-----------------------------------------------------------------------------------------
Network Name         Inference Latency         Time / FPS          
-----------------------------------------------------------------------------------------
mobilenet                  22.00 ms                  19.68 ms / 50.80 fps

Ryzen 3900X CPU:
-----------------------------------------------------------------------------------------
Network Name         Inference Latency         Time / FPS          
-----------------------------------------------------------------------------------------
mobilenet                  695.84 ms                 695.56 ms / 1.44 fps

And for those who prefer to see a working example, here is a CNN trained on MNIST in Keras using RX 5700XT:

[FiCacador]

ROCm-OpenCL-Runtime 3.1 changes introduced ROC_GFX10, that's Navi! It still isn't listed as supported on the official documentation, but neither is any non deprecated Linux kernel... Has anyone already tried 3.1 with Navi?

[jemzipx]

The Navi support seems to be a work in progress. There are no clear instructions on how to compile ROCm from scratch anyway. I downloaded all their sources codes using Google Repo tool and there are at least 40 different projects to compile separately! I couldn't find any easy way to automate the building process. There is a project that is supposed to automate the build process but has not been updated in the past few years. If the guys at AMD can illuminate us on how to automate ROCm compilation, it would be greatly appreciated.

[lrie]

There is no or no complete Navi support in ROCM 3.1 . Tested with Radeon 5700XT, ROCM 3.1 and Tensorflow 2.1

` gpu-user@gpu-server:~/benchmarks/scripts/tf_cnn_benchmarks$ python3 tf_cnn_benchmarks.py --num_gpus=1 --batch_size=64 --model=resnet50 WARNING:tensorflow:From /home/gpu-user/.local/lib/python3.6/site-packages/tensorflow_core/python/compat/v2_compat.py:88: disable_resource_variables (from tensorflow.python.ops.variable_scope) is deprecated and will be removed in a future version. Instructions for updating: non-resource variables are not supported in the long term 2020-03-10 19:45:39.130116: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE3 SSE4.1 SSE4.2 AVX AVX2 FMA 2020-03-10 19:45:39.155823: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 3099670000 Hz 2020-03-10 19:45:39.156223: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x4ed3e10 initialized for platform Host (this does not guarantee that XLA will be used). Devices: 2020-03-10 19:45:39.156330: I tensorflow/compiler/xla/service/service.cc:176] StreamExecutor device (0): Host, Default Version 2020-03-10 19:45:39.159225: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libhip_hcc.so 2020-03-10 19:45:39.193370: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1573] Found device 0 with properties: pciBusID: 0000:2b:00.0 name: Device 731f ROCm AMD GPU ISA: gfx1010 coreClock: 0.1GHz coreCount: 40 deviceMemorySize: 7.98GiB deviceMemoryBandwidth: -1B/s 2020-03-10 19:45:39.235125: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library librocblas.so 2020-03-10 19:45:39.236629: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libMIOpen.so 2020-03-10 19:45:39.237930: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library librocfft.so 2020-03-10 19:45:39.238294: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library librocrand.so 2020-03-10 19:45:39.238463: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1697] Adding visible gpu devices: 0 2020-03-10 19:45:39.238596: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1096] Device interconnect StreamExecutor with strength 1 edge matrix: 2020-03-10 19:45:39.238658: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1102] 0 2020-03-10 19:45:39.238695: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] 0: N 2020-03-10 19:45:39.238890: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1241] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 7524 MB memory) -> physical GPU (device: 0, name: Device 731f, pci bus id: 0000:2b:00.0) TensorFlow: 2.1 Model: resnet50 Dataset: imagenet (synthetic) Mode: training SingleSess: False Batch size: 64 global 64 per device Num batches: 100 Num epochs: 0.00 Devices: ['/gpu:0'] NUMA bind: False Data format: NCHW Optimizer: sgd Variables: parameter_server

Generating training model WARNING:tensorflow:From /home/gpu-user/benchmarks/scripts/tf_cnn_benchmarks/convnet_builder.py:134: conv2d (from tensorflow.python.layers.convolutional) is deprecated and will be removed in a future version. Instructions for updating: Use tf.keras.layers.Conv2D instead. W0310 19:45:39.282001 139982497593152 deprecation.py:323] From /home/gpu-user/benchmarks/scripts/tf_cnn_benchmarks/convnet_builder.py:134: conv2d (from tensorflow.python.layers.convolutional) is deprecated and will be removed in a future version. Instructions for updating: Use tf.keras.layers.Conv2D instead. WARNING:tensorflow:From /home/gpu-user/.local/lib/python3.6/site-packages/tensorflow_core/python/layers/convolutional.py:424: Layer.apply (from tensorflow.python.keras.engine.base_layer) is deprecated and will be removed in a future version. Instructions for updating: Please use layer.__call__ method instead. W0310 19:45:39.284210 139982497593152 deprecation.py:323] From /home/gpu-user/.local/lib/python3.6/site-packages/tensorflow_core/python/layers/convolutional.py:424: Layer.apply (from tensorflow.python.keras.engine.base_layer) is deprecated and will be removed in a future version. Instructions for updating: Please use layer.__call__ method instead. WARNING:tensorflow:From /home/gpu-user/benchmarks/scripts/tf_cnn_benchmarks/convnet_builder.py:266: max_pooling2d (from tensorflow.python.layers.pooling) is deprecated and will be removed in a future version. Instructions for updating: Use keras.layers.MaxPooling2D instead. W0310 19:45:39.321026 139982497593152 deprecation.py:323] From /home/gpu-user/benchmarks/scripts/tf_cnn_benchmarks/convnet_builder.py:266: max_pooling2d (from tensorflow.python.layers.pooling) is deprecated and will be removed in a future version. Instructions for updating: Use keras.layers.MaxPooling2D instead. Initializing graph WARNING:tensorflow:From /home/gpu-user/benchmarks/scripts/tf_cnn_benchmarks/benchmark_cnn.py:2267: Supervisor.init (from tensorflow.python.training.supervisor) is deprecated and will be removed in a future version. Instructions for updating: Please switch to tf.train.MonitoredTrainingSession W0310 19:45:41.986801 139982497593152 deprecation.py:323] From /home/gpu-user/benchmarks/scripts/tf_cnn_benchmarks/benchmark_cnn.py:2267: Supervisor.init (from tensorflow.python.training.supervisor) is deprecated and will be removed in a future version. Instructions for updating: Please switch to tf.train.MonitoredTrainingSession 2020-03-10 19:45:42.372816: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1573] Found device 0 with properties: pciBusID: 0000:2b:00.0 name: Device 731f ROCm AMD GPU ISA: gfx1010 coreClock: 0.1GHz coreCount: 40 deviceMemorySize: 7.98GiB deviceMemoryBandwidth: -1B/s 2020-03-10 19:45:42.373062: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library librocblas.so 2020-03-10 19:45:42.373111: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libMIOpen.so 2020-03-10 19:45:42.373155: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library librocfft.so 2020-03-10 19:45:42.373201: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library librocrand.so 2020-03-10 19:45:42.373313: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1697] Adding visible gpu devices: 0 2020-03-10 19:45:42.373361: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1096] Device interconnect StreamExecutor with strength 1 edge matrix: 2020-03-10 19:45:42.373398: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1102] 0 2020-03-10 19:45:42.373431: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] 0: N 2020-03-10 19:45:42.373560: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1241] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 7524 MB memory) -> physical GPU (device: 0, name: Device 731f, pci bus id: 0000:2b:00.0) terminate called after throwing an instance of 'std::runtime_error' what(): No device code available for function: _ZN10tensorflow7functor28FillPhiloxRandomKernelLaunchINS_6random27TruncatedNormalDistributionINS2_19SingleSampleAdapterINS2_12PhiloxRandomEEEfEEEEvS5_PNT17ResultElementTypeExS8, for agent: gfx1010 Fatal Python error: Aborted `

[Ge0rges]

Any update on the status of this issue?

[SlausB]

I'm not a ROCm user, but I tried installing it just to get OpenCL working because as I found somewhere that it's how ROCm works on AMD devices: through OpenCL. So ROCm supposed to set OpenCL up, but failed in my case. Eventually, I got OpenCL working on Ubuntu 18.04, but I think it should work on any distro (only kernel version matters): https://askubuntu.com/questions/1209725/how-to-get-opencl-support-for-navi10-gpus-from-amd <-- that's how I managed to achieve it. So I think if you guys set OpenCL up on your setup, ROCm should also work properly with RX5700 (navi10-12) I hope it helps ^^

[lrie]

@SlavMFM As mentioned, I was able to install ROCm 3.1 and thus OpenCL. clinfo results look fine. So there is Navi support, but it is not really functional. Look at the error above: No device code available for function: This seems to point to some not fully implemented libraries or missing assembler snippets. A bit like a car without engine.

[SlausB]

Interesting. Was you able to run some raw OpenCL examples to verify it's working properly? I once had "fake" OpenCL setup where clinfo was printing even GPU name (Radeon RX5700xt), but actual opencl functions from within C++ wasn't working. I just can't imagine how ROCm could utilize AMD GPUs rather just through OpenCL, so it it's working, ROCm should work too, hmm...

[jemzipx]

@SlavMFM As far as I know, AMD provides two seperate OpenCL 2.0 implementations. One in ROCm and another in AMDGPU PRO driver. I have tested the OpenCL 2.0 in AMDGPU PRO and can confirm it works fine (see my post above). Although, getting OpenCL to work does not guarantee that ROCm would work too.

[SlausB]

@jemzipx cool! I'm surprised amdgpu-pro drivers can be actually installed ^^ - you just get kernel 5.6 and install amdgpu-pro on top of it? Btw, there is at least another offician support: mesa drivers where AMD devs directly implement their GPUs support, for example navi14 was implemented half-a-year before GPUs release: https://lists.freedesktop.org/archives/mesa-dev/2019-August/222273.html - that's how I got OpenCL working in my case: with mesa-19.3 driver (which can stack on > kernel 5.4, but not 5.6).

[jemzipx]

@SlavMFM that is right. I installed amdgpu-pro on top of kernel 5.6rc in Manjaro (should work in Arch too). I tried Mesa before (actually that was the first thing I tried) but Mesa's OpenCL version was 1.2 which is quite old and kind of outdated. As far as I can tell, OpenCL 2.0 support for Navi can only be found in AMDGPU-PRO and ROCm.

[SlausB]

@jemzipx you think it'll work only on Manjaro and Arch and won't work on other distros like Ubuntu? Would like to try it on Ubuntu later. I'm currently having mesa-19.3 and it claims 2.1 platform and 2.0 device; not sure if it's practically 2.0 though.

[jemzipx]

@SlavMFM It should work on all distros. I have only tested Manjaro personally. I'm also curious to know how it plays on Ubuntu. BTW, have you tested Keras/PlaidML with your Mesa OpenCL on Ubuntu? Please give it a go and let us know the result.

[mritunjaymusale]

[Update] I was able to compile and install pytorch using the latest ROCm 3.1.1 using this guide. But I don't know how to install custom modules like DCNv2 that are C++ extensions of pytorch. There was some prompt I kept getting everytime I used pytorch, I can't remember, neither can I test it since I'm afk. All in all, things are looking better for Navi gpus, still no idea on when the official support will come through.

[Rmalavally]

AMD ROCm is validated for GPU compute hardware such as AMD Radeon Instinct GPUs. Other AMD cards may work, however, they are not officially supported at this time. We appreciate your feedback and we will consider it for future versions of ROCm.

Regards, Roopa

[FiCacador]

I don't think anyone here is asking for validation or even official support. All that Navi owners and potential buyers would like is for it to somehow work, now several months after it's release, just like it would work for them from day zero if they have got, say, a RTX 2070 using CUDA. People understand that the resources are limited, that it's open software, that it's a new architecture, that it's not focused on computing... but after so much time without support or even confirmation that it will ever come, it looks like individuals with interest in these kind of projects that don't require validation and also game on the side unfortunately have only one choice and it's not AMD. Which is understandable from a business perspective, only too bad for those customers to have less options on the market.

[foolnotion]

let's just call it AMD on the compute market is a joke at the moment and the only option for the consumer is to go nvidia.

[jemzipx]

@Rmalavally That is a very disappointing answer from an AMD's representative. The thing is that, Nvidia's success was due to the fact that every college gamer kid with their consumer-based GTX/RTX card could run deep learning algorithms as well. There are many of us that do not want to rely on Nvidia's propriety CUDA and would prefer to use open standards like OpenCL etc. This is a great business opportunity for AMD and is the one that matters more than designing graphic cards for next-gen xbox and playstation. We hope AMD change its approach to consumer-grade cards.

[valeriob01]

@Rmalavally That is a very disappointing answer from an AMD's representative. The thing is that, Nvidia's success was due to the fact that every college gamer kid with their consumer-based GTX/RTX card could run deep learning algorithms as well. There are many of us that do not want to rely on Nvidia's propriety CUDA and would prefer to use open standards like OpenCL etc. This is a great business opportunity for AMD and is the one that matters more than designing graphic cards for next-gen xbox and playstation. We hope AMD change its approach to consumer-grade cards.

This is not the first time we get disappointing answers from AMD, We ALL hope AMD change its approach to consumer-grade cards, and its approach to Radeon Instinct MI50 selling strategy for consumers too, not only for businesses.

[lrie]

Well let's face it: Jen-Hsun Huang said: "Nvidia is a Software Company," If you look at their stock chart, that may have been a wise decision.

If AMD decides to put all their effort on fancy 7nm tech, but does not want to spend the money for software developers, we just have to swallow that pill and move on.

[valeriob01]

Well let's face it: Jen-Hsun Huang said: "Nvidia is a Software Company," If you look at their stock chart, that may have been a wise decision.

If AMD decides to put all their effort on fancy 7nm tech, but does not want to spend the money for software developers, we just have to swallow that pill and move on.

it would be a bitter pill to swallow however...

[jemzipx]

They released ROCm 3.3.0 today but (as we all expected) there is no official support for Navi yet :( All it takes for them is to hire someone to work on this full-time for a few weeks but apparently AMD has more pressing issues at the moment. Thank you AMD for ignoring your loyal supporters.

[marcelocf]

PLEASE do complete navi10 support! I bought this card just because of opensource support (bought in release day). Not having it feels like I paid for the full price of the card, but I can only use half of it.

[himanshugoel2797]

Yes, I agree. Please support navi10. It's absurd and disappointing that the latest and current flagship hardware still doesn't support rocm 6+ months after release and doesn't even have a timeline for it. At this rate might just have to begrudgingly go with Nvidia next generation.

It doesn't need to be support in the workstation cards sense, but atleast in the "you can use this card for ROCm applications" sense.

[Dantali0n]

Hello, I think the only reasonable way we could expect widespread adoption of ROCm if it is supported by all AMD cards just like Nvidia's CUDA. The era of GPGPU compute is shifting more and more applications and developers are realizing when not gaming these cards can significantly improve the performance of many applications. ROCm can only be relied upon by developers if it is supported by all cards otherwise they would be better of writing OpenCL or something that generates to SPIR-V and than compiling to vulkan or similar.

By not supporting all cards you are effectively wasting the entire developing effort of ROCm and all the time of the employees and contributors who worked on it. Please, I encourage AMD to heavily reconsider their stance.

[Anton-Kyrychek]

@jemzipx Hey, I'm trying to make my rx5700xt work the same as you did(but on Ubuntu 18.04). I installed 5.7rc2 kernel, I have a question how did you install AMDGPU PRO? I find a lot of problems with dependencies when i tried to do it(which may or may not be connected to Ubuntu).

[jemzipx]

@Anton-Kyrychek I installed AMDGPU PRO through AUR (which is more specific to Arch Linux and it's derivatives like Manjaro). The installation was quite easy and the dependencies were resolved by Makepkg and Pacman automatically (nothing special was done on my side). I tried it only on 5.6 though. Not sure how it plays on 5.7rc.

[PhilipDeegan]

AMD ROCm is validated for GPU compute hardware such as AMD Radeon Instinct GPUs. Other AMD cards may work, however, they are not officially supported at this time. We appreciate your feedback and we will consider it for future versions of ROCm.

Regards, Roopa

You may want to update the docs: https://github.com/RadeonOpenCompute/ROCm

"Vega 10" chips, such as on the AMD Radeon RX Vega 64 and Radeon Instinct MI25

It does not suggest Navi support, but Vega is fine.

[Ge0rges]

@Anton-Kyrychek could you let us know if you are successful in using the RX 5700 XT card on ubuntu, and if so what steps you followed? Looking to use it with pytorch on ubuntu 19.

[Anton-Kyrychek]

@Ge0rges I was not, I'm not that smart with drivers.

[mritunjaymusale]

@Ge0rges try this method, this is what I have used with little success https://github.com/ROCmSoftwarePlatform/pytorch/issues/581

[Ge0rges]

@mritunjaymusale that worked for you on the 5700XT?

[mritunjaymusale]

@Ge0rges yes it worked for me.

[limapedro]

I've been following ROCm for a long time, I don't know a lot about the team behind, but it seems like AMD doesn't really care about us yet, GPU computing will be a huge part of the software stack, I see no reason to buy AMD GPUs at the moment, maybe they'll support Big Navi, unfortunately, AMD is years behind the competition in this area, imagine how many developers and consumers would benefit from AMD's adoption of GPU computing, anyway I'll continue to follow this thread and see if AMD change its ways in the near future.

[xinyazhang]

I really do not think this is the right strategy to market rocm by only focusing on Radeon Instinct.

The most common choices in my lab (and nearby labs) are GTX 1080Ti and 2080 RTX. The AI lab in our department has a cluster with Titan V GPUs, but nobody in my University, as far as I know, ever purchased the Tesla product line from NVIDIA for deep learning.

It's not difficult to realize the fact that deep learning researchers/developers always start with customer grade cards, and they may never consider the professional grade cards due to high costs. A professional-market-only approach will not bring new professional developers/researchers, ironically.

[Ge0rges]

@mritunjaymusale I've been able to compile pytorch with ROCm as per your link. However tests fail because HIP gives an hipErrorNoDevice error. Were you able to solve that?

[mritunjaymusale]

@Ge0rges I got the rocBlas error that's hinting to the device driver i believe this has to do with rocm not fully supporting the navi cards. I think we might have to compile rocm from scratch itself and see if it even accepts navi cards

[Ge0rges]

@mritunjaymusale Ah so it never worked for you I guess? I'll let you know if I have success compiling ROCm.

[Ge0rges]

I wonder if anyone has gotten Pytorch to work with RX 5700XT through OpenCL similarly to @jemzipx

[jemzipx]

@Ge0rges AFAIK, PyTorch does not support OpenCL. In fact, PyTorch developers have been quite hostile towards OpenCL so far: pytorch/pytorch#488

[KellieOwczarczak]

I am certainly struggling to really understand where this stands. I am looking to build an AMD machine for ROCm development purposes and am considering the RX 5700XT vs a Radeon VII. Really sounds like the VII is the way to go as there still is not Navi support. Do I have that right?

[KellieOwczarczak]

@mritunjaymusale Got the NVidia already. Wanting to do the AMD more for exploration, and understanding what is possible if that makes sense. Shoot, in a perfect world, I would do the ROCm in macOS, but truth is that's too much trouble to be worthwhile. But, thanks for more or less confirming that the VII is what I should go with for the AMD machine. ;)

[Djip007]

@ThemysciraData now supported card here: https://github.com/RadeonOpenCompute/ROCm#supported-gpus for low GPU RX580 ... for highend card Radeon VII (or Instinct card...) (RX 5700XT when supported will be slower than Radeon VII ) Note: new GCC10 (https://gcc.gnu.org/gcc-10/changes.html) support offoading with old Fiji or the never VEGA 10/VEGA 20 (no support for polaris now...)

[KellieOwczarczak]

@Djip007 I appreciate the link, but what makes you think I had not already looked at it?