AMDGPU-PRO 18.50 vs. ROCm 2.0 OpenCL Performance

[phoronix]

Phoronix: AMDGPU-PRO 18.50 vs. ROCm 2.0 OpenCL Performance

When recently publishing the PlaidML deep learning benchmarks and lczero chess neural network OpenCL tests, some Phoronix readers mentioned they were seeing vastly different results with using the PAL OpenCL driver in AMDGPU-PRO (Radeon Software) compared to using the ROCm compute stack. So for seeing how those two separate AMD OpenCL drivers compare, here are some benchmark results with a Vega GPU while testing ROCm 2.0 and AMDGPU-PRO 18.50.

http://www.phoronix.com/vr.php?view=27430

[tildearrow]

Typos:

The Radeon Software AMDGPU-PRO 18.50 PAL OpenCL driver was benchmarks followed

ROCm 2.0 was generally faster than the the OpenCL driver

more performance tuning as its clear from some of t

[tuxd3v]

That cl_mem test, explains a lot..

AMD should put more work on that, curiously the Bandwith was much higher on Rocm, which is good...
Once that cl_mem issue of read/write been solved, rocm will do a Big Jump in Performance.

My wish, for Rocm2.1
Solve the read/writes issue, and please provide a compatibility mode for Rx580( in mobos/CPUs without PCIe Atomics ),
Nvidia also does the same

[cb88]

That cl_mem test, explains a lot..

AMD should put more work on that, curiously the Bandwith was much higher on Rocm, which is good...
Once that cl_mem issue of read/write been solved, rocm will do a Big Jump in Performance.

My wish, for Rocm2.1
Solve the read/writes issue, and please provide a compatibility mode for Rx580( in mobos/CPUs without PCIe Atomics ),
Nvidia also does the same

Actually looks like the benchmark is probably just naive... ROCm 2.0 appears to be optimizing for kernel latency, which is potentially part of why it is faster in other benchs, and which is why read bandwidth is lower but write is unaffected with a naive bencmark (this goes in a queue anyway). Note however that the enqueued Read and Write are both much faster on ROCm.

[khnazile]

Is there a guide how to get ROCm working with mainline kernel without installing whole amdgpu-pro graphics stack? Some people claim it's possible, but I didn't manage to get it working. I tried installing pre-built binaries as https://rocm.github.io/ROCmInstall.html suggests, but it didn't work for me. Also, those binaries come unsigned for some reason...

[cb88]

Is there a guide how to get ROCm working with mainline kernel without installing whole amdgpu-pro graphics stack? Some people claim it's possible, but I didn't manage to get it working. I tried installing pre-built binaries as https://rocm.github.io/ROCmInstall.html suggests, but it didn't work for me. Also, those binaries come unsigned for some reason...

This is the wrong place to ask... Also ROCm has nothing to do with amdgpu-pro they are separate things. The easiest route is probably thier recommended rock-dkms if on Debian.

If you want to get ROCm working with an upstream kernel... then you'll just have to try several ones to see which one works probably 4.18 and 4.19 would be a good place to start. Note there are caveats to using an upstream kernel.

[Azpegath]

Michael, could you please add the Blender tests to this run? That's the only actual use I have for ROCm, and the reason I waited patiently and longingly for it.

[Michael]

Michael, could you please add the Blender tests to this run? That's the only actual use I have for ROCm, and the reason I waited patiently and longingly for it.

Had tried but currently was falling back to CPU rendering for some reason, didn't have time to explore to see if it was bug in ROCm or a Blender issue.

[Marc Driftmeyer]

Since AMD wrote the vast majority of the OpenCL stack in Blender one would think their ``tests'' at conferences would work with ROCm, but my guess is this HSA designed stack for HPC has no future on Blender.

If you'll notice Blender is only testing Pro Renderer on the new Vega VII test slides, not Cycles, never mind Evee.

[bridgman]

Is there a guide how to get ROCm working with mainline kernel without installing whole amdgpu-pro graphics stack? Some people claim it's possible, but I didn't manage to get it working. I tried installing pre-built binaries as https://rocm.github.io/ROCmInstall.html suggests, but it didn't work for me. Also, those binaries come unsigned for some reason...

You need 4.18 or newer upstream kernel; 4.19 if you have a Vega. Also need ROCm 1.9 or newer.

What hardware do you have ?