Unoptimized linux driver for Ryzen Embedded APU V1807b:
#opencl on ryzen#amd ryzen embedded v1000dipak
We are a startup in aviation industry and recently purchased Ryzen embedded V1807b APU from a reputed vendor, installed Ubuntu 18.04 on it. Downloaded one of the linux drivers - AMD Linux® 64-bit Driver 4.14 Kernel from AMD's website listed for this board. The install was successful, ./rocminfo and ./clinfo gave the correct information of the device hardware. But, we experienced issues from the driver. One of the example is, When running a simple OpenCL program on the system in a loop (1000 iterations), we noticed the program execution stalled for 2sec-3sec in between iterations and stopped execution after a certain iteration number without giving any error message or segfault. We contacted our vendor, shared our observation and test program with them. They asked us to install the other driver provided on the website AMD Linux® 64-bit Driver 4.19.8 Kernel. Did fresh install of 18.04, installed this new driver successfully. The simple OpenCL program completed successfully, but the problem of stalling of the program for 2-3sec still persisted.
We went ahead, installed various software packages necessary to run our image processing OpenCL software stack on this new Ryzen embedded APU. Prior to this we had Merlin Falcon RX421 BD embedded APU board with 8 compute units were running the same software stack on it at 15 frames per sec(fps). We were expecting increase in the performance and get higher frame rate from the Ryzen embedded APU. To our dismay, Ryzen embedded APU processed at 7 fps. We contacted our vendor again, they performed the standard LuxMark_v3.1 OpenCL benchmark on both Merlin falcon APU and Ryzen embbeded V1807b APU boards and shared the following results that correlate to the slowdown we see in our application.
LuxMark_v3.1 benchmark results for Windows 10 and Ubuntu 18.04 respectively:
OS :Windows 10
Merlin Falcon FP4-RX421BD:
Hotel Lobby: 349
Nueman TLM0102 SE: 1508
LuxBall HDR: 2290
Ryzen Embedded FP5-V1807:
Hotel Lobby: 767
Nueman TLM0102 SE: 2349
LuxBall HDR: 3484
OS: Ubuntu 18.04.1
Merlin Falcon FP4-RX421BD:
Hotel Lobby: 175
Nueman TLM0102 SE: 1474
LuxBall HDR: 2193
Ryzen Embedded FP5-V1807:
Hotel Lobby: 112
Nueman TLM0102 SE: 854
LuxBall HDR: 1423
As one can see, V1807 has higher performance than RX421 in Windows 10. The V1807 performs considerably slow and is only about 60% of RX421 performance in Linux.
We are baffled as to how AMD releases the drivers without actually testing or optimizing them for the boards and making customers scrabble for answers on various sites in vain leaving Linux users like us frustrated with their products. The Nvidia's embedded systems doesn't have any of these driver issues for Ubuntu OS and work out of the box.
And, before someone suggests, try ROCm n all that, we did that aswell. On a fresh Ubuntu 18.04 install(Ubuntu kernel 4.15), we installed ROCm 3.0 from the github page. While ./rocminfo gives the correct output, the ./clinfo executable resulted in segfault error making ROCm install useless for running the OpenCL programs.
Can We expect someone from AMD to respond to this problem and advice how to go about fixing this?
