OpenCL benchmark

Very well, I will attach some Android benchmarks along with temperature readings on everything,

Hi, @Electr1, any news about the OpenCL benchmarks?

Yep, I do, regarding openCL I am still working on the it, I got OpenCL working but clmark is taking a lot of time to get ready, sbc-bench is ok, in all we are good to go,

The OS is ubuntu-20.04 kernel-4.9 and is a new flash disk, I am waiting for my cpu fan, expected to arrive in about ETA 2 hours, you can expect to see the results by today evening :slight_smile: ,
have a good day!

this just in, recieved my parcel, the fan is really good, and the airflow is also quite high, super quiet, probably below 30 dB, highly recommended,


Will start the benchmarks !

1 Like

@bizcocho85 CLPeak is giving me a tough time, will procced to test hashcat bench marks…

clpeak gives about 70 GFLOPS with GPU alone with bandwidth at 7-8 Gbps.

for some reason when adding CPU to the MIX the performance drops to half 35-38 GFLOPS with GPU…

CPU I was using 4 A73 cores alone and left out other two cores

1 Like

Could you tell me how you did it ?,
I got confused at what to do with those .cmake files

I calculated this before,
But the estimated maximum GFLOPs of the GPU, is 166.06. GFLOPs

Amazing! Where did you buy it (or the name of it)? Would be interesting to know how much temperature it reduces and its consumption.

Regarding clpeak, in the github appear:

git clone https://github.com/krrishnarraj/clpeak
cd clpeak
git submodule update --init --recursive --remote
mkdir build
cd build
cmake ..
cmake --build .

If with the last cmake --build . it does not compile and build clpeak, you can run make.

Could you please dump the experiments? (program-repository, execution line + stdout/err).

Maybe they share the cache lines at the LLC (eg. L3) and the bandwidth is reduced.

What I would do is to run: only GPU, only CPU, both at the same time.
Then, if I see this behavior, I would execute GPU + 1 CPU core, GPU + 2 CPU cores, etc. To understand better.

Assuming you are not using buffers and drivers. In that case, you will add data transfer overheads.

ok thanks for the info,

Regarding the Fan I bought it from Amazon.in,
Temps are really good, The first time I used it Idle temps were about upper 30 °C, and below 40 °C, During heavy workload it went to a max of about 56 °C to 58 °C
here is the link if you want to buy internationally, I am not sure if they ship to places outside India though…

l did the things and it executed some things and i assume it built it successfully, how do I execute it ?
please forgive me I Kinda don’t know how to do some of this stuff, I am still learning some of this software :sweat_smile:

Never mind I got it working :stuck_out_tongue_winking_eye:

ok I got GPU compute results what can I do to get CPU compute results, @Archangel1235 a bit of help please,

GPU alone

Platform: ARM Platform
Device: Mali-G52
Driver version : 2.0 (Linux ARM64)
Compute units : 2
Clock frequency : 750 MHz

Global memory bandwidth (GBPS)
  float   : 6.26
  float2  : 6.90
  float4  : 7.41
  float8  : 3.49
  float16 : 1.72

Single-precision compute (GFLOPS)
  float   : 76.52
  float2  : 76.44
  float4  : 76.37
  float8  : 76.11
  float16 : 74.88

Half-precision compute (GFLOPS)
  half   : 76.46
  half2  : 151.87
  half4  : 151.41
  half8  : 150.55
  half16 : 149.02

No double precision support! Skipped

Integer compute (GIOPS)
  int   : 71.71
  int2  : 73.79
  int4  : 74.79
  int8  : 74.85
  int16 : 58.07

Transfer bandwidth (GBPS)
  enqueueWriteBuffer         : 4.73
  enqueueReadBuffer          : 4.72
  enqueueMapBuffer(for read) : 6254.91
    memcpy from mapped ptr   : 4.73
  enqueueUnmap(after write)  : 11289.36
    memcpy to mapped ptr     : 4.74

Kernel launch latency : 74.32 us

Interesting thing It shows only two compute units. A311D should have 4 Compute units… Maybe khadas team can elaborate on this

Also the clock speed is 50MHz lower that advertised

You have to build pocl for CPU support.

follow this post to build pocl

https://forum.odroid.com/viewtopic.php?t=34020#p290612

1 Like

Thanks mate, till now I have made a few benchmarks, both in Android and Linux, those pending are GLmark as well CLpeak for CPU,

Will try to somehow bring the results by today evening, yesterday I had some power issues in my home, and due to some improper electrical wiring Neutral was connected to ground and in my room and the power brick my VIM was connected to blew up, fortunately my VIM survived and nobody got hurt,

Benchmarks completed !
@bizcocho85 a new topic will be created listing all the benchmark results, will update, on CLpeak for CPU later,
Buenos dias !

Interesting, thanks. Can we contact them using this forum? Any recommended user?

Thank you very much.

When you will have OpenCL working in the CPU, I can suggest another program to test CPU+GPU co-execution.