Jump to content

Clpeak - a tool to measure peak capabilities of OpenCL devices


euqlaog

Recommended Posts

  1. Download clpeak
    https://github.com/krrishnarraj/clpeak/archive/master.zip
  2. Download and install CMake 3.1.2 http://www.cmake.org/files/v3.1/cmake-3.1.2-Darwin-x86_64.dmg
  3. Download cl.hpp https://www.khronos.org/registry/cl/api/1.2/cl.hpp and put the file into /System/Library/Frameworks/OpenCL.framework/Headers folder
  4. Open CMake by right mouse clicking it -> select Advanced -> Configure. If you don’t have Command Line Tools for Xcode installed, you will get a prompt window about it - in this case, let the program install it.
  5. Specify the generator: Unix Makefiles -> Use default native compilers
  6. Change the compiler CMAKE_CXX_COMPILER /usr/bin/clang++ (see the screenshot)
  7. Configure & Generate
  8. In Terminal, type "cd /Applications/clpeak" (the path where you wanted the program to be installed)
  9. Type “make” and press enter, you should get the following:
    [ 9%] Building CXX object CMakeFiles/clpeak.dir/src/common.cpp.o
    [ 18%] Building CXX object CMakeFiles/clpeak.dir/src/clpeak.cpp.o
    [ 27%] Building CXX object CMakeFiles/clpeak.dir/src/options.cpp.o
    [ 36%] Building CXX object CMakeFiles/clpeak.dir/src/logger.cpp.o
    [ 45%] Building CXX object CMakeFiles/clpeak.dir/src/global_bandwidth.cpp.o
    [ 54%] Building CXX object CMakeFiles/clpeak.dir/src/compute_sp.cpp.o
    [ 63%] Building CXX object CMakeFiles/clpeak.dir/src/compute_dp.cpp.o
    [ 72%] Building CXX object CMakeFiles/clpeak.dir/src/compute_integer.cpp.o
    [ 81%] Building CXX object CMakeFiles/clpeak.dir/src/transfer_bandwidth.cpp.o
    [ 90%] Building CXX object CMakeFiles/clpeak.dir/src/kernel_latency.cpp.o
    [100%] Building CXX object CMakeFiles/clpeak.dir/src/entry.cpp.o
    Linking CXX executable clpeak
    [100%] Built target clpeak

  10. To run clpeak, type “./clpeak” or double click the program in Finder

post-28870-14494999662862_thumb.png

Tests were done with a 2014 15” rMBP Iris Pro + AKiTiO on OS X 10.10.2 without an external monitor.

LuxMark v3.0 scores (OSX 10.10.2, luxball scene):

1. R9 280X 1070 MHz 3GB, 32 compute units (11739)

2. HD 7970 1010 MHz 3GB, 32 compute units (10628)

3. GTX 980 1215 MHz 3GB, 16 compute units (10577)

4. R9 270X 1120 MHz 2GB, 20 compute units (7697)

5. GTX 780 1019 MHz 6GB, 12 compute units (7013)

180€ R9 270X beats GTX 780. No doubt about AMD’s awesome OpenCL performance, and these AMDs are not higher-end cards.

post-28870-14494999662503_thumb.png

Clpeak gives interestingly different order. For some reason, double-precision compute test gave clCreateKernel (-46) error.

Clpeak Global memory bandwidth, float (GBPS)

1. R9 280X: 232.81

2. GTX 780: 229.04

3. HD 7970: 213.78

4. GTX 980: 191.09

5. R9 270X: 143.20

Clpeak Single-precision compute, float (GFLOPS)

1. GTX 980: 4519.79

2. R9 280X: 3871.81

3. GTX 780: 3270.67

4. HD 7970: 3039.50

5. R9 270X: 2715.69

These numbers seems to match pretty well Linux system results:

https://github.com/krrishnarraj/clpeak/blob/master/results/AMD_Accelerated_Parallel_Processing/Tahiti.log

https://github.com/krrishnarraj/clpeak/blob/master/results/NVIDIA_CUDA/GeForce_GTX_980.log

This is the output of R9 280X:

Device: AMD Radeon HD Tahiti XT Prototype Compute Engine

Driver version : 1.2 (Dec 23 2014 00:17:17) (Macintosh)

Compute units : 32

Clock frequency : 1070 MHz

Global memory bandwidth (GBPS)

float : 232.81

float2 : 240.90

float4 : 245.69

float8 : 129.68

float16 : 66.40

Single-precision compute (GFLOPS)

float : 3871.81

float2 : 3848.93

float4 : 3840.24

float8 : 3819.43

float16 : 3777.76

Double-precision compute (GFLOPS)

clCreateKernel (-46)

Tests skipped

Transfer bandwidth (GBPS)

enqueueWriteBuffer : 1.39

enqueueReadBuffer : 1.42

enqueueMapBuffer(for read) : 24.95

memcpy from mapped ptr : 5.65

enqueueUnmap(after write) : 3002.63

memcpy to mapped ptr : 6.32

Kernel launch latency : 7.65 us

post-28870-14494999628204_thumb.png

post-28870-14494999628599_thumb.png

post-28870-1449499962898_thumb.png

post-28870-14494999662043_thumb.png

  • Thumbs Up 2
Link to comment
Share on other sites

  • 2 weeks later...
×
×
  • Create New...

Important Information

By using this site, you agree to our Terms of Use. We have placed cookies on your device to help make this website better. You can adjust your cookie settings, otherwise we'll assume you're okay to continue.