#clpeak

 Reached a max global bandwidth of 131 GBPS with float/float2

 2600 GFLOPS with float4, hardware peak being 2700. The VLIW4 architecture requires vector-code to utilize peak capacity. With scalar code, you only achieve 1/4th of capacity

 Double doesn't require vector-code as 671 GFLOPS was achieved against hw peak of 675(1/4th of single fp) using scalar code

 Transfer bandwidth, it actually depends on the host m/c specs. Map/unmap is doing a zero-copy for a buffer created using CL_MEM_ALLOC_HOST_PTR flag (or it could a bug in code which makes it appear as zero-copy)

There is something interesting about intel cpus on intel platform

Platform: Intel(R) OpenCL                                 Platform: Intel(R) OpenCL   Device: Intel(R) Core(TM) i7-3630QM CPU @ 2.40GHz     Device: Intel(R) Core(TM) i7-3630QM CPU @ 2.40GHz     Driver version: 1.2 (Win32)                               Driver version: 1.2 (Win64)     Single-precision compute (GFLOPS)                         Single-precision compute (GFLOPS)       float   : 25.44                                             float   : 25.41       float2  : 50.81                                             float2  : 101.06       float4  : 51.61                                             float4  : 171.80       float8  : 52.55                                             float8  : 81.04       float16 : 51.92                                             float16 : 95.44     Double-precision compute (GFLOPS)                         Double-precision compute (GFLOPS)       double   : 25.45                                            double   : 25.41       double2  : 25.47                                            double2  : 87.65       double4  : 25.71                                            double4  : 34.09       double8  : 26.23                                            double8  : 30.03       double16 : 27.16                                            double16 : 86.77     Transfer bandwidth (GBPS)                                 Transfer bandwidth (GBPS)       enqueueWriteBuffer          : 2.30                         enqueueWriteBuffer         : 2.26       enqueueReadBuffer           : 6.47                        enqueueReadBuffer          : 7.56       enqueueMapBuffer(for read)  : 1.#J                         enqueueMapBuffer(for read) : 1.#J         memcpy from mapped ptr : 8.39                           memcpy from mapped ptr : 9.26       enqueueUnmap(after write)   : 1.#J                         enqueueUnmap(after write)  : 1.#J         memcpy to mapped ptr    : 8.13                           memcpy to mapped ptr   : 8.95

There is a huge compute performance difference in 32/64 modes. I got suspicious and went to intel forum software.intel.com/en-us/forums/topic/495379 . Compiler is exposed to more registers in x64. Running the same program in x64 is showing more than 3x performance!!! The hardware compute peak is 153.6 GFLOPS at 2.4GHz. Arik indicated that actual freq can go higher due to turbo mode. So 172 GFLOPS in x64 is possibly because of turbo mode

Coming to next device, Mali gpu in my laptop (yes yes its chromebook) exynos 5250 soc

Platform: ARM Platform   Device: Mali-T604     Driver version: 1.1 (Linux ARM)     Global memory bandwidth (GBPS)       float   : 1.56       float2  : 4.41       float4  : 5.75       float8  : Out of resources! Skipped     Single-precision compute (GFLOPS)       float   : 2.38       float2  : 16.40       float4  : 8.07       float8  : 21.84       float16 : 16.37     No double precision support! Skipped     Transfer bandwidth (GBPS)       enqueueWriteBuffer         : 5.84       enqueueReadBuffer          : 2.59       enqueueMapBuffer(for read) : 934.99         memcpy from mapped ptr   : 2.83       enqueueUnmap(after write)  : 1813.75         memcpy to mapped ptr     : 2.85     Kernel launch latency : 149.46 us

Highlights:

Max bandwidth of around 5.75 GBPS

Compute peak of around 22 GFLOPS. Not sure of what is the real hardware peak. Could be 32 something

Around 6 GBPS of host->device bandwidth. This is relatively higher than standard x86 based pcie systems. Map & unmap buffers are zero-copy

kernel latency measures time b/w when the kernel was queued at host and when it started executing on device. 150 us is acceptable for gpu

Lets take a device and profile 3 different opencl runtimes on it. Device was Intel i3-550 @ 3.2 GHz on linux x64. This is a nehalem based device having peak compute capacity of

51.2

GFLOPS pocl is an opensource implementation of opencl spec 

portablecl.org

 . Currently aims cpu devices. For this setup, pocl was compiled against llvm 3.2

Platform                           :    AMD            Intel           POCL Driver version                     :    1214.3          1.2.0.76921     0.9 Global memory bandwidth (GBPS)   float                             :    7.58            8.28            7.73   float2                            :    8.63            8.15            8.70   float4                            :    7.67            7.34            7.67   float8                            :    7.82            7.63            7.81   float16                           :    7.51            7.48            7.64 Single-precision compute (GFLOPS)   float                             :    3.64            14.56           3.64   float2                            :    7.26            29.07           7.30   float4                            :    14.45           46.65           14.58   float8                            :    28.50           28.68           28.51   float16                           :    18.31           45.49           49.67 Double-precision compute (GFLOPS)   double                           :    3.19            12.56           3.19   double2                          :    6.37            22.47           6.39   double4                          :    12.65           8.55            12.70   double8                          :    21.19           21.26           19.98   double16                         :    4.88            10.25           6.73 Transfer bandwidth (GBPS)   enqueueWriteBuffer               :    6.59            1.40            6.60   enqueueReadBuffer               :    4.16            4.13            4.05   enqueueMapBuffer(for read)      :    8159.13         1355.73         28256.36    memcpy from mapped ptr       :    4.17            4.13            4.05   enqueueUnmap(after write)        :   8801.16         1682.98         51622.20    memcpy to mapped ptr          :    4.31            4.26            4.24 Kernel launch latency (us)           :    11.36           5.33            5.12

Highlights:

Bandwidth wise, all 3 platforms have performed almost equally. Although theoretical peak being 21 GBPS, only 1/2 was achieved because only 1 RAM slot was used. You need to place RAM in both slots to realize peak bandwidth

single fp compute, AMD reached only half of hardware peak. Intel decently peaked at float4, while pocl peaked at float16

  Check

github.com/krrishnarraj/clpeak/tree/master/results