krrishnarraj / clpeak

A tool which profiles OpenCL devices to find their peak capacities
Apache License 2.0
409 stars 115 forks source link

Question concerning AMD Radeon RX 550 #117

Closed boustanihani closed 1 month ago

boustanihani commented 1 month ago

Are those values okay for an AMD Radeon RX 550:

(I am using NixOS 24.05 with MESA drivers)

Platform: Clover
  Device: AMD Radeon RX 550 / 550 Series (radeonsi, polaris12, LLVM 17.0.6, DRM 3.54, 6.6.51)
    Driver version  : 24.0.7 (Linux x64)
    Compute units   : 8
    Clock frequency : 1183 MHz

    Global memory bandwidth (GBPS)
      float   : 5.73
      float2  : 5.60
      float4  : 5.65
      float8  : 5.33
      float16 : 4.48

    Single-precision compute (GFLOPS)
      float   : 576.10
      float2  : 575.73
      float4  : 574.13
      float8  : 571.74
      float16 : 566.20

    No half precision support! Skipped

    Double-precision compute (GFLOPS)
      double   : 65.10
      double2  : 67.01
      double4  : 65.29
      double8  : 64.14
      double16 : 58.16

    Integer compute (GIOPS)
      int   : 228.10
      int2  : 227.43
      int4  : 225.38
      int8  : 221.10
      int16 : 218.38

    Transfer bandwidth (GBPS)
      enqueueWriteBuffer         : 7.92
      enqueueReadBuffer          : 8.20
      enqueueMapBuffer(for read) : 4802.07
        memcpy from mapped ptr   : 8.17
      enqueueUnmap(after write)  : 6963.31
        memcpy to mapped ptr     : 8.01

    Kernel launch latency : 84.36 us

Platform: rusticl
clCreateContextFromType (-1)
rjodinchr commented 1 month ago

Maybe you can try another OpenCL implementation to see if you get the same numbers: https://github.com/kpet/clvk clover is quite old. And rusticl seems not to support clCreateContextFromType. But clvk should be able to give you numbers you can trust.

boustanihani commented 1 month ago

Thanks :)