1 Reply Latest reply on Dec 5, 2010 4:46 PM by himanshu.gautam

    Very inconsistent ReadBuffer times

    Raistmer
      reading single int value longer than whole kernel run

      At the beginning of app:

      set_mem_kernel_cl_02F2B098 14 { 1 1 1} NULL 0,00581 0 2 0 0 1,00 6,00 0,00 1,00 0,00 0,00 0,03 6,00 23,33 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00
      PC_single_pulse_kernel_FFA_update_reduce0_02F2B058 15 { 32 12 256} { 1 1 256} 0,39519 0 34 0 2 1536,00 100,34 16,00 1,06 0,00 0,00 8,83 6,27 86,37 12288,00 0,00 89,18 66,89 0,00 1631,00 0,00 100,00 0,00 0,00
      ReadBuffer 16 0,59167 0,00
      PC_single_pulse_kernel_FFA_update_reduce1_02F2AFD8 17 { 32 12 16} NULL 0,04914 0 37 0 2 192,00 115,00 16,00 4,00 0,00 0,00 10,71 7,19 82,43 768,00 0,00 81,84 57,00 0,00 11,00 0,00 100,00 0,00 0,00
      ReadBuffer 18 0,46991 0,00


      Later:
      set_mem_kernel_cl_02F2B098 224 { 1 1 1} NULL 0,00303 0 2 0 0 1,00 6,00 0,00 1,00 0,00 0,00 0,04 6,00 23,33 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00 0,00
      PC_single_pulse_kernel_FFA_update_reduce0_02F2B058 225 { 32 12 256} { 1 1 256} 0,39184 0 34 0 2 1536,00 98,00 16,00 1,00 0,00 0,00 8,64 6,13 87,96 12288,00 0,00 89,30 66,96 0,00 1535,00 0,00 100,00 0,00 0,00
      ReadBuffer 226 0,19441 0,00
      PC_single_pulse_kernel_FFA_update_reduce1_02F2AFD8 227 { 32 12 16} NULL 0,03638 0 37 0 2 192,00 115,00 16,00 4,00 0,00 0,00 10,09 7,19 82,43 768,00 0,00 75,32 51,93 0,00 11,00 0,00 100,00 0,00 0,00
      ReadBuffer 228 1,03438 0,00


      That is, >1ms to read single sizeof(int) result flag !
      How it could be?
      Also, I have logs where it took almost second.
      Is it some bug in profiler that it reports such ridiculous numbers ?
        • Very inconsistent ReadBuffer times
          himanshu.gautam

          Ihave not seen such high times for clRead Buffer.

          It might happen if we queue up lot of commands in a command queue. They are then launched together as a batch to save on kernel launch overhead. In this case all the commands report similar start & end times which might be resulting in such high times.

          Do you see such numbers for a single command?