Can't compile internal While loop or For loop with break

Hi,

I used Aparapi before for GPU calculations Java on GPU: Pricing options with Monte Carlo simulation. I was excited to hear about Torando VM and wanted to repeat my calculations on FPGA.

I noticed an issue when moving from Aparapi to Tornado VM. I have a Monte Carlo simulation (n elements with m calculations/simulations per each). Seems that Tornado can’t compile simulations with breaks (while loop or for loop with break).

Maybe Tornado thinks that I modify incoming arrays, but this isn’t the case.

Do you have any suggestion about how to fix or debug this issue, please?

public static void mapCalcMc(float[] arrS, float[] arrB, int iterPerElement, int logicIterNum, float[] results) {
        int arrLength = arrS.length;
        int currentIter = 0;

        for (@Parallel int i = 0; i < arrLength; i++) {
            for (@Parallel int j = 0; j < iterPerElement; j++) {
                int k = 0;
                float curA = arrS[i];
                float b = arrB[i];
                while (k < logicIterNum
                        && curA > b//Condition that can't be compiled
                ) {
                    k++;
                }

                results[currentIter] = 0.0f;
                currentIter++;
            }
        }
    }

UPD:

1. I run code on AMD 4Gb GPU locally. Tornado works for me both on CPU and GPU with other tasks.
2. Code above is “Minimum Breaking Example”. I tried to remove all non-essential logic.
3. Here is the code to trigger task:

TaskSchedule task = new TaskSchedule("s0")
    .task("map", MonteCarloTornadoGPU::mapCalcMc, new float[10], new float[10],
      10,
      100, // It looks crazy but if I replace 100 wit 10 code works. 
           // This param is used for number of iterations in no-op loop 
       new float[100])
     .streamOut(gpuInterimMapResults);
task.execute();

I found a strange behavior, when I increase number of iterations in no-op while loop, app crashes (see comment to task code).

Attaching kernel out. kernel.txt

[AlexeyPirogov]

I’m sorry, seems I didn’t provide enough information in my original post. I extended post with extra details.

[gigiblender]

@AlexeyPirogov

Thank you for reporting this. It looks that this issue is now solved.

The kernel Tornado generates is printed below:

#pragma OPENCL EXTENSION cl_khr_fp64 : enable  
#pragma OPENCL EXTENSION cl_khr_int64_base_atomics : enable  
__kernel void mapCalcMc(__global uchar *_heap_base, ulong _frame_base, __constant uchar *_constant_region, __local uchar *_local_region, __global int *_atomics)
{
  long l_7, l_8, l_21, l_6, l_19, l_20; 
  ulong ul_11, ul_2, ul_1, ul_0, ul_22, ul_9; 
  int i_24, i_25, i_26, i_27, i_16, i_17, i_18, i_23, i_13, i_14, i_15, i_3, i_4, i_5; 
  float f_12, f_10; 

  __global ulong *_frame = (__global ulong *) &_heap_base[_frame_base];


  // BLOCK 0
  ul_0  =  (ulong) _frame[3];
  ul_1  =  (ulong) _frame[4];
  ul_2  =  (ulong) _frame[7];
  i_3  =  get_global_id(1);
  // BLOCK 1 MERGES [0 11 ]
  i_4  =  0;
  i_5  =  i_3;
  for(;i_5 < 10;)
  {
    // BLOCK 2
    l_6  =  (long) i_5;
    l_7  =  l_6 << 2;
    l_8  =  l_7 + 24L;
    ul_9  =  ul_1 + l_8;
    f_10  =  *((__global float *) ul_9);
    ul_11  =  ul_0 + l_8;
    f_12  =  *((__global float *) ul_11);
    i_13  =  get_global_id(0);
    // BLOCK 3 MERGES [2 10 ]
    i_14  =  i_4;
    i_15  =  i_13;
    for(;i_15 < 10;)
    {
      // BLOCK 4
      // BLOCK 5 MERGES [4 7 ]
      i_16  =  0;
      for(;i_16 < 100;)
      {
        // BLOCK 6
        i_17  =  isless(f_10, f_12);
        if(i_17 == 1)
        {
          // BLOCK 7
          i_18  =  i_16 + 1;
          i_16  =  i_18;
        }  // B7
        else
        {
          // BLOCK 8
          break;
        }  // B8
        // BLOCK 9
      }  // B9
      // BLOCK 10 MERGES [9 8 ]
      l_19  =  (long) i_14;
      l_20  =  l_19 << 2;
      l_21  =  l_20 + 24L;
      ul_22  =  ul_2 + l_21;
      *((__global float *) ul_22)  =  0.0F;
      i_23  =  get_global_size(0);
      i_24  =  i_23 + i_15;
      i_25  =  i_14 + 1;
      i_14  =  i_25;
      i_15  =  i_24;
    }  // B10
    // BLOCK 11
    i_26  =  get_global_size(1);
    i_27  =  i_26 + i_5;
    i_4  =  i_14;
    i_5  =  i_27;
  }  // B11
  // BLOCK 12
  return;
}  //  kernel

I will close this issue for now. If anything else shows up, please let us know with a new issue.