Pascal66/SSOZJCL2.java

19:25:57: Executing task 'SSOZJCL2.main()'...

> Task :compileJava UP-TO-DATE
> Task :processResources NO-SOURCE
> Task :classes UP-TO-DATE

> Task :SSOZJCL2.main()
Please enter an range of integer (comma or space separated): 
0 1e6
 Max threads = 8
Using Prime Generator parameters for given Pn 5
segment size = 33334 resgroups; seg array is [1 x 521] 64-bits
twinprime candidates = 100002 ; resgroups = 33334
each 3 threads has nextp[2 x 165] array
setup time = 0.01 secs
perform twinprimes ssoz sieve
typedef struct This_s{
   __constant long *Arestwins;
   long kmax;
   long Lmodpg;
   long Lend;
   long kmin;
   long Lstart;
   long K1;
   long KB;
   long Kn;
   __constant long *AprimesArray;
   int AprimesArray__javaArrayLength;
   __constant long *Aresinvrs;
   __local long *nextp;
   __local long *seg;
   __global long *lastwins;
   __global long *cnts;
   int passid;
}This;
int get_pass_id(This *this){
   return this->passid;
}
int fr_cridp_sieve_SSOZJCL2$twinSieve__bitCount(This *this, long i){
   i = i - ((((unsigned long)i) >> 1) & 6148914691236517205L);
   i = (i & 3689348814741910323L) + ((((unsigned long)i) >> 2) & 3689348814741910323L);
   i = (i + (((unsigned long)i) >> 4)) & 1085102592571150095L;
   i = i + (((unsigned long)i) >> 8);
   i = i + (((unsigned long)i) >> 16);
   i = i + (((unsigned long)i) >> 32);
   return(((int)i & 127));
}
long fr_cridp_sieve_SSOZJCL2$twinSieve__mod(This *this, long x, long y){
   return((((x >> 31) & (y - 1L)) + (x % y)));
}
__kernel void run(
   __constant long *Arestwins, 
   long kmax, 
   long Lmodpg, 
   long Lend, 
   long kmin, 
   long Lstart, 
   long K1, 
   long KB, 
   long Kn, 
   __constant long *AprimesArray, 
   int AprimesArray__javaArrayLength, 
   __constant long *Aresinvrs, 
   __local long *nextp, 
   __local long *seg, 
   __global long *lastwins, 
   __global long *cnts, 
   int passid
){
   This thisStruct;
   This* this=&thisStruct;
   this->Arestwins = Arestwins;
   this->kmax = kmax;
   this->Lmodpg = Lmodpg;
   this->Lend = Lend;
   this->kmin = kmin;
   this->Lstart = Lstart;
   this->K1 = K1;
   this->KB = KB;
   this->Kn = Kn;
   this->AprimesArray = AprimesArray;
   this->AprimesArray__javaArrayLength = AprimesArray__javaArrayLength;
   this->Aresinvrs = Aresinvrs;
   this->nextp = nextp;
   this->seg = seg;
   this->lastwins = lastwins;
   this->cnts = cnts;
   this->passid = passid;
   {
      int indx = get_global_id(0);
      long r_hi = this->Arestwins[indx];
      long r_lo = r_hi - 2L;
      long sum = 0L;
      if (((((this->kmax - 1L) * this->Lmodpg) + r_hi) - this->Lend)>0){
         this->kmax=this->kmax - 1L;
      }
      if ((((this->kmin * this->Lmodpg) + (r_hi - 2L)) - this->Lstart)<0){
         this->kmin=this->kmin + 1L;
      }
      long hi_tp = 0L;
      this->K1=this->kmin;
      for (; (this->kmin - this->kmax)<0; ){
         if ((this->KB - (this->kmax - this->kmin))>0){
            this->Kn=this->kmax - this->kmin;
         }
         for (int i = 0; i<this->AprimesArray__javaArrayLength; i++){
            if ((this->K1 - this->kmin)==0){
               long k = (this->AprimesArray[i] - 2L) / this->Lmodpg;
               int r = (int)(fr_cridp_sieve_SSOZJCL2$twinSieve__mod(this, (this->AprimesArray[i] - 2L), this->Lmodpg) + 2L);
               long r_inv = this->Aresinvrs[r];
               long ro = fr_cridp_sieve_SSOZJCL2$twinSieve__mod(this, ((r_lo * r_inv) - 2L), this->Lmodpg) + 2L;
               long ko = (k * (this->AprimesArray[i] + ro)) + ((((long)r * ro) - 2L) / this->Lmodpg);
               if ((ko - this->kmin)<0){
                  ko = fr_cridp_sieve_SSOZJCL2$twinSieve__mod(this, (this->kmin - ko), this->AprimesArray[i]);
                  if ((ko - 0L)>0){
                     ko = this->AprimesArray[i] - ko;
                  }
               } else {
                  ko = ko - this->kmin;
               }
               this->nextp[i << 1]  = ko;
               long ri = fr_cridp_sieve_SSOZJCL2$twinSieve__mod(this, ((r_hi * r_inv) - 2L), this->Lmodpg) + 2L;
               long ki = (k * (this->AprimesArray[i] + ri)) + ((((long)r * ri) - 2L) / this->Lmodpg);
               if ((ki - this->kmin)<0){
                  ki = fr_cridp_sieve_SSOZJCL2$twinSieve__mod(this, (this->kmin - ki), this->AprimesArray[i]);
                  if ((ki - 0L)>0){
                     ki = this->AprimesArray[i] - ki;
                  }
               } else {
                  ki = ki - this->kmin;
               }
               this->nextp[(i << 1) | 1]  = ki;
            }
            {
               long ko = this->nextp[(i << 1)];
               for (; (ko - this->Kn)<0; ko = ko + this->AprimesArray[i]){
                  this->seg[(int)(((unsigned long)ko) >> 6)]  = this->seg[(int)(((unsigned long)ko) >> 6)] | (1L << (int)(ko & 63L));
               }
               this->nextp[i << 1]  = ko - this->Kn;
               long ki = this->nextp[((i << 1) | 1)];
               for (; (ki - this->Kn)<0; ki = ki + this->AprimesArray[i]){
                  this->seg[(int)(((unsigned long)ki) >> 6)]  = this->seg[(int)(((unsigned long)ki) >> 6)] | (1L << (int)(ki & 63L));
               }
               this->nextp[(i << 1) | 1]  = ki - this->Kn;
            }
         }
         int upk = (int)(this->Kn - 1L);
         this->seg[((unsigned int)upk) >> 6]  = this->seg[((unsigned int)upk) >> 6] | -(2L << (int)((long)upk & 63L));
         int cnt = 0;
         for (int m = 0; m<=(((unsigned int)upk) >> 6); m++){
            cnt = cnt + (64 - fr_cridp_sieve_SSOZJCL2$twinSieve__bitCount(this, this->seg[m]));
         }
         if (cnt>0){
            sum = sum + (long)cnt;
            for (; ((this->seg[(((unsigned int)upk) >> 6)] & (1L << (int)((long)upk & 63L))) - 0L)!=0; upk--){
            }
            hi_tp = this->kmin + (long)upk;
         }
         this->kmin=this->kmin + this->KB;
         
         if ((this->kmin - this->kmax)<0){
            for (int b = 0; b<=(((unsigned int)upk) >> 6); b++){
               this->seg[b]  = 0L;
            }
         }
      }
      hi_tp = ((r_hi - this->Lend)>0)?0L:((hi_tp * this->Lmodpg) + r_hi);
      this->lastwins[indx]  = ((sum - 0L)==0)?1L:hi_tp;
      this->cnts[indx]  = sum;
      return;
   }
}

execution complete: twinSieve, modes=[GPU, CPU], current = GPU
Device 1964732344400
  vendor = Advanced Micro Devices, Inc.
  type:GPU
  maxComputeUnits=6
  maxWorkItemDimensions=3
  maxWorkItemSizes={256, 256, 256}
  maxWorkWorkGroupSize=256
  globalMemSize=1073741824
  localMemSize=32768
Range global:3 local:(derived)3
sieve time = 0.989 secs
last segment = 33334 resgroups; segment slices = 1
total twins = 2; last twin = -1+/-1
total time = 0.999 secs


Profiles by Kernel Subclass (mean elapsed times in milliseconds)

Device               Count    CLASS_MODEL_BUILT          INIT_JNI  OPENCL_GENERATED   OPENCL_COMPILED   PREPARE_EXECUTE          EXECUTED  Total
----------------- [[ twinSieve ]] --------------------------------------------------------------------------------------------------
AMD<GPU>             1                  354,874             0,135            14,256           215,798             0,476            77,296  662,835


BUILD SUCCESSFUL in 8s
2 actionable tasks: 1 executed, 1 up-to-date
19:26:06: Task execution finished 'SSOZJCL2.main()'.