voutfloatn.br

用于GPU通用计算的编程语言BrookGPU 0.4

   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         -0.5f,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else {
      output=input2[newindex];
   }
}

kernel void floatstreamCombine3(float input0[][],
                           float input1[][],
                           float input2[][],
                           float modulus,
                           float offset,
                           float length, 
                           float lengthmodmodulus,
                           float oldoutput<>,
                           out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         offset,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else {
      output=oldoutput;
   }
}

kernel void floatstreamCombine4f(float input0[][],
                            float input1[][],
                            float input2[][],
                            float input3[][],
                            float modulus,
                            float length, 
                            float lengthmodmodulus,
                            out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         -0.5f,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else {
      output=input3[newindex];
   }
}

kernel void floatstreamCombine4(float input0[][],
                           float input1[][],
                           float input2[][],
                           float input3[][],
                           float modulus,
                           float offset,
                           float length, 
                           float lengthmodmodulus,
                           float oldoutput<>,
                           out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         offset,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else if (whichmod==3) {
      output=input3[newindex];
   }else {
      output=oldoutput;
   }
}

kernel void floatstreamCombine5f(float input0[][],
                            float input1[][],
                            float input2[][],
                            float input3[][],
                            float input4[][],
                            float modulus,
                            float length, 
                            float lengthmodmodulus,
                            out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         -0.5f,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else if (whichmod==3) {
      output=input3[newindex];
   }else {
      output=input4[newindex];
   }
}
/*
kernel void floatstreamCombine5(float input0[][],
                           float input1[][],
                           float input2[][],
                           float input3[][],
                           float input4[][],
                           float modulus,
                           float offset,
                           float length, 
                           float lengthmodmodulus,
                           float oldoutput<>,
                           out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         offset,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else if (whichmod==3) {
      output=input3[newindex];
   }else if (whichmod==4) {
      output=input4[newindex];
   }else {
      output=oldoutput;
   }
}


kernel void floatstreamCombine6f(float input0[][],
                            float input1[][],
                            float input2[][],
                            float input3[][],
                            float input4[][],
                            float input5[][],
                            float modulus,
                            float length, 
                            float lengthmodmodulus,
                            out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         -0.5f,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else if (whichmod==3) {
      output=input3[newindex];
   }else if (whichmod==4) {
      output=input4[newindex];
   }else {
      output=input5[newindex];
   }
}

kernel void floatstreamCombine6(float input0[][],
                           float input1[][],
                           float input2[][],
                           float input3[][],
                           float input4[][],
                           float input5[][],
                           float modulus,
                           float offset,
                           float length, 
                           float lengthmodmodulus,
                           float oldoutput<>,
                           out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         offset,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else if (whichmod==3) {
      output=input3[newindex];
   }else if (whichmod==4) {
      output=input4[newindex];
   }else if (whichmod==5) {
      output=input5[newindex];
   }else {
      output=oldoutput;
   }
}

kernel void floatstreamCombine7f(float input0[][],
                            float input1[][],
                            float input2[][],
                            float input3[][],
                            float input4[][],
                            float input5[][],
                            float input6[][],
                            float modulus,
                            float length, 
                            float lengthmodmodulus,
                            out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         -0.5f,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else if (whichmod==3) {
      output=input3[newindex];
   }else if (whichmod==4) {
      output=input4[newindex];
   }else if (whichmod==5) {
      output=input5[newindex];
   }else {
      output=input6[newindex];
   }
}

kernel void floatstreamCombine7(float input0[][],
                           float input1[][],
                           float input2[][],
                           float input3[][],
                           float input4[][],
                           float input5[][],
                           float input6[][],
                           float modulus,
                           float offset,
                           float length, 
                           float lengthmodmodulus,
                           float oldoutput<>,
                           out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         offset,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else if (whichmod==3) {
      output=input3[newindex];
   }else if (whichmod==4) {
      output=input4[newindex];
   }else if (whichmod==5) {
      output=input5[newindex];
   }else if (whichmod==6) {
      output=input6[newindex];
   }else {
      output=oldoutput;
   }
}

kernel void floatstreamCombine8f(float input0[][],
                            float input1[][],
                            float input2[][],
                            float input3[][],
                            float input4[][],
                            float input5[][],
                            float input6[][],
                            float input7[][],
                            float modulus,
                            float length, 
                            float lengthmodmodulus,
                            out float output <>) {
   float2 newindex;
   float whichmod;
   calculateDividedIndex(indexof(output), modulus, length, newindex);
   calculateIndexModulus(indexof(output), 
                         modulus,
                         -0.5f,
                         lengthmodmodulus,
                         whichmod);
   if (whichmod==0||whichmod==modulus) {
      output=input0[newindex];
   }else if (whichmod==1) {
      output=input1[newindex];
   }else if (whichmod==2) {
      output=input2[newindex];
   }else if (whichmod==3) {
      output=input3[newindex];
   }else if (whichmod==4) {
      output=input4[newindex];
   }else if (whichmod==5) {
      output=input5[newindex];
   }else if (whichmod==6) {
      output=input6[newindex];
   }else {
      output=input7[newindex];
   }
}
*/
void combineStreamsHelperfloat (float (**streams) <>,
                                unsigned int num,
                                float output<>,
                                float modulus,
                                float offset,
                                float length,
                                float lengthmodmodulus) {
   switch (num) {
   case 1:
      floatstreamCombine1(*streams[0],
                     modulus,
                     offset,
                     length,
                     lengthmodmodulus,
                     output,
                     output);
      break;
   case 2:
      floatstreamCombine2(*streams[0],
                     *streams[1],
                     modulus,
                     offset,
                     length,
                     lengthmodmodulus,
                     output,
                     output);
      break;
   case 3:
      floatstreamCombine3(*streams[0],
                     *streams[1],
                     *streams[2],
                     modulus,
                     offset,
                     length,
                     lengthmodmodulus,