Enhancement: Symbolic Loop Merge Dims

[hikettei]

• Write tests
• Transformer Speed Improvement

[hikettei]

In this schedule, _gid1 and _gid2 is mergeable?

void fused_sumnode_softmax_maxreduce_attention_softmax2321523(float* val_593, float* val_610, const int64_t n, const int64_t s, const float* restrict val_599) {
  for (int _gid1=0; _gid1<12; _gid1+=1) {
    for (int _gid2=0; _gid2<s; _gid2+=1) {
      float val_363 = 0.0;
      float val_365 = _negative_infinity;
      for (int _gid3=0; _gid3<(n+s); _gid3+=1) {
        val_593[(((((n+s)*s)*_gid1)+((n+s)*_gid2))+_gid3)] = ((val_593[(((((n+s)*s)*_gid1)+((n+s)*_gid2))+_gid3)]*0.125)+val_599[(((n+s)*_gid2)+_gid3)]);
        val_365 = max(val_365, val_593[(((((n+s)*s)*_gid1)+((n+s)*_gid2))+_gid3)]);
      }
      for (int _gid3=0; _gid3<(n+s); _gid3+=1) {
        val_593[(((((n+s)*s)*_gid1)+((n+s)*_gid2))+_gid3)] = exp2(((val_593[(((((n+s)*s)*_gid1)+((n+s)*_gid2))+_gid3)]+-(val_365))*1.442695));
        val_363 = (val_363+val_593[(((((n+s)*s)*_gid1)+((n+s)*_gid2))+_gid3)]);
      }
      for (int _gid3=0; _gid3<(n+s); _gid3+=1) {
        val_610[(((((n+s)*s)*_gid1)+((n+s)*_gid2))+_gid3)] = (val_593[(((((n+s)*s)*_gid1)+((n+s)*_gid2))+_gid3)]*1.0/(val_363));
      }
    }
  }
}