ds_matmul.c 1.1 KB

123456789101112131415161718192021222324252627282930313233
  1. #include "stdlib.h"
  2. #include "util.h"
  3. #include "dataset.h"
  4. void __attribute__((noinline)) matmul(const int coreid, const int ncores, const int lda, const data_t A[], const data_t B[], data_t C[] )
  5. {
  6. int i,j,k,a,b,b1,a1,a2,a3,c,c1,c2,c3,b2,b3;
  7. for (j=coreid*4; j<lda; j+=4*ncores){
  8. a=j*lda;
  9. a1=(j+1)*lda;
  10. a2=(j+2)*lda;
  11. a3=(j+3)*lda;
  12. for (k=0;k<lda; k+=2)
  13. {
  14. b = k*lda;
  15. b1 = (k+1)*lda;
  16. for (i=0;i<lda;i++){
  17. c = B[b+i];
  18. c1 = B[b1+i];
  19. C[i+a]+=A[a+k]*c;
  20. C[i+a1]+=A[a1+k]*c;
  21. C[i+a2]+=A[a2+k]*c;
  22. C[i+a3]+=A[a3+k]*c;
  23. C[i+a]+=A[a+k+1]*c1;
  24. C[i+a1]+=A[a1+k+1]*c1;
  25. C[i+a2]+=A[a2+k+1]*c1;
  26. C[i+a3]+=A[a3+k+1]*c1;
  27. }
  28. }
  29. }
  30. }