examples/chemv.c

   1 /*
   2  * Copyright 2014      ARM Ltd.
   3  *
   4  * Use of this software is governed by the MIT license
   5  */
   6
   7 #include <stdio.h>
   8 #include <stdlib.h>
   9
  10 struct ComplexFloat
  11 {
  12         float Re;
  13         float Im;
  14 };
  15
  16 /* chemv - complex hermitian matrix-vector multiplication
  17  * The function body was taken from a VOBLA-generated BLAS library.
  18  */
  19 void chemv(int n, float alpha_re, float alpha_im,
  20         int ldAT, struct ComplexFloat AT[restrict const static n][ldAT],
  21         int incX, struct ComplexFloat X[restrict const static n][incX],
  22         float beta_re, float beta_im,
  23         int incY, struct ComplexFloat Y[restrict const static n][incY])
  24 {
  25 #pragma scop
  26         for (int i0 = 0; i0 <= (n-1); i0 += 1) {
  27                 float var5_Re;
  28                 float var5_Im;
  29                 var5_Re = ((Y[i0][0].Re*beta_re)-(Y[i0][0].Im*beta_im));
  30                 var5_Im = ((Y[i0][0].Im*beta_re)+(Y[i0][0].Re*beta_im));
  31                 Y[i0][0].Re = var5_Re;
  32                 Y[i0][0].Im = var5_Im;
  33         }
  34         for (int i1 = 0; i1 <= ((n-1)+1)-1; i1 += 1) {
  35                 float var2_Re;
  36                 float var3_Im;
  37                 float var2_Im;
  38                 float var4_Im;
  39                 float var4_Re;
  40                 float var3_Re;
  41                 var2_Re = (alpha_re*AT[i1][i1].Re);
  42                 var2_Im = (alpha_im*AT[i1][i1].Re);
  43                 var3_Re = ((var2_Re*X[i1][0].Re)-(var2_Im*X[i1][0].Im));
  44                 var3_Im = ((var2_Im*X[i1][0].Re)+(var2_Re*X[i1][0].Im));
  45                 var4_Re = (Y[i1][0].Re+var3_Re);
  46                 var4_Im = (Y[i1][0].Im+var3_Im);
  47                 Y[i1][0].Re = var4_Re;
  48                 Y[i1][0].Im = var4_Im;
  49         }
  50         for (int i2 = 0; i2 <= ((n-1)-1); i2 += 1) {
  51                 for (int i3 = 0; i3 <= (n-1)-(1+i2); i3 += 1) {
  52                         float var99_Re;
  53                         float var96_Re;
  54                         float var98_Im;
  55                         float var96_Im;
  56                         float var94_Im;
  57                         float var95_Im;
  58                         float var94_Re;
  59                         float var95_Re;
  60                         float var97_Im;
  61                         float var99_Im;
  62                         float var97_Re;
  63                         float var98_Re;
  64                         var94_Re = ((alpha_re*AT[i2][((1+i2)+i3)].Re)-
  65                                 (alpha_im*(-AT[i2][((1+i2)+i3)].Im)));
  66                         var94_Im = ((alpha_im*AT[i2][((1+i2)+i3)].Re)+
  67                                 (alpha_re*(-AT[i2][((1+i2)+i3)].Im)));
  68                         var95_Re = ((var94_Re*X[((i3+i2)+1)][0].Re)-
  69                                 (var94_Im*X[((i3+i2)+1)][0].Im));
  70                         var95_Im = ((var94_Im*X[((i3+i2)+1)][0].Re)+
  71                                 (var94_Re*X[((i3+i2)+1)][0].Im));
  72                         var96_Re = (Y[i2][0].Re+var95_Re);
  73                         var96_Im = (Y[i2][0].Im+var95_Im);
  74                         Y[i2][0].Re = var96_Re;
  75                         Y[i2][0].Im = var96_Im;
  76                         var97_Re = ((alpha_re*AT[i2][((1+i2)+i3)].Re)-
  77                                 (alpha_im*AT[i2][((1+i2)+i3)].Im));
  78                         var97_Im = ((alpha_im*AT[i2][((1+i2)+i3)].Re)+
  79                                 (alpha_re*AT[i2][((1+i2)+i3)].Im));
  80                         var98_Re = ((var97_Re*X[i2][0].Re)-
  81                                 (var97_Im*X[i2][0].Im));
  82                         var98_Im = ((var97_Im*X[i2][0].Re)+
  83                                 (var97_Re*X[i2][0].Im));
  84                         var99_Re = (Y[((i3+i2)+1)][0].Re+var98_Re);
  85                         var99_Im = (Y[((i3+i2)+1)][0].Im+var98_Im);
  86                         Y[((i3+i2)+1)][0].Re = var99_Re;
  87                         Y[((i3+i2)+1)][0].Im = var99_Im;
  88                 }
  89         }
  90 #pragma endscop
  91 }
  92
  93 int main()
  94 {
  95         const int n = 37;
  96         const int incX = 1;
  97         const int incY = 1;
  98         const int ldAT = n;
  99         struct ComplexFloat AT[n][ldAT];
 100         struct ComplexFloat X[n][incX];
 101         struct ComplexFloat Y[n][incY];
 102
 103         for (int i = 0; i < n; i++) {
 104                 X[i][0] = (struct ComplexFloat){i + 5, i * 2};
 105                 Y[i][0] = (struct ComplexFloat){i * 3, i + 7};
 106                 for (int j = 0; j < ldAT; j++) {
 107                         AT[i][j] = (struct ComplexFloat){i + j, i + 3};
 108                 }
 109         }
 110
 111         chemv(n, 3.14f, 1.59f, ldAT, AT, incX, X, 2.71f, 8.28f, incY, Y);
 112
 113         for (int i = 0; i < n; i++)
 114                 printf("%0.2f %0.2f\n", Y[i][0].Re, Y[i][0].Im);
 115
 116         return EXIT_SUCCESS;
 117 }