added fft for simple numbers

dspl/src/dspl_internal.h

@@ -7,6 +7,8 @@
 #define DSPL_RAND_MOD_X1            2147483647
 #define DSPL_RAND_MOD_X2            2145483479
 
+// sqrt(2^31)
+#define FFT_COMPOSITE_MAX           46340
 
 void transpose(double* a, int n, int m, double* b);
 void transpose_cmplx(complex_t* a, int n, int m, complex_t* b);

dspl/src/fft.c

@@ -115,8 +115,9 @@ composite FFT kernel
 *******************************************************************************/
 int fft_krn(complex_t* t0, complex_t* t1, fft_t* p, int n, int addr)
 {
-  int n1, n2, k;
+  int n1, n2, k, m, i;
   complex_t *pw = p->w+addr;
+  complex_t tmp;
   
   n1 = 1;
   if(n%16== 0) { n1 = 16;   goto label_size;  }
@@ -128,8 +129,22 @@ int fft_krn(complex_t* t0, complex_t* t1, fft_t* p, int n, int addr)
 
 label_size:
   if(n1 == 1)
-    return ERROR_FFT_SIZE;
-  
+  {
+    for(k = 0; k < n; k++)
+    {
+      RE(t1[k]) = IM(t1[k]) = 0.0;
+      for(m = 0; m < n; m++)
+      {
+        i = addr + ((k*m) % n);
+        RE(tmp) = CMRE(t0[m], pw[i]);
+        IM(tmp) = CMIM(t0[m], pw[i]);
+        RE(t1[k]) += RE(tmp);
+        IM(t1[k]) += IM(tmp);
+      }
+    }
+  }
+  else
+  {
     n2 = n / n1;
     
     if(n2>1)
@@ -179,6 +194,7 @@ label_size:
       }
       transpose_cmplx(t0, n2, n1, t1);
     }
+  }
   return RES_OK;
 
 }
@@ -192,7 +208,7 @@ FFT create for composite N
 int DSPL_API fft_create(fft_t *pfft, int n)
 {
 
-  int n1, n2, addr, s, k, m, nw;
+  int n1, n2, addr, s, k, m, nw, err;
   double phi;
   s = n;
   nw = addr = 0;
@@ -213,7 +229,26 @@ int DSPL_API fft_create(fft_t *pfft, int n)
 
 label_size:
     if(n2 == 1)
-      return ERROR_FFT_SIZE;
+    {
+      if(s > FFT_COMPOSITE_MAX)
+      {
+        err = ERROR_FFT_SIZE;
+        goto error_proc;
+      }
+      
+      nw += s;
+      pfft->w = pfft->w ? (complex_t*) realloc(pfft->w,  nw*sizeof(complex_t)):
+                          (complex_t*) malloc(           nw*sizeof(complex_t));
+      for(k = 0; k < s; k++)
+      {
+        phi = - M_2PI * (double)k / (double)s;
+        RE(pfft->w[addr]) = cos(phi);
+        IM(pfft->w[addr]) = sin(phi);
+        addr++;
+      }
+    }
+    else
+    {
       n1 = s / n2;
       nw += s;
       pfft->w = pfft->w ? (complex_t*) realloc(pfft->w,  nw*sizeof(complex_t)):
@@ -229,19 +264,25 @@ label_size:
           addr++;
         }
       }
+    }
     s /= n2;    
   }
 
   pfft->t0 = pfft->t0 ? (complex_t*) realloc(pfft->t0, n*sizeof(complex_t)):
                         (complex_t*) malloc(           n*sizeof(complex_t));
 
-
   pfft->t1 = pfft->t1 ? (complex_t*) realloc(pfft->t1, n*sizeof(complex_t)):
                         (complex_t*) malloc(           n*sizeof(complex_t));
                         
   pfft->n = n;
 
   return RES_OK;
+error_proc:
+  if(pfft->t0) free(pfft->t0);
+  if(pfft->t1) free(pfft->t1);
+  if(pfft->w)  free(pfft->w);
+  pfft->n = 0;
+  return err;
 }
 
 

dspl/src/fft_subkernel.c

@@ -380,7 +380,7 @@ void dft16 (complex_t *x,  complex_t* y)
   IM(t1[15]) =  RE(t1[15]) * DFT16_W2 - IM(t1[15]) * DFT16_W1;
   RE(t1[15]) = tmp;
   
-  transpose16x16(t1, t0);
+  transpose4x4(t1, t0);
   
   dft4(t0,    t1);
   dft4(t0+4,  t1+4);