NR SCFDMA changes for higher RB sizes (RB Sizes until 100Mhz)

openair1/PHY/MODULATION/nr_modulation.c

@@ -325,10 +325,10 @@ void nr_ue_layer_mapping(NR_UE_ULSCH_t **ulsch_ue,
 
 void nr_dft(int32_t *z, int32_t *d, uint32_t Msc_PUSCH)
 {
-#if defined(__x86_64__) || defined(__i386__)
-  __m128i dft_in128[1][1200], dft_out128[1][1200];
+#if defined(__x86_64__) || +defined(__i386__)
+  __m128i dft_in128[1][3240], dft_out128[1][3240];
 #elif defined(__arm__)
-  int16x8_t dft_in128[1][1200], dft_out128[1][1200];
+  int16x8_t dft_in128[1][3240], dft_out128[1][3240];
 #endif
   uint32_t *dft_in0 = (uint32_t*)dft_in128[0], *dft_out0 = (uint32_t*)dft_out128[0];
 
@@ -340,8 +340,10 @@ void nr_dft(int32_t *z, int32_t *d, uint32_t Msc_PUSCH)
   int16x8_t norm128;
 #endif
 
-  for (i = 0, ip = 0; i < Msc_PUSCH; i++, ip+=4) {
-    dft_in0[ip] = d[i];
+  if ((Msc_PUSCH % 1536) > 0) {
+    for (i = 0, ip = 0; i < Msc_PUSCH; i++, ip+=4) {
+      dft_in0[ip] = d[i];
+    }
   }
 
   switch (Msc_PUSCH) {
@@ -480,7 +482,7 @@ void nr_dft(int32_t *z, int32_t *d, uint32_t Msc_PUSCH)
       break;
 
     case 972:
-      dft(DFT_960,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      dft(DFT_972,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
       break;
 
     case 1080:
@@ -494,11 +496,98 @@ void nr_dft(int32_t *z, int32_t *d, uint32_t Msc_PUSCH)
     case 1200:
       dft(DFT_1200,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
       break;
+
+    case 1296:
+      dft(DFT_1296,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 1440:
+      dft(DFT_1440,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 1500:
+      dft(DFT_1500,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 1536:
+      //dft(DFT_1536,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      dft(DFT_1536,(int16_t*)d, (int16_t*)z, 1);
+      break;
+
+    case 1620:
+      dft(DFT_1620,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 1728:
+      dft(DFT_1728,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 1800:
+      dft(DFT_1800,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 1920:
+      dft(DFT_1920,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 1944:
+      dft(DFT_1944,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 2160:
+      dft(DFT_2160,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 2304:
+      dft(DFT_2304,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 2400:
+      dft(DFT_2400,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 2592:
+      dft(DFT_2592,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 2700:
+      dft(DFT_2700,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 2880:
+      dft(DFT_2880,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 2916:
+      dft(DFT_2916,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 3000:
+      dft(DFT_3000,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    case 3072:
+      //dft(DFT_3072,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      dft(DFT_3072,(int16_t*)d, (int16_t*)z, 1);
+      break;
+
+    case 3240:
+      dft(DFT_3240,(int16_t*)dft_in0, (int16_t*)dft_out0, 1);
+      break;
+
+    default:
+      // should not be reached
+      LOG_E( PHY, "Unsupported Msc_PUSCH value of %"PRIu16"\n", Msc_PUSCH );
+      return;
+
   }
 
-  for (i = 0, ip = 0; i < Msc_PUSCH; i++, ip+=4) {
-    z[i] = dft_out0[ip];
+
+  if ((Msc_PUSCH % 1536) > 0) {
+    for (i = 0, ip = 0; i < Msc_PUSCH; i++, ip+=4)
+      z[i] = dft_out0[ip];
   }
+
 }
 
 

openair1/PHY/NR_TRANSPORT/nr_ulsch_demodulation.c

@@ -16,10 +16,10 @@ void nr_idft(int32_t *z, uint32_t Msc_PUSCH)
 {
 
 #if defined(__x86_64__) || defined(__i386__)
-  __m128i idft_in128[1][1200], idft_out128[1][1200];
+  __m128i idft_in128[1][3240], idft_out128[1][3240];
   __m128i norm128;
 #elif defined(__arm__)
-  int16x8_t idft_in128[1][1200], idft_out128[1][1200];
+  int16x8_t idft_in128[1][3240], idft_out128[1][3240];
   int16x8_t norm128;
 #endif
   int16_t *idft_in0 = (int16_t*)idft_in128[0], *idft_out0 = (int16_t*)idft_out128[0];
@@ -28,20 +28,19 @@ void nr_idft(int32_t *z, uint32_t Msc_PUSCH)
 
   LOG_T(PHY,"Doing lte_idft for Msc_PUSCH %d\n",Msc_PUSCH);
 
-  // conjugate input
-  for (i = 0; i < (Msc_PUSCH>>2); i++) {
+  if ((Msc_PUSCH % 1536) > 0) {
+    // conjugate input
+    for (i = 0; i < (Msc_PUSCH>>2); i++) {
 #if defined(__x86_64__)||defined(__i386__)
-    *&(((__m128i*)z)[i]) = _mm_sign_epi16(*&(((__m128i*)z)[i]), *(__m128i*)&conjugate2[0]);
+      *&(((__m128i*)z)[i]) = _mm_sign_epi16(*&(((__m128i*)z)[i]), *(__m128i*)&conjugate2[0]);
 #elif defined(__arm__)
-    *&(((int16x8_t*)z)[i]) = vmulq_s16(*&(((int16x8_t*)z)[i]), *(int16x8_t*)&conjugate2[0]);
+      *&(((int16x8_t*)z)[i]) = vmulq_s16(*&(((int16x8_t*)z)[i]), *(int16x8_t*)&conjugate2[0]);
 #endif
+    }
+    for (i = 0, ip = 0; i < Msc_PUSCH; i++, ip+=4)
+      ((uint32_t*)idft_in0)[ip+0] = z[i];
   }
 
-  for (i=0,ip=0; i<Msc_PUSCH; i++, ip+=4) {
-    ((int32_t*)idft_in0)[ip+0] = z[i];
-  }
-
-
   switch (Msc_PUSCH) {
     case 12:
       dft(DFT_12,(int16_t *)idft_in0, (int16_t *)idft_out0,0);
@@ -194,25 +193,102 @@ void nr_idft(int32_t *z, uint32_t Msc_PUSCH)
       dft(DFT_1200,idft_in0, idft_out0, 1);
       break;
 
+    case 1296:
+      dft(DFT_1296,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 1440:
+      dft(DFT_1440,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 1500:
+      dft(DFT_1500,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 1536:
+      //dft(DFT_1536,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      idft(IDFT_1536,(int16_t*)z, (int16_t*)z, 1);
+      break;
+
+    case 1620:
+      dft(DFT_1620,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 1728:
+      dft(DFT_1728,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 1800:
+      dft(DFT_1800,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 1920:
+      dft(DFT_1920,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 1944:
+      dft(DFT_1944,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 2160:
+      dft(DFT_2160,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 2304:
+      dft(DFT_2304,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 2400:
+      dft(DFT_2400,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 2592:
+      dft(DFT_2592,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 2700:
+      dft(DFT_2700,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 2880:
+      dft(DFT_2880,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 2916:
+      dft(DFT_2916,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 3000:
+      dft(DFT_3000,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
+    case 3072:
+      //dft(DFT_3072,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      idft(IDFT_3072,(int16_t*)z, (int16_t*)z, 1);
+      break;
+
+    case 3240:
+      dft(DFT_3240,(int16_t*)idft_in0, (int16_t*)idft_out0, 1);
+      break;
+
     default:
       // should not be reached
       LOG_E( PHY, "Unsupported Msc_PUSCH value of %"PRIu16"\n", Msc_PUSCH );
       return;
   }
 
+  if ((Msc_PUSCH % 1536) > 0) {
+    for (i = 0, ip = 0; i < Msc_PUSCH; i++, ip+=4)
+      z[i] = ((uint32_t*)idft_out0)[ip];
 
-
-  for (i = 0, ip = 0; i < Msc_PUSCH; i++, ip+=4) {
-    z[i] = ((int32_t*)idft_out0)[ip];
-  }
-
-  // conjugate output
-  for (i = 0; i < (Msc_PUSCH>>2); i++) {
+    // conjugate output
+    for (i = 0; i < (Msc_PUSCH>>2); i++) {
 #if defined(__x86_64__) || defined(__i386__)
-    ((__m128i*)z)[i] = _mm_sign_epi16(((__m128i*)z)[i], *(__m128i*)&conjugate2[0]);
+      ((__m128i*)z)[i] = _mm_sign_epi16(((__m128i*)z)[i], *(__m128i*)&conjugate2[0]);
 #elif defined(__arm__)
-    *&(((int16x8_t*)z)[i]) = vmulq_s16(*&(((int16x8_t*)z)[i]), *(int16x8_t*)&conjugate2[0]);
+      *&(((int16x8_t*)z)[i]) = vmulq_s16(*&(((int16x8_t*)z)[i]), *(int16x8_t*)&conjugate2[0]);
 #endif
+    }
   }
 
 #if defined(__x86_64__) || defined(__i386__)
@@ -222,6 +298,7 @@ void nr_idft(int32_t *z, uint32_t Msc_PUSCH)
 
 }
 
+
 void nr_ulsch_extract_rbs_single(int32_t **rxdataF,
                                  NR_gNB_PUSCH *pusch_vars,
                                  unsigned char symbol,

openair1/PHY/TOOLS/oai_dfts.c

@@ -5952,8 +5952,6 @@ void idft32768(int16_t *x,int16_t *y,unsigned char scale)
 
 
 #endif
-
-
 int16_t twa1536[1024],twb1536[1024];
 
 // 512 x 3
@@ -9471,145 +9469,1058 @@ void dft1200(int16_t *x,int16_t *y,unsigned char scale_flag)
 
 }
 
-void init_rad4(int N,int16_t *tw) {
 
-  int16_t *twa = tw;
-  int16_t *twb = twa+(N/2);
-  int16_t *twc = twb+(N/2);
-  int i;
+static int16_t twa1296[431*2*4];
+static int16_t twb1296[431*2*4];
 
-  for (i=0;i<(N/4);i++) {
-    *twa = (int16_t)round(32767.0*cos(2*M_PI*i/N)); twa++;
-    *twa = -(int16_t)round(32767.0*sin(2*M_PI*i/N)); twa++;
-    *twb = (int16_t)round(32767.0*cos(2*M_PI*2*i/N)); twb++;
-    *twb = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N)); twb++;
-    *twc = (int16_t)round(32767.0*cos(2*M_PI*3*i/N)); twc++;
-    *twc = -(int16_t)round(32767.0*sin(2*M_PI*3*i/N)); twc++;
-  }
-}
-void init_rad4_rep(int N,int16_t *twa,int16_t *twb,int16_t *twc) {
+void dft1296(int16_t *x,int16_t *y,unsigned char scale_flag) //432 * 3
+{
 
   int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa1296[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb1296[0];
+  simd_q15_t x2128[1296];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[1296];//=&ytmp128array3[0];
 
-  for (i=1;i<(N/4);i++) {
-    twa[0] = (int16_t)round(32767.0*cos(2*M_PI*i/N)); 
-    twa[1] = -(int16_t)round(32767.0*sin(2*M_PI*i/N));
-    twb[0] = (int16_t)round(32767.0*cos(2*M_PI*2*i/N));
-    twb[1] = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N));
-    twc[0] = (int16_t)round(32767.0*cos(2*M_PI*3*i/N));
-    twc[1] = -(int16_t)round(32767.0*sin(2*M_PI*3*i/N));
-    for (j=1;j<4;j++) {
-      ((int32_t*)twa)[j]=((int32_t*)twa)[0];
-      ((int32_t*)twb)[j]=((int32_t*)twb)[0];
-      ((int32_t*)twc)[j]=((int32_t*)twc)[0];
-    }
-    twa+=8;
-    twb+=8;
-    twc+=8;
+
+
+  for (i=0,j=0; i<432; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+432] = x128[j+1];
+    x2128[i+864] = x128[j+2];
   }
-}
 
-void init_rad2(int N,int16_t *tw) {
+  dft432((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft432((int16_t *)(x2128+432),(int16_t *)(ytmp128+432),1);
+  dft432((int16_t *)(x2128+864),(int16_t *)(ytmp128+864),1);
 
-  int16_t *twa = tw;
-  int i;
+  bfly3_tw1(ytmp128,ytmp128+432,ytmp128+864,y128,y128+432,y128+864);
 
-  for (i=0;i<(N>>1);i++) {
-    *twa = (int16_t)round(32767.0*cos(2*M_PI*i/N)); twa++;
-    *twa = -(int16_t)round(32767.0*sin(2*M_PI*i/N)); twa++;
+  for (i=1,j=0; i<432; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+432+i,
+          ytmp128+864+i,
+          y128+i,
+          y128+432+i,
+          y128+864+i,
+          twa128+j,
+          twb128+j);
   }
-}
 
-void init_rad2_rep(int N,int16_t *twa) {
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<1296; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
 
+static int16_t twa1440[479*2*4];
+static int16_t twb1440[479*2*4];
+
+void dft1440(int16_t *x,int16_t *y,unsigned char scale_flag)  // 480 x 3
+{
   int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa1440[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb1440[0];
+  simd_q15_t x2128[1440];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[1440];//=&ytmp128array3[0];
 
-  for (i=1;i<(N/2);i++) {
-    twa[0] = (int16_t)round(32767.0*cos(2*M_PI*i/N)); 
-    twa[1] = -(int16_t)round(32767.0*sin(2*M_PI*i/N));
-    for (j=1;j<4;j++) {
-      ((int32_t*)twa)[j]=((int32_t*)twa)[0];
+
+
+  for (i=0,j=0; i<480; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+480] = x128[j+1];
+    x2128[i+960] = x128[j+2];
+  }
+
+  dft480((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft480((int16_t *)(x2128+480),(int16_t *)(ytmp128+480),1);
+  dft480((int16_t *)(x2128+960),(int16_t *)(ytmp128+960),1);
+
+  bfly3_tw1(ytmp128,ytmp128+480,ytmp128+960,y128,y128+480,y128+960);
+
+  for (i=1,j=0; i<480; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+480+i,
+          ytmp128+960+i,
+          y128+i,
+          y128+480+i,
+          y128+960+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<1440; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
     }
-    twa+=8;
   }
-}
 
-void init_rad3(int N,int16_t *twa,int16_t *twb) {
+  _mm_empty();
+  _m_empty();
 
-  int i;
+};
 
-  for (i=0;i<(N/3);i++) {
-    *twa = (int16_t)round(32767.0*cos(2*M_PI*i/N)); twa++;
-    *twa = -(int16_t)round(32767.0*sin(2*M_PI*i/N)); twa++;
-    *twb = (int16_t)round(32767.0*cos(2*M_PI*2*i/N)); twb++;
-    *twb = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N)); twb++;
+static int16_t twa1500[2392]__attribute__((aligned(32)));
+static int16_t twb1500[2392]__attribute__((aligned(32)));
+static int16_t twc1500[2392]__attribute__((aligned(32)));
+static int16_t twd1500[2392]__attribute__((aligned(32)));
+
+void dft1500(int16_t *x,int16_t *y,unsigned char scale_flag)
+{
+
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa1500[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb1500[0];
+  simd_q15_t *twc128=(simd_q15_t *)&twc1500[0];
+  simd_q15_t *twd128=(simd_q15_t *)&twd1500[0];
+  simd_q15_t x2128[1500];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[1500];//=&ytmp128array2[0];
+
+
+
+  for (i=0,j=0; i<300; i++,j+=5) {
+    x2128[i]    = x128[j];
+    x2128[i+300] = x128[j+1];
+    x2128[i+600] = x128[j+2];
+    x2128[i+900] = x128[j+3];
+    x2128[i+1200] = x128[j+4];
+  }
+
+  dft300((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft300((int16_t *)(x2128+300),(int16_t *)(ytmp128+300),1);
+  dft300((int16_t *)(x2128+600),(int16_t *)(ytmp128+600),1);
+  dft300((int16_t *)(x2128+900),(int16_t *)(ytmp128+900),1);
+  dft300((int16_t *)(x2128+1200),(int16_t *)(ytmp128+1200),1);
+
+  bfly5_tw1(ytmp128,ytmp128+300,ytmp128+600,ytmp128+900,ytmp128+1200,y128,y128+300,y128+600,y128+900,y128+1200);
+
+  for (i=1,j=0; i<300; i++,j++) {
+    bfly5(ytmp128+i,
+          ytmp128+300+i,
+          ytmp128+600+i,
+          ytmp128+900+i,
+          ytmp128+1200+i,
+          y128+i,
+          y128+300+i,
+          y128+600+i,
+          y128+900+i,
+          y128+1200+i,
+          twa128+j,
+          twb128+j,
+          twc128+j,
+          twd128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[15]);
+
+    for (i=0; i<1500; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
   }
+
+  _mm_empty();
+  _m_empty();
+
 }
 
-void init_rad3_rep(int N,int16_t *twa,int16_t *twb) {
+static int16_t twa1620[539*2*4];
+static int16_t twb1620[539*2*4];
 
+void dft1620(int16_t *x,int16_t *y,unsigned char scale_flag)  // 540 x 3
+{
   int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa1620[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb1620[0];
+  simd_q15_t x2128[1620];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[1620];//=&ytmp128array3[0];
 
-  for (i=1;i<(N/3);i++) {
-    twa[0] = (int16_t)round(32767.0*cos(2*M_PI*i/N)); 
-    twa[1] = -(int16_t)round(32767.0*sin(2*M_PI*i/N));
-    twb[0] = (int16_t)round(32767.0*cos(2*M_PI*2*i/N));
-    twb[1] = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N));
-    for (j=1;j<4;j++) {
-      ((int32_t*)twa)[j]=((int32_t*)twa)[0];
-      ((int32_t*)twb)[j]=((int32_t*)twb)[0];
+
+
+  for (i=0,j=0; i<540; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+540] = x128[j+1];
+    x2128[i+1080] = x128[j+2];
+  }
+
+  dft540((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft540((int16_t *)(x2128+540),(int16_t *)(ytmp128+540),1);
+  dft540((int16_t *)(x2128+1080),(int16_t *)(ytmp128+1080),1);
+
+  bfly3_tw1(ytmp128,ytmp128+540,ytmp128+1080,y128,y128+540,y128+1080);
+
+  for (i=1,j=0; i<540; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+540+i,
+          ytmp128+1080+i,
+          y128+i,
+          y128+540+i,
+          y128+1080+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<1620; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
     }
-    twa+=8;
-    twb+=8;
   }
-}
 
-void init_rad5_rep(int N,int16_t *twa,int16_t *twb,int16_t *twc,int16_t *twd) {
+  _mm_empty();
+  _m_empty();
+
+};
 
+static int16_t twa1728[575*2*4];
+static int16_t twb1728[575*2*4];
+
+void dft1728(int16_t *x,int16_t *y,unsigned char scale_flag)  // 576 x 3
+{
   int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa1728[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb1728[0];
+  simd_q15_t x2128[1728];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[1728];//=&ytmp128array3[0];
 
-  for (i=1;i<(N/5);i++) {
-    twa[0] = (int16_t)round(32767.0*cos(2*M_PI*i/N)); 
-    twa[1] = -(int16_t)round(32767.0*sin(2*M_PI*i/N));
-    twb[0] = (int16_t)round(32767.0*cos(2*M_PI*2*i/N));
-    twb[1] = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N));
-    twc[0] = (int16_t)round(32767.0*cos(2*M_PI*3*i/N));
-    twc[1] = -(int16_t)round(32767.0*sin(2*M_PI*3*i/N));
-    twd[0] = (int16_t)round(32767.0*cos(2*M_PI*4*i/N));
-    twd[1] = -(int16_t)round(32767.0*sin(2*M_PI*4*i/N));
-    for (j=1;j<4;j++) {
-      ((int32_t*)twa)[j]=((int32_t*)twa)[0];
-      ((int32_t*)twb)[j]=((int32_t*)twb)[0];
-      ((int32_t*)twc)[j]=((int32_t*)twc)[0];
-      ((int32_t*)twd)[j]=((int32_t*)twd)[0];
+
+
+  for (i=0,j=0; i<576; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+576] = x128[j+1];
+    x2128[i+1152] = x128[j+2];
+  }
+
+  dft576((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft576((int16_t *)(x2128+576),(int16_t *)(ytmp128+576),1);
+  dft576((int16_t *)(x2128+1152),(int16_t *)(ytmp128+1152),1);
+
+  bfly3_tw1(ytmp128,ytmp128+576,ytmp128+1152,y128,y128+576,y128+1152);
+
+  for (i=1,j=0; i<576; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+576+i,
+          ytmp128+1152+i,
+          y128+i,
+          y128+576+i,
+          y128+1152+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<1728; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
     }
-    twa+=8;
-    twb+=8;
-    twc+=8;
-    twd+=8;                  
   }
-}
-/*----------------------------------------------------------------*/
-/* dft library entry points:                                      */
 
-int dfts_autoinit(void)
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa1800[599*2*4];
+static int16_t twb1800[599*2*4];
+
+void dft1800(int16_t *x,int16_t *y,unsigned char scale_flag)  // 600 x 3
 {
-  init_rad4(1024,tw1024);
-  init_rad2(2048,tw2048);
-  init_rad4(4096,tw4096);
-  init_rad2(8192,tw8192);
-  init_rad4(16384,tw16384);
-  init_rad2(32768,tw32768);
-  
-  init_rad3(1536,twa1536,twb1536);
-  init_rad3(3072,twa3072,twb3072);
-  init_rad3(6144,twa6144,twb6144);
-  init_rad3(12288,twa12288,twb12288);
-  init_rad3(18432,twa18432,twb18432);
-  init_rad3(24576,twa24576,twb24576);
-  init_rad3(36864,twa36864,twb36864);
-  init_rad3(49152,twa49152,twb49152);
-  init_rad3(98304,twa98304,twb98304);
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa1800[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb1800[0];
+  simd_q15_t x2128[1800];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[1800];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<600; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+600] = x128[j+1];
+    x2128[i+1200] = x128[j+2];
+  }
+
+  dft600((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft600((int16_t *)(x2128+600),(int16_t *)(ytmp128+600),1);
+  dft600((int16_t *)(x2128+1200),(int16_t *)(ytmp128+1200),1);
+
+  bfly3_tw1(ytmp128,ytmp128+600,ytmp128+1200,y128,y128+600,y128+1200);
+
+  for (i=1,j=0; i<600; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+600+i,
+          ytmp128+1200+i,
+          y128+i,
+          y128+600+i,
+          y128+1200+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<1800; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa1920[479*2*4];
+static int16_t twb1920[479*2*4];
+static int16_t twc1920[479*2*4];
+
+void dft1920(int16_t *x,int16_t *y,unsigned char scale_flag)  // 480 x 4
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa1920[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb1920[0];
+  simd_q15_t *twc128=(simd_q15_t *)&twc1920[0];
+  simd_q15_t x2128[1920];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[1920];//=&ytmp128array2[0];
+
+
+
+  for (i=0,j=0; i<480; i++,j+=4) {
+    x2128[i]    = x128[j];
+    x2128[i+480] = x128[j+1];
+    x2128[i+960] = x128[j+2];
+    x2128[i+1440] = x128[j+3];
+  }
+
+  dft480((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft480((int16_t *)(x2128+480),(int16_t *)(ytmp128+480),1);
+  dft480((int16_t *)(x2128+960),(int16_t *)(ytmp128+960),1);
+  dft480((int16_t *)(x2128+1440),(int16_t *)(ytmp128+1440),1);
+
+  bfly4_tw1(ytmp128,ytmp128+480,ytmp128+960,ytmp128+1440,y128,y128+480,y128+960,y128+1440);
+
+  for (i=1,j=0; i<480; i++,j++) {
+    bfly4(ytmp128+i,
+          ytmp128+480+i,
+          ytmp128+960+i,
+          ytmp128+1440+i,
+          y128+i,
+          y128+480+i,
+          y128+960+i,
+          y128+1440+i,
+          twa128+j,
+          twb128+j,
+          twc128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[13]);
+    for (i=0; i<1920; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa1944[647*2*4];
+static int16_t twb1944[647*2*4];
+
+void dft1944(int16_t *x,int16_t *y,unsigned char scale_flag)  // 648 x 3
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa1944[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb1944[0];
+  simd_q15_t x2128[1944];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[1944];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<648; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+648] = x128[j+1];
+    x2128[i+1296] = x128[j+2];
+  }
+
+  dft648((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft648((int16_t *)(x2128+648),(int16_t *)(ytmp128+648),1);
+  dft648((int16_t *)(x2128+1296),(int16_t *)(ytmp128+1296),1);
+
+  bfly3_tw1(ytmp128,ytmp128+648,ytmp128+1296,y128,y128+648,y128+1296);
+
+  for (i=1,j=0; i<648; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+648+i,
+          ytmp128+1296+i,
+          y128+i,
+          y128+648+i,
+          y128+1296+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<1944; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa2160[719*2*4];
+static int16_t twb2160[719*2*4];
+
+void dft2160(int16_t *x,int16_t *y,unsigned char scale_flag)  // 720 x 3
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa2160[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb2160[0];
+  simd_q15_t x2128[2160];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[2160];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<720; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+720] = x128[j+1];
+    x2128[i+1440] = x128[j+2];
+  }
+
+  dft720((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft720((int16_t *)(x2128+720),(int16_t *)(ytmp128+720),1);
+  dft720((int16_t *)(x2128+1440),(int16_t *)(ytmp128+1440),1);
+
+  bfly3_tw1(ytmp128,ytmp128+720,ytmp128+1440,y128,y128+720,y128+1440);
+
+  for (i=1,j=0; i<720; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+720+i,
+          ytmp128+1440+i,
+          y128+i,
+          y128+720+i,
+          y128+1440+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<2160; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa2304[767*2*4];
+static int16_t twb2304[767*2*4];
+
+void dft2304(int16_t *x,int16_t *y,unsigned char scale_flag)  // 768 x 3
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa2304[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb2304[0];
+  simd_q15_t x2128[2304];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[2304];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<768; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+768] = x128[j+1];
+    x2128[i+1536] = x128[j+2];
+  }
+
+  dft768((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft768((int16_t *)(x2128+768),(int16_t *)(ytmp128+768),1);
+  dft768((int16_t *)(x2128+1536),(int16_t *)(ytmp128+1536),1);
+
+  bfly3_tw1(ytmp128,ytmp128+768,ytmp128+1536,y128,y128+768,y128+1536);
+
+  for (i=1,j=0; i<768; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+768+i,
+          ytmp128+1536+i,
+          y128+i,
+          y128+768+i,
+          y128+1536+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<2304; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa2400[599*2*4];
+static int16_t twb2400[599*2*4];
+static int16_t twc2400[599*2*4];
+
+void dft2400(int16_t *x,int16_t *y,unsigned char scale_flag)  // 600 x 4
+{
+
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa2400[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb2400[0];
+  simd_q15_t *twc128=(simd_q15_t *)&twc2400[0];
+  simd_q15_t x2128[2400];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[2400];//=&ytmp128array2[0];
+
+
+
+  for (i=0,j=0; i<600; i++,j+=4) {
+    x2128[i]    = x128[j];
+    x2128[i+600] = x128[j+1];
+    x2128[i+1200] = x128[j+2];
+    x2128[i+1800] = x128[j+3];
+  }
+
+  dft600((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft600((int16_t *)(x2128+600),(int16_t *)(ytmp128+600),1);
+  dft600((int16_t *)(x2128+1200),(int16_t *)(ytmp128+1200),1);
+  dft600((int16_t *)(x2128+1800),(int16_t *)(ytmp128+1800),1);
+
+  bfly4_tw1(ytmp128,ytmp128+600,ytmp128+1200,ytmp128+1800,y128,y128+600,y128+1200,y128+1800);
+
+  for (i=1,j=0; i<600; i++,j++) {
+    bfly4(ytmp128+i,
+          ytmp128+600+i,
+          ytmp128+1200+i,
+          ytmp128+1800+i,
+          y128+i,
+          y128+600+i,
+          y128+1200+i,
+          y128+1800+i,
+          twa128+j,
+          twb128+j,
+          twc128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[13]);
+    for (i=0; i<2400; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa2592[863*2*4];
+static int16_t twb2592[863*2*4];
+
+void dft2592(int16_t *x,int16_t *y,unsigned char scale_flag)  // 864 x 3
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa2592[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb2592[0];
+  simd_q15_t x2128[2592];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[2592];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<864; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+864] = x128[j+1];
+    x2128[i+1728] = x128[j+2];
+  }
+
+  dft864((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft864((int16_t *)(x2128+864),(int16_t *)(ytmp128+864),1);
+  dft864((int16_t *)(x2128+1728),(int16_t *)(ytmp128+1728),1);
+
+  bfly3_tw1(ytmp128,ytmp128+864,ytmp128+1728,y128,y128+864,y128+1728);
+
+  for (i=1,j=0; i<864; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+864+i,
+          ytmp128+1728+i,
+          y128+i,
+          y128+864+i,
+          y128+1728+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<2592; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa2700[899*2*4];
+static int16_t twb2700[899*2*4];
+
+void dft2700(int16_t *x,int16_t *y,unsigned char scale_flag)  // 900 x 3
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa2700[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb2700[0];
+  simd_q15_t x2128[2700];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[2700];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<900; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+900] = x128[j+1];
+    x2128[i+1800] = x128[j+2];
+  }
+
+  dft900((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft900((int16_t *)(x2128+900),(int16_t *)(ytmp128+900),1);
+  dft900((int16_t *)(x2128+1800),(int16_t *)(ytmp128+1800),1);
+
+  bfly3_tw1(ytmp128,ytmp128+900,ytmp128+1800,y128,y128+900,y128+1800);
+
+  for (i=1,j=0; i<900; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+900+i,
+          ytmp128+1800+i,
+          y128+i,
+          y128+900+i,
+          y128+1800+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<2700; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa2880[959*2*4];
+static int16_t twb2880[959*2*4];
+
+void dft2880(int16_t *x,int16_t *y,unsigned char scale_flag)  // 960 x 3
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa2880[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb2880[0];
+  simd_q15_t x2128[2880];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[2880];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<960; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+960] = x128[j+1];
+    x2128[i+1920] = x128[j+2];
+  }
+
+  dft960((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft960((int16_t *)(x2128+960),(int16_t *)(ytmp128+960),1);
+  dft960((int16_t *)(x2128+1920),(int16_t *)(ytmp128+1920),1);
+
+  bfly3_tw1(ytmp128,ytmp128+960,ytmp128+1920,y128,y128+960,y128+1920);
+
+  for (i=1,j=0; i<960; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+960+i,
+          ytmp128+1920+i,
+          y128+i,
+          y128+960+i,
+          y128+1920+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<2880; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa2916[971*2*4];
+static int16_t twb2916[971*2*4];
+
+void dft2916(int16_t *x,int16_t *y,unsigned char scale_flag)  // 972 x 3
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa2916[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb2916[0];
+  simd_q15_t x2128[2916];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[2916];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<972; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+972] = x128[j+1];
+    x2128[i+1944] = x128[j+2];
+  }
+
+  dft972((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft972((int16_t *)(x2128+972),(int16_t *)(ytmp128+972),1);
+  dft972((int16_t *)(x2128+1944),(int16_t *)(ytmp128+1944),1);
+
+  bfly3_tw1(ytmp128,ytmp128+972,ytmp128+1944,y128,y128+972,y128+1944);
+
+  for (i=1,j=0; i<972; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+972+i,
+          ytmp128+1944+i,
+          y128+i,
+          y128+972+i,
+          y128+1944+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<2916; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+static int16_t twa3000[599*8]__attribute__((aligned(32)));
+static int16_t twb3000[599*8]__attribute__((aligned(32)));
+static int16_t twc3000[599*8]__attribute__((aligned(32)));
+static int16_t twd3000[599*8]__attribute__((aligned(32)));
+
+void dft3000(int16_t *x,int16_t *y,unsigned char scale_flag) // 600 * 5
+{
+
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa3000[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb3000[0];
+  simd_q15_t *twc128=(simd_q15_t *)&twc3000[0];
+  simd_q15_t *twd128=(simd_q15_t *)&twd3000[0];
+  simd_q15_t x2128[3000];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[3000];//=&ytmp128array2[0];
+
+
+
+  for (i=0,j=0; i<600; i++,j+=5) {
+    x2128[i]    = x128[j];
+    x2128[i+600] = x128[j+1];
+    x2128[i+1200] = x128[j+2];
+    x2128[i+1800] = x128[j+3];
+    x2128[i+2400] = x128[j+4];
+  }
+
+  dft600((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft600((int16_t *)(x2128+600),(int16_t *)(ytmp128+600),1);
+  dft600((int16_t *)(x2128+1200),(int16_t *)(ytmp128+1200),1);
+  dft600((int16_t *)(x2128+1800),(int16_t *)(ytmp128+1800),1);
+  dft600((int16_t *)(x2128+2400),(int16_t *)(ytmp128+2400),1);
+
+  bfly5_tw1(ytmp128,ytmp128+600,ytmp128+1200,ytmp128+1800,ytmp128+2400,y128,y128+600,y128+1200,y128+1800,y128+2400);
+
+  for (i=1,j=0; i<600; i++,j++) {
+    bfly5(ytmp128+i,
+          ytmp128+600+i,
+          ytmp128+1200+i,
+          ytmp128+1800+i,
+          ytmp128+2400+i,
+          y128+i,
+          y128+600+i,
+          y128+1200+i,
+          y128+1800+i,
+          y128+2400+i,
+          twa128+j,
+          twb128+j,
+          twc128+j,
+          twd128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[15]);
+
+    for (i=0; i<3000; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+}
+
+static int16_t twa3240[1079*2*4];
+static int16_t twb3240[1079*2*4];
+
+void dft3240(int16_t *x,int16_t *y,unsigned char scale_flag)  // 1080 x 3
+{
+  int i,j;
+  simd_q15_t *x128=(simd_q15_t *)x;
+  simd_q15_t *y128=(simd_q15_t *)y;
+  simd_q15_t *twa128=(simd_q15_t *)&twa3240[0];
+  simd_q15_t *twb128=(simd_q15_t *)&twb3240[0];
+  simd_q15_t x2128[3240];// = (simd_q15_t *)&x2128array[0];
+  simd_q15_t ytmp128[3240];//=&ytmp128array3[0];
+
+
+
+  for (i=0,j=0; i<1080; i++,j+=3) {
+    x2128[i]    = x128[j];
+    x2128[i+1080] = x128[j+1];
+    x2128[i+2160] = x128[j+2];
+  }
+
+  dft1080((int16_t *)x2128,(int16_t *)ytmp128,1);
+  dft1080((int16_t *)(x2128+1080),(int16_t *)(ytmp128+1080),1);
+  dft1080((int16_t *)(x2128+2160),(int16_t *)(ytmp128+2160),1);
+
+  bfly3_tw1(ytmp128,ytmp128+1080,ytmp128+2160,y128,y128+1080,y128+2160);
+
+  for (i=1,j=0; i<1080; i++,j++) {
+    bfly3(ytmp128+i,
+          ytmp128+1080+i,
+          ytmp128+2160+i,
+          y128+i,
+          y128+1080+i,
+          y128+2160+i,
+          twa128+j,
+          twb128+j);
+  }
+
+  if (scale_flag==1) {
+    norm128 = set1_int16(dft_norm_table[14]);
+
+    for (i=0; i<3240; i++) {
+      y128[i] = mulhi_int16(y128[i],norm128);
+    }
+  }
+
+  _mm_empty();
+  _m_empty();
+
+};
+
+void init_rad4(int N,int16_t *tw) {
+
+  int16_t *twa = tw;
+  int16_t *twb = twa+(N/2);
+  int16_t *twc = twb+(N/2);
+  int i;
+
+  for (i=0;i<(N/4);i++) {
+    *twa = (int16_t)round(32767.0*cos(2*M_PI*i/N)); twa++;
+    *twa = -(int16_t)round(32767.0*sin(2*M_PI*i/N)); twa++;
+    *twb = (int16_t)round(32767.0*cos(2*M_PI*2*i/N)); twb++;
+    *twb = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N)); twb++;
+    *twc = (int16_t)round(32767.0*cos(2*M_PI*3*i/N)); twc++;
+    *twc = -(int16_t)round(32767.0*sin(2*M_PI*3*i/N)); twc++;
+  }
+}
+void init_rad4_rep(int N,int16_t *twa,int16_t *twb,int16_t *twc) {
+
+  int i,j;
+
+  for (i=1;i<(N/4);i++) {
+    twa[0] = (int16_t)round(32767.0*cos(2*M_PI*i/N));
+    twa[1] = -(int16_t)round(32767.0*sin(2*M_PI*i/N));
+    twb[0] = (int16_t)round(32767.0*cos(2*M_PI*2*i/N));
+    twb[1] = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N));
+    twc[0] = (int16_t)round(32767.0*cos(2*M_PI*3*i/N));
+    twc[1] = -(int16_t)round(32767.0*sin(2*M_PI*3*i/N));
+    for (j=1;j<4;j++) {
+      ((int32_t*)twa)[j]=((int32_t*)twa)[0];
+      ((int32_t*)twb)[j]=((int32_t*)twb)[0];
+      ((int32_t*)twc)[j]=((int32_t*)twc)[0];
+    }
+    twa+=8;
+    twb+=8;
+    twc+=8;
+  }
+}
+
+void init_rad2(int N,int16_t *tw) {
+
+  int16_t *twa = tw;
+  int i;
+
+  for (i=0;i<(N>>1);i++) {
+    *twa = (int16_t)round(32767.0*cos(2*M_PI*i/N)); twa++;
+    *twa = -(int16_t)round(32767.0*sin(2*M_PI*i/N)); twa++;
+  }
+}
+
+void init_rad2_rep(int N,int16_t *twa) {
+
+  int i,j;
+
+  for (i=1;i<(N/2);i++) {
+    twa[0] = (int16_t)round(32767.0*cos(2*M_PI*i/N));
+    twa[1] = -(int16_t)round(32767.0*sin(2*M_PI*i/N));
+    for (j=1;j<4;j++) {
+      ((int32_t*)twa)[j]=((int32_t*)twa)[0];
+    }
+    twa+=8;
+  }
+}
+
+void init_rad3(int N,int16_t *twa,int16_t *twb) {
+
+  int i;
+
+  for (i=0;i<(N/3);i++) {
+    *twa = (int16_t)round(32767.0*cos(2*M_PI*i/N)); twa++;
+    *twa = -(int16_t)round(32767.0*sin(2*M_PI*i/N)); twa++;
+    *twb = (int16_t)round(32767.0*cos(2*M_PI*2*i/N)); twb++;
+    *twb = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N)); twb++;
+  }
+}
+
+void init_rad3_rep(int N,int16_t *twa,int16_t *twb) {
+
+  int i,j;
+
+  for (i=1;i<(N/3);i++) {
+    twa[0] = (int16_t)round(32767.0*cos(2*M_PI*i/N));
+    twa[1] = -(int16_t)round(32767.0*sin(2*M_PI*i/N));
+    twb[0] = (int16_t)round(32767.0*cos(2*M_PI*2*i/N));
+    twb[1] = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N));
+    for (j=1;j<4;j++) {
+      ((int32_t*)twa)[j]=((int32_t*)twa)[0];
+      ((int32_t*)twb)[j]=((int32_t*)twb)[0];
+    }
+    twa+=8;
+    twb+=8;
+  }
+}
+
+void init_rad5_rep(int N,int16_t *twa,int16_t *twb,int16_t *twc,int16_t *twd) {
+
+  int i,j;
+
+  for (i=1;i<(N/5);i++) {
+    twa[0] = (int16_t)round(32767.0*cos(2*M_PI*i/N));
+    twa[1] = -(int16_t)round(32767.0*sin(2*M_PI*i/N));
+    twb[0] = (int16_t)round(32767.0*cos(2*M_PI*2*i/N));
+    twb[1] = -(int16_t)round(32767.0*sin(2*M_PI*2*i/N));
+    twc[0] = (int16_t)round(32767.0*cos(2*M_PI*3*i/N));
+    twc[1] = -(int16_t)round(32767.0*sin(2*M_PI*3*i/N));
+    twd[0] = (int16_t)round(32767.0*cos(2*M_PI*4*i/N));
+    twd[1] = -(int16_t)round(32767.0*sin(2*M_PI*4*i/N));
+    for (j=1;j<4;j++) {
+      ((int32_t*)twa)[j]=((int32_t*)twa)[0];
+      ((int32_t*)twb)[j]=((int32_t*)twb)[0];
+      ((int32_t*)twc)[j]=((int32_t*)twc)[0];
+      ((int32_t*)twd)[j]=((int32_t*)twd)[0];
+    }
+    twa+=8;
+    twb+=8;
+    twc+=8;
+    twd+=8;
+  }
+}
+/*----------------------------------------------------------------*/
+/* dft library entry points:                                      */
+
+int dfts_autoinit(void)
+{
+  init_rad4(1024,tw1024);
+  init_rad2(2048,tw2048);
+  init_rad4(4096,tw4096);
+  init_rad2(8192,tw8192);
+  init_rad4(16384,tw16384);
+  init_rad2(32768,tw32768);
+
+  init_rad3(1536,twa1536,twb1536);
+  init_rad3(3072,twa3072,twb3072);
+  init_rad3(6144,twa6144,twb6144);
+  init_rad3(12288,twa12288,twb12288);
+  init_rad3(18432,twa18432,twb18432);
+  init_rad3(24576,twa24576,twb24576);
+  init_rad3(36864,twa36864,twb36864);
+  init_rad3(49152,twa49152,twb49152);
+  init_rad3(98304,twa98304,twb98304);
 
 
   init_rad2_rep(24,tw24);
@@ -9645,6 +10556,24 @@ int dfts_autoinit(void)
   init_rad3_rep(1080,twa1080,twb1080);
   init_rad4_rep(1152,twa1152,twb1152,twc1152);
   init_rad4_rep(1200,twa1200,twb1200,twc1200);
+  init_rad3_rep(1296,twa1296,twb1296);
+  init_rad3_rep(1440,twa1440,twb1440);
+  init_rad5_rep(1500,twa1500,twb1500,twc1500,twd1500);
+  init_rad3_rep(1620,twa1620,twb1620);
+  init_rad3_rep(1728,twa1728,twb1728);
+  init_rad3_rep(1800,twa1800,twb1800);
+  init_rad4_rep(1920,twa1920,twb1920, twc1920);
+  init_rad3_rep(1944,twa1944,twb1944);
+  init_rad3_rep(2160,twa2160,twb2160);
+  init_rad3_rep(2304,twa2304,twb2304);
+  init_rad4_rep(2400,twa2400,twb2400,twc2400);
+  init_rad3_rep(2592,twa2592,twb2592);
+  init_rad3_rep(2700,twa2700,twb2700);
+  init_rad3_rep(2880,twa2880,twb2880);
+  init_rad3_rep(2916,twa2916,twb2916);
+  init_rad5_rep(3000,twa3000,twb3000,twc3000,twd3000);
+  init_rad3_rep(3240,twa3240,twb3240);
+
   return 0;
 }
 

openair1/PHY/TOOLS/tools_defs.h

@@ -226,9 +226,26 @@ void dft1024(int16_t *x,int16_t *y,uint8_t scale_flag);
 void dft1080(int16_t *x,int16_t *y,uint8_t scale_flag);
 void dft1152(int16_t *x,int16_t *y,uint8_t scale_flag);
 void dft1200(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft1296(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft1440(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft1500(int16_t *x,int16_t *y,uint8_t scale_flag);
 void dft1536(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
+void dft1620(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft1728(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft1800(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft1920(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft1944(int16_t *x,int16_t *y,uint8_t scale_flag);
 void dft2048(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft2160(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft2304(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft2400(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft2592(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft2700(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft2880(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft2916(int16_t *x,int16_t *y,uint8_t scale_flag);
+void dft3000(int16_t *x,int16_t *y,uint8_t scale_flag);
 void dft3072(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
+void dft3240(int16_t *x,int16_t *y,uint8_t scale_flag);
 void dft4096(int16_t *x,int16_t *y,uint8_t scale_flag);
 void dft6144(int16_t *sigF,int16_t *sig,uint8_t scale_flag);
 void dft8192(int16_t *x,int16_t *y,uint8_t scale_flag);
@@ -283,21 +300,25 @@ typedef enum DFT_size_idx {
 	DFT_108,   DFT_120,   DFT_128,  DFT_144,    DFT_180,  DFT_192,  DFT_216,   DFT_240,
 	DFT_256,   DFT_288,   DFT_300,  DFT_324,    DFT_360,  DFT_384,  DFT_432,   DFT_480,
 	DFT_512,   DFT_540,   DFT_576,  DFT_600,    DFT_648,  DFT_720,  DFT_768,   DFT_864,
-	DFT_900,   DFT_960,   DFT_972,  DFT_1024,   DFT_1080, DFT_1152, DFT_1200,  DFT_1536,
-	DFT_2048,  DFT_3072,  DFT_4096, DFT_6144,   DFT_8192, DFT_9216, DFT_12288, DFT_18432, 
-	DFT_24576, DFT_36864, DFT_49152, DFT_73728, DFT_98304,
+	DFT_900,   DFT_960,   DFT_972,  DFT_1024,   DFT_1080, DFT_1152, DFT_1200,  DFT_1296,
+	DFT_1440,  DFT_1500,  DFT_1536, DFT_1620,   DFT_1728, DFT_1800, DFT_1920,  DFT_1944,
+	DFT_2048,  DFT_2160,  DFT_2304, DFT_2400,   DFT_2592, DFT_2700, DFT_2880,  DFT_2916,
+	DFT_3000,  DFT_3072,  DFT_3240, DFT_4096,   DFT_6144, DFT_8192, DFT_9216,  DFT_12288,
+	DFT_18432, DFT_24576, DFT_36864, DFT_49152, DFT_73728, DFT_98304,
 	DFT_SIZE_IDXTABLESIZE
 } dft_size_idx_t;
 
 #ifdef OAIDFTS_MAIN
 adftfunc_t dft_ftab[]={
 	dft12,    dft24,    dft36,    dft48,    dft60,   dft72,   dft96,
-	dft108,   dft120,   dft128,   dft144,   dft180,  dft192,  dft216,   dft240,   
+	dft108,   dft120,   dft128,   dft144,   dft180,  dft192,  dft216,   dft240,
 	dft256,   dft288,   dft300,   dft324,   dft360,  dft384,  dft432,   dft480,
 	dft512,   dft540,   dft576,   dft600,   dft648,  dft720,  dft768,   dft864,
-	dft900,   dft960,   dft972,   dft1024,  dft1080, dft1152, dft1200,  dft1536,
-	dft2048,  dft3072,  dft4096,  dft6144,  dft8192, dft9216, dft12288, dft18432, 
-	dft24576, dft36864, dft49152, dft73728, dft98304
+	dft900,   dft960,   dft972,   dft1024,  dft1080, dft1152, dft1200,  dft1296,
+	dft1440,  dft1500,  dft1536,  dft1620,  dft1728, dft1800, dft1920,  dft1944,
+	dft2048,  dft2160,  dft2304,  dft2400,  dft2592, dft2700, dft2880,  dft2916,
+	dft3000,  dft3072,  dft3240,  dft4096,  dft6144, dft8192, dft9216,  dft12288,
+	dft18432, dft24576, dft36864, dft49152, dft73728, dft98304
 };
 #endif
 
@@ -309,9 +330,9 @@ typedef enum idft_size_idx {
 } idft_size_idx_t;
 #ifdef OAIDFTS_MAIN
 aidftfunc_t idft_ftab[]={
-        idft128,   idft256,  idft512,   idft1024,  idft1536,  idft2048,  idft3072,  idft4096,
-	    idft6144,  idft8192, idft9216,  idft12288, idft18432, idft24576, idft36864, idft49152, 
-	    idft73728, idft98304
+	idft128,   idft256,  idft512,   idft1024,  idft1536,  idft2048,  idft3072,  idft4096,
+	idft6144,  idft8192, idft9216,  idft12288, idft18432, idft24576, idft36864, idft49152,
+	idft73728, idft98304
 };
 #endif