arm_cfft_radix8_f32.c

00001 /* ----------------------------------------------------------------------
00002  * Project:      CMSIS DSP Library
00003  * Title:        arm_cfft_radix8_f32.c
00004  * Description:  Radix-8 Decimation in Frequency CFFT & CIFFT Floating point processing function
00005  *
00006  * $Date:        27. January 2017
00007  * $Revision:    V.1.5.1
00008  *
00009  * Target Processor: Cortex-M cores
00010  * -------------------------------------------------------------------- */
00011 /*
00012  * Copyright (C) 2010-2017 ARM Limited or its affiliates. All rights reserved.
00013  *
00014  * SPDX-License-Identifier: Apache-2.0
00015  *
00016  * Licensed under the Apache License, Version 2.0 (the License); you may
00017  * not use this file except in compliance with the License.
00018  * You may obtain a copy of the License at
00019  *
00020  * www.apache.org/licenses/LICENSE-2.0
00021  *
00022  * Unless required by applicable law or agreed to in writing, software
00023  * distributed under the License is distributed on an AS IS BASIS, WITHOUT
00024  * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
00025  * See the License for the specific language governing permissions and
00026  * limitations under the License.
00027  */
00028 
00029 #include "arm_math.h"
00030 
00031 
00032 /* ----------------------------------------------------------------------
00033  * Internal helper function used by the FFTs
00034  * -------------------------------------------------------------------- */
00035 
00036 /*
00037 * @brief  Core function for the floating-point CFFT butterfly process.
00038 * @param[in, out] *pSrc            points to the in-place buffer of floating-point data type.
00039 * @param[in]      fftLen           length of the FFT.
00040 * @param[in]      *pCoef           points to the twiddle coefficient buffer.
00041 * @param[in]      twidCoefModifier twiddle coefficient modifier that supports different size FFTs with the same twiddle factor table.
00042 * @return none.
00043 */
00044 
00045 void arm_radix8_butterfly_f32(
00046 float32_t * pSrc,
00047 uint16_t fftLen,
00048 const float32_t * pCoef,
00049 uint16_t twidCoefModifier)
00050 {
00051    uint32_t ia1, ia2, ia3, ia4, ia5, ia6, ia7;
00052    uint32_t i1, i2, i3, i4, i5, i6, i7, i8;
00053    uint32_t id;
00054    uint32_t n1, n2, j;
00055 
00056    float32_t r1, r2, r3, r4, r5, r6, r7, r8;
00057    float32_t t1, t2;
00058    float32_t s1, s2, s3, s4, s5, s6, s7, s8;
00059    float32_t p1, p2, p3, p4;
00060    float32_t co2, co3, co4, co5, co6, co7, co8;
00061    float32_t si2, si3, si4, si5, si6, si7, si8;
00062    const float32_t C81 = 0.70710678118f;
00063 
00064    n2 = fftLen;
00065 
00066    do
00067    {
00068       n1 = n2;
00069       n2 = n2 >> 3;
00070       i1 = 0;
00071 
00072       do
00073       {
00074          i2 = i1 + n2;
00075          i3 = i2 + n2;
00076          i4 = i3 + n2;
00077          i5 = i4 + n2;
00078          i6 = i5 + n2;
00079          i7 = i6 + n2;
00080          i8 = i7 + n2;
00081          r1 = pSrc[2 * i1] + pSrc[2 * i5];
00082          r5 = pSrc[2 * i1] - pSrc[2 * i5];
00083          r2 = pSrc[2 * i2] + pSrc[2 * i6];
00084          r6 = pSrc[2 * i2] - pSrc[2 * i6];
00085          r3 = pSrc[2 * i3] + pSrc[2 * i7];
00086          r7 = pSrc[2 * i3] - pSrc[2 * i7];
00087          r4 = pSrc[2 * i4] + pSrc[2 * i8];
00088          r8 = pSrc[2 * i4] - pSrc[2 * i8];
00089          t1 = r1 - r3;
00090          r1 = r1 + r3;
00091          r3 = r2 - r4;
00092          r2 = r2 + r4;
00093          pSrc[2 * i1] = r1 + r2;
00094          pSrc[2 * i5] = r1 - r2;
00095          r1 = pSrc[2 * i1 + 1] + pSrc[2 * i5 + 1];
00096          s5 = pSrc[2 * i1 + 1] - pSrc[2 * i5 + 1];
00097          r2 = pSrc[2 * i2 + 1] + pSrc[2 * i6 + 1];
00098          s6 = pSrc[2 * i2 + 1] - pSrc[2 * i6 + 1];
00099          s3 = pSrc[2 * i3 + 1] + pSrc[2 * i7 + 1];
00100          s7 = pSrc[2 * i3 + 1] - pSrc[2 * i7 + 1];
00101          r4 = pSrc[2 * i4 + 1] + pSrc[2 * i8 + 1];
00102          s8 = pSrc[2 * i4 + 1] - pSrc[2 * i8 + 1];
00103          t2 = r1 - s3;
00104          r1 = r1 + s3;
00105          s3 = r2 - r4;
00106          r2 = r2 + r4;
00107          pSrc[2 * i1 + 1] = r1 + r2;
00108          pSrc[2 * i5 + 1] = r1 - r2;
00109          pSrc[2 * i3]     = t1 + s3;
00110          pSrc[2 * i7]     = t1 - s3;
00111          pSrc[2 * i3 + 1] = t2 - r3;
00112          pSrc[2 * i7 + 1] = t2 + r3;
00113          r1 = (r6 - r8) * C81;
00114          r6 = (r6 + r8) * C81;
00115          r2 = (s6 - s8) * C81;
00116          s6 = (s6 + s8) * C81;
00117          t1 = r5 - r1;
00118          r5 = r5 + r1;
00119          r8 = r7 - r6;
00120          r7 = r7 + r6;
00121          t2 = s5 - r2;
00122          s5 = s5 + r2;
00123          s8 = s7 - s6;
00124          s7 = s7 + s6;
00125          pSrc[2 * i2]     = r5 + s7;
00126          pSrc[2 * i8]     = r5 - s7;
00127          pSrc[2 * i6]     = t1 + s8;
00128          pSrc[2 * i4]     = t1 - s8;
00129          pSrc[2 * i2 + 1] = s5 - r7;
00130          pSrc[2 * i8 + 1] = s5 + r7;
00131          pSrc[2 * i6 + 1] = t2 - r8;
00132          pSrc[2 * i4 + 1] = t2 + r8;
00133 
00134          i1 += n1;
00135       } while (i1 < fftLen);
00136 
00137       if (n2 < 8)
00138          break;
00139 
00140       ia1 = 0;
00141       j = 1;
00142 
00143       do
00144       {
00145          /*  index calculation for the coefficients */
00146          id  = ia1 + twidCoefModifier;
00147          ia1 = id;
00148          ia2 = ia1 + id;
00149          ia3 = ia2 + id;
00150          ia4 = ia3 + id;
00151          ia5 = ia4 + id;
00152          ia6 = ia5 + id;
00153          ia7 = ia6 + id;
00154 
00155          co2 = pCoef[2 * ia1];
00156          co3 = pCoef[2 * ia2];
00157          co4 = pCoef[2 * ia3];
00158          co5 = pCoef[2 * ia4];
00159          co6 = pCoef[2 * ia5];
00160          co7 = pCoef[2 * ia6];
00161          co8 = pCoef[2 * ia7];
00162          si2 = pCoef[2 * ia1 + 1];
00163          si3 = pCoef[2 * ia2 + 1];
00164          si4 = pCoef[2 * ia3 + 1];
00165          si5 = pCoef[2 * ia4 + 1];
00166          si6 = pCoef[2 * ia5 + 1];
00167          si7 = pCoef[2 * ia6 + 1];
00168          si8 = pCoef[2 * ia7 + 1];
00169 
00170          i1 = j;
00171 
00172          do
00173          {
00174             /*  index calculation for the input */
00175             i2 = i1 + n2;
00176             i3 = i2 + n2;
00177             i4 = i3 + n2;
00178             i5 = i4 + n2;
00179             i6 = i5 + n2;
00180             i7 = i6 + n2;
00181             i8 = i7 + n2;
00182             r1 = pSrc[2 * i1] + pSrc[2 * i5];
00183             r5 = pSrc[2 * i1] - pSrc[2 * i5];
00184             r2 = pSrc[2 * i2] + pSrc[2 * i6];
00185             r6 = pSrc[2 * i2] - pSrc[2 * i6];
00186             r3 = pSrc[2 * i3] + pSrc[2 * i7];
00187             r7 = pSrc[2 * i3] - pSrc[2 * i7];
00188             r4 = pSrc[2 * i4] + pSrc[2 * i8];
00189             r8 = pSrc[2 * i4] - pSrc[2 * i8];
00190             t1 = r1 - r3;
00191             r1 = r1 + r3;
00192             r3 = r2 - r4;
00193             r2 = r2 + r4;
00194             pSrc[2 * i1] = r1 + r2;
00195             r2 = r1 - r2;
00196             s1 = pSrc[2 * i1 + 1] + pSrc[2 * i5 + 1];
00197             s5 = pSrc[2 * i1 + 1] - pSrc[2 * i5 + 1];
00198             s2 = pSrc[2 * i2 + 1] + pSrc[2 * i6 + 1];
00199             s6 = pSrc[2 * i2 + 1] - pSrc[2 * i6 + 1];
00200             s3 = pSrc[2 * i3 + 1] + pSrc[2 * i7 + 1];
00201             s7 = pSrc[2 * i3 + 1] - pSrc[2 * i7 + 1];
00202             s4 = pSrc[2 * i4 + 1] + pSrc[2 * i8 + 1];
00203             s8 = pSrc[2 * i4 + 1] - pSrc[2 * i8 + 1];
00204             t2 = s1 - s3;
00205             s1 = s1 + s3;
00206             s3 = s2 - s4;
00207             s2 = s2 + s4;
00208             r1 = t1 + s3;
00209             t1 = t1 - s3;
00210             pSrc[2 * i1 + 1] = s1 + s2;
00211             s2 = s1 - s2;
00212             s1 = t2 - r3;
00213             t2 = t2 + r3;
00214             p1 = co5 * r2;
00215             p2 = si5 * s2;
00216             p3 = co5 * s2;
00217             p4 = si5 * r2;
00218             pSrc[2 * i5]     = p1 + p2;
00219             pSrc[2 * i5 + 1] = p3 - p4;
00220             p1 = co3 * r1;
00221             p2 = si3 * s1;
00222             p3 = co3 * s1;
00223             p4 = si3 * r1;
00224             pSrc[2 * i3]     = p1 + p2;
00225             pSrc[2 * i3 + 1] = p3 - p4;
00226             p1 = co7 * t1;
00227             p2 = si7 * t2;
00228             p3 = co7 * t2;
00229             p4 = si7 * t1;
00230             pSrc[2 * i7]     = p1 + p2;
00231             pSrc[2 * i7 + 1] = p3 - p4;
00232             r1 = (r6 - r8) * C81;
00233             r6 = (r6 + r8) * C81;
00234             s1 = (s6 - s8) * C81;
00235             s6 = (s6 + s8) * C81;
00236             t1 = r5 - r1;
00237             r5 = r5 + r1;
00238             r8 = r7 - r6;
00239             r7 = r7 + r6;
00240             t2 = s5 - s1;
00241             s5 = s5 + s1;
00242             s8 = s7 - s6;
00243             s7 = s7 + s6;
00244             r1 = r5 + s7;
00245             r5 = r5 - s7;
00246             r6 = t1 + s8;
00247             t1 = t1 - s8;
00248             s1 = s5 - r7;
00249             s5 = s5 + r7;
00250             s6 = t2 - r8;
00251             t2 = t2 + r8;
00252             p1 = co2 * r1;
00253             p2 = si2 * s1;
00254             p3 = co2 * s1;
00255             p4 = si2 * r1;
00256             pSrc[2 * i2]     = p1 + p2;
00257             pSrc[2 * i2 + 1] = p3 - p4;
00258             p1 = co8 * r5;
00259             p2 = si8 * s5;
00260             p3 = co8 * s5;
00261             p4 = si8 * r5;
00262             pSrc[2 * i8]     = p1 + p2;
00263             pSrc[2 * i8 + 1] = p3 - p4;
00264             p1 = co6 * r6;
00265             p2 = si6 * s6;
00266             p3 = co6 * s6;
00267             p4 = si6 * r6;
00268             pSrc[2 * i6]     = p1 + p2;
00269             pSrc[2 * i6 + 1] = p3 - p4;
00270             p1 = co4 * t1;
00271             p2 = si4 * t2;
00272             p3 = co4 * t2;
00273             p4 = si4 * t1;
00274             pSrc[2 * i4]     = p1 + p2;
00275             pSrc[2 * i4 + 1] = p3 - p4;
00276 
00277             i1 += n1;
00278          } while (i1 < fftLen);
00279 
00280          j++;
00281       } while (j < n2);
00282 
00283       twidCoefModifier <<= 3;
00284    } while (n2 > 7);
00285 }
00286