arm_shift_q31.c
Go to the documentation of this file.
00001 /* ---------------------------------------------------------------------- 00002 * Copyright (C) 2011 ARM Limited. All rights reserved. 00003 * 00004 * $Date: 15. December 2011 00005 * $Revision: V2.0.0 00006 * 00007 * Project: Cortex-R DSP Library 00008 * Title: arm_shift_q31.c 00009 * 00010 * Description: Shifts the elements of a Q31 vector by a specified number of bits. 00011 * 00012 * Target Processor: Cortex-R4/R5 00013 * 00014 * Version 1.0.0 2011/03/08 00015 * Alpha release. 00016 * 00017 * Version 1.0.1 2011/09/30 00018 * Beta release. 00019 * 00020 * Version 2.0.0 2011/12/15 00021 * Final release. 00022 * 00023 * -------------------------------------------------------------------- */ 00024 #include "arm_math.h" 00025 00064 void arm_shift_q31( 00065 q31_t * pSrc, 00066 int8_t shiftBits, 00067 q31_t * pDst, 00068 uint32_t blockSize) 00069 { 00070 uint32_t blkCnt; /* loop counter */ 00071 uint8_t sign; /* Sign of shiftBits */ 00072 q31_t in1, in2, in3, in4; /* Temporary input variables */ 00073 q31_t out1, out2, out3, out4; /* Temporary output variables */ 00074 00075 /*loop Unrolling */ 00076 blkCnt = blockSize >> 3u; 00077 00078 /* Getting the sign of shiftBits */ 00079 sign = (shiftBits & 0x80); 00080 00081 if(sign == 0u) 00082 { 00083 /* First part of the processing with loop unrolling. Compute 8 outputs at a time. 00084 ** a second loop below computes the remaining 1 to 7 samples. */ 00085 while(blkCnt > 0u) 00086 { 00087 /* C = A << shiftBits */ 00088 /* Shift the input and then store the results in the destination buffer. */ 00089 in1 = *pSrc; 00090 in2 = *(pSrc + 1); 00091 out1 = in1 << shiftBits; 00092 in3 = *(pSrc + 2); 00093 out2 = in2 << shiftBits; 00094 in4 = *(pSrc + 3); 00095 if(in1 != (out1 >> shiftBits)) 00096 out1 = 0x7FFFFFFF ^(in1 >> 31); 00097 00098 if(in2 != (out2 >> shiftBits)) 00099 out2 = 0x7FFFFFFF ^(in2 >> 31); 00100 00101 *pDst = out1; 00102 out3 = in3 << shiftBits; 00103 *(pDst + 1) = out2; 00104 out4 = in4 << shiftBits; 00105 00106 if(in3 != (out3 >> shiftBits)) 00107 out3 = 0x7FFFFFFF ^(in3 >> 31); 00108 00109 if(in4 != (out4 >> shiftBits)) 00110 out4 = 0x7FFFFFFF ^(in4 >> 31); 00111 00112 *(pDst + 2) = out3; 00113 *(pDst + 3) = out4; 00114 00115 in1 = *(pSrc + 4); 00116 in2 = *(pSrc + 5); 00117 out1 = in1 << shiftBits; 00118 in3 = *(pSrc + 6); 00119 out2 = in2 << shiftBits; 00120 in4 = *(pSrc + 7); 00121 if(in1 != (out1 >> shiftBits)) 00122 out1 = 0x7FFFFFFF ^(in1 >> 31); 00123 00124 if(in2 != (out2 >> shiftBits)) 00125 out2 = 0x7FFFFFFF ^(in2 >> 31); 00126 00127 *(pDst + 4) = out1; 00128 out3 = in3 << shiftBits; 00129 *(pDst + 5) = out2; 00130 out4 = in4 << shiftBits; 00131 00132 if(in3 != (out3 >> shiftBits)) 00133 out3 = 0x7FFFFFFF ^(in3 >> 31); 00134 00135 if(in4 != (out4 >> shiftBits)) 00136 out4 = 0x7FFFFFFF ^(in4 >> 31); 00137 00138 *(pDst + 6) = out3; 00139 /* Update source pointer to process next sampels */ 00140 pSrc += 8u; 00141 *(pDst + 7) = out4; 00142 00143 /* Update destination pointer to process next sampels */ 00144 pDst += 8u; 00145 00146 /* Decrement the loop counter */ 00147 blkCnt--; 00148 } 00149 00150 /* If the blockSize is not a multiple of 8, compute any remaining output samples here. 00151 ** No loop unrolling is used. */ 00152 blkCnt = blockSize % 0x8u; 00153 00154 while(blkCnt > 0u) 00155 { 00156 /* C = A (>> or <<) shiftBits */ 00157 /* Shift the input and then store the result in the destination buffer. */ 00158 in1 = *pSrc++; 00159 out1 = in1 << shiftBits; 00160 if(in1 != (out1 >> shiftBits)) 00161 out1 = 0x7FFFFFFF ^(in1 >> 31); 00162 00163 *pDst++ = out1; 00164 00165 /* Decrement the loop counter */ 00166 blkCnt--; 00167 } 00168 } 00169 else 00170 { 00171 shiftBits = -shiftBits; 00172 00173 /* First part of the processing with loop unrolling. Compute 8 outputs at a time. 00174 ** a second loop below computes the remaining 1 to 7 samples. */ 00175 while(blkCnt > 0u) 00176 { 00177 /* C = A >> shiftBits */ 00178 /* Shift the input and then store the results in the destination buffer. */ 00179 in1 = *pSrc; 00180 in2 = *(pSrc + 1); 00181 in3 = *(pSrc + 2); 00182 in4 = *(pSrc + 3); 00183 00184 *pDst = (in1 >> shiftBits); 00185 *(pDst + 1) = (in2 >> shiftBits); 00186 *(pDst + 2) = (in3 >> shiftBits); 00187 *(pDst + 3) = (in4 >> shiftBits); 00188 00189 in1 = *(pSrc + 4); 00190 in2 = *(pSrc + 5); 00191 in3 = *(pSrc + 6); 00192 in4 = *(pSrc + 7); 00193 00194 *(pDst + 4) = (in1 >> shiftBits); 00195 *(pDst + 5) = (in2 >> shiftBits); 00196 *(pDst + 6) = (in3 >> shiftBits); 00197 *(pDst + 7) = (in4 >> shiftBits); 00198 00199 pSrc += 8u; 00200 pDst += 8u; 00201 00202 blkCnt--; 00203 } 00204 00205 /* If the blockSize is not a multiple of 8, compute any remaining output samples here. 00206 ** No loop unrolling is used. */ 00207 blkCnt = blockSize % 0x8u; 00208 00209 while(blkCnt > 0u) 00210 { 00211 /* C = A (>> or <<) shiftBits */ 00212 /* Shift the input and then store the result in the destination buffer. */ 00213 in1 = *pSrc++; 00214 *pDst++ = (in1 >> shiftBits); 00215 00216 /* Decrement the loop counter */ 00217 blkCnt--; 00218 } 00219 } 00220 00221 } 00222
