arm_scale_q7.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_scale_q7.c 00009 * 00010 * Description: Multiplies a Q7 vector by a scalar. 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 00054 void arm_scale_q7( 00055 q7_t * pSrc, 00056 q7_t scaleFract, 00057 int8_t shift, 00058 q7_t * pDst, 00059 uint32_t blockSize) 00060 { 00061 int8_t kShift = 7 - shift; /* shift to apply after scaling */ 00062 uint32_t blkCnt; /* loop counter */ 00063 q7_t in1, in2, in3, in4, in5, in6; /* temporary input variables */ 00064 q15_t out1, out2, out3, out4; /* Temporary variables to store input & output */ 00065 00066 /*loop Unrolling */ 00067 blkCnt = blockSize >> 3u; 00068 00069 /* First part of the processing with loop unrolling. Compute 8 outputs at a time. 00070 ** a second loop below computes the remaining 1 to 7 samples. */ 00071 while(blkCnt > 0u) 00072 { 00073 /* Reading 8 inputs from memory */ 00074 /* read inputs from source */ 00075 in1 = pSrc[0]; 00076 in2 = pSrc[1]; 00077 00078 /* multiply with scaling factor */ 00079 out1 = (q15_t)((q7_t) (in1) * (scaleFract)); 00080 00081 /* read input from source */ 00082 in3 = pSrc[2]; 00083 00084 /* multiply with scaling factor */ 00085 out2 = (q15_t)((q7_t) (in2) * (scaleFract)); 00086 00087 /* read input from source */ 00088 in4 = pSrc[3]; 00089 00090 /* multiply with scaling factor */ 00091 out3 = (q15_t)((q7_t) (in3) * (scaleFract)); 00092 out4 = (q15_t)((q7_t) (in4) * (scaleFract)); 00093 00094 /* apply shifting */ 00095 out1 = out1 >> kShift; 00096 out2 = out2 >> kShift; 00097 00098 /* read input from source */ 00099 in1 = pSrc[4]; 00100 00101 /* apply shifting */ 00102 out3 = out3 >> kShift; 00103 out4 = out4 >> kShift; 00104 00105 /* saturate the output */ 00106 #ifdef CCS 00107 00108 in3 = (q7_t) __SSATA(out1, 0, 8); 00109 in4 = (q7_t) __SSATA(out2, 0, 8); 00110 in5 = (q7_t) __SSATA(out3, 0, 8); 00111 in6 = (q7_t) __SSATA(out4, 0, 8); 00112 00113 #else 00114 00115 in3 = (q7_t) __SSAT(out1, 8); 00116 in4 = (q7_t) __SSAT(out2, 8); 00117 in5 = (q7_t) __SSAT(out3, 8); 00118 in6 = (q7_t) __SSAT(out4, 8); 00119 00120 #endif 00121 00122 /* Store the results of 4 inputs in the destination buffer in single cycle by packing */ 00123 *__SIMD32(pDst)++ = __PACKq7(in3, in4, in5, in6); 00124 00125 /* read input from source */ 00126 in2 = pSrc[5]; 00127 00128 /* multiply with scaling factor */ 00129 out1 = (q15_t)((q7_t) (in1) * (scaleFract)); 00130 00131 /* read input from source */ 00132 in3 = pSrc[6]; 00133 00134 /* multiply with scaling factor */ 00135 out2 = (q15_t)((q7_t) (in2) * (scaleFract)); 00136 00137 /* read input from source */ 00138 in4 = pSrc[7]; 00139 00140 /* multiply with scaling factor */ 00141 out3 = (q15_t)((q7_t) (in3) * (scaleFract)); 00142 out4 = (q15_t)((q7_t) (in4) * (scaleFract)); 00143 00144 /* apply shifting */ 00145 out1 = out1 >> kShift; 00146 out2 = out2 >> kShift; 00147 out3 = out3 >> kShift; 00148 out4 = out4 >> kShift; 00149 00150 /* saturate the output */ 00151 #ifdef CCS 00152 00153 in1 = (q7_t) __SSATA(out1, 0, 8); 00154 in2 = (q7_t) __SSATA(out2, 0, 8); 00155 in3 = (q7_t) __SSATA(out3, 0, 8); 00156 in4 = (q7_t) __SSATA(out4, 0, 8); 00157 00158 #else 00159 00160 in1 = (q7_t) __SSAT(out1, 8); 00161 in2 = (q7_t) __SSAT(out2, 8); 00162 in3 = (q7_t) __SSAT(out3, 8); 00163 in4 = (q7_t) __SSAT(out4, 8); 00164 00165 #endif 00166 00167 /* Store the results of 4 inputs in the destination buffer in single cycle by packing */ 00168 *__SIMD32(pDst)++ = __PACKq7(in1, in2, in3, in4); 00169 00170 /* Update source pointer to process next sampels */ 00171 pSrc += 8u; 00172 00173 /* Decrement the loop counter */ 00174 blkCnt--; 00175 } 00176 00177 /* If the blockSize is not a multiple of 8, compute any remaining output samples here. 00178 ** No loop unrolling is used. */ 00179 blkCnt = blockSize % 0x8u; 00180 00181 while(blkCnt > 0u) 00182 { 00183 /* C = A * scale */ 00184 /* Scale the input and then store the result in the destination buffer. */ 00185 #ifdef CCS 00186 *pDst++ = (q7_t) (__SSATA(((*pSrc++) * scaleFract) >> kShift, 0, 8)); 00187 #else 00188 *pDst++ = (q7_t) (__SSAT(((*pSrc++) * scaleFract) >> kShift, 8)); 00189 #endif 00190 /* Decrement the loop counter */ 00191 blkCnt--; 00192 } 00193 } 00194
