arm_cmplx_mult_real_q15.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_cmplx_mult_real_q15.c 00009 * 00010 * Description: Q15 complex by real multiplication 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_cmplx_mult_real_q15( 00055 q15_t * pSrcCmplx, 00056 q15_t * pSrcReal, 00057 q15_t * pCmplxDst, 00058 uint32_t numSamples) 00059 { 00060 q15_t in; /* Temporary variable to store input value */ 00061 uint32_t blkCnt; /* loop counters */ 00062 q31_t inA1, inA2; /* Temporary variables to hold input data */ 00063 q31_t inB1; /* Temporary variables to hold input data */ 00064 q15_t out1, out2, out3, out4; /* Temporary variables to hold output data */ 00065 q31_t mul1, mul2, mul3, mul4; /* Temporary variables to hold intermediate data */ 00066 00067 /* loop Unrolling */ 00068 blkCnt = numSamples >> 2u; 00069 00070 /* First part of the processing with loop unrolling. Compute 4 outputs at a time. 00071 ** a second loop below computes the remaining 1 to 3 samples. */ 00072 while(blkCnt > 0u) 00073 { 00074 /* C[2 * i] = A[2 * i] * B[i]. */ 00075 /* C[2 * i + 1] = A[2 * i + 1] * B[i]. */ 00076 /* read complex number both real and imaginary from complex input buffer */ 00077 inA1 = _SIMD32_OFFSET(pSrcCmplx); 00078 /* read two real values at a time from real input buffer */ 00079 inB1 = _SIMD32_OFFSET(pSrcReal); 00080 /* read complex number both real and imaginary from complex input buffer */ 00081 inA2 = _SIMD32_OFFSET(pSrcCmplx + 2); 00082 00083 /* multiply complex number with real numbers */ 00084 #ifndef ARM_MATH_BIG_ENDIAN 00085 00086 mul1 = (q31_t)((q15_t)(inA1)* (q15_t)(inB1)); 00087 mul2 = (q31_t)((q15_t)(inA1 >> 16) * (q15_t)(inB1)); 00088 mul3 = (q31_t)((q15_t)(inA2)* (q15_t)(inB1 >> 16)); 00089 mul4 = (q31_t)((q15_t)(inA2 >> 16) * (q15_t)(inB1 >> 16)); 00090 00091 #else 00092 00093 mul2 = (q31_t)((q15_t)(inA1 >> 16)* (q15_t)(inB1>> 16)); 00094 mul1 = (q31_t)((q15_t)inA1 * (q15_t)(inB1>> 16)); 00095 mul4 = (q31_t)((q15_t)(inA2 >> 16)* (q15_t)inB1); 00096 mul3 = (q31_t)((q15_t)inA2 * (q15_t)inB1); 00097 00098 #endif // #ifndef ARM_MATH_BIG_ENDIAN 00099 00100 /* saturate the result */ 00101 #ifdef CCS 00102 00103 out1 = (q15_t) __SSATA(mul1, 15u, 16); 00104 out2 = (q15_t) __SSATA(mul2, 15u, 16); 00105 out3 = (q15_t) __SSATA(mul3, 15u, 16); 00106 out4 = (q15_t) __SSATA(mul4, 15u, 16); 00107 00108 #else 00109 00110 out1 = (q15_t) __SSAT(mul1 >> 15u, 16); 00111 out2 = (q15_t) __SSAT(mul2 >> 15u, 16); 00112 out3 = (q15_t) __SSAT(mul3 >> 15u, 16); 00113 out4 = (q15_t) __SSAT(mul4 >> 15u, 16); 00114 00115 #endif // #ifdef CCS 00116 00117 /* pack real and imaginary outputs and store them to destination */ 00118 00119 _SIMD32_OFFSET(pCmplxDst) = __PKHBT(out1, out2, 16); 00120 _SIMD32_OFFSET(pCmplxDst + 2) = __PKHBT(out3, out4, 16); 00121 00122 inA1 = _SIMD32_OFFSET(pSrcCmplx + 4); 00123 inB1 = _SIMD32_OFFSET(pSrcReal + 2); 00124 inA2 = _SIMD32_OFFSET(pSrcCmplx + 6); 00125 00126 #ifndef ARM_MATH_BIG_ENDIAN 00127 00128 mul1 = (q31_t)((q15_t)(inA1)* (q15_t)(inB1)); 00129 mul2 = (q31_t)((q15_t)(inA1 >> 16) * (q15_t)(inB1)); 00130 mul3 = (q31_t)((q15_t)(inA2) * (q15_t)(inB1 >> 16)); 00131 mul4 = (q31_t)((q15_t)(inA2 >> 16) * (q15_t)(inB1 >> 16)); 00132 00133 #else 00134 00135 mul2 = (q31_t)((q15_t)(inA1 >> 16)* (q15_t)(inB1>> 16)); 00136 mul1 = (q31_t)((q15_t)inA1 * (q15_t)(inB1>> 16)); 00137 mul4 = (q31_t)((q15_t)(inA2 >> 16)* (q15_t)inB1); 00138 mul3 = (q31_t)((q15_t)inA2 * (q15_t)inB1); 00139 00140 #endif // #ifndef ARM_MATH_BIG_ENDIAN 00141 00142 #ifdef CCS 00143 00144 out1 = (q15_t) __SSATA(mul1, 15u, 16); 00145 out2 = (q15_t) __SSATA(mul2, 15u, 16); 00146 out3 = (q15_t) __SSATA(mul3, 15u, 16); 00147 out4 = (q15_t) __SSATA(mul4, 15u, 16); 00148 00149 #else 00150 00151 out1 = (q15_t) __SSAT(mul1 >> 15u, 16); 00152 out2 = (q15_t) __SSAT(mul2 >> 15u, 16); 00153 out3 = (q15_t) __SSAT(mul3 >> 15u, 16); 00154 out4 = (q15_t) __SSAT(mul4 >> 15u, 16); 00155 00156 #endif // #ifdef CCS 00157 00158 00159 _SIMD32_OFFSET(pCmplxDst + 4) = __PKHBT(out1, out2, 16); 00160 _SIMD32_OFFSET(pCmplxDst + 6) = __PKHBT(out3, out4, 16); 00161 00162 /* update pointer to process next samples */ 00163 pSrcCmplx += 8u; 00164 pSrcReal += 4u; 00165 pCmplxDst += 8u; 00166 00167 /* Decrement the numSamples loop counter */ 00168 blkCnt--; 00169 } 00170 00171 /* If the numSamples is not a multiple of 4, compute any remaining output samples here. 00172 ** No loop unrolling is used. */ 00173 blkCnt = numSamples % 0x4u; 00174 00175 while(blkCnt > 0u) 00176 { 00177 /* C[2 * i] = A[2 * i] * B[i]. */ 00178 /* C[2 * i + 1] = A[2 * i + 1] * B[i]. */ 00179 in = *pSrcReal++; 00180 /* store the result in the destination buffer. */ 00181 #ifdef CCS 00182 *pCmplxDst++ = (q15_t) __SSATA(((q31_t) (*pSrcCmplx++) * (in)), 15, 16); 00183 *pCmplxDst++ = (q15_t) __SSATA(((q31_t) (*pSrcCmplx++) * (in)), 15, 16); 00184 #else 00185 *pCmplxDst++ = (q15_t) __SSAT((((q31_t) (*pSrcCmplx++) * (in)) >> 15), 16); 00186 *pCmplxDst++ = (q15_t) __SSAT((((q31_t) (*pSrcCmplx++) * (in)) >> 15), 16); 00187 #endif 00188 00189 /* Decrement the numSamples loop counter */ 00190 blkCnt--; 00191 } 00192 } 00193
