// Name: ippsFIRSRGetSize, ippsFIRSRInit_32f, ippsFIRSRInit_64f // ippsFIRSR_32f, ippsFIRSR_64f // Purpose: Get sizes of the FIR spec structure and temporary buffer // initialize FIR spec structure - set taps and delay line // perform FIR filtering // Parameters: // pTaps - pointer to the filter coefficients // tapsLen - number of coefficients // tapsType - type of coefficients (ipp32f or ipp64f) // pSpecSize - pointer to the size of FIR spec // pBufSize - pointer to the size of temporal buffer // algType - mask for the algorithm type definition (direct, fft, auto) // pDlySrc - pointer to the input delay line values, can be NULL // pDlyDst - pointer to the output delay line values, can be NULL // pSpec - pointer to the constant internal structure // pSrc - pointer to the source vector. // pDst - pointer to the destination vector // numIters - length of the destination vector // pBuf - pointer to the work buffer // Return: // status - status value returned, its value are // ippStsNullPtrErr - one of the specified pointer is NULL // ippStsFIRLenErr - tapsLen <= 0 // ippStsContextMatchErr - wrong state identifier // ippStsNoErr - OK // ippStsSizeErr - numIters is not positive // ippStsAlgTypeErr - unsupported algorithm type // ippStsMismatch - not effective algorithm. */ IppStatus ippsFIRSRGetSize (int tapsLen, IppDataType tapsType , int* pSpecSize, int* pBufSize ) IppStatus ippsFIRSRInit_32f( const Ipp32f* pTaps, int tapsLen, IppAlgType algType, IppsFIRSpec_32f* pSpec ) IppStatus ippsFIRSR_32f (const Ipp32f* pSrc, Ipp32f* pDst, int numIters, IppsFIRSpec_32f* pSpec, const Ipp32f* pDlySrc, Ipp32f* pDlyDst, Ipp8u* pBuf) 

 void fir_omp(Ipp32f* src, Ipp32f* dst, int len, int order, IppsFIRSpec_32f* pSpec, Ipp32f* pDlySrc, Ipp32f* pDlyDst, Ipp8u* pBuffer) { int tlen, ttail; tlen = len / NTHREADS; ttail = len % NTHREADS; #pragma omp parallel num_threads(NTHREADS) { int id = omp_get_thread_num(); Ipp32f* s = src + id*tlen; Ipp32f* d = dst + id*tlen; int len = tlen + ((id == NTHREADS-1) ? ttail : 0); Ipp8u* b = pBuffer + id*bufSize; if (id == 0) ippsFIRSR_32f(s, d, len, pSpec, pDlySrc, NULL, b); else if (id == NTHREADS - 1) ippsFIRSR_32f(s, d, len, pSpec, s - (order - 1), pDlyDst, b); else ippsFIRSR_32f(s, d, len, pSpec, s - (order - 1), NULL, b); } } 

 #include <stdio.h> #include <math.h> #include <omp.h> #include "ippcore.h" #include "ipps.h" #include "bmp.h" void save_csv(Ipp32f* pSrc, int len, char* fName) { FILE *fp; int i; if((fp=fopen(fName, "w"))==NULL) { printf("Cannot open %s\n", fName); return; } for (i = 0; i < len; i++){ fprintf(fp, "%.3f\n", pSrc[i]); } fclose(fp); } Ipp32f* pSrc; Ipp32f* pDft; Ipp32f* pDst; Ipp32f* pTaps; Ipp64f rFreq = 0.2; int bufSize; int NTHREADS = 1; IppAlgType algType = ippAlgDirect; void firgenlowpass(int order) { IppStatus status; Ipp8u* pBuffer; Ipp64f* pTaps_64f; int size; int i; status = ippsFIRGenGetBufferSize(order, &size); pBuffer = ippsMalloc_8u(size); pTaps_64f = ippsMalloc_64f(order); ippsFIRGenLowpass_64f(rFreq, pTaps_64f, order, ippWinBartlett, ippTrue, pBuffer); for (i = 0; i < order;i++) { pTaps[i] = pTaps_64f[i]; } ippsFree(pTaps_64f); } void fir_omp(Ipp32f* src, Ipp32f* dst, int len, int order, IppsFIRSpec_32f* pSpec, Ipp32f* pDlySrc, Ipp32f* pDlyDst, Ipp8u* pBuffer) { int tlen, ttail; tlen = len / NTHREADS; ttail = len % NTHREADS; #pragma omp parallel num_threads(NTHREADS) { int id = omp_get_thread_num(); Ipp32f* s = src + id*tlen; Ipp32f* d = dst + id*tlen; int len = tlen + ((id == NTHREADS-1) ? ttail : 0); Ipp8u* b = pBuffer + id*bufSize; if (id == 0) ippsFIRSR_32f(s, d, len, pSpec, pDlySrc, NULL, b); else if (id == NTHREADS - 1) ippsFIRSR_32f(s, d, len, pSpec, s - (order - 1), pDlyDst, b); else ippsFIRSR_32f(s, d, len, pSpec, s - (order - 1), NULL, b); } } void perf(int len, int order, float* cpMAC) { IppStatus status; IppsFIRSpec_32f* pSpec; Ipp8u* pBuffer; int specSize; Ipp32f* pDlySrc = NULL;/*initialize delay line with "0"*/ Ipp32f* pDlyDst = NULL;/*don't write output delay line*/ __int64 beg=0, end=0; int i, loop = 10000; /*allocate memory for input and output vectors*/ pSrc = ippsMalloc_32f(len); pDst = ippsMalloc_32f(len); pTaps = ippsMalloc_32f(order); /*create special vector Jaehne*/ ippsVectorJaehne_32f(pSrc, len, 128); /*get lowpass filter coeffs*/ firgenlowpass(order); /*get necessary buffer sizes for pSpec and for pBuffer*/ status = ippsFIRSRGetSize(order, ipp32f, &specSize, &bufSize); /*allocate memory for pSpec*/ pSpec = (IppsFIRSpec_32f*)ippsMalloc_8u(specSize); /*for N threads bufSize should be multiplied by N*/ /*allocate bufSize*NTHREADS bytes*/ pBuffer = ippsMalloc_8u(bufSize*NTHREADS); /*initalize pSpec*/ status = ippsFIRSRInit_32f(pTaps, order, algType, pSpec); /*apply FIR filter*/ /*start measurement for sinle threaded*/ if (NTHREADS == 1){ ippsFIRSR_32f(pSrc, pDst, len, pSpec, pDlySrc, pDlyDst, pBuffer); beg = __rdtsc(); for (int i = 0; i < loop; i++) { ippsFIRSR_32f(pSrc, pDst, len, pSpec, pDlySrc, pDlyDst, pBuffer); } end = __rdtsc(); } else { fir_omp(pSrc, pDst, len, order, pSpec, pDlySrc, pDlyDst, pBuffer); beg = __rdtsc(); for (int i = 0; i < loop; i++) { fir_omp(pSrc, pDst, len, order, pSpec, pDlySrc, pDlyDst, pBuffer); } end = __rdtsc(); } *cpMAC = ((double)(end - beg) / ((double)loop * (double)len * (double)order)); printf("%5d, %5d, %3.3f\n", len, order, *cpMAC); ippsFree(pSrc); ippsFree(pDst); ippsFree(pTaps); ippsFree(pSpec); ippsFree(pBuffer); } int main() { int len = 32768; int order; float cpMAC; NTHREADS = 1; algType = ippAlgDirect; //algType = ippAlgFFT; len = 128; printf("\nthreads: %d\n", NTHREADS); printf("len, order, cpMAC\n\n"); for (order = 1; order <= 512; order++){ perf(len, order, &cpMAC); } return 0; }