__m128 test_mm_cvtpd_ps(__m128d A) { // DAG-LABEL: test_mm_cvtpd_ps // DAG: call <4 x float> @llvm.x86.sse2.cvtpd2ps // // ASM-LABEL: test_mm_cvtpd_ps // ASM: cvtpd2ps return _mm_cvtpd_ps(A); }
// Conversions SIMDValue SIMDFloat32x4Operation::OpFromFloat64x2(const SIMDValue& value) { X86SIMDValue x86Result; X86SIMDValue v = X86SIMDValue::ToX86SIMDValue(value); // Converts the two double-precision, floating-point values of v.m128d_value // to single-precision, floating-point values. x86Result.m128_value = _mm_cvtpd_ps(v.m128d_value); return X86SIMDValue::ToSIMDValue(x86Result); }
static inline __m128d my_invrsq_pd(__m128d x) { const __m128d three = {3.0, 3.0}; const __m128d half = {0.5, 0.5}; __m128 t = _mm_rsqrt_ps(_mm_cvtpd_ps(x)); /* Convert to single precision and do _mm_rsqrt_ps() */ __m128d t1 = _mm_cvtps_pd(t); /* Convert back to double precision */ /* First Newton-Rapson step, accuracy is now 24 bits */ __m128d t2 = _mm_mul_pd(half,_mm_mul_pd(t1,_mm_sub_pd(three,_mm_mul_pd(x,_mm_mul_pd(t1,t1))))); /* Return second Newton-Rapson step, accuracy 48 bits */ return _mm_mul_pd(half,_mm_mul_pd(t2,_mm_sub_pd(three,_mm_mul_pd(x,_mm_mul_pd(t2,t2))))); }
// ============================================================================= // // sse3_vChirpData // version by: Alex Kan // http://tbp.berkeley.edu/~alexkan/seti/ // int sse3_ChirpData_ak( sah_complex * cx_DataArray, sah_complex * cx_ChirpDataArray, int chirp_rate_ind, double chirp_rate, int ul_NumDataPoints, double sample_rate ) { int i; if (chirp_rate_ind == 0) { memcpy(cx_ChirpDataArray, cx_DataArray, (int)ul_NumDataPoints * sizeof(sah_complex) ); return 0; } int vEnd; double srate = chirp_rate * 0.5 / (sample_rate * sample_rate); __m128d rate = _mm_set1_pd(chirp_rate * 0.5 / (sample_rate * sample_rate)); __m128d roundVal = _mm_set1_pd(srate >= 0.0 ? TWO_TO_52 : -TWO_TO_52); // main vectorised loop vEnd = ul_NumDataPoints - (ul_NumDataPoints & 3); for (i = 0; i < vEnd; i += 4) { const float *data = (const float *) (cx_DataArray + i); float *chirped = (float *) (cx_ChirpDataArray + i); __m128d di = _mm_set1_pd(i); __m128d a1 = _mm_add_pd(_mm_set_pd(1.0, 0.0), di); __m128d a2 = _mm_add_pd(_mm_set_pd(3.0, 2.0), di); __m128 d1, d2; __m128 cd1, cd2; __m128 td1, td2; __m128 x; __m128 y; __m128 s; __m128 c; __m128 m; // load the signal to be chirped prefetchnta((const void *)( data+32 )); d1 = _mm_load_ps(data); d2 = _mm_load_ps(data+4); // calculate the input angle a1 = _mm_mul_pd(_mm_mul_pd(a1, a1), rate); a2 = _mm_mul_pd(_mm_mul_pd(a2, a2), rate); // reduce the angle to the range (-0.5, 0.5) a1 = _mm_sub_pd(a1, _mm_sub_pd(_mm_add_pd(a1, roundVal), roundVal)); a2 = _mm_sub_pd(a2, _mm_sub_pd(_mm_add_pd(a2, roundVal), roundVal)); // convert pair of packed double into packed single x = _mm_movelh_ps(_mm_cvtpd_ps(a1), _mm_cvtpd_ps(a2)); // square to the range [0, 0.25) y = _mm_mul_ps(x, x); // perform the initial polynomial approximations s = _mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(y, SS4), SS3), y), SS2), y), SS1), x); c = _mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(y, CC3), CC2), y), CC1), y), ONE); // perform first angle doubling x = _mm_sub_ps(_mm_mul_ps(c, c), _mm_mul_ps(s, s)); y = _mm_mul_ps(_mm_mul_ps(s, c), TWO); // calculate scaling factor to correct the magnitude // m1 = vec_nmsub(y1, y1, vec_nmsub(x1, x1, TWO)); // m2 = vec_nmsub(y2, y2, vec_nmsub(x2, x2, TWO)); m = vec_recip3(_mm_add_ps(_mm_mul_ps(x, x), _mm_mul_ps(y, y))); // perform second angle doubling c = _mm_sub_ps(_mm_mul_ps(x, x), _mm_mul_ps(y, y)); s = _mm_mul_ps(_mm_mul_ps(y, x), TWO); // correct the magnitude (final sine / cosine approximations) s = _mm_mul_ps(s, m); c = _mm_mul_ps(c, m); // chirp the data cd1 = _mm_shuffle_ps(c, c, 0x50); cd2 = _mm_shuffle_ps(c, c, 0xfa); cd1 = _mm_mul_ps(cd1, d1); cd2 = _mm_mul_ps(cd2, d2); d1 = _mm_shuffle_ps(d1, d1, 0xb1); d2 = _mm_shuffle_ps(d2, d2, 0xb1); td1 = _mm_shuffle_ps(s, s, 0x50); td2 = _mm_shuffle_ps(s, s, 0xfa); td1 = _mm_mul_ps(td1, d1); td2 = _mm_mul_ps(td2, d2); cd1 = _mm_addsub_ps(cd1, td1); cd2 = _mm_addsub_ps(cd2, td2); // store chirped values _mm_stream_ps(chirped, cd1); _mm_stream_ps(chirped+4, cd2); } _mm_sfence(); // handle tail elements with scalar code for ( ; i < ul_NumDataPoints; ++i) { double angle = srate * i * i * 0.5; double s = sin(angle); double c = cos(angle); float re = cx_DataArray[i][0]; float im = cx_DataArray[i][1]; cx_ChirpDataArray[i][0] = re * c - im * s; cx_ChirpDataArray[i][1] = re * s + im * c; } analysis_state.FLOP_counter+=12.0*ul_NumDataPoints; return 0; }
static inline __m128d my_invrsq_pd(__m128d x) { const __m128d three = (const __m128d) {3.0f, 3.0f}; const __m128d half = (const __m128d) {0.5f, 0.5f}; __m128 t = _mm_rsqrt_ps(_mm_cvtpd_ps(x)); /* Convert to single precision and do _mm_rsqrt_ps() */ __m128d t1 = _mm_cvtps_pd(t); /* Convert back to double precision */ /* First Newton-Rapson step, accuracy is now 24 bits */ __m128d t2 = _mm_mul_pd(half,_mm_mul_pd(t1,_mm_sub_pd(three,_mm_mul_pd(x,_mm_mul_pd(t1,t1))))); /* Return second Newton-Rapson step, accuracy 48 bits */ return (__m128d) _mm_mul_pd(half,_mm_mul_pd(t2,_mm_sub_pd(three,_mm_mul_pd(x,_mm_mul_pd(t2,t2))))); } /* to extract single integers from a __m128i datatype */ #define _mm_extract_epi64(x, imm) \ _mm_cvtsi128_si32(_mm_srli_si128((x), 4 * (imm))) void nb_kernel400_x86_64_sse2(int * p_nri, int * iinr, int * jindex, int * jjnr, int * shift, double * shiftvec, double * fshift, int * gid, double * pos, double * faction, double * charge, double * p_facel, double * p_krf, double * p_crf, double * Vc, int * type, int * p_ntype, double * vdwparam, double * Vvdw, double * p_tabscale, double * VFtab, double * invsqrta, double * dvda, double * p_gbtabscale, double * GBtab, int * p_nthreads, int * count, void * mtx, int * outeriter, int * inneriter, double * work) { int nri,ntype,nthreads,offset; int n,ii,is3,ii3,k,nj0,nj1,jnr1,jnr2,j13,j23,ggid; double facel,krf,crf,tabscl,gbtabscl,vct,vgbt; double shX,shY,shZ,isai_d,dva; gmx_gbdata_t *gbdata; float * gpol; __m128d ix,iy,iz,jx,jy,jz; __m128d dx,dy,dz,t1,t2,t3; __m128d fix,fiy,fiz,rsq11,rinv,r,fscal,rt,eps,eps2; __m128d q,iq,qq,isai,isaj,isaprod,vcoul,gbscale,dvdai,dvdaj; __m128d Y,F,G,H,Fp,VV,FF,vgb,fijC,dvdatmp,dvdasum,vctot,vgbtot,n0d; __m128d xmm0,xmm1,xmm2,xmm3,xmm4,xmm5,xmm6,xmm7,xmm8; __m128d fac,tabscale,gbtabscale; __m128i n0,nnn; const __m128d neg = {-1.0f,-1.0f}; const __m128d zero = {0.0f,0.0f}; const __m128d half = {0.5f,0.5f}; const __m128d two = {2.0f,2.0f}; const __m128d three = {3.0f,3.0f}; gbdata = (gmx_gbdata_t *)work; gpol = gbdata->gpol; nri = *p_nri; ntype = *p_ntype; nthreads = *p_nthreads; facel = (*p_facel) * (1.0 - (1.0/gbdata->gb_epsilon_solvent)); krf = *p_krf; crf = *p_crf; tabscl = *p_tabscale; gbtabscl = *p_gbtabscale; nj1 = 0; /* Splat variables */ fac = _mm_load1_pd(&facel); tabscale = _mm_load1_pd(&tabscl); gbtabscale = _mm_load1_pd(&gbtabscl); /* Keep compiler happy */ dvdatmp = _mm_setzero_pd(); vgb = _mm_setzero_pd(); dvdaj = _mm_setzero_pd(); isaj = _mm_setzero_pd(); vcoul = _mm_setzero_pd(); t1 = _mm_setzero_pd(); t2 = _mm_setzero_pd(); t3 = _mm_setzero_pd(); jnr1=jnr2=0; j13=j23=0; for(n=0;n<nri;n++) { is3 = 3*shift[n]; shX = shiftvec[is3]; shY = shiftvec[is3+1]; shZ = shiftvec[is3+2]; nj0 = jindex[n]; nj1 = jindex[n+1]; offset = (nj1-nj0)%2; ii = iinr[n]; ii3 = ii*3; ix = _mm_set1_pd(shX+pos[ii3+0]); iy = _mm_set1_pd(shX+pos[ii3+1]); iz = _mm_set1_pd(shX+pos[ii3+2]); q = _mm_set1_pd(charge[ii]); iq = _mm_mul_pd(fac,q); isai_d = invsqrta[ii]; isai = _mm_load1_pd(&isai_d); fix = _mm_setzero_pd(); fiy = _mm_setzero_pd(); fiz = _mm_setzero_pd(); dvdasum = _mm_setzero_pd(); vctot = _mm_setzero_pd(); vgbtot = _mm_setzero_pd(); for(k=nj0;k<nj1-offset; k+=2) { jnr1 = jjnr[k]; jnr2 = jjnr[k+1]; j13 = jnr1 * 3; j23 = jnr2 * 3; /* Load coordinates */ xmm1 = _mm_loadu_pd(pos+j13); /* x1 y1 */ xmm2 = _mm_loadu_pd(pos+j23); /* x2 y2 */ xmm5 = _mm_load_sd(pos+j13+2); /* z1 - */ xmm6 = _mm_load_sd(pos+j23+2); /* z2 - */ /* transpose */ jx = _mm_shuffle_pd(xmm1,xmm2,_MM_SHUFFLE2(0,0)); jy = _mm_shuffle_pd(xmm1,xmm2,_MM_SHUFFLE2(1,1)); jz = _mm_shuffle_pd(xmm5,xmm6,_MM_SHUFFLE2(0,0)); /* distances */ dx = _mm_sub_pd(ix,jx); dy = _mm_sub_pd(iy,jy); dz = _mm_sub_pd(iz,jz); rsq11 = _mm_add_pd( _mm_add_pd( _mm_mul_pd(dx,dx) , _mm_mul_pd(dy,dy) ) , _mm_mul_pd(dz,dz) ); rinv = my_invrsq_pd(rsq11); /* Load invsqrta */ isaj = _mm_loadl_pd(isaj,invsqrta+jnr1); isaj = _mm_loadh_pd(isaj,invsqrta+jnr2); isaprod = _mm_mul_pd(isai,isaj); /* Load charges */ q = _mm_loadl_pd(q,charge+jnr1); q = _mm_loadh_pd(q,charge+jnr2); qq = _mm_mul_pd(iq,q); vcoul = _mm_mul_pd(qq,rinv); fscal = _mm_mul_pd(vcoul,rinv); qq = _mm_mul_pd(isaprod,qq); qq = _mm_mul_pd(qq,neg); gbscale = _mm_mul_pd(isaprod,gbtabscale); /* Load dvdaj */ dvdaj = _mm_loadl_pd(dvdaj, dvda+jnr1); dvdaj = _mm_loadh_pd(dvdaj, dvda+jnr2); r = _mm_mul_pd(rsq11,rinv); rt = _mm_mul_pd(r,gbscale); n0 = _mm_cvttpd_epi32(rt); n0d = _mm_cvtepi32_pd(n0); eps = _mm_sub_pd(rt,n0d); eps2 = _mm_mul_pd(eps,eps); nnn = _mm_slli_epi64(n0,2); xmm1 = _mm_load_pd(GBtab+(_mm_extract_epi64(nnn,0))); /* Y1 F1 */ xmm2 = _mm_load_pd(GBtab+(_mm_extract_epi64(nnn,1))); /* Y2 F2 */ xmm3 = _mm_load_pd(GBtab+(_mm_extract_epi64(nnn,0))+2); /* G1 H1 */ xmm4 = _mm_load_pd(GBtab+(_mm_extract_epi64(nnn,1))+2); /* G2 H2 */ Y = _mm_shuffle_pd(xmm1,xmm2,_MM_SHUFFLE2(0,0)); /* Y1 Y2 */ F = _mm_shuffle_pd(xmm1,xmm2,_MM_SHUFFLE2(1,1)); /* F1 F2 */ G = _mm_shuffle_pd(xmm3,xmm4,_MM_SHUFFLE2(0,0)); /* G1 G2 */ H = _mm_shuffle_pd(xmm3,xmm4,_MM_SHUFFLE2(1,1)); /* H1 H2 */ G = _mm_mul_pd(G,eps); H = _mm_mul_pd(H,eps2); Fp = _mm_add_pd(F,G); Fp = _mm_add_pd(Fp,H); VV = _mm_mul_pd(Fp,eps); VV = _mm_add_pd(Y,VV); H = _mm_mul_pd(two,H); FF = _mm_add_pd(Fp,G); FF = _mm_add_pd(FF,H); vgb = _mm_mul_pd(qq,VV); fijC = _mm_mul_pd(qq,FF); fijC = _mm_mul_pd(fijC,gbscale); dvdatmp = _mm_mul_pd(fijC,r); dvdatmp = _mm_add_pd(vgb,dvdatmp); dvdatmp = _mm_mul_pd(dvdatmp,neg); dvdatmp = _mm_mul_pd(dvdatmp,half); dvdasum = _mm_add_pd(dvdasum,dvdatmp); xmm1 = _mm_mul_pd(dvdatmp,isaj); xmm1 = _mm_mul_pd(xmm1,isaj); dvdaj = _mm_add_pd(dvdaj,xmm1); /* store dvda */ _mm_storel_pd(dvda+jnr1,dvdaj); _mm_storeh_pd(dvda+jnr2,dvdaj); vctot = _mm_add_pd(vctot,vcoul); vgbtot = _mm_add_pd(vgbtot,vgb); fscal = _mm_sub_pd(fijC,fscal); fscal = _mm_mul_pd(fscal,neg); fscal = _mm_mul_pd(fscal,rinv); /* calculate partial force terms */ t1 = _mm_mul_pd(fscal,dx); t2 = _mm_mul_pd(fscal,dy); t3 = _mm_mul_pd(fscal,dz); /* update the i force */ fix = _mm_add_pd(fix,t1); fiy = _mm_add_pd(fiy,t2); fiz = _mm_add_pd(fiz,t3); /* accumulate forces from memory */ xmm1 = _mm_loadu_pd(faction+j13); /* fx1 fy1 */ xmm2 = _mm_loadu_pd(faction+j23); /* fx2 fy2 */ xmm5 = _mm_load1_pd(faction+j13+2); /* fz1 fz1 */ xmm6 = _mm_load1_pd(faction+j23+2); /* fz2 fz2 */ /* transpose */ xmm7 = _mm_shuffle_pd(xmm5,xmm6,_MM_SHUFFLE2(0,0)); /* fz1 fz2 */ xmm5 = _mm_shuffle_pd(xmm1,xmm2,_MM_SHUFFLE2(0,0)); /* fx1 fx2 */ xmm6 = _mm_shuffle_pd(xmm1,xmm2,_MM_SHUFFLE2(1,1)); /* fy1 fy2 */ /* subtract partial forces */ xmm5 = _mm_sub_pd(xmm5,t1); xmm6 = _mm_sub_pd(xmm6,t2); xmm7 = _mm_sub_pd(xmm7,t3); xmm1 = _mm_shuffle_pd(xmm5,xmm6,_MM_SHUFFLE2(0,0)); /* fx1 fy1 */ xmm2 = _mm_shuffle_pd(xmm5,xmm6,_MM_SHUFFLE2(1,1)); /* fy1 fy2 */ /* store fx and fy */ _mm_storeu_pd(faction+j13,xmm1); _mm_storeu_pd(faction+j23,xmm2); /* .. then fz */ _mm_storel_pd(faction+j13+2,xmm7); _mm_storel_pd(faction+j23+2,xmm7); } /* In double precision, offset can only be either 0 or 1 */ if(offset!=0) { jnr1 = jjnr[k]; j13 = jnr1*3; jx = _mm_load_sd(pos+j13); jy = _mm_load_sd(pos+j13+1); jz = _mm_load_sd(pos+j13+2); isaj = _mm_load_sd(invsqrta+jnr1); isaprod = _mm_mul_sd(isai,isaj); dvdaj = _mm_load_sd(dvda+jnr1); q = _mm_load_sd(charge+jnr1); qq = _mm_mul_sd(iq,q); dx = _mm_sub_sd(ix,jx); dy = _mm_sub_sd(iy,jy); dz = _mm_sub_sd(iz,jz); rsq11 = _mm_add_pd( _mm_add_pd( _mm_mul_pd(dx,dx) , _mm_mul_pd(dy,dy) ) , _mm_mul_pd(dz,dz) ); rinv = my_invrsq_pd(rsq11); vcoul = _mm_mul_sd(qq,rinv); fscal = _mm_mul_sd(vcoul,rinv); qq = _mm_mul_sd(isaprod,qq); qq = _mm_mul_sd(qq,neg); gbscale = _mm_mul_sd(isaprod,gbtabscale); r = _mm_mul_sd(rsq11,rinv); rt = _mm_mul_sd(r,gbscale); n0 = _mm_cvttpd_epi32(rt); n0d = _mm_cvtepi32_pd(n0); eps = _mm_sub_sd(rt,n0d); eps2 = _mm_mul_sd(eps,eps); nnn = _mm_slli_epi64(n0,2); xmm1 = _mm_load_pd(GBtab+(_mm_extract_epi64(nnn,0))); xmm2 = _mm_load_pd(GBtab+(_mm_extract_epi64(nnn,1))); xmm3 = _mm_load_pd(GBtab+(_mm_extract_epi64(nnn,0))+2); xmm4 = _mm_load_pd(GBtab+(_mm_extract_epi64(nnn,1))+2); Y = _mm_shuffle_pd(xmm1,xmm2,_MM_SHUFFLE2(0,0)); F = _mm_shuffle_pd(xmm1,xmm2,_MM_SHUFFLE2(1,1)); G = _mm_shuffle_pd(xmm3,xmm4,_MM_SHUFFLE2(0,0)); H = _mm_shuffle_pd(xmm3,xmm4,_MM_SHUFFLE2(1,1)); G = _mm_mul_sd(G,eps); H = _mm_mul_sd(H,eps2); Fp = _mm_add_sd(F,G); Fp = _mm_add_sd(Fp,H); VV = _mm_mul_sd(Fp,eps); VV = _mm_add_sd(Y,VV); H = _mm_mul_sd(two,H); FF = _mm_add_sd(Fp,G); FF = _mm_add_sd(FF,H); vgb = _mm_mul_sd(qq,VV); fijC = _mm_mul_sd(qq,FF); fijC = _mm_mul_sd(fijC,gbscale); dvdatmp = _mm_mul_sd(fijC,r); dvdatmp = _mm_add_sd(vgb,dvdatmp); dvdatmp = _mm_mul_sd(dvdatmp,neg); dvdatmp = _mm_mul_sd(dvdatmp,half); dvdasum = _mm_add_sd(dvdasum,dvdatmp); xmm1 = _mm_mul_sd(dvdatmp,isaj); xmm1 = _mm_mul_sd(xmm1,isaj); dvdaj = _mm_add_sd(dvdaj,xmm1); /* store dvda */ _mm_storel_pd(dvda+jnr1,dvdaj); vctot = _mm_add_sd(vctot,vcoul); vgbtot = _mm_add_sd(vgbtot,vgb); fscal = _mm_sub_sd(fijC,fscal); fscal = _mm_mul_sd(fscal,neg); fscal = _mm_mul_sd(fscal,rinv); /* calculate partial force terms */ t1 = _mm_mul_sd(fscal,dx); t2 = _mm_mul_sd(fscal,dy); t3 = _mm_mul_sd(fscal,dz); /* update the i force */ fix = _mm_add_sd(fix,t1); fiy = _mm_add_sd(fiy,t2); fiz = _mm_add_sd(fiz,t3); /* accumulate forces from memory */ xmm5 = _mm_load_sd(faction+j13); /* fx */ xmm6 = _mm_load_sd(faction+j13+1); /* fy */ xmm7 = _mm_load_sd(faction+j13+2); /* fz */ /* subtract partial forces */ xmm5 = _mm_sub_sd(xmm5,t1); xmm6 = _mm_sub_sd(xmm6,t2); xmm7 = _mm_sub_sd(xmm7,t3); /* store forces */ _mm_store_sd(faction+j13,xmm5); _mm_store_sd(faction+j13+1,xmm6); _mm_store_sd(faction+j13+2,xmm7); } /* fix/fiy/fiz now contain four partial terms, that all should be * added to the i particle forces */ t1 = _mm_unpacklo_pd(t1,fix); t2 = _mm_unpacklo_pd(t2,fiy); t3 = _mm_unpacklo_pd(t3,fiz); fix = _mm_add_pd(fix,t1); fiy = _mm_add_pd(fiy,t2); fiz = _mm_add_pd(fiz,t3); fix = _mm_shuffle_pd(fix,fix,_MM_SHUFFLE2(1,1)); fiy = _mm_shuffle_pd(fiy,fiy,_MM_SHUFFLE2(1,1)); fiz = _mm_shuffle_pd(fiz,fiz,_MM_SHUFFLE2(1,1)); /* Load i forces from memory */ xmm1 = _mm_load_sd(faction+ii3); xmm2 = _mm_load_sd(faction+ii3+1); xmm3 = _mm_load_sd(faction+ii3+2); /* Add to i force */ fix = _mm_add_sd(fix,xmm1); fiy = _mm_add_sd(fiy,xmm2); fiz = _mm_add_sd(fiz,xmm3); /* store i forces to memory */ _mm_store_sd(faction+ii3,fix); _mm_store_sd(faction+ii3+1,fiy); _mm_store_sd(faction+ii3+2,fiz); /* now do dvda */ dvdatmp = _mm_unpacklo_pd(dvdatmp,dvdasum); dvdasum = _mm_add_pd(dvdasum,dvdatmp); _mm_storeh_pd(&dva,dvdasum); dvda[ii] = dvda[ii] + dva*isai_d*isai_d; ggid = gid[n]; /* Coulomb potential */ vcoul = _mm_unpacklo_pd(vcoul,vctot); vctot = _mm_add_pd(vctot,vcoul); _mm_storeh_pd(&vct,vctot); Vc[ggid] = Vc[ggid] + vct; /* GB potential */ vgb = _mm_unpacklo_pd(vgb,vgbtot); vgbtot = _mm_add_pd(vgbtot,vgb); _mm_storeh_pd(&vgbt,vgbtot); gpol[ggid] = gpol[ggid] + vgbt; } *outeriter = nri; *inneriter = nj1; }
tag::cpu_, Dummy> : callable { template<class Sig> struct result; template<class This,class A0> struct result<This(A0, A0)> { typedef typename meta::scalar_of<A0>::type stype; typedef typename meta::downgrade<stype>::type sftype; typedef simd::native<sftype,tag::sse_> type; }; NT2_FUNCTOR_CALL(2) { typedef typename NT2_RETURN_TYPE(2)::type rtype; typedef typename meta::as_integer<rtype>::type itype; rtype z = {_mm_cvtpd_ps(a1)}; itype iz = simd::native_cast<itype>(z); iz = _mm_slli_si128(iz, 8); rtype r = simd::native_cast<rtype>(iz); return b_or(r, simd::native_cast<rtype>(_mm_cvtpd_ps(a0))); } }; } } ///////////////////////////////////////////////////////////////////////////// // Implementation when type A0 is int32_t ///////////////////////////////////////////////////////////////////////////// NT2_REGISTER_DISPATCH(tag::group_, tag::cpu_, (A0), ((simd_<ints32_<A0>,tag::sse_>)) ((simd_<ints32_<A0>,tag::sse_>))
tag::cpu_, Dummy> : callable { template<class Sig> struct result; template<class This,class A0> struct result<This(A0, A0)> { typedef typename meta::scalar_of<A0>::type stype; typedef typename meta::float__<stype>::type sftype; typedef simd::native<sftype,tag::sse_> type; }; NT2_FUNCTOR_CALL(2) { typedef typename NT2_RETURN_TYPE(2)::type rtype; typedef typename meta::as_integer<rtype>::type itype; rtype r = simd::native_cast<rtype>(_mm_slli_si128(simd::native_cast<itype >(_mm_cvtpd_ps(a1)), 8)); return b_or(r, simd::native_cast<rtype>(_mm_cvtpd_ps(a0))); } }; } } ///////////////////////////////////////////////////////////////////////////// // Implementation when type A0 is int32_t ///////////////////////////////////////////////////////////////////////////// NT2_REGISTER_DISPATCH(tag::group_, tag::cpu_, (A0), ((simd_<ints32_<A0>,tag::sse_>)) ((simd_<ints32_<A0>,tag::sse_>)) ); namespace nt2 { namespace ext
// ============================================================================= // // sse2_vChirpData // version by: Alex Kan - SSE2 mods (haddsum removal) BH // http://tbp.berkeley.edu/~alexkan/seti/ // int sse2_ChirpData_ak( sah_complex * cx_DataArray, sah_complex * cx_ChirpDataArray, int chirp_rate_ind, double chirp_rate, int ul_NumDataPoints, double sample_rate ) { int i; if (chirp_rate_ind == 0) { memcpy(cx_ChirpDataArray, cx_DataArray, (int)ul_NumDataPoints * sizeof(sah_complex) ); return 0; } int vEnd; double srate = chirp_rate * 0.5 / (sample_rate * sample_rate); __m128d rate = _mm_set1_pd(chirp_rate * 0.5 / (sample_rate * sample_rate)); __m128d roundVal = _mm_set1_pd(srate >= 0.0 ? TWO_TO_52 : -TWO_TO_52); // main vectorised loop vEnd = ul_NumDataPoints - (ul_NumDataPoints & 3); for (i = 0; i < vEnd; i += 4) { const float *data = (const float *) (cx_DataArray + i); float *chirped = (float *) (cx_ChirpDataArray + i); __m128d di = _mm_set1_pd(i); __m128d a1 = _mm_add_pd(_mm_set_pd(1.0, 0.0), di); __m128d a2 = _mm_add_pd(_mm_set_pd(3.0, 2.0), di); __m128d x1, y1; __m128 d1, d2; __m128 cd1, cd2; __m128 td1, td2; __m128 x; __m128 y; __m128 s; __m128 c; __m128 m; // load the signal to be chirped prefetchnta((const void *)( data+32 )); d1 = _mm_load_ps(data); d2 = _mm_load_ps(data+4); // calculate the input angle a1 = _mm_mul_pd(a1, a1); a2 = _mm_mul_pd(a2, a2); a1 = _mm_mul_pd(a1, rate); a2 = _mm_mul_pd(a2, rate); // reduce the angle to the range (-0.5, 0.5) x1 = _mm_add_pd(a1, roundVal); y1 = _mm_add_pd(a2, roundVal); x1 = _mm_sub_pd(x1, roundVal); y1 = _mm_sub_pd(y1, roundVal); a1 = _mm_sub_pd(a1, x1); a2 = _mm_sub_pd(a2, y1); // convert pair of packed double into packed single x = _mm_movelh_ps(_mm_cvtpd_ps(a1), _mm_cvtpd_ps(a2)); // square to the range [0, 0.25) y = _mm_mul_ps(x, x); // perform the initial polynomial approximations s = _mm_mul_ps(y, SS4); c = _mm_mul_ps(y, CC3); s = _mm_add_ps(s, SS3); c = _mm_add_ps(c, CC2); s = _mm_mul_ps(s, y); c = _mm_mul_ps(c, y); s = _mm_add_ps(s, SS2); c = _mm_add_ps(c, CC1); s = _mm_mul_ps(s, y); c = _mm_mul_ps(c, y); s = _mm_add_ps(s, SS1); s = _mm_mul_ps(s, x); c = _mm_add_ps(c, ONE); // perform first angle doubling x = _mm_sub_ps(_mm_mul_ps(c, c), _mm_mul_ps(s, s)); y = _mm_mul_ps(_mm_mul_ps(s, c), TWO); // calculate scaling factor to correct the magnitude // m1 = vec_nmsub(y1, y1, vec_nmsub(x1, x1, TWO)); // m2 = vec_nmsub(y2, y2, vec_nmsub(x2, x2, TWO)); m = vec_recip2(_mm_add_ps(_mm_mul_ps(x, x), _mm_mul_ps(y, y))); // perform second angle doubling c = _mm_sub_ps(_mm_mul_ps(x, x), _mm_mul_ps(y, y)); s = _mm_mul_ps(_mm_mul_ps(y, x), TWO); // correct the magnitude (final sine / cosine approximations) c = _mm_mul_ps(c, m); s = _mm_mul_ps(s, m); /* c1 c2 c3 c4 s1 s2 s3 s4 R1 i1 R2 I2 R3 i3 R4 i4 R1 * c1 + (i1 * s1 * -1) i1 * c1 + R1 * s1 R2 * c2 + (i2 * s2 * -1) i2 * c2 + R2 * s2 */ x = d1; y = d2; x = _mm_shuffle_ps(x, x, 0xB1); y = _mm_shuffle_ps(y, y, 0xB1); x = _mm_mul_ps(x, R_NEG); y = _mm_mul_ps(y, R_NEG); cd1 = _mm_shuffle_ps(c, c, 0x50); // 01 01 00 00 AaBb => BBbb => c3c3c4c4 cd2 = _mm_shuffle_ps(c, c, 0xfa); // 11 11 10 10 AaBb => AAaa => c1c1c2c2 td1 = _mm_shuffle_ps(s, s, 0x50); td2 = _mm_shuffle_ps(s, s, 0xfa); cd1 = _mm_mul_ps(cd1, d1); cd2 = _mm_mul_ps(cd2, d2); td1 = _mm_mul_ps(td1, x); td2 = _mm_mul_ps(td2, y); cd1 = _mm_add_ps(cd1, td1); cd2 = _mm_add_ps(cd2, td2); // store chirped values _mm_stream_ps(chirped+0, cd1); _mm_stream_ps(chirped+4, cd2); } _mm_sfence(); if( i < ul_NumDataPoints) { // use original routine to finish up any tailings (max stride-1 elements) v_ChirpData(cx_DataArray+i, cx_ChirpDataArray+i , chirp_rate_ind, chirp_rate, ul_NumDataPoints-i, sample_rate); } analysis_state.FLOP_counter+=12.0*ul_NumDataPoints; return 0; }
int sse3_ChirpData_ak8( sah_complex * cx_DataArray, sah_complex * cx_ChirpDataArray, int chirp_rate_ind, double chirp_rate, int ul_NumDataPoints, double sample_rate ) { #ifdef USE_MANUAL_CALLSTACK call_stack.enter("sse3_ChirpData_ak8()"); #endif int i; if (chirp_rate_ind == 0) { memcpy(cx_ChirpDataArray, cx_DataArray, (int)ul_NumDataPoints * sizeof(sah_complex) ); #ifdef USE_MANUAL_CALLSTACK call_stack.exit(); #endif return 0; } int vEnd; double srate = chirp_rate * 0.5 / (sample_rate * sample_rate); __m128d rate = _mm_set1_pd(chirp_rate * 0.5 / (sample_rate * sample_rate)); __m128d roundVal = _mm_set1_pd(srate >= 0.0 ? TWO_TO_52 : -TWO_TO_52); __m128d DFOUR = _mm_set_pd(4.0, 4.0); // main vectorised loop vEnd = ul_NumDataPoints - (ul_NumDataPoints & 3); __m128d di1 = _mm_set_pd(2.0, 0.0); // set time patterns for eventual moveldup/movehdup __m128d di2 = _mm_set_pd(3.0, 1.0); for (i = 0; i < vEnd; i += 4) { const float *d = (const float *) (cx_DataArray + i); float *cd = (float *) (cx_ChirpDataArray + i); __m128d a1, a2; __m128 d1, d2; __m128 cd1, cd2; __m128 td1, td2; __m128 x; __m128 y; __m128 z; __m128 s; __m128 c; __m128 m; // load the signal to be chirped d1 = _mm_load_ps(d); d2 = _mm_load_ps(d+4); // calculate the input angle a1 = _mm_mul_pd(_mm_mul_pd(di1, di1), rate); a2 = _mm_mul_pd(_mm_mul_pd(di2, di2), rate); // update times for next di1 = _mm_add_pd(di1, DFOUR); di2 = _mm_add_pd(di2, DFOUR); // reduce the angle to the range (-0.5, 0.5) a1 = _mm_sub_pd(a1, _mm_sub_pd(_mm_add_pd(a1, roundVal), roundVal)); a2 = _mm_sub_pd(a2, _mm_sub_pd(_mm_add_pd(a2, roundVal), roundVal)); // convert pair of packed double into packed single x = _mm_movelh_ps(_mm_cvtpd_ps(a1), _mm_cvtpd_ps(a2)); // 3 1 2 0 // square to the range [0, 0.25) y = _mm_mul_ps(x, x); // perform the initial polynomial approximations, Estrin's method z = _mm_mul_ps(y, y); s = _mm_mul_ps(_mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(y, SS4F), SS3F), z), _mm_add_ps(_mm_mul_ps(y, SS2F), SS1F)), x); c = _mm_add_ps(_mm_mul_ps(_mm_add_ps(_mm_mul_ps(y, CC3F), CC2F), z), _mm_add_ps(_mm_mul_ps(y, CC1F), ONE)); // perform first angle doubling x = _mm_sub_ps(_mm_mul_ps(c, c), _mm_mul_ps(s, s)); y = _mm_mul_ps(_mm_mul_ps(s, c), TWO); // calculate scaling factor to correct the magnitude m = _mm_sub_ps(_mm_sub_ps(TWO, _mm_mul_ps(x, x)), _mm_mul_ps(y, y)); // perform second angle doubling c = _mm_sub_ps(_mm_mul_ps(x, x), _mm_mul_ps(y, y)); s = _mm_mul_ps(_mm_mul_ps(y, x), TWO); // correct the magnitude (final sine / cosine approximations) c = _mm_mul_ps(c, m); // c3 c1 c2 c0 s = _mm_mul_ps(s, m); // chirp the data cd1 = _mm_moveldup_ps(c); // c1 c1 c0 c0 cd2 = _mm_movehdup_ps(c); // c3 c3 c2 c2 cd1 = _mm_mul_ps(cd1, d1); // c1.i1 c1.r1 c0.i0 c0.r0 cd2 = _mm_mul_ps(cd2, d2); // c3.i3 c3.r3 c2.i2 c2.r2 d1 = _mm_shuffle_ps(d1, d1, 0xb1); d2 = _mm_shuffle_ps(d2, d2, 0xb1); td1 = _mm_moveldup_ps(s); td2 = _mm_movehdup_ps(s); td1 = _mm_mul_ps(td1, d1); td2 = _mm_mul_ps(td2, d2); cd1 = _mm_addsub_ps(cd1, td1); cd2 = _mm_addsub_ps(cd2, td2); // store chirped values _mm_stream_ps(cd, cd1); _mm_stream_ps(cd+4, cd2); } // handle tail elements with scalar code for (; i < ul_NumDataPoints; ++i) { double angle = srate * i * i * 0.5; double s = sin(angle); double c = cos(angle); float re = cx_DataArray[i][0]; float im = cx_DataArray[i][1]; cx_ChirpDataArray[i][0] = re * c - im * s; cx_ChirpDataArray[i][1] = re * s + im * c; } analysis_state.FLOP_counter+=12.0*ul_NumDataPoints; #ifdef USE_MANUAL_CALLSTACK call_stack.exit(); #endif return 0; }