Fulguro: /home/christophe/src/fulguro/trunk/src/arith/flgrArithSSE.c Source File

00001 /****************************************************************
00002  * Fulguro
00003  * Copyright (C) 2004 Christophe Clienti
00004  *
00005  * This library is free software; you can redistribute it and/or
00006  * modify it under the terms of the GNU Lesser General Public
00007  * License as published by the Free Software Foundation; either
00008  * version 2 of the License, or (at your option) any later version.
00009  *
00010  * This library is distributed in the hope that it will be useful,
00011  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00012  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
00013  * Lesser General Public License for more details.
00014  *
00015  * You should have received a copy of the GNU Lesser General Public
00016  * License along with this library; if not, write to the
00017  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
00018  * Boston, MA 02111-1307, USA.
00019  ***************************************************************/
00020 
00021 #include <string.h>
00022 #include <flgrCoreVector.h>
00023 #include <flgrCoreDispatch.h>
00024 
00025 #ifdef __SSE__
00026 #include <mmintrin.h>
00027 #include <xmmintrin.h>
00028 #endif
00029 
00030 #include "flgrArithSSE.h"
00031 
00032 #define FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(instr,dtype,set_vec_const) \
00033   __m64 r1,r2;                          \
00034   void *psrc1 = data1->array;                   \
00035   void *pdest = datdest->array;                 \
00036   int j,c = (sizeof(__m64)*8)/data1->bps;           \
00037   int length = datdest->length;                 \
00038                                 \
00039                     \
00040                                 \
00041   _mm_empty();                          \
00042   r2=set_vec_const(*((dtype*) (vec_const->array)));     \
00043                                 \
00044                                 \
00045   for(j=0;j<length;j+=c) {                  \
00046     r1 = *((__m64*) psrc1);                 \
00047     r1 = instr(r1,r2);                      \
00048     *((__m64*) pdest) = r1;                 \
00049                                 \
00050     psrc1+=sizeof(__m64);                   \
00051     pdest+=sizeof(__m64);                   \
00052   }                             \
00053                                 \
00054   _mm_empty();                          \
00055                                 \
00056   return
00057 
00058 
00059 #define FLGR_MACRO_ARITH_1D_DIADIC_SSE(instr)   \
00060   __m64 r1,r2;                  \
00061   void *psrc1 = data1->array;           \
00062   void *psrc2 = data2->array;           \
00063   void *pdest = datdest->array;         \
00064   int j,c = (sizeof(__m64)*8)/data1->bps;   \
00065   int length = datdest->length*datdest->spp;    \
00066                         \
00067             \
00068                         \
00069   _mm_empty();                  \
00070                         \
00071   for(j=0;j<length;j+=c) {          \
00072     r1 = *((__m64*) psrc1);         \
00073     r2 = *((__m64*) psrc2);         \
00074     r1 = instr(r1,r2);              \
00075     *((__m64*) pdest) = r1;         \
00076                         \
00077     psrc1+=sizeof(__m64);           \
00078     psrc2+=sizeof(__m64);           \
00079     pdest+=sizeof(__m64);           \
00080   }                     \
00081                         \
00082   _mm_empty();                  \
00083   return
00084 
00085 
00086 void flgr1d_arith_inf_SSE_fgBIT(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00087 #ifdef __SSE__
00088   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_and_si64);
00089 #endif
00090 }
00091 void flgr1d_arith_inf_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00092 #ifdef __SSE__
00093   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_min_pu8);
00094 #endif
00095 }
00096 void flgr1d_arith_inf_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00097 #ifdef __SSE__
00098   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_min_pi16);
00099 #endif
00100 }
00101 
00102 
00103 void flgr1d_arith_sup_SSE_fgBIT(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00104 #ifdef __SSE__
00105   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_or_si64);
00106 #endif
00107 }
00108 void flgr1d_arith_sup_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00109 #ifdef __SSE__
00110   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_max_pu8);
00111 #endif
00112 }
00113 void flgr1d_arith_sup_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00114 #ifdef __SSE__
00115   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_max_pi16);
00116 #endif
00117 }
00118 
00119 
00120 
00121 void flgr1d_arith_and_SSE_fgBIT(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00122 #ifdef __SSE__
00123   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_and_si64);
00124 #endif
00125 }
00126 void flgr1d_arith_and_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00127 #ifdef __SSE__
00128   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_and_si64);
00129 #endif
00130 }
00131 void flgr1d_arith_and_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00132 #ifdef __SSE__
00133   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_and_si64);
00134 #endif
00135 }
00136 void flgr1d_arith_and_SSE_fgUINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00137 #ifdef __SSE__
00138   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_and_si64);
00139 #endif
00140 }
00141 void flgr1d_arith_and_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00142 #ifdef __SSE__
00143   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_and_si64);
00144 #endif
00145 }
00146 void flgr1d_arith_and_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00147 #ifdef __SSE__
00148   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_and_si64);
00149 #endif
00150 }
00151 void flgr1d_arith_and_SSE_fgINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00152 #ifdef __SSE__
00153   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_and_si64);
00154 #endif
00155 }
00156 
00157 
00158 
00159 
00160 void flgr1d_arith_or_SSE_fgBIT(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00161 #ifdef __SSE__
00162   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_or_si64);
00163 #endif
00164 }
00165 void flgr1d_arith_or_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00166 #ifdef __SSE__
00167   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_or_si64);
00168 #endif
00169 }
00170 void flgr1d_arith_or_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00171 #ifdef __SSE__
00172   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_or_si64);
00173 #endif
00174 }
00175 void flgr1d_arith_or_SSE_fgUINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00176 #ifdef __SSE__
00177   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_or_si64);
00178 #endif
00179 }
00180 void flgr1d_arith_or_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00181 #ifdef __SSE__
00182   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_or_si64);
00183 #endif
00184 }
00185 void flgr1d_arith_or_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00186 #ifdef __SSE__
00187   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_or_si64);
00188 #endif
00189 }
00190 void flgr1d_arith_or_SSE_fgINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00191 #ifdef __SSE__
00192   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_or_si64);
00193 #endif
00194 }
00195 
00196 
00197 
00198 
00199 void flgr1d_arith_xor_SSE_fgBIT(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00200 #ifdef __SSE__
00201   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00202 #endif
00203 }
00204 void flgr1d_arith_xor_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00205 #ifdef __SSE__
00206   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00207 #endif
00208 }
00209 void flgr1d_arith_xor_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00210 #ifdef __SSE__
00211   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00212 #endif
00213 }
00214 void flgr1d_arith_xor_SSE_fgUINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00215 #ifdef __SSE__
00216   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00217 #endif
00218 }
00219 void flgr1d_arith_xor_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00220 #ifdef __SSE__
00221   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00222 #endif
00223 }
00224 void flgr1d_arith_xor_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00225 #ifdef __SSE__
00226   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00227 #endif
00228 }
00229 void flgr1d_arith_xor_SSE_fgINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00230 #ifdef __SSE__
00231   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00232 #endif
00233 }
00234 
00235 
00236 
00237 void flgr1d_arith_add_SSE_fgBIT(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00238 #ifdef __SSE__
00239   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00240 #endif
00241 }
00242 void flgr1d_arith_add_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00243 #ifdef __SSE__
00244   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_add_pi8);
00245 #endif
00246 }
00247 void flgr1d_arith_add_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00248 #ifdef __SSE__
00249   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_add_pi16);
00250 #endif
00251 }
00252 void flgr1d_arith_add_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00253 #ifdef __SSE__
00254   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_add_pi8);
00255 #endif
00256 }
00257 void flgr1d_arith_add_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00258 #ifdef __SSE__
00259   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_add_pi16);
00260 #endif
00261 }
00262 
00263 
00264 
00265 void flgr1d_arith_addsat_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00266 #ifdef __SSE__
00267   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_adds_pu8);
00268 #endif
00269 }
00270 void flgr1d_arith_addsat_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00271 #ifdef __SSE__
00272   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_adds_pu16);
00273 #endif
00274 }
00275 void flgr1d_arith_addsat_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00276 #ifdef __SSE__
00277   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_adds_pi16);
00278 #endif
00279 }
00280 
00281 
00282 
00283 void flgr1d_arith_sub_SSE_fgBIT(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00284 #ifdef __SSE__
00285   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_xor_si64);
00286 #endif
00287 }
00288 void flgr1d_arith_sub_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00289 #ifdef __SSE__
00290   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_sub_pi8);
00291 #endif
00292 }
00293 void flgr1d_arith_sub_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00294 #ifdef __SSE__
00295   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_sub_pi16);
00296 #endif
00297 }
00298 void flgr1d_arith_sub_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00299 #ifdef __SSE__
00300   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_sub_pi8);
00301 #endif
00302 }
00303 void flgr1d_arith_sub_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00304 #ifdef __SSE__
00305   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_sub_pi16);
00306 #endif
00307 }
00308 
00309 
00310 
00311 void flgr1d_arith_subsat_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00312 #ifdef __SSE__
00313   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_subs_pu8);
00314 #endif
00315 }
00316 void flgr1d_arith_subsat_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00317 #ifdef __SSE__
00318   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_subs_pu16);
00319 #endif
00320 }
00321 void flgr1d_arith_subsat_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Data1D *data2) {
00322 #ifdef __SSE__
00323   FLGR_MACRO_ARITH_1D_DIADIC_SSE(_mm_subs_pi16);
00324 #endif
00325 }
00326 
00327 
00328 
00329 
00330 
00331 
00332 
00333 void flgr1d_arith_const_inf_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00334 #ifdef __SSE__
00335   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_min_pu8,fgUINT8,_mm_set1_pi8);
00336 #endif
00337 }
00338 void flgr1d_arith_const_inf_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00339 #ifdef __SSE__
00340   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_min_pi16,fgINT16,_mm_set1_pi16);
00341 #endif
00342 }
00343 
00344 
00345 
00346 void flgr1d_arith_const_sup_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00347 #ifdef __SSE__
00348   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_max_pu8,fgUINT8,_mm_set1_pi8);
00349 #endif
00350 }
00351 void flgr1d_arith_const_sup_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00352 #ifdef __SSE__
00353   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_max_pi16,fgINT16,_mm_set1_pi16);
00354 #endif
00355 }
00356 
00357 
00358 
00359 void flgr1d_arith_const_and_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00360 #ifdef __SSE__
00361   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_and_si64,fgUINT8,_mm_set1_pi8);
00362 #endif
00363 }
00364 void flgr1d_arith_const_and_spp1_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00365 #ifdef __SSE__
00366   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_and_si64,fgUINT16,_mm_set1_pi16);
00367 #endif
00368 }
00369 void flgr1d_arith_const_and_spp1_SSE_fgUINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00370 #ifdef __SSE__
00371   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_and_si64,fgUINT32,_mm_set1_pi32);
00372 #endif
00373 }
00374 void flgr1d_arith_const_and_spp1_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00375 #ifdef __SSE__
00376   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_and_si64,fgINT8,_mm_set1_pi8);
00377 #endif
00378 }
00379 void flgr1d_arith_const_and_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00380 #ifdef __SSE__
00381   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_and_si64,fgINT16,_mm_set1_pi16);
00382 #endif
00383 }
00384 void flgr1d_arith_const_and_spp1_SSE_fgINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00385 #ifdef __SSE__
00386   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_and_si64,fgINT32,_mm_set1_pi32);
00387 #endif
00388 }
00389 
00390 
00391 
00392 void flgr1d_arith_const_or_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00393 #ifdef __SSE__
00394   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_or_si64,fgUINT8,_mm_set1_pi8);
00395 #endif
00396 }
00397 void flgr1d_arith_const_or_spp1_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00398 #ifdef __SSE__
00399   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_or_si64,fgUINT16,_mm_set1_pi16);
00400 #endif
00401 }
00402 void flgr1d_arith_const_or_spp1_SSE_fgUINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00403 #ifdef __SSE__
00404   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_or_si64,fgUINT32,_mm_set1_pi32);
00405 #endif
00406 }
00407 void flgr1d_arith_const_or_spp1_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00408 #ifdef __SSE__
00409   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_or_si64,fgINT8,_mm_set1_pi8);
00410 #endif
00411 }
00412 void flgr1d_arith_const_or_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00413 #ifdef __SSE__
00414   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_or_si64,fgINT16,_mm_set1_pi16);
00415 #endif
00416 }
00417 void flgr1d_arith_const_or_spp1_SSE_fgINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00418 #ifdef __SSE__
00419   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_or_si64,fgINT32,_mm_set1_pi32);
00420 #endif
00421 }
00422 
00423 
00424 
00425 void flgr1d_arith_const_xor_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00426 #ifdef __SSE__
00427   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_xor_si64,fgUINT8,_mm_set1_pi8);
00428 #endif
00429 }
00430 void flgr1d_arith_const_xor_spp1_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00431 #ifdef __SSE__
00432   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_xor_si64,fgUINT16,_mm_set1_pi16);
00433 #endif
00434 }
00435 void flgr1d_arith_const_xor_spp1_SSE_fgUINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00436 #ifdef __SSE__
00437   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_xor_si64,fgUINT32,_mm_set1_pi32);
00438 #endif
00439 }
00440 void flgr1d_arith_const_xor_spp1_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00441 #ifdef __SSE__
00442   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_xor_si64,fgINT8,_mm_set1_pi8);
00443 #endif
00444 }
00445 void flgr1d_arith_const_xor_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00446 #ifdef __SSE__
00447   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_xor_si64,fgINT16,_mm_set1_pi16);
00448 #endif
00449 }
00450 void flgr1d_arith_const_xor_spp1_SSE_fgINT32(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00451 #ifdef __SSE__
00452   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_xor_si64,fgINT32,_mm_set1_pi32);
00453 #endif
00454 }
00455 
00456 
00457 
00458 void flgr1d_arith_const_add_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00459 #ifdef __SSE__
00460   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_add_pi8,fgUINT8,_mm_set1_pi8);
00461 #endif
00462 }
00463 void flgr1d_arith_const_add_spp1_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00464 #ifdef __SSE__
00465   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_add_pi16,fgUINT16,_mm_set1_pi16);
00466 #endif
00467 }
00468 void flgr1d_arith_const_add_spp1_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00469 #ifdef __SSE__
00470   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_add_pi8,fgINT8,_mm_set1_pi8);
00471 #endif
00472 }
00473 void flgr1d_arith_const_add_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00474 #ifdef __SSE__
00475   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_add_pi16,fgINT16,_mm_set1_pi16);
00476 #endif
00477 }
00478 
00479 
00480 
00481 void flgr1d_arith_const_addsat_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00482 #ifdef __SSE__
00483   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_adds_pu8,fgUINT8,_mm_set1_pi8);
00484 #endif
00485 }
00486 void flgr1d_arith_const_addsat_spp1_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00487 #ifdef __SSE__
00488   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_adds_pu16,fgUINT16,_mm_set1_pi16);
00489 #endif
00490 }
00491 void flgr1d_arith_const_addsat_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00492 #ifdef __SSE__
00493   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_adds_pi16,fgINT16,_mm_set1_pi16);
00494 #endif
00495 }
00496 
00497 
00498 
00499 void flgr1d_arith_const_sub_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00500 #ifdef __SSE__
00501  FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_sub_pi8,fgUINT8,_mm_set1_pi8);
00502 #endif
00503 }
00504 void flgr1d_arith_const_sub_spp1_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00505 #ifdef __SSE__
00506  FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_sub_pi16,fgUINT16,_mm_set1_pi16);
00507 #endif
00508 }
00509 void flgr1d_arith_const_sub_spp1_SSE_fgINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00510 #ifdef __SSE__
00511  FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_sub_pi8,fgINT8,_mm_set1_pi8);
00512 #endif
00513 }
00514 void flgr1d_arith_const_sub_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00515 #ifdef __SSE__
00516  FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_sub_pi16,fgINT16,_mm_set1_pi16);
00517 #endif
00518 }
00519 
00520 
00521 
00522 void flgr1d_arith_const_subsat_spp1_SSE_fgUINT8(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00523 #ifdef __SSE__
00524   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_subs_pu8,fgUINT8,_mm_set1_pi8);
00525 #endif
00526 }
00527 void flgr1d_arith_const_subsat_spp1_SSE_fgUINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00528 #ifdef __SSE__
00529   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_subs_pu16,fgUINT16,_mm_set1_pi16);
00530 #endif
00531 }
00532 void flgr1d_arith_const_subsat_spp1_SSE_fgINT16(FLGR_Data1D *datdest, FLGR_Data1D *data1, FLGR_Vector *vec_const) {
00533 #ifdef __SSE__
00534   FLGR_MACRO_ARITH_1D_DIADIC_CONST_SPP1_SSE(_mm_subs_pi16,fgINT16,_mm_set1_pi16);
00535 #endif
00536 }