Fulguro: /home/christophe/src/fulguro/trunk/src/core/flgrCoreReplaceFast.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 <stdio.h>
00022 #include <stdlib.h>
00023 #include <string.h>
00024 
00025 #ifdef __SSE2__
00026 #include <emmintrin.h>
00027 #include <xmmintrin.h>
00028 #endif
00029 
00030 #include "flgrCoreDispatch.h"
00031 #include "flgrCoreReplaceFast.h"
00032 
00033 
00034 void flgr1d_replace_I_LE_I_I_I_fast_fgUINT8(FLGR_Data1D *datout,
00035                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00036                         FLGR_Data1D *dattrue, FLGR_Data1D *datfalse) {
00037 #ifdef __SSE2__
00038   int j;
00039   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00040   fgUINT8 *rowptr_true, *rowptr_false;
00041 
00042   __m128i ptst1, ptst2, ptrue, pfalse;
00043   __m128i mask,one;
00044 
00045   
00046 
00047   one = _mm_set1_epi8((char) 0xFF);
00048 
00049   rowptr_out = (fgUINT8 *) datout->array;
00050   rowptr_t1 = (fgUINT8 *) dattest1->array;
00051   rowptr_t2 = (fgUINT8 *) dattest2->array;
00052   rowptr_true = (fgUINT8 *) dattrue->array;
00053   rowptr_false = (fgUINT8 *) datfalse->array;
00054 
00055 
00056   for(j=0 ; j<datout->length ; j+=16) {
00057     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00058     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00059     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00060     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00061 
00062     ptst1 = _mm_max_epu8(ptst1,ptst2);
00063     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00064     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00065     //_mm_mfence();
00066     mask = _mm_andnot_si128(mask,one);
00067     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00068 
00069     rowptr_false+=16;
00070     rowptr_true+=16;
00071     rowptr_t2+=16;
00072     rowptr_t1+=16;
00073     rowptr_out+=16;
00074   }
00075 #endif
00076 }
00077 void flgr1d_replace_I_EQ_I_I_I_fast_fgUINT8(FLGR_Data1D *datout,
00078                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00079                         FLGR_Data1D *dattrue, FLGR_Data1D *datfalse) {
00080 #ifdef __SSE2__
00081   int j;
00082   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00083   fgUINT8 *rowptr_true, *rowptr_false;
00084 
00085   __m128i ptst1, ptst2, ptrue, pfalse;
00086   __m128i mask,one;
00087 
00088   
00089 
00090   one = _mm_set1_epi8((char) 0xFF);
00091 
00092   rowptr_out = (fgUINT8 *) datout->array;
00093   rowptr_t1 = (fgUINT8 *) dattest1->array;
00094   rowptr_t2 = (fgUINT8 *) dattest2->array;
00095   rowptr_true = (fgUINT8 *) dattrue->array;
00096   rowptr_false = (fgUINT8 *) datfalse->array;
00097 
00098   for(j=0 ; j<datout->length ; j+=16) {
00099     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00100     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00101     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00102     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00103 
00104 
00105     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00106     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00107     //_mm_mfence();
00108     mask = _mm_andnot_si128(mask,one);
00109     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00110 
00111     rowptr_false+=16;
00112     rowptr_true+=16;
00113     rowptr_t2+=16;
00114     rowptr_t1+=16;
00115     rowptr_out+=16;
00116   }
00117 #endif
00118 }
00119 void flgr1d_replace_I_NE_I_I_I_fast_fgUINT8(FLGR_Data1D *datout,
00120                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00121                         FLGR_Data1D *dattrue, FLGR_Data1D *datfalse) {
00122 #ifdef __SSE2__
00123   int j;
00124   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00125   fgUINT8 *rowptr_true, *rowptr_false;
00126 
00127   __m128i ptst1, ptst2, ptrue, pfalse;
00128   __m128i mask,one;
00129 
00130   
00131 
00132   one = _mm_set1_epi8((char) 0xFF);
00133 
00134   rowptr_out = (fgUINT8 *) datout->array;
00135   rowptr_t1 = (fgUINT8 *) dattest1->array;
00136   rowptr_t2 = (fgUINT8 *) dattest2->array;
00137   rowptr_true = (fgUINT8 *) dattrue->array;
00138   rowptr_false = (fgUINT8 *) datfalse->array;
00139 
00140 
00141   for(j=0 ; j<datout->length ; j+=16) {
00142     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00143     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00144     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00145     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00146 
00147     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00148     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00149     //_mm_mfence();
00150     mask = _mm_andnot_si128(mask,one);
00151     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00152 
00153 
00154     rowptr_false+=16;
00155     rowptr_true+=16;
00156     rowptr_t2+=16;
00157     rowptr_t1+=16;
00158     rowptr_out+=16;
00159   }
00160 #endif
00161 }
00162 void flgr1d_replace_I_GE_I_I_I_fast_fgUINT8(FLGR_Data1D *datout,
00163                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00164                         FLGR_Data1D *dattrue, FLGR_Data1D *datfalse) {
00165 #ifdef __SSE2__
00166   int j;
00167   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00168   fgUINT8 *rowptr_true, *rowptr_false;
00169 
00170   __m128i ptst1, ptst2, ptrue, pfalse;
00171   __m128i mask,one;
00172 
00173   
00174 
00175   one = _mm_set1_epi8((char) 0xFF);
00176 
00177   rowptr_out = (fgUINT8 *) datout->array;
00178   rowptr_t1 = (fgUINT8 *) dattest1->array;
00179   rowptr_t2 = (fgUINT8 *) dattest2->array;
00180   rowptr_true = (fgUINT8 *) dattrue->array;
00181   rowptr_false = (fgUINT8 *) datfalse->array;
00182 
00183 
00184   for(j=0 ; j<datout->length ; j+=16) {
00185     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00186     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00187     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00188     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00189 
00190     ptst1 = _mm_min_epu8(ptst1,ptst2);
00191     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00192     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00193     //_mm_mfence();
00194     mask = _mm_andnot_si128(mask,one);
00195     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00196 
00197 
00198     rowptr_false+=16;
00199     rowptr_true+=16;
00200     rowptr_t2+=16;
00201     rowptr_t1+=16;
00202     rowptr_out+=16;
00203   }
00204 #endif
00205 }
00206 
00207 
00208 
00209 
00210 
00211 
00212 
00213 
00214 void flgr1d_replace_I_LE_I_I_C_fast_fgUINT8(FLGR_Data1D *datout,
00215                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00216                         FLGR_Data1D *dattrue, fgUINT8 vfalse) {
00217 #ifdef __SSE2__
00218   int j;
00219   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00220   fgUINT8 *rowptr_true;
00221 
00222   __m128i ptst1, ptst2, ptrue, pfalse;
00223   __m128i mask,one;
00224 
00225   
00226 
00227   one = _mm_set1_epi8((char) 0xFF);
00228   pfalse = _mm_set1_epi8((char) vfalse);
00229 
00230   rowptr_out = (fgUINT8 *) datout->array;
00231   rowptr_t1 = (fgUINT8 *) dattest1->array;
00232   rowptr_t2 = (fgUINT8 *) dattest2->array;
00233   rowptr_true = (fgUINT8 *) dattrue->array;
00234 
00235 
00236   for(j=0 ; j<datout->length ; j+=16) {
00237     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00238     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00239     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00240 
00241     ptst1 = _mm_max_epu8(ptst1,ptst2);
00242     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00243     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00244     //_mm_mfence();
00245     mask = _mm_andnot_si128(mask,one);
00246     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00247 
00248 
00249     rowptr_true+=16;
00250     rowptr_t2+=16;
00251     rowptr_t1+=16;
00252     rowptr_out+=16;
00253   }
00254 #endif
00255 }
00256 void flgr1d_replace_I_EQ_I_I_C_fast_fgUINT8(FLGR_Data1D *datout,
00257                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00258                         FLGR_Data1D *dattrue, fgUINT8 vfalse) {
00259 #ifdef __SSE2__
00260   int j;
00261   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00262   fgUINT8 *rowptr_true;
00263 
00264   __m128i ptst1, ptst2, ptrue, pfalse;
00265   __m128i mask,one;
00266 
00267   
00268 
00269   one = _mm_set1_epi8((char) 0xFF);
00270   pfalse = _mm_set1_epi8((char) vfalse);
00271 
00272   rowptr_out = (fgUINT8 *) datout->array;
00273   rowptr_t1 = (fgUINT8 *) dattest1->array;
00274   rowptr_t2 = (fgUINT8 *) dattest2->array;
00275   rowptr_true = (fgUINT8 *) dattrue->array;
00276 
00277 
00278   for(j=0 ; j<datout->length ; j+=16) {
00279     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00280     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00281     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00282 
00283     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00284     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00285     //_mm_mfence();
00286     mask = _mm_andnot_si128(mask,one);
00287     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00288 
00289 
00290     rowptr_true+=16;
00291     rowptr_t2+=16;
00292     rowptr_t1+=16;
00293     rowptr_out+=16;
00294   }
00295 #endif
00296 }
00297 void flgr1d_replace_I_NE_I_I_C_fast_fgUINT8(FLGR_Data1D *datout,
00298                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00299                         FLGR_Data1D *dattrue, fgUINT8 vfalse) {
00300 #ifdef __SSE2__
00301   int j;
00302   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00303   fgUINT8 *rowptr_true;
00304 
00305   __m128i ptst1, ptst2, ptrue, pfalse;
00306   __m128i mask,one;
00307 
00308   
00309 
00310   one = _mm_set1_epi8((char) 0xFF);
00311   pfalse = _mm_set1_epi8((char) vfalse);
00312 
00313   rowptr_out = (fgUINT8 *) datout->array;
00314   rowptr_t1 = (fgUINT8 *) dattest1->array;
00315   rowptr_t2 = (fgUINT8 *) dattest2->array;
00316   rowptr_true = (fgUINT8 *) dattrue->array;
00317 
00318 
00319   for(j=0 ; j<datout->length ; j+=16) {
00320     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00321     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00322     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00323 
00324     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00325     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00326     //_mm_mfence();
00327     mask = _mm_andnot_si128(mask,one);
00328     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00329 
00330 
00331     rowptr_true+=16;
00332     rowptr_t2+=16;
00333     rowptr_t1+=16;
00334     rowptr_out+=16;
00335   }
00336 #endif
00337 }
00338 void flgr1d_replace_I_GE_I_I_C_fast_fgUINT8(FLGR_Data1D *datout,
00339                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00340                         FLGR_Data1D *dattrue, fgUINT8 vfalse) {
00341 #ifdef __SSE2__
00342   int j;
00343   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00344   fgUINT8 *rowptr_true;
00345 
00346   __m128i ptst1, ptst2, ptrue, pfalse;
00347   __m128i mask,one;
00348 
00349   
00350 
00351   one = _mm_set1_epi8((char) 0xFF);
00352   pfalse = _mm_set1_epi8((char) vfalse);
00353 
00354   rowptr_out = (fgUINT8 *) datout->array;
00355   rowptr_t1 = (fgUINT8 *) dattest1->array;
00356   rowptr_t2 = (fgUINT8 *) dattest2->array;
00357   rowptr_true = (fgUINT8 *) dattrue->array;
00358 
00359 
00360   for(j=0 ; j<datout->length ; j+=16) {
00361     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00362     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00363     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00364 
00365     ptst1 = _mm_min_epu8(ptst1,ptst2);
00366     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00367     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00368     //_mm_mfence();
00369     mask = _mm_andnot_si128(mask,one);
00370     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00371 
00372 
00373     rowptr_true+=16;
00374     rowptr_t2+=16;
00375     rowptr_t1+=16;
00376     rowptr_out+=16;
00377   }
00378 #endif
00379 }
00380 
00381 
00382 
00383 
00384 
00385 
00386 
00387 
00388 
00389 
00390 void flgr1d_replace_I_LE_I_C_I_fast_fgUINT8(FLGR_Data1D *datout,
00391                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00392                         fgUINT8 vtrue, FLGR_Data1D *datfalse) {
00393 #ifdef __SSE2__
00394   int j;
00395   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00396   fgUINT8 *rowptr_false;
00397 
00398   __m128i ptst1, ptst2, ptrue, pfalse;
00399   __m128i mask,one;
00400 
00401   
00402 
00403   one = _mm_set1_epi8((char) 0xFF);
00404   ptrue = _mm_set1_epi8((char) vtrue);
00405 
00406   rowptr_out = (fgUINT8 *) datout->array;
00407   rowptr_t1 = (fgUINT8 *) dattest1->array;
00408   rowptr_t2 = (fgUINT8 *) dattest2->array;
00409   rowptr_false = (fgUINT8 *) datfalse->array;
00410 
00411 
00412   for(j=0 ; j<datout->length ; j+=16) {
00413     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00414     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00415     pfalse  = _mm_load_si128((__m128i*) rowptr_false);
00416 
00417     ptst1 = _mm_max_epu8(ptst1,ptst2);
00418     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00419     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00420     //_mm_mfence();
00421     mask = _mm_andnot_si128(mask,one);
00422     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00423 
00424 
00425     rowptr_false+=16;
00426     rowptr_t2+=16;
00427     rowptr_t1+=16;
00428     rowptr_out+=16;
00429   }
00430 #endif
00431 }
00432 void flgr1d_replace_I_EQ_I_C_I_fast_fgUINT8(FLGR_Data1D *datout,
00433                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00434                         fgUINT8 vtrue, FLGR_Data1D *datfalse) {
00435 #ifdef __SSE2__
00436   int j;
00437   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00438   fgUINT8 *rowptr_false;
00439 
00440   __m128i ptst1, ptst2, ptrue, pfalse;
00441   __m128i mask,one;
00442 
00443   
00444 
00445   one = _mm_set1_epi8((char) 0xFF);
00446   ptrue = _mm_set1_epi8((char) vtrue);
00447 
00448   rowptr_out = (fgUINT8 *) datout->array;
00449   rowptr_t1 = (fgUINT8 *) dattest1->array;
00450   rowptr_t2 = (fgUINT8 *) dattest2->array;
00451   rowptr_false = (fgUINT8 *) datfalse->array;
00452 
00453 
00454   for(j=0 ; j<datout->length ; j+=16) {
00455     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00456     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00457     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00458 
00459     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00460     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00461     //_mm_mfence();
00462     mask = _mm_andnot_si128(mask,one);
00463     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00464 
00465 
00466     rowptr_false+=16;
00467     rowptr_t2+=16;
00468     rowptr_t1+=16;
00469     rowptr_out+=16;
00470   }
00471 #endif
00472 }
00473 void flgr1d_replace_I_NE_I_C_I_fast_fgUINT8(FLGR_Data1D *datout,
00474                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00475                         fgUINT8 vtrue, FLGR_Data1D *datfalse) {
00476 #ifdef __SSE2__
00477   int j;
00478   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00479   fgUINT8 *rowptr_false;
00480 
00481   __m128i ptst1, ptst2, ptrue, pfalse;
00482   __m128i mask,one;
00483 
00484   
00485 
00486   one = _mm_set1_epi8((char) 0xFF);
00487   ptrue = _mm_set1_epi8((char) vtrue);
00488 
00489   rowptr_out = (fgUINT8 *) datout->array;
00490   rowptr_t1 = (fgUINT8 *) dattest1->array;
00491   rowptr_t2 = (fgUINT8 *) dattest2->array;
00492   rowptr_false = (fgUINT8 *) datfalse->array;
00493 
00494 
00495   for(j=0 ; j<datout->length ; j+=16) {
00496     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00497     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00498     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00499 
00500     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00501     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00502     //_mm_mfence();
00503     mask = _mm_andnot_si128(mask,one);
00504     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00505 
00506 
00507     rowptr_false+=16;
00508     rowptr_t2+=16;
00509     rowptr_t1+=16;
00510     rowptr_out+=16;
00511   }
00512 #endif
00513 }
00514 void flgr1d_replace_I_GE_I_C_I_fast_fgUINT8(FLGR_Data1D *datout,
00515                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00516                         fgUINT8 vtrue, FLGR_Data1D *datfalse) {
00517 #ifdef __SSE2__
00518   int j;
00519   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00520   fgUINT8 *rowptr_false;
00521 
00522   __m128i ptst1, ptst2, ptrue, pfalse;
00523   __m128i mask,one;
00524 
00525   
00526 
00527   one = _mm_set1_epi8((char) 0xFF);
00528   ptrue = _mm_set1_epi8((char) vtrue);
00529 
00530   rowptr_out = (fgUINT8 *) datout->array;
00531   rowptr_t1 = (fgUINT8 *) dattest1->array;
00532   rowptr_t2 = (fgUINT8 *) dattest2->array;
00533   rowptr_false = (fgUINT8 *) datfalse->array;
00534 
00535 
00536   for(j=0 ; j<datout->length ; j+=16) {
00537     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00538     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00539     pfalse  = _mm_load_si128((__m128i*) rowptr_false);
00540 
00541     ptst1 = _mm_min_epu8(ptst1,ptst2);
00542     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00543     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00544     //_mm_mfence();
00545     mask = _mm_andnot_si128(mask,one);
00546     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00547 
00548 
00549     rowptr_false+=16;
00550     rowptr_t2+=16;
00551     rowptr_t1+=16;
00552     rowptr_out+=16;
00553   }
00554 #endif
00555 }
00556 
00557 
00558 
00559 
00560 
00561 
00562 
00563 
00564 
00565 void flgr1d_replace_I_LE_I_C_C_fast_fgUINT8(FLGR_Data1D *datout,
00566                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00567                         fgUINT8 vtrue, fgUINT8 vfalse) {
00568 #ifdef __SSE2__
00569   int j;
00570   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00571 
00572   __m128i ptst1, ptst2, ptrue, pfalse;
00573   __m128i mask,one;
00574 
00575   
00576 
00577   one = _mm_set1_epi8((char) 0xFF);
00578   ptrue = _mm_set1_epi8((char) vtrue);
00579   pfalse = _mm_set1_epi8((char) vfalse);
00580 
00581   rowptr_out = (fgUINT8 *) datout->array;
00582   rowptr_t1 = (fgUINT8 *) dattest1->array;
00583   rowptr_t2 = (fgUINT8 *) dattest2->array;
00584 
00585 
00586   for(j=0 ; j<datout->length ; j+=16) {
00587     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00588     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00589 
00590     ptst1 = _mm_max_epu8(ptst1,ptst2);
00591     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00592     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00593     //_mm_mfence();
00594     mask = _mm_andnot_si128(mask,one);
00595     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00596 
00597 
00598     rowptr_t2+=16;
00599     rowptr_t1+=16;
00600     rowptr_out+=16;
00601   }
00602 #endif
00603 }
00604 void flgr1d_replace_I_EQ_I_C_C_fast_fgUINT8(FLGR_Data1D *datout,
00605                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00606                         fgUINT8 vtrue, fgUINT8 vfalse) {
00607 #ifdef __SSE2__
00608   int j;
00609   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00610 
00611   __m128i ptst1, ptst2, ptrue, pfalse;
00612   __m128i mask,one;
00613 
00614   
00615 
00616   one = _mm_set1_epi8((char) 0xFF);
00617   ptrue = _mm_set1_epi8((char) vtrue);
00618   pfalse = _mm_set1_epi8((char) vfalse);
00619 
00620   rowptr_out = (fgUINT8 *) datout->array;
00621   rowptr_t1 = (fgUINT8 *) dattest1->array;
00622   rowptr_t2 = (fgUINT8 *) dattest2->array;
00623 
00624 
00625   for(j=0 ; j<datout->length ; j+=16) {
00626     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00627     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00628 
00629     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00630     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00631     //_mm_mfence();
00632     mask = _mm_andnot_si128(mask,one);
00633     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00634 
00635 
00636     rowptr_t2+=16;
00637     rowptr_t1+=16;
00638     rowptr_out+=16;
00639   }
00640 #endif
00641 }
00642 void flgr1d_replace_I_NE_I_C_C_fast_fgUINT8(FLGR_Data1D *datout,
00643                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00644                         fgUINT8 vtrue, fgUINT8 vfalse) {
00645 #ifdef __SSE2__
00646   int j;
00647   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00648 
00649   __m128i ptst1, ptst2, ptrue, pfalse;
00650   __m128i mask,one;
00651 
00652   
00653 
00654   one = _mm_set1_epi8((char) 0xFF);
00655   ptrue = _mm_set1_epi8((char) vtrue);
00656   pfalse = _mm_set1_epi8((char) vfalse);
00657 
00658   rowptr_out = (fgUINT8 *) datout->array;
00659   rowptr_t1 = (fgUINT8 *) dattest1->array;
00660   rowptr_t2 = (fgUINT8 *) dattest2->array;
00661 
00662 
00663   for(j=0 ; j<datout->length ; j+=16) {
00664     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00665     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00666 
00667     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00668     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00669     //_mm_mfence();
00670     mask = _mm_andnot_si128(mask,one);
00671     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00672 
00673 
00674     rowptr_t2+=16;
00675     rowptr_t1+=16;
00676     rowptr_out+=16;
00677   }
00678 #endif
00679 }
00680 void flgr1d_replace_I_GE_I_C_C_fast_fgUINT8(FLGR_Data1D *datout,
00681                         FLGR_Data1D *dattest1, FLGR_Data1D *dattest2,
00682                         fgUINT8 vtrue, fgUINT8 vfalse) {
00683 #ifdef __SSE2__
00684   int j;
00685   fgUINT8 *rowptr_out,*rowptr_t1,*rowptr_t2;
00686 
00687   __m128i ptst1, ptst2, ptrue, pfalse;
00688   __m128i mask,one;
00689 
00690   
00691 
00692   one = _mm_set1_epi8((char) 0xFF);
00693   ptrue = _mm_set1_epi8((char) vtrue);
00694   pfalse = _mm_set1_epi8((char) vfalse);
00695 
00696   rowptr_out = (fgUINT8 *) datout->array;
00697   rowptr_t1 = (fgUINT8 *) dattest1->array;
00698   rowptr_t2 = (fgUINT8 *) dattest2->array;
00699 
00700 
00701   for(j=0 ; j<datout->length ; j+=16) {
00702     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00703     ptst2  = _mm_load_si128((__m128i*) rowptr_t2);
00704 
00705     ptst1 = _mm_min_epu8(ptst1,ptst2);
00706     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00707     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00708     //_mm_mfence();
00709     mask = _mm_andnot_si128(mask,one);
00710     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00711 
00712 
00713     rowptr_t2+=16;
00714     rowptr_t1+=16;
00715     rowptr_out+=16;
00716   }
00717 #endif
00718 }
00719 
00720 
00721 
00722 
00723 
00724 
00725 
00726 
00727 
00728 
00729 
00730 void flgr1d_replace_I_LE_C_I_I_fast_fgUINT8(FLGR_Data1D *datout,
00731                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
00732                         FLGR_Data1D *dattrue, FLGR_Data1D *datfalse) {
00733 #ifdef __SSE2__
00734   int j;
00735   fgUINT8 *rowptr_out,*rowptr_t1;
00736   fgUINT8 *rowptr_true, *rowptr_false;
00737 
00738   __m128i ptst1, ptst2, ptrue, pfalse;
00739   __m128i mask,one;
00740 
00741   
00742 
00743   one = _mm_set1_epi8((char) 0xFF);
00744   ptst2 = _mm_set1_epi8((char) vtest2);
00745 
00746   rowptr_out = (fgUINT8 *) datout->array;
00747   rowptr_t1 = (fgUINT8 *) dattest1->array;
00748   rowptr_true = (fgUINT8 *) dattrue->array;
00749   rowptr_false = (fgUINT8 *) datfalse->array;
00750 
00751 
00752   for(j=0 ; j<datout->length ; j+=16) {
00753     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00754     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00755     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00756 
00757     ptst1 = _mm_max_epu8(ptst1,ptst2);
00758     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00759     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00760     //_mm_mfence();
00761     mask = _mm_andnot_si128(mask,one);
00762     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00763 
00764 
00765     rowptr_false+=16;
00766     rowptr_true+=16;
00767     rowptr_t1+=16;
00768     rowptr_out+=16;
00769   }
00770 #endif
00771 }
00772 void flgr1d_replace_I_EQ_C_I_I_fast_fgUINT8(FLGR_Data1D *datout,
00773                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
00774                         FLGR_Data1D *dattrue, FLGR_Data1D *datfalse) {
00775 #ifdef __SSE2__
00776   int j;
00777   fgUINT8 *rowptr_out,*rowptr_t1;
00778   fgUINT8 *rowptr_true, *rowptr_false;
00779 
00780   __m128i ptst1, ptst2, ptrue, pfalse;
00781   __m128i mask,one;
00782 
00783   
00784 
00785   one = _mm_set1_epi8((char) 0xFF);
00786   ptst2 = _mm_set1_epi8((char) vtest2);
00787 
00788   rowptr_out = (fgUINT8 *) datout->array;
00789   rowptr_t1 = (fgUINT8 *) dattest1->array;
00790   rowptr_true = (fgUINT8 *) dattrue->array;
00791   rowptr_false = (fgUINT8 *) datfalse->array;
00792 
00793 
00794   for(j=0 ; j<datout->length ; j+=16) {
00795     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00796     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00797     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00798 
00799     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00800     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00801     //_mm_mfence();
00802     mask = _mm_andnot_si128(mask,one);
00803     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00804 
00805 
00806     rowptr_false+=16;
00807     rowptr_true+=16;
00808     rowptr_t1+=16;
00809     rowptr_out+=16;
00810   }
00811 #endif
00812 }
00813 void flgr1d_replace_I_NE_C_I_I_fast_fgUINT8(FLGR_Data1D *datout,
00814                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
00815                         FLGR_Data1D *dattrue, FLGR_Data1D *datfalse) {
00816 #ifdef __SSE2__
00817   int j;
00818   fgUINT8 *rowptr_out,*rowptr_t1;
00819   fgUINT8 *rowptr_true, *rowptr_false;
00820 
00821   __m128i ptst1, ptst2, ptrue, pfalse;
00822   __m128i mask,one;
00823 
00824   
00825 
00826   one = _mm_set1_epi8((char) 0xFF);
00827   ptst2 = _mm_set1_epi8((char) vtest2);
00828 
00829   rowptr_out = (fgUINT8 *) datout->array;
00830   rowptr_t1 = (fgUINT8 *) dattest1->array;
00831   rowptr_true = (fgUINT8 *) dattrue->array;
00832   rowptr_false = (fgUINT8 *) datfalse->array;
00833 
00834 
00835   for(j=0 ; j<datout->length ; j+=16) {
00836     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00837     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00838     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00839 
00840     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00841     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00842     //_mm_mfence();
00843     mask = _mm_andnot_si128(mask,one);
00844     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00845 
00846 
00847     rowptr_false+=16;
00848     rowptr_true+=16;
00849     rowptr_t1+=16;
00850     rowptr_out+=16;
00851   }
00852 #endif
00853 }
00854 void flgr1d_replace_I_GE_C_I_I_fast_fgUINT8(FLGR_Data1D *datout,
00855                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
00856                         FLGR_Data1D *dattrue, FLGR_Data1D *datfalse) {
00857 #ifdef __SSE2__
00858   int j;
00859   fgUINT8 *rowptr_out,*rowptr_t1;
00860   fgUINT8 *rowptr_true, *rowptr_false;
00861 
00862   __m128i ptst1, ptst2, ptrue, pfalse;
00863   __m128i mask,one;
00864 
00865   
00866 
00867   one = _mm_set1_epi8((char) 0xFF);
00868   ptst2 = _mm_set1_epi8((char) vtest2);
00869 
00870   rowptr_out = (fgUINT8 *) datout->array;
00871   rowptr_t1 = (fgUINT8 *) dattest1->array;
00872   rowptr_true = (fgUINT8 *) dattrue->array;
00873   rowptr_false = (fgUINT8 *) datfalse->array;
00874 
00875 
00876   for(j=0 ; j<datout->length ; j+=16) {
00877     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00878     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00879     pfalse = _mm_load_si128((__m128i*) rowptr_false);
00880 
00881     ptst1 = _mm_min_epu8(ptst1,ptst2);
00882     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00883     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00884     //_mm_mfence();
00885     mask = _mm_andnot_si128(mask,one);
00886     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00887 
00888 
00889     rowptr_false+=16;
00890     rowptr_true+=16;
00891     rowptr_t1+=16;
00892     rowptr_out+=16;
00893   }
00894 #endif
00895 }
00896 
00897 
00898 
00899 
00900 
00901 
00902 
00903 
00904 void flgr1d_replace_I_LE_C_I_C_fast_fgUINT8(FLGR_Data1D *datout,
00905                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
00906                         FLGR_Data1D *dattrue, fgUINT8 vfalse) {
00907 #ifdef __SSE2__
00908   int j;
00909   fgUINT8 *rowptr_out,*rowptr_t1;
00910   fgUINT8 *rowptr_true;
00911 
00912   __m128i ptst1, ptst2, ptrue, pfalse;
00913   __m128i mask,one;
00914 
00915   
00916 
00917   one = _mm_set1_epi8((char) 0xFF);
00918 
00919   ptst2 = _mm_set1_epi8((char) vtest2);
00920   pfalse = _mm_set1_epi8((char) vfalse);
00921 
00922   rowptr_out = (fgUINT8 *) datout->array;
00923   rowptr_t1 = (fgUINT8 *) dattest1->array;
00924   rowptr_true = (fgUINT8 *) dattrue->array;
00925 
00926 
00927   for(j=0 ; j<datout->length ; j+=16) {
00928     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00929     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00930 
00931     ptst1 = _mm_max_epu8(ptst1,ptst2);
00932     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00933     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00934     //_mm_mfence();
00935     mask = _mm_andnot_si128(mask,one);
00936     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00937 
00938 
00939     rowptr_true+=16;
00940     rowptr_t1+=16;
00941     rowptr_out+=16;
00942   }
00943 #endif
00944 }
00945 void flgr1d_replace_I_EQ_C_I_C_fast_fgUINT8(FLGR_Data1D *datout,
00946                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
00947                         FLGR_Data1D *dattrue, fgUINT8 vfalse) {
00948 #ifdef __SSE2__
00949   int j;
00950   fgUINT8 *rowptr_out,*rowptr_t1;
00951   fgUINT8 *rowptr_true;
00952 
00953   __m128i ptst1, ptst2, ptrue, pfalse;
00954   __m128i mask,one;
00955 
00956   
00957 
00958   one = _mm_set1_epi8((char) 0xFF);
00959 
00960   ptst2 = _mm_set1_epi8((char) vtest2);
00961   pfalse = _mm_set1_epi8((char) vfalse);
00962 
00963   rowptr_out = (fgUINT8 *) datout->array;
00964   rowptr_t1 = (fgUINT8 *) dattest1->array;
00965   rowptr_true = (fgUINT8 *) dattrue->array;
00966 
00967 
00968   for(j=0 ; j<datout->length ; j+=16) {
00969     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
00970     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
00971 
00972     mask = _mm_cmpeq_epi8(ptst1,ptst2);
00973     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
00974     //_mm_mfence();
00975     mask = _mm_andnot_si128(mask,one);
00976     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
00977 
00978 
00979     rowptr_true+=16;
00980     rowptr_t1+=16;
00981     rowptr_out+=16;
00982   }
00983 #endif
00984 }
00985 void flgr1d_replace_I_NE_C_I_C_fast_fgUINT8(FLGR_Data1D *datout,
00986                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
00987                         FLGR_Data1D *dattrue, fgUINT8 vfalse) {
00988 #ifdef __SSE2__
00989   int j;
00990   fgUINT8 *rowptr_out,*rowptr_t1;
00991   fgUINT8 *rowptr_true;
00992 
00993   __m128i ptst1, ptst2, ptrue, pfalse;
00994   __m128i mask,one;
00995 
00996   
00997 
00998   one = _mm_set1_epi8((char) 0xFF);
00999 
01000   ptst2 = _mm_set1_epi8((char) vtest2);
01001   pfalse = _mm_set1_epi8((char) vfalse);
01002 
01003   rowptr_out = (fgUINT8 *) datout->array;
01004   rowptr_t1 = (fgUINT8 *) dattest1->array;
01005   rowptr_true = (fgUINT8 *) dattrue->array;
01006 
01007 
01008   for(j=0 ; j<datout->length ; j+=16) {
01009     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01010     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
01011 
01012     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01013     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01014     //_mm_mfence();
01015     mask = _mm_andnot_si128(mask,one);
01016     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01017 
01018 
01019     rowptr_true+=16;
01020     rowptr_t1+=16;
01021     rowptr_out+=16;
01022   }
01023 #endif
01024 }
01025 void flgr1d_replace_I_GE_C_I_C_fast_fgUINT8(FLGR_Data1D *datout,
01026                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01027                         FLGR_Data1D *dattrue, fgUINT8 vfalse) {
01028 #ifdef __SSE2__
01029   int j;
01030   fgUINT8 *rowptr_out,*rowptr_t1;
01031   fgUINT8 *rowptr_true;
01032 
01033   __m128i ptst1, ptst2, ptrue, pfalse;
01034   __m128i mask,one;
01035 
01036   
01037 
01038   one = _mm_set1_epi8((char) 0xFF);
01039 
01040   ptst2 = _mm_set1_epi8((char) vtest2);
01041   pfalse = _mm_set1_epi8((char) vfalse);
01042 
01043   rowptr_out = (fgUINT8 *) datout->array;
01044   rowptr_t1 = (fgUINT8 *) dattest1->array;
01045   rowptr_true = (fgUINT8 *) dattrue->array;
01046 
01047 
01048   for(j=0 ; j<datout->length ; j+=16) {
01049     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01050     ptrue  = _mm_load_si128((__m128i*) rowptr_true);
01051 
01052     ptst1 = _mm_min_epu8(ptst1,ptst2);
01053     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01054     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01055     //_mm_mfence();
01056     mask = _mm_andnot_si128(mask,one);
01057     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01058 
01059 
01060     rowptr_true+=16;
01061     rowptr_t1+=16;
01062     rowptr_out+=16;
01063   }
01064 #endif
01065 }
01066 
01067 
01068 
01069 
01070 
01071 
01072 
01073 
01074 
01075 
01076 void flgr1d_replace_I_LE_C_C_I_fast_fgUINT8(FLGR_Data1D *datout,
01077                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01078                         fgUINT8 vtrue, FLGR_Data1D *datfalse) {
01079 #ifdef __SSE2__
01080   int j;
01081   fgUINT8 *rowptr_out,*rowptr_t1;
01082   fgUINT8 *rowptr_false;
01083 
01084   __m128i ptst1, ptst2, ptrue, pfalse;
01085   __m128i mask,one;
01086 
01087   
01088 
01089   one = _mm_set1_epi8((char) 0xFF);
01090 
01091   ptst2 = _mm_set1_epi8((char) vtest2);
01092   ptrue = _mm_set1_epi8((char) vtrue);
01093 
01094   rowptr_out = (fgUINT8 *) datout->array;
01095   rowptr_t1 = (fgUINT8 *) dattest1->array;
01096   rowptr_false = (fgUINT8 *) datfalse->array;
01097 
01098 
01099   for(j=0 ; j<datout->length ; j+=16) {
01100     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01101     pfalse  = _mm_load_si128((__m128i*) rowptr_false);
01102 
01103     ptst1 = _mm_max_epu8(ptst1,ptst2);
01104     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01105     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01106     //_mm_mfence();
01107     mask = _mm_andnot_si128(mask,one);
01108     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01109 
01110 
01111     rowptr_false+=16;
01112     rowptr_t1+=16;
01113     rowptr_out+=16;
01114   }
01115 #endif
01116 }
01117 void flgr1d_replace_I_EQ_C_C_I_fast_fgUINT8(FLGR_Data1D *datout,
01118                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01119                         fgUINT8 vtrue, FLGR_Data1D *datfalse) {
01120 #ifdef __SSE2__
01121   int j;
01122   fgUINT8 *rowptr_out,*rowptr_t1;
01123   fgUINT8 *rowptr_false;
01124 
01125   __m128i ptst1, ptst2, ptrue, pfalse;
01126   __m128i mask,one;
01127 
01128   
01129 
01130   one = _mm_set1_epi8((char) 0xFF);
01131 
01132   ptst2 = _mm_set1_epi8((char) vtest2);
01133   ptrue = _mm_set1_epi8((char) vtrue);
01134 
01135   rowptr_out = (fgUINT8 *) datout->array;
01136   rowptr_t1 = (fgUINT8 *) dattest1->array;
01137   rowptr_false = (fgUINT8 *) datfalse->array;
01138 
01139 
01140   for(j=0 ; j<datout->length ; j+=16) {
01141     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01142     pfalse = _mm_load_si128((__m128i*) rowptr_false);
01143 
01144     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01145     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01146     //_mm_mfence();
01147     mask = _mm_andnot_si128(mask,one);
01148     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01149 
01150 
01151     rowptr_false+=16;
01152     rowptr_t1+=16;
01153     rowptr_out+=16;
01154   }
01155 #endif
01156 }
01157 void flgr1d_replace_I_NE_C_C_I_fast_fgUINT8(FLGR_Data1D *datout,
01158                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01159                         fgUINT8 vtrue, FLGR_Data1D *datfalse) {
01160 #ifdef __SSE2__
01161   int j;
01162   fgUINT8 *rowptr_out,*rowptr_t1;
01163   fgUINT8 *rowptr_false;
01164 
01165   __m128i ptst1, ptst2, ptrue, pfalse;
01166   __m128i mask,one;
01167 
01168   
01169 
01170   one = _mm_set1_epi8((char) 0xFF);
01171 
01172   ptst2 = _mm_set1_epi8((char) vtest2);
01173   ptrue = _mm_set1_epi8((char) vtrue);
01174 
01175   rowptr_out = (fgUINT8 *) datout->array;
01176   rowptr_t1 = (fgUINT8 *) dattest1->array;
01177   rowptr_false = (fgUINT8 *) datfalse->array;
01178 
01179 
01180   for(j=0 ; j<datout->length ; j+=16) {
01181     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01182     pfalse = _mm_load_si128((__m128i*) rowptr_false);
01183 
01184     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01185     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01186     //_mm_mfence();
01187     mask = _mm_andnot_si128(mask,one);
01188     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01189 
01190 
01191     rowptr_false+=16;
01192     rowptr_t1+=16;
01193     rowptr_out+=16;
01194   }
01195 #endif
01196 }
01197 void flgr1d_replace_I_GE_C_C_I_fast_fgUINT8(FLGR_Data1D *datout,
01198                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01199                         fgUINT8 vtrue, FLGR_Data1D *datfalse) {
01200 #ifdef __SSE2__
01201   int j;
01202   fgUINT8 *rowptr_out,*rowptr_t1;
01203   fgUINT8 *rowptr_false;
01204 
01205   __m128i ptst1, ptst2, ptrue, pfalse;
01206   __m128i mask,one;
01207 
01208   
01209 
01210   one = _mm_set1_epi8((char) 0xFF);
01211 
01212   ptst2 = _mm_set1_epi8((char) vtest2);
01213   ptrue = _mm_set1_epi8((char) vtrue);
01214 
01215   rowptr_out = (fgUINT8 *) datout->array;
01216   rowptr_t1 = (fgUINT8 *) dattest1->array;
01217   rowptr_false = (fgUINT8 *) datfalse->array;
01218 
01219 
01220   for(j=0 ; j<datout->length ; j+=16) {
01221     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01222     pfalse  = _mm_load_si128((__m128i*) rowptr_false);
01223 
01224     ptst1 = _mm_min_epu8(ptst1,ptst2);
01225     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01226     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01227     //_mm_mfence();
01228     mask = _mm_andnot_si128(mask,one);
01229     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01230 
01231 
01232     rowptr_false+=16;
01233     rowptr_t1+=16;
01234     rowptr_out+=16;
01235   }
01236 #endif
01237 }
01238 
01239 
01240 
01241 
01242 
01243 
01244 
01245 
01246 
01247 void flgr1d_replace_I_LE_C_C_C_fast_fgUINT8(FLGR_Data1D *datout,
01248                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01249                         fgUINT8 vtrue, fgUINT8 vfalse) {
01250 #ifdef __SSE2__
01251   int j;
01252   fgUINT8 *rowptr_out,*rowptr_t1;
01253 
01254   __m128i ptst1, ptst2, ptrue, pfalse;
01255   __m128i mask,one;
01256 
01257   
01258 
01259   one = _mm_set1_epi8((char) 0xFF);
01260 
01261   ptst2 = _mm_set1_epi8((char) vtest2);
01262   ptrue = _mm_set1_epi8((char) vtrue);
01263   pfalse = _mm_set1_epi8((char) vfalse);
01264 
01265   rowptr_out = (fgUINT8 *) datout->array;
01266   rowptr_t1 = (fgUINT8 *) dattest1->array;
01267 
01268 
01269   for(j=0 ; j<datout->length ; j+=16) {
01270     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01271 
01272     ptst1 = _mm_max_epu8(ptst1,ptst2);
01273     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01274     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01275     //_mm_mfence();
01276     mask = _mm_andnot_si128(mask,one);
01277     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01278 
01279 
01280     rowptr_t1+=16;
01281     rowptr_out+=16;
01282   }
01283 #endif
01284 }
01285 void flgr1d_replace_I_EQ_C_C_C_fast_fgUINT8(FLGR_Data1D *datout,
01286                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01287                         fgUINT8 vtrue, fgUINT8 vfalse) {
01288 #ifdef __SSE2__
01289   int j;
01290   fgUINT8 *rowptr_out,*rowptr_t1;
01291 
01292   __m128i ptst1, ptst2, ptrue, pfalse;
01293   __m128i mask,one;
01294 
01295   
01296 
01297   one = _mm_set1_epi8((char) 0xFF);
01298 
01299   ptst2 = _mm_set1_epi8((char) vtest2);
01300   ptrue = _mm_set1_epi8((char) vtrue);
01301   pfalse = _mm_set1_epi8((char) vfalse);
01302 
01303   rowptr_out = (fgUINT8 *) datout->array;
01304   rowptr_t1 = (fgUINT8 *) dattest1->array;
01305 
01306 
01307   for(j=0 ; j<datout->length ; j+=16) {
01308     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01309 
01310     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01311     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01312     //_mm_mfence();
01313     mask = _mm_andnot_si128(mask,one);
01314     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01315 
01316 
01317     rowptr_t1+=16;
01318     rowptr_out+=16;
01319   }
01320 #endif
01321 }
01322 void flgr1d_replace_I_NE_C_C_C_fast_fgUINT8(FLGR_Data1D *datout,
01323                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01324                         fgUINT8 vtrue, fgUINT8 vfalse) {
01325 #ifdef __SSE2__
01326   int j;
01327   fgUINT8 *rowptr_out,*rowptr_t1;
01328 
01329   __m128i ptst1, ptst2, ptrue, pfalse;
01330   __m128i mask,one;
01331 
01332   
01333 
01334   one = _mm_set1_epi8((char) 0xFF);
01335 
01336   ptst2 = _mm_set1_epi8((char) vtest2);
01337   ptrue = _mm_set1_epi8((char) vtrue);
01338   pfalse = _mm_set1_epi8((char) vfalse);
01339 
01340   rowptr_out = (fgUINT8 *) datout->array;
01341   rowptr_t1 = (fgUINT8 *) dattest1->array;
01342 
01343 
01344   for(j=0 ; j<datout->length ; j+=16) {
01345     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01346 
01347     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01348     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01349     //_mm_mfence();
01350     mask = _mm_andnot_si128(mask,one);
01351     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01352 
01353 
01354     rowptr_t1+=16;
01355     rowptr_out+=16;
01356   }
01357 #endif
01358 }
01359 void flgr1d_replace_I_GE_C_C_C_fast_fgUINT8(FLGR_Data1D *datout,
01360                         FLGR_Data1D *dattest1,fgUINT8 vtest2,
01361                         fgUINT8 vtrue, fgUINT8 vfalse) {
01362 #ifdef __SSE2__
01363   int j;
01364   fgUINT8 *rowptr_out,*rowptr_t1;
01365 
01366   __m128i ptst1, ptst2, ptrue, pfalse;
01367   __m128i mask,one;
01368 
01369   
01370 
01371   one = _mm_set1_epi8((char) 0xFF);
01372 
01373   ptst2 = _mm_set1_epi8((char) vtest2);
01374   ptrue = _mm_set1_epi8((char) vtrue);
01375   pfalse = _mm_set1_epi8((char) vfalse);
01376 
01377   rowptr_out = (fgUINT8 *) datout->array;
01378   rowptr_t1 = (fgUINT8 *) dattest1->array;
01379 
01380 
01381   for(j=0 ; j<datout->length ; j+=16) {
01382     ptst1  = _mm_load_si128((__m128i*) rowptr_t1);
01383 
01384     ptst1 = _mm_min_epu8(ptst1,ptst2);
01385     mask = _mm_cmpeq_epi8(ptst1,ptst2);
01386     _mm_maskmoveu_si128(ptrue,mask, (char*) rowptr_out);
01387     //_mm_mfence();
01388     mask = _mm_andnot_si128(mask,one);
01389     _mm_maskmoveu_si128(pfalse,mask, (char*) rowptr_out);
01390 
01391 
01392     rowptr_t1+=16;
01393     rowptr_out+=16;
01394   }
01395 #endif
01396 }
01397 
01398 
01399 
01400 
01401 
01402 
01403 
01404 
01405 
01406 
01407