Fulguro: /home/christophe/src/fulguro/trunk/src/core/flgrCoreTransposeFastUtil.h 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 #ifdef __cplusplus
00022 extern "C" {
00023 #endif
00024 
00025 #ifndef __FLGR_CORE_TRANSFORM_FAST_UTIL_H
00026 #define __FLGR_CORE_TRANSFORM_FAST_UTIL_H
00027  
00028 
00029 /*
00030   Use SSE instructions
00031   Get Row r1,r2,r3,r4,r5,r6,r7,r8 and do transposition
00032   Transposed Matrix Row are ordered as follow :
00033   r8,r4,r6,r2,r7,r3,r5,r1
00034 */
00035 #define FLGR_MACRO_TRANSPOSE_BLOCK_8x8_BYTE(tmp, r1,r2,r3,r4,r5,r6,r7,r8) \
00036   /*FIRST STEP GROUP Byte To Word*/                 \
00037   tmp1 = _m_punpckhbw(r1,r2);                       \
00038   r2  = _m_punpcklbw(r1,r2);                        \
00039   r1 = tmp1;                                \
00040                                     \
00041   tmp1 = _m_punpckhbw(r3,r4);                       \
00042   r4  = _m_punpcklbw(r3,r4);                        \
00043   r3 = tmp1;                                \
00044                                     \
00045   tmp1 = _m_punpckhbw(r5,r6);                       \
00046   r6  = _m_punpcklbw(r5,r6);                        \
00047   r5 = tmp1;                                \
00048                                     \
00049   tmp1 = _m_punpckhbw(r7,r8);                       \
00050   r8  = _m_punpcklbw(r7,r8);                        \
00051   r7 = tmp1;                                \
00052                                     \
00053   /*SECOND STEP GROUP Word to DWord*/                   \
00054   tmp1 = _m_punpckhwd(r1,r3);                       \
00055   r3  = _m_punpcklwd(r1,r3);                        \
00056   r1 = tmp1;                                \
00057                                     \
00058   tmp1 = _m_punpckhwd(r2,r4);                       \
00059   r4  = _m_punpcklwd(r2,r4);                        \
00060   r2 = tmp1;                                \
00061                                     \
00062   tmp1 = _m_punpckhwd(r5,r7);                       \
00063   r7  = _m_punpcklwd(r5,r7);                        \
00064   r5 = tmp1;                                \
00065                                     \
00066   tmp1 = _m_punpckhwd(r6,r8);                       \
00067   r8  = _m_punpcklwd(r6,r8);                        \
00068   r6 = tmp1;                                \
00069                                     \
00070   /*THIRD STEP GROUP DWord to QWord */                  \
00071   tmp1 = _m_punpckhdq(r1,r5);                       \
00072   r5  = _m_punpckldq(r1,r5);                        \
00073   r1 = tmp1;                                \
00074                                     \
00075   tmp1 = _m_punpckhdq(r3,r7);                       \
00076   r7  = _m_punpckldq(r3,r7);                        \
00077   r3 = tmp1;                                \
00078                                     \
00079   tmp1 = _m_punpckhdq(r2,r6);                       \
00080   r6  = _m_punpckldq(r2,r6);                        \
00081   r2 = tmp1;                                \
00082                                     \
00083   tmp1 = _m_punpckhdq(r4,r8);                       \
00084   r8  = _m_punpckldq(r4,r8);                        \
00085   r4 = tmp1;                                \
00086 
00087 
00088 
00089 #define PUNPCKHBW(a,b) _mm_unpackhi_epi8(a,b)
00090 #define PUNPCKLBW(a,b) _mm_unpacklo_epi8(a,b)
00091 
00092 #define PUNPCKHWD(a,b) _mm_unpackhi_epi16(a,b)
00093 #define PUNPCKLWD(a,b) _mm_unpacklo_epi16(a,b)
00094 
00095 #define PUNPCKHDQ(a,b) _mm_unpackhi_epi32(a,b)
00096 #define PUNPCKLDQ(a,b) _mm_unpacklo_epi32(a,b)
00097 
00098 #define PUNPCKHQDQ(a,b) _mm_unpackhi_epi64(a,b)
00099 #define PUNPCKLQDQ(a,b) _mm_unpacklo_epi64(a,b)
00100 
00101 
00102 
00103 /*
00104   Use SSE2 instructions
00105   Get Row r1,r2,r3,r4,r5,r6,r7,r8,r9,r10,r11,r12,r13,r14,r15,r16 and do transposition
00106   Transposed Matrix Row are ordered as follow :
00107   r16,r8,r12,r4,r14,r6,r10,r2,r15,r7,r11,r3,r13,r5,r9,r1
00108 */
00109 
00110 #define FLGR_MACRO_TRANSPOSE_BLOCK_16x16_BYTE(r1,r2,r3,r4,r5,r6,r7,r8,  \
00111                           r9,r10,r11,r12,r13,r14,r15,r16,tmp1) \
00112   /*FIRST STEP GROUP Byte To Word*/                 \
00113   tmp1 = PUNPCKHBW(r1,r2);                      \
00114   r2  = PUNPCKLBW(r1,r2);                       \
00115   r1 = tmp1;                                \
00116                                     \
00117   tmp1 = PUNPCKHBW(r3,r4);                      \
00118   r4  = PUNPCKLBW(r3,r4);                       \
00119   r3 = tmp1;                                \
00120                                     \
00121   tmp1 = PUNPCKHBW(r5,r6);                      \
00122   r6  = PUNPCKLBW(r5,r6);                       \
00123   r5 = tmp1;                                \
00124                                     \
00125   tmp1 = PUNPCKHBW(r7,r8);                      \
00126   r8  = PUNPCKLBW(r7,r8);                       \
00127   r7 = tmp1;                                \
00128                                     \
00129   tmp1 = PUNPCKHBW(r9,r10);                     \
00130   r10 = PUNPCKLBW(r9,r10);                      \
00131   r9 = tmp1;                                \
00132                                     \
00133   tmp1 = PUNPCKHBW(r11,r12);                        \
00134   r12 = PUNPCKLBW(r11,r12);                     \
00135   r11 = tmp1;                               \
00136                                     \
00137   tmp1 = PUNPCKHBW(r13,r14);                        \
00138   r14 = PUNPCKLBW(r13,r14);                     \
00139   r13 = tmp1;                               \
00140                                     \
00141   tmp1 = PUNPCKHBW(r15,r16);                        \
00142   r16 = PUNPCKLBW(r15,r16);                     \
00143   r15 = tmp1;                               \
00144                                     \
00145   /*SECOND STEP GROUP Word to DWord*/                   \
00146   tmp1 = PUNPCKHWD(r1,r3);                      \
00147   r3  = PUNPCKLWD(r1,r3);                       \
00148   r1 = tmp1;                                \
00149                                     \
00150   tmp1 = PUNPCKHWD(r2,r4);                      \
00151   r4  = PUNPCKLWD(r2,r4);                       \
00152   r2 = tmp1;                                \
00153                                     \
00154   tmp1 = PUNPCKHWD(r5,r7);                      \
00155   r7  = PUNPCKLWD(r5,r7);                       \
00156   r5 = tmp1;                                \
00157                                     \
00158   tmp1 = PUNPCKHWD(r6,r8);                      \
00159   r8  = PUNPCKLWD(r6,r8);                       \
00160   r6 = tmp1;                                \
00161                                     \
00162   tmp1 = PUNPCKHWD(r9,r11);                     \
00163   r11  = PUNPCKLWD(r9,r11);                     \
00164   r9 = tmp1;                                \
00165                                     \
00166   tmp1 = PUNPCKHWD(r10,r12);                        \
00167   r12  = PUNPCKLWD(r10,r12);                        \
00168   r10 = tmp1;                               \
00169                                     \
00170   tmp1 = PUNPCKHWD(r13,r15);                        \
00171   r15  = PUNPCKLWD(r13,r15);                        \
00172   r13 = tmp1;                               \
00173                                     \
00174   tmp1 = PUNPCKHWD(r14,r16);                        \
00175   r16  = PUNPCKLWD(r14,r16);                        \
00176   r14 = tmp1;                               \
00177                                     \
00178   /*THIRD STEP GROUP DWord to QWord */                  \
00179   tmp1 = PUNPCKHDQ(r1,r5);                      \
00180   r5  = PUNPCKLDQ(r1,r5);                       \
00181   r1 = tmp1;                                \
00182                                     \
00183   tmp1 = PUNPCKHDQ(r2,r6);                      \
00184   r6  = PUNPCKLDQ(r2,r6);                       \
00185   r2 = tmp1;                                \
00186                                     \
00187   tmp1 = PUNPCKHDQ(r3,r7);                      \
00188   r7  = PUNPCKLDQ(r3,r7);                       \
00189   r3 = tmp1;                                \
00190                                     \
00191   tmp1 = PUNPCKHDQ(r4,r8);                      \
00192   r8  = PUNPCKLDQ(r4,r8);                       \
00193   r4 = tmp1;                                \
00194                                     \
00195   tmp1 = PUNPCKHDQ(r9,r13);                     \
00196   r13  = PUNPCKLDQ(r9,r13);                     \
00197   r9 = tmp1;                                \
00198                                     \
00199   tmp1 = PUNPCKHDQ(r10,r14);                        \
00200   r14  = PUNPCKLDQ(r10,r14);                        \
00201   r10 = tmp1;                               \
00202                                     \
00203   tmp1 = PUNPCKHDQ(r11,r15);                        \
00204   r15  = PUNPCKLDQ(r11,r15);                        \
00205   r11 = tmp1;                               \
00206                                     \
00207   tmp1 = PUNPCKHDQ(r12,r16);                        \
00208   r16  = PUNPCKLDQ(r12,r16);                        \
00209   r12 = tmp1;                               \
00210                                     \
00211   /*FOURTH STEP GROUP QWord to DQWord */                \
00212   tmp1 = PUNPCKHQDQ(r1,r9);                     \
00213   r9  = PUNPCKLQDQ(r1,r9);                      \
00214   r1 = tmp1;                                \
00215                                     \
00216   tmp1 = PUNPCKHQDQ(r2,r10);                        \
00217   r10  = PUNPCKLQDQ(r2,r10);                        \
00218   r2 = tmp1;                                \
00219                                     \
00220   tmp1 = PUNPCKHQDQ(r3,r11);                        \
00221   r11  = PUNPCKLQDQ(r3,r11);                        \
00222   r3 = tmp1;                                \
00223                                     \
00224   tmp1 = PUNPCKHQDQ(r4,r12);                        \
00225   r12  = PUNPCKLQDQ(r4,r12);                        \
00226   r4 = tmp1;                                \
00227                                     \
00228   tmp1 = PUNPCKHQDQ(r5,r13);                        \
00229   r13  = PUNPCKLQDQ(r5,r13);                        \
00230   r5 = tmp1;                                \
00231                                     \
00232   tmp1 = PUNPCKHQDQ(r6,r14);                        \
00233   r14  = PUNPCKLQDQ(r6,r14);                        \
00234   r6 = tmp1;                                \
00235                                     \
00236   tmp1 = PUNPCKHQDQ(r7,r15);                        \
00237   r15  = PUNPCKLQDQ(r7,r15);                        \
00238   r7 = tmp1;                                \
00239                                     \
00240   tmp1 = PUNPCKHQDQ(r8,r16);                        \
00241   r16  = PUNPCKLQDQ(r8,r16);                        \
00242   r8 = tmp1
00243   
00244 
00245 
00246 /*
00247   Get Row r1,r2,r3,r4,r5,r6,r7,r8 and do transposition
00248   Transposed Matrix Row are ordered as follow :
00249   r8,r4,r6,r2,r7,r3,r5,r1
00250 */
00251 #define FLGR_MACRO_TRANSPOSE_BLOCK_8x8_WORD(r1,r2,r3,r4,r5,r6,r7,r8, tmp1) \
00252   /*FIRST STEP GROUP Word To DWord*/                    \
00253   tmp1 = PUNPCKHWD(r1,r2);                      \
00254   r2  = PUNPCKLWD(r1,r2);                       \
00255   r1 = tmp1;                                \
00256                                     \
00257   tmp1 = PUNPCKHWD(r3,r4);                      \
00258   r4  = PUNPCKLWD(r3,r4);                       \
00259   r3 = tmp1;                                \
00260                                     \
00261   tmp1 = PUNPCKHWD(r5,r6);                      \
00262   r6  = PUNPCKLWD(r5,r6);                       \
00263   r5 = tmp1;                                \
00264                                     \
00265   tmp1 = PUNPCKHWD(r7,r8);                      \
00266   r8  = PUNPCKLWD(r7,r8);                       \
00267   r7 = tmp1;                                \
00268                                     \
00269   /*SECOND STEP GROUP DWord to QWord*/                  \
00270   tmp1 = PUNPCKHDQ(r1,r3);                      \
00271   r3  = PUNPCKLDQ(r1,r3);                       \
00272   r1 = tmp1;                                \
00273                                     \
00274   tmp1 = PUNPCKHDQ(r2,r4);                      \
00275   r4  = PUNPCKLDQ(r2,r4);                       \
00276   r2 = tmp1;                                \
00277                                     \
00278   tmp1 = PUNPCKHDQ(r5,r7);                      \
00279   r7  = PUNPCKLDQ(r5,r7);                       \
00280   r5 = tmp1;                                \
00281                                     \
00282   tmp1 = PUNPCKHDQ(r6,r8);                      \
00283   r8  = PUNPCKLDQ(r6,r8);                       \
00284   r6 = tmp1;                                \
00285                                     \
00286   /*THIRD STEP GROUP QWord to DQWord */                 \
00287   tmp1 = PUNPCKHQDQ(r1,r5);                     \
00288   r5  = PUNPCKLQDQ(r1,r5);                      \
00289   r1 = tmp1;                                \
00290                                     \
00291   tmp1 = PUNPCKHQDQ(r3,r7);                     \
00292   r7  = PUNPCKLQDQ(r3,r7);                      \
00293   r3 = tmp1;                                \
00294                                     \
00295   tmp1 = PUNPCKHQDQ(r2,r6);                     \
00296   r6  = PUNPCKLQDQ(r2,r6);                      \
00297   r2 = tmp1;                                \
00298                                     \
00299   tmp1 = PUNPCKHQDQ(r4,r8);                     \
00300   r8  = PUNPCKLQDQ(r4,r8);                      \
00301   r4 = tmp1                     
00302 
00303 /*
00304   Get Row r1,r2,r3,r4,r5,r6,r7,r8 and do transposition
00305   Transposed Matrix Row are ordered as follow :
00306   r8,r4,r6,r2,r7,r3,r5,r1
00307 */
00308 #define FLGR_MACRO_TRANSPOSE_BLOCK_4x4_DWORD(r1,r2,r3,r4,tmp1)  \
00309   /*FIRST STEP GROUP Word To DWord*/                \
00310   tmp1 = PUNPCKHDQ(r1,r2);                  \
00311   r2  = PUNPCKLDQ(r1,r2);                   \
00312   r1 = tmp1;                            \
00313                                 \
00314   tmp1 = PUNPCKHDQ(r3,r4);                  \
00315   r4  = PUNPCKLDQ(r3,r4);                   \
00316   r3 = tmp1;                            \
00317                                 \
00318   /*SECOND STEP GROUP DWord to QWord*/              \
00319   tmp1 = PUNPCKHQDQ(r1,r3);                 \
00320   r3  = PUNPCKLQDQ(r1,r3);                  \
00321   r1 = tmp1;                            \
00322                                 \
00323   tmp1 = PUNPCKHQDQ(r2,r4);                 \
00324   r4  = PUNPCKLQDQ(r2,r4);                  \
00325   r2 = tmp1
00326 
00327 
00328 
00329 #endif
00330 
00331 #ifdef __cplusplus
00332 }
00333 #endif