1 ;*****************************************************************************
2 ;* dct-64.asm: h264 encoder library
3 ;*****************************************************************************
4 ;* Copyright (C) 2003-2008 x264 project
6 ;* Authors: Laurent Aimar <fenrir@via.ecp.fr> (initial version)
7 ;* Loren Merritt <lorenm@u.washington.edu> (dct8, misc)
8 ;* Min Chen <chenm001.163.com> (converted to nasm)
10 ;* This program is free software; you can redistribute it and/or modify
11 ;* it under the terms of the GNU General Public License as published by
12 ;* the Free Software Foundation; either version 2 of the License, or
13 ;* (at your option) any later version.
15 ;* This program is distributed in the hope that it will be useful,
16 ;* but WITHOUT ANY WARRANTY; without even the implied warranty of
17 ;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 ;* GNU General Public License for more details.
20 ;* You should have received a copy of the GNU General Public License
21 ;* along with this program; if not, write to the Free Software
22 ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
23 ;*****************************************************************************
55 %macro TRANSPOSE8x8W 9
56 SBUTTERFLY wd, %1, %2, %9
57 SBUTTERFLY wd, %3, %4, %9
58 SBUTTERFLY wd, %5, %6, %9
59 SBUTTERFLY wd, %7, %8, %9
60 SBUTTERFLY dq, %1, %3, %9
61 SBUTTERFLY dq, %2, %4, %9
62 SBUTTERFLY dq, %5, %7, %9
63 SBUTTERFLY dq, %6, %8, %9
64 SBUTTERFLY qdq, %1, %5, %9
65 SBUTTERFLY qdq, %2, %6, %9
66 SBUTTERFLY qdq, %3, %7, %9
67 SBUTTERFLY qdq, %4, %8, %9
72 %macro STORE_DIFF_8P 4
84 SUMSUB_BA m%8, m%1 ; %8=s07, %1=d07
85 SUMSUB_BA m%7, m%2 ; %7=s16, %2=d16
86 SUMSUB_BA m%6, m%3 ; %6=s25, %3=d25
87 SUMSUB_BA m%5, m%4 ; %5=s34, %4=d34
89 SUMSUB_BA m%5, m%8 ; %5=a0, %8=a2
90 SUMSUB_BA m%6, m%7 ; %6=a1, %7=a3
96 paddw m%9, m%3 ; %9=a4
102 psubw m%10, m%3 ; %10=a7
109 psubw m%1, m%3 ; %1=a5
110 psubw m%4, m%2 ; %4=a6
112 SUMSUB_BA m%6, m%5 ; %6=b0, %5=b4
116 paddw m%2, m%9 ; %2=b1
118 psubw m%9, m%10 ; %9=b7
122 paddw m%3, m%8 ; %3=b2
124 psubw m%8, m%7 ; %8=b6
128 paddw m%7, m%1 ; %7=b3
130 psubw m%4, m%1 ; %4=b5
132 SWAP %1, %6, %4, %7, %8, %9
135 ;-----------------------------------------------------------------------------
136 ; void x264_sub8x8_dct8_sse2( int16_t dct[8][8], uint8_t *pix1, uint8_t *pix2 )
137 ;-----------------------------------------------------------------------------
138 cglobal x264_sub8x8_dct8_sse2
139 LOAD_DIFF_8P m0, m8, m9, [r1+0*FENC_STRIDE], [r2+0*FDEC_STRIDE]
140 LOAD_DIFF_8P m1, m8, m9, [r1+1*FENC_STRIDE], [r2+1*FDEC_STRIDE]
141 LOAD_DIFF_8P m2, m8, m9, [r1+2*FENC_STRIDE], [r2+2*FDEC_STRIDE]
142 LOAD_DIFF_8P m3, m8, m9, [r1+3*FENC_STRIDE], [r2+3*FDEC_STRIDE]
143 LOAD_DIFF_8P m4, m8, m9, [r1+4*FENC_STRIDE], [r2+4*FDEC_STRIDE]
144 LOAD_DIFF_8P m5, m8, m9, [r1+5*FENC_STRIDE], [r2+5*FDEC_STRIDE]
145 LOAD_DIFF_8P m6, m8, m9, [r1+6*FENC_STRIDE], [r2+6*FDEC_STRIDE]
146 LOAD_DIFF_8P m7, m8, m9, [r1+7*FENC_STRIDE], [r2+7*FDEC_STRIDE]
148 DCT8_1D 0,1,2,3,4,5,6,7,8,9
149 TRANSPOSE8x8W 0,1,2,3,4,5,6,7,8
150 DCT8_1D 0,1,2,3,4,5,6,7,8,9
164 SUMSUB_BA m%5, m%1 ; %5=a0, %1=a2
167 psubw m%3, m%7 ; %3=a4
169 paddw m%7, m%10 ; %7=a6
175 paddw m%9, m%6 ; %9=a7
181 psubw m%10, m%2 ; %10=a5
189 psubw m%2, m%4 ; %2=a3
190 psubw m%6, m%8 ; %6=a1
192 SUMSUB_BA m%7, m%5 ; %7=b0, %5=b6
193 SUMSUB_BA m%3, m%1 ; %3=b2, %1=b4
197 paddw m%4, m%6 ; %4=b1
199 psubw m%9, m%6 ; %9=b7
203 paddw m%8, m%2 ; %8=b3
205 psubw m%2, m%10 ; %2=b5
207 SUMSUB_BA m%9, m%7 ; %9=c0, %7=c7
208 SUMSUB_BA m%2, m%3 ; %2=c1, %3=c6
209 SUMSUB_BA m%8, m%1 ; %8=c2, %1=c5
210 SUMSUB_BA m%4, m%5 ; %4=c3, %5=c4
216 ;-----------------------------------------------------------------------------
217 ; void x264_add8x8_idct8_sse2( uint8_t *p_dst, int16_t dct[8][8] )
218 ;-----------------------------------------------------------------------------
219 cglobal x264_add8x8_idct8_sse2
229 IDCT8_1D 0,1,2,3,4,5,6,7,8,9
230 TRANSPOSE8x8W 0,1,2,3,4,5,6,7,8
231 paddw m0, [pw_32 GLOBAL] ; rounding for the >>6 at the end
232 IDCT8_1D 0,1,2,3,4,5,6,7,8,9
235 STORE_DIFF_8P m0, m8, m9, [r0+0*FDEC_STRIDE]
236 STORE_DIFF_8P m1, m8, m9, [r0+1*FDEC_STRIDE]
237 STORE_DIFF_8P m2, m8, m9, [r0+2*FDEC_STRIDE]
238 STORE_DIFF_8P m3, m8, m9, [r0+3*FDEC_STRIDE]
239 STORE_DIFF_8P m4, m8, m9, [r0+4*FDEC_STRIDE]
240 STORE_DIFF_8P m5, m8, m9, [r0+5*FDEC_STRIDE]
241 STORE_DIFF_8P m6, m8, m9, [r0+6*FDEC_STRIDE]
242 STORE_DIFF_8P m7, m8, m9, [r0+7*FDEC_STRIDE]