| 1 | ;***************************************************************************** |
| 2 | ;* MMX/SSE2/AVX-optimized 10-bit H.264 weighted prediction code |
| 3 | ;***************************************************************************** |
| 4 | ;* Copyright (C) 2005-2011 x264 project |
| 5 | ;* |
| 6 | ;* Authors: Daniel Kang <daniel.d.kang@gmail.com> |
| 7 | ;* |
| 8 | ;* This file is part of FFmpeg. |
| 9 | ;* |
| 10 | ;* FFmpeg is free software; you can redistribute it and/or |
| 11 | ;* modify it under the terms of the GNU Lesser General Public |
| 12 | ;* License as published by the Free Software Foundation; either |
| 13 | ;* version 2.1 of the License, or (at your option) any later version. |
| 14 | ;* |
| 15 | ;* FFmpeg 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 GNU |
| 18 | ;* Lesser General Public License for more details. |
| 19 | ;* |
| 20 | ;* You should have received a copy of the GNU Lesser General Public |
| 21 | ;* License along with FFmpeg; if not, write to the Free Software |
| 22 | ;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 23 | ;****************************************************************************** |
| 24 | |
| 25 | %include "libavutil/x86/x86util.asm" |
| 26 | |
| 27 | SECTION_RODATA 32 |
| 28 | |
| 29 | pw_pixel_max: times 8 dw ((1 << 10)-1) |
| 30 | sq_1: dq 1 |
| 31 | dq 0 |
| 32 | |
| 33 | cextern pw_1 |
| 34 | |
| 35 | SECTION .text |
| 36 | |
| 37 | ;----------------------------------------------------------------------------- |
| 38 | ; void ff_h264_weight_16_10(uint8_t *dst, int stride, int height, |
| 39 | ; int log2_denom, int weight, int offset); |
| 40 | ;----------------------------------------------------------------------------- |
| 41 | %macro WEIGHT_PROLOGUE 0 |
| 42 | .prologue: |
| 43 | PROLOGUE 0,6,8 |
| 44 | movifnidn r0, r0mp |
| 45 | movifnidn r1d, r1m |
| 46 | movifnidn r2d, r2m |
| 47 | movifnidn r4d, r4m |
| 48 | movifnidn r5d, r5m |
| 49 | %endmacro |
| 50 | |
| 51 | %macro WEIGHT_SETUP 0 |
| 52 | mova m0, [pw_1] |
| 53 | movd m2, r3m |
| 54 | pslld m0, m2 ; 1<<log2_denom |
| 55 | SPLATW m0, m0 |
| 56 | shl r5, 19 ; *8, move to upper half of dword |
| 57 | lea r5, [r5+r4*2+0x10000] |
| 58 | movd m3, r5d ; weight<<1 | 1+(offset<<(3)) |
| 59 | pshufd m3, m3, 0 |
| 60 | mova m4, [pw_pixel_max] |
| 61 | paddw m2, [sq_1] ; log2_denom+1 |
| 62 | %if notcpuflag(sse4) |
| 63 | pxor m7, m7 |
| 64 | %endif |
| 65 | %endmacro |
| 66 | |
| 67 | %macro WEIGHT_OP 1-2 |
| 68 | %if %0==1 |
| 69 | mova m5, [r0+%1] |
| 70 | punpckhwd m6, m5, m0 |
| 71 | punpcklwd m5, m0 |
| 72 | %else |
| 73 | movq m5, [r0+%1] |
| 74 | movq m6, [r0+%2] |
| 75 | punpcklwd m5, m0 |
| 76 | punpcklwd m6, m0 |
| 77 | %endif |
| 78 | pmaddwd m5, m3 |
| 79 | pmaddwd m6, m3 |
| 80 | psrad m5, m2 |
| 81 | psrad m6, m2 |
| 82 | %if cpuflag(sse4) |
| 83 | packusdw m5, m6 |
| 84 | pminsw m5, m4 |
| 85 | %else |
| 86 | packssdw m5, m6 |
| 87 | CLIPW m5, m7, m4 |
| 88 | %endif |
| 89 | %endmacro |
| 90 | |
| 91 | %macro WEIGHT_FUNC_DBL 0 |
| 92 | cglobal h264_weight_16_10 |
| 93 | WEIGHT_PROLOGUE |
| 94 | WEIGHT_SETUP |
| 95 | .nextrow: |
| 96 | WEIGHT_OP 0 |
| 97 | mova [r0 ], m5 |
| 98 | WEIGHT_OP 16 |
| 99 | mova [r0+16], m5 |
| 100 | add r0, r1 |
| 101 | dec r2d |
| 102 | jnz .nextrow |
| 103 | REP_RET |
| 104 | %endmacro |
| 105 | |
| 106 | INIT_XMM sse2 |
| 107 | WEIGHT_FUNC_DBL |
| 108 | INIT_XMM sse4 |
| 109 | WEIGHT_FUNC_DBL |
| 110 | |
| 111 | |
| 112 | %macro WEIGHT_FUNC_MM 0 |
| 113 | cglobal h264_weight_8_10 |
| 114 | WEIGHT_PROLOGUE |
| 115 | WEIGHT_SETUP |
| 116 | .nextrow: |
| 117 | WEIGHT_OP 0 |
| 118 | mova [r0], m5 |
| 119 | add r0, r1 |
| 120 | dec r2d |
| 121 | jnz .nextrow |
| 122 | REP_RET |
| 123 | %endmacro |
| 124 | |
| 125 | INIT_XMM sse2 |
| 126 | WEIGHT_FUNC_MM |
| 127 | INIT_XMM sse4 |
| 128 | WEIGHT_FUNC_MM |
| 129 | |
| 130 | |
| 131 | %macro WEIGHT_FUNC_HALF_MM 0 |
| 132 | cglobal h264_weight_4_10 |
| 133 | WEIGHT_PROLOGUE |
| 134 | sar r2d, 1 |
| 135 | WEIGHT_SETUP |
| 136 | lea r3, [r1*2] |
| 137 | .nextrow: |
| 138 | WEIGHT_OP 0, r1 |
| 139 | movh [r0], m5 |
| 140 | movhps [r0+r1], m5 |
| 141 | add r0, r3 |
| 142 | dec r2d |
| 143 | jnz .nextrow |
| 144 | REP_RET |
| 145 | %endmacro |
| 146 | |
| 147 | INIT_XMM sse2 |
| 148 | WEIGHT_FUNC_HALF_MM |
| 149 | INIT_XMM sse4 |
| 150 | WEIGHT_FUNC_HALF_MM |
| 151 | |
| 152 | |
| 153 | ;----------------------------------------------------------------------------- |
| 154 | ; void ff_h264_biweight_16_10(uint8_t *dst, uint8_t *src, int stride, |
| 155 | ; int height, int log2_denom, int weightd, |
| 156 | ; int weights, int offset); |
| 157 | ;----------------------------------------------------------------------------- |
| 158 | %if ARCH_X86_32 |
| 159 | DECLARE_REG_TMP 3 |
| 160 | %else |
| 161 | DECLARE_REG_TMP 7 |
| 162 | %endif |
| 163 | |
| 164 | %macro BIWEIGHT_PROLOGUE 0 |
| 165 | .prologue: |
| 166 | PROLOGUE 0,8,8 |
| 167 | movifnidn r0, r0mp |
| 168 | movifnidn r1, r1mp |
| 169 | movifnidn r2d, r2m |
| 170 | movifnidn r5d, r5m |
| 171 | movifnidn r6d, r6m |
| 172 | movifnidn t0d, r7m |
| 173 | %endmacro |
| 174 | |
| 175 | %macro BIWEIGHT_SETUP 0 |
| 176 | lea t0, [t0*4+1] ; (offset<<2)+1 |
| 177 | or t0, 1 |
| 178 | shl r6, 16 |
| 179 | or r5, r6 |
| 180 | movd m4, r5d ; weightd | weights |
| 181 | movd m5, t0d ; (offset+1)|1 |
| 182 | movd m6, r4m ; log2_denom |
| 183 | pslld m5, m6 ; (((offset<<2)+1)|1)<<log2_denom |
| 184 | paddd m6, [sq_1] |
| 185 | pshufd m4, m4, 0 |
| 186 | pshufd m5, m5, 0 |
| 187 | mova m3, [pw_pixel_max] |
| 188 | movifnidn r3d, r3m |
| 189 | %if notcpuflag(sse4) |
| 190 | pxor m7, m7 |
| 191 | %endif |
| 192 | %endmacro |
| 193 | |
| 194 | %macro BIWEIGHT 1-2 |
| 195 | %if %0==1 |
| 196 | mova m0, [r0+%1] |
| 197 | mova m1, [r1+%1] |
| 198 | punpckhwd m2, m0, m1 |
| 199 | punpcklwd m0, m1 |
| 200 | %else |
| 201 | movq m0, [r0+%1] |
| 202 | movq m1, [r1+%1] |
| 203 | punpcklwd m0, m1 |
| 204 | movq m2, [r0+%2] |
| 205 | movq m1, [r1+%2] |
| 206 | punpcklwd m2, m1 |
| 207 | %endif |
| 208 | pmaddwd m0, m4 |
| 209 | pmaddwd m2, m4 |
| 210 | paddd m0, m5 |
| 211 | paddd m2, m5 |
| 212 | psrad m0, m6 |
| 213 | psrad m2, m6 |
| 214 | %if cpuflag(sse4) |
| 215 | packusdw m0, m2 |
| 216 | pminsw m0, m3 |
| 217 | %else |
| 218 | packssdw m0, m2 |
| 219 | CLIPW m0, m7, m3 |
| 220 | %endif |
| 221 | %endmacro |
| 222 | |
| 223 | %macro BIWEIGHT_FUNC_DBL 0 |
| 224 | cglobal h264_biweight_16_10 |
| 225 | BIWEIGHT_PROLOGUE |
| 226 | BIWEIGHT_SETUP |
| 227 | .nextrow: |
| 228 | BIWEIGHT 0 |
| 229 | mova [r0 ], m0 |
| 230 | BIWEIGHT 16 |
| 231 | mova [r0+16], m0 |
| 232 | add r0, r2 |
| 233 | add r1, r2 |
| 234 | dec r3d |
| 235 | jnz .nextrow |
| 236 | REP_RET |
| 237 | %endmacro |
| 238 | |
| 239 | INIT_XMM sse2 |
| 240 | BIWEIGHT_FUNC_DBL |
| 241 | INIT_XMM sse4 |
| 242 | BIWEIGHT_FUNC_DBL |
| 243 | |
| 244 | %macro BIWEIGHT_FUNC 0 |
| 245 | cglobal h264_biweight_8_10 |
| 246 | BIWEIGHT_PROLOGUE |
| 247 | BIWEIGHT_SETUP |
| 248 | .nextrow: |
| 249 | BIWEIGHT 0 |
| 250 | mova [r0], m0 |
| 251 | add r0, r2 |
| 252 | add r1, r2 |
| 253 | dec r3d |
| 254 | jnz .nextrow |
| 255 | REP_RET |
| 256 | %endmacro |
| 257 | |
| 258 | INIT_XMM sse2 |
| 259 | BIWEIGHT_FUNC |
| 260 | INIT_XMM sse4 |
| 261 | BIWEIGHT_FUNC |
| 262 | |
| 263 | %macro BIWEIGHT_FUNC_HALF 0 |
| 264 | cglobal h264_biweight_4_10 |
| 265 | BIWEIGHT_PROLOGUE |
| 266 | BIWEIGHT_SETUP |
| 267 | sar r3d, 1 |
| 268 | lea r4, [r2*2] |
| 269 | .nextrow: |
| 270 | BIWEIGHT 0, r2 |
| 271 | movh [r0 ], m0 |
| 272 | movhps [r0+r2], m0 |
| 273 | add r0, r4 |
| 274 | add r1, r4 |
| 275 | dec r3d |
| 276 | jnz .nextrow |
| 277 | REP_RET |
| 278 | %endmacro |
| 279 | |
| 280 | INIT_XMM sse2 |
| 281 | BIWEIGHT_FUNC_HALF |
| 282 | INIT_XMM sse4 |
| 283 | BIWEIGHT_FUNC_HALF |