Mercurial > libavcodec.hg
comparison arm/h264idct_neon.S @ 8462:0ca0e3c98ed5 libavcodec
ARM: add new h264 idct functions
| author | mru |
|---|---|
| date | Thu, 25 Dec 2008 23:13:43 +0000 |
| parents | 9281a8a9387a |
| children | 779a9c93bf61 |
comparison
equal
deleted
inserted
replaced
| 8461:11307ea31e57 | 8462:0ca0e3c98ed5 |
|---|---|
| 18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | 18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
| 19 */ | 19 */ |
| 20 | 20 |
| 21 #include "asm.S" | 21 #include "asm.S" |
| 22 | 22 |
| 23 preserve8 | |
| 23 .fpu neon | 24 .fpu neon |
| 24 | 25 |
| 25 .text | 26 .text |
| 26 | 27 |
| 27 function ff_h264_idct_add_neon, export=1 | 28 function ff_h264_idct_add_neon, export=1 |
| 92 vst1.32 {d0[1]}, [r0,:32], r2 | 93 vst1.32 {d0[1]}, [r0,:32], r2 |
| 93 vst1.32 {d1[0]}, [r0,:32], r2 | 94 vst1.32 {d1[0]}, [r0,:32], r2 |
| 94 vst1.32 {d1[1]}, [r0,:32], r2 | 95 vst1.32 {d1[1]}, [r0,:32], r2 |
| 95 bx lr | 96 bx lr |
| 96 .endfunc | 97 .endfunc |
| 98 | |
| 99 function ff_h264_idct_add16_neon, export=1 | |
| 100 push {r4-r8,lr} | |
| 101 mov r4, r0 | |
| 102 mov r5, r1 | |
| 103 mov r1, r2 | |
| 104 mov r2, r3 | |
| 105 ldr r6, [sp, #24] | |
| 106 movw r7, #:lower16:scan8 | |
| 107 movt r7, #:upper16:scan8 | |
| 108 mov ip, #16 | |
| 109 1: ldrb r8, [r7], #1 | |
| 110 ldr r0, [r5], #4 | |
| 111 ldrb r8, [r6, r8] | |
| 112 subs r8, r8, #1 | |
| 113 blt 2f | |
| 114 ldrsh lr, [r1] | |
| 115 add r0, r0, r4 | |
| 116 movne lr, #0 | |
| 117 cmp lr, #0 | |
| 118 adrne lr, ff_h264_idct_dc_add_neon | |
| 119 adreq lr, ff_h264_idct_add_neon | |
| 120 blx lr | |
| 121 2: subs ip, ip, #1 | |
| 122 add r1, r1, #32 | |
| 123 bne 1b | |
| 124 pop {r4-r8,pc} | |
| 125 .endfunc | |
| 126 | |
| 127 function ff_h264_idct_add16intra_neon, export=1 | |
| 128 push {r4-r8,lr} | |
| 129 mov r4, r0 | |
| 130 mov r5, r1 | |
| 131 mov r1, r2 | |
| 132 mov r2, r3 | |
| 133 ldr r6, [sp, #24] | |
| 134 movw r7, #:lower16:scan8 | |
| 135 movt r7, #:upper16:scan8 | |
| 136 mov ip, #16 | |
| 137 1: ldrb r8, [r7], #1 | |
| 138 ldr r0, [r5], #4 | |
| 139 ldrb r8, [r6, r8] | |
| 140 add r0, r0, r4 | |
| 141 cmp r8, #0 | |
| 142 ldrsh r8, [r1] | |
| 143 adrne lr, ff_h264_idct_add_neon | |
| 144 adreq lr, ff_h264_idct_dc_add_neon | |
| 145 cmpeq r8, #0 | |
| 146 blxne lr | |
| 147 subs ip, ip, #1 | |
| 148 add r1, r1, #32 | |
| 149 bne 1b | |
| 150 pop {r4-r8,pc} | |
| 151 .endfunc | |
| 152 | |
| 153 function ff_h264_idct_add8_neon, export=1 | |
| 154 push {r4-r10,lr} | |
| 155 ldm r0, {r4,r9} | |
| 156 add r5, r1, #16*4 | |
| 157 add r1, r2, #16*32 | |
| 158 mov r2, r3 | |
| 159 ldr r6, [sp, #32] | |
| 160 movw r7, #:lower16:scan8+16 | |
| 161 movt r7, #:upper16:scan8+16 | |
| 162 mov ip, #8 | |
| 163 1: ldrb r8, [r7], #1 | |
| 164 ldr r0, [r5], #4 | |
| 165 ldrb r8, [r6, r8] | |
| 166 tst ip, #4 | |
| 167 addeq r0, r0, r4 | |
| 168 addne r0, r0, r9 | |
| 169 cmp r8, #0 | |
| 170 ldrsh r8, [r1] | |
| 171 adrne lr, ff_h264_idct_add_neon | |
| 172 adreq lr, ff_h264_idct_dc_add_neon | |
| 173 cmpeq r8, #0 | |
| 174 blxne lr | |
| 175 subs ip, ip, #1 | |
| 176 add r1, r1, #32 | |
| 177 bne 1b | |
| 178 pop {r4-r10,pc} | |
| 179 .endfunc | |
| 180 | |
| 181 .section .rodata | |
| 182 scan8: .byte 4+1*8, 5+1*8, 4+2*8, 5+2*8 | |
| 183 .byte 6+1*8, 7+1*8, 6+2*8, 7+2*8 | |
| 184 .byte 4+3*8, 5+3*8, 4+4*8, 5+4*8 | |
| 185 .byte 6+3*8, 7+3*8, 6+4*8, 7+4*8 | |
| 186 .byte 1+1*8, 2+1*8 | |
| 187 .byte 1+2*8, 2+2*8 | |
| 188 .byte 1+4*8, 2+4*8 | |
| 189 .byte 1+5*8, 2+5*8 |