media/libvpx/vp8/encoder/x86/subtract_sse2.asm

changeset 0
6474c204b198
     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/media/libvpx/vp8/encoder/x86/subtract_sse2.asm	Wed Dec 31 06:09:35 2014 +0100
     1.3 @@ -0,0 +1,245 @@
     1.4 +;
     1.5 +;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
     1.6 +;
     1.7 +;  Use of this source code is governed by a BSD-style license
     1.8 +;  that can be found in the LICENSE file in the root of the source
     1.9 +;  tree. An additional intellectual property rights grant can be found
    1.10 +;  in the file PATENTS.  All contributing project authors may
    1.11 +;  be found in the AUTHORS file in the root of the source tree.
    1.12 +;
    1.13 +
    1.14 +
    1.15 +%include "vpx_ports/x86_abi_support.asm"
    1.16 +
    1.17 +;void vp8_subtract_b_sse2_impl(unsigned char *z,  int src_stride,
    1.18 +;                            short *diff, unsigned char *Predictor,
    1.19 +;                            int pitch);
    1.20 +global sym(vp8_subtract_b_sse2_impl) PRIVATE
    1.21 +sym(vp8_subtract_b_sse2_impl):
    1.22 +    push        rbp
    1.23 +    mov         rbp, rsp
    1.24 +    SHADOW_ARGS_TO_STACK 5
    1.25 +    GET_GOT     rbx
    1.26 +    push rsi
    1.27 +    push rdi
    1.28 +    ; end prolog
    1.29 +
    1.30 +        mov     rdi,        arg(2) ;diff
    1.31 +        mov     rax,        arg(3) ;Predictor
    1.32 +        mov     rsi,        arg(0) ;z
    1.33 +        movsxd  rdx,        dword ptr arg(1);src_stride;
    1.34 +        movsxd  rcx,        dword ptr arg(4);pitch
    1.35 +        pxor    mm7,        mm7
    1.36 +
    1.37 +        movd    mm0,        [rsi]
    1.38 +        movd    mm1,        [rax]
    1.39 +        punpcklbw   mm0,    mm7
    1.40 +        punpcklbw   mm1,    mm7
    1.41 +        psubw   mm0,        mm1
    1.42 +        movq    MMWORD PTR [rdi],      mm0
    1.43 +
    1.44 +        movd    mm0,        [rsi+rdx]
    1.45 +        movd    mm1,        [rax+rcx]
    1.46 +        punpcklbw   mm0,    mm7
    1.47 +        punpcklbw   mm1,    mm7
    1.48 +        psubw   mm0,        mm1
    1.49 +        movq    MMWORD PTR [rdi+rcx*2], mm0
    1.50 +
    1.51 +        movd    mm0,        [rsi+rdx*2]
    1.52 +        movd    mm1,        [rax+rcx*2]
    1.53 +        punpcklbw   mm0,    mm7
    1.54 +        punpcklbw   mm1,    mm7
    1.55 +        psubw   mm0,        mm1
    1.56 +        movq    MMWORD PTR [rdi+rcx*4], mm0
    1.57 +
    1.58 +        lea     rsi,        [rsi+rdx*2]
    1.59 +        lea     rcx,        [rcx+rcx*2]
    1.60 +
    1.61 +        movd    mm0,        [rsi+rdx]
    1.62 +        movd    mm1,        [rax+rcx]
    1.63 +        punpcklbw   mm0,    mm7
    1.64 +        punpcklbw   mm1,    mm7
    1.65 +        psubw   mm0,        mm1
    1.66 +        movq    MMWORD PTR [rdi+rcx*2], mm0
    1.67 +
    1.68 +    ; begin epilog
    1.69 +    pop rdi
    1.70 +    pop rsi
    1.71 +    RESTORE_GOT
    1.72 +    UNSHADOW_ARGS
    1.73 +    pop         rbp
    1.74 +    ret
    1.75 +
    1.76 +
    1.77 +;void vp8_subtract_mby_sse2(short *diff, unsigned char *src, int src_stride,
    1.78 +;unsigned char *pred, int pred_stride)
    1.79 +global sym(vp8_subtract_mby_sse2) PRIVATE
    1.80 +sym(vp8_subtract_mby_sse2):
    1.81 +    push        rbp
    1.82 +    mov         rbp, rsp
    1.83 +    SHADOW_ARGS_TO_STACK 5
    1.84 +    GET_GOT     rbx
    1.85 +    push rsi
    1.86 +    push rdi
    1.87 +    ; end prolog
    1.88 +
    1.89 +    mov         rdi,        arg(0)          ;diff
    1.90 +    mov         rsi,        arg(1)          ;src
    1.91 +    movsxd      rdx,        dword ptr arg(2);src_stride
    1.92 +    mov         rax,        arg(3)          ;pred
    1.93 +    movdqa      xmm4,       [GLOBAL(t80)]
    1.94 +    push        rbx
    1.95 +    mov         rcx,        8               ; do two lines at one time
    1.96 +    movsxd      rbx,        dword ptr arg(4);pred_stride
    1.97 +
    1.98 +.submby_loop:
    1.99 +    movdqa      xmm0,       [rsi]           ; src
   1.100 +    movdqa      xmm1,       [rax]           ; pred
   1.101 +
   1.102 +    movdqa      xmm2,       xmm0
   1.103 +    psubb       xmm0,       xmm1
   1.104 +
   1.105 +    pxor        xmm1,       xmm4            ;convert to signed values
   1.106 +    pxor        xmm2,       xmm4
   1.107 +    pcmpgtb     xmm1,       xmm2            ; obtain sign information
   1.108 +
   1.109 +    movdqa      xmm2,       xmm0
   1.110 +    punpcklbw   xmm0,       xmm1            ; put sign back to subtraction
   1.111 +    punpckhbw   xmm2,       xmm1            ; put sign back to subtraction
   1.112 +
   1.113 +    movdqa      xmm3,       [rsi + rdx]
   1.114 +    movdqa      xmm5,       [rax + rbx]
   1.115 +
   1.116 +    lea         rsi,        [rsi+rdx*2]
   1.117 +    lea         rax,        [rax+rbx*2]
   1.118 +
   1.119 +    movdqa      [rdi],      xmm0
   1.120 +    movdqa      [rdi +16],  xmm2
   1.121 +
   1.122 +    movdqa      xmm1,       xmm3
   1.123 +    psubb       xmm3,       xmm5
   1.124 +
   1.125 +    pxor        xmm5,       xmm4            ;convert to signed values
   1.126 +    pxor        xmm1,       xmm4
   1.127 +    pcmpgtb     xmm5,       xmm1            ; obtain sign information
   1.128 +
   1.129 +    movdqa      xmm1,       xmm3
   1.130 +    punpcklbw   xmm3,       xmm5            ; put sign back to subtraction
   1.131 +    punpckhbw   xmm1,       xmm5            ; put sign back to subtraction
   1.132 +
   1.133 +    movdqa      [rdi +32],  xmm3
   1.134 +    movdqa      [rdi +48],  xmm1
   1.135 +
   1.136 +    add         rdi,        64
   1.137 +    dec         rcx
   1.138 +    jnz         .submby_loop
   1.139 +
   1.140 +    pop rbx
   1.141 +    pop rdi
   1.142 +    pop rsi
   1.143 +    ; begin epilog
   1.144 +    RESTORE_GOT
   1.145 +    UNSHADOW_ARGS
   1.146 +    pop         rbp
   1.147 +    ret
   1.148 +
   1.149 +;vp8_subtract_mbuv_sse2(short *diff, unsigned char *usrc, unsigned char *vsrc,
   1.150 +;                         int src_stride, unsigned char *upred,
   1.151 +;                         unsigned char *vpred, int pred_stride)
   1.152 +global sym(vp8_subtract_mbuv_sse2) PRIVATE
   1.153 +sym(vp8_subtract_mbuv_sse2):
   1.154 +    push        rbp
   1.155 +    mov         rbp, rsp
   1.156 +    SHADOW_ARGS_TO_STACK 7
   1.157 +    GET_GOT     rbx
   1.158 +    push rsi
   1.159 +    push rdi
   1.160 +    ; end prolog
   1.161 +
   1.162 +    movdqa      xmm4,       [GLOBAL(t80)]
   1.163 +    mov         rdi,        arg(0)          ;diff
   1.164 +    mov         rsi,        arg(1)          ;usrc
   1.165 +    movsxd      rdx,        dword ptr arg(3);src_stride;
   1.166 +    mov         rax,        arg(4)          ;upred
   1.167 +    add         rdi,        256*2           ;diff = diff + 256 (shorts)
   1.168 +    mov         rcx,        4
   1.169 +    push        rbx
   1.170 +    movsxd      rbx,        dword ptr arg(6);pred_stride
   1.171 +
   1.172 +    ;u
   1.173 +.submbu_loop:
   1.174 +    movq        xmm0,       [rsi]           ; src
   1.175 +    movq        xmm2,       [rsi+rdx]       ; src -- next line
   1.176 +    movq        xmm1,       [rax]           ; pred
   1.177 +    movq        xmm3,       [rax+rbx]       ; pred -- next line
   1.178 +    lea         rsi,        [rsi + rdx*2]
   1.179 +    lea         rax,        [rax + rbx*2]
   1.180 +
   1.181 +    punpcklqdq  xmm0,       xmm2
   1.182 +    punpcklqdq  xmm1,       xmm3
   1.183 +
   1.184 +    movdqa      xmm2,       xmm0
   1.185 +    psubb       xmm0,       xmm1            ; subtraction with sign missed
   1.186 +
   1.187 +    pxor        xmm1,       xmm4            ;convert to signed values
   1.188 +    pxor        xmm2,       xmm4
   1.189 +    pcmpgtb     xmm1,       xmm2            ; obtain sign information
   1.190 +
   1.191 +    movdqa      xmm2,       xmm0
   1.192 +    movdqa      xmm3,       xmm1
   1.193 +    punpcklbw   xmm0,       xmm1            ; put sign back to subtraction
   1.194 +    punpckhbw   xmm2,       xmm3            ; put sign back to subtraction
   1.195 +
   1.196 +    movdqa      [rdi],      xmm0            ; store difference
   1.197 +    movdqa      [rdi +16],  xmm2            ; store difference
   1.198 +    add         rdi,        32
   1.199 +    sub         rcx, 1
   1.200 +    jnz         .submbu_loop
   1.201 +
   1.202 +    mov         rsi,        arg(2)          ;vsrc
   1.203 +    mov         rax,        arg(5)          ;vpred
   1.204 +    mov         rcx,        4
   1.205 +
   1.206 +    ;v
   1.207 +.submbv_loop:
   1.208 +    movq        xmm0,       [rsi]           ; src
   1.209 +    movq        xmm2,       [rsi+rdx]       ; src -- next line
   1.210 +    movq        xmm1,       [rax]           ; pred
   1.211 +    movq        xmm3,       [rax+rbx]       ; pred -- next line
   1.212 +    lea         rsi,        [rsi + rdx*2]
   1.213 +    lea         rax,        [rax + rbx*2]
   1.214 +
   1.215 +    punpcklqdq  xmm0,       xmm2
   1.216 +    punpcklqdq  xmm1,       xmm3
   1.217 +
   1.218 +    movdqa      xmm2,       xmm0
   1.219 +    psubb       xmm0,       xmm1            ; subtraction with sign missed
   1.220 +
   1.221 +    pxor        xmm1,       xmm4            ;convert to signed values
   1.222 +    pxor        xmm2,       xmm4
   1.223 +    pcmpgtb     xmm1,       xmm2            ; obtain sign information
   1.224 +
   1.225 +    movdqa      xmm2,       xmm0
   1.226 +    movdqa      xmm3,       xmm1
   1.227 +    punpcklbw   xmm0,       xmm1            ; put sign back to subtraction
   1.228 +    punpckhbw   xmm2,       xmm3            ; put sign back to subtraction
   1.229 +
   1.230 +    movdqa      [rdi],      xmm0            ; store difference
   1.231 +    movdqa      [rdi +16],  xmm2            ; store difference
   1.232 +    add         rdi,        32
   1.233 +    sub         rcx, 1
   1.234 +    jnz         .submbv_loop
   1.235 +
   1.236 +    pop         rbx
   1.237 +    ; begin epilog
   1.238 +    pop rdi
   1.239 +    pop rsi
   1.240 +    RESTORE_GOT
   1.241 +    UNSHADOW_ARGS
   1.242 +    pop         rbp
   1.243 +    ret
   1.244 +
   1.245 +SECTION_RODATA
   1.246 +align 16
   1.247 +t80:
   1.248 +    times 16 db 0x80

mercurial