media/libvpx/vp8/common/x86/variance_impl_ssse3.asm

changeset 0
6474c204b198
     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/media/libvpx/vp8/common/x86/variance_impl_ssse3.asm	Wed Dec 31 06:09:35 2014 +0100
     1.3 @@ -0,0 +1,364 @@
     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 +%define xmm_filter_shift            7
    1.18 +
    1.19 +
    1.20 +;void vp8_filter_block2d_bil_var_ssse3
    1.21 +;(
    1.22 +;    unsigned char *ref_ptr,
    1.23 +;    int ref_pixels_per_line,
    1.24 +;    unsigned char *src_ptr,
    1.25 +;    int src_pixels_per_line,
    1.26 +;    unsigned int Height,
    1.27 +;    int  xoffset,
    1.28 +;    int  yoffset,
    1.29 +;    int *sum,
    1.30 +;    unsigned int *sumsquared;;
    1.31 +;
    1.32 +;)
    1.33 +;Note: The filter coefficient at offset=0 is 128. Since the second register
    1.34 +;for Pmaddubsw is signed bytes, we must calculate zero offset seperately.
    1.35 +global sym(vp8_filter_block2d_bil_var_ssse3) PRIVATE
    1.36 +sym(vp8_filter_block2d_bil_var_ssse3):
    1.37 +    push        rbp
    1.38 +    mov         rbp, rsp
    1.39 +    SHADOW_ARGS_TO_STACK 9
    1.40 +    SAVE_XMM 7
    1.41 +    GET_GOT     rbx
    1.42 +    push rsi
    1.43 +    push rdi
    1.44 +    ; end prolog
    1.45 +
    1.46 +        pxor            xmm6,           xmm6
    1.47 +        pxor            xmm7,           xmm7
    1.48 +
    1.49 +        lea             rcx,            [GLOBAL(vp8_bilinear_filters_ssse3)]
    1.50 +        movsxd          rax,            dword ptr arg(5)     ; xoffset
    1.51 +
    1.52 +        cmp             rax,            0                    ; skip first_pass filter if xoffset=0
    1.53 +        je              .filter_block2d_bil_var_ssse3_sp_only
    1.54 +
    1.55 +        shl             rax,            4                    ; point to filter coeff with xoffset
    1.56 +        lea             rax,            [rax + rcx]          ; HFilter
    1.57 +
    1.58 +        movsxd          rdx,            dword ptr arg(6)     ; yoffset
    1.59 +
    1.60 +        cmp             rdx,            0                    ; skip second_pass filter if yoffset=0
    1.61 +        je              .filter_block2d_bil_var_ssse3_fp_only
    1.62 +
    1.63 +        shl             rdx,            4
    1.64 +        lea             rdx,            [rdx + rcx]          ; VFilter
    1.65 +
    1.66 +        mov             rsi,            arg(0)               ;ref_ptr
    1.67 +        mov             rdi,            arg(2)               ;src_ptr
    1.68 +        movsxd          rcx,            dword ptr arg(4)     ;Height
    1.69 +
    1.70 +        movdqu          xmm0,           XMMWORD PTR [rsi]
    1.71 +        movdqu          xmm1,           XMMWORD PTR [rsi+1]
    1.72 +        movdqa          xmm2,           xmm0
    1.73 +
    1.74 +        punpcklbw       xmm0,           xmm1
    1.75 +        punpckhbw       xmm2,           xmm1
    1.76 +        pmaddubsw       xmm0,           [rax]
    1.77 +        pmaddubsw       xmm2,           [rax]
    1.78 +
    1.79 +        paddw           xmm0,           [GLOBAL(xmm_bi_rd)]
    1.80 +        paddw           xmm2,           [GLOBAL(xmm_bi_rd)]
    1.81 +        psraw           xmm0,           xmm_filter_shift
    1.82 +        psraw           xmm2,           xmm_filter_shift
    1.83 +
    1.84 +        packuswb        xmm0,           xmm2
    1.85 +
    1.86 +%if ABI_IS_32BIT
    1.87 +        add             rsi,            dword ptr arg(1) ;ref_pixels_per_line
    1.88 +%else
    1.89 +        movsxd          r8,             dword ptr arg(1) ;ref_pixels_per_line
    1.90 +        movsxd          r9,             dword ptr arg(3) ;src_pixels_per_line
    1.91 +        lea             rsi,            [rsi + r8]
    1.92 +%endif
    1.93 +
    1.94 +.filter_block2d_bil_var_ssse3_loop:
    1.95 +        movdqu          xmm1,           XMMWORD PTR [rsi]
    1.96 +        movdqu          xmm2,           XMMWORD PTR [rsi+1]
    1.97 +        movdqa          xmm3,           xmm1
    1.98 +
    1.99 +        punpcklbw       xmm1,           xmm2
   1.100 +        punpckhbw       xmm3,           xmm2
   1.101 +        pmaddubsw       xmm1,           [rax]
   1.102 +        pmaddubsw       xmm3,           [rax]
   1.103 +
   1.104 +        paddw           xmm1,           [GLOBAL(xmm_bi_rd)]
   1.105 +        paddw           xmm3,           [GLOBAL(xmm_bi_rd)]
   1.106 +        psraw           xmm1,           xmm_filter_shift
   1.107 +        psraw           xmm3,           xmm_filter_shift
   1.108 +        packuswb        xmm1,           xmm3
   1.109 +
   1.110 +        movdqa          xmm2,           xmm0
   1.111 +        movdqa          xmm0,           xmm1
   1.112 +        movdqa          xmm3,           xmm2
   1.113 +
   1.114 +        punpcklbw       xmm2,           xmm1
   1.115 +        punpckhbw       xmm3,           xmm1
   1.116 +        pmaddubsw       xmm2,           [rdx]
   1.117 +        pmaddubsw       xmm3,           [rdx]
   1.118 +
   1.119 +        paddw           xmm2,           [GLOBAL(xmm_bi_rd)]
   1.120 +        paddw           xmm3,           [GLOBAL(xmm_bi_rd)]
   1.121 +        psraw           xmm2,           xmm_filter_shift
   1.122 +        psraw           xmm3,           xmm_filter_shift
   1.123 +
   1.124 +        movq            xmm1,           QWORD PTR [rdi]
   1.125 +        pxor            xmm4,           xmm4
   1.126 +        punpcklbw       xmm1,           xmm4
   1.127 +        movq            xmm5,           QWORD PTR [rdi+8]
   1.128 +        punpcklbw       xmm5,           xmm4
   1.129 +
   1.130 +        psubw           xmm2,           xmm1
   1.131 +        psubw           xmm3,           xmm5
   1.132 +        paddw           xmm6,           xmm2
   1.133 +        paddw           xmm6,           xmm3
   1.134 +        pmaddwd         xmm2,           xmm2
   1.135 +        pmaddwd         xmm3,           xmm3
   1.136 +        paddd           xmm7,           xmm2
   1.137 +        paddd           xmm7,           xmm3
   1.138 +
   1.139 +%if ABI_IS_32BIT
   1.140 +        add             rsi,            dword ptr arg(1)     ;ref_pixels_per_line
   1.141 +        add             rdi,            dword ptr arg(3)     ;src_pixels_per_line
   1.142 +%else
   1.143 +        lea             rsi,            [rsi + r8]
   1.144 +        lea             rdi,            [rdi + r9]
   1.145 +%endif
   1.146 +
   1.147 +        sub             rcx,            1
   1.148 +        jnz             .filter_block2d_bil_var_ssse3_loop
   1.149 +
   1.150 +        jmp             .filter_block2d_bil_variance
   1.151 +
   1.152 +.filter_block2d_bil_var_ssse3_sp_only:
   1.153 +        movsxd          rdx,            dword ptr arg(6)     ; yoffset
   1.154 +
   1.155 +        cmp             rdx,            0                    ; Both xoffset =0 and yoffset=0
   1.156 +        je              .filter_block2d_bil_var_ssse3_full_pixel
   1.157 +
   1.158 +        shl             rdx,            4
   1.159 +        lea             rdx,            [rdx + rcx]          ; VFilter
   1.160 +
   1.161 +        mov             rsi,            arg(0)               ;ref_ptr
   1.162 +        mov             rdi,            arg(2)               ;src_ptr
   1.163 +        movsxd          rcx,            dword ptr arg(4)     ;Height
   1.164 +        movsxd          rax,            dword ptr arg(1)     ;ref_pixels_per_line
   1.165 +
   1.166 +        movdqu          xmm1,           XMMWORD PTR [rsi]
   1.167 +        movdqa          xmm0,           xmm1
   1.168 +
   1.169 +%if ABI_IS_32BIT=0
   1.170 +        movsxd          r9,             dword ptr arg(3) ;src_pixels_per_line
   1.171 +%endif
   1.172 +
   1.173 +        lea             rsi,            [rsi + rax]
   1.174 +
   1.175 +.filter_block2d_bil_sp_only_loop:
   1.176 +        movdqu          xmm3,           XMMWORD PTR [rsi]
   1.177 +        movdqa          xmm2,           xmm1
   1.178 +        movdqa          xmm0,           xmm3
   1.179 +
   1.180 +        punpcklbw       xmm1,           xmm3
   1.181 +        punpckhbw       xmm2,           xmm3
   1.182 +        pmaddubsw       xmm1,           [rdx]
   1.183 +        pmaddubsw       xmm2,           [rdx]
   1.184 +
   1.185 +        paddw           xmm1,           [GLOBAL(xmm_bi_rd)]
   1.186 +        paddw           xmm2,           [GLOBAL(xmm_bi_rd)]
   1.187 +        psraw           xmm1,           xmm_filter_shift
   1.188 +        psraw           xmm2,           xmm_filter_shift
   1.189 +
   1.190 +        movq            xmm3,           QWORD PTR [rdi]
   1.191 +        pxor            xmm4,           xmm4
   1.192 +        punpcklbw       xmm3,           xmm4
   1.193 +        movq            xmm5,           QWORD PTR [rdi+8]
   1.194 +        punpcklbw       xmm5,           xmm4
   1.195 +
   1.196 +        psubw           xmm1,           xmm3
   1.197 +        psubw           xmm2,           xmm5
   1.198 +        paddw           xmm6,           xmm1
   1.199 +        paddw           xmm6,           xmm2
   1.200 +        pmaddwd         xmm1,           xmm1
   1.201 +        pmaddwd         xmm2,           xmm2
   1.202 +        paddd           xmm7,           xmm1
   1.203 +        paddd           xmm7,           xmm2
   1.204 +
   1.205 +        movdqa          xmm1,           xmm0
   1.206 +        lea             rsi,            [rsi + rax]          ;ref_pixels_per_line
   1.207 +
   1.208 +%if ABI_IS_32BIT
   1.209 +        add             rdi,            dword ptr arg(3)     ;src_pixels_per_line
   1.210 +%else
   1.211 +        lea             rdi,            [rdi + r9]
   1.212 +%endif
   1.213 +
   1.214 +        sub             rcx,            1
   1.215 +        jnz             .filter_block2d_bil_sp_only_loop
   1.216 +
   1.217 +        jmp             .filter_block2d_bil_variance
   1.218 +
   1.219 +.filter_block2d_bil_var_ssse3_full_pixel:
   1.220 +        mov             rsi,            arg(0)               ;ref_ptr
   1.221 +        mov             rdi,            arg(2)               ;src_ptr
   1.222 +        movsxd          rcx,            dword ptr arg(4)     ;Height
   1.223 +        movsxd          rax,            dword ptr arg(1)     ;ref_pixels_per_line
   1.224 +        movsxd          rdx,            dword ptr arg(3)     ;src_pixels_per_line
   1.225 +        pxor            xmm0,           xmm0
   1.226 +
   1.227 +.filter_block2d_bil_full_pixel_loop:
   1.228 +        movq            xmm1,           QWORD PTR [rsi]
   1.229 +        punpcklbw       xmm1,           xmm0
   1.230 +        movq            xmm2,           QWORD PTR [rsi+8]
   1.231 +        punpcklbw       xmm2,           xmm0
   1.232 +
   1.233 +        movq            xmm3,           QWORD PTR [rdi]
   1.234 +        punpcklbw       xmm3,           xmm0
   1.235 +        movq            xmm4,           QWORD PTR [rdi+8]
   1.236 +        punpcklbw       xmm4,           xmm0
   1.237 +
   1.238 +        psubw           xmm1,           xmm3
   1.239 +        psubw           xmm2,           xmm4
   1.240 +        paddw           xmm6,           xmm1
   1.241 +        paddw           xmm6,           xmm2
   1.242 +        pmaddwd         xmm1,           xmm1
   1.243 +        pmaddwd         xmm2,           xmm2
   1.244 +        paddd           xmm7,           xmm1
   1.245 +        paddd           xmm7,           xmm2
   1.246 +
   1.247 +        lea             rsi,            [rsi + rax]          ;ref_pixels_per_line
   1.248 +        lea             rdi,            [rdi + rdx]          ;src_pixels_per_line
   1.249 +        sub             rcx,            1
   1.250 +        jnz             .filter_block2d_bil_full_pixel_loop
   1.251 +
   1.252 +        jmp             .filter_block2d_bil_variance
   1.253 +
   1.254 +.filter_block2d_bil_var_ssse3_fp_only:
   1.255 +        mov             rsi,            arg(0)               ;ref_ptr
   1.256 +        mov             rdi,            arg(2)               ;src_ptr
   1.257 +        movsxd          rcx,            dword ptr arg(4)     ;Height
   1.258 +        movsxd          rdx,            dword ptr arg(1)     ;ref_pixels_per_line
   1.259 +
   1.260 +        pxor            xmm0,           xmm0
   1.261 +
   1.262 +%if ABI_IS_32BIT=0
   1.263 +        movsxd          r9,             dword ptr arg(3) ;src_pixels_per_line
   1.264 +%endif
   1.265 +
   1.266 +.filter_block2d_bil_fp_only_loop:
   1.267 +        movdqu          xmm1,           XMMWORD PTR [rsi]
   1.268 +        movdqu          xmm2,           XMMWORD PTR [rsi+1]
   1.269 +        movdqa          xmm3,           xmm1
   1.270 +
   1.271 +        punpcklbw       xmm1,           xmm2
   1.272 +        punpckhbw       xmm3,           xmm2
   1.273 +        pmaddubsw       xmm1,           [rax]
   1.274 +        pmaddubsw       xmm3,           [rax]
   1.275 +
   1.276 +        paddw           xmm1,           [GLOBAL(xmm_bi_rd)]
   1.277 +        paddw           xmm3,           [GLOBAL(xmm_bi_rd)]
   1.278 +        psraw           xmm1,           xmm_filter_shift
   1.279 +        psraw           xmm3,           xmm_filter_shift
   1.280 +
   1.281 +        movq            xmm2,           XMMWORD PTR [rdi]
   1.282 +        pxor            xmm4,           xmm4
   1.283 +        punpcklbw       xmm2,           xmm4
   1.284 +        movq            xmm5,           QWORD PTR [rdi+8]
   1.285 +        punpcklbw       xmm5,           xmm4
   1.286 +
   1.287 +        psubw           xmm1,           xmm2
   1.288 +        psubw           xmm3,           xmm5
   1.289 +        paddw           xmm6,           xmm1
   1.290 +        paddw           xmm6,           xmm3
   1.291 +        pmaddwd         xmm1,           xmm1
   1.292 +        pmaddwd         xmm3,           xmm3
   1.293 +        paddd           xmm7,           xmm1
   1.294 +        paddd           xmm7,           xmm3
   1.295 +
   1.296 +        lea             rsi,            [rsi + rdx]
   1.297 +%if ABI_IS_32BIT
   1.298 +        add             rdi,            dword ptr arg(3)     ;src_pixels_per_line
   1.299 +%else
   1.300 +        lea             rdi,            [rdi + r9]
   1.301 +%endif
   1.302 +
   1.303 +        sub             rcx,            1
   1.304 +        jnz             .filter_block2d_bil_fp_only_loop
   1.305 +
   1.306 +        jmp             .filter_block2d_bil_variance
   1.307 +
   1.308 +.filter_block2d_bil_variance:
   1.309 +        pxor        xmm0,           xmm0
   1.310 +        pxor        xmm1,           xmm1
   1.311 +        pxor        xmm5,           xmm5
   1.312 +
   1.313 +        punpcklwd   xmm0,           xmm6
   1.314 +        punpckhwd   xmm1,           xmm6
   1.315 +        psrad       xmm0,           16
   1.316 +        psrad       xmm1,           16
   1.317 +        paddd       xmm0,           xmm1
   1.318 +        movdqa      xmm1,           xmm0
   1.319 +
   1.320 +        movdqa      xmm6,           xmm7
   1.321 +        punpckldq   xmm6,           xmm5
   1.322 +        punpckhdq   xmm7,           xmm5
   1.323 +        paddd       xmm6,           xmm7
   1.324 +
   1.325 +        punpckldq   xmm0,           xmm5
   1.326 +        punpckhdq   xmm1,           xmm5
   1.327 +        paddd       xmm0,           xmm1
   1.328 +
   1.329 +        movdqa      xmm7,           xmm6
   1.330 +        movdqa      xmm1,           xmm0
   1.331 +
   1.332 +        psrldq      xmm7,           8
   1.333 +        psrldq      xmm1,           8
   1.334 +
   1.335 +        paddd       xmm6,           xmm7
   1.336 +        paddd       xmm0,           xmm1
   1.337 +
   1.338 +        mov         rsi,            arg(7) ;[Sum]
   1.339 +        mov         rdi,            arg(8) ;[SSE]
   1.340 +
   1.341 +        movd        [rsi],       xmm0
   1.342 +        movd        [rdi],       xmm6
   1.343 +
   1.344 +    ; begin epilog
   1.345 +    pop rdi
   1.346 +    pop rsi
   1.347 +    RESTORE_GOT
   1.348 +    RESTORE_XMM
   1.349 +    UNSHADOW_ARGS
   1.350 +    pop         rbp
   1.351 +    ret
   1.352 +
   1.353 +
   1.354 +SECTION_RODATA
   1.355 +align 16
   1.356 +xmm_bi_rd:
   1.357 +    times 8 dw 64
   1.358 +align 16
   1.359 +vp8_bilinear_filters_ssse3:
   1.360 +    times 8 db 128, 0
   1.361 +    times 8 db 112, 16
   1.362 +    times 8 db 96,  32
   1.363 +    times 8 db 80,  48
   1.364 +    times 8 db 64,  64
   1.365 +    times 8 db 48,  80
   1.366 +    times 8 db 32,  96
   1.367 +    times 8 db 16,  112

mercurial