media/libvpx/vp8/common/x86/vp8_asm_stubs.c

changeset 0
6474c204b198
     1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/media/libvpx/vp8/common/x86/vp8_asm_stubs.c	Wed Dec 31 06:09:35 2014 +0100
     1.3 @@ -0,0 +1,625 @@
     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_config.h"
    1.16 +#include "vp8_rtcd.h"
    1.17 +#include "vpx_ports/mem.h"
    1.18 +#include "filter_x86.h"
    1.19 +
    1.20 +extern const short vp8_six_tap_mmx[8][6*8];
    1.21 +
    1.22 +extern void vp8_filter_block1d_h6_mmx
    1.23 +(
    1.24 +    unsigned char   *src_ptr,
    1.25 +    unsigned short  *output_ptr,
    1.26 +    unsigned int    src_pixels_per_line,
    1.27 +    unsigned int    pixel_step,
    1.28 +    unsigned int    output_height,
    1.29 +    unsigned int    output_width,
    1.30 +    const short      *vp8_filter
    1.31 +);
    1.32 +extern void vp8_filter_block1dc_v6_mmx
    1.33 +(
    1.34 +    unsigned short *src_ptr,
    1.35 +    unsigned char  *output_ptr,
    1.36 +    int             output_pitch,
    1.37 +    unsigned int    pixels_per_line,
    1.38 +    unsigned int    pixel_step,
    1.39 +    unsigned int    output_height,
    1.40 +    unsigned int    output_width,
    1.41 +    const short    *vp8_filter
    1.42 +);
    1.43 +extern void vp8_filter_block1d8_h6_sse2
    1.44 +(
    1.45 +    unsigned char  *src_ptr,
    1.46 +    unsigned short *output_ptr,
    1.47 +    unsigned int    src_pixels_per_line,
    1.48 +    unsigned int    pixel_step,
    1.49 +    unsigned int    output_height,
    1.50 +    unsigned int    output_width,
    1.51 +    const short    *vp8_filter
    1.52 +);
    1.53 +extern void vp8_filter_block1d16_h6_sse2
    1.54 +(
    1.55 +    unsigned char  *src_ptr,
    1.56 +    unsigned short *output_ptr,
    1.57 +    unsigned int    src_pixels_per_line,
    1.58 +    unsigned int    pixel_step,
    1.59 +    unsigned int    output_height,
    1.60 +    unsigned int    output_width,
    1.61 +    const short    *vp8_filter
    1.62 +);
    1.63 +extern void vp8_filter_block1d8_v6_sse2
    1.64 +(
    1.65 +    unsigned short *src_ptr,
    1.66 +    unsigned char *output_ptr,
    1.67 +    int dst_ptich,
    1.68 +    unsigned int pixels_per_line,
    1.69 +    unsigned int pixel_step,
    1.70 +    unsigned int output_height,
    1.71 +    unsigned int output_width,
    1.72 +    const short    *vp8_filter
    1.73 +);
    1.74 +extern void vp8_filter_block1d16_v6_sse2
    1.75 +(
    1.76 +    unsigned short *src_ptr,
    1.77 +    unsigned char *output_ptr,
    1.78 +    int dst_ptich,
    1.79 +    unsigned int pixels_per_line,
    1.80 +    unsigned int pixel_step,
    1.81 +    unsigned int output_height,
    1.82 +    unsigned int output_width,
    1.83 +    const short    *vp8_filter
    1.84 +);
    1.85 +extern void vp8_unpack_block1d16_h6_sse2
    1.86 +(
    1.87 +    unsigned char  *src_ptr,
    1.88 +    unsigned short *output_ptr,
    1.89 +    unsigned int    src_pixels_per_line,
    1.90 +    unsigned int    output_height,
    1.91 +    unsigned int    output_width
    1.92 +);
    1.93 +extern void vp8_filter_block1d8_h6_only_sse2
    1.94 +(
    1.95 +    unsigned char  *src_ptr,
    1.96 +    unsigned int    src_pixels_per_line,
    1.97 +    unsigned char  *output_ptr,
    1.98 +    int dst_ptich,
    1.99 +    unsigned int    output_height,
   1.100 +    const short    *vp8_filter
   1.101 +);
   1.102 +extern void vp8_filter_block1d16_h6_only_sse2
   1.103 +(
   1.104 +    unsigned char  *src_ptr,
   1.105 +    unsigned int    src_pixels_per_line,
   1.106 +    unsigned char  *output_ptr,
   1.107 +    int dst_ptich,
   1.108 +    unsigned int    output_height,
   1.109 +    const short    *vp8_filter
   1.110 +);
   1.111 +extern void vp8_filter_block1d8_v6_only_sse2
   1.112 +(
   1.113 +    unsigned char *src_ptr,
   1.114 +    unsigned int   src_pixels_per_line,
   1.115 +    unsigned char *output_ptr,
   1.116 +    int dst_ptich,
   1.117 +    unsigned int   output_height,
   1.118 +    const short   *vp8_filter
   1.119 +);
   1.120 +
   1.121 +
   1.122 +#if HAVE_MMX
   1.123 +void vp8_sixtap_predict4x4_mmx
   1.124 +(
   1.125 +    unsigned char  *src_ptr,
   1.126 +    int   src_pixels_per_line,
   1.127 +    int  xoffset,
   1.128 +    int  yoffset,
   1.129 +    unsigned char *dst_ptr,
   1.130 +    int dst_pitch
   1.131 +)
   1.132 +{
   1.133 +    DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 16*16);  /* Temp data bufffer used in filtering */
   1.134 +    const short *HFilter, *VFilter;
   1.135 +    HFilter = vp8_six_tap_mmx[xoffset];
   1.136 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line), FData2, src_pixels_per_line, 1, 9, 8, HFilter);
   1.137 +    VFilter = vp8_six_tap_mmx[yoffset];
   1.138 +    vp8_filter_block1dc_v6_mmx(FData2 + 8, dst_ptr, dst_pitch, 8, 4 , 4, 4, VFilter);
   1.139 +
   1.140 +}
   1.141 +
   1.142 +
   1.143 +void vp8_sixtap_predict16x16_mmx
   1.144 +(
   1.145 +    unsigned char  *src_ptr,
   1.146 +    int   src_pixels_per_line,
   1.147 +    int  xoffset,
   1.148 +    int  yoffset,
   1.149 +    unsigned char *dst_ptr,
   1.150 +    int dst_pitch
   1.151 +)
   1.152 +{
   1.153 +
   1.154 +    DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 24*24);  /* Temp data bufffer used in filtering */
   1.155 +
   1.156 +    const short *HFilter, *VFilter;
   1.157 +
   1.158 +
   1.159 +    HFilter = vp8_six_tap_mmx[xoffset];
   1.160 +
   1.161 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 21, 32, HFilter);
   1.162 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 21, 32, HFilter);
   1.163 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 8,  FData2 + 8, src_pixels_per_line, 1, 21, 32, HFilter);
   1.164 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 12, FData2 + 12, src_pixels_per_line, 1, 21, 32, HFilter);
   1.165 +
   1.166 +    VFilter = vp8_six_tap_mmx[yoffset];
   1.167 +    vp8_filter_block1dc_v6_mmx(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, 16, VFilter);
   1.168 +    vp8_filter_block1dc_v6_mmx(FData2 + 36, dst_ptr + 4, dst_pitch, 32, 16 , 16, 16, VFilter);
   1.169 +    vp8_filter_block1dc_v6_mmx(FData2 + 40, dst_ptr + 8, dst_pitch, 32, 16 , 16, 16, VFilter);
   1.170 +    vp8_filter_block1dc_v6_mmx(FData2 + 44, dst_ptr + 12, dst_pitch, 32, 16 , 16, 16, VFilter);
   1.171 +
   1.172 +}
   1.173 +
   1.174 +
   1.175 +void vp8_sixtap_predict8x8_mmx
   1.176 +(
   1.177 +    unsigned char  *src_ptr,
   1.178 +    int   src_pixels_per_line,
   1.179 +    int  xoffset,
   1.180 +    int  yoffset,
   1.181 +    unsigned char *dst_ptr,
   1.182 +    int dst_pitch
   1.183 +)
   1.184 +{
   1.185 +
   1.186 +    DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 256);    /* Temp data bufffer used in filtering */
   1.187 +
   1.188 +    const short *HFilter, *VFilter;
   1.189 +
   1.190 +    HFilter = vp8_six_tap_mmx[xoffset];
   1.191 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 13, 16, HFilter);
   1.192 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 13, 16, HFilter);
   1.193 +
   1.194 +    VFilter = vp8_six_tap_mmx[yoffset];
   1.195 +    vp8_filter_block1dc_v6_mmx(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 8, 8, VFilter);
   1.196 +    vp8_filter_block1dc_v6_mmx(FData2 + 20, dst_ptr + 4, dst_pitch, 16, 8 , 8, 8, VFilter);
   1.197 +
   1.198 +}
   1.199 +
   1.200 +
   1.201 +void vp8_sixtap_predict8x4_mmx
   1.202 +(
   1.203 +    unsigned char  *src_ptr,
   1.204 +    int   src_pixels_per_line,
   1.205 +    int  xoffset,
   1.206 +    int  yoffset,
   1.207 +    unsigned char *dst_ptr,
   1.208 +    int dst_pitch
   1.209 +)
   1.210 +{
   1.211 +
   1.212 +    DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 256);    /* Temp data bufffer used in filtering */
   1.213 +
   1.214 +    const short *HFilter, *VFilter;
   1.215 +
   1.216 +    HFilter = vp8_six_tap_mmx[xoffset];
   1.217 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line),    FData2,   src_pixels_per_line, 1, 9, 16, HFilter);
   1.218 +    vp8_filter_block1d_h6_mmx(src_ptr - (2 * src_pixels_per_line) + 4,  FData2 + 4, src_pixels_per_line, 1, 9, 16, HFilter);
   1.219 +
   1.220 +    VFilter = vp8_six_tap_mmx[yoffset];
   1.221 +    vp8_filter_block1dc_v6_mmx(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 4, 8, VFilter);
   1.222 +    vp8_filter_block1dc_v6_mmx(FData2 + 20, dst_ptr + 4, dst_pitch, 16, 8 , 4, 8, VFilter);
   1.223 +
   1.224 +}
   1.225 +
   1.226 +
   1.227 +
   1.228 +void vp8_bilinear_predict16x16_mmx
   1.229 +(
   1.230 +    unsigned char  *src_ptr,
   1.231 +    int   src_pixels_per_line,
   1.232 +    int  xoffset,
   1.233 +    int  yoffset,
   1.234 +    unsigned char *dst_ptr,
   1.235 +    int dst_pitch
   1.236 +)
   1.237 +{
   1.238 +    vp8_bilinear_predict8x8_mmx(src_ptr,   src_pixels_per_line, xoffset, yoffset, dst_ptr,   dst_pitch);
   1.239 +    vp8_bilinear_predict8x8_mmx(src_ptr + 8, src_pixels_per_line, xoffset, yoffset, dst_ptr + 8, dst_pitch);
   1.240 +    vp8_bilinear_predict8x8_mmx(src_ptr + 8 * src_pixels_per_line,   src_pixels_per_line, xoffset, yoffset, dst_ptr + dst_pitch * 8,   dst_pitch);
   1.241 +    vp8_bilinear_predict8x8_mmx(src_ptr + 8 * src_pixels_per_line + 8, src_pixels_per_line, xoffset, yoffset, dst_ptr + dst_pitch * 8 + 8, dst_pitch);
   1.242 +}
   1.243 +#endif
   1.244 +
   1.245 +
   1.246 +#if HAVE_SSE2
   1.247 +void vp8_sixtap_predict16x16_sse2
   1.248 +(
   1.249 +    unsigned char  *src_ptr,
   1.250 +    int   src_pixels_per_line,
   1.251 +    int  xoffset,
   1.252 +    int  yoffset,
   1.253 +    unsigned char *dst_ptr,
   1.254 +    int dst_pitch
   1.255 +
   1.256 +)
   1.257 +{
   1.258 +    DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 24*24);    /* Temp data bufffer used in filtering */
   1.259 +
   1.260 +    const short *HFilter, *VFilter;
   1.261 +
   1.262 +    if (xoffset)
   1.263 +    {
   1.264 +        if (yoffset)
   1.265 +        {
   1.266 +            HFilter = vp8_six_tap_mmx[xoffset];
   1.267 +            vp8_filter_block1d16_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 21, 32, HFilter);
   1.268 +            VFilter = vp8_six_tap_mmx[yoffset];
   1.269 +            vp8_filter_block1d16_v6_sse2(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, dst_pitch, VFilter);
   1.270 +        }
   1.271 +        else
   1.272 +        {
   1.273 +            /* First-pass only */
   1.274 +            HFilter = vp8_six_tap_mmx[xoffset];
   1.275 +            vp8_filter_block1d16_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 16, HFilter);
   1.276 +        }
   1.277 +    }
   1.278 +    else
   1.279 +    {
   1.280 +        /* Second-pass only */
   1.281 +        VFilter = vp8_six_tap_mmx[yoffset];
   1.282 +        vp8_unpack_block1d16_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 21, 32);
   1.283 +        vp8_filter_block1d16_v6_sse2(FData2 + 32, dst_ptr,   dst_pitch, 32, 16 , 16, dst_pitch, VFilter);
   1.284 +    }
   1.285 +}
   1.286 +
   1.287 +
   1.288 +void vp8_sixtap_predict8x8_sse2
   1.289 +(
   1.290 +    unsigned char  *src_ptr,
   1.291 +    int   src_pixels_per_line,
   1.292 +    int  xoffset,
   1.293 +    int  yoffset,
   1.294 +    unsigned char *dst_ptr,
   1.295 +    int dst_pitch
   1.296 +)
   1.297 +{
   1.298 +    DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 256);  /* Temp data bufffer used in filtering */
   1.299 +    const short *HFilter, *VFilter;
   1.300 +
   1.301 +    if (xoffset)
   1.302 +    {
   1.303 +        if (yoffset)
   1.304 +        {
   1.305 +            HFilter = vp8_six_tap_mmx[xoffset];
   1.306 +            vp8_filter_block1d8_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 13, 16, HFilter);
   1.307 +            VFilter = vp8_six_tap_mmx[yoffset];
   1.308 +            vp8_filter_block1d8_v6_sse2(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 8, dst_pitch, VFilter);
   1.309 +        }
   1.310 +        else
   1.311 +        {
   1.312 +            /* First-pass only */
   1.313 +            HFilter = vp8_six_tap_mmx[xoffset];
   1.314 +            vp8_filter_block1d8_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 8, HFilter);
   1.315 +        }
   1.316 +    }
   1.317 +    else
   1.318 +    {
   1.319 +        /* Second-pass only */
   1.320 +        VFilter = vp8_six_tap_mmx[yoffset];
   1.321 +        vp8_filter_block1d8_v6_only_sse2(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 8, VFilter);
   1.322 +    }
   1.323 +}
   1.324 +
   1.325 +
   1.326 +void vp8_sixtap_predict8x4_sse2
   1.327 +(
   1.328 +    unsigned char  *src_ptr,
   1.329 +    int   src_pixels_per_line,
   1.330 +    int  xoffset,
   1.331 +    int  yoffset,
   1.332 +    unsigned char *dst_ptr,
   1.333 +    int dst_pitch
   1.334 +)
   1.335 +{
   1.336 +    DECLARE_ALIGNED_ARRAY(16, unsigned short, FData2, 256);  /* Temp data bufffer used in filtering */
   1.337 +    const short *HFilter, *VFilter;
   1.338 +
   1.339 +    if (xoffset)
   1.340 +    {
   1.341 +        if (yoffset)
   1.342 +        {
   1.343 +            HFilter = vp8_six_tap_mmx[xoffset];
   1.344 +            vp8_filter_block1d8_h6_sse2(src_ptr - (2 * src_pixels_per_line), FData2,   src_pixels_per_line, 1, 9, 16, HFilter);
   1.345 +            VFilter = vp8_six_tap_mmx[yoffset];
   1.346 +            vp8_filter_block1d8_v6_sse2(FData2 + 16, dst_ptr,   dst_pitch, 16, 8 , 4, dst_pitch, VFilter);
   1.347 +        }
   1.348 +        else
   1.349 +        {
   1.350 +            /* First-pass only */
   1.351 +            HFilter = vp8_six_tap_mmx[xoffset];
   1.352 +            vp8_filter_block1d8_h6_only_sse2(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch, 4, HFilter);
   1.353 +        }
   1.354 +    }
   1.355 +    else
   1.356 +    {
   1.357 +        /* Second-pass only */
   1.358 +        VFilter = vp8_six_tap_mmx[yoffset];
   1.359 +        vp8_filter_block1d8_v6_only_sse2(src_ptr - (2 * src_pixels_per_line), src_pixels_per_line, dst_ptr, dst_pitch, 4, VFilter);
   1.360 +    }
   1.361 +}
   1.362 +
   1.363 +#endif
   1.364 +
   1.365 +#if HAVE_SSSE3
   1.366 +
   1.367 +extern void vp8_filter_block1d8_h6_ssse3
   1.368 +(
   1.369 +    unsigned char  *src_ptr,
   1.370 +    unsigned int    src_pixels_per_line,
   1.371 +    unsigned char  *output_ptr,
   1.372 +    unsigned int    output_pitch,
   1.373 +    unsigned int    output_height,
   1.374 +    unsigned int    vp8_filter_index
   1.375 +);
   1.376 +
   1.377 +extern void vp8_filter_block1d16_h6_ssse3
   1.378 +(
   1.379 +    unsigned char  *src_ptr,
   1.380 +    unsigned int    src_pixels_per_line,
   1.381 +    unsigned char  *output_ptr,
   1.382 +    unsigned int    output_pitch,
   1.383 +    unsigned int    output_height,
   1.384 +    unsigned int    vp8_filter_index
   1.385 +);
   1.386 +
   1.387 +extern void vp8_filter_block1d16_v6_ssse3
   1.388 +(
   1.389 +    unsigned char *src_ptr,
   1.390 +    unsigned int   src_pitch,
   1.391 +    unsigned char *output_ptr,
   1.392 +    unsigned int   out_pitch,
   1.393 +    unsigned int   output_height,
   1.394 +    unsigned int   vp8_filter_index
   1.395 +);
   1.396 +
   1.397 +extern void vp8_filter_block1d8_v6_ssse3
   1.398 +(
   1.399 +    unsigned char *src_ptr,
   1.400 +    unsigned int   src_pitch,
   1.401 +    unsigned char *output_ptr,
   1.402 +    unsigned int   out_pitch,
   1.403 +    unsigned int   output_height,
   1.404 +    unsigned int   vp8_filter_index
   1.405 +);
   1.406 +
   1.407 +extern void vp8_filter_block1d4_h6_ssse3
   1.408 +(
   1.409 +    unsigned char  *src_ptr,
   1.410 +    unsigned int    src_pixels_per_line,
   1.411 +    unsigned char  *output_ptr,
   1.412 +    unsigned int    output_pitch,
   1.413 +    unsigned int    output_height,
   1.414 +    unsigned int    vp8_filter_index
   1.415 +);
   1.416 +
   1.417 +extern void vp8_filter_block1d4_v6_ssse3
   1.418 +(
   1.419 +    unsigned char *src_ptr,
   1.420 +    unsigned int   src_pitch,
   1.421 +    unsigned char *output_ptr,
   1.422 +    unsigned int   out_pitch,
   1.423 +    unsigned int   output_height,
   1.424 +    unsigned int   vp8_filter_index
   1.425 +);
   1.426 +
   1.427 +void vp8_sixtap_predict16x16_ssse3
   1.428 +(
   1.429 +    unsigned char  *src_ptr,
   1.430 +    int   src_pixels_per_line,
   1.431 +    int  xoffset,
   1.432 +    int  yoffset,
   1.433 +    unsigned char *dst_ptr,
   1.434 +    int dst_pitch
   1.435 +
   1.436 +)
   1.437 +{
   1.438 +    DECLARE_ALIGNED_ARRAY(16, unsigned char, FData2, 24*24);
   1.439 +
   1.440 +    if (xoffset)
   1.441 +    {
   1.442 +        if (yoffset)
   1.443 +        {
   1.444 +            vp8_filter_block1d16_h6_ssse3(src_ptr - (2 * src_pixels_per_line),
   1.445 +                                          src_pixels_per_line, FData2,
   1.446 +                                          16, 21, xoffset);
   1.447 +            vp8_filter_block1d16_v6_ssse3(FData2 , 16, dst_ptr, dst_pitch,
   1.448 +                                          16, yoffset);
   1.449 +        }
   1.450 +        else
   1.451 +        {
   1.452 +            /* First-pass only */
   1.453 +            vp8_filter_block1d16_h6_ssse3(src_ptr, src_pixels_per_line,
   1.454 +                                          dst_ptr, dst_pitch, 16, xoffset);
   1.455 +        }
   1.456 +    }
   1.457 +    else
   1.458 +    {
   1.459 +        if (yoffset)
   1.460 +        {
   1.461 +            /* Second-pass only */
   1.462 +            vp8_filter_block1d16_v6_ssse3(src_ptr - (2 * src_pixels_per_line),
   1.463 +                                          src_pixels_per_line,
   1.464 +                                          dst_ptr, dst_pitch, 16, yoffset);
   1.465 +        }
   1.466 +        else
   1.467 +        {
   1.468 +            /* ssse3 second-pass only function couldn't handle (xoffset==0 &&
   1.469 +             * yoffset==0) case correctly. Add copy function here to guarantee
   1.470 +             * six-tap function handles all possible offsets. */
   1.471 +            vp8_copy_mem16x16(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch);
   1.472 +        }
   1.473 +    }
   1.474 +}
   1.475 +
   1.476 +void vp8_sixtap_predict8x8_ssse3
   1.477 +(
   1.478 +    unsigned char  *src_ptr,
   1.479 +    int   src_pixels_per_line,
   1.480 +    int  xoffset,
   1.481 +    int  yoffset,
   1.482 +    unsigned char *dst_ptr,
   1.483 +    int dst_pitch
   1.484 +)
   1.485 +{
   1.486 +    DECLARE_ALIGNED_ARRAY(16, unsigned char, FData2, 256);
   1.487 +
   1.488 +    if (xoffset)
   1.489 +    {
   1.490 +        if (yoffset)
   1.491 +        {
   1.492 +            vp8_filter_block1d8_h6_ssse3(src_ptr - (2 * src_pixels_per_line),
   1.493 +                                         src_pixels_per_line, FData2,
   1.494 +                                         8, 13, xoffset);
   1.495 +            vp8_filter_block1d8_v6_ssse3(FData2, 8, dst_ptr, dst_pitch,
   1.496 +                                         8, yoffset);
   1.497 +        }
   1.498 +        else
   1.499 +        {
   1.500 +            vp8_filter_block1d8_h6_ssse3(src_ptr, src_pixels_per_line,
   1.501 +                                         dst_ptr, dst_pitch, 8, xoffset);
   1.502 +        }
   1.503 +    }
   1.504 +    else
   1.505 +    {
   1.506 +        if (yoffset)
   1.507 +        {
   1.508 +            /* Second-pass only */
   1.509 +            vp8_filter_block1d8_v6_ssse3(src_ptr - (2 * src_pixels_per_line),
   1.510 +                                         src_pixels_per_line,
   1.511 +                                         dst_ptr, dst_pitch, 8, yoffset);
   1.512 +        }
   1.513 +        else
   1.514 +        {
   1.515 +            /* ssse3 second-pass only function couldn't handle (xoffset==0 &&
   1.516 +             * yoffset==0) case correctly. Add copy function here to guarantee
   1.517 +             * six-tap function handles all possible offsets. */
   1.518 +            vp8_copy_mem8x8(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch);
   1.519 +        }
   1.520 +    }
   1.521 +}
   1.522 +
   1.523 +
   1.524 +void vp8_sixtap_predict8x4_ssse3
   1.525 +(
   1.526 +    unsigned char  *src_ptr,
   1.527 +    int   src_pixels_per_line,
   1.528 +    int  xoffset,
   1.529 +    int  yoffset,
   1.530 +    unsigned char *dst_ptr,
   1.531 +    int dst_pitch
   1.532 +)
   1.533 +{
   1.534 +    DECLARE_ALIGNED_ARRAY(16, unsigned char, FData2, 256);
   1.535 +
   1.536 +    if (xoffset)
   1.537 +    {
   1.538 +        if (yoffset)
   1.539 +        {
   1.540 +            vp8_filter_block1d8_h6_ssse3(src_ptr - (2 * src_pixels_per_line),
   1.541 +                                         src_pixels_per_line, FData2,
   1.542 +                                         8, 9, xoffset);
   1.543 +            vp8_filter_block1d8_v6_ssse3(FData2, 8, dst_ptr, dst_pitch,
   1.544 +                                         4, yoffset);
   1.545 +        }
   1.546 +        else
   1.547 +        {
   1.548 +            /* First-pass only */
   1.549 +            vp8_filter_block1d8_h6_ssse3(src_ptr, src_pixels_per_line,
   1.550 +                                         dst_ptr, dst_pitch, 4, xoffset);
   1.551 +        }
   1.552 +    }
   1.553 +    else
   1.554 +    {
   1.555 +        if (yoffset)
   1.556 +        {
   1.557 +            /* Second-pass only */
   1.558 +            vp8_filter_block1d8_v6_ssse3(src_ptr - (2 * src_pixels_per_line),
   1.559 +                                         src_pixels_per_line,
   1.560 +                                         dst_ptr, dst_pitch, 4, yoffset);
   1.561 +        }
   1.562 +        else
   1.563 +        {
   1.564 +            /* ssse3 second-pass only function couldn't handle (xoffset==0 &&
   1.565 +             * yoffset==0) case correctly. Add copy function here to guarantee
   1.566 +             * six-tap function handles all possible offsets. */
   1.567 +            vp8_copy_mem8x4(src_ptr, src_pixels_per_line, dst_ptr, dst_pitch);
   1.568 +        }
   1.569 +    }
   1.570 +}
   1.571 +
   1.572 +void vp8_sixtap_predict4x4_ssse3
   1.573 +(
   1.574 +    unsigned char  *src_ptr,
   1.575 +    int   src_pixels_per_line,
   1.576 +    int  xoffset,
   1.577 +    int  yoffset,
   1.578 +    unsigned char *dst_ptr,
   1.579 +    int dst_pitch
   1.580 +)
   1.581 +{
   1.582 +  DECLARE_ALIGNED_ARRAY(16, unsigned char, FData2, 4*9);
   1.583 +
   1.584 +  if (xoffset)
   1.585 +  {
   1.586 +      if (yoffset)
   1.587 +      {
   1.588 +          vp8_filter_block1d4_h6_ssse3(src_ptr - (2 * src_pixels_per_line),
   1.589 +                                       src_pixels_per_line,
   1.590 +                                       FData2, 4, 9, xoffset);
   1.591 +          vp8_filter_block1d4_v6_ssse3(FData2, 4, dst_ptr, dst_pitch,
   1.592 +                                       4, yoffset);
   1.593 +      }
   1.594 +      else
   1.595 +      {
   1.596 +          vp8_filter_block1d4_h6_ssse3(src_ptr, src_pixels_per_line,
   1.597 +                                       dst_ptr, dst_pitch, 4, xoffset);
   1.598 +      }
   1.599 +  }
   1.600 +  else
   1.601 +  {
   1.602 +      if (yoffset)
   1.603 +      {
   1.604 +          vp8_filter_block1d4_v6_ssse3(src_ptr - (2 * src_pixels_per_line),
   1.605 +                                       src_pixels_per_line,
   1.606 +                                       dst_ptr, dst_pitch, 4, yoffset);
   1.607 +      }
   1.608 +      else
   1.609 +      {
   1.610 +        /* ssse3 second-pass only function couldn't handle (xoffset==0 &&
   1.611 +          * yoffset==0) case correctly. Add copy function here to guarantee
   1.612 +          * six-tap function handles all possible offsets. */
   1.613 +          int r;
   1.614 +
   1.615 +          for (r = 0; r < 4; r++)
   1.616 +          {
   1.617 +            dst_ptr[0]  = src_ptr[0];
   1.618 +            dst_ptr[1]  = src_ptr[1];
   1.619 +            dst_ptr[2]  = src_ptr[2];
   1.620 +            dst_ptr[3]  = src_ptr[3];
   1.621 +            dst_ptr     += dst_pitch;
   1.622 +            src_ptr     += src_pixels_per_line;
   1.623 +          }
   1.624 +      }
   1.625 +  }
   1.626 +}
   1.627 +
   1.628 +#endif

mercurial