1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/libavcodec/ppc/dsputil_altivec.c	Tue Sep 25 15:55:33 2012 +0200
     1.3 @@ -0,0 +1,619 @@
     1.4 +/*
     1.5 + * Copyright (c) 2002 Brian Foley
     1.6 + * Copyright (c) 2002 Dieter Shirley
     1.7 + * Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
     1.8 + *
     1.9 + * This file is part of FFmpeg.
    1.10 + *
    1.11 + * FFmpeg is free software; you can redistribute it and/or
    1.12 + * modify it under the terms of the GNU Lesser General Public
    1.13 + * License as published by the Free Software Foundation; either
    1.14 + * version 2.1 of the License, or (at your option) any later version.
    1.15 + *
    1.16 + * FFmpeg is distributed in the hope that it will be useful,
    1.17 + * but WITHOUT ANY WARRANTY; without even the implied warranty of
    1.18 + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
    1.19 + * Lesser General Public License for more details.
    1.20 + *
    1.21 + * You should have received a copy of the GNU Lesser General Public
    1.22 + * License along with FFmpeg; if not, write to the Free Software
    1.23 + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
    1.24 + */
    1.25 +
    1.26 +#include "config.h"
    1.27 +#if HAVE_ALTIVEC_H
    1.28 +#include <altivec.h>
    1.29 +#endif
    1.30 +#include "libavcodec/dsputil.h"
    1.31 +#include "dsputil_ppc.h"
    1.32 +#include "util_altivec.h"
    1.33 +#include "types_altivec.h"
    1.34 +#include "dsputil_altivec.h"
    1.35 +
    1.36 +
    1.37 +static void get_pixels_altivec(DCTELEM *restrict block, const uint8_t *pixels, int line_size)
    1.38 +{
    1.39 +    int i;
    1.40 +    vector unsigned char perm, bytes, *pixv;
    1.41 +    const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
    1.42 +    vector signed short shorts;
    1.43 +
    1.44 +    for (i = 0; i < 8; i++) {
    1.45 +        // Read potentially unaligned pixels.
    1.46 +        // We're reading 16 pixels, and actually only want 8,
    1.47 +        // but we simply ignore the extras.
    1.48 +        perm = vec_lvsl(0, pixels);
    1.49 +        pixv = (vector unsigned char *) pixels;
    1.50 +        bytes = vec_perm(pixv[0], pixv[1], perm);
    1.51 +
    1.52 +        // convert the bytes into shorts
    1.53 +        shorts = (vector signed short)vec_mergeh(zero, bytes);
    1.54 +
    1.55 +        // save the data to the block, we assume the block is 16-byte aligned
    1.56 +        vec_st(shorts, i*16, (vector signed short*)block);
    1.57 +
    1.58 +        pixels += line_size;
    1.59 +    }
    1.60 +}
    1.61 +
    1.62 +static void diff_pixels_altivec(DCTELEM *restrict block, const uint8_t *s1,
    1.63 +        const uint8_t *s2, int stride)
    1.64 +{
    1.65 +    int i;
    1.66 +    vector unsigned char perm, bytes, *pixv;
    1.67 +    const vector unsigned char zero = (const vector unsigned char)vec_splat_u8(0);
    1.68 +    vector signed short shorts1, shorts2;
    1.69 +
    1.70 +    for (i = 0; i < 4; i++) {
    1.71 +        // Read potentially unaligned pixels
    1.72 +        // We're reading 16 pixels, and actually only want 8,
    1.73 +        // but we simply ignore the extras.
    1.74 +        perm = vec_lvsl(0, s1);
    1.75 +        pixv = (vector unsigned char *) s1;
    1.76 +        bytes = vec_perm(pixv[0], pixv[1], perm);
    1.77 +
    1.78 +        // convert the bytes into shorts
    1.79 +        shorts1 = (vector signed short)vec_mergeh(zero, bytes);
    1.80 +
    1.81 +        // Do the same for the second block of pixels
    1.82 +        perm = vec_lvsl(0, s2);
    1.83 +        pixv = (vector unsigned char *) s2;
    1.84 +        bytes = vec_perm(pixv[0], pixv[1], perm);
    1.85 +
    1.86 +        // convert the bytes into shorts
    1.87 +        shorts2 = (vector signed short)vec_mergeh(zero, bytes);
    1.88 +
    1.89 +        // Do the subtraction
    1.90 +        shorts1 = vec_sub(shorts1, shorts2);
    1.91 +
    1.92 +        // save the data to the block, we assume the block is 16-byte aligned
    1.93 +        vec_st(shorts1, 0, (vector signed short*)block);
    1.94 +
    1.95 +        s1 += stride;
    1.96 +        s2 += stride;
    1.97 +        block += 8;
    1.98 +
    1.99 +
   1.100 +        // The code below is a copy of the code above... This is a manual
   1.101 +        // unroll.
   1.102 +
   1.103 +        // Read potentially unaligned pixels
   1.104 +        // We're reading 16 pixels, and actually only want 8,
   1.105 +        // but we simply ignore the extras.
   1.106 +        perm = vec_lvsl(0, s1);
   1.107 +        pixv = (vector unsigned char *) s1;
   1.108 +        bytes = vec_perm(pixv[0], pixv[1], perm);
   1.109 +
   1.110 +        // convert the bytes into shorts
   1.111 +        shorts1 = (vector signed short)vec_mergeh(zero, bytes);
   1.112 +
   1.113 +        // Do the same for the second block of pixels
   1.114 +        perm = vec_lvsl(0, s2);
   1.115 +        pixv = (vector unsigned char *) s2;
   1.116 +        bytes = vec_perm(pixv[0], pixv[1], perm);
   1.117 +
   1.118 +        // convert the bytes into shorts
   1.119 +        shorts2 = (vector signed short)vec_mergeh(zero, bytes);
   1.120 +
   1.121 +        // Do the subtraction
   1.122 +        shorts1 = vec_sub(shorts1, shorts2);
   1.123 +
   1.124 +        // save the data to the block, we assume the block is 16-byte aligned
   1.125 +        vec_st(shorts1, 0, (vector signed short*)block);
   1.126 +
   1.127 +        s1 += stride;
   1.128 +        s2 += stride;
   1.129 +        block += 8;
   1.130 +    }
   1.131 +}
   1.132 +
   1.133 +
   1.134 +static void clear_block_altivec(DCTELEM *block) {
   1.135 +    LOAD_ZERO;
   1.136 +    vec_st(zero_s16v,   0, block);
   1.137 +    vec_st(zero_s16v,  16, block);
   1.138 +    vec_st(zero_s16v,  32, block);
   1.139 +    vec_st(zero_s16v,  48, block);
   1.140 +    vec_st(zero_s16v,  64, block);
   1.141 +    vec_st(zero_s16v,  80, block);
   1.142 +    vec_st(zero_s16v,  96, block);
   1.143 +    vec_st(zero_s16v, 112, block);
   1.144 +}
   1.145 +
   1.146 +
   1.147 +
   1.148 +/* next one assumes that ((line_size % 16) == 0) */
   1.149 +void put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
   1.150 +{
   1.151 +POWERPC_PERF_DECLARE(altivec_put_pixels16_num, 1);
   1.152 +    register vector unsigned char pixelsv1, pixelsv2;
   1.153 +    register vector unsigned char pixelsv1B, pixelsv2B;
   1.154 +    register vector unsigned char pixelsv1C, pixelsv2C;
   1.155 +    register vector unsigned char pixelsv1D, pixelsv2D;
   1.156 +
   1.157 +    register vector unsigned char perm = vec_lvsl(0, pixels);
   1.158 +    int i;
   1.159 +    register int line_size_2 = line_size << 1;
   1.160 +    register int line_size_3 = line_size + line_size_2;
   1.161 +    register int line_size_4 = line_size << 2;
   1.162 +
   1.163 +POWERPC_PERF_START_COUNT(altivec_put_pixels16_num, 1);
   1.164 +// hand-unrolling the loop by 4 gains about 15%
   1.165 +// mininum execution time goes from 74 to 60 cycles
   1.166 +// it's faster than -funroll-loops, but using
   1.167 +// -funroll-loops w/ this is bad - 74 cycles again.
   1.168 +// all this is on a 7450, tuning for the 7450
   1.169 +#if 0
   1.170 +    for (i = 0; i < h; i++) {
   1.171 +        pixelsv1 = vec_ld(0, pixels);
   1.172 +        pixelsv2 = vec_ld(16, pixels);
   1.173 +        vec_st(vec_perm(pixelsv1, pixelsv2, perm),
   1.174 +               0, block);
   1.175 +        pixels+=line_size;
   1.176 +        block +=line_size;
   1.177 +    }
   1.178 +#else
   1.179 +    for (i = 0; i < h; i += 4) {
   1.180 +        pixelsv1  = vec_ld( 0, pixels);
   1.181 +        pixelsv2  = vec_ld(15, pixels);
   1.182 +        pixelsv1B = vec_ld(line_size, pixels);
   1.183 +        pixelsv2B = vec_ld(15 + line_size, pixels);
   1.184 +        pixelsv1C = vec_ld(line_size_2, pixels);
   1.185 +        pixelsv2C = vec_ld(15 + line_size_2, pixels);
   1.186 +        pixelsv1D = vec_ld(line_size_3, pixels);
   1.187 +        pixelsv2D = vec_ld(15 + line_size_3, pixels);
   1.188 +        vec_st(vec_perm(pixelsv1, pixelsv2, perm),
   1.189 +               0, (unsigned char*)block);
   1.190 +        vec_st(vec_perm(pixelsv1B, pixelsv2B, perm),
   1.191 +               line_size, (unsigned char*)block);
   1.192 +        vec_st(vec_perm(pixelsv1C, pixelsv2C, perm),
   1.193 +               line_size_2, (unsigned char*)block);
   1.194 +        vec_st(vec_perm(pixelsv1D, pixelsv2D, perm),
   1.195 +               line_size_3, (unsigned char*)block);
   1.196 +        pixels+=line_size_4;
   1.197 +        block +=line_size_4;
   1.198 +    }
   1.199 +#endif
   1.200 +POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_num, 1);
   1.201 +}
   1.202 +
   1.203 +/* next one assumes that ((line_size % 16) == 0) */
   1.204 +#define op_avg(a,b)  a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
   1.205 +void avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
   1.206 +{
   1.207 +POWERPC_PERF_DECLARE(altivec_avg_pixels16_num, 1);
   1.208 +    register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
   1.209 +    register vector unsigned char perm = vec_lvsl(0, pixels);
   1.210 +    int i;
   1.211 +
   1.212 +POWERPC_PERF_START_COUNT(altivec_avg_pixels16_num, 1);
   1.213 +
   1.214 +    for (i = 0; i < h; i++) {
   1.215 +        pixelsv1 = vec_ld( 0, pixels);
   1.216 +        pixelsv2 = vec_ld(16,pixels);
   1.217 +        blockv = vec_ld(0, block);
   1.218 +        pixelsv = vec_perm(pixelsv1, pixelsv2, perm);
   1.219 +        blockv = vec_avg(blockv,pixelsv);
   1.220 +        vec_st(blockv, 0, (unsigned char*)block);
   1.221 +        pixels+=line_size;
   1.222 +        block +=line_size;
   1.223 +    }
   1.224 +
   1.225 +POWERPC_PERF_STOP_COUNT(altivec_avg_pixels16_num, 1);
   1.226 +}
   1.227 +
   1.228 +/* next one assumes that ((line_size % 8) == 0) */
   1.229 +static void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
   1.230 +{
   1.231 +POWERPC_PERF_DECLARE(altivec_avg_pixels8_num, 1);
   1.232 +    register vector unsigned char pixelsv1, pixelsv2, pixelsv, blockv;
   1.233 +    int i;
   1.234 +
   1.235 +POWERPC_PERF_START_COUNT(altivec_avg_pixels8_num, 1);
   1.236 +
   1.237 +   for (i = 0; i < h; i++) {
   1.238 +       /* block is 8 bytes-aligned, so we're either in the
   1.239 +          left block (16 bytes-aligned) or in the right block (not) */
   1.240 +       int rightside = ((unsigned long)block & 0x0000000F);
   1.241 +
   1.242 +       blockv = vec_ld(0, block);
   1.243 +       pixelsv1 = vec_ld( 0, pixels);
   1.244 +       pixelsv2 = vec_ld(16, pixels);
   1.245 +       pixelsv = vec_perm(pixelsv1, pixelsv2, vec_lvsl(0, pixels));
   1.246 +
   1.247 +       if (rightside) {
   1.248 +           pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
   1.249 +       } else {
   1.250 +           pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
   1.251 +       }
   1.252 +
   1.253 +       blockv = vec_avg(blockv, pixelsv);
   1.254 +
   1.255 +       vec_st(blockv, 0, block);
   1.256 +
   1.257 +       pixels += line_size;
   1.258 +       block += line_size;
   1.259 +   }
   1.260 +
   1.261 +POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_num, 1);
   1.262 +}
   1.263 +
   1.264 +/* next one assumes that ((line_size % 8) == 0) */
   1.265 +static void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
   1.266 +{
   1.267 +POWERPC_PERF_DECLARE(altivec_put_pixels8_xy2_num, 1);
   1.268 +    register int i;
   1.269 +    register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
   1.270 +    register vector unsigned char blockv, temp1, temp2;
   1.271 +    register vector unsigned short pixelssum1, pixelssum2, temp3;
   1.272 +    register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
   1.273 +    register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
   1.274 +
   1.275 +    temp1 = vec_ld(0, pixels);
   1.276 +    temp2 = vec_ld(16, pixels);
   1.277 +    pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
   1.278 +    if ((((unsigned long)pixels) & 0x0000000F) ==  0x0000000F) {
   1.279 +        pixelsv2 = temp2;
   1.280 +    } else {
   1.281 +        pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
   1.282 +    }
   1.283 +    pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.284 +    pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.285 +    pixelssum1 = vec_add((vector unsigned short)pixelsv1,
   1.286 +                         (vector unsigned short)pixelsv2);
   1.287 +    pixelssum1 = vec_add(pixelssum1, vctwo);
   1.288 +
   1.289 +POWERPC_PERF_START_COUNT(altivec_put_pixels8_xy2_num, 1);
   1.290 +    for (i = 0; i < h ; i++) {
   1.291 +        int rightside = ((unsigned long)block & 0x0000000F);
   1.292 +        blockv = vec_ld(0, block);
   1.293 +
   1.294 +        temp1 = vec_ld(line_size, pixels);
   1.295 +        temp2 = vec_ld(line_size + 16, pixels);
   1.296 +        pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
   1.297 +        if (((((unsigned long)pixels) + line_size) & 0x0000000F) ==  0x0000000F) {
   1.298 +            pixelsv2 = temp2;
   1.299 +        } else {
   1.300 +            pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
   1.301 +        }
   1.302 +
   1.303 +        pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.304 +        pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.305 +        pixelssum2 = vec_add((vector unsigned short)pixelsv1,
   1.306 +                             (vector unsigned short)pixelsv2);
   1.307 +        temp3 = vec_add(pixelssum1, pixelssum2);
   1.308 +        temp3 = vec_sra(temp3, vctwo);
   1.309 +        pixelssum1 = vec_add(pixelssum2, vctwo);
   1.310 +        pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
   1.311 +
   1.312 +        if (rightside) {
   1.313 +            blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
   1.314 +        } else {
   1.315 +            blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
   1.316 +        }
   1.317 +
   1.318 +        vec_st(blockv, 0, block);
   1.319 +
   1.320 +        block += line_size;
   1.321 +        pixels += line_size;
   1.322 +    }
   1.323 +
   1.324 +POWERPC_PERF_STOP_COUNT(altivec_put_pixels8_xy2_num, 1);
   1.325 +}
   1.326 +
   1.327 +/* next one assumes that ((line_size % 8) == 0) */
   1.328 +static void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
   1.329 +{
   1.330 +POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels8_xy2_num, 1);
   1.331 +    register int i;
   1.332 +    register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
   1.333 +    register vector unsigned char blockv, temp1, temp2;
   1.334 +    register vector unsigned short pixelssum1, pixelssum2, temp3;
   1.335 +    register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
   1.336 +    register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
   1.337 +    register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
   1.338 +
   1.339 +    temp1 = vec_ld(0, pixels);
   1.340 +    temp2 = vec_ld(16, pixels);
   1.341 +    pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
   1.342 +    if ((((unsigned long)pixels) & 0x0000000F) ==  0x0000000F) {
   1.343 +        pixelsv2 = temp2;
   1.344 +    } else {
   1.345 +        pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
   1.346 +    }
   1.347 +    pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.348 +    pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.349 +    pixelssum1 = vec_add((vector unsigned short)pixelsv1,
   1.350 +                         (vector unsigned short)pixelsv2);
   1.351 +    pixelssum1 = vec_add(pixelssum1, vcone);
   1.352 +
   1.353 +POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
   1.354 +    for (i = 0; i < h ; i++) {
   1.355 +        int rightside = ((unsigned long)block & 0x0000000F);
   1.356 +        blockv = vec_ld(0, block);
   1.357 +
   1.358 +        temp1 = vec_ld(line_size, pixels);
   1.359 +        temp2 = vec_ld(line_size + 16, pixels);
   1.360 +        pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
   1.361 +        if (((((unsigned long)pixels) + line_size) & 0x0000000F) ==  0x0000000F) {
   1.362 +            pixelsv2 = temp2;
   1.363 +        } else {
   1.364 +            pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
   1.365 +        }
   1.366 +
   1.367 +        pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.368 +        pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.369 +        pixelssum2 = vec_add((vector unsigned short)pixelsv1,
   1.370 +                             (vector unsigned short)pixelsv2);
   1.371 +        temp3 = vec_add(pixelssum1, pixelssum2);
   1.372 +        temp3 = vec_sra(temp3, vctwo);
   1.373 +        pixelssum1 = vec_add(pixelssum2, vcone);
   1.374 +        pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
   1.375 +
   1.376 +        if (rightside) {
   1.377 +            blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
   1.378 +        } else {
   1.379 +            blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
   1.380 +        }
   1.381 +
   1.382 +        vec_st(blockv, 0, block);
   1.383 +
   1.384 +        block += line_size;
   1.385 +        pixels += line_size;
   1.386 +    }
   1.387 +
   1.388 +POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels8_xy2_num, 1);
   1.389 +}
   1.390 +
   1.391 +/* next one assumes that ((line_size % 16) == 0) */
   1.392 +static void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
   1.393 +{
   1.394 +POWERPC_PERF_DECLARE(altivec_put_pixels16_xy2_num, 1);
   1.395 +    register int i;
   1.396 +    register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
   1.397 +    register vector unsigned char blockv, temp1, temp2;
   1.398 +    register vector unsigned short temp3, temp4,
   1.399 +        pixelssum1, pixelssum2, pixelssum3, pixelssum4;
   1.400 +    register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
   1.401 +    register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
   1.402 +
   1.403 +POWERPC_PERF_START_COUNT(altivec_put_pixels16_xy2_num, 1);
   1.404 +
   1.405 +    temp1 = vec_ld(0, pixels);
   1.406 +    temp2 = vec_ld(16, pixels);
   1.407 +    pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
   1.408 +    if ((((unsigned long)pixels) & 0x0000000F) ==  0x0000000F) {
   1.409 +        pixelsv2 = temp2;
   1.410 +    } else {
   1.411 +        pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
   1.412 +    }
   1.413 +    pixelsv3 = vec_mergel(vczero, pixelsv1);
   1.414 +    pixelsv4 = vec_mergel(vczero, pixelsv2);
   1.415 +    pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.416 +    pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.417 +    pixelssum3 = vec_add((vector unsigned short)pixelsv3,
   1.418 +                         (vector unsigned short)pixelsv4);
   1.419 +    pixelssum3 = vec_add(pixelssum3, vctwo);
   1.420 +    pixelssum1 = vec_add((vector unsigned short)pixelsv1,
   1.421 +                         (vector unsigned short)pixelsv2);
   1.422 +    pixelssum1 = vec_add(pixelssum1, vctwo);
   1.423 +
   1.424 +    for (i = 0; i < h ; i++) {
   1.425 +        blockv = vec_ld(0, block);
   1.426 +
   1.427 +        temp1 = vec_ld(line_size, pixels);
   1.428 +        temp2 = vec_ld(line_size + 16, pixels);
   1.429 +        pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
   1.430 +        if (((((unsigned long)pixels) + line_size) & 0x0000000F) ==  0x0000000F) {
   1.431 +            pixelsv2 = temp2;
   1.432 +        } else {
   1.433 +            pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
   1.434 +        }
   1.435 +
   1.436 +        pixelsv3 = vec_mergel(vczero, pixelsv1);
   1.437 +        pixelsv4 = vec_mergel(vczero, pixelsv2);
   1.438 +        pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.439 +        pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.440 +
   1.441 +        pixelssum4 = vec_add((vector unsigned short)pixelsv3,
   1.442 +                             (vector unsigned short)pixelsv4);
   1.443 +        pixelssum2 = vec_add((vector unsigned short)pixelsv1,
   1.444 +                             (vector unsigned short)pixelsv2);
   1.445 +        temp4 = vec_add(pixelssum3, pixelssum4);
   1.446 +        temp4 = vec_sra(temp4, vctwo);
   1.447 +        temp3 = vec_add(pixelssum1, pixelssum2);
   1.448 +        temp3 = vec_sra(temp3, vctwo);
   1.449 +
   1.450 +        pixelssum3 = vec_add(pixelssum4, vctwo);
   1.451 +        pixelssum1 = vec_add(pixelssum2, vctwo);
   1.452 +
   1.453 +        blockv = vec_packsu(temp3, temp4);
   1.454 +
   1.455 +        vec_st(blockv, 0, block);
   1.456 +
   1.457 +        block += line_size;
   1.458 +        pixels += line_size;
   1.459 +    }
   1.460 +
   1.461 +POWERPC_PERF_STOP_COUNT(altivec_put_pixels16_xy2_num, 1);
   1.462 +}
   1.463 +
   1.464 +/* next one assumes that ((line_size % 16) == 0) */
   1.465 +static void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, int line_size, int h)
   1.466 +{
   1.467 +POWERPC_PERF_DECLARE(altivec_put_no_rnd_pixels16_xy2_num, 1);
   1.468 +    register int i;
   1.469 +    register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
   1.470 +    register vector unsigned char blockv, temp1, temp2;
   1.471 +    register vector unsigned short temp3, temp4,
   1.472 +        pixelssum1, pixelssum2, pixelssum3, pixelssum4;
   1.473 +    register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
   1.474 +    register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
   1.475 +    register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
   1.476 +
   1.477 +POWERPC_PERF_START_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
   1.478 +
   1.479 +    temp1 = vec_ld(0, pixels);
   1.480 +    temp2 = vec_ld(16, pixels);
   1.481 +    pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
   1.482 +    if ((((unsigned long)pixels) & 0x0000000F) ==  0x0000000F) {
   1.483 +        pixelsv2 = temp2;
   1.484 +    } else {
   1.485 +        pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
   1.486 +    }
   1.487 +    pixelsv3 = vec_mergel(vczero, pixelsv1);
   1.488 +    pixelsv4 = vec_mergel(vczero, pixelsv2);
   1.489 +    pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.490 +    pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.491 +    pixelssum3 = vec_add((vector unsigned short)pixelsv3,
   1.492 +                         (vector unsigned short)pixelsv4);
   1.493 +    pixelssum3 = vec_add(pixelssum3, vcone);
   1.494 +    pixelssum1 = vec_add((vector unsigned short)pixelsv1,
   1.495 +                         (vector unsigned short)pixelsv2);
   1.496 +    pixelssum1 = vec_add(pixelssum1, vcone);
   1.497 +
   1.498 +    for (i = 0; i < h ; i++) {
   1.499 +        blockv = vec_ld(0, block);
   1.500 +
   1.501 +        temp1 = vec_ld(line_size, pixels);
   1.502 +        temp2 = vec_ld(line_size + 16, pixels);
   1.503 +        pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
   1.504 +        if (((((unsigned long)pixels) + line_size) & 0x0000000F) ==  0x0000000F) {
   1.505 +            pixelsv2 = temp2;
   1.506 +        } else {
   1.507 +            pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
   1.508 +        }
   1.509 +
   1.510 +        pixelsv3 = vec_mergel(vczero, pixelsv1);
   1.511 +        pixelsv4 = vec_mergel(vczero, pixelsv2);
   1.512 +        pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.513 +        pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.514 +
   1.515 +        pixelssum4 = vec_add((vector unsigned short)pixelsv3,
   1.516 +                             (vector unsigned short)pixelsv4);
   1.517 +        pixelssum2 = vec_add((vector unsigned short)pixelsv1,
   1.518 +                             (vector unsigned short)pixelsv2);
   1.519 +        temp4 = vec_add(pixelssum3, pixelssum4);
   1.520 +        temp4 = vec_sra(temp4, vctwo);
   1.521 +        temp3 = vec_add(pixelssum1, pixelssum2);
   1.522 +        temp3 = vec_sra(temp3, vctwo);
   1.523 +
   1.524 +        pixelssum3 = vec_add(pixelssum4, vcone);
   1.525 +        pixelssum1 = vec_add(pixelssum2, vcone);
   1.526 +
   1.527 +        blockv = vec_packsu(temp3, temp4);
   1.528 +
   1.529 +        vec_st(blockv, 0, block);
   1.530 +
   1.531 +        block += line_size;
   1.532 +        pixels += line_size;
   1.533 +    }
   1.534 +
   1.535 +POWERPC_PERF_STOP_COUNT(altivec_put_no_rnd_pixels16_xy2_num, 1);
   1.536 +}
   1.537 +
   1.538 +/* next one assumes that ((line_size % 8) == 0) */
   1.539 +static void avg_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, int line_size, int h)
   1.540 +{
   1.541 +POWERPC_PERF_DECLARE(altivec_avg_pixels8_xy2_num, 1);
   1.542 +    register int i;
   1.543 +    register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
   1.544 +    register vector unsigned char blockv, temp1, temp2, blocktemp;
   1.545 +    register vector unsigned short pixelssum1, pixelssum2, temp3;
   1.546 +
   1.547 +    register const vector unsigned char vczero = (const vector unsigned char)
   1.548 +                                        vec_splat_u8(0);
   1.549 +    register const vector unsigned short vctwo = (const vector unsigned short)
   1.550 +                                        vec_splat_u16(2);
   1.551 +
   1.552 +    temp1 = vec_ld(0, pixels);
   1.553 +    temp2 = vec_ld(16, pixels);
   1.554 +    pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(0, pixels));
   1.555 +    if ((((unsigned long)pixels) & 0x0000000F) ==  0x0000000F) {
   1.556 +        pixelsv2 = temp2;
   1.557 +    } else {
   1.558 +        pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(1, pixels));
   1.559 +    }
   1.560 +    pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.561 +    pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.562 +    pixelssum1 = vec_add((vector unsigned short)pixelsv1,
   1.563 +                         (vector unsigned short)pixelsv2);
   1.564 +    pixelssum1 = vec_add(pixelssum1, vctwo);
   1.565 +
   1.566 +POWERPC_PERF_START_COUNT(altivec_avg_pixels8_xy2_num, 1);
   1.567 +    for (i = 0; i < h ; i++) {
   1.568 +        int rightside = ((unsigned long)block & 0x0000000F);
   1.569 +        blockv = vec_ld(0, block);
   1.570 +
   1.571 +        temp1 = vec_ld(line_size, pixels);
   1.572 +        temp2 = vec_ld(line_size + 16, pixels);
   1.573 +        pixelsv1 = vec_perm(temp1, temp2, vec_lvsl(line_size, pixels));
   1.574 +        if (((((unsigned long)pixels) + line_size) & 0x0000000F) ==  0x0000000F) {
   1.575 +            pixelsv2 = temp2;
   1.576 +        } else {
   1.577 +            pixelsv2 = vec_perm(temp1, temp2, vec_lvsl(line_size + 1, pixels));
   1.578 +        }
   1.579 +
   1.580 +        pixelsv1 = vec_mergeh(vczero, pixelsv1);
   1.581 +        pixelsv2 = vec_mergeh(vczero, pixelsv2);
   1.582 +        pixelssum2 = vec_add((vector unsigned short)pixelsv1,
   1.583 +                             (vector unsigned short)pixelsv2);
   1.584 +        temp3 = vec_add(pixelssum1, pixelssum2);
   1.585 +        temp3 = vec_sra(temp3, vctwo);
   1.586 +        pixelssum1 = vec_add(pixelssum2, vctwo);
   1.587 +        pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
   1.588 +
   1.589 +        if (rightside) {
   1.590 +            blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
   1.591 +        } else {
   1.592 +            blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
   1.593 +        }
   1.594 +
   1.595 +        blockv = vec_avg(blocktemp, blockv);
   1.596 +        vec_st(blockv, 0, block);
   1.597 +
   1.598 +        block += line_size;
   1.599 +        pixels += line_size;
   1.600 +    }
   1.601 +
   1.602 +POWERPC_PERF_STOP_COUNT(altivec_avg_pixels8_xy2_num, 1);
   1.603 +}
   1.604 +
   1.605 +void dsputil_init_altivec(DSPContext* c)
   1.606 +{
   1.607 +    c->diff_pixels = diff_pixels_altivec;
   1.608 +    c->get_pixels = get_pixels_altivec;
   1.609 +    c->clear_block = clear_block_altivec;
   1.610 +
   1.611 +    c->put_pixels_tab[0][0] = put_pixels16_altivec;
   1.612 +    /* the two functions do the same thing, so use the same code */
   1.613 +    c->put_no_rnd_pixels_tab[0][0] = put_pixels16_altivec;
   1.614 +    c->avg_pixels_tab[0][0] = avg_pixels16_altivec;
   1.615 +    c->avg_pixels_tab[1][0] = avg_pixels8_altivec;
   1.616 +    c->avg_pixels_tab[1][3] = avg_pixels8_xy2_altivec;
   1.617 +    c->put_pixels_tab[1][3] = put_pixels8_xy2_altivec;
   1.618 +    c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_altivec;
   1.619 +    c->put_pixels_tab[0][3] = put_pixels16_xy2_altivec;
   1.620 +    c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_altivec;
   1.621 +
   1.622 +}