| 1 |
/* filter_vsx_intrinsics.c - PowerPC optimised filter functions |
| 2 |
* |
| 3 |
* Copyright (c) 2018 Cosmin Truta |
| 4 |
* Copyright (c) 2017 Glenn Randers-Pehrson |
| 5 |
* Written by Vadim Barkov, 2017. |
| 6 |
* |
| 7 |
* This code is released under the libpng license. |
| 8 |
* For conditions of distribution and use, see the disclaimer |
| 9 |
* and license in png.h |
| 10 |
*/ |
| 11 |
|
| 12 |
#include <stdio.h> |
| 13 |
#include <stdint.h> |
| 14 |
#include "../pngpriv.h" |
| 15 |
|
| 16 |
#ifdef PNG_READ_SUPPORTED |
| 17 |
|
| 18 |
/* This code requires -maltivec and -mvsx on the command line: */ |
| 19 |
#if PNG_POWERPC_VSX_IMPLEMENTATION == 1 /* intrinsics code from pngpriv.h */ |
| 20 |
|
| 21 |
#include <altivec.h> |
| 22 |
|
| 23 |
#if PNG_POWERPC_VSX_OPT > 0 |
| 24 |
|
| 25 |
#ifndef __VSX__ |
| 26 |
# error "This code requires VSX support (POWER7 and later). Please provide -mvsx compiler flag." |
| 27 |
#endif |
| 28 |
|
| 29 |
#define vec_ld_unaligned(vec,data) vec = vec_vsx_ld(0,data) |
| 30 |
#define vec_st_unaligned(vec,data) vec_vsx_st(vec,0,data) |
| 31 |
|
| 32 |
|
| 33 |
/* Functions in this file look at most 3 pixels (a,b,c) to predict the 4th (d). |
| 34 |
* They're positioned like this: |
| 35 |
* prev: c b |
| 36 |
* row: a d |
| 37 |
* The Sub filter predicts d=a, Avg d=(a+b)/2, and Paeth predicts d to be |
| 38 |
* whichever of a, b, or c is closest to p=a+b-c. |
| 39 |
* ( this is taken from ../intel/filter_sse2_intrinsics.c ) |
| 40 |
*/ |
| 41 |
|
| 42 |
#define vsx_declare_common_vars(row_info,row,prev_row,offset) \ |
| 43 |
png_byte i;\ |
| 44 |
png_bytep rp = row + offset;\ |
| 45 |
png_const_bytep pp = prev_row;\ |
| 46 |
size_t unaligned_top = 16 - (((size_t)rp % 16));\ |
| 47 |
size_t istop;\ |
| 48 |
if(unaligned_top == 16)\ |
| 49 |
unaligned_top = 0;\ |
| 50 |
istop = row_info->rowbytes;\ |
| 51 |
if((unaligned_top < istop))\ |
| 52 |
istop -= unaligned_top;\ |
| 53 |
else{\ |
| 54 |
unaligned_top = istop;\ |
| 55 |
istop = 0;\ |
| 56 |
} |
| 57 |
|
| 58 |
void png_read_filter_row_up_vsx(png_row_infop row_info, png_bytep row, |
| 59 |
png_const_bytep prev_row) |
| 60 |
{ |
| 61 |
vector unsigned char rp_vec; |
| 62 |
vector unsigned char pp_vec; |
| 63 |
vsx_declare_common_vars(row_info,row,prev_row,0) |
| 64 |
|
| 65 |
/* Altivec operations require 16-byte aligned data |
| 66 |
* but input can be unaligned. So we calculate |
| 67 |
* unaligned part as usual. |
| 68 |
*/ |
| 69 |
for (i = 0; i < unaligned_top; i++) |
| 70 |
{ |
| 71 |
*rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff); |
| 72 |
rp++; |
| 73 |
} |
| 74 |
|
| 75 |
/* Using SIMD while we can */ |
| 76 |
while( istop >= 16 ) |
| 77 |
{ |
| 78 |
rp_vec = vec_ld(0,rp); |
| 79 |
vec_ld_unaligned(pp_vec,pp); |
| 80 |
|
| 81 |
rp_vec = vec_add(rp_vec,pp_vec); |
| 82 |
|
| 83 |
vec_st(rp_vec,0,rp); |
| 84 |
|
| 85 |
pp += 16; |
| 86 |
rp += 16; |
| 87 |
istop -= 16; |
| 88 |
} |
| 89 |
|
| 90 |
if(istop > 0) |
| 91 |
{ |
| 92 |
/* If byte count of row is not divisible by 16 |
| 93 |
* we will process remaining part as usual |
| 94 |
*/ |
| 95 |
for (i = 0; i < istop; i++) |
| 96 |
{ |
| 97 |
*rp = (png_byte)(((int)(*rp) + (int)(*pp++)) & 0xff); |
| 98 |
rp++; |
| 99 |
} |
| 100 |
} |
| 101 |
|
| 102 |
} |
| 103 |
|
| 104 |
static const vector unsigned char VSX_LEFTSHIFTED1_4 = {16,16,16,16, 0, 1, 2, 3,16,16,16,16,16,16,16,16}; |
| 105 |
static const vector unsigned char VSX_LEFTSHIFTED2_4 = {16,16,16,16,16,16,16,16, 4, 5, 6, 7,16,16,16,16}; |
| 106 |
static const vector unsigned char VSX_LEFTSHIFTED3_4 = {16,16,16,16,16,16,16,16,16,16,16,16, 8, 9,10,11}; |
| 107 |
|
| 108 |
static const vector unsigned char VSX_LEFTSHIFTED1_3 = {16,16,16, 0, 1, 2,16,16,16,16,16,16,16,16,16,16}; |
| 109 |
static const vector unsigned char VSX_LEFTSHIFTED2_3 = {16,16,16,16,16,16, 3, 4, 5,16,16,16,16,16,16,16}; |
| 110 |
static const vector unsigned char VSX_LEFTSHIFTED3_3 = {16,16,16,16,16,16,16,16,16, 6, 7, 8,16,16,16,16}; |
| 111 |
static const vector unsigned char VSX_LEFTSHIFTED4_3 = {16,16,16,16,16,16,16,16,16,16,16,16, 9,10,11,16}; |
| 112 |
|
| 113 |
static const vector unsigned char VSX_NOT_SHIFTED1_4 = {16,16,16,16, 4, 5, 6, 7,16,16,16,16,16,16,16,16}; |
| 114 |
static const vector unsigned char VSX_NOT_SHIFTED2_4 = {16,16,16,16,16,16,16,16, 8, 9,10,11,16,16,16,16}; |
| 115 |
static const vector unsigned char VSX_NOT_SHIFTED3_4 = {16,16,16,16,16,16,16,16,16,16,16,16,12,13,14,15}; |
| 116 |
|
| 117 |
static const vector unsigned char VSX_NOT_SHIFTED1_3 = {16,16,16, 3, 4, 5,16,16,16,16,16,16,16,16,16,16}; |
| 118 |
static const vector unsigned char VSX_NOT_SHIFTED2_3 = {16,16,16,16,16,16, 6, 7, 8,16,16,16,16,16,16,16}; |
| 119 |
static const vector unsigned char VSX_NOT_SHIFTED3_3 = {16,16,16,16,16,16,16,16,16, 9,10,11,16,16,16,16}; |
| 120 |
static const vector unsigned char VSX_NOT_SHIFTED4_3 = {16,16,16,16,16,16,16,16,16,16,16,16,12,13,14,16}; |
| 121 |
|
| 122 |
static const vector unsigned char VSX_CHAR_ZERO = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0}; |
| 123 |
#ifdef __LITTLE_ENDIAN__ |
| 124 |
|
| 125 |
static const vector unsigned char VSX_CHAR_TO_SHORT1_4 = { 4,16, 5,16, 6,16, 7,16,16,16,16,16,16,16,16,16}; |
| 126 |
static const vector unsigned char VSX_CHAR_TO_SHORT2_4 = { 8,16, 9,16,10,16,11,16,16,16,16,16,16,16,16,16}; |
| 127 |
static const vector unsigned char VSX_CHAR_TO_SHORT3_4 = {12,16,13,16,14,16,15,16,16,16,16,16,16,16,16,16}; |
| 128 |
|
| 129 |
static const vector unsigned char VSX_SHORT_TO_CHAR1_4 = {16,16,16,16, 0, 2, 4, 6,16,16,16,16,16,16,16,16}; |
| 130 |
static const vector unsigned char VSX_SHORT_TO_CHAR2_4 = {16,16,16,16,16,16,16,16, 0, 2, 4, 6,16,16,16,16}; |
| 131 |
static const vector unsigned char VSX_SHORT_TO_CHAR3_4 = {16,16,16,16,16,16,16,16,16,16,16,16, 0, 2, 4, 6}; |
| 132 |
|
| 133 |
static const vector unsigned char VSX_CHAR_TO_SHORT1_3 = { 3,16, 4,16, 5,16,16,16,16,16,16,16,16,16,16,16}; |
| 134 |
static const vector unsigned char VSX_CHAR_TO_SHORT2_3 = { 6,16, 7,16, 8,16,16,16,16,16,16,16,16,16,16,16}; |
| 135 |
static const vector unsigned char VSX_CHAR_TO_SHORT3_3 = { 9,16,10,16,11,16,16,16,16,16,16,16,16,16,16,16}; |
| 136 |
static const vector unsigned char VSX_CHAR_TO_SHORT4_3 = {12,16,13,16,14,16,16,16,16,16,16,16,16,16,16,16}; |
| 137 |
|
| 138 |
static const vector unsigned char VSX_SHORT_TO_CHAR1_3 = {16,16,16, 0, 2, 4,16,16,16,16,16,16,16,16,16,16}; |
| 139 |
static const vector unsigned char VSX_SHORT_TO_CHAR2_3 = {16,16,16,16,16,16, 0, 2, 4,16,16,16,16,16,16,16}; |
| 140 |
static const vector unsigned char VSX_SHORT_TO_CHAR3_3 = {16,16,16,16,16,16,16,16,16, 0, 2, 4,16,16,16,16}; |
| 141 |
static const vector unsigned char VSX_SHORT_TO_CHAR4_3 = {16,16,16,16,16,16,16,16,16,16,16,16, 0, 2, 4,16}; |
| 142 |
|
| 143 |
#elif defined(__BIG_ENDIAN__) |
| 144 |
|
| 145 |
static const vector unsigned char VSX_CHAR_TO_SHORT1_4 = {16, 4,16, 5,16, 6,16, 7,16,16,16,16,16,16,16,16}; |
| 146 |
static const vector unsigned char VSX_CHAR_TO_SHORT2_4 = {16, 8,16, 9,16,10,16,11,16,16,16,16,16,16,16,16}; |
| 147 |
static const vector unsigned char VSX_CHAR_TO_SHORT3_4 = {16,12,16,13,16,14,16,15,16,16,16,16,16,16,16,16}; |
| 148 |
|
| 149 |
static const vector unsigned char VSX_SHORT_TO_CHAR1_4 = {16,16,16,16, 1, 3, 5, 7,16,16,16,16,16,16,16,16}; |
| 150 |
static const vector unsigned char VSX_SHORT_TO_CHAR2_4 = {16,16,16,16,16,16,16,16, 1, 3, 5, 7,16,16,16,16}; |
| 151 |
static const vector unsigned char VSX_SHORT_TO_CHAR3_4 = {16,16,16,16,16,16,16,16,16,16,16,16, 1, 3, 5, 7}; |
| 152 |
|
| 153 |
static const vector unsigned char VSX_CHAR_TO_SHORT1_3 = {16, 3,16, 4,16, 5,16,16,16,16,16,16,16,16,16,16}; |
| 154 |
static const vector unsigned char VSX_CHAR_TO_SHORT2_3 = {16, 6,16, 7,16, 8,16,16,16,16,16,16,16,16,16,16}; |
| 155 |
static const vector unsigned char VSX_CHAR_TO_SHORT3_3 = {16, 9,16,10,16,11,16,16,16,16,16,16,16,16,16,16}; |
| 156 |
static const vector unsigned char VSX_CHAR_TO_SHORT4_3 = {16,12,16,13,16,14,16,16,16,16,16,16,16,16,16,16}; |
| 157 |
|
| 158 |
static const vector unsigned char VSX_SHORT_TO_CHAR1_3 = {16,16,16, 1, 3, 5,16,16,16,16,16,16,16,16,16,16}; |
| 159 |
static const vector unsigned char VSX_SHORT_TO_CHAR2_3 = {16,16,16,16,16,16, 1, 3, 5,16,16,16,16,16,16,16}; |
| 160 |
static const vector unsigned char VSX_SHORT_TO_CHAR3_3 = {16,16,16,16,16,16,16,16,16, 1, 3, 5,16,16,16,16}; |
| 161 |
static const vector unsigned char VSX_SHORT_TO_CHAR4_3 = {16,16,16,16,16,16,16,16,16,16,16,16, 1, 3, 5,16}; |
| 162 |
|
| 163 |
#endif |
| 164 |
|
| 165 |
#define vsx_char_to_short(vec,offset,bpp) (vector unsigned short)vec_perm((vec),VSX_CHAR_ZERO,VSX_CHAR_TO_SHORT##offset##_##bpp) |
| 166 |
#define vsx_short_to_char(vec,offset,bpp) vec_perm(((vector unsigned char)(vec)),VSX_CHAR_ZERO,VSX_SHORT_TO_CHAR##offset##_##bpp) |
| 167 |
|
| 168 |
#ifdef PNG_USE_ABS |
| 169 |
# define vsx_abs(number) abs(number) |
| 170 |
#else |
| 171 |
# define vsx_abs(number) (number > 0) ? (number) : -(number) |
| 172 |
#endif |
| 173 |
|
| 174 |
void png_read_filter_row_sub4_vsx(png_row_infop row_info, png_bytep row, |
| 175 |
png_const_bytep prev_row) |
| 176 |
{ |
| 177 |
png_byte bpp = 4; |
| 178 |
|
| 179 |
vector unsigned char rp_vec; |
| 180 |
vector unsigned char part_vec; |
| 181 |
|
| 182 |
vsx_declare_common_vars(row_info,row,prev_row,bpp) |
| 183 |
|
| 184 |
PNG_UNUSED(pp) |
| 185 |
|
| 186 |
/* Altivec operations require 16-byte aligned data |
| 187 |
* but input can be unaligned. So we calculate |
| 188 |
* unaligned part as usual. |
| 189 |
*/ |
| 190 |
for (i = 0; i < unaligned_top; i++) |
| 191 |
{ |
| 192 |
*rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff); |
| 193 |
rp++; |
| 194 |
} |
| 195 |
|
| 196 |
/* Using SIMD while we can */ |
| 197 |
while( istop >= 16 ) |
| 198 |
{ |
| 199 |
for(i=0;i < bpp ; i++) |
| 200 |
{ |
| 201 |
*rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff); |
| 202 |
rp++; |
| 203 |
} |
| 204 |
rp -= bpp; |
| 205 |
|
| 206 |
rp_vec = vec_ld(0,rp); |
| 207 |
part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED1_4); |
| 208 |
rp_vec = vec_add(rp_vec,part_vec); |
| 209 |
|
| 210 |
part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED2_4); |
| 211 |
rp_vec = vec_add(rp_vec,part_vec); |
| 212 |
|
| 213 |
part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED3_4); |
| 214 |
rp_vec = vec_add(rp_vec,part_vec); |
| 215 |
|
| 216 |
vec_st(rp_vec,0,rp); |
| 217 |
|
| 218 |
rp += 16; |
| 219 |
istop -= 16; |
| 220 |
} |
| 221 |
|
| 222 |
if(istop > 0) |
| 223 |
for (i = 0; i < istop % 16; i++) |
| 224 |
{ |
| 225 |
*rp = (png_byte)(((int)(*rp) + (int)(*(rp - bpp))) & 0xff); |
| 226 |
rp++; |
| 227 |
} |
| 228 |
|
| 229 |
} |
| 230 |
|
| 231 |
void png_read_filter_row_sub3_vsx(png_row_infop row_info, png_bytep row, |
| 232 |
png_const_bytep prev_row) |
| 233 |
{ |
| 234 |
png_byte bpp = 3; |
| 235 |
|
| 236 |
vector unsigned char rp_vec; |
| 237 |
vector unsigned char part_vec; |
| 238 |
|
| 239 |
vsx_declare_common_vars(row_info,row,prev_row,bpp) |
| 240 |
|
| 241 |
PNG_UNUSED(pp) |
| 242 |
|
| 243 |
/* Altivec operations require 16-byte aligned data |
| 244 |
* but input can be unaligned. So we calculate |
| 245 |
* unaligned part as usual. |
| 246 |
*/ |
| 247 |
for (i = 0; i < unaligned_top; i++) |
| 248 |
{ |
| 249 |
*rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff); |
| 250 |
rp++; |
| 251 |
} |
| 252 |
|
| 253 |
/* Using SIMD while we can */ |
| 254 |
while( istop >= 16 ) |
| 255 |
{ |
| 256 |
for(i=0;i < bpp ; i++) |
| 257 |
{ |
| 258 |
*rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff); |
| 259 |
rp++; |
| 260 |
} |
| 261 |
rp -= bpp; |
| 262 |
|
| 263 |
rp_vec = vec_ld(0,rp); |
| 264 |
part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED1_3); |
| 265 |
rp_vec = vec_add(rp_vec,part_vec); |
| 266 |
|
| 267 |
part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED2_3); |
| 268 |
rp_vec = vec_add(rp_vec,part_vec); |
| 269 |
|
| 270 |
part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED3_3); |
| 271 |
rp_vec = vec_add(rp_vec,part_vec); |
| 272 |
|
| 273 |
part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED4_3); |
| 274 |
rp_vec = vec_add(rp_vec,part_vec); |
| 275 |
|
| 276 |
vec_st(rp_vec,0,rp); |
| 277 |
rp += 15; |
| 278 |
istop -= 16; |
| 279 |
|
| 280 |
/* Since 16 % bpp = 16 % 3 = 1, last element of array must |
| 281 |
* be proceeded manually |
| 282 |
*/ |
| 283 |
*rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff); |
| 284 |
rp++; |
| 285 |
} |
| 286 |
|
| 287 |
if(istop > 0) |
| 288 |
for (i = 0; i < istop % 16; i++) |
| 289 |
{ |
| 290 |
*rp = (png_byte)(((int)(*rp) + (int)(*(rp-bpp))) & 0xff); |
| 291 |
rp++; |
| 292 |
} |
| 293 |
} |
| 294 |
|
| 295 |
void png_read_filter_row_avg4_vsx(png_row_infop row_info, png_bytep row, |
| 296 |
png_const_bytep prev_row) |
| 297 |
{ |
| 298 |
png_byte bpp = 4; |
| 299 |
|
| 300 |
vector unsigned char rp_vec; |
| 301 |
vector unsigned char pp_vec; |
| 302 |
vector unsigned char pp_part_vec; |
| 303 |
vector unsigned char rp_part_vec; |
| 304 |
vector unsigned char avg_vec; |
| 305 |
|
| 306 |
vsx_declare_common_vars(row_info,row,prev_row,bpp) |
| 307 |
rp -= bpp; |
| 308 |
if(istop >= bpp) |
| 309 |
istop -= bpp; |
| 310 |
|
| 311 |
for (i = 0; i < bpp; i++) |
| 312 |
{ |
| 313 |
*rp = (png_byte)(((int)(*rp) + |
| 314 |
((int)(*pp++) / 2 )) & 0xff); |
| 315 |
|
| 316 |
rp++; |
| 317 |
} |
| 318 |
|
| 319 |
/* Altivec operations require 16-byte aligned data |
| 320 |
* but input can be unaligned. So we calculate |
| 321 |
* unaligned part as usual. |
| 322 |
*/ |
| 323 |
for (i = 0; i < unaligned_top; i++) |
| 324 |
{ |
| 325 |
*rp = (png_byte)(((int)(*rp) + |
| 326 |
(int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff); |
| 327 |
|
| 328 |
rp++; |
| 329 |
} |
| 330 |
|
| 331 |
/* Using SIMD while we can */ |
| 332 |
while( istop >= 16 ) |
| 333 |
{ |
| 334 |
for(i=0;i < bpp ; i++) |
| 335 |
{ |
| 336 |
*rp = (png_byte)(((int)(*rp) + |
| 337 |
(int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff); |
| 338 |
|
| 339 |
rp++; |
| 340 |
} |
| 341 |
rp -= bpp; |
| 342 |
pp -= bpp; |
| 343 |
|
| 344 |
vec_ld_unaligned(pp_vec,pp); |
| 345 |
rp_vec = vec_ld(0,rp); |
| 346 |
|
| 347 |
rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED1_4); |
| 348 |
pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED1_4); |
| 349 |
avg_vec = vec_avg(rp_part_vec,pp_part_vec); |
| 350 |
avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1))); |
| 351 |
rp_vec = vec_add(rp_vec,avg_vec); |
| 352 |
|
| 353 |
rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED2_4); |
| 354 |
pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED2_4); |
| 355 |
avg_vec = vec_avg(rp_part_vec,pp_part_vec); |
| 356 |
avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1))); |
| 357 |
rp_vec = vec_add(rp_vec,avg_vec); |
| 358 |
|
| 359 |
rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED3_4); |
| 360 |
pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED3_4); |
| 361 |
avg_vec = vec_avg(rp_part_vec,pp_part_vec); |
| 362 |
avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1))); |
| 363 |
rp_vec = vec_add(rp_vec,avg_vec); |
| 364 |
|
| 365 |
vec_st(rp_vec,0,rp); |
| 366 |
|
| 367 |
rp += 16; |
| 368 |
pp += 16; |
| 369 |
istop -= 16; |
| 370 |
} |
| 371 |
|
| 372 |
if(istop > 0) |
| 373 |
for (i = 0; i < istop % 16; i++) |
| 374 |
{ |
| 375 |
*rp = (png_byte)(((int)(*rp) + |
| 376 |
(int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff); |
| 377 |
|
| 378 |
rp++; |
| 379 |
} |
| 380 |
} |
| 381 |
|
| 382 |
void png_read_filter_row_avg3_vsx(png_row_infop row_info, png_bytep row, |
| 383 |
png_const_bytep prev_row) |
| 384 |
{ |
| 385 |
png_byte bpp = 3; |
| 386 |
|
| 387 |
vector unsigned char rp_vec; |
| 388 |
vector unsigned char pp_vec; |
| 389 |
vector unsigned char pp_part_vec; |
| 390 |
vector unsigned char rp_part_vec; |
| 391 |
vector unsigned char avg_vec; |
| 392 |
|
| 393 |
vsx_declare_common_vars(row_info,row,prev_row,bpp) |
| 394 |
rp -= bpp; |
| 395 |
if(istop >= bpp) |
| 396 |
istop -= bpp; |
| 397 |
|
| 398 |
for (i = 0; i < bpp; i++) |
| 399 |
{ |
| 400 |
*rp = (png_byte)(((int)(*rp) + |
| 401 |
((int)(*pp++) / 2 )) & 0xff); |
| 402 |
|
| 403 |
rp++; |
| 404 |
} |
| 405 |
|
| 406 |
/* Altivec operations require 16-byte aligned data |
| 407 |
* but input can be unaligned. So we calculate |
| 408 |
* unaligned part as usual. |
| 409 |
*/ |
| 410 |
for (i = 0; i < unaligned_top; i++) |
| 411 |
{ |
| 412 |
*rp = (png_byte)(((int)(*rp) + |
| 413 |
(int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff); |
| 414 |
|
| 415 |
rp++; |
| 416 |
} |
| 417 |
|
| 418 |
/* Using SIMD while we can */ |
| 419 |
while( istop >= 16 ) |
| 420 |
{ |
| 421 |
for(i=0;i < bpp ; i++) |
| 422 |
{ |
| 423 |
*rp = (png_byte)(((int)(*rp) + |
| 424 |
(int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff); |
| 425 |
|
| 426 |
rp++; |
| 427 |
} |
| 428 |
rp -= bpp; |
| 429 |
pp -= bpp; |
| 430 |
|
| 431 |
vec_ld_unaligned(pp_vec,pp); |
| 432 |
rp_vec = vec_ld(0,rp); |
| 433 |
|
| 434 |
rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED1_3); |
| 435 |
pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED1_3); |
| 436 |
avg_vec = vec_avg(rp_part_vec,pp_part_vec); |
| 437 |
avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1))); |
| 438 |
rp_vec = vec_add(rp_vec,avg_vec); |
| 439 |
|
| 440 |
rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED2_3); |
| 441 |
pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED2_3); |
| 442 |
avg_vec = vec_avg(rp_part_vec,pp_part_vec); |
| 443 |
avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1))); |
| 444 |
rp_vec = vec_add(rp_vec,avg_vec); |
| 445 |
|
| 446 |
rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED3_3); |
| 447 |
pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED3_3); |
| 448 |
avg_vec = vec_avg(rp_part_vec,pp_part_vec); |
| 449 |
avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1))); |
| 450 |
rp_vec = vec_add(rp_vec,avg_vec); |
| 451 |
|
| 452 |
rp_part_vec = vec_perm(rp_vec,VSX_CHAR_ZERO,VSX_LEFTSHIFTED4_3); |
| 453 |
pp_part_vec = vec_perm(pp_vec,VSX_CHAR_ZERO,VSX_NOT_SHIFTED4_3); |
| 454 |
avg_vec = vec_avg(rp_part_vec,pp_part_vec); |
| 455 |
avg_vec = vec_sub(avg_vec, vec_and(vec_xor(rp_part_vec,pp_part_vec),vec_splat_u8(1))); |
| 456 |
rp_vec = vec_add(rp_vec,avg_vec); |
| 457 |
|
| 458 |
vec_st(rp_vec,0,rp); |
| 459 |
|
| 460 |
rp += 15; |
| 461 |
pp += 15; |
| 462 |
istop -= 16; |
| 463 |
|
| 464 |
/* Since 16 % bpp = 16 % 3 = 1, last element of array must |
| 465 |
* be proceeded manually |
| 466 |
*/ |
| 467 |
*rp = (png_byte)(((int)(*rp) + |
| 468 |
(int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff); |
| 469 |
rp++; |
| 470 |
} |
| 471 |
|
| 472 |
if(istop > 0) |
| 473 |
for (i = 0; i < istop % 16; i++) |
| 474 |
{ |
| 475 |
*rp = (png_byte)(((int)(*rp) + |
| 476 |
(int)(*pp++ + *(rp-bpp)) / 2 ) & 0xff); |
| 477 |
|
| 478 |
rp++; |
| 479 |
} |
| 480 |
} |
| 481 |
|
| 482 |
/* Bytewise c ? t : e. */ |
| 483 |
#define if_then_else(c,t,e) vec_sel(e,t,c) |
| 484 |
|
| 485 |
#define vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) {\ |
| 486 |
c = *(pp - bpp);\ |
| 487 |
a = *(rp - bpp);\ |
| 488 |
b = *pp++;\ |
| 489 |
p = b - c;\ |
| 490 |
pc = a - c;\ |
| 491 |
pa = vsx_abs(p);\ |
| 492 |
pb = vsx_abs(pc);\ |
| 493 |
pc = vsx_abs(p + pc);\ |
| 494 |
if (pb < pa) pa = pb, a = b;\ |
| 495 |
if (pc < pa) a = c;\ |
| 496 |
a += *rp;\ |
| 497 |
*rp++ = (png_byte)a;\ |
| 498 |
} |
| 499 |
|
| 500 |
void png_read_filter_row_paeth4_vsx(png_row_infop row_info, png_bytep row, |
| 501 |
png_const_bytep prev_row) |
| 502 |
{ |
| 503 |
png_byte bpp = 4; |
| 504 |
|
| 505 |
int a, b, c, pa, pb, pc, p; |
| 506 |
vector unsigned char rp_vec; |
| 507 |
vector unsigned char pp_vec; |
| 508 |
vector unsigned short a_vec,b_vec,c_vec,nearest_vec; |
| 509 |
vector signed short pa_vec,pb_vec,pc_vec,smallest_vec; |
| 510 |
|
| 511 |
vsx_declare_common_vars(row_info,row,prev_row,bpp) |
| 512 |
rp -= bpp; |
| 513 |
if(istop >= bpp) |
| 514 |
istop -= bpp; |
| 515 |
|
| 516 |
/* Process the first pixel in the row completely (this is the same as 'up' |
| 517 |
* because there is only one candidate predictor for the first row). |
| 518 |
*/ |
| 519 |
for(i = 0; i < bpp ; i++) |
| 520 |
{ |
| 521 |
*rp = (png_byte)( *rp + *pp); |
| 522 |
rp++; |
| 523 |
pp++; |
| 524 |
} |
| 525 |
|
| 526 |
for(i = 0; i < unaligned_top ; i++) |
| 527 |
{ |
| 528 |
vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) |
| 529 |
} |
| 530 |
|
| 531 |
while( istop >= 16) |
| 532 |
{ |
| 533 |
for(i = 0; i < bpp ; i++) |
| 534 |
{ |
| 535 |
vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) |
| 536 |
} |
| 537 |
|
| 538 |
rp -= bpp; |
| 539 |
pp -= bpp; |
| 540 |
rp_vec = vec_ld(0,rp); |
| 541 |
vec_ld_unaligned(pp_vec,pp); |
| 542 |
|
| 543 |
a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED1_4),1,4); |
| 544 |
b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED1_4),1,4); |
| 545 |
c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED1_4),1,4); |
| 546 |
pa_vec = (vector signed short) vec_sub(b_vec,c_vec); |
| 547 |
pb_vec = (vector signed short) vec_sub(a_vec , c_vec); |
| 548 |
pc_vec = vec_add(pa_vec,pb_vec); |
| 549 |
pa_vec = vec_abs(pa_vec); |
| 550 |
pb_vec = vec_abs(pb_vec); |
| 551 |
pc_vec = vec_abs(pc_vec); |
| 552 |
smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec)); |
| 553 |
nearest_vec = if_then_else( |
| 554 |
vec_cmpeq(pa_vec,smallest_vec), |
| 555 |
a_vec, |
| 556 |
if_then_else( |
| 557 |
vec_cmpeq(pb_vec,smallest_vec), |
| 558 |
b_vec, |
| 559 |
c_vec |
| 560 |
) |
| 561 |
); |
| 562 |
rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,1,4))); |
| 563 |
|
| 564 |
a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED2_4),2,4); |
| 565 |
b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED2_4),2,4); |
| 566 |
c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED2_4),2,4); |
| 567 |
pa_vec = (vector signed short) vec_sub(b_vec,c_vec); |
| 568 |
pb_vec = (vector signed short) vec_sub(a_vec , c_vec); |
| 569 |
pc_vec = vec_add(pa_vec,pb_vec); |
| 570 |
pa_vec = vec_abs(pa_vec); |
| 571 |
pb_vec = vec_abs(pb_vec); |
| 572 |
pc_vec = vec_abs(pc_vec); |
| 573 |
smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec)); |
| 574 |
nearest_vec = if_then_else( |
| 575 |
vec_cmpeq(pa_vec,smallest_vec), |
| 576 |
a_vec, |
| 577 |
if_then_else( |
| 578 |
vec_cmpeq(pb_vec,smallest_vec), |
| 579 |
b_vec, |
| 580 |
c_vec |
| 581 |
) |
| 582 |
); |
| 583 |
rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,2,4))); |
| 584 |
|
| 585 |
a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED3_4),3,4); |
| 586 |
b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED3_4),3,4); |
| 587 |
c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED3_4),3,4); |
| 588 |
pa_vec = (vector signed short) vec_sub(b_vec,c_vec); |
| 589 |
pb_vec = (vector signed short) vec_sub(a_vec , c_vec); |
| 590 |
pc_vec = vec_add(pa_vec,pb_vec); |
| 591 |
pa_vec = vec_abs(pa_vec); |
| 592 |
pb_vec = vec_abs(pb_vec); |
| 593 |
pc_vec = vec_abs(pc_vec); |
| 594 |
smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec)); |
| 595 |
nearest_vec = if_then_else( |
| 596 |
vec_cmpeq(pa_vec,smallest_vec), |
| 597 |
a_vec, |
| 598 |
if_then_else( |
| 599 |
vec_cmpeq(pb_vec,smallest_vec), |
| 600 |
b_vec, |
| 601 |
c_vec |
| 602 |
) |
| 603 |
); |
| 604 |
rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,3,4))); |
| 605 |
|
| 606 |
vec_st(rp_vec,0,rp); |
| 607 |
|
| 608 |
rp += 16; |
| 609 |
pp += 16; |
| 610 |
istop -= 16; |
| 611 |
} |
| 612 |
|
| 613 |
if(istop > 0) |
| 614 |
for (i = 0; i < istop % 16; i++) |
| 615 |
{ |
| 616 |
vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) |
| 617 |
} |
| 618 |
} |
| 619 |
|
| 620 |
void png_read_filter_row_paeth3_vsx(png_row_infop row_info, png_bytep row, |
| 621 |
png_const_bytep prev_row) |
| 622 |
{ |
| 623 |
png_byte bpp = 3; |
| 624 |
|
| 625 |
int a, b, c, pa, pb, pc, p; |
| 626 |
vector unsigned char rp_vec; |
| 627 |
vector unsigned char pp_vec; |
| 628 |
vector unsigned short a_vec,b_vec,c_vec,nearest_vec; |
| 629 |
vector signed short pa_vec,pb_vec,pc_vec,smallest_vec; |
| 630 |
|
| 631 |
vsx_declare_common_vars(row_info,row,prev_row,bpp) |
| 632 |
rp -= bpp; |
| 633 |
if(istop >= bpp) |
| 634 |
istop -= bpp; |
| 635 |
|
| 636 |
/* Process the first pixel in the row completely (this is the same as 'up' |
| 637 |
* because there is only one candidate predictor for the first row). |
| 638 |
*/ |
| 639 |
for(i = 0; i < bpp ; i++) |
| 640 |
{ |
| 641 |
*rp = (png_byte)( *rp + *pp); |
| 642 |
rp++; |
| 643 |
pp++; |
| 644 |
} |
| 645 |
|
| 646 |
for(i = 0; i < unaligned_top ; i++) |
| 647 |
{ |
| 648 |
vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) |
| 649 |
} |
| 650 |
|
| 651 |
while( istop >= 16) |
| 652 |
{ |
| 653 |
for(i = 0; i < bpp ; i++) |
| 654 |
{ |
| 655 |
vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) |
| 656 |
} |
| 657 |
|
| 658 |
rp -= bpp; |
| 659 |
pp -= bpp; |
| 660 |
rp_vec = vec_ld(0,rp); |
| 661 |
vec_ld_unaligned(pp_vec,pp); |
| 662 |
|
| 663 |
a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED1_3),1,3); |
| 664 |
b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED1_3),1,3); |
| 665 |
c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED1_3),1,3); |
| 666 |
pa_vec = (vector signed short) vec_sub(b_vec,c_vec); |
| 667 |
pb_vec = (vector signed short) vec_sub(a_vec , c_vec); |
| 668 |
pc_vec = vec_add(pa_vec,pb_vec); |
| 669 |
pa_vec = vec_abs(pa_vec); |
| 670 |
pb_vec = vec_abs(pb_vec); |
| 671 |
pc_vec = vec_abs(pc_vec); |
| 672 |
smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec)); |
| 673 |
nearest_vec = if_then_else( |
| 674 |
vec_cmpeq(pa_vec,smallest_vec), |
| 675 |
a_vec, |
| 676 |
if_then_else( |
| 677 |
vec_cmpeq(pb_vec,smallest_vec), |
| 678 |
b_vec, |
| 679 |
c_vec |
| 680 |
) |
| 681 |
); |
| 682 |
rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,1,3))); |
| 683 |
|
| 684 |
a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED2_3),2,3); |
| 685 |
b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED2_3),2,3); |
| 686 |
c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED2_3),2,3); |
| 687 |
pa_vec = (vector signed short) vec_sub(b_vec,c_vec); |
| 688 |
pb_vec = (vector signed short) vec_sub(a_vec , c_vec); |
| 689 |
pc_vec = vec_add(pa_vec,pb_vec); |
| 690 |
pa_vec = vec_abs(pa_vec); |
| 691 |
pb_vec = vec_abs(pb_vec); |
| 692 |
pc_vec = vec_abs(pc_vec); |
| 693 |
smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec)); |
| 694 |
nearest_vec = if_then_else( |
| 695 |
vec_cmpeq(pa_vec,smallest_vec), |
| 696 |
a_vec, |
| 697 |
if_then_else( |
| 698 |
vec_cmpeq(pb_vec,smallest_vec), |
| 699 |
b_vec, |
| 700 |
c_vec |
| 701 |
) |
| 702 |
); |
| 703 |
rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,2,3))); |
| 704 |
|
| 705 |
a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED3_3),3,3); |
| 706 |
b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED3_3),3,3); |
| 707 |
c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED3_3),3,3); |
| 708 |
pa_vec = (vector signed short) vec_sub(b_vec,c_vec); |
| 709 |
pb_vec = (vector signed short) vec_sub(a_vec , c_vec); |
| 710 |
pc_vec = vec_add(pa_vec,pb_vec); |
| 711 |
pa_vec = vec_abs(pa_vec); |
| 712 |
pb_vec = vec_abs(pb_vec); |
| 713 |
pc_vec = vec_abs(pc_vec); |
| 714 |
smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec)); |
| 715 |
nearest_vec = if_then_else( |
| 716 |
vec_cmpeq(pa_vec,smallest_vec), |
| 717 |
a_vec, |
| 718 |
if_then_else( |
| 719 |
vec_cmpeq(pb_vec,smallest_vec), |
| 720 |
b_vec, |
| 721 |
c_vec |
| 722 |
) |
| 723 |
); |
| 724 |
rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,3,3))); |
| 725 |
|
| 726 |
a_vec = vsx_char_to_short(vec_perm(rp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED4_3),4,3); |
| 727 |
b_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_NOT_SHIFTED4_3),4,3); |
| 728 |
c_vec = vsx_char_to_short(vec_perm(pp_vec , VSX_CHAR_ZERO , VSX_LEFTSHIFTED4_3),4,3); |
| 729 |
pa_vec = (vector signed short) vec_sub(b_vec,c_vec); |
| 730 |
pb_vec = (vector signed short) vec_sub(a_vec , c_vec); |
| 731 |
pc_vec = vec_add(pa_vec,pb_vec); |
| 732 |
pa_vec = vec_abs(pa_vec); |
| 733 |
pb_vec = vec_abs(pb_vec); |
| 734 |
pc_vec = vec_abs(pc_vec); |
| 735 |
smallest_vec = vec_min(pc_vec, vec_min(pa_vec,pb_vec)); |
| 736 |
nearest_vec = if_then_else( |
| 737 |
vec_cmpeq(pa_vec,smallest_vec), |
| 738 |
a_vec, |
| 739 |
if_then_else( |
| 740 |
vec_cmpeq(pb_vec,smallest_vec), |
| 741 |
b_vec, |
| 742 |
c_vec |
| 743 |
) |
| 744 |
); |
| 745 |
rp_vec = vec_add(rp_vec,(vsx_short_to_char(nearest_vec,4,3))); |
| 746 |
|
| 747 |
vec_st(rp_vec,0,rp); |
| 748 |
|
| 749 |
rp += 15; |
| 750 |
pp += 15; |
| 751 |
istop -= 16; |
| 752 |
|
| 753 |
/* Since 16 % bpp = 16 % 3 = 1, last element of array must |
| 754 |
* be proceeded manually |
| 755 |
*/ |
| 756 |
vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) |
| 757 |
} |
| 758 |
|
| 759 |
if(istop > 0) |
| 760 |
for (i = 0; i < istop % 16; i++) |
| 761 |
{ |
| 762 |
vsx_paeth_process(rp,pp,a,b,c,pa,pb,pc,bpp) |
| 763 |
} |
| 764 |
} |
| 765 |
|
| 766 |
#endif /* PNG_POWERPC_VSX_OPT > 0 */ |
| 767 |
#endif /* PNG_POWERPC_VSX_IMPLEMENTATION == 1 (intrinsics) */ |
| 768 |
#endif /* READ */ |