1/* 2 * Copyright (C) 2010-2011 Intel Corporation 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17#define JPEG_INTERNALS 18#include "jinclude.h" 19#include "jpeglib.h" 20#include "jdct.h" /* Private declarations for DCT subsystem */ 21 22#ifdef ANDROID_INTELSSE2_IDCT 23#include <emmintrin.h> 24 25#if DCTSIZE != 8 26 Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ 27#endif 28 29#define BITS_INV_ACC 4 30#define SHIFT_INV_ROW 12 31#define SHIFT_INV_COL 5 32const short RND_INV_ROW = 2048; 33const short RND_INV_COL = 16; 34const short RND_INV_CORR = 15; 35 36static const short __attribute__ ((aligned(16))) M128_one_corr[8] = {1,1,1,1,1,1,1,1}; 37static const short __attribute__ ((aligned(16))) M128_round_inv_row[8] = {2048,0,2048,0,2048,0,2048,0}; 38static const short __attribute__ ((aligned(16))) M128_round_inv_col[8] = {16,16,16,16,16,16,16,16}; 39static const short __attribute__ ((aligned(16))) M128_round_inv_corr[8] = {15,15,15,15,15,15,15,15}; 40 41static const short __attribute__ ((aligned(16))) M128_tg_1_16[8] = {13036, 13036, 13036, 13036, 13036, 13036, 13036, 13036}; 42static const short __attribute__ ((aligned(16))) M128_tg_2_16[8] = {27146, 27146, 27146, 27146, 27146, 27146, 27146, 27146}; 43static const short __attribute__ ((aligned(16))) M128_tg_3_16[8] = {-21746, -21746, -21746, -21746, -21746, -21746, -21746, -21746}; 44static const short __attribute__ ((aligned(16))) M128_cos_4_16[8] = {-19195, -19195, -19195, -19195, -19195, -19195, -19195, -19195}; 45 46static const short __attribute__ ((aligned(16))) jpeg_adjust[8] = {128, 128, 128, 128, 128, 128, 128, 128}; 47 48// Table for rows 0,4 49static const short __attribute__ ((aligned(16))) M128_tab_i_04[32] = { 5016384, 21407, 16384, 8867, 5116384, -8867, 16384, -21407, 5216384, 8867, -16384, -21407, 53-16384, 21407, 16384, -8867, 5422725, 19266, 19266, -4520, 5512873, -22725, 4520, -12873, 5612873, 4520, -22725, -12873, 574520, 19266, 19266, -22725 58}; 59 60// Table for rows 1,7 61static const short __attribute__ ((aligned(16))) M128_tab_i_17[32] = { 6222725, 29692, 22725, 12299, 6322725, -12299, 22725, -29692, 6422725, 12299, -22725, -29692, 65-22725, 29692, 22725, -12299, 6631521, 26722, 26722, -6270, 6717855, -31521, 6270, -17855, 6817855, 6270, -31521, -17855, 696270, 26722, 26722, -31521 70}; 71 72// Table for rows 2,6 73static const short __attribute__ ((aligned(16))) M128_tab_i_26[32] = { 7421407, 27969, 21407, 11585, 7521407, -11585, 21407, -27969, 7621407, 11585, -21407, -27969, 77-21407, 27969, 21407, -11585, 7829692, 25172, 25172, -5906, 7916819, -29692, 5906, -16819, 8016819, 5906, -29692, -16819, 815906, 25172, 25172, -29692 82}; 83 84// Table for rows 3,5 85static const short __attribute__ ((aligned(16))) M128_tab_i_35[32] = { 8619266, 25172, 19266, 10426, 8719266, -10426, 19266, -25172, 8819266, 10426, -19266, -25172, 89-19266, 25172, 19266, -10426, 9026722, 22654, 22654, -5315, 9115137, -26722, 5315, -15137, 9215137, 5315, -26722, -15137, 935315, 22654, 22654, -26722 94}; 95 96 97/* 98 * Perform dequantization and inverse DCT on one block of coefficients by SSE. 99 */ 100 101GLOBAL(void) 102jpeg_idct_intelsse (j_decompress_ptr cinfo, jpeg_component_info * compptr, 103 JCOEFPTR coef_block, 104 JSAMPARRAY output_buf, JDIMENSION output_col) 105{ 106 __m128i row0, tmp1, tmp2, tmp3, row2, tmp5, tmp6, tmp7; 107 int ctr; 108 JSAMPROW outptrTemp; 109 JSAMPLE *range_limit = IDCT_range_limit(cinfo); 110 short __attribute__((aligned(16))) quantptrSSE[DCTSIZE2]; 111 short __attribute__((aligned(16))) workspaceSSE[DCTSIZE2]; 112 short __attribute__((aligned(16))) coef_blockSSE[DCTSIZE2]; 113 __m128i x0, x1, x2, x3, x4, x5, x6, x7; 114 __m128i* tg3, *tg1, *tg2, *cos4; 115 __m128i tm765, tp765, tm465, tp465, tp03, tm03, tp12, tm12, tp65, tm65; 116 __m128i t0, t1, t2, t3, t4, t5, t6, t7; 117 __m128i temp, temp2; 118 short * wsptr; 119 unsigned char * outptr; 120 121#define iDCT_8_2ROWs(table1, table2) \ 122 row0 = _mm_shufflelo_epi16(row0, 0xD8); /*x7, x6, x5, x4, x3, x1, x2, x0*/ \ 123 row2 = _mm_shufflelo_epi16(row2, 0xD8); \ 124 tmp1 = _mm_shuffle_epi32(row0, 0); /*x2, x0, x2, x0, x2, x0, x2, x0*/ \ 125 tmp5 = _mm_shuffle_epi32(row2, 0); \ 126 \ 127 tmp3 = _mm_shuffle_epi32(row0, 0x55); /*x3, x1, x3, x1, x3, x1, x3, x1*/ \ 128 tmp7 = _mm_shuffle_epi32(row2, 0x55); \ 129 row0 = _mm_shufflehi_epi16(row0, 0xD8); /*x7, x5, x6, x4, x3, x1, x2, x0*/ \ 130 row2 = _mm_shufflehi_epi16(row2, 0xD8); \ 131 \ 132 tmp1 = _mm_madd_epi16(tmp1, * ( __m128i*)table1); /*x2*w13+x0*w12, x2*w9+x0*w8, x2*w5+x0*w4, x2*w1+x0*w0*/ \ 133 tmp5 = _mm_madd_epi16(tmp5, * ( __m128i*)table2); \ 134 \ 135 tmp2 = _mm_shuffle_epi32(row0, 0xAA); /*x6, x4, x6, x4, x6, x4, x6, x4*/ \ 136 tmp6 = _mm_shuffle_epi32(row2, 0xAA); \ 137 row0 = _mm_shuffle_epi32(row0, 0xFF); /*x7, x5, x7, x5, x7, x5, x7, x5*/ \ 138 row2 = _mm_shuffle_epi32(row2, 0xFF); \ 139\ 140 tmp3 = _mm_madd_epi16(tmp3, * ( __m128i*)(table1+16)); /*x3*w29+x1*w28, x3*w25+x1*w24, x3*w21+x1*w20, x3*w17+x1*w16*/ \ 141 tmp7 = _mm_madd_epi16(tmp7, * ( __m128i*)(table2+16) ); \ 142 row0 = _mm_madd_epi16(row0, * ( __m128i*)(table1+24)); /*x7*w31+x5*w30, x7*w27+x5*w26, x7*w23+x5*w22, x7*w19+x5*w18*/ \ 143 row2 = _mm_madd_epi16(row2, * ( __m128i*)(table2+24) ); \ 144 tmp2 = _mm_madd_epi16(tmp2, * ( __m128i*)(table1+8) ); /*x6*w15+x4*w14, x6*w11+x4*w10, x6*w7+x4*w6, x6*w3+x4*w2*/ \ 145 tmp6 = _mm_madd_epi16(tmp6, * ( __m128i*)(table2+8) ); \ 146 \ 147 tmp1 = _mm_add_epi32(tmp1, * ( __m128i*)M128_round_inv_row); \ 148 tmp5 = _mm_add_epi32(tmp5, * ( __m128i*)M128_round_inv_row); \ 149 row0 = _mm_add_epi32(row0, tmp3); /*b3, b2, b1, b0*/ \ 150 row2 = _mm_add_epi32(row2, tmp7); \ 151 tmp1 = _mm_add_epi32(tmp1, tmp2); /*a3, a2, a1, a0*/ \ 152 tmp5 = _mm_add_epi32(tmp5, tmp6); \ 153 \ 154 tmp2 = tmp1; \ 155 tmp6 = tmp5; \ 156 tmp2 = _mm_sub_epi32(tmp2, row0); /*for row0. y4= a3-b3, y5=a2-b2, y6=a1-b1, y7=a0-b0 */ \ 157 tmp6 = _mm_sub_epi32(tmp6, row2); \ 158 row0 = _mm_add_epi32(row0, tmp1); /*y3=a3+b3,y2=a2+b2,y1=a1+b1,y0=a0+b0*/ \ 159 row2 = _mm_add_epi32(row2, tmp5); \ 160 tmp2 = _mm_srai_epi32(tmp2, SHIFT_INV_ROW); \ 161 tmp6 = _mm_srai_epi32(tmp6, SHIFT_INV_ROW); \ 162 row0 = _mm_srai_epi32(row0, SHIFT_INV_ROW); \ 163 row2 = _mm_srai_epi32(row2, SHIFT_INV_ROW); \ 164 tmp2 = _mm_shuffle_epi32(tmp2, 0x1B); /*y7, y6, y5, y4*/ \ 165 tmp6 = _mm_shuffle_epi32(tmp6, 0x1B); \ 166 row0 = _mm_packs_epi32(row0, tmp2); /*row0 = y7,y6,y5,y4,y3,y2,y1,y0*/ \ 167 row2 = _mm_packs_epi32(row2, tmp6); /*row2 = y7,...y0*/ 168 169 170#define iDCT_8_COL() \ 171 x3 = _mm_load_si128(( __m128i*)(wsptr+24));\ 172 x1 = _mm_load_si128(( __m128i*)(wsptr+8));\ 173 x5 = row0;\ 174 x7 = row2;\ 175\ 176 tg3 = ( __m128i*)(M128_tg_3_16);\ 177 tg1 = ( __m128i*)(M128_tg_1_16);\ 178 tg2 = ( __m128i*)(M128_tg_2_16);\ 179 cos4 =(__m128i*)(M128_cos_4_16);\ 180\ 181 temp = _mm_mulhi_epi16(x5, *tg3); /*row5*tg3*/ \ 182 temp2 = _mm_mulhi_epi16(x3, *tg3);\ 183 temp = _mm_adds_epi16(temp, x5); /*coef adjustment*/ \ 184 temp2 = _mm_adds_epi16(temp2, x3);\ 185 tm765 = _mm_adds_epi16(temp, x3);\ 186 tm465 = _mm_subs_epi16(x5, temp2);\ 187\ 188 temp = _mm_mulhi_epi16(x7, *tg1); /*row7*tg1*/ \ 189 temp2 = _mm_mulhi_epi16(x1, *tg1);\ 190 tp765 = _mm_adds_epi16(temp, x1);\ 191 tp465 = _mm_subs_epi16(temp2, x7); /*row1*tg1 - row7*/ \ 192\ 193 t7 = _mm_adds_epi16(tp765, tm765);\ 194 t7 = _mm_adds_epi16(t7, *( __m128i*)M128_one_corr);\ 195 tp65 = _mm_subs_epi16(tp765, tm765);\ 196 t4 = _mm_adds_epi16(tp465, tm465);\ 197 tm65 = _mm_subs_epi16(tp465, tm465);\ 198 tm65 = _mm_adds_epi16(tm65, *( __m128i*)M128_one_corr);\ 199\ 200 x0 = _mm_load_si128(( __m128i*)(wsptr));\ 201 x4 = _mm_load_si128(( __m128i*)(wsptr+32));\ 202 x2 = _mm_load_si128(( __m128i*)(wsptr+16));\ 203 x6 = _mm_load_si128(( __m128i*)(wsptr+48));\ 204\ 205 /*t6 = ( tp65 + tm65 ) * cos_4_16;*/ \ 206 temp = _mm_adds_epi16(tp65, tm65);\ 207 temp2 = _mm_subs_epi16(tp65, tm65);\ 208 t6 = _mm_mulhi_epi16(temp, *cos4);\ 209 t5 = _mm_mulhi_epi16(temp2, *cos4);\ 210 t6 = _mm_adds_epi16(t6, temp);\ 211 t6 = _mm_or_si128(t6, *( __m128i*)M128_one_corr);\ 212 t5 = _mm_adds_epi16(t5, temp2);\ 213 t5 = _mm_or_si128(t5, *( __m128i*)M128_one_corr);\ 214\ 215 tp03 = _mm_adds_epi16(x0, x4);\ 216 tp12 = _mm_subs_epi16(x0, x4);\ 217\ 218 temp = _mm_mulhi_epi16(x6, *tg2);\ 219 temp2 = _mm_mulhi_epi16(x2, *tg2);\ 220 tm03 = _mm_adds_epi16(temp, x2);\ 221 tm12 = _mm_subs_epi16(temp2, x6);\ 222\ 223 t0 = _mm_adds_epi16(tp03, tm03);\ 224 t0 = _mm_adds_epi16(t0, *( __m128i*)M128_round_inv_col);\ 225 t3 = _mm_subs_epi16(tp03, tm03);\ 226 t3 = _mm_adds_epi16(t3, *( __m128i*)M128_round_inv_corr);\ 227 t1 = _mm_adds_epi16(tp12, tm12);\ 228 t1 = _mm_adds_epi16(t1, *( __m128i*)M128_round_inv_col);\ 229 t2 = _mm_subs_epi16(tp12, tm12);\ 230 t2 = _mm_adds_epi16(t2, *( __m128i*)M128_round_inv_corr);\ 231\ 232 temp = _mm_adds_epi16(t0, t7); /*y0*/ \ 233 temp2 = _mm_adds_epi16(t1, t6); /*y1*/ \ 234 temp = _mm_srai_epi16(temp, SHIFT_INV_COL);\ 235 temp2 = _mm_srai_epi16(temp2, SHIFT_INV_COL);\ 236 temp = _mm_adds_epi16(temp, *( __m128i*)jpeg_adjust); /*Add 128 for jpeg decoding*/ \ 237 temp2 = _mm_adds_epi16(temp2, *( __m128i*)jpeg_adjust);\ 238\ 239 temp = _mm_packus_epi16(temp, temp2);\ 240 _mm_store_si128(( __m128i*)(outptr), temp); /*store y0, y1*/ \ 241\ 242 temp = _mm_adds_epi16(t2, t5);\ 243 temp2 = _mm_adds_epi16(t3, t4);\ 244 temp = _mm_srai_epi16(temp, SHIFT_INV_COL);\ 245 temp2 = _mm_srai_epi16(temp2, SHIFT_INV_COL);\ 246 temp = _mm_adds_epi16(temp, *( __m128i*)jpeg_adjust);\ 247 temp2 = _mm_adds_epi16(temp2, *( __m128i*)jpeg_adjust);\ 248\ 249 temp = _mm_packus_epi16(temp, temp2);\ 250 _mm_store_si128(( __m128i*)(outptr+16), temp); /*store y2, y3*/ \ 251\ 252 temp = _mm_subs_epi16(t3, t4);\ 253 temp2 = _mm_subs_epi16(t2, t5);\ 254 temp = _mm_srai_epi16(temp, SHIFT_INV_COL);\ 255 temp2 = _mm_srai_epi16(temp2, SHIFT_INV_COL);\ 256 temp = _mm_adds_epi16(temp, *( __m128i*)jpeg_adjust);\ 257 temp2 = _mm_adds_epi16(temp2, *( __m128i*)jpeg_adjust);\ 258\ 259 temp = _mm_packus_epi16(temp, temp2);\ 260 _mm_store_si128(( __m128i*)(outptr+32), temp); /*store y4, y5*/ \ 261\ 262 temp = _mm_subs_epi16(t1, t6);\ 263 temp2 = _mm_subs_epi16(t0, t7);\ 264 temp = _mm_srai_epi16(temp, SHIFT_INV_COL);\ 265 temp2 = _mm_srai_epi16(temp2, SHIFT_INV_COL);\ 266 temp = _mm_adds_epi16(temp, *( __m128i*)jpeg_adjust);\ 267 temp2 = _mm_adds_epi16(temp2, *( __m128i*)jpeg_adjust);\ 268\ 269 temp = _mm_packus_epi16(temp, temp2);\ 270 _mm_store_si128(( __m128i*)(outptr+48), temp); /*store y6, y7*/ 271 272 273 /*Memcpy to do 16byte alignment. */ 274 memcpy((char*)quantptrSSE, (char*)compptr->dct_table, sizeof(quantptrSSE)); 275 memcpy((char*)coef_blockSSE, (char*)coef_block, sizeof(coef_blockSSE)); 276 277 wsptr = (short *)workspaceSSE; 278 outptr = (unsigned char*)workspaceSSE; 279 280 // row 0 and row 2 281 row0 = _mm_load_si128((__m128i const*)(coef_blockSSE)); 282 row2 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*2)); 283 row0 = _mm_mullo_epi16( row0, *(__m128i const*)quantptrSSE ); 284 row2 = _mm_mullo_epi16( row2, *(__m128i const*)(quantptrSSE+8*2) ); 285 286 iDCT_8_2ROWs(M128_tab_i_04, M128_tab_i_26); 287 288 _mm_store_si128((__m128i*)(wsptr), row0); 289 _mm_store_si128((__m128i*)(wsptr+8*2), row2); 290 291 // row 4 and row 6 292 row0 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*4)); 293 row2 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*6)); 294 row0 = _mm_mullo_epi16(row0, *(__m128i const*)(quantptrSSE+8*4) ); 295 row2 = _mm_mullo_epi16(row2, *(__m128i const*)(quantptrSSE+8*6) ); 296 297 iDCT_8_2ROWs(M128_tab_i_04, M128_tab_i_26); 298 299 _mm_store_si128((__m128i*)(wsptr+32), row0); 300 _mm_store_si128((__m128i*)(wsptr+48), row2); 301 302 // row 3 and row 1 303 row0 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*3)); 304 row2 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*1)); 305 row0 = _mm_mullo_epi16(row0, *(__m128i const*)(quantptrSSE+24) ); 306 row2 = _mm_mullo_epi16(row2, *(__m128i const*)(quantptrSSE+8) ); 307 308 iDCT_8_2ROWs(M128_tab_i_35, M128_tab_i_17); 309 310 _mm_store_si128((__m128i*)(wsptr+24), row0); 311 _mm_store_si128((__m128i*)(wsptr+8), row2); 312 313 // row 5 and row 7 314 row0 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*5)); 315 row2 = _mm_load_si128((__m128i const*)(coef_blockSSE+8*7)); 316 row0 = _mm_mullo_epi16(row0, *(__m128i const*)(quantptrSSE+40) ); 317 row2 = _mm_mullo_epi16(row2, *(__m128i const*)(quantptrSSE+56)); 318 319 iDCT_8_2ROWs( M128_tab_i_35, M128_tab_i_17); 320 321 iDCT_8_COL(); 322 323 for(ctr = 0; ctr < DCTSIZE; ctr++) 324 { 325 outptrTemp = output_buf[ctr] + output_col; 326 memcpy(outptrTemp, outptr, DCTSIZE); 327 outptr += DCTSIZE; /* advance pointer to next row */ 328 } 329 330 return; 331} 332#endif /* ANDROID_INTELSSE2_IDCT */ 333