jidctred-sse2.asm (21813B)
1 ; 2 ; jidctred.asm - reduced-size IDCT (64-bit SSE2) 3 ; 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5 ; Copyright (C) 2009, 2016, 2024, D. R. Commander. 6 ; Copyright (C) 2018, Matthias Räncker. 7 ; Copyright (C) 2023, Aliaksiej Kandracienka. 8 ; 9 ; Based on the x86 SIMD extension for IJG JPEG library 10 ; Copyright (C) 1999-2006, MIYASAKA Masaru. 11 ; For conditions of distribution and use, see copyright notice in jsimdext.inc 12 ; 13 ; This file should be assembled with NASM (Netwide Assembler) or Yasm. 14 ; 15 ; This file contains inverse-DCT routines that produce reduced-size 16 ; output: either 4x4 or 2x2 pixels from an 8x8 DCT block. 17 ; The following code is based directly on the IJG's original jidctred.c; 18 ; see the jidctred.c for more details. 19 20 %include "jsimdext.inc" 21 %include "jdct.inc" 22 23 ; -------------------------------------------------------------------------- 24 25 %define CONST_BITS 13 26 %define PASS1_BITS 2 27 28 %define DESCALE_P1_4 (CONST_BITS - PASS1_BITS + 1) 29 %define DESCALE_P2_4 (CONST_BITS + PASS1_BITS + 3 + 1) 30 %define DESCALE_P1_2 (CONST_BITS - PASS1_BITS + 2) 31 %define DESCALE_P2_2 (CONST_BITS + PASS1_BITS + 3 + 2) 32 33 %if CONST_BITS == 13 34 F_0_211 equ 1730 ; FIX(0.211164243) 35 F_0_509 equ 4176 ; FIX(0.509795579) 36 F_0_601 equ 4926 ; FIX(0.601344887) 37 F_0_720 equ 5906 ; FIX(0.720959822) 38 F_0_765 equ 6270 ; FIX(0.765366865) 39 F_0_850 equ 6967 ; FIX(0.850430095) 40 F_0_899 equ 7373 ; FIX(0.899976223) 41 F_1_061 equ 8697 ; FIX(1.061594337) 42 F_1_272 equ 10426 ; FIX(1.272758580) 43 F_1_451 equ 11893 ; FIX(1.451774981) 44 F_1_847 equ 15137 ; FIX(1.847759065) 45 F_2_172 equ 17799 ; FIX(2.172734803) 46 F_2_562 equ 20995 ; FIX(2.562915447) 47 F_3_624 equ 29692 ; FIX(3.624509785) 48 %else 49 ; NASM cannot do compile-time arithmetic on floating-point constants. 50 %define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 51 F_0_211 equ DESCALE( 226735879, 30 - CONST_BITS) ; FIX(0.211164243) 52 F_0_509 equ DESCALE( 547388834, 30 - CONST_BITS) ; FIX(0.509795579) 53 F_0_601 equ DESCALE( 645689155, 30 - CONST_BITS) ; FIX(0.601344887) 54 F_0_720 equ DESCALE( 774124714, 30 - CONST_BITS) ; FIX(0.720959822) 55 F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS) ; FIX(0.765366865) 56 F_0_850 equ DESCALE( 913142361, 30 - CONST_BITS) ; FIX(0.850430095) 57 F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS) ; FIX(0.899976223) 58 F_1_061 equ DESCALE(1139878239, 30 - CONST_BITS) ; FIX(1.061594337) 59 F_1_272 equ DESCALE(1366614119, 30 - CONST_BITS) ; FIX(1.272758580) 60 F_1_451 equ DESCALE(1558831516, 30 - CONST_BITS) ; FIX(1.451774981) 61 F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065) 62 F_2_172 equ DESCALE(2332956230, 30 - CONST_BITS) ; FIX(2.172734803) 63 F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS) ; FIX(2.562915447) 64 F_3_624 equ DESCALE(3891787747, 30 - CONST_BITS) ; FIX(3.624509785) 65 %endif 66 67 ; -------------------------------------------------------------------------- 68 SECTION SEG_CONST 69 70 ALIGNZ 32 71 GLOBAL_DATA(jconst_idct_red_sse2) 72 73 EXTN(jconst_idct_red_sse2): 74 75 PW_F184_MF076 times 4 dw F_1_847, -F_0_765 76 PW_F256_F089 times 4 dw F_2_562, F_0_899 77 PW_F106_MF217 times 4 dw F_1_061, -F_2_172 78 PW_MF060_MF050 times 4 dw -F_0_601, -F_0_509 79 PW_F145_MF021 times 4 dw F_1_451, -F_0_211 80 PW_F362_MF127 times 4 dw F_3_624, -F_1_272 81 PW_F085_MF072 times 4 dw F_0_850, -F_0_720 82 PD_DESCALE_P1_4 times 4 dd 1 << (DESCALE_P1_4 - 1) 83 PD_DESCALE_P2_4 times 4 dd 1 << (DESCALE_P2_4 - 1) 84 PD_DESCALE_P1_2 times 4 dd 1 << (DESCALE_P1_2 - 1) 85 PD_DESCALE_P2_2 times 4 dd 1 << (DESCALE_P2_2 - 1) 86 PB_CENTERJSAMP times 16 db CENTERJSAMPLE 87 88 ALIGNZ 32 89 90 ; -------------------------------------------------------------------------- 91 SECTION SEG_TEXT 92 BITS 64 93 ; 94 ; Perform dequantization and inverse DCT on one block of coefficients, 95 ; producing a reduced-size 4x4 output block. 96 ; 97 ; GLOBAL(void) 98 ; jsimd_idct_4x4_sse2(void *dct_table, JCOEFPTR coef_block, 99 ; JSAMPARRAY output_buf, JDIMENSION output_col) 100 ; 101 102 ; r10 = void *dct_table 103 ; r11 = JCOEFPTR coef_block 104 ; r12 = JSAMPARRAY output_buf 105 ; r13d = JDIMENSION output_col 106 107 %define wk(i) r15 - (WK_NUM - (i)) * SIZEOF_XMMWORD 108 ; xmmword wk[WK_NUM] 109 %define WK_NUM 2 110 111 align 32 112 GLOBAL_FUNCTION(jsimd_idct_4x4_sse2) 113 114 EXTN(jsimd_idct_4x4_sse2): 115 ENDBR64 116 push rbp 117 mov rbp, rsp 118 push r15 119 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 120 ; Allocate stack space for wk array. r15 is used to access it. 121 mov r15, rsp 122 sub rsp, byte (SIZEOF_XMMWORD * WK_NUM) 123 COLLECT_ARGS 4 124 125 ; ---- Pass 1: process columns from input. 126 127 mov rdx, r10 ; quantptr 128 mov rsi, r11 ; inptr 129 130 %ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2 131 mov eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)] 132 or eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)] 133 jnz short .columnDCT 134 135 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)] 136 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)] 137 por xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)] 138 por xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)] 139 por xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)] 140 por xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)] 141 por xmm0, xmm1 142 packsswb xmm0, xmm0 143 packsswb xmm0, xmm0 144 movd eax, xmm0 145 test rax, rax 146 jnz short .columnDCT 147 148 ; -- AC terms all zero 149 150 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)] 151 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 152 153 psllw xmm0, PASS1_BITS 154 155 movdqa xmm3, xmm0 ; xmm0=in0=(00 01 02 03 04 05 06 07) 156 punpcklwd xmm0, xmm0 ; xmm0=(00 00 01 01 02 02 03 03) 157 punpckhwd xmm3, xmm3 ; xmm3=(04 04 05 05 06 06 07 07) 158 159 pshufd xmm1, xmm0, 0x50 ; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01) 160 pshufd xmm0, xmm0, 0xFA ; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03) 161 pshufd xmm6, xmm3, 0x50 ; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05) 162 pshufd xmm3, xmm3, 0xFA ; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07) 163 164 jmp near .column_end 165 %endif 166 .columnDCT: 167 168 ; -- Odd part 169 170 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)] 171 movdqa xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)] 172 pmullw xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 173 pmullw xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 174 movdqa xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)] 175 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)] 176 pmullw xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 177 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 178 179 movdqa xmm4, xmm0 180 movdqa xmm5, xmm0 181 punpcklwd xmm4, xmm1 182 punpckhwd xmm5, xmm1 183 movdqa xmm0, xmm4 184 movdqa xmm1, xmm5 185 pmaddwd xmm4, [rel PW_F256_F089] ; xmm4=(tmp2L) 186 pmaddwd xmm5, [rel PW_F256_F089] ; xmm5=(tmp2H) 187 pmaddwd xmm0, [rel PW_F106_MF217] ; xmm0=(tmp0L) 188 pmaddwd xmm1, [rel PW_F106_MF217] ; xmm1=(tmp0H) 189 190 movdqa xmm6, xmm2 191 movdqa xmm7, xmm2 192 punpcklwd xmm6, xmm3 193 punpckhwd xmm7, xmm3 194 movdqa xmm2, xmm6 195 movdqa xmm3, xmm7 196 pmaddwd xmm6, [rel PW_MF060_MF050] ; xmm6=(tmp2L) 197 pmaddwd xmm7, [rel PW_MF060_MF050] ; xmm7=(tmp2H) 198 pmaddwd xmm2, [rel PW_F145_MF021] ; xmm2=(tmp0L) 199 pmaddwd xmm3, [rel PW_F145_MF021] ; xmm3=(tmp0H) 200 201 paddd xmm6, xmm4 ; xmm6=tmp2L 202 paddd xmm7, xmm5 ; xmm7=tmp2H 203 paddd xmm2, xmm0 ; xmm2=tmp0L 204 paddd xmm3, xmm1 ; xmm3=tmp0H 205 206 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=tmp0L 207 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=tmp0H 208 209 ; -- Even part 210 211 movdqa xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)] 212 movdqa xmm5, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)] 213 movdqa xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)] 214 pmullw xmm4, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 215 pmullw xmm5, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 216 pmullw xmm0, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 217 218 pxor xmm1, xmm1 219 pxor xmm2, xmm2 220 punpcklwd xmm1, xmm4 ; xmm1=tmp0L 221 punpckhwd xmm2, xmm4 ; xmm2=tmp0H 222 psrad xmm1, (16-CONST_BITS-1) ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1 223 psrad xmm2, (16-CONST_BITS-1) ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1 224 225 movdqa xmm3, xmm5 ; xmm5=in2=z2 226 punpcklwd xmm5, xmm0 ; xmm0=in6=z3 227 punpckhwd xmm3, xmm0 228 pmaddwd xmm5, [rel PW_F184_MF076] ; xmm5=tmp2L 229 pmaddwd xmm3, [rel PW_F184_MF076] ; xmm3=tmp2H 230 231 movdqa xmm4, xmm1 232 movdqa xmm0, xmm2 233 paddd xmm1, xmm5 ; xmm1=tmp10L 234 paddd xmm2, xmm3 ; xmm2=tmp10H 235 psubd xmm4, xmm5 ; xmm4=tmp12L 236 psubd xmm0, xmm3 ; xmm0=tmp12H 237 238 ; -- Final output stage 239 240 movdqa xmm5, xmm1 241 movdqa xmm3, xmm2 242 paddd xmm1, xmm6 ; xmm1=data0L 243 paddd xmm2, xmm7 ; xmm2=data0H 244 psubd xmm5, xmm6 ; xmm5=data3L 245 psubd xmm3, xmm7 ; xmm3=data3H 246 247 movdqa xmm6, [rel PD_DESCALE_P1_4] ; xmm6=[rel PD_DESCALE_P1_4] 248 249 paddd xmm1, xmm6 250 paddd xmm2, xmm6 251 psrad xmm1, DESCALE_P1_4 252 psrad xmm2, DESCALE_P1_4 253 paddd xmm5, xmm6 254 paddd xmm3, xmm6 255 psrad xmm5, DESCALE_P1_4 256 psrad xmm3, DESCALE_P1_4 257 258 packssdw xmm1, xmm2 ; xmm1=data0=(00 01 02 03 04 05 06 07) 259 packssdw xmm5, xmm3 ; xmm5=data3=(30 31 32 33 34 35 36 37) 260 261 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp0L 262 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp0H 263 264 movdqa xmm2, xmm4 265 movdqa xmm3, xmm0 266 paddd xmm4, xmm7 ; xmm4=data1L 267 paddd xmm0, xmm6 ; xmm0=data1H 268 psubd xmm2, xmm7 ; xmm2=data2L 269 psubd xmm3, xmm6 ; xmm3=data2H 270 271 movdqa xmm7, [rel PD_DESCALE_P1_4] ; xmm7=[rel PD_DESCALE_P1_4] 272 273 paddd xmm4, xmm7 274 paddd xmm0, xmm7 275 psrad xmm4, DESCALE_P1_4 276 psrad xmm0, DESCALE_P1_4 277 paddd xmm2, xmm7 278 paddd xmm3, xmm7 279 psrad xmm2, DESCALE_P1_4 280 psrad xmm3, DESCALE_P1_4 281 282 packssdw xmm4, xmm0 ; xmm4=data1=(10 11 12 13 14 15 16 17) 283 packssdw xmm2, xmm3 ; xmm2=data2=(20 21 22 23 24 25 26 27) 284 285 movdqa xmm6, xmm1 ; transpose coefficients(phase 1) 286 punpcklwd xmm1, xmm4 ; xmm1=(00 10 01 11 02 12 03 13) 287 punpckhwd xmm6, xmm4 ; xmm6=(04 14 05 15 06 16 07 17) 288 movdqa xmm7, xmm2 ; transpose coefficients(phase 1) 289 punpcklwd xmm2, xmm5 ; xmm2=(20 30 21 31 22 32 23 33) 290 punpckhwd xmm7, xmm5 ; xmm7=(24 34 25 35 26 36 27 37) 291 292 movdqa xmm0, xmm1 ; transpose coefficients(phase 2) 293 punpckldq xmm1, xmm2 ; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31) 294 punpckhdq xmm0, xmm2 ; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33) 295 movdqa xmm3, xmm6 ; transpose coefficients(phase 2) 296 punpckldq xmm6, xmm7 ; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35) 297 punpckhdq xmm3, xmm7 ; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37) 298 .column_end: 299 300 ; -- Prefetch the next coefficient block 301 302 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32] 303 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32] 304 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32] 305 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32] 306 307 ; ---- Pass 2: process rows, store into output array. 308 309 mov rdi, r12 ; (JSAMPROW *) 310 mov eax, r13d 311 312 ; -- Even part 313 314 pxor xmm4, xmm4 315 punpcklwd xmm4, xmm1 ; xmm4=tmp0 316 psrad xmm4, (16-CONST_BITS-1) ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1 317 318 ; -- Odd part 319 320 punpckhwd xmm1, xmm0 321 punpckhwd xmm6, xmm3 322 movdqa xmm5, xmm1 323 movdqa xmm2, xmm6 324 pmaddwd xmm1, [rel PW_F256_F089] ; xmm1=(tmp2) 325 pmaddwd xmm6, [rel PW_MF060_MF050] ; xmm6=(tmp2) 326 pmaddwd xmm5, [rel PW_F106_MF217] ; xmm5=(tmp0) 327 pmaddwd xmm2, [rel PW_F145_MF021] ; xmm2=(tmp0) 328 329 paddd xmm6, xmm1 ; xmm6=tmp2 330 paddd xmm2, xmm5 ; xmm2=tmp0 331 332 ; -- Even part 333 334 punpcklwd xmm0, xmm3 335 pmaddwd xmm0, [rel PW_F184_MF076] ; xmm0=tmp2 336 337 movdqa xmm7, xmm4 338 paddd xmm4, xmm0 ; xmm4=tmp10 339 psubd xmm7, xmm0 ; xmm7=tmp12 340 341 ; -- Final output stage 342 343 movdqa xmm1, [rel PD_DESCALE_P2_4] ; xmm1=[rel PD_DESCALE_P2_4] 344 345 movdqa xmm5, xmm4 346 movdqa xmm3, xmm7 347 paddd xmm4, xmm6 ; xmm4=data0=(00 10 20 30) 348 paddd xmm7, xmm2 ; xmm7=data1=(01 11 21 31) 349 psubd xmm5, xmm6 ; xmm5=data3=(03 13 23 33) 350 psubd xmm3, xmm2 ; xmm3=data2=(02 12 22 32) 351 352 paddd xmm4, xmm1 353 paddd xmm7, xmm1 354 psrad xmm4, DESCALE_P2_4 355 psrad xmm7, DESCALE_P2_4 356 paddd xmm5, xmm1 357 paddd xmm3, xmm1 358 psrad xmm5, DESCALE_P2_4 359 psrad xmm3, DESCALE_P2_4 360 361 packssdw xmm4, xmm3 ; xmm4=(00 10 20 30 02 12 22 32) 362 packssdw xmm7, xmm5 ; xmm7=(01 11 21 31 03 13 23 33) 363 364 movdqa xmm0, xmm4 ; transpose coefficients(phase 1) 365 punpcklwd xmm4, xmm7 ; xmm4=(00 01 10 11 20 21 30 31) 366 punpckhwd xmm0, xmm7 ; xmm0=(02 03 12 13 22 23 32 33) 367 368 movdqa xmm6, xmm4 ; transpose coefficients(phase 2) 369 punpckldq xmm4, xmm0 ; xmm4=(00 01 02 03 10 11 12 13) 370 punpckhdq xmm6, xmm0 ; xmm6=(20 21 22 23 30 31 32 33) 371 372 packsswb xmm4, xmm6 ; xmm4=(00 01 02 03 10 11 12 13 20 ..) 373 paddb xmm4, [rel PB_CENTERJSAMP] 374 375 pshufd xmm2, xmm4, 0x39 ; xmm2=(10 11 12 13 20 21 22 23 30 ..) 376 pshufd xmm1, xmm4, 0x4E ; xmm1=(20 21 22 23 30 31 32 33 00 ..) 377 pshufd xmm3, xmm4, 0x93 ; xmm3=(30 31 32 33 00 01 02 03 10 ..) 378 379 mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] 380 mov rsip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] 381 movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4 382 movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2 383 mov rdxp, JSAMPROW [rdi+2*SIZEOF_JSAMPROW] 384 mov rsip, JSAMPROW [rdi+3*SIZEOF_JSAMPROW] 385 movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1 386 movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3 387 388 UNCOLLECT_ARGS 4 389 lea rsp, [rbp-8] 390 pop r15 391 pop rbp 392 ret 393 394 ; -------------------------------------------------------------------------- 395 ; 396 ; Perform dequantization and inverse DCT on one block of coefficients, 397 ; producing a reduced-size 2x2 output block. 398 ; 399 ; GLOBAL(void) 400 ; jsimd_idct_2x2_sse2(void *dct_table, JCOEFPTR coef_block, 401 ; JSAMPARRAY output_buf, JDIMENSION output_col) 402 ; 403 404 ; r10 = void *dct_table 405 ; r11 = JCOEFPTR coef_block 406 ; r12 = JSAMPARRAY output_buf 407 ; r13d = JDIMENSION output_col 408 409 align 32 410 GLOBAL_FUNCTION(jsimd_idct_2x2_sse2) 411 412 EXTN(jsimd_idct_2x2_sse2): 413 ENDBR64 414 push rbp 415 mov rbp, rsp 416 COLLECT_ARGS 4 417 push rbx 418 419 ; ---- Pass 1: process columns from input. 420 421 mov rdx, r10 ; quantptr 422 mov rsi, r11 ; inptr 423 424 ; | input: | result: | 425 ; | 00 01 ** 03 ** 05 ** 07 | | 426 ; | 10 11 ** 13 ** 15 ** 17 | | 427 ; | ** ** ** ** ** ** ** ** | | 428 ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 | 429 ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 | 430 ; | 50 51 ** 53 ** 55 ** 57 | | 431 ; | ** ** ** ** ** ** ** ** | | 432 ; | 70 71 ** 73 ** 75 ** 77 | | 433 434 ; -- Odd part 435 436 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)] 437 movdqa xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)] 438 pmullw xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 439 pmullw xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 440 movdqa xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)] 441 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)] 442 pmullw xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 443 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 444 445 ; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37) 446 ; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77) 447 448 pcmpeqd xmm7, xmm7 449 pslld xmm7, WORD_BIT ; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..} 450 451 movdqa xmm4, xmm0 ; xmm4=(10 11 ** 13 ** 15 ** 17) 452 movdqa xmm5, xmm2 ; xmm5=(50 51 ** 53 ** 55 ** 57) 453 punpcklwd xmm4, xmm1 ; xmm4=(10 30 11 31 ** ** 13 33) 454 punpcklwd xmm5, xmm3 ; xmm5=(50 70 51 71 ** ** 53 73) 455 pmaddwd xmm4, [rel PW_F362_MF127] 456 pmaddwd xmm5, [rel PW_F085_MF072] 457 458 psrld xmm0, WORD_BIT ; xmm0=(11 -- 13 -- 15 -- 17 --) 459 pand xmm1, xmm7 ; xmm1=(-- 31 -- 33 -- 35 -- 37) 460 psrld xmm2, WORD_BIT ; xmm2=(51 -- 53 -- 55 -- 57 --) 461 pand xmm3, xmm7 ; xmm3=(-- 71 -- 73 -- 75 -- 77) 462 por xmm0, xmm1 ; xmm0=(11 31 13 33 15 35 17 37) 463 por xmm2, xmm3 ; xmm2=(51 71 53 73 55 75 57 77) 464 pmaddwd xmm0, [rel PW_F362_MF127] 465 pmaddwd xmm2, [rel PW_F085_MF072] 466 467 paddd xmm4, xmm5 ; xmm4=tmp0[col0 col1 **** col3] 468 paddd xmm0, xmm2 ; xmm0=tmp0[col1 col3 col5 col7] 469 470 ; -- Even part 471 472 movdqa xmm6, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)] 473 pmullw xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 474 475 ; xmm6=(00 01 ** 03 ** 05 ** 07) 476 477 movdqa xmm1, xmm6 ; xmm1=(00 01 ** 03 ** 05 ** 07) 478 pslld xmm6, WORD_BIT ; xmm6=(-- 00 -- ** -- ** -- **) 479 pand xmm1, xmm7 ; xmm1=(-- 01 -- 03 -- 05 -- 07) 480 psrad xmm6, (WORD_BIT-CONST_BITS-2) ; xmm6=tmp10[col0 **** **** ****] 481 psrad xmm1, (WORD_BIT-CONST_BITS-2) ; xmm1=tmp10[col1 col3 col5 col7] 482 483 ; -- Final output stage 484 485 movdqa xmm3, xmm6 486 movdqa xmm5, xmm1 487 paddd xmm6, xmm4 ; xmm6=data0[col0 **** **** ****]=(A0 ** ** **) 488 paddd xmm1, xmm0 ; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7) 489 psubd xmm3, xmm4 ; xmm3=data1[col0 **** **** ****]=(B0 ** ** **) 490 psubd xmm5, xmm0 ; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7) 491 492 movdqa xmm2, [rel PD_DESCALE_P1_2] ; xmm2=[rel PD_DESCALE_P1_2] 493 494 punpckldq xmm6, xmm3 ; xmm6=(A0 B0 ** **) 495 496 movdqa xmm7, xmm1 497 punpcklqdq xmm1, xmm5 ; xmm1=(A1 A3 B1 B3) 498 punpckhqdq xmm7, xmm5 ; xmm7=(A5 A7 B5 B7) 499 500 paddd xmm6, xmm2 501 psrad xmm6, DESCALE_P1_2 502 503 paddd xmm1, xmm2 504 paddd xmm7, xmm2 505 psrad xmm1, DESCALE_P1_2 506 psrad xmm7, DESCALE_P1_2 507 508 ; -- Prefetch the next coefficient block 509 510 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32] 511 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32] 512 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32] 513 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32] 514 515 ; ---- Pass 2: process rows, store into output array. 516 517 mov rdi, r12 ; (JSAMPROW *) 518 mov eax, r13d 519 520 ; | input:| result:| 521 ; | A0 B0 | | 522 ; | A1 B1 | C0 C1 | 523 ; | A3 B3 | D0 D1 | 524 ; | A5 B5 | | 525 ; | A7 B7 | | 526 527 ; -- Odd part 528 529 packssdw xmm1, xmm1 ; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3) 530 packssdw xmm7, xmm7 ; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7) 531 pmaddwd xmm1, [rel PW_F362_MF127] 532 pmaddwd xmm7, [rel PW_F085_MF072] 533 534 paddd xmm1, xmm7 ; xmm1=tmp0[row0 row1 row0 row1] 535 536 ; -- Even part 537 538 pslld xmm6, (CONST_BITS+2) ; xmm6=tmp10[row0 row1 **** ****] 539 540 ; -- Final output stage 541 542 movdqa xmm4, xmm6 543 paddd xmm6, xmm1 ; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **) 544 psubd xmm4, xmm1 ; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **) 545 546 punpckldq xmm6, xmm4 ; xmm6=(C0 D0 C1 D1) 547 548 paddd xmm6, [rel PD_DESCALE_P2_2] 549 psrad xmm6, DESCALE_P2_2 550 551 packssdw xmm6, xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1) 552 packsswb xmm6, xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..) 553 paddb xmm6, [rel PB_CENTERJSAMP] 554 555 pextrw ebx, xmm6, 0x00 ; ebx=(C0 D0 -- --) 556 pextrw ecx, xmm6, 0x01 ; ecx=(C1 D1 -- --) 557 558 mov rdxp, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] 559 mov rsip, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] 560 mov word [rdx+rax*SIZEOF_JSAMPLE], bx 561 mov word [rsi+rax*SIZEOF_JSAMPLE], cx 562 563 pop rbx 564 UNCOLLECT_ARGS 4 565 pop rbp 566 ret 567 568 ; For some reason, the OS X linker does not honor the request to align the 569 ; segment unless we do this. 570 align 32