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, D. R. Commander. 6; 7; Based on the x86 SIMD extension for IJG JPEG library 8; Copyright (C) 1999-2006, MIYASAKA Masaru. 9; For conditions of distribution and use, see copyright notice in jsimdext.inc 10; 11; This file should be assembled with NASM (Netwide Assembler), 12; can *not* be assembled with Microsoft's MASM or any compatible 13; assembler (including Borland's Turbo Assembler). 14; NASM is available from http://nasm.sourceforge.net/ or 15; http://sourceforge.net/project/showfiles.php?group_id=6208 16; 17; This file contains inverse-DCT routines that produce reduced-size 18; output: either 4x4 or 2x2 pixels from an 8x8 DCT block. 19; The following code is based directly on the IJG's original jidctred.c; 20; see the jidctred.c for more details. 21 22%include "jsimdext.inc" 23%include "jdct.inc" 24 25; -------------------------------------------------------------------------- 26 27%define CONST_BITS 13 28%define PASS1_BITS 2 29 30%define DESCALE_P1_4 (CONST_BITS - PASS1_BITS + 1) 31%define DESCALE_P2_4 (CONST_BITS + PASS1_BITS + 3 + 1) 32%define DESCALE_P1_2 (CONST_BITS - PASS1_BITS + 2) 33%define DESCALE_P2_2 (CONST_BITS + PASS1_BITS + 3 + 2) 34 35%if CONST_BITS == 13 36F_0_211 equ 1730 ; FIX(0.211164243) 37F_0_509 equ 4176 ; FIX(0.509795579) 38F_0_601 equ 4926 ; FIX(0.601344887) 39F_0_720 equ 5906 ; FIX(0.720959822) 40F_0_765 equ 6270 ; FIX(0.765366865) 41F_0_850 equ 6967 ; FIX(0.850430095) 42F_0_899 equ 7373 ; FIX(0.899976223) 43F_1_061 equ 8697 ; FIX(1.061594337) 44F_1_272 equ 10426 ; FIX(1.272758580) 45F_1_451 equ 11893 ; FIX(1.451774981) 46F_1_847 equ 15137 ; FIX(1.847759065) 47F_2_172 equ 17799 ; FIX(2.172734803) 48F_2_562 equ 20995 ; FIX(2.562915447) 49F_3_624 equ 29692 ; FIX(3.624509785) 50%else 51; NASM cannot do compile-time arithmetic on floating-point constants. 52%define DESCALE(x, n) (((x) + (1 << ((n) - 1))) >> (n)) 53F_0_211 equ DESCALE( 226735879, 30 - CONST_BITS) ; FIX(0.211164243) 54F_0_509 equ DESCALE( 547388834, 30 - CONST_BITS) ; FIX(0.509795579) 55F_0_601 equ DESCALE( 645689155, 30 - CONST_BITS) ; FIX(0.601344887) 56F_0_720 equ DESCALE( 774124714, 30 - CONST_BITS) ; FIX(0.720959822) 57F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS) ; FIX(0.765366865) 58F_0_850 equ DESCALE( 913142361, 30 - CONST_BITS) ; FIX(0.850430095) 59F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS) ; FIX(0.899976223) 60F_1_061 equ DESCALE(1139878239, 30 - CONST_BITS) ; FIX(1.061594337) 61F_1_272 equ DESCALE(1366614119, 30 - CONST_BITS) ; FIX(1.272758580) 62F_1_451 equ DESCALE(1558831516, 30 - CONST_BITS) ; FIX(1.451774981) 63F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS) ; FIX(1.847759065) 64F_2_172 equ DESCALE(2332956230, 30 - CONST_BITS) ; FIX(2.172734803) 65F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS) ; FIX(2.562915447) 66F_3_624 equ DESCALE(3891787747, 30 - CONST_BITS) ; FIX(3.624509785) 67%endif 68 69; -------------------------------------------------------------------------- 70 SECTION SEG_CONST 71 72 alignz 32 73 GLOBAL_DATA(jconst_idct_red_sse2) 74 75EXTN(jconst_idct_red_sse2): 76 77PW_F184_MF076 times 4 dw F_1_847, -F_0_765 78PW_F256_F089 times 4 dw F_2_562, F_0_899 79PW_F106_MF217 times 4 dw F_1_061, -F_2_172 80PW_MF060_MF050 times 4 dw -F_0_601, -F_0_509 81PW_F145_MF021 times 4 dw F_1_451, -F_0_211 82PW_F362_MF127 times 4 dw F_3_624, -F_1_272 83PW_F085_MF072 times 4 dw F_0_850, -F_0_720 84PD_DESCALE_P1_4 times 4 dd 1 << (DESCALE_P1_4 - 1) 85PD_DESCALE_P2_4 times 4 dd 1 << (DESCALE_P2_4 - 1) 86PD_DESCALE_P1_2 times 4 dd 1 << (DESCALE_P1_2 - 1) 87PD_DESCALE_P2_2 times 4 dd 1 << (DESCALE_P2_2 - 1) 88PB_CENTERJSAMP times 16 db CENTERJSAMPLE 89 90 alignz 32 91 92; -------------------------------------------------------------------------- 93 SECTION SEG_TEXT 94 BITS 64 95; 96; Perform dequantization and inverse DCT on one block of coefficients, 97; producing a reduced-size 4x4 output block. 98; 99; GLOBAL(void) 100; jsimd_idct_4x4_sse2(void *dct_table, JCOEFPTR coef_block, 101; JSAMPARRAY output_buf, JDIMENSION output_col) 102; 103 104; r10 = void *dct_table 105; r11 = JCOEFPTR coef_block 106; r12 = JSAMPARRAY output_buf 107; r13d = JDIMENSION output_col 108 109%define original_rbp rbp + 0 110%define wk(i) rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD 111 ; xmmword wk[WK_NUM] 112%define WK_NUM 2 113 114 align 32 115 GLOBAL_FUNCTION(jsimd_idct_4x4_sse2) 116 117EXTN(jsimd_idct_4x4_sse2): 118 push rbp 119 mov rax, rsp ; rax = original rbp 120 sub rsp, byte 4 121 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 122 mov [rsp], rax 123 mov rbp, rsp ; rbp = aligned rbp 124 lea rsp, [wk(0)] 125 collect_args 4 126 127 ; ---- Pass 1: process columns from input. 128 129 mov rdx, r10 ; quantptr 130 mov rsi, r11 ; inptr 131 132%ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2 133 mov eax, dword [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)] 134 or eax, dword [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)] 135 jnz short .columnDCT 136 137 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)] 138 movdqa xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)] 139 por xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)] 140 por xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)] 141 por xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)] 142 por xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)] 143 por xmm0, xmm1 144 packsswb xmm0, xmm0 145 packsswb xmm0, xmm0 146 movd eax, xmm0 147 test rax, rax 148 jnz short .columnDCT 149 150 ; -- AC terms all zero 151 152 movdqa xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)] 153 pmullw xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 154 155 psllw xmm0, PASS1_BITS 156 157 movdqa xmm3, xmm0 ; xmm0=in0=(00 01 02 03 04 05 06 07) 158 punpcklwd xmm0, xmm0 ; xmm0=(00 00 01 01 02 02 03 03) 159 punpckhwd xmm3, xmm3 ; xmm3=(04 04 05 05 06 06 07 07) 160 161 pshufd xmm1, xmm0, 0x50 ; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01) 162 pshufd xmm0, xmm0, 0xFA ; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03) 163 pshufd xmm6, xmm3, 0x50 ; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05) 164 pshufd xmm3, xmm3, 0xFA ; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07) 165 166 jmp near .column_end 167%endif 168.columnDCT: 169 170 ; -- Odd part 171 172 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)] 173 movdqa xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)] 174 pmullw xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 175 pmullw xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 176 movdqa xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)] 177 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)] 178 pmullw xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 179 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 180 181 movdqa xmm4, xmm0 182 movdqa xmm5, xmm0 183 punpcklwd xmm4, xmm1 184 punpckhwd xmm5, xmm1 185 movdqa xmm0, xmm4 186 movdqa xmm1, xmm5 187 pmaddwd xmm4, [rel PW_F256_F089] ; xmm4=(tmp2L) 188 pmaddwd xmm5, [rel PW_F256_F089] ; xmm5=(tmp2H) 189 pmaddwd xmm0, [rel PW_F106_MF217] ; xmm0=(tmp0L) 190 pmaddwd xmm1, [rel PW_F106_MF217] ; xmm1=(tmp0H) 191 192 movdqa xmm6, xmm2 193 movdqa xmm7, xmm2 194 punpcklwd xmm6, xmm3 195 punpckhwd xmm7, xmm3 196 movdqa xmm2, xmm6 197 movdqa xmm3, xmm7 198 pmaddwd xmm6, [rel PW_MF060_MF050] ; xmm6=(tmp2L) 199 pmaddwd xmm7, [rel PW_MF060_MF050] ; xmm7=(tmp2H) 200 pmaddwd xmm2, [rel PW_F145_MF021] ; xmm2=(tmp0L) 201 pmaddwd xmm3, [rel PW_F145_MF021] ; xmm3=(tmp0H) 202 203 paddd xmm6, xmm4 ; xmm6=tmp2L 204 paddd xmm7, xmm5 ; xmm7=tmp2H 205 paddd xmm2, xmm0 ; xmm2=tmp0L 206 paddd xmm3, xmm1 ; xmm3=tmp0H 207 208 movdqa XMMWORD [wk(0)], xmm2 ; wk(0)=tmp0L 209 movdqa XMMWORD [wk(1)], xmm3 ; wk(1)=tmp0H 210 211 ; -- Even part 212 213 movdqa xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)] 214 movdqa xmm5, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)] 215 movdqa xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)] 216 pmullw xmm4, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 217 pmullw xmm5, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 218 pmullw xmm0, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 219 220 pxor xmm1, xmm1 221 pxor xmm2, xmm2 222 punpcklwd xmm1, xmm4 ; xmm1=tmp0L 223 punpckhwd xmm2, xmm4 ; xmm2=tmp0H 224 psrad xmm1, (16-CONST_BITS-1) ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1 225 psrad xmm2, (16-CONST_BITS-1) ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1 226 227 movdqa xmm3, xmm5 ; xmm5=in2=z2 228 punpcklwd xmm5, xmm0 ; xmm0=in6=z3 229 punpckhwd xmm3, xmm0 230 pmaddwd xmm5, [rel PW_F184_MF076] ; xmm5=tmp2L 231 pmaddwd xmm3, [rel PW_F184_MF076] ; xmm3=tmp2H 232 233 movdqa xmm4, xmm1 234 movdqa xmm0, xmm2 235 paddd xmm1, xmm5 ; xmm1=tmp10L 236 paddd xmm2, xmm3 ; xmm2=tmp10H 237 psubd xmm4, xmm5 ; xmm4=tmp12L 238 psubd xmm0, xmm3 ; xmm0=tmp12H 239 240 ; -- Final output stage 241 242 movdqa xmm5, xmm1 243 movdqa xmm3, xmm2 244 paddd xmm1, xmm6 ; xmm1=data0L 245 paddd xmm2, xmm7 ; xmm2=data0H 246 psubd xmm5, xmm6 ; xmm5=data3L 247 psubd xmm3, xmm7 ; xmm3=data3H 248 249 movdqa xmm6, [rel PD_DESCALE_P1_4] ; xmm6=[rel PD_DESCALE_P1_4] 250 251 paddd xmm1, xmm6 252 paddd xmm2, xmm6 253 psrad xmm1, DESCALE_P1_4 254 psrad xmm2, DESCALE_P1_4 255 paddd xmm5, xmm6 256 paddd xmm3, xmm6 257 psrad xmm5, DESCALE_P1_4 258 psrad xmm3, DESCALE_P1_4 259 260 packssdw xmm1, xmm2 ; xmm1=data0=(00 01 02 03 04 05 06 07) 261 packssdw xmm5, xmm3 ; xmm5=data3=(30 31 32 33 34 35 36 37) 262 263 movdqa xmm7, XMMWORD [wk(0)] ; xmm7=tmp0L 264 movdqa xmm6, XMMWORD [wk(1)] ; xmm6=tmp0H 265 266 movdqa xmm2, xmm4 267 movdqa xmm3, xmm0 268 paddd xmm4, xmm7 ; xmm4=data1L 269 paddd xmm0, xmm6 ; xmm0=data1H 270 psubd xmm2, xmm7 ; xmm2=data2L 271 psubd xmm3, xmm6 ; xmm3=data2H 272 273 movdqa xmm7, [rel PD_DESCALE_P1_4] ; xmm7=[rel PD_DESCALE_P1_4] 274 275 paddd xmm4, xmm7 276 paddd xmm0, xmm7 277 psrad xmm4, DESCALE_P1_4 278 psrad xmm0, DESCALE_P1_4 279 paddd xmm2, xmm7 280 paddd xmm3, xmm7 281 psrad xmm2, DESCALE_P1_4 282 psrad xmm3, DESCALE_P1_4 283 284 packssdw xmm4, xmm0 ; xmm4=data1=(10 11 12 13 14 15 16 17) 285 packssdw xmm2, xmm3 ; xmm2=data2=(20 21 22 23 24 25 26 27) 286 287 movdqa xmm6, xmm1 ; transpose coefficients(phase 1) 288 punpcklwd xmm1, xmm4 ; xmm1=(00 10 01 11 02 12 03 13) 289 punpckhwd xmm6, xmm4 ; xmm6=(04 14 05 15 06 16 07 17) 290 movdqa xmm7, xmm2 ; transpose coefficients(phase 1) 291 punpcklwd xmm2, xmm5 ; xmm2=(20 30 21 31 22 32 23 33) 292 punpckhwd xmm7, xmm5 ; xmm7=(24 34 25 35 26 36 27 37) 293 294 movdqa xmm0, xmm1 ; transpose coefficients(phase 2) 295 punpckldq xmm1, xmm2 ; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31) 296 punpckhdq xmm0, xmm2 ; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33) 297 movdqa xmm3, xmm6 ; transpose coefficients(phase 2) 298 punpckldq xmm6, xmm7 ; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35) 299 punpckhdq xmm3, xmm7 ; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37) 300.column_end: 301 302 ; -- Prefetch the next coefficient block 303 304 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32] 305 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32] 306 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32] 307 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32] 308 309 ; ---- Pass 2: process rows, store into output array. 310 311 mov rax, [original_rbp] 312 mov rdi, r12 ; (JSAMPROW *) 313 mov eax, r13d 314 315 ; -- Even part 316 317 pxor xmm4, xmm4 318 punpcklwd xmm4, xmm1 ; xmm4=tmp0 319 psrad xmm4, (16-CONST_BITS-1) ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1 320 321 ; -- Odd part 322 323 punpckhwd xmm1, xmm0 324 punpckhwd xmm6, xmm3 325 movdqa xmm5, xmm1 326 movdqa xmm2, xmm6 327 pmaddwd xmm1, [rel PW_F256_F089] ; xmm1=(tmp2) 328 pmaddwd xmm6, [rel PW_MF060_MF050] ; xmm6=(tmp2) 329 pmaddwd xmm5, [rel PW_F106_MF217] ; xmm5=(tmp0) 330 pmaddwd xmm2, [rel PW_F145_MF021] ; xmm2=(tmp0) 331 332 paddd xmm6, xmm1 ; xmm6=tmp2 333 paddd xmm2, xmm5 ; xmm2=tmp0 334 335 ; -- Even part 336 337 punpcklwd xmm0, xmm3 338 pmaddwd xmm0, [rel PW_F184_MF076] ; xmm0=tmp2 339 340 movdqa xmm7, xmm4 341 paddd xmm4, xmm0 ; xmm4=tmp10 342 psubd xmm7, xmm0 ; xmm7=tmp12 343 344 ; -- Final output stage 345 346 movdqa xmm1, [rel PD_DESCALE_P2_4] ; xmm1=[rel PD_DESCALE_P2_4] 347 348 movdqa xmm5, xmm4 349 movdqa xmm3, xmm7 350 paddd xmm4, xmm6 ; xmm4=data0=(00 10 20 30) 351 paddd xmm7, xmm2 ; xmm7=data1=(01 11 21 31) 352 psubd xmm5, xmm6 ; xmm5=data3=(03 13 23 33) 353 psubd xmm3, xmm2 ; xmm3=data2=(02 12 22 32) 354 355 paddd xmm4, xmm1 356 paddd xmm7, xmm1 357 psrad xmm4, DESCALE_P2_4 358 psrad xmm7, DESCALE_P2_4 359 paddd xmm5, xmm1 360 paddd xmm3, xmm1 361 psrad xmm5, DESCALE_P2_4 362 psrad xmm3, DESCALE_P2_4 363 364 packssdw xmm4, xmm3 ; xmm4=(00 10 20 30 02 12 22 32) 365 packssdw xmm7, xmm5 ; xmm7=(01 11 21 31 03 13 23 33) 366 367 movdqa xmm0, xmm4 ; transpose coefficients(phase 1) 368 punpcklwd xmm4, xmm7 ; xmm4=(00 01 10 11 20 21 30 31) 369 punpckhwd xmm0, xmm7 ; xmm0=(02 03 12 13 22 23 32 33) 370 371 movdqa xmm6, xmm4 ; transpose coefficients(phase 2) 372 punpckldq xmm4, xmm0 ; xmm4=(00 01 02 03 10 11 12 13) 373 punpckhdq xmm6, xmm0 ; xmm6=(20 21 22 23 30 31 32 33) 374 375 packsswb xmm4, xmm6 ; xmm4=(00 01 02 03 10 11 12 13 20 ..) 376 paddb xmm4, [rel PB_CENTERJSAMP] 377 378 pshufd xmm2, xmm4, 0x39 ; xmm2=(10 11 12 13 20 21 22 23 30 ..) 379 pshufd xmm1, xmm4, 0x4E ; xmm1=(20 21 22 23 30 31 32 33 00 ..) 380 pshufd xmm3, xmm4, 0x93 ; xmm3=(30 31 32 33 00 01 02 03 10 ..) 381 382 mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] 383 mov rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] 384 movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4 385 movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2 386 mov rdx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW] 387 mov rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW] 388 movd XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1 389 movd XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3 390 391 uncollect_args 4 392 mov rsp, rbp ; rsp <- aligned rbp 393 pop rsp ; rsp <- original rbp 394 pop rbp 395 ret 396 397; -------------------------------------------------------------------------- 398; 399; Perform dequantization and inverse DCT on one block of coefficients, 400; producing a reduced-size 2x2 output block. 401; 402; GLOBAL(void) 403; jsimd_idct_2x2_sse2(void *dct_table, JCOEFPTR coef_block, 404; JSAMPARRAY output_buf, JDIMENSION output_col) 405; 406 407; r10 = void *dct_table 408; r11 = JCOEFPTR coef_block 409; r12 = JSAMPARRAY output_buf 410; r13d = JDIMENSION output_col 411 412 align 32 413 GLOBAL_FUNCTION(jsimd_idct_2x2_sse2) 414 415EXTN(jsimd_idct_2x2_sse2): 416 push rbp 417 mov rax, rsp 418 mov rbp, rsp 419 collect_args 4 420 push rbx 421 422 ; ---- Pass 1: process columns from input. 423 424 mov rdx, r10 ; quantptr 425 mov rsi, r11 ; inptr 426 427 ; | input: | result: | 428 ; | 00 01 ** 03 ** 05 ** 07 | | 429 ; | 10 11 ** 13 ** 15 ** 17 | | 430 ; | ** ** ** ** ** ** ** ** | | 431 ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 | 432 ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 | 433 ; | 50 51 ** 53 ** 55 ** 57 | | 434 ; | ** ** ** ** ** ** ** ** | | 435 ; | 70 71 ** 73 ** 75 ** 77 | | 436 437 ; -- Odd part 438 439 movdqa xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)] 440 movdqa xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)] 441 pmullw xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 442 pmullw xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 443 movdqa xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)] 444 movdqa xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)] 445 pmullw xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 446 pmullw xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 447 448 ; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37) 449 ; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77) 450 451 pcmpeqd xmm7, xmm7 452 pslld xmm7, WORD_BIT ; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..} 453 454 movdqa xmm4, xmm0 ; xmm4=(10 11 ** 13 ** 15 ** 17) 455 movdqa xmm5, xmm2 ; xmm5=(50 51 ** 53 ** 55 ** 57) 456 punpcklwd xmm4, xmm1 ; xmm4=(10 30 11 31 ** ** 13 33) 457 punpcklwd xmm5, xmm3 ; xmm5=(50 70 51 71 ** ** 53 73) 458 pmaddwd xmm4, [rel PW_F362_MF127] 459 pmaddwd xmm5, [rel PW_F085_MF072] 460 461 psrld xmm0, WORD_BIT ; xmm0=(11 -- 13 -- 15 -- 17 --) 462 pand xmm1, xmm7 ; xmm1=(-- 31 -- 33 -- 35 -- 37) 463 psrld xmm2, WORD_BIT ; xmm2=(51 -- 53 -- 55 -- 57 --) 464 pand xmm3, xmm7 ; xmm3=(-- 71 -- 73 -- 75 -- 77) 465 por xmm0, xmm1 ; xmm0=(11 31 13 33 15 35 17 37) 466 por xmm2, xmm3 ; xmm2=(51 71 53 73 55 75 57 77) 467 pmaddwd xmm0, [rel PW_F362_MF127] 468 pmaddwd xmm2, [rel PW_F085_MF072] 469 470 paddd xmm4, xmm5 ; xmm4=tmp0[col0 col1 **** col3] 471 paddd xmm0, xmm2 ; xmm0=tmp0[col1 col3 col5 col7] 472 473 ; -- Even part 474 475 movdqa xmm6, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)] 476 pmullw xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)] 477 478 ; xmm6=(00 01 ** 03 ** 05 ** 07) 479 480 movdqa xmm1, xmm6 ; xmm1=(00 01 ** 03 ** 05 ** 07) 481 pslld xmm6, WORD_BIT ; xmm6=(-- 00 -- ** -- ** -- **) 482 pand xmm1, xmm7 ; xmm1=(-- 01 -- 03 -- 05 -- 07) 483 psrad xmm6, (WORD_BIT-CONST_BITS-2) ; xmm6=tmp10[col0 **** **** ****] 484 psrad xmm1, (WORD_BIT-CONST_BITS-2) ; xmm1=tmp10[col1 col3 col5 col7] 485 486 ; -- Final output stage 487 488 movdqa xmm3, xmm6 489 movdqa xmm5, xmm1 490 paddd xmm6, xmm4 ; xmm6=data0[col0 **** **** ****]=(A0 ** ** **) 491 paddd xmm1, xmm0 ; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7) 492 psubd xmm3, xmm4 ; xmm3=data1[col0 **** **** ****]=(B0 ** ** **) 493 psubd xmm5, xmm0 ; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7) 494 495 movdqa xmm2, [rel PD_DESCALE_P1_2] ; xmm2=[rel PD_DESCALE_P1_2] 496 497 punpckldq xmm6, xmm3 ; xmm6=(A0 B0 ** **) 498 499 movdqa xmm7, xmm1 500 punpcklqdq xmm1, xmm5 ; xmm1=(A1 A3 B1 B3) 501 punpckhqdq xmm7, xmm5 ; xmm7=(A5 A7 B5 B7) 502 503 paddd xmm6, xmm2 504 psrad xmm6, DESCALE_P1_2 505 506 paddd xmm1, xmm2 507 paddd xmm7, xmm2 508 psrad xmm1, DESCALE_P1_2 509 psrad xmm7, DESCALE_P1_2 510 511 ; -- Prefetch the next coefficient block 512 513 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32] 514 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32] 515 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32] 516 prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32] 517 518 ; ---- Pass 2: process rows, store into output array. 519 520 mov rdi, r12 ; (JSAMPROW *) 521 mov eax, r13d 522 523 ; | input:| result:| 524 ; | A0 B0 | | 525 ; | A1 B1 | C0 C1 | 526 ; | A3 B3 | D0 D1 | 527 ; | A5 B5 | | 528 ; | A7 B7 | | 529 530 ; -- Odd part 531 532 packssdw xmm1, xmm1 ; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3) 533 packssdw xmm7, xmm7 ; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7) 534 pmaddwd xmm1, [rel PW_F362_MF127] 535 pmaddwd xmm7, [rel PW_F085_MF072] 536 537 paddd xmm1, xmm7 ; xmm1=tmp0[row0 row1 row0 row1] 538 539 ; -- Even part 540 541 pslld xmm6, (CONST_BITS+2) ; xmm6=tmp10[row0 row1 **** ****] 542 543 ; -- Final output stage 544 545 movdqa xmm4, xmm6 546 paddd xmm6, xmm1 ; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **) 547 psubd xmm4, xmm1 ; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **) 548 549 punpckldq xmm6, xmm4 ; xmm6=(C0 D0 C1 D1) 550 551 paddd xmm6, [rel PD_DESCALE_P2_2] 552 psrad xmm6, DESCALE_P2_2 553 554 packssdw xmm6, xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1) 555 packsswb xmm6, xmm6 ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..) 556 paddb xmm6, [rel PB_CENTERJSAMP] 557 558 pextrw ebx, xmm6, 0x00 ; ebx=(C0 D0 -- --) 559 pextrw ecx, xmm6, 0x01 ; ecx=(C1 D1 -- --) 560 561 mov rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW] 562 mov rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW] 563 mov word [rdx+rax*SIZEOF_JSAMPLE], bx 564 mov word [rsi+rax*SIZEOF_JSAMPLE], cx 565 566 pop rbx 567 uncollect_args 4 568 pop rbp 569 ret 570 571; For some reason, the OS X linker does not honor the request to align the 572; segment unless we do this. 573 align 32 574