1; 2; jchuff-sse2.asm - Huffman entropy encoding (64-bit SSE2) 3; 4; Copyright (C) 2009-2011, 2014-2016, 2019, 2021, D. R. Commander. 5; Copyright (C) 2015, Matthieu Darbois. 6; Copyright (C) 2018, Matthias Räncker. 7; 8; Based on the x86 SIMD extension for IJG JPEG library 9; Copyright (C) 1999-2006, MIYASAKA Masaru. 10; For conditions of distribution and use, see copyright notice in jsimdext.inc 11; 12; This file should be assembled with NASM (Netwide Assembler), 13; can *not* be assembled with Microsoft's MASM or any compatible 14; assembler (including Borland's Turbo Assembler). 15; NASM is available from http://nasm.sourceforge.net/ or 16; http://sourceforge.net/project/showfiles.php?group_id=6208 17; 18; This file contains an SSE2 implementation for Huffman coding of one block. 19; The following code is based on jchuff.c; see jchuff.c for more details. 20 21%include "jsimdext.inc" 22 23struc working_state 24.next_output_byte: resp 1 ; => next byte to write in buffer 25.free_in_buffer: resp 1 ; # of byte spaces remaining in buffer 26.cur.put_buffer.simd resq 1 ; current bit accumulation buffer 27.cur.free_bits resd 1 ; # of bits available in it 28.cur.last_dc_val resd 4 ; last DC coef for each component 29.cinfo: resp 1 ; dump_buffer needs access to this 30endstruc 31 32struc c_derived_tbl 33.ehufco: resd 256 ; code for each symbol 34.ehufsi: resb 256 ; length of code for each symbol 35; If no code has been allocated for a symbol S, ehufsi[S] contains 0 36endstruc 37 38; -------------------------------------------------------------------------- 39 SECTION SEG_CONST 40 41 alignz 32 42 GLOBAL_DATA(jconst_huff_encode_one_block) 43 44EXTN(jconst_huff_encode_one_block): 45 46jpeg_mask_bits dd 0x0000, 0x0001, 0x0003, 0x0007 47 dd 0x000f, 0x001f, 0x003f, 0x007f 48 dd 0x00ff, 0x01ff, 0x03ff, 0x07ff 49 dd 0x0fff, 0x1fff, 0x3fff, 0x7fff 50 51 alignz 32 52 53times 1 << 14 db 15 54times 1 << 13 db 14 55times 1 << 12 db 13 56times 1 << 11 db 12 57times 1 << 10 db 11 58times 1 << 9 db 10 59times 1 << 8 db 9 60times 1 << 7 db 8 61times 1 << 6 db 7 62times 1 << 5 db 6 63times 1 << 4 db 5 64times 1 << 3 db 4 65times 1 << 2 db 3 66times 1 << 1 db 2 67times 1 << 0 db 1 68times 1 db 0 69jpeg_nbits_table: 70times 1 db 0 71times 1 << 0 db 1 72times 1 << 1 db 2 73times 1 << 2 db 3 74times 1 << 3 db 4 75times 1 << 4 db 5 76times 1 << 5 db 6 77times 1 << 6 db 7 78times 1 << 7 db 8 79times 1 << 8 db 9 80times 1 << 9 db 10 81times 1 << 10 db 11 82times 1 << 11 db 12 83times 1 << 12 db 13 84times 1 << 13 db 14 85times 1 << 14 db 15 86times 1 << 15 db 16 87 88 alignz 32 89 90%define NBITS(x) nbits_base + x 91%define MASK_BITS(x) NBITS((x) * 4) + (jpeg_mask_bits - jpeg_nbits_table) 92 93; -------------------------------------------------------------------------- 94 SECTION SEG_TEXT 95 BITS 64 96 97; Shorthand used to describe SIMD operations: 98; wN: xmmN treated as eight signed 16-bit values 99; wN[i]: perform the same operation on all eight signed 16-bit values, i=0..7 100; bN: xmmN treated as 16 unsigned 8-bit values 101; bN[i]: perform the same operation on all 16 unsigned 8-bit values, i=0..15 102; Contents of SIMD registers are shown in memory order. 103 104; Fill the bit buffer to capacity with the leading bits from code, then output 105; the bit buffer and put the remaining bits from code into the bit buffer. 106; 107; Usage: 108; code - contains the bits to shift into the bit buffer (LSB-aligned) 109; %1 - the label to which to jump when the macro completes 110; %2 (optional) - extra instructions to execute after nbits has been set 111; 112; Upon completion, free_bits will be set to the number of remaining bits from 113; code, and put_buffer will contain those remaining bits. temp and code will 114; be clobbered. 115; 116; This macro encodes any 0xFF bytes as 0xFF 0x00, as does the EMIT_BYTE() 117; macro in jchuff.c. 118 119%macro EMIT_QWORD 1-2 120 add nbitsb, free_bitsb ; nbits += free_bits; 121 neg free_bitsb ; free_bits = -free_bits; 122 mov tempd, code ; temp = code; 123 shl put_buffer, nbitsb ; put_buffer <<= nbits; 124 mov nbitsb, free_bitsb ; nbits = free_bits; 125 neg free_bitsb ; free_bits = -free_bits; 126 shr tempd, nbitsb ; temp >>= nbits; 127 or tempq, put_buffer ; temp |= put_buffer; 128 movq xmm0, tempq ; xmm0.u64 = { temp, 0 }; 129 bswap tempq ; temp = htonl(temp); 130 mov put_buffer, codeq ; put_buffer = code; 131 pcmpeqb xmm0, xmm1 ; b0[i] = (b0[i] == 0xFF ? 0xFF : 0); 132 %2 133 pmovmskb code, xmm0 ; code = 0; code |= ((b0[i] >> 7) << i); 134 mov qword [buffer], tempq ; memcpy(buffer, &temp, 8); 135 ; (speculative; will be overwritten if 136 ; code contains any 0xFF bytes) 137 add free_bitsb, 64 ; free_bits += 64; 138 add bufferp, 8 ; buffer += 8; 139 test code, code ; if (code == 0) /* No 0xFF bytes */ 140 jz %1 ; return; 141 ; Execute the equivalent of the EMIT_BYTE() macro in jchuff.c for all 8 142 ; bytes in the qword. 143 cmp tempb, 0xFF ; Set CF if temp[0] < 0xFF 144 mov byte [buffer-7], 0 ; buffer[-7] = 0; 145 sbb bufferp, 6 ; buffer -= (6 + (temp[0] < 0xFF ? 1 : 0)); 146 mov byte [buffer], temph ; buffer[0] = temp[1]; 147 cmp temph, 0xFF ; Set CF if temp[1] < 0xFF 148 mov byte [buffer+1], 0 ; buffer[1] = 0; 149 sbb bufferp, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0)); 150 shr tempq, 16 ; temp >>= 16; 151 mov byte [buffer], tempb ; buffer[0] = temp[0]; 152 cmp tempb, 0xFF ; Set CF if temp[0] < 0xFF 153 mov byte [buffer+1], 0 ; buffer[1] = 0; 154 sbb bufferp, -2 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0)); 155 mov byte [buffer], temph ; buffer[0] = temp[1]; 156 cmp temph, 0xFF ; Set CF if temp[1] < 0xFF 157 mov byte [buffer+1], 0 ; buffer[1] = 0; 158 sbb bufferp, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0)); 159 shr tempq, 16 ; temp >>= 16; 160 mov byte [buffer], tempb ; buffer[0] = temp[0]; 161 cmp tempb, 0xFF ; Set CF if temp[0] < 0xFF 162 mov byte [buffer+1], 0 ; buffer[1] = 0; 163 sbb bufferp, -2 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0)); 164 mov byte [buffer], temph ; buffer[0] = temp[1]; 165 cmp temph, 0xFF ; Set CF if temp[1] < 0xFF 166 mov byte [buffer+1], 0 ; buffer[1] = 0; 167 sbb bufferp, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0)); 168 shr tempd, 16 ; temp >>= 16; 169 mov byte [buffer], tempb ; buffer[0] = temp[0]; 170 cmp tempb, 0xFF ; Set CF if temp[0] < 0xFF 171 mov byte [buffer+1], 0 ; buffer[1] = 0; 172 sbb bufferp, -2 ; buffer -= (-2 + (temp[0] < 0xFF ? 1 : 0)); 173 mov byte [buffer], temph ; buffer[0] = temp[1]; 174 cmp temph, 0xFF ; Set CF if temp[1] < 0xFF 175 mov byte [buffer+1], 0 ; buffer[1] = 0; 176 sbb bufferp, -2 ; buffer -= (-2 + (temp[1] < 0xFF ? 1 : 0)); 177 jmp %1 ; return; 178%endmacro 179 180; 181; Encode a single block's worth of coefficients. 182; 183; GLOBAL(JOCTET *) 184; jsimd_huff_encode_one_block_sse2(working_state *state, JOCTET *buffer, 185; JCOEFPTR block, int last_dc_val, 186; c_derived_tbl *dctbl, c_derived_tbl *actbl) 187; 188; NOTES: 189; When shuffling data, we try to avoid pinsrw as much as possible, since it is 190; slow on many CPUs. Its reciprocal throughput (issue latency) is 1 even on 191; modern CPUs, so chains of pinsrw instructions (even with different outputs) 192; can limit performance. pinsrw is a VectorPath instruction on AMD K8 and 193; requires 2 µops (with memory operand) on Intel. In either case, only one 194; pinsrw instruction can be decoded per cycle (and nothing else if they are 195; back-to-back), so out-of-order execution cannot be used to work around long 196; pinsrw chains (though for Sandy Bridge and later, this may be less of a 197; problem if the code runs from the µop cache.) 198; 199; We use tzcnt instead of bsf without checking for support. The instruction is 200; executed as bsf on CPUs that don't support tzcnt (encoding is equivalent to 201; rep bsf.) The destination (first) operand of bsf (and tzcnt on some CPUs) is 202; an input dependency (although the behavior is not formally defined, Intel 203; CPUs usually leave the destination unmodified if the source is zero.) This 204; can prevent out-of-order execution, so we clear the destination before 205; invoking tzcnt. 206; 207; Initial register allocation 208; rax - buffer 209; rbx - temp 210; rcx - nbits 211; rdx - block --> free_bits 212; rsi - nbits_base 213; rdi - t 214; rbp - code 215; r8 - dctbl --> code_temp 216; r9 - actbl 217; r10 - state 218; r11 - index 219; r12 - put_buffer 220 221%define buffer rax 222%ifdef WIN64 223%define bufferp rax 224%else 225%define bufferp raxp 226%endif 227%define tempq rbx 228%define tempd ebx 229%define tempb bl 230%define temph bh 231%define nbitsq rcx 232%define nbits ecx 233%define nbitsb cl 234%define block rdx 235%define nbits_base rsi 236%define t rdi 237%define td edi 238%define codeq rbp 239%define code ebp 240%define dctbl r8 241%define actbl r9 242%define state r10 243%define index r11 244%define indexd r11d 245%define put_buffer r12 246%define put_bufferd r12d 247 248; Step 1: Re-arrange input data according to jpeg_natural_order 249; xx 01 02 03 04 05 06 07 xx 01 08 16 09 02 03 10 250; 08 09 10 11 12 13 14 15 17 24 32 25 18 11 04 05 251; 16 17 18 19 20 21 22 23 12 19 26 33 40 48 41 34 252; 24 25 26 27 28 29 30 31 ==> 27 20 13 06 07 14 21 28 253; 32 33 34 35 36 37 38 39 35 42 49 56 57 50 43 36 254; 40 41 42 43 44 45 46 47 29 22 15 23 30 37 44 51 255; 48 49 50 51 52 53 54 55 58 59 52 45 38 31 39 46 256; 56 57 58 59 60 61 62 63 53 60 61 54 47 55 62 63 257 258 align 32 259 GLOBAL_FUNCTION(jsimd_huff_encode_one_block_sse2) 260 261EXTN(jsimd_huff_encode_one_block_sse2): 262 263%ifdef WIN64 264 265; rcx = working_state *state 266; rdx = JOCTET *buffer 267; r8 = JCOEFPTR block 268; r9 = int last_dc_val 269; [rax+48] = c_derived_tbl *dctbl 270; [rax+56] = c_derived_tbl *actbl 271 272 ;X: X = code stream 273 mov buffer, rdx 274 mov block, r8 275 movups xmm3, XMMWORD [block + 0 * SIZEOF_WORD] ;D: w3 = xx 01 02 03 04 05 06 07 276 push rbx 277 push rbp 278 movdqa xmm0, xmm3 ;A: w0 = xx 01 02 03 04 05 06 07 279 push rsi 280 push rdi 281 push r12 282 movups xmm1, XMMWORD [block + 8 * SIZEOF_WORD] ;B: w1 = 08 09 10 11 12 13 14 15 283 mov state, rcx 284 movsx code, word [block] ;Z: code = block[0]; 285 pxor xmm4, xmm4 ;A: w4[i] = 0; 286 sub code, r9d ;Z: code -= last_dc_val; 287 mov dctbl, POINTER [rsp+6*8+4*8] 288 mov actbl, POINTER [rsp+6*8+5*8] 289 punpckldq xmm0, xmm1 ;A: w0 = xx 01 08 09 02 03 10 11 290 lea nbits_base, [rel jpeg_nbits_table] 291 add rsp, -DCTSIZE2 * SIZEOF_WORD 292 mov t, rsp 293 294%else 295 296; rdi = working_state *state 297; rsi = JOCTET *buffer 298; rdx = JCOEFPTR block 299; rcx = int last_dc_val 300; r8 = c_derived_tbl *dctbl 301; r9 = c_derived_tbl *actbl 302 303 ;X: X = code stream 304 movups xmm3, XMMWORD [block + 0 * SIZEOF_WORD] ;D: w3 = xx 01 02 03 04 05 06 07 305 push rbx 306 push rbp 307 movdqa xmm0, xmm3 ;A: w0 = xx 01 02 03 04 05 06 07 308 push r12 309 mov state, rdi 310 mov buffer, rsi 311 movups xmm1, XMMWORD [block + 8 * SIZEOF_WORD] ;B: w1 = 08 09 10 11 12 13 14 15 312 movsx codeq, word [block] ;Z: code = block[0]; 313 lea nbits_base, [rel jpeg_nbits_table] 314 pxor xmm4, xmm4 ;A: w4[i] = 0; 315 sub codeq, rcx ;Z: code -= last_dc_val; 316 punpckldq xmm0, xmm1 ;A: w0 = xx 01 08 09 02 03 10 11 317 lea t, [rsp - DCTSIZE2 * SIZEOF_WORD] ; use red zone for t_ 318 319%endif 320 321 pshuflw xmm0, xmm0, 11001001b ;A: w0 = 01 08 xx 09 02 03 10 11 322 pinsrw xmm0, word [block + 16 * SIZEOF_WORD], 2 ;A: w0 = 01 08 16 09 02 03 10 11 323 punpckhdq xmm3, xmm1 ;D: w3 = 04 05 12 13 06 07 14 15 324 punpcklqdq xmm1, xmm3 ;B: w1 = 08 09 10 11 04 05 12 13 325 pinsrw xmm0, word [block + 17 * SIZEOF_WORD], 7 ;A: w0 = 01 08 16 09 02 03 10 17 326 ;A: (Row 0, offset 1) 327 pcmpgtw xmm4, xmm0 ;A: w4[i] = (w0[i] < 0 ? -1 : 0); 328 paddw xmm0, xmm4 ;A: w0[i] += w4[i]; 329 movaps XMMWORD [t + 0 * SIZEOF_WORD], xmm0 ;A: t[i] = w0[i]; 330 331 movq xmm2, qword [block + 24 * SIZEOF_WORD] ;B: w2 = 24 25 26 27 -- -- -- -- 332 pshuflw xmm2, xmm2, 11011000b ;B: w2 = 24 26 25 27 -- -- -- -- 333 pslldq xmm1, 1 * SIZEOF_WORD ;B: w1 = -- 08 09 10 11 04 05 12 334 movups xmm5, XMMWORD [block + 48 * SIZEOF_WORD] ;H: w5 = 48 49 50 51 52 53 54 55 335 movsd xmm1, xmm2 ;B: w1 = 24 26 25 27 11 04 05 12 336 punpcklqdq xmm2, xmm5 ;C: w2 = 24 26 25 27 48 49 50 51 337 pinsrw xmm1, word [block + 32 * SIZEOF_WORD], 1 ;B: w1 = 24 32 25 27 11 04 05 12 338 pxor xmm4, xmm4 ;A: w4[i] = 0; 339 psrldq xmm3, 2 * SIZEOF_WORD ;D: w3 = 12 13 06 07 14 15 -- -- 340 pcmpeqw xmm0, xmm4 ;A: w0[i] = (w0[i] == 0 ? -1 : 0); 341 pinsrw xmm1, word [block + 18 * SIZEOF_WORD], 3 ;B: w1 = 24 32 25 18 11 04 05 12 342 ; (Row 1, offset 1) 343 pcmpgtw xmm4, xmm1 ;B: w4[i] = (w1[i] < 0 ? -1 : 0); 344 paddw xmm1, xmm4 ;B: w1[i] += w4[i]; 345 movaps XMMWORD [t + 8 * SIZEOF_WORD], xmm1 ;B: t[i+8] = w1[i]; 346 pxor xmm4, xmm4 ;B: w4[i] = 0; 347 pcmpeqw xmm1, xmm4 ;B: w1[i] = (w1[i] == 0 ? -1 : 0); 348 349 packsswb xmm0, xmm1 ;AB: b0[i] = w0[i], b0[i+8] = w1[i] 350 ; w/ signed saturation 351 352 pinsrw xmm3, word [block + 20 * SIZEOF_WORD], 0 ;D: w3 = 20 13 06 07 14 15 -- -- 353 pinsrw xmm3, word [block + 21 * SIZEOF_WORD], 5 ;D: w3 = 20 13 06 07 14 21 -- -- 354 pinsrw xmm3, word [block + 28 * SIZEOF_WORD], 6 ;D: w3 = 20 13 06 07 14 21 28 -- 355 pinsrw xmm3, word [block + 35 * SIZEOF_WORD], 7 ;D: w3 = 20 13 06 07 14 21 28 35 356 ; (Row 3, offset 1) 357 pcmpgtw xmm4, xmm3 ;D: w4[i] = (w3[i] < 0 ? -1 : 0); 358 paddw xmm3, xmm4 ;D: w3[i] += w4[i]; 359 movaps XMMWORD [t + 24 * SIZEOF_WORD], xmm3 ;D: t[i+24] = w3[i]; 360 pxor xmm4, xmm4 ;D: w4[i] = 0; 361 pcmpeqw xmm3, xmm4 ;D: w3[i] = (w3[i] == 0 ? -1 : 0); 362 363 pinsrw xmm2, word [block + 19 * SIZEOF_WORD], 0 ;C: w2 = 19 26 25 27 48 49 50 51 364 cmp code, 1 << 31 ;Z: Set CF if code < 0x80000000, 365 ;Z: i.e. if code is positive 366 pinsrw xmm2, word [block + 33 * SIZEOF_WORD], 2 ;C: w2 = 19 26 33 27 48 49 50 51 367 pinsrw xmm2, word [block + 40 * SIZEOF_WORD], 3 ;C: w2 = 19 26 33 40 48 49 50 51 368 adc code, -1 ;Z: code += -1 + (code >= 0 ? 1 : 0); 369 pinsrw xmm2, word [block + 41 * SIZEOF_WORD], 5 ;C: w2 = 19 26 33 40 48 41 50 51 370 pinsrw xmm2, word [block + 34 * SIZEOF_WORD], 6 ;C: w2 = 19 26 33 40 48 41 34 51 371 movsxd codeq, code ;Z: sign extend code 372 pinsrw xmm2, word [block + 27 * SIZEOF_WORD], 7 ;C: w2 = 19 26 33 40 48 41 34 27 373 ; (Row 2, offset 1) 374 pcmpgtw xmm4, xmm2 ;C: w4[i] = (w2[i] < 0 ? -1 : 0); 375 paddw xmm2, xmm4 ;C: w2[i] += w4[i]; 376 movaps XMMWORD [t + 16 * SIZEOF_WORD], xmm2 ;C: t[i+16] = w2[i]; 377 pxor xmm4, xmm4 ;C: w4[i] = 0; 378 pcmpeqw xmm2, xmm4 ;C: w2[i] = (w2[i] == 0 ? -1 : 0); 379 380 packsswb xmm2, xmm3 ;CD: b2[i] = w2[i], b2[i+8] = w3[i] 381 ; w/ signed saturation 382 383 movzx nbitsq, byte [NBITS(codeq)] ;Z: nbits = JPEG_NBITS(code); 384 movdqa xmm3, xmm5 ;H: w3 = 48 49 50 51 52 53 54 55 385 pmovmskb tempd, xmm2 ;Z: temp = 0; temp |= ((b2[i] >> 7) << i); 386 pmovmskb put_bufferd, xmm0 ;Z: put_buffer = 0; put_buffer |= ((b0[i] >> 7) << i); 387 movups xmm0, XMMWORD [block + 56 * SIZEOF_WORD] ;H: w0 = 56 57 58 59 60 61 62 63 388 punpckhdq xmm3, xmm0 ;H: w3 = 52 53 60 61 54 55 62 63 389 shl tempd, 16 ;Z: temp <<= 16; 390 psrldq xmm3, 1 * SIZEOF_WORD ;H: w3 = 53 60 61 54 55 62 63 -- 391 pxor xmm2, xmm2 ;H: w2[i] = 0; 392 or put_bufferd, tempd ;Z: put_buffer |= temp; 393 pshuflw xmm3, xmm3, 00111001b ;H: w3 = 60 61 54 53 55 62 63 -- 394 movq xmm1, qword [block + 44 * SIZEOF_WORD] ;G: w1 = 44 45 46 47 -- -- -- -- 395 unpcklps xmm5, xmm0 ;E: w5 = 48 49 56 57 50 51 58 59 396 pxor xmm0, xmm0 ;H: w0[i] = 0; 397 pinsrw xmm3, word [block + 47 * SIZEOF_WORD], 3 ;H: w3 = 60 61 54 47 55 62 63 -- 398 ; (Row 7, offset 1) 399 pcmpgtw xmm2, xmm3 ;H: w2[i] = (w3[i] < 0 ? -1 : 0); 400 paddw xmm3, xmm2 ;H: w3[i] += w2[i]; 401 movaps XMMWORD [t + 56 * SIZEOF_WORD], xmm3 ;H: t[i+56] = w3[i]; 402 movq xmm4, qword [block + 36 * SIZEOF_WORD] ;G: w4 = 36 37 38 39 -- -- -- -- 403 pcmpeqw xmm3, xmm0 ;H: w3[i] = (w3[i] == 0 ? -1 : 0); 404 punpckldq xmm4, xmm1 ;G: w4 = 36 37 44 45 38 39 46 47 405 mov tempd, [dctbl + c_derived_tbl.ehufco + nbitsq * 4] 406 ;Z: temp = dctbl->ehufco[nbits]; 407 movdqa xmm1, xmm4 ;F: w1 = 36 37 44 45 38 39 46 47 408 psrldq xmm4, 1 * SIZEOF_WORD ;G: w4 = 37 44 45 38 39 46 47 -- 409 shufpd xmm1, xmm5, 10b ;F: w1 = 36 37 44 45 50 51 58 59 410 and code, dword [MASK_BITS(nbitsq)] ;Z: code &= (1 << nbits) - 1; 411 pshufhw xmm4, xmm4, 11010011b ;G: w4 = 37 44 45 38 -- 39 46 -- 412 pslldq xmm1, 1 * SIZEOF_WORD ;F: w1 = -- 36 37 44 45 50 51 58 413 shl tempq, nbitsb ;Z: temp <<= nbits; 414 pinsrw xmm4, word [block + 59 * SIZEOF_WORD], 0 ;G: w4 = 59 44 45 38 -- 39 46 -- 415 pshufd xmm1, xmm1, 11011000b ;F: w1 = -- 36 45 50 37 44 51 58 416 pinsrw xmm4, word [block + 52 * SIZEOF_WORD], 1 ;G: w4 = 59 52 45 38 -- 39 46 -- 417 or code, tempd ;Z: code |= temp; 418 movlps xmm1, qword [block + 20 * SIZEOF_WORD] ;F: w1 = 20 21 22 23 37 44 51 58 419 pinsrw xmm4, word [block + 31 * SIZEOF_WORD], 4 ;G: w4 = 59 52 45 38 31 39 46 -- 420 pshuflw xmm1, xmm1, 01110010b ;F: w1 = 22 20 23 21 37 44 51 58 421 pinsrw xmm4, word [block + 53 * SIZEOF_WORD], 7 ;G: w4 = 59 52 45 38 31 39 46 53 422 ; (Row 6, offset 1) 423 pxor xmm2, xmm2 ;G: w2[i] = 0; 424 pcmpgtw xmm0, xmm4 ;G: w0[i] = (w4[i] < 0 ? -1 : 0); 425 pinsrw xmm1, word [block + 15 * SIZEOF_WORD], 1 ;F: w1 = 22 15 23 21 37 44 51 58 426 paddw xmm4, xmm0 ;G: w4[i] += w0[i]; 427 movaps XMMWORD [t + 48 * SIZEOF_WORD], xmm4 ;G: t[48+i] = w4[i]; 428 pinsrw xmm1, word [block + 30 * SIZEOF_WORD], 3 ;F: w1 = 22 15 23 30 37 44 51 58 429 ; (Row 5, offset 1) 430 pcmpeqw xmm4, xmm2 ;G: w4[i] = (w4[i] == 0 ? -1 : 0); 431 pinsrw xmm5, word [block + 42 * SIZEOF_WORD], 0 ;E: w5 = 42 49 56 57 50 51 58 59 432 433 packsswb xmm4, xmm3 ;GH: b4[i] = w4[i], b4[i+8] = w3[i] 434 ; w/ signed saturation 435 436 pxor xmm0, xmm0 ;F: w0[i] = 0; 437 pinsrw xmm5, word [block + 43 * SIZEOF_WORD], 5 ;E: w5 = 42 49 56 57 50 43 58 59 438 pcmpgtw xmm2, xmm1 ;F: w2[i] = (w1[i] < 0 ? -1 : 0); 439 pmovmskb tempd, xmm4 ;Z: temp = 0; temp |= ((b4[i] >> 7) << i); 440 pinsrw xmm5, word [block + 36 * SIZEOF_WORD], 6 ;E: w5 = 42 49 56 57 50 43 36 59 441 paddw xmm1, xmm2 ;F: w1[i] += w2[i]; 442 movaps XMMWORD [t + 40 * SIZEOF_WORD], xmm1 ;F: t[40+i] = w1[i]; 443 pinsrw xmm5, word [block + 29 * SIZEOF_WORD], 7 ;E: w5 = 42 49 56 57 50 43 36 29 444 ; (Row 4, offset 1) 445%undef block 446%define free_bitsq rdx 447%define free_bitsd edx 448%define free_bitsb dl 449 pcmpeqw xmm1, xmm0 ;F: w1[i] = (w1[i] == 0 ? -1 : 0); 450 shl tempq, 48 ;Z: temp <<= 48; 451 pxor xmm2, xmm2 ;E: w2[i] = 0; 452 pcmpgtw xmm0, xmm5 ;E: w0[i] = (w5[i] < 0 ? -1 : 0); 453 paddw xmm5, xmm0 ;E: w5[i] += w0[i]; 454 or tempq, put_buffer ;Z: temp |= put_buffer; 455 movaps XMMWORD [t + 32 * SIZEOF_WORD], xmm5 ;E: t[32+i] = w5[i]; 456 lea t, [dword t - 2] ;Z: t = &t[-1]; 457 pcmpeqw xmm5, xmm2 ;E: w5[i] = (w5[i] == 0 ? -1 : 0); 458 459 packsswb xmm5, xmm1 ;EF: b5[i] = w5[i], b5[i+8] = w1[i] 460 ; w/ signed saturation 461 462 add nbitsb, byte [dctbl + c_derived_tbl.ehufsi + nbitsq] 463 ;Z: nbits += dctbl->ehufsi[nbits]; 464%undef dctbl 465%define code_temp r8d 466 pmovmskb indexd, xmm5 ;Z: index = 0; index |= ((b5[i] >> 7) << i); 467 mov free_bitsd, [state+working_state.cur.free_bits] 468 ;Z: free_bits = state->cur.free_bits; 469 pcmpeqw xmm1, xmm1 ;Z: b1[i] = 0xFF; 470 shl index, 32 ;Z: index <<= 32; 471 mov put_buffer, [state+working_state.cur.put_buffer.simd] 472 ;Z: put_buffer = state->cur.put_buffer.simd; 473 or index, tempq ;Z: index |= temp; 474 not index ;Z: index = ~index; 475 sub free_bitsb, nbitsb ;Z: if ((free_bits -= nbits) >= 0) 476 jnl .ENTRY_SKIP_EMIT_CODE ;Z: goto .ENTRY_SKIP_EMIT_CODE; 477 align 16 478.EMIT_CODE: ;Z: .EMIT_CODE: 479 EMIT_QWORD .BLOOP_COND ;Z: insert code, flush buffer, goto .BLOOP_COND 480 481; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 482 483 align 16 484.BRLOOP: ; do { 485 lea code_temp, [nbitsq - 16] ; code_temp = nbits - 16; 486 movzx nbits, byte [actbl + c_derived_tbl.ehufsi + 0xf0] 487 ; nbits = actbl->ehufsi[0xf0]; 488 mov code, [actbl + c_derived_tbl.ehufco + 0xf0 * 4] 489 ; code = actbl->ehufco[0xf0]; 490 sub free_bitsb, nbitsb ; if ((free_bits -= nbits) <= 0) 491 jle .EMIT_BRLOOP_CODE ; goto .EMIT_BRLOOP_CODE; 492 shl put_buffer, nbitsb ; put_buffer <<= nbits; 493 mov nbits, code_temp ; nbits = code_temp; 494 or put_buffer, codeq ; put_buffer |= code; 495 cmp nbits, 16 ; if (nbits <= 16) 496 jle .ERLOOP ; break; 497 jmp .BRLOOP ; } while (1); 498 499; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 500 501 align 16 502 times 5 nop 503.ENTRY_SKIP_EMIT_CODE: ; .ENTRY_SKIP_EMIT_CODE: 504 shl put_buffer, nbitsb ; put_buffer <<= nbits; 505 or put_buffer, codeq ; put_buffer |= code; 506.BLOOP_COND: ; .BLOOP_COND: 507 test index, index ; if (index != 0) 508 jz .ELOOP ; { 509.BLOOP: ; do { 510 xor nbits, nbits ; nbits = 0; /* kill tzcnt input dependency */ 511 tzcnt nbitsq, index ; nbits = # of trailing 0 bits in index 512 inc nbits ; ++nbits; 513 lea t, [t + nbitsq * 2] ; t = &t[nbits]; 514 shr index, nbitsb ; index >>= nbits; 515.EMIT_BRLOOP_CODE_END: ; .EMIT_BRLOOP_CODE_END: 516 cmp nbits, 16 ; if (nbits > 16) 517 jg .BRLOOP ; goto .BRLOOP; 518.ERLOOP: ; .ERLOOP: 519 movsx codeq, word [t] ; code = *t; 520 lea tempd, [nbitsq * 2] ; temp = nbits * 2; 521 movzx nbits, byte [NBITS(codeq)] ; nbits = JPEG_NBITS(code); 522 lea tempd, [nbitsq + tempq * 8] ; temp = temp * 8 + nbits; 523 mov code_temp, [actbl + c_derived_tbl.ehufco + (tempq - 16) * 4] 524 ; code_temp = actbl->ehufco[temp-16]; 525 shl code_temp, nbitsb ; code_temp <<= nbits; 526 and code, dword [MASK_BITS(nbitsq)] ; code &= (1 << nbits) - 1; 527 add nbitsb, [actbl + c_derived_tbl.ehufsi + (tempq - 16)] 528 ; free_bits -= actbl->ehufsi[temp-16]; 529 or code, code_temp ; code |= code_temp; 530 sub free_bitsb, nbitsb ; if ((free_bits -= nbits) <= 0) 531 jle .EMIT_CODE ; goto .EMIT_CODE; 532 shl put_buffer, nbitsb ; put_buffer <<= nbits; 533 or put_buffer, codeq ; put_buffer |= code; 534 test index, index 535 jnz .BLOOP ; } while (index != 0); 536.ELOOP: ; } /* index != 0 */ 537 sub td, esp ; t -= (WIN64: &t_[0], UNIX: &t_[64]); 538%ifdef WIN64 539 cmp td, (DCTSIZE2 - 2) * SIZEOF_WORD ; if (t != 62) 540%else 541 cmp td, -2 * SIZEOF_WORD ; if (t != -2) 542%endif 543 je .EFN ; { 544 movzx nbits, byte [actbl + c_derived_tbl.ehufsi + 0] 545 ; nbits = actbl->ehufsi[0]; 546 mov code, [actbl + c_derived_tbl.ehufco + 0] ; code = actbl->ehufco[0]; 547 sub free_bitsb, nbitsb ; if ((free_bits -= nbits) <= 0) 548 jg .EFN_SKIP_EMIT_CODE ; { 549 EMIT_QWORD .EFN ; insert code, flush buffer 550 align 16 551.EFN_SKIP_EMIT_CODE: ; } else { 552 shl put_buffer, nbitsb ; put_buffer <<= nbits; 553 or put_buffer, codeq ; put_buffer |= code; 554.EFN: ; } } 555 mov [state + working_state.cur.put_buffer.simd], put_buffer 556 ; state->cur.put_buffer.simd = put_buffer; 557 mov byte [state + working_state.cur.free_bits], free_bitsb 558 ; state->cur.free_bits = free_bits; 559%ifdef WIN64 560 sub rsp, -DCTSIZE2 * SIZEOF_WORD 561 pop r12 562 pop rdi 563 pop rsi 564 pop rbp 565 pop rbx 566%else 567 pop r12 568 pop rbp 569 pop rbx 570%endif 571 ret 572 573; ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ 574 575 align 16 576.EMIT_BRLOOP_CODE: 577 EMIT_QWORD .EMIT_BRLOOP_CODE_END, { mov nbits, code_temp } 578 ; insert code, flush buffer, 579 ; nbits = code_temp, goto .EMIT_BRLOOP_CODE_END 580 581; For some reason, the OS X linker does not honor the request to align the 582; segment unless we do this. 583 align 32 584