1; 2; jchuff-sse2-64.asm - Huffman entropy encoding (64-bit SSE2) 3; 4; Copyright 2009-2011, 2014-2016 D. R. Commander. 5; Copyright 2015 Matthieu Darbois 6; 7; Based on 8; 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 directly on jchuff.c; see jchuff.c for more 20; details. 21; 22; [TAB8] 23 24%include "jsimdext.inc" 25 26; -------------------------------------------------------------------------- 27 SECTION SEG_CONST 28 29 alignz 16 30 global EXTN(jconst_huff_encode_one_block) PRIVATE 31 32EXTN(jconst_huff_encode_one_block): 33 34%include "jpeg_nbits_table.inc" 35 36 alignz 16 37 38; -------------------------------------------------------------------------- 39 SECTION SEG_TEXT 40 BITS 64 41 42; These macros perform the same task as the emit_bits() function in the 43; original libjpeg code. In addition to reducing overhead by explicitly 44; inlining the code, additional performance is achieved by taking into 45; account the size of the bit buffer and waiting until it is almost full 46; before emptying it. This mostly benefits 64-bit platforms, since 6 47; bytes can be stored in a 64-bit bit buffer before it has to be emptied. 48 49%macro EMIT_BYTE 0 50 sub put_bits, 8 ; put_bits -= 8; 51 mov rdx, put_buffer 52 mov ecx, put_bits 53 shr rdx, cl ; c = (JOCTET)GETJOCTET(put_buffer >> put_bits); 54 mov byte [buffer], dl ; *buffer++ = c; 55 add buffer, 1 56 cmp dl, 0xFF ; need to stuff a zero byte? 57 jne %%.EMIT_BYTE_END 58 mov byte [buffer], 0 ; *buffer++ = 0; 59 add buffer, 1 60%%.EMIT_BYTE_END: 61%endmacro 62 63%macro PUT_BITS 1 64 add put_bits, ecx ; put_bits += size; 65 shl put_buffer, cl ; put_buffer = (put_buffer << size); 66 or put_buffer, %1 67%endmacro 68 69%macro CHECKBUF31 0 70 cmp put_bits, 32 ; if (put_bits > 31) { 71 jl %%.CHECKBUF31_END 72 EMIT_BYTE 73 EMIT_BYTE 74 EMIT_BYTE 75 EMIT_BYTE 76%%.CHECKBUF31_END: 77%endmacro 78 79%macro CHECKBUF47 0 80 cmp put_bits, 48 ; if (put_bits > 47) { 81 jl %%.CHECKBUF47_END 82 EMIT_BYTE 83 EMIT_BYTE 84 EMIT_BYTE 85 EMIT_BYTE 86 EMIT_BYTE 87 EMIT_BYTE 88%%.CHECKBUF47_END: 89%endmacro 90 91%macro EMIT_BITS 2 92 CHECKBUF47 93 mov ecx, %2 94 PUT_BITS %1 95%endmacro 96 97%macro kloop_prepare 37 ;(ko, jno0, ..., jno31, xmm0, xmm1, xmm2, xmm3) 98 pxor xmm8, xmm8 ; __m128i neg = _mm_setzero_si128(); 99 pxor xmm9, xmm9 ; __m128i neg = _mm_setzero_si128(); 100 pxor xmm10, xmm10 ; __m128i neg = _mm_setzero_si128(); 101 pxor xmm11, xmm11 ; __m128i neg = _mm_setzero_si128(); 102 pinsrw %34, word [r12 + %2 * SIZEOF_WORD], 0 ; xmm_shadow[0] = block[jno0]; 103 pinsrw %35, word [r12 + %10 * SIZEOF_WORD], 0 ; xmm_shadow[8] = block[jno8]; 104 pinsrw %36, word [r12 + %18 * SIZEOF_WORD], 0 ; xmm_shadow[16] = block[jno16]; 105 pinsrw %37, word [r12 + %26 * SIZEOF_WORD], 0 ; xmm_shadow[24] = block[jno24]; 106 pinsrw %34, word [r12 + %3 * SIZEOF_WORD], 1 ; xmm_shadow[1] = block[jno1]; 107 pinsrw %35, word [r12 + %11 * SIZEOF_WORD], 1 ; xmm_shadow[9] = block[jno9]; 108 pinsrw %36, word [r12 + %19 * SIZEOF_WORD], 1 ; xmm_shadow[17] = block[jno17]; 109 pinsrw %37, word [r12 + %27 * SIZEOF_WORD], 1 ; xmm_shadow[25] = block[jno25]; 110 pinsrw %34, word [r12 + %4 * SIZEOF_WORD], 2 ; xmm_shadow[2] = block[jno2]; 111 pinsrw %35, word [r12 + %12 * SIZEOF_WORD], 2 ; xmm_shadow[10] = block[jno10]; 112 pinsrw %36, word [r12 + %20 * SIZEOF_WORD], 2 ; xmm_shadow[18] = block[jno18]; 113 pinsrw %37, word [r12 + %28 * SIZEOF_WORD], 2 ; xmm_shadow[26] = block[jno26]; 114 pinsrw %34, word [r12 + %5 * SIZEOF_WORD], 3 ; xmm_shadow[3] = block[jno3]; 115 pinsrw %35, word [r12 + %13 * SIZEOF_WORD], 3 ; xmm_shadow[11] = block[jno11]; 116 pinsrw %36, word [r12 + %21 * SIZEOF_WORD], 3 ; xmm_shadow[19] = block[jno19]; 117 pinsrw %37, word [r12 + %29 * SIZEOF_WORD], 3 ; xmm_shadow[27] = block[jno27]; 118 pinsrw %34, word [r12 + %6 * SIZEOF_WORD], 4 ; xmm_shadow[4] = block[jno4]; 119 pinsrw %35, word [r12 + %14 * SIZEOF_WORD], 4 ; xmm_shadow[12] = block[jno12]; 120 pinsrw %36, word [r12 + %22 * SIZEOF_WORD], 4 ; xmm_shadow[20] = block[jno20]; 121 pinsrw %37, word [r12 + %30 * SIZEOF_WORD], 4 ; xmm_shadow[28] = block[jno28]; 122 pinsrw %34, word [r12 + %7 * SIZEOF_WORD], 5 ; xmm_shadow[5] = block[jno5]; 123 pinsrw %35, word [r12 + %15 * SIZEOF_WORD], 5 ; xmm_shadow[13] = block[jno13]; 124 pinsrw %36, word [r12 + %23 * SIZEOF_WORD], 5 ; xmm_shadow[21] = block[jno21]; 125 pinsrw %37, word [r12 + %31 * SIZEOF_WORD], 5 ; xmm_shadow[29] = block[jno29]; 126 pinsrw %34, word [r12 + %8 * SIZEOF_WORD], 6 ; xmm_shadow[6] = block[jno6]; 127 pinsrw %35, word [r12 + %16 * SIZEOF_WORD], 6 ; xmm_shadow[14] = block[jno14]; 128 pinsrw %36, word [r12 + %24 * SIZEOF_WORD], 6 ; xmm_shadow[22] = block[jno22]; 129 pinsrw %37, word [r12 + %32 * SIZEOF_WORD], 6 ; xmm_shadow[30] = block[jno30]; 130 pinsrw %34, word [r12 + %9 * SIZEOF_WORD], 7 ; xmm_shadow[7] = block[jno7]; 131 pinsrw %35, word [r12 + %17 * SIZEOF_WORD], 7 ; xmm_shadow[15] = block[jno15]; 132 pinsrw %36, word [r12 + %25 * SIZEOF_WORD], 7 ; xmm_shadow[23] = block[jno23]; 133%if %1 != 32 134 pinsrw %37, word [r12 + %33 * SIZEOF_WORD], 7 ; xmm_shadow[31] = block[jno31]; 135%else 136 pinsrw %37, ebx, 7 ; xmm_shadow[31] = block[jno31]; 137%endif 138 pcmpgtw xmm8, %34 ; neg = _mm_cmpgt_epi16(neg, x1); 139 pcmpgtw xmm9, %35 ; neg = _mm_cmpgt_epi16(neg, x1); 140 pcmpgtw xmm10, %36 ; neg = _mm_cmpgt_epi16(neg, x1); 141 pcmpgtw xmm11, %37 ; neg = _mm_cmpgt_epi16(neg, x1); 142 paddw %34, xmm8 ; x1 = _mm_add_epi16(x1, neg); 143 paddw %35, xmm9 ; x1 = _mm_add_epi16(x1, neg); 144 paddw %36, xmm10 ; x1 = _mm_add_epi16(x1, neg); 145 paddw %37, xmm11 ; x1 = _mm_add_epi16(x1, neg); 146 pxor %34, xmm8 ; x1 = _mm_xor_si128(x1, neg); 147 pxor %35, xmm9 ; x1 = _mm_xor_si128(x1, neg); 148 pxor %36, xmm10 ; x1 = _mm_xor_si128(x1, neg); 149 pxor %37, xmm11 ; x1 = _mm_xor_si128(x1, neg); 150 pxor xmm8, %34 ; neg = _mm_xor_si128(neg, x1); 151 pxor xmm9, %35 ; neg = _mm_xor_si128(neg, x1); 152 pxor xmm10, %36 ; neg = _mm_xor_si128(neg, x1); 153 pxor xmm11, %37 ; neg = _mm_xor_si128(neg, x1); 154 movdqa XMMWORD [t1 + %1 * SIZEOF_WORD], %34 ; _mm_storeu_si128((__m128i *)(t1 + ko), x1); 155 movdqa XMMWORD [t1 + (%1 + 8) * SIZEOF_WORD], %35 ; _mm_storeu_si128((__m128i *)(t1 + ko + 8), x1); 156 movdqa XMMWORD [t1 + (%1 + 16) * SIZEOF_WORD], %36 ; _mm_storeu_si128((__m128i *)(t1 + ko + 16), x1); 157 movdqa XMMWORD [t1 + (%1 + 24) * SIZEOF_WORD], %37 ; _mm_storeu_si128((__m128i *)(t1 + ko + 24), x1); 158 movdqa XMMWORD [t2 + %1 * SIZEOF_WORD], xmm8 ; _mm_storeu_si128((__m128i *)(t2 + ko), neg); 159 movdqa XMMWORD [t2 + (%1 + 8) * SIZEOF_WORD], xmm9 ; _mm_storeu_si128((__m128i *)(t2 + ko + 8), neg); 160 movdqa XMMWORD [t2 + (%1 + 16) * SIZEOF_WORD], xmm10 ; _mm_storeu_si128((__m128i *)(t2 + ko + 16), neg); 161 movdqa XMMWORD [t2 + (%1 + 24) * SIZEOF_WORD], xmm11 ; _mm_storeu_si128((__m128i *)(t2 + ko + 24), neg); 162%endmacro 163 164; 165; Encode a single block's worth of coefficients. 166; 167; GLOBAL(JOCTET*) 168; jsimd_huff_encode_one_block_sse2 (working_state *state, JOCTET *buffer, 169; JCOEFPTR block, int last_dc_val, 170; c_derived_tbl *dctbl, c_derived_tbl *actbl) 171; 172 173; r10 = working_state *state 174; r11 = JOCTET *buffer 175; r12 = JCOEFPTR block 176; r13 = int last_dc_val 177; r14 = c_derived_tbl *dctbl 178; r15 = c_derived_tbl *actbl 179 180%define t1 rbp-(DCTSIZE2*SIZEOF_WORD) 181%define t2 t1-(DCTSIZE2*SIZEOF_WORD) 182%define put_buffer r8 183%define put_bits r9d 184%define buffer rax 185 186 align 16 187 global EXTN(jsimd_huff_encode_one_block_sse2) PRIVATE 188 189EXTN(jsimd_huff_encode_one_block_sse2): 190 push rbp 191 mov rax,rsp ; rax = original rbp 192 sub rsp, byte 4 193 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits 194 mov [rsp],rax 195 mov rbp,rsp ; rbp = aligned rbp 196 lea rsp, [t2] 197 collect_args 198%ifdef WIN64 199 movaps XMMWORD [rsp-1*SIZEOF_XMMWORD], xmm8 200 movaps XMMWORD [rsp-2*SIZEOF_XMMWORD], xmm9 201 movaps XMMWORD [rsp-3*SIZEOF_XMMWORD], xmm10 202 movaps XMMWORD [rsp-4*SIZEOF_XMMWORD], xmm11 203 sub rsp, 4*SIZEOF_XMMWORD 204%endif 205 push rbx 206 207 mov buffer, r11 ; r11 is now sratch 208 209 mov put_buffer, MMWORD [r10+16] ; put_buffer = state->cur.put_buffer; 210 mov put_bits, DWORD [r10+24] ; put_bits = state->cur.put_bits; 211 push r10 ; r10 is now scratch 212 213 ; Encode the DC coefficient difference per section F.1.2.1 214 movsx edi, word [r12] ; temp = temp2 = block[0] - last_dc_val; 215 sub edi, r13d ; r13 is not used anymore 216 mov ebx, edi 217 218 ; This is a well-known technique for obtaining the absolute value 219 ; without a branch. It is derived from an assembly language technique 220 ; presented in "How to Optimize for the Pentium Processors", 221 ; Copyright (c) 1996, 1997 by Agner Fog. 222 mov esi, edi 223 sar esi, 31 ; temp3 = temp >> (CHAR_BIT * sizeof(int) - 1); 224 xor edi, esi ; temp ^= temp3; 225 sub edi, esi ; temp -= temp3; 226 227 ; For a negative input, want temp2 = bitwise complement of abs(input) 228 ; This code assumes we are on a two's complement machine 229 add ebx, esi ; temp2 += temp3; 230 231 ; Find the number of bits needed for the magnitude of the coefficient 232 lea r11, [rel jpeg_nbits_table] 233 movzx rdi, byte [r11 + rdi] ; nbits = JPEG_NBITS(temp); 234 ; Emit the Huffman-coded symbol for the number of bits 235 mov r11d, INT [r14 + rdi * 4] ; code = dctbl->ehufco[nbits]; 236 movzx esi, byte [r14 + rdi + 1024] ; size = dctbl->ehufsi[nbits]; 237 EMIT_BITS r11, esi ; EMIT_BITS(code, size) 238 239 ; Mask off any extra bits in code 240 mov esi, 1 241 mov ecx, edi 242 shl esi, cl 243 dec esi 244 and ebx, esi ; temp2 &= (((JLONG) 1)<<nbits) - 1; 245 246 ; Emit that number of bits of the value, if positive, 247 ; or the complement of its magnitude, if negative. 248 EMIT_BITS rbx, edi ; EMIT_BITS(temp2, nbits) 249 250 ; Prepare data 251 xor ebx, ebx 252 kloop_prepare 0, 1, 8, 16, 9, 2, 3, 10, 17, 24, 32, 25, \ 253 18, 11, 4, 5, 12, 19, 26, 33, 40, 48, 41, 34, \ 254 27, 20, 13, 6, 7, 14, 21, 28, 35, \ 255 xmm0, xmm1, xmm2, xmm3 256 kloop_prepare 32, 42, 49, 56, 57, 50, 43, 36, 29, 22, 15, 23, \ 257 30, 37, 44, 51, 58, 59, 52, 45, 38, 31, 39, 46, \ 258 53, 60, 61, 54, 47, 55, 62, 63, 63, \ 259 xmm4, xmm5, xmm6, xmm7 260 261 pxor xmm8, xmm8 262 pcmpeqw xmm0, xmm8 ; tmp0 = _mm_cmpeq_epi16(tmp0, zero); 263 pcmpeqw xmm1, xmm8 ; tmp1 = _mm_cmpeq_epi16(tmp1, zero); 264 pcmpeqw xmm2, xmm8 ; tmp2 = _mm_cmpeq_epi16(tmp2, zero); 265 pcmpeqw xmm3, xmm8 ; tmp3 = _mm_cmpeq_epi16(tmp3, zero); 266 pcmpeqw xmm4, xmm8 ; tmp4 = _mm_cmpeq_epi16(tmp4, zero); 267 pcmpeqw xmm5, xmm8 ; tmp5 = _mm_cmpeq_epi16(tmp5, zero); 268 pcmpeqw xmm6, xmm8 ; tmp6 = _mm_cmpeq_epi16(tmp6, zero); 269 pcmpeqw xmm7, xmm8 ; tmp7 = _mm_cmpeq_epi16(tmp7, zero); 270 packsswb xmm0, xmm1 ; tmp0 = _mm_packs_epi16(tmp0, tmp1); 271 packsswb xmm2, xmm3 ; tmp2 = _mm_packs_epi16(tmp2, tmp3); 272 packsswb xmm4, xmm5 ; tmp4 = _mm_packs_epi16(tmp4, tmp5); 273 packsswb xmm6, xmm7 ; tmp6 = _mm_packs_epi16(tmp6, tmp7); 274 pmovmskb r11d, xmm0 ; index = ((uint64_t)_mm_movemask_epi8(tmp0)) << 0; 275 pmovmskb r12d, xmm2 ; index = ((uint64_t)_mm_movemask_epi8(tmp2)) << 16; 276 pmovmskb r13d, xmm4 ; index = ((uint64_t)_mm_movemask_epi8(tmp4)) << 32; 277 pmovmskb r14d, xmm6 ; index = ((uint64_t)_mm_movemask_epi8(tmp6)) << 48; 278 shl r12, 16 279 shl r14, 16 280 or r11, r12 281 or r13, r14 282 shl r13, 32 283 or r11, r13 284 not r11 ; index = ~index; 285 286 ;mov MMWORD [ t1 + DCTSIZE2 * SIZEOF_WORD ], r11 287 ;jmp .EFN 288 289 mov r13d, INT [r15 + 240 * 4] ; code_0xf0 = actbl->ehufco[0xf0]; 290 movzx r14d, byte [r15 + 1024 + 240] ; size_0xf0 = actbl->ehufsi[0xf0]; 291 lea rsi, [t1] 292.BLOOP: 293 bsf r12, r11 ; r = __builtin_ctzl(index); 294 jz .ELOOP 295 mov rcx, r12 296 lea rsi, [rsi+r12*2] ; k += r; 297 shr r11, cl ; index >>= r; 298 movzx rdi, word [rsi] ; temp = t1[k]; 299 lea rbx, [rel jpeg_nbits_table] 300 movzx rdi, byte [rbx + rdi] ; nbits = JPEG_NBITS(temp); 301.BRLOOP: 302 cmp r12, 16 ; while (r > 15) { 303 jl .ERLOOP 304 EMIT_BITS r13, r14d ; EMIT_BITS(code_0xf0, size_0xf0) 305 sub r12, 16 ; r -= 16; 306 jmp .BRLOOP 307.ERLOOP: 308 ; Emit Huffman symbol for run length / number of bits 309 CHECKBUF31 ; uses rcx, rdx 310 311 shl r12, 4 ; temp3 = (r << 4) + nbits; 312 add r12, rdi 313 mov ebx, INT [r15 + r12 * 4] ; code = actbl->ehufco[temp3]; 314 movzx ecx, byte [r15 + r12 + 1024] ; size = actbl->ehufsi[temp3]; 315 PUT_BITS rbx 316 317 ;EMIT_CODE(code, size) 318 319 movsx ebx, word [rsi-DCTSIZE2*2] ; temp2 = t2[k]; 320 ; Mask off any extra bits in code 321 mov rcx, rdi 322 mov rdx, 1 323 shl rdx, cl 324 dec rdx 325 and rbx, rdx ; temp2 &= (((JLONG) 1)<<nbits) - 1; 326 PUT_BITS rbx ; PUT_BITS(temp2, nbits) 327 328 shr r11, 1 ; index >>= 1; 329 add rsi, 2 ; ++k; 330 jmp .BLOOP 331.ELOOP: 332 ; If the last coef(s) were zero, emit an end-of-block code 333 lea rdi, [t1 + (DCTSIZE2-1) * 2] ; r = DCTSIZE2-1-k; 334 cmp rdi, rsi ; if (r > 0) { 335 je .EFN 336 mov ebx, INT [r15] ; code = actbl->ehufco[0]; 337 movzx r12d, byte [r15 + 1024] ; size = actbl->ehufsi[0]; 338 EMIT_BITS rbx, r12d 339.EFN: 340 pop r10 341 ; Save put_buffer & put_bits 342 mov MMWORD [r10+16], put_buffer ; state->cur.put_buffer = put_buffer; 343 mov DWORD [r10+24], put_bits ; state->cur.put_bits = put_bits; 344 345 pop rbx 346%ifdef WIN64 347 movaps xmm11, XMMWORD [rsp+0*SIZEOF_XMMWORD] 348 movaps xmm10, XMMWORD [rsp+1*SIZEOF_XMMWORD] 349 movaps xmm9, XMMWORD [rsp+2*SIZEOF_XMMWORD] 350 movaps xmm8, XMMWORD [rsp+3*SIZEOF_XMMWORD] 351 add rsp, 4*SIZEOF_XMMWORD 352%endif 353 uncollect_args 354 mov rsp,rbp ; rsp <- aligned rbp 355 pop rsp ; rsp <- original rbp 356 pop rbp 357 ret 358 359; For some reason, the OS X linker does not honor the request to align the 360; segment unless we do this. 361 align 16 362