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