jdsample-avx2.asm (31013B)
1 ; 2 ; jdsample.asm - upsampling (AVX2) 3 ; 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5 ; Copyright (C) 2015, Intel Corporation. 6 ; Copyright (C) 2016, 2024, D. R. Commander. 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) or Yasm. 13 14 %include "jsimdext.inc" 15 16 ; -------------------------------------------------------------------------- 17 SECTION SEG_CONST 18 19 ALIGNZ 32 20 GLOBAL_DATA(jconst_fancy_upsample_avx2) 21 22 EXTN(jconst_fancy_upsample_avx2): 23 24 PW_ONE times 16 dw 1 25 PW_TWO times 16 dw 2 26 PW_THREE times 16 dw 3 27 PW_SEVEN times 16 dw 7 28 PW_EIGHT times 16 dw 8 29 30 ALIGNZ 32 31 32 ; -------------------------------------------------------------------------- 33 SECTION SEG_TEXT 34 BITS 32 35 ; 36 ; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical. 37 ; 38 ; The upsampling algorithm is linear interpolation between pixel centers, 39 ; also known as a "triangle filter". This is a good compromise between 40 ; speed and visual quality. The centers of the output pixels are 1/4 and 3/4 41 ; of the way between input pixel centers. 42 ; 43 ; GLOBAL(void) 44 ; jsimd_h2v1_fancy_upsample_avx2(int max_v_samp_factor, 45 ; JDIMENSION downsampled_width, 46 ; JSAMPARRAY input_data, 47 ; JSAMPARRAY *output_data_ptr); 48 ; 49 50 %define max_v_samp(b) (b) + 8 ; int max_v_samp_factor 51 %define downsamp_width(b) (b) + 12 ; JDIMENSION downsampled_width 52 %define input_data(b) (b) + 16 ; JSAMPARRAY input_data 53 %define output_data_ptr(b) (b) + 20 ; JSAMPARRAY *output_data_ptr 54 55 align 32 56 GLOBAL_FUNCTION(jsimd_h2v1_fancy_upsample_avx2) 57 58 EXTN(jsimd_h2v1_fancy_upsample_avx2): 59 push ebp 60 mov ebp, esp 61 PUSHPIC ebx 62 ; push ecx ; need not be preserved 63 ; push edx ; need not be preserved 64 push esi 65 push edi 66 67 GET_GOT ebx ; get GOT address 68 69 mov eax, JDIMENSION [downsamp_width(ebp)] ; colctr 70 test eax, eax 71 jz near .return 72 73 mov ecx, INT [max_v_samp(ebp)] ; rowctr 74 test ecx, ecx 75 jz near .return 76 77 mov esi, JSAMPARRAY [input_data(ebp)] ; input_data 78 mov edi, POINTER [output_data_ptr(ebp)] 79 mov edi, JSAMPARRAY [edi] ; output_data 80 ALIGNX 16, 7 81 .rowloop: 82 push eax ; colctr 83 push edi 84 push esi 85 86 mov esi, JSAMPROW [esi] ; inptr 87 mov edi, JSAMPROW [edi] ; outptr 88 89 test eax, SIZEOF_YMMWORD-1 90 jz short .skip 91 mov dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE] 92 mov JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl ; insert a dummy sample 93 .skip: 94 vpxor ymm0, ymm0, ymm0 ; ymm0=(all 0's) 95 vpcmpeqb xmm7, xmm7, xmm7 96 vpsrldq xmm7, xmm7, (SIZEOF_XMMWORD-1) ; (ff -- -- -- ... -- --) LSB is ff 97 vpand ymm7, ymm7, YMMWORD [esi+0*SIZEOF_YMMWORD] 98 99 add eax, byte SIZEOF_YMMWORD-1 100 and eax, byte -SIZEOF_YMMWORD 101 cmp eax, byte SIZEOF_YMMWORD 102 ja short .columnloop 103 ALIGNX 16, 7 104 105 .columnloop_last: 106 vpcmpeqb xmm6, xmm6, xmm6 107 vpslldq xmm6, xmm6, (SIZEOF_XMMWORD-1) 108 vperm2i128 ymm6, ymm6, ymm6, 1 ; (---- ---- ... ---- ---- ff) MSB is ff 109 vpand ymm6, ymm6, YMMWORD [esi+0*SIZEOF_YMMWORD] 110 jmp short .upsample 111 ALIGNX 16, 7 112 113 .columnloop: 114 vmovdqu ymm6, YMMWORD [esi+1*SIZEOF_YMMWORD] 115 vperm2i128 ymm6, ymm0, ymm6, 0x20 116 vpslldq ymm6, ymm6, 15 117 118 .upsample: 119 vmovdqu ymm1, YMMWORD [esi+0*SIZEOF_YMMWORD] ; ymm1=( 0 1 2 ... 29 30 31) 120 121 vperm2i128 ymm2, ymm0, ymm1, 0x20 122 vpalignr ymm2, ymm1, ymm2, 15 ; ymm2=(-- 0 1 ... 28 29 30) 123 vperm2i128 ymm4, ymm0, ymm1, 0x03 124 vpalignr ymm3, ymm4, ymm1, 1 ; ymm3=( 1 2 3 ... 30 31 --) 125 126 vpor ymm2, ymm2, ymm7 ; ymm2=(-1 0 1 ... 28 29 30) 127 vpor ymm3, ymm3, ymm6 ; ymm3=( 1 2 3 ... 30 31 32) 128 129 vpsrldq ymm7, ymm4, (SIZEOF_XMMWORD-1) ; ymm7=(31 -- -- ... -- -- --) 130 131 vpunpckhbw ymm4, ymm1, ymm0 ; ymm4=( 8 9 10 11 12 13 14 15 24 25 26 27 28 29 30 31) 132 vpunpcklbw ymm5, ymm1, ymm0 ; ymm5=( 0 1 2 3 4 5 6 7 16 17 18 19 20 21 22 23) 133 vperm2i128 ymm1, ymm5, ymm4, 0x20 ; ymm1=( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 134 vperm2i128 ymm4, ymm5, ymm4, 0x31 ; ymm4=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 135 136 vpunpckhbw ymm5, ymm2, ymm0 ; ymm5=( 7 8 9 10 11 12 13 14 23 24 25 26 27 28 29 30) 137 vpunpcklbw ymm6, ymm2, ymm0 ; ymm6=(-1 0 1 2 3 4 5 6 15 16 17 18 19 20 21 22) 138 vperm2i128 ymm2, ymm6, ymm5, 0x20 ; ymm2=(-1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14) 139 vperm2i128 ymm5, ymm6, ymm5, 0x31 ; ymm5=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30) 140 141 vpunpckhbw ymm6, ymm3, ymm0 ; ymm6=( 1 2 3 4 5 6 7 8 17 18 19 20 21 22 23 24) 142 vpunpcklbw ymm0, ymm3, ymm0 ; ymm0=( 9 10 11 12 13 14 15 16 25 26 27 28 29 30 31 32) 143 vperm2i128 ymm3, ymm0, ymm6, 0x20 ; ymm3=( 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16) 144 vperm2i128 ymm6, ymm0, ymm6, 0x31 ; ymm6=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32) 145 146 vpxor ymm0, ymm0, ymm0 ; ymm0=(all 0's) 147 148 vpmullw ymm1, ymm1, [GOTOFF(ebx,PW_THREE)] 149 vpmullw ymm4, ymm4, [GOTOFF(ebx,PW_THREE)] 150 vpaddw ymm2, ymm2, [GOTOFF(ebx,PW_ONE)] 151 vpaddw ymm5, ymm5, [GOTOFF(ebx,PW_ONE)] 152 vpaddw ymm3, ymm3, [GOTOFF(ebx,PW_TWO)] 153 vpaddw ymm6, ymm6, [GOTOFF(ebx,PW_TWO)] 154 155 vpaddw ymm2, ymm2, ymm1 156 vpaddw ymm5, ymm5, ymm4 157 vpsrlw ymm2, ymm2, 2 ; ymm2=OutLE=( 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30) 158 vpsrlw ymm5, ymm5, 2 ; ymm5=OutHE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62) 159 vpaddw ymm3, ymm3, ymm1 160 vpaddw ymm6, ymm6, ymm4 161 vpsrlw ymm3, ymm3, 2 ; ymm3=OutLO=( 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31) 162 vpsrlw ymm6, ymm6, 2 ; ymm6=OutHO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63) 163 164 vpsllw ymm3, ymm3, BYTE_BIT 165 vpsllw ymm6, ymm6, BYTE_BIT 166 vpor ymm2, ymm2, ymm3 ; ymm2=OutL=( 0 1 2 ... 29 30 31) 167 vpor ymm5, ymm5, ymm6 ; ymm5=OutH=(32 33 34 ... 61 62 63) 168 169 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymm2 170 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymm5 171 172 sub eax, byte SIZEOF_YMMWORD 173 add esi, byte 1*SIZEOF_YMMWORD ; inptr 174 add edi, byte 2*SIZEOF_YMMWORD ; outptr 175 cmp eax, byte SIZEOF_YMMWORD 176 ja near .columnloop 177 test eax, eax 178 jnz near .columnloop_last 179 180 pop esi 181 pop edi 182 pop eax 183 184 add esi, byte SIZEOF_JSAMPROW ; input_data 185 add edi, byte SIZEOF_JSAMPROW ; output_data 186 dec ecx ; rowctr 187 jg near .rowloop 188 189 .return: 190 vzeroupper 191 pop edi 192 pop esi 193 ; pop edx ; need not be preserved 194 ; pop ecx ; need not be preserved 195 POPPIC ebx 196 pop ebp 197 ret 198 199 ; -------------------------------------------------------------------------- 200 ; 201 ; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical. 202 ; Again a triangle filter; see comments for h2v1 case, above. 203 ; 204 ; GLOBAL(void) 205 ; jsimd_h2v2_fancy_upsample_avx2(int max_v_samp_factor, 206 ; JDIMENSION downsampled_width, 207 ; JSAMPARRAY input_data, 208 ; JSAMPARRAY *output_data_ptr); 209 ; 210 211 %define max_v_samp(b) (b) + 8 ; int max_v_samp_factor 212 %define downsamp_width(b) (b) + 12 ; JDIMENSION downsampled_width 213 %define input_data(b) (b) + 16 ; JSAMPARRAY input_data 214 %define output_data_ptr(b) (b) + 20 ; JSAMPARRAY *output_data_ptr 215 216 %define original_ebp ebp + 0 217 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD 218 ; ymmword wk[WK_NUM] 219 %define WK_NUM 4 220 %define gotptr wk(0) - SIZEOF_POINTER ; void *gotptr 221 222 align 32 223 GLOBAL_FUNCTION(jsimd_h2v2_fancy_upsample_avx2) 224 225 EXTN(jsimd_h2v2_fancy_upsample_avx2): 226 push ebp 227 mov eax, esp ; eax = original ebp 228 sub esp, byte 4 229 and esp, byte (-SIZEOF_YMMWORD) ; align to 256 bits 230 mov [esp], eax 231 mov ebp, esp ; ebp = aligned ebp 232 lea esp, [wk(0)] 233 PUSHPIC eax ; make a room for GOT address 234 push ebx 235 ; push ecx ; need not be preserved 236 ; push edx ; need not be preserved 237 push esi 238 push edi 239 240 GET_GOT ebx ; get GOT address 241 MOVPIC POINTER [gotptr], ebx ; save GOT address 242 243 mov edx, eax ; edx = original ebp 244 mov eax, JDIMENSION [downsamp_width(edx)] ; colctr 245 test eax, eax 246 jz near .return 247 248 mov ecx, INT [max_v_samp(edx)] ; rowctr 249 test ecx, ecx 250 jz near .return 251 252 mov esi, JSAMPARRAY [input_data(edx)] ; input_data 253 mov edi, POINTER [output_data_ptr(edx)] 254 mov edi, JSAMPARRAY [edi] ; output_data 255 ALIGNX 16, 7 256 .rowloop: 257 push eax ; colctr 258 push ecx 259 push edi 260 push esi 261 262 mov ecx, JSAMPROW [esi-1*SIZEOF_JSAMPROW] ; inptr1(above) 263 mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; inptr0 264 mov esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; inptr1(below) 265 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] ; outptr0 266 mov edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW] ; outptr1 267 268 test eax, SIZEOF_YMMWORD-1 269 jz short .skip 270 push edx 271 mov dl, JSAMPLE [ecx+(eax-1)*SIZEOF_JSAMPLE] 272 mov JSAMPLE [ecx+eax*SIZEOF_JSAMPLE], dl 273 mov dl, JSAMPLE [ebx+(eax-1)*SIZEOF_JSAMPLE] 274 mov JSAMPLE [ebx+eax*SIZEOF_JSAMPLE], dl 275 mov dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE] 276 mov JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl ; insert a dummy sample 277 pop edx 278 .skip: 279 ; -- process the first column block 280 281 vmovdqu ymm0, YMMWORD [ebx+0*SIZEOF_YMMWORD] ; ymm0=row[ 0][0] 282 vmovdqu ymm1, YMMWORD [ecx+0*SIZEOF_YMMWORD] ; ymm1=row[-1][0] 283 vmovdqu ymm2, YMMWORD [esi+0*SIZEOF_YMMWORD] ; ymm2=row[+1][0] 284 285 PUSHPIC ebx 286 MOVPIC ebx, POINTER [gotptr] ; load GOT address 287 288 vpxor ymm3, ymm3, ymm3 ; ymm3=(all 0's) 289 290 vpunpckhbw ymm4, ymm0, ymm3 ; ymm4=row[ 0]( 8 9 10 11 12 13 14 15 24 25 26 27 28 29 30 31) 291 vpunpcklbw ymm5, ymm0, ymm3 ; ymm5=row[ 0]( 0 1 2 3 4 5 6 7 16 17 18 19 20 21 22 23) 292 vperm2i128 ymm0, ymm5, ymm4, 0x20 ; ymm0=row[ 0]( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 293 vperm2i128 ymm4, ymm5, ymm4, 0x31 ; ymm4=row[ 0](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 294 295 vpunpckhbw ymm5, ymm1, ymm3 ; ymm5=row[-1]( 8 9 10 11 12 13 14 15 24 25 26 27 28 29 30 31) 296 vpunpcklbw ymm6, ymm1, ymm3 ; ymm6=row[-1]( 0 1 2 3 4 5 6 7 16 17 18 19 20 21 22 23) 297 vperm2i128 ymm1, ymm6, ymm5, 0x20 ; ymm1=row[-1]( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 298 vperm2i128 ymm5, ymm6, ymm5, 0x31 ; ymm5=row[-1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 299 300 vpunpckhbw ymm6, ymm2, ymm3 ; ymm6=row[+1]( 8 9 10 11 12 13 14 15 24 25 26 27 28 29 30 31) 301 vpunpcklbw ymm3, ymm2, ymm3 ; ymm3=row[+1]( 0 1 2 3 4 5 6 7 16 17 18 19 20 21 22 23) 302 vperm2i128 ymm2, ymm3, ymm6, 0x20 ; ymm2=row[+1]( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 303 vperm2i128 ymm6, ymm3, ymm6, 0x31 ; ymm6=row[+1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 304 305 vpmullw ymm0, ymm0, [GOTOFF(ebx,PW_THREE)] 306 vpmullw ymm4, ymm4, [GOTOFF(ebx,PW_THREE)] 307 308 vpcmpeqb xmm7, xmm7, xmm7 309 vpsrldq xmm7, xmm7, (SIZEOF_XMMWORD-2) ; (ffff ---- ---- ... ---- ----) LSB is ffff 310 311 vpaddw ymm1, ymm1, ymm0 ; ymm1=Int0L=( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 312 vpaddw ymm5, ymm5, ymm4 ; ymm5=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 313 vpaddw ymm2, ymm2, ymm0 ; ymm2=Int1L=( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 314 vpaddw ymm6, ymm6, ymm4 ; ymm6=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 315 316 vmovdqu YMMWORD [edx+0*SIZEOF_YMMWORD], ymm1 ; temporarily save 317 vmovdqu YMMWORD [edx+1*SIZEOF_YMMWORD], ymm5 ; the intermediate data 318 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymm2 319 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymm6 320 321 vpand ymm1, ymm1, ymm7 ; ymm1=( 0 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --) 322 vpand ymm2, ymm2, ymm7 ; ymm2=( 0 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --) 323 324 vmovdqa YMMWORD [wk(0)], ymm1 325 vmovdqa YMMWORD [wk(1)], ymm2 326 327 POPPIC ebx 328 329 add eax, byte SIZEOF_YMMWORD-1 330 and eax, byte -SIZEOF_YMMWORD 331 cmp eax, byte SIZEOF_YMMWORD 332 ja short .columnloop 333 ALIGNX 16, 7 334 335 .columnloop_last: 336 ; -- process the last column block 337 338 PUSHPIC ebx 339 MOVPIC ebx, POINTER [gotptr] ; load GOT address 340 341 vpcmpeqb xmm1, xmm1, xmm1 342 vpslldq xmm1, xmm1, (SIZEOF_XMMWORD-2) 343 vperm2i128 ymm1, ymm1, ymm1, 1 ; (---- ---- ... ---- ---- ffff) MSB is ffff 344 345 vpand ymm2, ymm1, YMMWORD [edi+1*SIZEOF_YMMWORD] 346 vpand ymm1, ymm1, YMMWORD [edx+1*SIZEOF_YMMWORD] 347 348 vmovdqa YMMWORD [wk(2)], ymm1 ; ymm1=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 31) 349 vmovdqa YMMWORD [wk(3)], ymm2 ; ymm2=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 31) 350 351 jmp near .upsample 352 ALIGNX 16, 7 353 354 .columnloop: 355 ; -- process the next column block 356 357 vmovdqu ymm0, YMMWORD [ebx+1*SIZEOF_YMMWORD] ; ymm0=row[ 0][1] 358 vmovdqu ymm1, YMMWORD [ecx+1*SIZEOF_YMMWORD] ; ymm1=row[-1][1] 359 vmovdqu ymm2, YMMWORD [esi+1*SIZEOF_YMMWORD] ; ymm2=row[+1][1] 360 361 PUSHPIC ebx 362 MOVPIC ebx, POINTER [gotptr] ; load GOT address 363 364 vpxor ymm3, ymm3, ymm3 ; ymm3=(all 0's) 365 366 vpunpckhbw ymm4, ymm0, ymm3 ; ymm4=row[ 0]( 8 9 10 11 12 13 14 15 24 25 26 27 28 29 30 31) 367 vpunpcklbw ymm5, ymm0, ymm3 ; ymm5=row[ 0]( 0 1 2 3 4 5 6 7 16 17 18 19 20 21 22 23) 368 vperm2i128 ymm0, ymm5, ymm4, 0x20 ; ymm0=row[ 0]( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 369 vperm2i128 ymm4, ymm5, ymm4, 0x31 ; ymm4=row[ 0](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 370 371 vpunpckhbw ymm5, ymm1, ymm3 ; ymm5=row[-1]( 8 9 10 11 12 13 14 15 24 25 26 27 28 29 30 31) 372 vpunpcklbw ymm6, ymm1, ymm3 ; ymm6=row[-1]( 0 1 2 3 4 5 6 7 16 17 18 19 20 21 22 23) 373 vperm2i128 ymm1, ymm6, ymm5, 0x20 ; ymm1=row[-1]( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 374 vperm2i128 ymm5, ymm6, ymm5, 0x31 ; ymm5=row[-1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 375 376 vpunpckhbw ymm6, ymm2, ymm3 ; ymm6=row[+1]( 8 9 10 11 12 13 14 15 24 25 26 27 28 29 30 31) 377 vpunpcklbw ymm7, ymm2, ymm3 ; ymm7=row[+1]( 0 1 2 3 4 5 6 7 16 17 18 19 20 21 22 23) 378 vperm2i128 ymm2, ymm7, ymm6, 0x20 ; ymm2=row[+1]( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 379 vperm2i128 ymm6, ymm7, ymm6, 0x31 ; ymm6=row[+1](16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 380 381 vpmullw ymm0, ymm0, [GOTOFF(ebx,PW_THREE)] 382 vpmullw ymm4, ymm4, [GOTOFF(ebx,PW_THREE)] 383 384 vpaddw ymm1, ymm1, ymm0 ; ymm1=Int0L=( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 385 vpaddw ymm5, ymm5, ymm4 ; ymm5=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 386 vpaddw ymm2, ymm2, ymm0 ; ymm2=Int1L=( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 387 vpaddw ymm6, ymm6, ymm4 ; ymm6=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 388 389 vmovdqu YMMWORD [edx+2*SIZEOF_YMMWORD], ymm1 ; temporarily save 390 vmovdqu YMMWORD [edx+3*SIZEOF_YMMWORD], ymm5 ; the intermediate data 391 vmovdqu YMMWORD [edi+2*SIZEOF_YMMWORD], ymm2 392 vmovdqu YMMWORD [edi+3*SIZEOF_YMMWORD], ymm6 393 394 vperm2i128 ymm1, ymm3, ymm1, 0x20 395 vpslldq ymm1, ymm1, 14 ; ymm1=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 0) 396 vperm2i128 ymm2, ymm3, ymm2, 0x20 397 vpslldq ymm2, ymm2, 14 ; ymm2=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 0) 398 399 vmovdqa YMMWORD [wk(2)], ymm1 400 vmovdqa YMMWORD [wk(3)], ymm2 401 402 .upsample: 403 ; -- process the upper row 404 405 vmovdqu ymm7, YMMWORD [edx+0*SIZEOF_YMMWORD] ; ymm7=Int0L=( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 406 vmovdqu ymm3, YMMWORD [edx+1*SIZEOF_YMMWORD] ; ymm3=Int0H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 407 408 vpxor ymm1, ymm1, ymm1 ; ymm1=(all 0's) 409 410 vperm2i128 ymm0, ymm1, ymm7, 0x03 411 vpalignr ymm0, ymm0, ymm7, 2 ; ymm0=( 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 --) 412 vperm2i128 ymm4, ymm1, ymm3, 0x20 413 vpslldq ymm4, ymm4, 14 ; ymm4=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 16) 414 415 vperm2i128 ymm5, ymm1, ymm7, 0x03 416 vpsrldq ymm5, ymm5, 14 ; ymm5=(15 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --) 417 vperm2i128 ymm6, ymm1, ymm3, 0x20 418 vpalignr ymm6, ymm3, ymm6, 14 ; ymm6=(-- 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30) 419 420 vpor ymm0, ymm0, ymm4 ; ymm0=( 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16) 421 vpor ymm5, ymm5, ymm6 ; ymm5=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30) 422 423 vperm2i128 ymm2, ymm1, ymm3, 0x03 424 vpalignr ymm2, ymm2, ymm3, 2 ; ymm2=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 --) 425 vperm2i128 ymm4, ymm1, ymm3, 0x03 426 vpsrldq ymm4, ymm4, 14 ; ymm4=(31 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --) 427 vperm2i128 ymm1, ymm1, ymm7, 0x20 428 vpalignr ymm1, ymm7, ymm1, 14 ; ymm1=(-- 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14) 429 430 vpor ymm1, ymm1, YMMWORD [wk(0)] ; ymm1=(-1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14) 431 vpor ymm2, ymm2, YMMWORD [wk(2)] ; ymm2=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32) 432 433 vmovdqa YMMWORD [wk(0)], ymm4 434 435 vpmullw ymm7, ymm7, [GOTOFF(ebx,PW_THREE)] 436 vpmullw ymm3, ymm3, [GOTOFF(ebx,PW_THREE)] 437 vpaddw ymm1, ymm1, [GOTOFF(ebx,PW_EIGHT)] 438 vpaddw ymm5, ymm5, [GOTOFF(ebx,PW_EIGHT)] 439 vpaddw ymm0, ymm0, [GOTOFF(ebx,PW_SEVEN)] 440 vpaddw ymm2, [GOTOFF(ebx,PW_SEVEN)] 441 442 vpaddw ymm1, ymm1, ymm7 443 vpaddw ymm5, ymm5, ymm3 444 vpsrlw ymm1, ymm1, 4 ; ymm1=Out0LE=( 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30) 445 vpsrlw ymm5, ymm5, 4 ; ymm5=Out0HE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62) 446 vpaddw ymm0, ymm0, ymm7 447 vpaddw ymm2, ymm2, ymm3 448 vpsrlw ymm0, ymm0, 4 ; ymm0=Out0LO=( 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31) 449 vpsrlw ymm2, ymm2, 4 ; ymm2=Out0HO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63) 450 451 vpsllw ymm0, ymm0, BYTE_BIT 452 vpsllw ymm2, ymm2, BYTE_BIT 453 vpor ymm1, ymm1, ymm0 ; ymm1=Out0L=( 0 1 2 ... 29 30 31) 454 vpor ymm5, ymm5, ymm2 ; ymm5=Out0H=(32 33 34 ... 61 62 63) 455 456 vmovdqu YMMWORD [edx+0*SIZEOF_YMMWORD], ymm1 457 vmovdqu YMMWORD [edx+1*SIZEOF_YMMWORD], ymm5 458 459 ; -- process the lower row 460 461 vmovdqu ymm6, YMMWORD [edi+0*SIZEOF_YMMWORD] ; ymm6=Int1L=( 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15) 462 vmovdqu ymm4, YMMWORD [edi+1*SIZEOF_YMMWORD] ; ymm4=Int1H=(16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31) 463 464 vpxor ymm1, ymm1, ymm1 ; ymm1=(all 0's) 465 466 vperm2i128 ymm7, ymm1, ymm6, 0x03 467 vpalignr ymm7, ymm7, ymm6, 2 ; ymm7=( 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 --) 468 vperm2i128 ymm3, ymm1, ymm4, 0x20 469 vpslldq ymm3, ymm3, 14 ; ymm3=(-- -- -- -- -- -- -- -- -- -- -- -- -- -- -- 16) 470 471 vperm2i128 ymm0, ymm1, ymm6, 0x03 472 vpsrldq ymm0, ymm0, 14 ; ymm0=(15 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --) 473 vperm2i128 ymm2, ymm1, ymm4, 0x20 474 vpalignr ymm2, ymm4, ymm2, 14 ; ymm2=(-- 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30) 475 476 vpor ymm7, ymm7, ymm3 ; ymm7=( 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16) 477 vpor ymm0, ymm0, ymm2 ; ymm0=(15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30) 478 479 vperm2i128 ymm5, ymm1, ymm4, 0x03 480 vpalignr ymm5, ymm5, ymm4, 2 ; ymm5=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 --) 481 vperm2i128 ymm3, ymm1, ymm4, 0x03 482 vpsrldq ymm3, ymm3, 14 ; ymm3=(31 -- -- -- -- -- -- -- -- -- -- -- -- -- -- --) 483 vperm2i128 ymm1, ymm1, ymm6, 0x20 484 vpalignr ymm1, ymm6, ymm1, 14 ; ymm1=(-- 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14) 485 486 vpor ymm1, ymm1, YMMWORD [wk(1)] ; ymm1=(-1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14) 487 vpor ymm5, ymm5, YMMWORD [wk(3)] ; ymm5=(17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32) 488 489 vmovdqa YMMWORD [wk(1)], ymm3 490 491 vpmullw ymm6, ymm6, [GOTOFF(ebx,PW_THREE)] 492 vpmullw ymm4, ymm4, [GOTOFF(ebx,PW_THREE)] 493 vpaddw ymm1, ymm1, [GOTOFF(ebx,PW_EIGHT)] 494 vpaddw ymm0, ymm0, [GOTOFF(ebx,PW_EIGHT)] 495 vpaddw ymm7, ymm7, [GOTOFF(ebx,PW_SEVEN)] 496 vpaddw ymm5, ymm5, [GOTOFF(ebx,PW_SEVEN)] 497 498 vpaddw ymm1, ymm1, ymm6 499 vpaddw ymm0, ymm0, ymm4 500 vpsrlw ymm1, ymm1, 4 ; ymm1=Out1LE=( 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30) 501 vpsrlw ymm0, ymm0, 4 ; ymm0=Out1HE=(32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62) 502 vpaddw ymm7, ymm7, ymm6 503 vpaddw ymm5, ymm5, ymm4 504 vpsrlw ymm7, ymm7, 4 ; ymm7=Out1LO=( 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31) 505 vpsrlw ymm5, ymm5, 4 ; ymm5=Out1HO=(33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63) 506 507 vpsllw ymm7, ymm7, BYTE_BIT 508 vpsllw ymm5, ymm5, BYTE_BIT 509 vpor ymm1, ymm1, ymm7 ; ymm1=Out1L=( 0 1 2 ... 29 30 31) 510 vpor ymm0, ymm0, ymm5 ; ymm0=Out1H=(32 33 34 ... 61 62 63) 511 512 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymm1 513 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymm0 514 515 POPPIC ebx 516 517 sub eax, byte SIZEOF_YMMWORD 518 add ecx, byte 1*SIZEOF_YMMWORD ; inptr1(above) 519 add ebx, byte 1*SIZEOF_YMMWORD ; inptr0 520 add esi, byte 1*SIZEOF_YMMWORD ; inptr1(below) 521 add edx, byte 2*SIZEOF_YMMWORD ; outptr0 522 add edi, byte 2*SIZEOF_YMMWORD ; outptr1 523 cmp eax, byte SIZEOF_YMMWORD 524 ja near .columnloop 525 test eax, eax 526 jnz near .columnloop_last 527 528 pop esi 529 pop edi 530 pop ecx 531 pop eax 532 533 add esi, byte 1*SIZEOF_JSAMPROW ; input_data 534 add edi, byte 2*SIZEOF_JSAMPROW ; output_data 535 sub ecx, byte 2 ; rowctr 536 jg near .rowloop 537 538 .return: 539 vzeroupper 540 pop edi 541 pop esi 542 ; pop edx ; need not be preserved 543 ; pop ecx ; need not be preserved 544 pop ebx 545 mov esp, ebp ; esp <- aligned ebp 546 pop esp ; esp <- original ebp 547 pop ebp 548 ret 549 550 ; -------------------------------------------------------------------------- 551 ; 552 ; Fast processing for the common case of 2:1 horizontal and 1:1 vertical. 553 ; It's still a box filter. 554 ; 555 ; GLOBAL(void) 556 ; jsimd_h2v1_upsample_avx2(int max_v_samp_factor, JDIMENSION output_width, 557 ; JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr); 558 ; 559 560 %define max_v_samp(b) (b) + 8 ; int max_v_samp_factor 561 %define output_width(b) (b) + 12 ; JDIMENSION output_width 562 %define input_data(b) (b) + 16 ; JSAMPARRAY input_data 563 %define output_data_ptr(b) (b) + 20 ; JSAMPARRAY *output_data_ptr 564 565 align 32 566 GLOBAL_FUNCTION(jsimd_h2v1_upsample_avx2) 567 568 EXTN(jsimd_h2v1_upsample_avx2): 569 push ebp 570 mov ebp, esp 571 ; push ebx ; unused 572 ; push ecx ; need not be preserved 573 ; push edx ; need not be preserved 574 push esi 575 push edi 576 577 mov edx, JDIMENSION [output_width(ebp)] 578 add edx, byte (SIZEOF_YMMWORD-1) 579 and edx, -SIZEOF_YMMWORD 580 jz short .return 581 582 mov ecx, INT [max_v_samp(ebp)] ; rowctr 583 test ecx, ecx 584 jz short .return 585 586 mov esi, JSAMPARRAY [input_data(ebp)] ; input_data 587 mov edi, POINTER [output_data_ptr(ebp)] 588 mov edi, JSAMPARRAY [edi] ; output_data 589 ALIGNX 16, 7 590 .rowloop: 591 push edi 592 push esi 593 594 mov esi, JSAMPROW [esi] ; inptr 595 mov edi, JSAMPROW [edi] ; outptr 596 mov eax, edx ; colctr 597 ALIGNX 16, 7 598 .columnloop: 599 600 cmp eax, byte SIZEOF_YMMWORD 601 ja near .above_16 602 603 vmovdqu xmm0, XMMWORD [esi+0*SIZEOF_YMMWORD] 604 vpunpckhbw xmm1, xmm0, xmm0 605 vpunpcklbw xmm0, xmm0, xmm0 606 607 vmovdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0 608 vmovdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmm1 609 610 jmp short .nextrow 611 612 .above_16: 613 vmovdqu ymm0, YMMWORD [esi+0*SIZEOF_YMMWORD] 614 615 vpermq ymm0, ymm0, 0xd8 616 vpunpckhbw ymm1, ymm0, ymm0 617 vpunpcklbw ymm0, ymm0, ymm0 618 619 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymm0 620 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymm1 621 622 sub eax, byte 2*SIZEOF_YMMWORD 623 jz short .nextrow 624 625 add esi, byte SIZEOF_YMMWORD ; inptr 626 add edi, byte 2*SIZEOF_YMMWORD ; outptr 627 jmp short .columnloop 628 ALIGNX 16, 7 629 630 .nextrow: 631 pop esi 632 pop edi 633 634 add esi, byte SIZEOF_JSAMPROW ; input_data 635 add edi, byte SIZEOF_JSAMPROW ; output_data 636 dec ecx ; rowctr 637 jg short .rowloop 638 639 .return: 640 vzeroupper 641 pop edi 642 pop esi 643 ; pop edx ; need not be preserved 644 ; pop ecx ; need not be preserved 645 ; pop ebx ; unused 646 pop ebp 647 ret 648 649 ; -------------------------------------------------------------------------- 650 ; 651 ; Fast processing for the common case of 2:1 horizontal and 2:1 vertical. 652 ; It's still a box filter. 653 ; 654 ; GLOBAL(void) 655 ; jsimd_h2v2_upsample_avx2(int max_v_samp_factor, JDIMENSION output_width, 656 ; JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr); 657 ; 658 659 %define max_v_samp(b) (b) + 8 ; int max_v_samp_factor 660 %define output_width(b) (b) + 12 ; JDIMENSION output_width 661 %define input_data(b) (b) + 16 ; JSAMPARRAY input_data 662 %define output_data_ptr(b) (b) + 20 ; JSAMPARRAY *output_data_ptr 663 664 align 32 665 GLOBAL_FUNCTION(jsimd_h2v2_upsample_avx2) 666 667 EXTN(jsimd_h2v2_upsample_avx2): 668 push ebp 669 mov ebp, esp 670 push ebx 671 ; push ecx ; need not be preserved 672 ; push edx ; need not be preserved 673 push esi 674 push edi 675 676 mov edx, JDIMENSION [output_width(ebp)] 677 add edx, byte (SIZEOF_YMMWORD-1) 678 and edx, -SIZEOF_YMMWORD 679 jz near .return 680 681 mov ecx, INT [max_v_samp(ebp)] ; rowctr 682 test ecx, ecx 683 jz near .return 684 685 mov esi, JSAMPARRAY [input_data(ebp)] ; input_data 686 mov edi, POINTER [output_data_ptr(ebp)] 687 mov edi, JSAMPARRAY [edi] ; output_data 688 ALIGNX 16, 7 689 .rowloop: 690 push edi 691 push esi 692 693 mov esi, JSAMPROW [esi] ; inptr 694 mov ebx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] ; outptr0 695 mov edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW] ; outptr1 696 mov eax, edx ; colctr 697 ALIGNX 16, 7 698 .columnloop: 699 700 cmp eax, byte SIZEOF_YMMWORD 701 ja short .above_16 702 703 vmovdqu xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD] 704 vpunpckhbw xmm1, xmm0, xmm0 705 vpunpcklbw xmm0, xmm0, xmm0 706 707 vmovdqu XMMWORD [ebx+0*SIZEOF_XMMWORD], xmm0 708 vmovdqu XMMWORD [ebx+1*SIZEOF_XMMWORD], xmm1 709 vmovdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0 710 vmovdqu XMMWORD [edi+1*SIZEOF_XMMWORD], xmm1 711 712 jmp near .nextrow 713 714 .above_16: 715 vmovdqu ymm0, YMMWORD [esi+0*SIZEOF_YMMWORD] 716 717 vpermq ymm0, ymm0, 0xd8 718 vpunpckhbw ymm1, ymm0, ymm0 719 vpunpcklbw ymm0, ymm0, ymm0 720 721 vmovdqu YMMWORD [ebx+0*SIZEOF_YMMWORD], ymm0 722 vmovdqu YMMWORD [ebx+1*SIZEOF_YMMWORD], ymm1 723 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymm0 724 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymm1 725 726 sub eax, byte 2*SIZEOF_YMMWORD 727 jz short .nextrow 728 729 add esi, byte SIZEOF_YMMWORD ; inptr 730 add ebx, 2*SIZEOF_YMMWORD ; outptr0 731 add edi, 2*SIZEOF_YMMWORD ; outptr1 732 jmp short .columnloop 733 ALIGNX 16, 7 734 735 .nextrow: 736 pop esi 737 pop edi 738 739 add esi, byte 1*SIZEOF_JSAMPROW ; input_data 740 add edi, byte 2*SIZEOF_JSAMPROW ; output_data 741 sub ecx, byte 2 ; rowctr 742 jg near .rowloop 743 744 .return: 745 vzeroupper 746 pop edi 747 pop esi 748 ; pop edx ; need not be preserved 749 ; pop ecx ; need not be preserved 750 pop ebx 751 pop ebp 752 ret 753 754 ; For some reason, the OS X linker does not honor the request to align the 755 ; segment unless we do this. 756 align 32