1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jdsamss2.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,729 @@
1.4 +;
1.5 +; jdsamss2.asm - upsampling (SSE2)
1.6 +;
1.7 +; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
1.8 +;
1.9 +; Based on
1.10 +; x86 SIMD extension for IJG JPEG library
1.11 +; Copyright (C) 1999-2006, MIYASAKA Masaru.
1.12 +; For conditions of distribution and use, see copyright notice in jsimdext.inc
1.13 +;
1.14 +; This file should be assembled with NASM (Netwide Assembler),
1.15 +; can *not* be assembled with Microsoft's MASM or any compatible
1.16 +; assembler (including Borland's Turbo Assembler).
1.17 +; NASM is available from http://nasm.sourceforge.net/ or
1.18 +; http://sourceforge.net/project/showfiles.php?group_id=6208
1.19 +;
1.20 +; [TAB8]
1.21 +
1.22 +%include "jsimdext.inc"
1.23 +
1.24 +; --------------------------------------------------------------------------
1.25 + SECTION SEG_CONST
1.26 +
1.27 + alignz 16
1.28 + global EXTN(jconst_fancy_upsample_sse2)
1.29 +
1.30 +EXTN(jconst_fancy_upsample_sse2):
1.31 +
1.32 +PW_ONE times 8 dw 1
1.33 +PW_TWO times 8 dw 2
1.34 +PW_THREE times 8 dw 3
1.35 +PW_SEVEN times 8 dw 7
1.36 +PW_EIGHT times 8 dw 8
1.37 +
1.38 + alignz 16
1.39 +
1.40 +; --------------------------------------------------------------------------
1.41 + SECTION SEG_TEXT
1.42 + BITS 32
1.43 +;
1.44 +; Fancy processing for the common case of 2:1 horizontal and 1:1 vertical.
1.45 +;
1.46 +; The upsampling algorithm is linear interpolation between pixel centers,
1.47 +; also known as a "triangle filter". This is a good compromise between
1.48 +; speed and visual quality. The centers of the output pixels are 1/4 and 3/4
1.49 +; of the way between input pixel centers.
1.50 +;
1.51 +; GLOBAL(void)
1.52 +; jsimd_h2v1_fancy_upsample_sse2 (int max_v_samp_factor,
1.53 +; JDIMENSION downsampled_width,
1.54 +; JSAMPARRAY input_data,
1.55 +; JSAMPARRAY * output_data_ptr);
1.56 +;
1.57 +
1.58 +%define max_v_samp(b) (b)+8 ; int max_v_samp_factor
1.59 +%define downsamp_width(b) (b)+12 ; JDIMENSION downsampled_width
1.60 +%define input_data(b) (b)+16 ; JSAMPARRAY input_data
1.61 +%define output_data_ptr(b) (b)+20 ; JSAMPARRAY * output_data_ptr
1.62 +
1.63 + align 16
1.64 + global EXTN(jsimd_h2v1_fancy_upsample_sse2)
1.65 +
1.66 +EXTN(jsimd_h2v1_fancy_upsample_sse2):
1.67 + push ebp
1.68 + mov ebp,esp
1.69 + pushpic ebx
1.70 +; push ecx ; need not be preserved
1.71 +; push edx ; need not be preserved
1.72 + push esi
1.73 + push edi
1.74 +
1.75 + get_GOT ebx ; get GOT address
1.76 +
1.77 + mov eax, JDIMENSION [downsamp_width(ebp)] ; colctr
1.78 + test eax,eax
1.79 + jz near .return
1.80 +
1.81 + mov ecx, INT [max_v_samp(ebp)] ; rowctr
1.82 + test ecx,ecx
1.83 + jz near .return
1.84 +
1.85 + mov esi, JSAMPARRAY [input_data(ebp)] ; input_data
1.86 + mov edi, POINTER [output_data_ptr(ebp)]
1.87 + mov edi, JSAMPARRAY [edi] ; output_data
1.88 + alignx 16,7
1.89 +.rowloop:
1.90 + push eax ; colctr
1.91 + push edi
1.92 + push esi
1.93 +
1.94 + mov esi, JSAMPROW [esi] ; inptr
1.95 + mov edi, JSAMPROW [edi] ; outptr
1.96 +
1.97 + test eax, SIZEOF_XMMWORD-1
1.98 + jz short .skip
1.99 + mov dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE]
1.100 + mov JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl ; insert a dummy sample
1.101 +.skip:
1.102 + pxor xmm0,xmm0 ; xmm0=(all 0's)
1.103 + pcmpeqb xmm7,xmm7
1.104 + psrldq xmm7,(SIZEOF_XMMWORD-1)
1.105 + pand xmm7, XMMWORD [esi+0*SIZEOF_XMMWORD]
1.106 +
1.107 + add eax, byte SIZEOF_XMMWORD-1
1.108 + and eax, byte -SIZEOF_XMMWORD
1.109 + cmp eax, byte SIZEOF_XMMWORD
1.110 + ja short .columnloop
1.111 + alignx 16,7
1.112 +
1.113 +.columnloop_last:
1.114 + pcmpeqb xmm6,xmm6
1.115 + pslldq xmm6,(SIZEOF_XMMWORD-1)
1.116 + pand xmm6, XMMWORD [esi+0*SIZEOF_XMMWORD]
1.117 + jmp short .upsample
1.118 + alignx 16,7
1.119 +
1.120 +.columnloop:
1.121 + movdqa xmm6, XMMWORD [esi+1*SIZEOF_XMMWORD]
1.122 + pslldq xmm6,(SIZEOF_XMMWORD-1)
1.123 +
1.124 +.upsample:
1.125 + movdqa xmm1, XMMWORD [esi+0*SIZEOF_XMMWORD]
1.126 + movdqa xmm2,xmm1
1.127 + movdqa xmm3,xmm1 ; xmm1=( 0 1 2 ... 13 14 15)
1.128 + pslldq xmm2,1 ; xmm2=(-- 0 1 ... 12 13 14)
1.129 + psrldq xmm3,1 ; xmm3=( 1 2 3 ... 14 15 --)
1.130 +
1.131 + por xmm2,xmm7 ; xmm2=(-1 0 1 ... 12 13 14)
1.132 + por xmm3,xmm6 ; xmm3=( 1 2 3 ... 14 15 16)
1.133 +
1.134 + movdqa xmm7,xmm1
1.135 + psrldq xmm7,(SIZEOF_XMMWORD-1) ; xmm7=(15 -- -- ... -- -- --)
1.136 +
1.137 + movdqa xmm4,xmm1
1.138 + punpcklbw xmm1,xmm0 ; xmm1=( 0 1 2 3 4 5 6 7)
1.139 + punpckhbw xmm4,xmm0 ; xmm4=( 8 9 10 11 12 13 14 15)
1.140 + movdqa xmm5,xmm2
1.141 + punpcklbw xmm2,xmm0 ; xmm2=(-1 0 1 2 3 4 5 6)
1.142 + punpckhbw xmm5,xmm0 ; xmm5=( 7 8 9 10 11 12 13 14)
1.143 + movdqa xmm6,xmm3
1.144 + punpcklbw xmm3,xmm0 ; xmm3=( 1 2 3 4 5 6 7 8)
1.145 + punpckhbw xmm6,xmm0 ; xmm6=( 9 10 11 12 13 14 15 16)
1.146 +
1.147 + pmullw xmm1,[GOTOFF(ebx,PW_THREE)]
1.148 + pmullw xmm4,[GOTOFF(ebx,PW_THREE)]
1.149 + paddw xmm2,[GOTOFF(ebx,PW_ONE)]
1.150 + paddw xmm5,[GOTOFF(ebx,PW_ONE)]
1.151 + paddw xmm3,[GOTOFF(ebx,PW_TWO)]
1.152 + paddw xmm6,[GOTOFF(ebx,PW_TWO)]
1.153 +
1.154 + paddw xmm2,xmm1
1.155 + paddw xmm5,xmm4
1.156 + psrlw xmm2,2 ; xmm2=OutLE=( 0 2 4 6 8 10 12 14)
1.157 + psrlw xmm5,2 ; xmm5=OutHE=(16 18 20 22 24 26 28 30)
1.158 + paddw xmm3,xmm1
1.159 + paddw xmm6,xmm4
1.160 + psrlw xmm3,2 ; xmm3=OutLO=( 1 3 5 7 9 11 13 15)
1.161 + psrlw xmm6,2 ; xmm6=OutHO=(17 19 21 23 25 27 29 31)
1.162 +
1.163 + psllw xmm3,BYTE_BIT
1.164 + psllw xmm6,BYTE_BIT
1.165 + por xmm2,xmm3 ; xmm2=OutL=( 0 1 2 ... 13 14 15)
1.166 + por xmm5,xmm6 ; xmm5=OutH=(16 17 18 ... 29 30 31)
1.167 +
1.168 + movdqa XMMWORD [edi+0*SIZEOF_XMMWORD], xmm2
1.169 + movdqa XMMWORD [edi+1*SIZEOF_XMMWORD], xmm5
1.170 +
1.171 + sub eax, byte SIZEOF_XMMWORD
1.172 + add esi, byte 1*SIZEOF_XMMWORD ; inptr
1.173 + add edi, byte 2*SIZEOF_XMMWORD ; outptr
1.174 + cmp eax, byte SIZEOF_XMMWORD
1.175 + ja near .columnloop
1.176 + test eax,eax
1.177 + jnz near .columnloop_last
1.178 +
1.179 + pop esi
1.180 + pop edi
1.181 + pop eax
1.182 +
1.183 + add esi, byte SIZEOF_JSAMPROW ; input_data
1.184 + add edi, byte SIZEOF_JSAMPROW ; output_data
1.185 + dec ecx ; rowctr
1.186 + jg near .rowloop
1.187 +
1.188 +.return:
1.189 + pop edi
1.190 + pop esi
1.191 +; pop edx ; need not be preserved
1.192 +; pop ecx ; need not be preserved
1.193 + poppic ebx
1.194 + pop ebp
1.195 + ret
1.196 +
1.197 +; --------------------------------------------------------------------------
1.198 +;
1.199 +; Fancy processing for the common case of 2:1 horizontal and 2:1 vertical.
1.200 +; Again a triangle filter; see comments for h2v1 case, above.
1.201 +;
1.202 +; GLOBAL(void)
1.203 +; jsimd_h2v2_fancy_upsample_sse2 (int max_v_samp_factor,
1.204 +; JDIMENSION downsampled_width,
1.205 +; JSAMPARRAY input_data,
1.206 +; JSAMPARRAY * output_data_ptr);
1.207 +;
1.208 +
1.209 +%define max_v_samp(b) (b)+8 ; int max_v_samp_factor
1.210 +%define downsamp_width(b) (b)+12 ; JDIMENSION downsampled_width
1.211 +%define input_data(b) (b)+16 ; JSAMPARRAY input_data
1.212 +%define output_data_ptr(b) (b)+20 ; JSAMPARRAY * output_data_ptr
1.213 +
1.214 +%define original_ebp ebp+0
1.215 +%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
1.216 +%define WK_NUM 4
1.217 +%define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr
1.218 +
1.219 + align 16
1.220 + global EXTN(jsimd_h2v2_fancy_upsample_sse2)
1.221 +
1.222 +EXTN(jsimd_h2v2_fancy_upsample_sse2):
1.223 + push ebp
1.224 + mov eax,esp ; eax = original ebp
1.225 + sub esp, byte 4
1.226 + and esp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
1.227 + mov [esp],eax
1.228 + mov ebp,esp ; ebp = aligned ebp
1.229 + lea esp, [wk(0)]
1.230 + pushpic eax ; make a room for GOT address
1.231 + push ebx
1.232 +; push ecx ; need not be preserved
1.233 +; push edx ; need not be preserved
1.234 + push esi
1.235 + push edi
1.236 +
1.237 + get_GOT ebx ; get GOT address
1.238 + movpic POINTER [gotptr], ebx ; save GOT address
1.239 +
1.240 + mov edx,eax ; edx = original ebp
1.241 + mov eax, JDIMENSION [downsamp_width(edx)] ; colctr
1.242 + test eax,eax
1.243 + jz near .return
1.244 +
1.245 + mov ecx, INT [max_v_samp(edx)] ; rowctr
1.246 + test ecx,ecx
1.247 + jz near .return
1.248 +
1.249 + mov esi, JSAMPARRAY [input_data(edx)] ; input_data
1.250 + mov edi, POINTER [output_data_ptr(edx)]
1.251 + mov edi, JSAMPARRAY [edi] ; output_data
1.252 + alignx 16,7
1.253 +.rowloop:
1.254 + push eax ; colctr
1.255 + push ecx
1.256 + push edi
1.257 + push esi
1.258 +
1.259 + mov ecx, JSAMPROW [esi-1*SIZEOF_JSAMPROW] ; inptr1(above)
1.260 + mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; inptr0
1.261 + mov esi, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; inptr1(below)
1.262 + mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] ; outptr0
1.263 + mov edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW] ; outptr1
1.264 +
1.265 + test eax, SIZEOF_XMMWORD-1
1.266 + jz short .skip
1.267 + push edx
1.268 + mov dl, JSAMPLE [ecx+(eax-1)*SIZEOF_JSAMPLE]
1.269 + mov JSAMPLE [ecx+eax*SIZEOF_JSAMPLE], dl
1.270 + mov dl, JSAMPLE [ebx+(eax-1)*SIZEOF_JSAMPLE]
1.271 + mov JSAMPLE [ebx+eax*SIZEOF_JSAMPLE], dl
1.272 + mov dl, JSAMPLE [esi+(eax-1)*SIZEOF_JSAMPLE]
1.273 + mov JSAMPLE [esi+eax*SIZEOF_JSAMPLE], dl ; insert a dummy sample
1.274 + pop edx
1.275 +.skip:
1.276 + ; -- process the first column block
1.277 +
1.278 + movdqa xmm0, XMMWORD [ebx+0*SIZEOF_XMMWORD] ; xmm0=row[ 0][0]
1.279 + movdqa xmm1, XMMWORD [ecx+0*SIZEOF_XMMWORD] ; xmm1=row[-1][0]
1.280 + movdqa xmm2, XMMWORD [esi+0*SIZEOF_XMMWORD] ; xmm2=row[+1][0]
1.281 +
1.282 + pushpic ebx
1.283 + movpic ebx, POINTER [gotptr] ; load GOT address
1.284 +
1.285 + pxor xmm3,xmm3 ; xmm3=(all 0's)
1.286 + movdqa xmm4,xmm0
1.287 + punpcklbw xmm0,xmm3 ; xmm0=row[ 0]( 0 1 2 3 4 5 6 7)
1.288 + punpckhbw xmm4,xmm3 ; xmm4=row[ 0]( 8 9 10 11 12 13 14 15)
1.289 + movdqa xmm5,xmm1
1.290 + punpcklbw xmm1,xmm3 ; xmm1=row[-1]( 0 1 2 3 4 5 6 7)
1.291 + punpckhbw xmm5,xmm3 ; xmm5=row[-1]( 8 9 10 11 12 13 14 15)
1.292 + movdqa xmm6,xmm2
1.293 + punpcklbw xmm2,xmm3 ; xmm2=row[+1]( 0 1 2 3 4 5 6 7)
1.294 + punpckhbw xmm6,xmm3 ; xmm6=row[+1]( 8 9 10 11 12 13 14 15)
1.295 +
1.296 + pmullw xmm0,[GOTOFF(ebx,PW_THREE)]
1.297 + pmullw xmm4,[GOTOFF(ebx,PW_THREE)]
1.298 +
1.299 + pcmpeqb xmm7,xmm7
1.300 + psrldq xmm7,(SIZEOF_XMMWORD-2)
1.301 +
1.302 + paddw xmm1,xmm0 ; xmm1=Int0L=( 0 1 2 3 4 5 6 7)
1.303 + paddw xmm5,xmm4 ; xmm5=Int0H=( 8 9 10 11 12 13 14 15)
1.304 + paddw xmm2,xmm0 ; xmm2=Int1L=( 0 1 2 3 4 5 6 7)
1.305 + paddw xmm6,xmm4 ; xmm6=Int1H=( 8 9 10 11 12 13 14 15)
1.306 +
1.307 + movdqa XMMWORD [edx+0*SIZEOF_XMMWORD], xmm1 ; temporarily save
1.308 + movdqa XMMWORD [edx+1*SIZEOF_XMMWORD], xmm5 ; the intermediate data
1.309 + movdqa XMMWORD [edi+0*SIZEOF_XMMWORD], xmm2
1.310 + movdqa XMMWORD [edi+1*SIZEOF_XMMWORD], xmm6
1.311 +
1.312 + pand xmm1,xmm7 ; xmm1=( 0 -- -- -- -- -- -- --)
1.313 + pand xmm2,xmm7 ; xmm2=( 0 -- -- -- -- -- -- --)
1.314 +
1.315 + movdqa XMMWORD [wk(0)], xmm1
1.316 + movdqa XMMWORD [wk(1)], xmm2
1.317 +
1.318 + poppic ebx
1.319 +
1.320 + add eax, byte SIZEOF_XMMWORD-1
1.321 + and eax, byte -SIZEOF_XMMWORD
1.322 + cmp eax, byte SIZEOF_XMMWORD
1.323 + ja short .columnloop
1.324 + alignx 16,7
1.325 +
1.326 +.columnloop_last:
1.327 + ; -- process the last column block
1.328 +
1.329 + pushpic ebx
1.330 + movpic ebx, POINTER [gotptr] ; load GOT address
1.331 +
1.332 + pcmpeqb xmm1,xmm1
1.333 + pslldq xmm1,(SIZEOF_XMMWORD-2)
1.334 + movdqa xmm2,xmm1
1.335 +
1.336 + pand xmm1, XMMWORD [edx+1*SIZEOF_XMMWORD]
1.337 + pand xmm2, XMMWORD [edi+1*SIZEOF_XMMWORD]
1.338 +
1.339 + movdqa XMMWORD [wk(2)], xmm1 ; xmm1=(-- -- -- -- -- -- -- 15)
1.340 + movdqa XMMWORD [wk(3)], xmm2 ; xmm2=(-- -- -- -- -- -- -- 15)
1.341 +
1.342 + jmp near .upsample
1.343 + alignx 16,7
1.344 +
1.345 +.columnloop:
1.346 + ; -- process the next column block
1.347 +
1.348 + movdqa xmm0, XMMWORD [ebx+1*SIZEOF_XMMWORD] ; xmm0=row[ 0][1]
1.349 + movdqa xmm1, XMMWORD [ecx+1*SIZEOF_XMMWORD] ; xmm1=row[-1][1]
1.350 + movdqa xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD] ; xmm2=row[+1][1]
1.351 +
1.352 + pushpic ebx
1.353 + movpic ebx, POINTER [gotptr] ; load GOT address
1.354 +
1.355 + pxor xmm3,xmm3 ; xmm3=(all 0's)
1.356 + movdqa xmm4,xmm0
1.357 + punpcklbw xmm0,xmm3 ; xmm0=row[ 0]( 0 1 2 3 4 5 6 7)
1.358 + punpckhbw xmm4,xmm3 ; xmm4=row[ 0]( 8 9 10 11 12 13 14 15)
1.359 + movdqa xmm5,xmm1
1.360 + punpcklbw xmm1,xmm3 ; xmm1=row[-1]( 0 1 2 3 4 5 6 7)
1.361 + punpckhbw xmm5,xmm3 ; xmm5=row[-1]( 8 9 10 11 12 13 14 15)
1.362 + movdqa xmm6,xmm2
1.363 + punpcklbw xmm2,xmm3 ; xmm2=row[+1]( 0 1 2 3 4 5 6 7)
1.364 + punpckhbw xmm6,xmm3 ; xmm6=row[+1]( 8 9 10 11 12 13 14 15)
1.365 +
1.366 + pmullw xmm0,[GOTOFF(ebx,PW_THREE)]
1.367 + pmullw xmm4,[GOTOFF(ebx,PW_THREE)]
1.368 +
1.369 + paddw xmm1,xmm0 ; xmm1=Int0L=( 0 1 2 3 4 5 6 7)
1.370 + paddw xmm5,xmm4 ; xmm5=Int0H=( 8 9 10 11 12 13 14 15)
1.371 + paddw xmm2,xmm0 ; xmm2=Int1L=( 0 1 2 3 4 5 6 7)
1.372 + paddw xmm6,xmm4 ; xmm6=Int1H=( 8 9 10 11 12 13 14 15)
1.373 +
1.374 + movdqa XMMWORD [edx+2*SIZEOF_XMMWORD], xmm1 ; temporarily save
1.375 + movdqa XMMWORD [edx+3*SIZEOF_XMMWORD], xmm5 ; the intermediate data
1.376 + movdqa XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
1.377 + movdqa XMMWORD [edi+3*SIZEOF_XMMWORD], xmm6
1.378 +
1.379 + pslldq xmm1,(SIZEOF_XMMWORD-2) ; xmm1=(-- -- -- -- -- -- -- 0)
1.380 + pslldq xmm2,(SIZEOF_XMMWORD-2) ; xmm2=(-- -- -- -- -- -- -- 0)
1.381 +
1.382 + movdqa XMMWORD [wk(2)], xmm1
1.383 + movdqa XMMWORD [wk(3)], xmm2
1.384 +
1.385 +.upsample:
1.386 + ; -- process the upper row
1.387 +
1.388 + movdqa xmm7, XMMWORD [edx+0*SIZEOF_XMMWORD]
1.389 + movdqa xmm3, XMMWORD [edx+1*SIZEOF_XMMWORD]
1.390 +
1.391 + movdqa xmm0,xmm7 ; xmm7=Int0L=( 0 1 2 3 4 5 6 7)
1.392 + movdqa xmm4,xmm3 ; xmm3=Int0H=( 8 9 10 11 12 13 14 15)
1.393 + psrldq xmm0,2 ; xmm0=( 1 2 3 4 5 6 7 --)
1.394 + pslldq xmm4,(SIZEOF_XMMWORD-2) ; xmm4=(-- -- -- -- -- -- -- 8)
1.395 + movdqa xmm5,xmm7
1.396 + movdqa xmm6,xmm3
1.397 + psrldq xmm5,(SIZEOF_XMMWORD-2) ; xmm5=( 7 -- -- -- -- -- -- --)
1.398 + pslldq xmm6,2 ; xmm6=(-- 8 9 10 11 12 13 14)
1.399 +
1.400 + por xmm0,xmm4 ; xmm0=( 1 2 3 4 5 6 7 8)
1.401 + por xmm5,xmm6 ; xmm5=( 7 8 9 10 11 12 13 14)
1.402 +
1.403 + movdqa xmm1,xmm7
1.404 + movdqa xmm2,xmm3
1.405 + pslldq xmm1,2 ; xmm1=(-- 0 1 2 3 4 5 6)
1.406 + psrldq xmm2,2 ; xmm2=( 9 10 11 12 13 14 15 --)
1.407 + movdqa xmm4,xmm3
1.408 + psrldq xmm4,(SIZEOF_XMMWORD-2) ; xmm4=(15 -- -- -- -- -- -- --)
1.409 +
1.410 + por xmm1, XMMWORD [wk(0)] ; xmm1=(-1 0 1 2 3 4 5 6)
1.411 + por xmm2, XMMWORD [wk(2)] ; xmm2=( 9 10 11 12 13 14 15 16)
1.412 +
1.413 + movdqa XMMWORD [wk(0)], xmm4
1.414 +
1.415 + pmullw xmm7,[GOTOFF(ebx,PW_THREE)]
1.416 + pmullw xmm3,[GOTOFF(ebx,PW_THREE)]
1.417 + paddw xmm1,[GOTOFF(ebx,PW_EIGHT)]
1.418 + paddw xmm5,[GOTOFF(ebx,PW_EIGHT)]
1.419 + paddw xmm0,[GOTOFF(ebx,PW_SEVEN)]
1.420 + paddw xmm2,[GOTOFF(ebx,PW_SEVEN)]
1.421 +
1.422 + paddw xmm1,xmm7
1.423 + paddw xmm5,xmm3
1.424 + psrlw xmm1,4 ; xmm1=Out0LE=( 0 2 4 6 8 10 12 14)
1.425 + psrlw xmm5,4 ; xmm5=Out0HE=(16 18 20 22 24 26 28 30)
1.426 + paddw xmm0,xmm7
1.427 + paddw xmm2,xmm3
1.428 + psrlw xmm0,4 ; xmm0=Out0LO=( 1 3 5 7 9 11 13 15)
1.429 + psrlw xmm2,4 ; xmm2=Out0HO=(17 19 21 23 25 27 29 31)
1.430 +
1.431 + psllw xmm0,BYTE_BIT
1.432 + psllw xmm2,BYTE_BIT
1.433 + por xmm1,xmm0 ; xmm1=Out0L=( 0 1 2 ... 13 14 15)
1.434 + por xmm5,xmm2 ; xmm5=Out0H=(16 17 18 ... 29 30 31)
1.435 +
1.436 + movdqa XMMWORD [edx+0*SIZEOF_XMMWORD], xmm1
1.437 + movdqa XMMWORD [edx+1*SIZEOF_XMMWORD], xmm5
1.438 +
1.439 + ; -- process the lower row
1.440 +
1.441 + movdqa xmm6, XMMWORD [edi+0*SIZEOF_XMMWORD]
1.442 + movdqa xmm4, XMMWORD [edi+1*SIZEOF_XMMWORD]
1.443 +
1.444 + movdqa xmm7,xmm6 ; xmm6=Int1L=( 0 1 2 3 4 5 6 7)
1.445 + movdqa xmm3,xmm4 ; xmm4=Int1H=( 8 9 10 11 12 13 14 15)
1.446 + psrldq xmm7,2 ; xmm7=( 1 2 3 4 5 6 7 --)
1.447 + pslldq xmm3,(SIZEOF_XMMWORD-2) ; xmm3=(-- -- -- -- -- -- -- 8)
1.448 + movdqa xmm0,xmm6
1.449 + movdqa xmm2,xmm4
1.450 + psrldq xmm0,(SIZEOF_XMMWORD-2) ; xmm0=( 7 -- -- -- -- -- -- --)
1.451 + pslldq xmm2,2 ; xmm2=(-- 8 9 10 11 12 13 14)
1.452 +
1.453 + por xmm7,xmm3 ; xmm7=( 1 2 3 4 5 6 7 8)
1.454 + por xmm0,xmm2 ; xmm0=( 7 8 9 10 11 12 13 14)
1.455 +
1.456 + movdqa xmm1,xmm6
1.457 + movdqa xmm5,xmm4
1.458 + pslldq xmm1,2 ; xmm1=(-- 0 1 2 3 4 5 6)
1.459 + psrldq xmm5,2 ; xmm5=( 9 10 11 12 13 14 15 --)
1.460 + movdqa xmm3,xmm4
1.461 + psrldq xmm3,(SIZEOF_XMMWORD-2) ; xmm3=(15 -- -- -- -- -- -- --)
1.462 +
1.463 + por xmm1, XMMWORD [wk(1)] ; xmm1=(-1 0 1 2 3 4 5 6)
1.464 + por xmm5, XMMWORD [wk(3)] ; xmm5=( 9 10 11 12 13 14 15 16)
1.465 +
1.466 + movdqa XMMWORD [wk(1)], xmm3
1.467 +
1.468 + pmullw xmm6,[GOTOFF(ebx,PW_THREE)]
1.469 + pmullw xmm4,[GOTOFF(ebx,PW_THREE)]
1.470 + paddw xmm1,[GOTOFF(ebx,PW_EIGHT)]
1.471 + paddw xmm0,[GOTOFF(ebx,PW_EIGHT)]
1.472 + paddw xmm7,[GOTOFF(ebx,PW_SEVEN)]
1.473 + paddw xmm5,[GOTOFF(ebx,PW_SEVEN)]
1.474 +
1.475 + paddw xmm1,xmm6
1.476 + paddw xmm0,xmm4
1.477 + psrlw xmm1,4 ; xmm1=Out1LE=( 0 2 4 6 8 10 12 14)
1.478 + psrlw xmm0,4 ; xmm0=Out1HE=(16 18 20 22 24 26 28 30)
1.479 + paddw xmm7,xmm6
1.480 + paddw xmm5,xmm4
1.481 + psrlw xmm7,4 ; xmm7=Out1LO=( 1 3 5 7 9 11 13 15)
1.482 + psrlw xmm5,4 ; xmm5=Out1HO=(17 19 21 23 25 27 29 31)
1.483 +
1.484 + psllw xmm7,BYTE_BIT
1.485 + psllw xmm5,BYTE_BIT
1.486 + por xmm1,xmm7 ; xmm1=Out1L=( 0 1 2 ... 13 14 15)
1.487 + por xmm0,xmm5 ; xmm0=Out1H=(16 17 18 ... 29 30 31)
1.488 +
1.489 + movdqa XMMWORD [edi+0*SIZEOF_XMMWORD], xmm1
1.490 + movdqa XMMWORD [edi+1*SIZEOF_XMMWORD], xmm0
1.491 +
1.492 + poppic ebx
1.493 +
1.494 + sub eax, byte SIZEOF_XMMWORD
1.495 + add ecx, byte 1*SIZEOF_XMMWORD ; inptr1(above)
1.496 + add ebx, byte 1*SIZEOF_XMMWORD ; inptr0
1.497 + add esi, byte 1*SIZEOF_XMMWORD ; inptr1(below)
1.498 + add edx, byte 2*SIZEOF_XMMWORD ; outptr0
1.499 + add edi, byte 2*SIZEOF_XMMWORD ; outptr1
1.500 + cmp eax, byte SIZEOF_XMMWORD
1.501 + ja near .columnloop
1.502 + test eax,eax
1.503 + jnz near .columnloop_last
1.504 +
1.505 + pop esi
1.506 + pop edi
1.507 + pop ecx
1.508 + pop eax
1.509 +
1.510 + add esi, byte 1*SIZEOF_JSAMPROW ; input_data
1.511 + add edi, byte 2*SIZEOF_JSAMPROW ; output_data
1.512 + sub ecx, byte 2 ; rowctr
1.513 + jg near .rowloop
1.514 +
1.515 +.return:
1.516 + pop edi
1.517 + pop esi
1.518 +; pop edx ; need not be preserved
1.519 +; pop ecx ; need not be preserved
1.520 + pop ebx
1.521 + mov esp,ebp ; esp <- aligned ebp
1.522 + pop esp ; esp <- original ebp
1.523 + pop ebp
1.524 + ret
1.525 +
1.526 +; --------------------------------------------------------------------------
1.527 +;
1.528 +; Fast processing for the common case of 2:1 horizontal and 1:1 vertical.
1.529 +; It's still a box filter.
1.530 +;
1.531 +; GLOBAL(void)
1.532 +; jsimd_h2v1_upsample_sse2 (int max_v_samp_factor,
1.533 +; JDIMENSION output_width,
1.534 +; JSAMPARRAY input_data,
1.535 +; JSAMPARRAY * output_data_ptr);
1.536 +;
1.537 +
1.538 +%define max_v_samp(b) (b)+8 ; int max_v_samp_factor
1.539 +%define output_width(b) (b)+12 ; JDIMENSION output_width
1.540 +%define input_data(b) (b)+16 ; JSAMPARRAY input_data
1.541 +%define output_data_ptr(b) (b)+20 ; JSAMPARRAY * output_data_ptr
1.542 +
1.543 + align 16
1.544 + global EXTN(jsimd_h2v1_upsample_sse2)
1.545 +
1.546 +EXTN(jsimd_h2v1_upsample_sse2):
1.547 + push ebp
1.548 + mov ebp,esp
1.549 +; push ebx ; unused
1.550 +; push ecx ; need not be preserved
1.551 +; push edx ; need not be preserved
1.552 + push esi
1.553 + push edi
1.554 +
1.555 + mov edx, JDIMENSION [output_width(ebp)]
1.556 + add edx, byte (2*SIZEOF_XMMWORD)-1
1.557 + and edx, byte -(2*SIZEOF_XMMWORD)
1.558 + jz short .return
1.559 +
1.560 + mov ecx, INT [max_v_samp(ebp)] ; rowctr
1.561 + test ecx,ecx
1.562 + jz short .return
1.563 +
1.564 + mov esi, JSAMPARRAY [input_data(ebp)] ; input_data
1.565 + mov edi, POINTER [output_data_ptr(ebp)]
1.566 + mov edi, JSAMPARRAY [edi] ; output_data
1.567 + alignx 16,7
1.568 +.rowloop:
1.569 + push edi
1.570 + push esi
1.571 +
1.572 + mov esi, JSAMPROW [esi] ; inptr
1.573 + mov edi, JSAMPROW [edi] ; outptr
1.574 + mov eax,edx ; colctr
1.575 + alignx 16,7
1.576 +.columnloop:
1.577 +
1.578 + movdqa xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
1.579 +
1.580 + movdqa xmm1,xmm0
1.581 + punpcklbw xmm0,xmm0
1.582 + punpckhbw xmm1,xmm1
1.583 +
1.584 + movdqa XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
1.585 + movdqa XMMWORD [edi+1*SIZEOF_XMMWORD], xmm1
1.586 +
1.587 + sub eax, byte 2*SIZEOF_XMMWORD
1.588 + jz short .nextrow
1.589 +
1.590 + movdqa xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD]
1.591 +
1.592 + movdqa xmm3,xmm2
1.593 + punpcklbw xmm2,xmm2
1.594 + punpckhbw xmm3,xmm3
1.595 +
1.596 + movdqa XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
1.597 + movdqa XMMWORD [edi+3*SIZEOF_XMMWORD], xmm3
1.598 +
1.599 + sub eax, byte 2*SIZEOF_XMMWORD
1.600 + jz short .nextrow
1.601 +
1.602 + add esi, byte 2*SIZEOF_XMMWORD ; inptr
1.603 + add edi, byte 4*SIZEOF_XMMWORD ; outptr
1.604 + jmp short .columnloop
1.605 + alignx 16,7
1.606 +
1.607 +.nextrow:
1.608 + pop esi
1.609 + pop edi
1.610 +
1.611 + add esi, byte SIZEOF_JSAMPROW ; input_data
1.612 + add edi, byte SIZEOF_JSAMPROW ; output_data
1.613 + dec ecx ; rowctr
1.614 + jg short .rowloop
1.615 +
1.616 +.return:
1.617 + pop edi
1.618 + pop esi
1.619 +; pop edx ; need not be preserved
1.620 +; pop ecx ; need not be preserved
1.621 +; pop ebx ; unused
1.622 + pop ebp
1.623 + ret
1.624 +
1.625 +; --------------------------------------------------------------------------
1.626 +;
1.627 +; Fast processing for the common case of 2:1 horizontal and 2:1 vertical.
1.628 +; It's still a box filter.
1.629 +;
1.630 +; GLOBAL(void)
1.631 +; jsimd_h2v2_upsample_sse2 (nt max_v_samp_factor,
1.632 +; JDIMENSION output_width,
1.633 +; JSAMPARRAY input_data,
1.634 +; JSAMPARRAY * output_data_ptr);
1.635 +;
1.636 +
1.637 +%define max_v_samp(b) (b)+8 ; int max_v_samp_factor
1.638 +%define output_width(b) (b)+12 ; JDIMENSION output_width
1.639 +%define input_data(b) (b)+16 ; JSAMPARRAY input_data
1.640 +%define output_data_ptr(b) (b)+20 ; JSAMPARRAY * output_data_ptr
1.641 +
1.642 + align 16
1.643 + global EXTN(jsimd_h2v2_upsample_sse2)
1.644 +
1.645 +EXTN(jsimd_h2v2_upsample_sse2):
1.646 + push ebp
1.647 + mov ebp,esp
1.648 + push ebx
1.649 +; push ecx ; need not be preserved
1.650 +; push edx ; need not be preserved
1.651 + push esi
1.652 + push edi
1.653 +
1.654 + mov edx, JDIMENSION [output_width(ebp)]
1.655 + add edx, byte (2*SIZEOF_XMMWORD)-1
1.656 + and edx, byte -(2*SIZEOF_XMMWORD)
1.657 + jz near .return
1.658 +
1.659 + mov ecx, INT [max_v_samp(ebp)] ; rowctr
1.660 + test ecx,ecx
1.661 + jz near .return
1.662 +
1.663 + mov esi, JSAMPARRAY [input_data(ebp)] ; input_data
1.664 + mov edi, POINTER [output_data_ptr(ebp)]
1.665 + mov edi, JSAMPARRAY [edi] ; output_data
1.666 + alignx 16,7
1.667 +.rowloop:
1.668 + push edi
1.669 + push esi
1.670 +
1.671 + mov esi, JSAMPROW [esi] ; inptr
1.672 + mov ebx, JSAMPROW [edi+0*SIZEOF_JSAMPROW] ; outptr0
1.673 + mov edi, JSAMPROW [edi+1*SIZEOF_JSAMPROW] ; outptr1
1.674 + mov eax,edx ; colctr
1.675 + alignx 16,7
1.676 +.columnloop:
1.677 +
1.678 + movdqa xmm0, XMMWORD [esi+0*SIZEOF_XMMWORD]
1.679 +
1.680 + movdqa xmm1,xmm0
1.681 + punpcklbw xmm0,xmm0
1.682 + punpckhbw xmm1,xmm1
1.683 +
1.684 + movdqa XMMWORD [ebx+0*SIZEOF_XMMWORD], xmm0
1.685 + movdqa XMMWORD [ebx+1*SIZEOF_XMMWORD], xmm1
1.686 + movdqa XMMWORD [edi+0*SIZEOF_XMMWORD], xmm0
1.687 + movdqa XMMWORD [edi+1*SIZEOF_XMMWORD], xmm1
1.688 +
1.689 + sub eax, byte 2*SIZEOF_XMMWORD
1.690 + jz short .nextrow
1.691 +
1.692 + movdqa xmm2, XMMWORD [esi+1*SIZEOF_XMMWORD]
1.693 +
1.694 + movdqa xmm3,xmm2
1.695 + punpcklbw xmm2,xmm2
1.696 + punpckhbw xmm3,xmm3
1.697 +
1.698 + movdqa XMMWORD [ebx+2*SIZEOF_XMMWORD], xmm2
1.699 + movdqa XMMWORD [ebx+3*SIZEOF_XMMWORD], xmm3
1.700 + movdqa XMMWORD [edi+2*SIZEOF_XMMWORD], xmm2
1.701 + movdqa XMMWORD [edi+3*SIZEOF_XMMWORD], xmm3
1.702 +
1.703 + sub eax, byte 2*SIZEOF_XMMWORD
1.704 + jz short .nextrow
1.705 +
1.706 + add esi, byte 2*SIZEOF_XMMWORD ; inptr
1.707 + add ebx, byte 4*SIZEOF_XMMWORD ; outptr0
1.708 + add edi, byte 4*SIZEOF_XMMWORD ; outptr1
1.709 + jmp short .columnloop
1.710 + alignx 16,7
1.711 +
1.712 +.nextrow:
1.713 + pop esi
1.714 + pop edi
1.715 +
1.716 + add esi, byte 1*SIZEOF_JSAMPROW ; input_data
1.717 + add edi, byte 2*SIZEOF_JSAMPROW ; output_data
1.718 + sub ecx, byte 2 ; rowctr
1.719 + jg short .rowloop
1.720 +
1.721 +.return:
1.722 + pop edi
1.723 + pop esi
1.724 +; pop edx ; need not be preserved
1.725 +; pop ecx ; need not be preserved
1.726 + pop ebx
1.727 + pop ebp
1.728 + ret
1.729 +
1.730 +; For some reason, the OS X linker does not honor the request to align the
1.731 +; segment unless we do this.
1.732 + align 16
