1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jdclrss2-64.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,441 @@
1.4 +;
1.5 +; jdclrss2-64.asm - colorspace conversion (64-bit SSE2)
1.6 +;
1.7 +; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
1.8 +; Copyright 2009, 2012 D. R. Commander
1.9 +;
1.10 +; Based on
1.11 +; x86 SIMD extension for IJG JPEG library
1.12 +; Copyright (C) 1999-2006, MIYASAKA Masaru.
1.13 +; For conditions of distribution and use, see copyright notice in jsimdext.inc
1.14 +;
1.15 +; This file should be assembled with NASM (Netwide Assembler),
1.16 +; can *not* be assembled with Microsoft's MASM or any compatible
1.17 +; assembler (including Borland's Turbo Assembler).
1.18 +; NASM is available from http://nasm.sourceforge.net/ or
1.19 +; http://sourceforge.net/project/showfiles.php?group_id=6208
1.20 +;
1.21 +; [TAB8]
1.22 +
1.23 +%include "jcolsamp.inc"
1.24 +
1.25 +; --------------------------------------------------------------------------
1.26 +;
1.27 +; Convert some rows of samples to the output colorspace.
1.28 +;
1.29 +; GLOBAL(void)
1.30 +; jsimd_ycc_rgb_convert_sse2 (JDIMENSION out_width,
1.31 +; JSAMPIMAGE input_buf, JDIMENSION input_row,
1.32 +; JSAMPARRAY output_buf, int num_rows)
1.33 +;
1.34 +
1.35 +; r10 = JDIMENSION out_width
1.36 +; r11 = JSAMPIMAGE input_buf
1.37 +; r12 = JDIMENSION input_row
1.38 +; r13 = JSAMPARRAY output_buf
1.39 +; r14 = int num_rows
1.40 +
1.41 +%define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
1.42 +%define WK_NUM 2
1.43 +
1.44 + align 16
1.45 + global EXTN(jsimd_ycc_rgb_convert_sse2)
1.46 +
1.47 +EXTN(jsimd_ycc_rgb_convert_sse2):
1.48 + push rbp
1.49 + mov rax,rsp ; rax = original rbp
1.50 + sub rsp, byte 4
1.51 + and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
1.52 + mov [rsp],rax
1.53 + mov rbp,rsp ; rbp = aligned rbp
1.54 + lea rsp, [wk(0)]
1.55 + collect_args
1.56 + push rbx
1.57 +
1.58 + mov rcx, r10 ; num_cols
1.59 + test rcx,rcx
1.60 + jz near .return
1.61 +
1.62 + push rcx
1.63 +
1.64 + mov rdi, r11
1.65 + mov rcx, r12
1.66 + mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
1.67 + mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
1.68 + mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
1.69 + lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
1.70 + lea rbx, [rbx+rcx*SIZEOF_JSAMPROW]
1.71 + lea rdx, [rdx+rcx*SIZEOF_JSAMPROW]
1.72 +
1.73 + pop rcx
1.74 +
1.75 + mov rdi, r13
1.76 + mov eax, r14d
1.77 + test rax,rax
1.78 + jle near .return
1.79 +.rowloop:
1.80 + push rax
1.81 + push rdi
1.82 + push rdx
1.83 + push rbx
1.84 + push rsi
1.85 + push rcx ; col
1.86 +
1.87 + mov rsi, JSAMPROW [rsi] ; inptr0
1.88 + mov rbx, JSAMPROW [rbx] ; inptr1
1.89 + mov rdx, JSAMPROW [rdx] ; inptr2
1.90 + mov rdi, JSAMPROW [rdi] ; outptr
1.91 +.columnloop:
1.92 +
1.93 + movdqa xmm5, XMMWORD [rbx] ; xmm5=Cb(0123456789ABCDEF)
1.94 + movdqa xmm1, XMMWORD [rdx] ; xmm1=Cr(0123456789ABCDEF)
1.95 +
1.96 + pcmpeqw xmm4,xmm4
1.97 + pcmpeqw xmm7,xmm7
1.98 + psrlw xmm4,BYTE_BIT
1.99 + psllw xmm7,7 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
1.100 + movdqa xmm0,xmm4 ; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
1.101 +
1.102 + pand xmm4,xmm5 ; xmm4=Cb(02468ACE)=CbE
1.103 + psrlw xmm5,BYTE_BIT ; xmm5=Cb(13579BDF)=CbO
1.104 + pand xmm0,xmm1 ; xmm0=Cr(02468ACE)=CrE
1.105 + psrlw xmm1,BYTE_BIT ; xmm1=Cr(13579BDF)=CrO
1.106 +
1.107 + paddw xmm4,xmm7
1.108 + paddw xmm5,xmm7
1.109 + paddw xmm0,xmm7
1.110 + paddw xmm1,xmm7
1.111 +
1.112 + ; (Original)
1.113 + ; R = Y + 1.40200 * Cr
1.114 + ; G = Y - 0.34414 * Cb - 0.71414 * Cr
1.115 + ; B = Y + 1.77200 * Cb
1.116 + ;
1.117 + ; (This implementation)
1.118 + ; R = Y + 0.40200 * Cr + Cr
1.119 + ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
1.120 + ; B = Y - 0.22800 * Cb + Cb + Cb
1.121 +
1.122 + movdqa xmm2,xmm4 ; xmm2=CbE
1.123 + movdqa xmm3,xmm5 ; xmm3=CbO
1.124 + paddw xmm4,xmm4 ; xmm4=2*CbE
1.125 + paddw xmm5,xmm5 ; xmm5=2*CbO
1.126 + movdqa xmm6,xmm0 ; xmm6=CrE
1.127 + movdqa xmm7,xmm1 ; xmm7=CrO
1.128 + paddw xmm0,xmm0 ; xmm0=2*CrE
1.129 + paddw xmm1,xmm1 ; xmm1=2*CrO
1.130 +
1.131 + pmulhw xmm4,[rel PW_MF0228] ; xmm4=(2*CbE * -FIX(0.22800))
1.132 + pmulhw xmm5,[rel PW_MF0228] ; xmm5=(2*CbO * -FIX(0.22800))
1.133 + pmulhw xmm0,[rel PW_F0402] ; xmm0=(2*CrE * FIX(0.40200))
1.134 + pmulhw xmm1,[rel PW_F0402] ; xmm1=(2*CrO * FIX(0.40200))
1.135 +
1.136 + paddw xmm4,[rel PW_ONE]
1.137 + paddw xmm5,[rel PW_ONE]
1.138 + psraw xmm4,1 ; xmm4=(CbE * -FIX(0.22800))
1.139 + psraw xmm5,1 ; xmm5=(CbO * -FIX(0.22800))
1.140 + paddw xmm0,[rel PW_ONE]
1.141 + paddw xmm1,[rel PW_ONE]
1.142 + psraw xmm0,1 ; xmm0=(CrE * FIX(0.40200))
1.143 + psraw xmm1,1 ; xmm1=(CrO * FIX(0.40200))
1.144 +
1.145 + paddw xmm4,xmm2
1.146 + paddw xmm5,xmm3
1.147 + paddw xmm4,xmm2 ; xmm4=(CbE * FIX(1.77200))=(B-Y)E
1.148 + paddw xmm5,xmm3 ; xmm5=(CbO * FIX(1.77200))=(B-Y)O
1.149 + paddw xmm0,xmm6 ; xmm0=(CrE * FIX(1.40200))=(R-Y)E
1.150 + paddw xmm1,xmm7 ; xmm1=(CrO * FIX(1.40200))=(R-Y)O
1.151 +
1.152 + movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=(B-Y)E
1.153 + movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(B-Y)O
1.154 +
1.155 + movdqa xmm4,xmm2
1.156 + movdqa xmm5,xmm3
1.157 + punpcklwd xmm2,xmm6
1.158 + punpckhwd xmm4,xmm6
1.159 + pmaddwd xmm2,[rel PW_MF0344_F0285]
1.160 + pmaddwd xmm4,[rel PW_MF0344_F0285]
1.161 + punpcklwd xmm3,xmm7
1.162 + punpckhwd xmm5,xmm7
1.163 + pmaddwd xmm3,[rel PW_MF0344_F0285]
1.164 + pmaddwd xmm5,[rel PW_MF0344_F0285]
1.165 +
1.166 + paddd xmm2,[rel PD_ONEHALF]
1.167 + paddd xmm4,[rel PD_ONEHALF]
1.168 + psrad xmm2,SCALEBITS
1.169 + psrad xmm4,SCALEBITS
1.170 + paddd xmm3,[rel PD_ONEHALF]
1.171 + paddd xmm5,[rel PD_ONEHALF]
1.172 + psrad xmm3,SCALEBITS
1.173 + psrad xmm5,SCALEBITS
1.174 +
1.175 + packssdw xmm2,xmm4 ; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
1.176 + packssdw xmm3,xmm5 ; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
1.177 + psubw xmm2,xmm6 ; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
1.178 + psubw xmm3,xmm7 ; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
1.179 +
1.180 + movdqa xmm5, XMMWORD [rsi] ; xmm5=Y(0123456789ABCDEF)
1.181 +
1.182 + pcmpeqw xmm4,xmm4
1.183 + psrlw xmm4,BYTE_BIT ; xmm4={0xFF 0x00 0xFF 0x00 ..}
1.184 + pand xmm4,xmm5 ; xmm4=Y(02468ACE)=YE
1.185 + psrlw xmm5,BYTE_BIT ; xmm5=Y(13579BDF)=YO
1.186 +
1.187 + paddw xmm0,xmm4 ; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
1.188 + paddw xmm1,xmm5 ; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
1.189 + packuswb xmm0,xmm0 ; xmm0=R(02468ACE********)
1.190 + packuswb xmm1,xmm1 ; xmm1=R(13579BDF********)
1.191 +
1.192 + paddw xmm2,xmm4 ; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
1.193 + paddw xmm3,xmm5 ; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
1.194 + packuswb xmm2,xmm2 ; xmm2=G(02468ACE********)
1.195 + packuswb xmm3,xmm3 ; xmm3=G(13579BDF********)
1.196 +
1.197 + paddw xmm4, XMMWORD [wk(0)] ; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
1.198 + paddw xmm5, XMMWORD [wk(1)] ; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
1.199 + packuswb xmm4,xmm4 ; xmm4=B(02468ACE********)
1.200 + packuswb xmm5,xmm5 ; xmm5=B(13579BDF********)
1.201 +
1.202 +%if RGB_PIXELSIZE == 3 ; ---------------
1.203 +
1.204 + ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
1.205 + ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
1.206 + ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
1.207 + ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
1.208 +
1.209 + punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
1.210 + punpcklbw xmmE,xmmB ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
1.211 + punpcklbw xmmD,xmmF ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
1.212 +
1.213 + movdqa xmmG,xmmA
1.214 + movdqa xmmH,xmmA
1.215 + punpcklwd xmmA,xmmE ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
1.216 + punpckhwd xmmG,xmmE ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
1.217 +
1.218 + psrldq xmmH,2 ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
1.219 + psrldq xmmE,2 ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
1.220 +
1.221 + movdqa xmmC,xmmD
1.222 + movdqa xmmB,xmmD
1.223 + punpcklwd xmmD,xmmH ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
1.224 + punpckhwd xmmC,xmmH ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
1.225 +
1.226 + psrldq xmmB,2 ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
1.227 +
1.228 + movdqa xmmF,xmmE
1.229 + punpcklwd xmmE,xmmB ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
1.230 + punpckhwd xmmF,xmmB ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
1.231 +
1.232 + pshufd xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
1.233 + movdqa xmmB,xmmE
1.234 + punpckldq xmmA,xmmD ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
1.235 + punpckldq xmmE,xmmH ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
1.236 + punpckhdq xmmD,xmmB ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
1.237 +
1.238 + pshufd xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
1.239 + movdqa xmmB,xmmF
1.240 + punpckldq xmmG,xmmC ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
1.241 + punpckldq xmmF,xmmH ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
1.242 + punpckhdq xmmC,xmmB ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
1.243 +
1.244 + punpcklqdq xmmA,xmmE ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
1.245 + punpcklqdq xmmD,xmmG ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
1.246 + punpcklqdq xmmF,xmmC ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
1.247 +
1.248 + cmp rcx, byte SIZEOF_XMMWORD
1.249 + jb short .column_st32
1.250 +
1.251 + test rdi, SIZEOF_XMMWORD-1
1.252 + jnz short .out1
1.253 + ; --(aligned)-------------------
1.254 + movntdq XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
1.255 + movntdq XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
1.256 + movntdq XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
1.257 + jmp short .out0
1.258 +.out1: ; --(unaligned)-----------------
1.259 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
1.260 + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
1.261 + movdqu XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
1.262 +.out0:
1.263 + add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr
1.264 + sub rcx, byte SIZEOF_XMMWORD
1.265 + jz near .nextrow
1.266 +
1.267 + add rsi, byte SIZEOF_XMMWORD ; inptr0
1.268 + add rbx, byte SIZEOF_XMMWORD ; inptr1
1.269 + add rdx, byte SIZEOF_XMMWORD ; inptr2
1.270 + jmp near .columnloop
1.271 +
1.272 +.column_st32:
1.273 + lea rcx, [rcx+rcx*2] ; imul ecx, RGB_PIXELSIZE
1.274 + cmp rcx, byte 2*SIZEOF_XMMWORD
1.275 + jb short .column_st16
1.276 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
1.277 + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
1.278 + add rdi, byte 2*SIZEOF_XMMWORD ; outptr
1.279 + movdqa xmmA,xmmF
1.280 + sub rcx, byte 2*SIZEOF_XMMWORD
1.281 + jmp short .column_st15
1.282 +.column_st16:
1.283 + cmp rcx, byte SIZEOF_XMMWORD
1.284 + jb short .column_st15
1.285 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
1.286 + add rdi, byte SIZEOF_XMMWORD ; outptr
1.287 + movdqa xmmA,xmmD
1.288 + sub rcx, byte SIZEOF_XMMWORD
1.289 +.column_st15:
1.290 + ; Store the lower 8 bytes of xmmA to the output when it has enough
1.291 + ; space.
1.292 + cmp rcx, byte SIZEOF_MMWORD
1.293 + jb short .column_st7
1.294 + movq XMM_MMWORD [rdi], xmmA
1.295 + add rdi, byte SIZEOF_MMWORD
1.296 + sub rcx, byte SIZEOF_MMWORD
1.297 + psrldq xmmA, SIZEOF_MMWORD
1.298 +.column_st7:
1.299 + ; Store the lower 4 bytes of xmmA to the output when it has enough
1.300 + ; space.
1.301 + cmp rcx, byte SIZEOF_DWORD
1.302 + jb short .column_st3
1.303 + movd XMM_DWORD [rdi], xmmA
1.304 + add rdi, byte SIZEOF_DWORD
1.305 + sub rcx, byte SIZEOF_DWORD
1.306 + psrldq xmmA, SIZEOF_DWORD
1.307 +.column_st3:
1.308 + ; Store the lower 2 bytes of rax to the output when it has enough
1.309 + ; space.
1.310 + movd eax, xmmA
1.311 + cmp rcx, byte SIZEOF_WORD
1.312 + jb short .column_st1
1.313 + mov WORD [rdi], ax
1.314 + add rdi, byte SIZEOF_WORD
1.315 + sub rcx, byte SIZEOF_WORD
1.316 + shr rax, 16
1.317 +.column_st1:
1.318 + ; Store the lower 1 byte of rax to the output when it has enough
1.319 + ; space.
1.320 + test rcx, rcx
1.321 + jz short .nextrow
1.322 + mov BYTE [rdi], al
1.323 +
1.324 +%else ; RGB_PIXELSIZE == 4 ; -----------
1.325 +
1.326 +%ifdef RGBX_FILLER_0XFF
1.327 + pcmpeqb xmm6,xmm6 ; xmm6=XE=X(02468ACE********)
1.328 + pcmpeqb xmm7,xmm7 ; xmm7=XO=X(13579BDF********)
1.329 +%else
1.330 + pxor xmm6,xmm6 ; xmm6=XE=X(02468ACE********)
1.331 + pxor xmm7,xmm7 ; xmm7=XO=X(13579BDF********)
1.332 +%endif
1.333 + ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
1.334 + ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
1.335 + ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
1.336 + ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
1.337 +
1.338 + punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
1.339 + punpcklbw xmmE,xmmG ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
1.340 + punpcklbw xmmB,xmmD ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
1.341 + punpcklbw xmmF,xmmH ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
1.342 +
1.343 + movdqa xmmC,xmmA
1.344 + punpcklwd xmmA,xmmE ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
1.345 + punpckhwd xmmC,xmmE ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
1.346 + movdqa xmmG,xmmB
1.347 + punpcklwd xmmB,xmmF ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
1.348 + punpckhwd xmmG,xmmF ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
1.349 +
1.350 + movdqa xmmD,xmmA
1.351 + punpckldq xmmA,xmmB ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
1.352 + punpckhdq xmmD,xmmB ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
1.353 + movdqa xmmH,xmmC
1.354 + punpckldq xmmC,xmmG ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
1.355 + punpckhdq xmmH,xmmG ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
1.356 +
1.357 + cmp rcx, byte SIZEOF_XMMWORD
1.358 + jb short .column_st32
1.359 +
1.360 + test rdi, SIZEOF_XMMWORD-1
1.361 + jnz short .out1
1.362 + ; --(aligned)-------------------
1.363 + movntdq XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
1.364 + movntdq XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
1.365 + movntdq XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
1.366 + movntdq XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
1.367 + jmp short .out0
1.368 +.out1: ; --(unaligned)-----------------
1.369 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
1.370 + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
1.371 + movdqu XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
1.372 + movdqu XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
1.373 +.out0:
1.374 + add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr
1.375 + sub rcx, byte SIZEOF_XMMWORD
1.376 + jz near .nextrow
1.377 +
1.378 + add rsi, byte SIZEOF_XMMWORD ; inptr0
1.379 + add rbx, byte SIZEOF_XMMWORD ; inptr1
1.380 + add rdx, byte SIZEOF_XMMWORD ; inptr2
1.381 + jmp near .columnloop
1.382 +
1.383 +.column_st32:
1.384 + cmp rcx, byte SIZEOF_XMMWORD/2
1.385 + jb short .column_st16
1.386 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
1.387 + movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
1.388 + add rdi, byte 2*SIZEOF_XMMWORD ; outptr
1.389 + movdqa xmmA,xmmC
1.390 + movdqa xmmD,xmmH
1.391 + sub rcx, byte SIZEOF_XMMWORD/2
1.392 +.column_st16:
1.393 + cmp rcx, byte SIZEOF_XMMWORD/4
1.394 + jb short .column_st15
1.395 + movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
1.396 + add rdi, byte SIZEOF_XMMWORD ; outptr
1.397 + movdqa xmmA,xmmD
1.398 + sub rcx, byte SIZEOF_XMMWORD/4
1.399 +.column_st15:
1.400 + ; Store two pixels (8 bytes) of xmmA to the output when it has enough
1.401 + ; space.
1.402 + cmp rcx, byte SIZEOF_XMMWORD/8
1.403 + jb short .column_st7
1.404 + movq MMWORD [rdi], xmmA
1.405 + add rdi, byte SIZEOF_XMMWORD/8*4
1.406 + sub rcx, byte SIZEOF_XMMWORD/8
1.407 + psrldq xmmA, SIZEOF_XMMWORD/8*4
1.408 +.column_st7:
1.409 + ; Store one pixel (4 bytes) of xmmA to the output when it has enough
1.410 + ; space.
1.411 + test rcx, rcx
1.412 + jz short .nextrow
1.413 + movd XMM_DWORD [rdi], xmmA
1.414 +
1.415 +%endif ; RGB_PIXELSIZE ; ---------------
1.416 +
1.417 +.nextrow:
1.418 + pop rcx
1.419 + pop rsi
1.420 + pop rbx
1.421 + pop rdx
1.422 + pop rdi
1.423 + pop rax
1.424 +
1.425 + add rsi, byte SIZEOF_JSAMPROW
1.426 + add rbx, byte SIZEOF_JSAMPROW
1.427 + add rdx, byte SIZEOF_JSAMPROW
1.428 + add rdi, byte SIZEOF_JSAMPROW ; output_buf
1.429 + dec rax ; num_rows
1.430 + jg near .rowloop
1.431 +
1.432 + sfence ; flush the write buffer
1.433 +
1.434 +.return:
1.435 + pop rbx
1.436 + uncollect_args
1.437 + mov rsp,rbp ; rsp <- aligned rbp
1.438 + pop rsp ; rsp <- original rbp
1.439 + pop rbp
1.440 + ret
1.441 +
1.442 +; For some reason, the OS X linker does not honor the request to align the
1.443 +; segment unless we do this.
1.444 + align 16
