The Tor Browser / annotate

media/libjpeg/simd/jcqnts2i-64.asm@b8a032363ba2 (annotated)

media/libjpeg/simd/jcqnts2i-64.asm

Thu, 22 Jan 2015 13:21:57 +0100

author
Michael Schloh von Bennewitz <michael@schloh.com>
date
Thu, 22 Jan 2015 13:21:57 +0100
branch
TOR_BUG_9701
changeset 15
b8a032363ba2
permissions
-rw-r--r--

Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6

michael@0 1 ;
michael@0 2 ; jcqnts2i-64.asm - sample data conversion and quantization (64-bit SSE2)
michael@0 3 ;
michael@0 4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0 5 ; Copyright 2009 D. R. Commander
michael@0 6 ;
michael@0 7 ; Based on
michael@0 8 ; x86 SIMD extension for IJG JPEG library
michael@0 9 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
michael@0 10 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
michael@0 11 ;
michael@0 12 ; This file should be assembled with NASM (Netwide Assembler),
michael@0 13 ; can *not* be assembled with Microsoft's MASM or any compatible
michael@0 14 ; assembler (including Borland's Turbo Assembler).
michael@0 15 ; NASM is available from http://nasm.sourceforge.net/ or
michael@0 16 ; http://sourceforge.net/project/showfiles.php?group_id=6208
michael@0 17 ;
michael@0 18 ; [TAB8]
michael@0 19
michael@0 20 %include "jsimdext.inc"
michael@0 21 %include "jdct.inc"
michael@0 22
michael@0 23 ; --------------------------------------------------------------------------
michael@0 24 SECTION SEG_TEXT
michael@0 25 BITS 64
michael@0 26 ;
michael@0 27 ; Load data into workspace, applying unsigned->signed conversion
michael@0 28 ;
michael@0 29 ; GLOBAL(void)
michael@0 30 ; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
michael@0 31 ; DCTELEM * workspace);
michael@0 32 ;
michael@0 33
michael@0 34 ; r10 = JSAMPARRAY sample_data
michael@0 35 ; r11 = JDIMENSION start_col
michael@0 36 ; r12 = DCTELEM * workspace
michael@0 37
michael@0 38 align 16
michael@0 39 global EXTN(jsimd_convsamp_sse2)
michael@0 40
michael@0 41 EXTN(jsimd_convsamp_sse2):
michael@0 42 push rbp
michael@0 43 mov rax,rsp
michael@0 44 mov rbp,rsp
michael@0 45 collect_args
michael@0 46 push rbx
michael@0 47
michael@0 48 pxor xmm6,xmm6 ; xmm6=(all 0's)
michael@0 49 pcmpeqw xmm7,xmm7
michael@0 50 psllw xmm7,7 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
michael@0 51
michael@0 52 mov rsi, r10
michael@0 53 mov rax, r11
michael@0 54 mov rdi, r12
michael@0 55 mov rcx, DCTSIZE/4
michael@0 56 .convloop:
michael@0 57 mov rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
michael@0 58 mov rdx, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
michael@0 59
michael@0 60 movq xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE] ; xmm0=(01234567)
michael@0 61 movq xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE] ; xmm1=(89ABCDEF)
michael@0 62
michael@0 63 mov rbx, JSAMPROW [rsi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
michael@0 64 mov rdx, JSAMPROW [rsi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
michael@0 65
michael@0 66 movq xmm2, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE] ; xmm2=(GHIJKLMN)
michael@0 67 movq xmm3, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE] ; xmm3=(OPQRSTUV)
michael@0 68
michael@0 69 punpcklbw xmm0,xmm6 ; xmm0=(01234567)
michael@0 70 punpcklbw xmm1,xmm6 ; xmm1=(89ABCDEF)
michael@0 71 paddw xmm0,xmm7
michael@0 72 paddw xmm1,xmm7
michael@0 73 punpcklbw xmm2,xmm6 ; xmm2=(GHIJKLMN)
michael@0 74 punpcklbw xmm3,xmm6 ; xmm3=(OPQRSTUV)
michael@0 75 paddw xmm2,xmm7
michael@0 76 paddw xmm3,xmm7
michael@0 77
michael@0 78 movdqa XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0
michael@0 79 movdqa XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1
michael@0 80 movdqa XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2
michael@0 81 movdqa XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3
michael@0 82
michael@0 83 add rsi, byte 4*SIZEOF_JSAMPROW
michael@0 84 add rdi, byte 4*DCTSIZE*SIZEOF_DCTELEM
michael@0 85 dec rcx
michael@0 86 jnz short .convloop
michael@0 87
michael@0 88 pop rbx
michael@0 89 uncollect_args
michael@0 90 pop rbp
michael@0 91 ret
michael@0 92
michael@0 93 ; --------------------------------------------------------------------------
michael@0 94 ;
michael@0 95 ; Quantize/descale the coefficients, and store into coef_block
michael@0 96 ;
michael@0 97 ; This implementation is based on an algorithm described in
michael@0 98 ; "How to optimize for the Pentium family of microprocessors"
michael@0 99 ; (http://www.agner.org/assem/).
michael@0 100 ;
michael@0 101 ; GLOBAL(void)
michael@0 102 ; jsimd_quantize_sse2 (JCOEFPTR coef_block, DCTELEM * divisors,
michael@0 103 ; DCTELEM * workspace);
michael@0 104 ;
michael@0 105
michael@0 106 %define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM)
michael@0 107 %define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM)
michael@0 108 %define SCALE(m,n,b) XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM)
michael@0 109
michael@0 110 ; r10 = JCOEFPTR coef_block
michael@0 111 ; r11 = DCTELEM * divisors
michael@0 112 ; r12 = DCTELEM * workspace
michael@0 113
michael@0 114 align 16
michael@0 115 global EXTN(jsimd_quantize_sse2)
michael@0 116
michael@0 117 EXTN(jsimd_quantize_sse2):
michael@0 118 push rbp
michael@0 119 mov rax,rsp
michael@0 120 mov rbp,rsp
michael@0 121 collect_args
michael@0 122
michael@0 123 mov rsi, r12
michael@0 124 mov rdx, r11
michael@0 125 mov rdi, r10
michael@0 126 mov rax, DCTSIZE2/32
michael@0 127 .quantloop:
michael@0 128 movdqa xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_DCTELEM)]
michael@0 129 movdqa xmm5, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_DCTELEM)]
michael@0 130 movdqa xmm6, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_DCTELEM)]
michael@0 131 movdqa xmm7, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_DCTELEM)]
michael@0 132 movdqa xmm0,xmm4
michael@0 133 movdqa xmm1,xmm5
michael@0 134 movdqa xmm2,xmm6
michael@0 135 movdqa xmm3,xmm7
michael@0 136 psraw xmm4,(WORD_BIT-1)
michael@0 137 psraw xmm5,(WORD_BIT-1)
michael@0 138 psraw xmm6,(WORD_BIT-1)
michael@0 139 psraw xmm7,(WORD_BIT-1)
michael@0 140 pxor xmm0,xmm4
michael@0 141 pxor xmm1,xmm5
michael@0 142 pxor xmm2,xmm6
michael@0 143 pxor xmm3,xmm7
michael@0 144 psubw xmm0,xmm4 ; if (xmm0 < 0) xmm0 = -xmm0;
michael@0 145 psubw xmm1,xmm5 ; if (xmm1 < 0) xmm1 = -xmm1;
michael@0 146 psubw xmm2,xmm6 ; if (xmm2 < 0) xmm2 = -xmm2;
michael@0 147 psubw xmm3,xmm7 ; if (xmm3 < 0) xmm3 = -xmm3;
michael@0 148
michael@0 149 paddw xmm0, XMMWORD [CORRECTION(0,0,rdx)] ; correction + roundfactor
michael@0 150 paddw xmm1, XMMWORD [CORRECTION(1,0,rdx)]
michael@0 151 paddw xmm2, XMMWORD [CORRECTION(2,0,rdx)]
michael@0 152 paddw xmm3, XMMWORD [CORRECTION(3,0,rdx)]
michael@0 153 pmulhuw xmm0, XMMWORD [RECIPROCAL(0,0,rdx)] ; reciprocal
michael@0 154 pmulhuw xmm1, XMMWORD [RECIPROCAL(1,0,rdx)]
michael@0 155 pmulhuw xmm2, XMMWORD [RECIPROCAL(2,0,rdx)]
michael@0 156 pmulhuw xmm3, XMMWORD [RECIPROCAL(3,0,rdx)]
michael@0 157 pmulhuw xmm0, XMMWORD [SCALE(0,0,rdx)] ; scale
michael@0 158 pmulhuw xmm1, XMMWORD [SCALE(1,0,rdx)]
michael@0 159 pmulhuw xmm2, XMMWORD [SCALE(2,0,rdx)]
michael@0 160 pmulhuw xmm3, XMMWORD [SCALE(3,0,rdx)]
michael@0 161
michael@0 162 pxor xmm0,xmm4
michael@0 163 pxor xmm1,xmm5
michael@0 164 pxor xmm2,xmm6
michael@0 165 pxor xmm3,xmm7
michael@0 166 psubw xmm0,xmm4
michael@0 167 psubw xmm1,xmm5
michael@0 168 psubw xmm2,xmm6
michael@0 169 psubw xmm3,xmm7
michael@0 170 movdqa XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0
michael@0 171 movdqa XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1
michael@0 172 movdqa XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2
michael@0 173 movdqa XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3
michael@0 174
michael@0 175 add rsi, byte 32*SIZEOF_DCTELEM
michael@0 176 add rdx, byte 32*SIZEOF_DCTELEM
michael@0 177 add rdi, byte 32*SIZEOF_JCOEF
michael@0 178 dec rax
michael@0 179 jnz near .quantloop
michael@0 180
michael@0 181 uncollect_args
michael@0 182 pop rbp
michael@0 183 ret
michael@0 184
michael@0 185 ; For some reason, the OS X linker does not honor the request to align the
michael@0 186 ; segment unless we do this.
michael@0 187 align 16

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial