1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jcqnts2i.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,200 @@
1.4 +;
1.5 +; jcqnts2i.asm - sample data conversion and quantization (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 +%include "jdct.inc"
1.24 +
1.25 +; --------------------------------------------------------------------------
1.26 + SECTION SEG_TEXT
1.27 + BITS 32
1.28 +;
1.29 +; Load data into workspace, applying unsigned->signed conversion
1.30 +;
1.31 +; GLOBAL(void)
1.32 +; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
1.33 +; DCTELEM * workspace);
1.34 +;
1.35 +
1.36 +%define sample_data ebp+8 ; JSAMPARRAY sample_data
1.37 +%define start_col ebp+12 ; JDIMENSION start_col
1.38 +%define workspace ebp+16 ; DCTELEM * workspace
1.39 +
1.40 + align 16
1.41 + global EXTN(jsimd_convsamp_sse2)
1.42 +
1.43 +EXTN(jsimd_convsamp_sse2):
1.44 + push ebp
1.45 + mov ebp,esp
1.46 + push ebx
1.47 +; push ecx ; need not be preserved
1.48 +; push edx ; need not be preserved
1.49 + push esi
1.50 + push edi
1.51 +
1.52 + pxor xmm6,xmm6 ; xmm6=(all 0's)
1.53 + pcmpeqw xmm7,xmm7
1.54 + psllw xmm7,7 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
1.55 +
1.56 + mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *)
1.57 + mov eax, JDIMENSION [start_col]
1.58 + mov edi, POINTER [workspace] ; (DCTELEM *)
1.59 + mov ecx, DCTSIZE/4
1.60 + alignx 16,7
1.61 +.convloop:
1.62 + mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
1.63 + mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
1.64 +
1.65 + movq xmm0, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE] ; xmm0=(01234567)
1.66 + movq xmm1, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE] ; xmm1=(89ABCDEF)
1.67 +
1.68 + mov ebx, JSAMPROW [esi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
1.69 + mov edx, JSAMPROW [esi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
1.70 +
1.71 + movq xmm2, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE] ; xmm2=(GHIJKLMN)
1.72 + movq xmm3, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE] ; xmm3=(OPQRSTUV)
1.73 +
1.74 + punpcklbw xmm0,xmm6 ; xmm0=(01234567)
1.75 + punpcklbw xmm1,xmm6 ; xmm1=(89ABCDEF)
1.76 + paddw xmm0,xmm7
1.77 + paddw xmm1,xmm7
1.78 + punpcklbw xmm2,xmm6 ; xmm2=(GHIJKLMN)
1.79 + punpcklbw xmm3,xmm6 ; xmm3=(OPQRSTUV)
1.80 + paddw xmm2,xmm7
1.81 + paddw xmm3,xmm7
1.82 +
1.83 + movdqa XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_DCTELEM)], xmm0
1.84 + movdqa XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_DCTELEM)], xmm1
1.85 + movdqa XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_DCTELEM)], xmm2
1.86 + movdqa XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_DCTELEM)], xmm3
1.87 +
1.88 + add esi, byte 4*SIZEOF_JSAMPROW
1.89 + add edi, byte 4*DCTSIZE*SIZEOF_DCTELEM
1.90 + dec ecx
1.91 + jnz short .convloop
1.92 +
1.93 + pop edi
1.94 + pop esi
1.95 +; pop edx ; need not be preserved
1.96 +; pop ecx ; need not be preserved
1.97 + pop ebx
1.98 + pop ebp
1.99 + ret
1.100 +
1.101 +; --------------------------------------------------------------------------
1.102 +;
1.103 +; Quantize/descale the coefficients, and store into coef_block
1.104 +;
1.105 +; This implementation is based on an algorithm described in
1.106 +; "How to optimize for the Pentium family of microprocessors"
1.107 +; (http://www.agner.org/assem/).
1.108 +;
1.109 +; GLOBAL(void)
1.110 +; jsimd_quantize_sse2 (JCOEFPTR coef_block, DCTELEM * divisors,
1.111 +; DCTELEM * workspace);
1.112 +;
1.113 +
1.114 +%define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM)
1.115 +%define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM)
1.116 +%define SCALE(m,n,b) XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM)
1.117 +
1.118 +%define coef_block ebp+8 ; JCOEFPTR coef_block
1.119 +%define divisors ebp+12 ; DCTELEM * divisors
1.120 +%define workspace ebp+16 ; DCTELEM * workspace
1.121 +
1.122 + align 16
1.123 + global EXTN(jsimd_quantize_sse2)
1.124 +
1.125 +EXTN(jsimd_quantize_sse2):
1.126 + push ebp
1.127 + mov ebp,esp
1.128 +; push ebx ; unused
1.129 +; push ecx ; unused
1.130 +; push edx ; need not be preserved
1.131 + push esi
1.132 + push edi
1.133 +
1.134 + mov esi, POINTER [workspace]
1.135 + mov edx, POINTER [divisors]
1.136 + mov edi, JCOEFPTR [coef_block]
1.137 + mov eax, DCTSIZE2/32
1.138 + alignx 16,7
1.139 +.quantloop:
1.140 + movdqa xmm4, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_DCTELEM)]
1.141 + movdqa xmm5, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_DCTELEM)]
1.142 + movdqa xmm6, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_DCTELEM)]
1.143 + movdqa xmm7, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_DCTELEM)]
1.144 + movdqa xmm0,xmm4
1.145 + movdqa xmm1,xmm5
1.146 + movdqa xmm2,xmm6
1.147 + movdqa xmm3,xmm7
1.148 + psraw xmm4,(WORD_BIT-1)
1.149 + psraw xmm5,(WORD_BIT-1)
1.150 + psraw xmm6,(WORD_BIT-1)
1.151 + psraw xmm7,(WORD_BIT-1)
1.152 + pxor xmm0,xmm4
1.153 + pxor xmm1,xmm5
1.154 + pxor xmm2,xmm6
1.155 + pxor xmm3,xmm7
1.156 + psubw xmm0,xmm4 ; if (xmm0 < 0) xmm0 = -xmm0;
1.157 + psubw xmm1,xmm5 ; if (xmm1 < 0) xmm1 = -xmm1;
1.158 + psubw xmm2,xmm6 ; if (xmm2 < 0) xmm2 = -xmm2;
1.159 + psubw xmm3,xmm7 ; if (xmm3 < 0) xmm3 = -xmm3;
1.160 +
1.161 + paddw xmm0, XMMWORD [CORRECTION(0,0,edx)] ; correction + roundfactor
1.162 + paddw xmm1, XMMWORD [CORRECTION(1,0,edx)]
1.163 + paddw xmm2, XMMWORD [CORRECTION(2,0,edx)]
1.164 + paddw xmm3, XMMWORD [CORRECTION(3,0,edx)]
1.165 + pmulhuw xmm0, XMMWORD [RECIPROCAL(0,0,edx)] ; reciprocal
1.166 + pmulhuw xmm1, XMMWORD [RECIPROCAL(1,0,edx)]
1.167 + pmulhuw xmm2, XMMWORD [RECIPROCAL(2,0,edx)]
1.168 + pmulhuw xmm3, XMMWORD [RECIPROCAL(3,0,edx)]
1.169 + pmulhuw xmm0, XMMWORD [SCALE(0,0,edx)] ; scale
1.170 + pmulhuw xmm1, XMMWORD [SCALE(1,0,edx)]
1.171 + pmulhuw xmm2, XMMWORD [SCALE(2,0,edx)]
1.172 + pmulhuw xmm3, XMMWORD [SCALE(3,0,edx)]
1.173 +
1.174 + pxor xmm0,xmm4
1.175 + pxor xmm1,xmm5
1.176 + pxor xmm2,xmm6
1.177 + pxor xmm3,xmm7
1.178 + psubw xmm0,xmm4
1.179 + psubw xmm1,xmm5
1.180 + psubw xmm2,xmm6
1.181 + psubw xmm3,xmm7
1.182 + movdqa XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_DCTELEM)], xmm0
1.183 + movdqa XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_DCTELEM)], xmm1
1.184 + movdqa XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_DCTELEM)], xmm2
1.185 + movdqa XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_DCTELEM)], xmm3
1.186 +
1.187 + add esi, byte 32*SIZEOF_DCTELEM
1.188 + add edx, byte 32*SIZEOF_DCTELEM
1.189 + add edi, byte 32*SIZEOF_JCOEF
1.190 + dec eax
1.191 + jnz near .quantloop
1.192 +
1.193 + pop edi
1.194 + pop esi
1.195 +; pop edx ; need not be preserved
1.196 +; pop ecx ; unused
1.197 +; pop ebx ; unused
1.198 + pop ebp
1.199 + ret
1.200 +
1.201 +; For some reason, the OS X linker does not honor the request to align the
1.202 +; segment unless we do this.
1.203 + align 16
