1; 2; jcqnts2i.asm - sample data conversion and quantization (SSE2) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; 6; Based on 7; x86 SIMD extension for IJG JPEG library 8; Copyright (C) 1999-2006, MIYASAKA Masaru. 9; For conditions of distribution and use, see copyright notice in jsimdext.inc 10; 11; This file should be assembled with NASM (Netwide Assembler), 12; can *not* be assembled with Microsoft's MASM or any compatible 13; assembler (including Borland's Turbo Assembler). 14; NASM is available from http://nasm.sourceforge.net/ or 15; http://sourceforge.net/project/showfiles.php?group_id=6208 16; 17; [TAB8] 18 19%include "jsimdext.inc" 20%include "jdct.inc" 21 22; -------------------------------------------------------------------------- 23 SECTION SEG_TEXT 24 BITS 32 25; 26; Load data into workspace, applying unsigned->signed conversion 27; 28; GLOBAL(void) 29; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col, 30; DCTELEM * workspace); 31; 32 33%define sample_data ebp+8 ; JSAMPARRAY sample_data 34%define start_col ebp+12 ; JDIMENSION start_col 35%define workspace ebp+16 ; DCTELEM * workspace 36 37 align 16 38 global EXTN(jsimd_convsamp_sse2) PRIVATE 39 40EXTN(jsimd_convsamp_sse2): 41 push ebp 42 mov ebp,esp 43 push ebx 44; push ecx ; need not be preserved 45; push edx ; need not be preserved 46 push esi 47 push edi 48 49 pxor xmm6,xmm6 ; xmm6=(all 0's) 50 pcmpeqw xmm7,xmm7 51 psllw xmm7,7 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..} 52 53 mov esi, JSAMPARRAY [sample_data] ; (JSAMPROW *) 54 mov eax, JDIMENSION [start_col] 55 mov edi, POINTER [workspace] ; (DCTELEM *) 56 mov ecx, DCTSIZE/4 57 alignx 16,7 58.convloop: 59 mov ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) 60 mov edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) 61 62 movq xmm0, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE] ; xmm0=(01234567) 63 movq xmm1, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE] ; xmm1=(89ABCDEF) 64 65 mov ebx, JSAMPROW [esi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *) 66 mov edx, JSAMPROW [esi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *) 67 68 movq xmm2, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE] ; xmm2=(GHIJKLMN) 69 movq xmm3, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE] ; xmm3=(OPQRSTUV) 70 71 punpcklbw xmm0,xmm6 ; xmm0=(01234567) 72 punpcklbw xmm1,xmm6 ; xmm1=(89ABCDEF) 73 paddw xmm0,xmm7 74 paddw xmm1,xmm7 75 punpcklbw xmm2,xmm6 ; xmm2=(GHIJKLMN) 76 punpcklbw xmm3,xmm6 ; xmm3=(OPQRSTUV) 77 paddw xmm2,xmm7 78 paddw xmm3,xmm7 79 80 movdqa XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_DCTELEM)], xmm0 81 movdqa XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_DCTELEM)], xmm1 82 movdqa XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_DCTELEM)], xmm2 83 movdqa XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_DCTELEM)], xmm3 84 85 add esi, byte 4*SIZEOF_JSAMPROW 86 add edi, byte 4*DCTSIZE*SIZEOF_DCTELEM 87 dec ecx 88 jnz short .convloop 89 90 pop edi 91 pop esi 92; pop edx ; need not be preserved 93; pop ecx ; need not be preserved 94 pop ebx 95 pop ebp 96 ret 97 98; -------------------------------------------------------------------------- 99; 100; Quantize/descale the coefficients, and store into coef_block 101; 102; This implementation is based on an algorithm described in 103; "How to optimize for the Pentium family of microprocessors" 104; (http://www.agner.org/assem/). 105; 106; GLOBAL(void) 107; jsimd_quantize_sse2 (JCOEFPTR coef_block, DCTELEM * divisors, 108; DCTELEM * workspace); 109; 110 111%define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM) 112%define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM) 113%define SCALE(m,n,b) XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM) 114 115%define coef_block ebp+8 ; JCOEFPTR coef_block 116%define divisors ebp+12 ; DCTELEM * divisors 117%define workspace ebp+16 ; DCTELEM * workspace 118 119 align 16 120 global EXTN(jsimd_quantize_sse2) PRIVATE 121 122EXTN(jsimd_quantize_sse2): 123 push ebp 124 mov ebp,esp 125; push ebx ; unused 126; push ecx ; unused 127; push edx ; need not be preserved 128 push esi 129 push edi 130 131 mov esi, POINTER [workspace] 132 mov edx, POINTER [divisors] 133 mov edi, JCOEFPTR [coef_block] 134 mov eax, DCTSIZE2/32 135 alignx 16,7 136.quantloop: 137 movdqa xmm4, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_DCTELEM)] 138 movdqa xmm5, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_DCTELEM)] 139 movdqa xmm6, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_DCTELEM)] 140 movdqa xmm7, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_DCTELEM)] 141 movdqa xmm0,xmm4 142 movdqa xmm1,xmm5 143 movdqa xmm2,xmm6 144 movdqa xmm3,xmm7 145 psraw xmm4,(WORD_BIT-1) 146 psraw xmm5,(WORD_BIT-1) 147 psraw xmm6,(WORD_BIT-1) 148 psraw xmm7,(WORD_BIT-1) 149 pxor xmm0,xmm4 150 pxor xmm1,xmm5 151 pxor xmm2,xmm6 152 pxor xmm3,xmm7 153 psubw xmm0,xmm4 ; if (xmm0 < 0) xmm0 = -xmm0; 154 psubw xmm1,xmm5 ; if (xmm1 < 0) xmm1 = -xmm1; 155 psubw xmm2,xmm6 ; if (xmm2 < 0) xmm2 = -xmm2; 156 psubw xmm3,xmm7 ; if (xmm3 < 0) xmm3 = -xmm3; 157 158 paddw xmm0, XMMWORD [CORRECTION(0,0,edx)] ; correction + roundfactor 159 paddw xmm1, XMMWORD [CORRECTION(1,0,edx)] 160 paddw xmm2, XMMWORD [CORRECTION(2,0,edx)] 161 paddw xmm3, XMMWORD [CORRECTION(3,0,edx)] 162 pmulhuw xmm0, XMMWORD [RECIPROCAL(0,0,edx)] ; reciprocal 163 pmulhuw xmm1, XMMWORD [RECIPROCAL(1,0,edx)] 164 pmulhuw xmm2, XMMWORD [RECIPROCAL(2,0,edx)] 165 pmulhuw xmm3, XMMWORD [RECIPROCAL(3,0,edx)] 166 pmulhuw xmm0, XMMWORD [SCALE(0,0,edx)] ; scale 167 pmulhuw xmm1, XMMWORD [SCALE(1,0,edx)] 168 pmulhuw xmm2, XMMWORD [SCALE(2,0,edx)] 169 pmulhuw xmm3, XMMWORD [SCALE(3,0,edx)] 170 171 pxor xmm0,xmm4 172 pxor xmm1,xmm5 173 pxor xmm2,xmm6 174 pxor xmm3,xmm7 175 psubw xmm0,xmm4 176 psubw xmm1,xmm5 177 psubw xmm2,xmm6 178 psubw xmm3,xmm7 179 movdqa XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_DCTELEM)], xmm0 180 movdqa XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_DCTELEM)], xmm1 181 movdqa XMMWORD [XMMBLOCK(2,0,edi,SIZEOF_DCTELEM)], xmm2 182 movdqa XMMWORD [XMMBLOCK(3,0,edi,SIZEOF_DCTELEM)], xmm3 183 184 add esi, byte 32*SIZEOF_DCTELEM 185 add edx, byte 32*SIZEOF_DCTELEM 186 add edi, byte 32*SIZEOF_JCOEF 187 dec eax 188 jnz near .quantloop 189 190 pop edi 191 pop esi 192; pop edx ; need not be preserved 193; pop ecx ; unused 194; pop ebx ; unused 195 pop ebp 196 ret 197 198; For some reason, the OS X linker does not honor the request to align the 199; segment unless we do this. 200 align 16 201