1; 2; jquanti.asm - sample data conversion and quantization (64-bit SSE2) 3; 4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; Copyright 2009 D. R. Commander 6; 7; Based on 8; x86 SIMD extension for IJG JPEG library 9; Copyright (C) 1999-2006, MIYASAKA Masaru. 10; For conditions of distribution and use, see copyright notice in jsimdext.inc 11; 12; This file should be assembled with NASM (Netwide Assembler), 13; can *not* be assembled with Microsoft's MASM or any compatible 14; assembler (including Borland's Turbo Assembler). 15; NASM is available from http://nasm.sourceforge.net/ or 16; http://sourceforge.net/project/showfiles.php?group_id=6208 17; 18; [TAB8] 19 20%include "jsimdext.inc" 21%include "jdct.inc" 22 23; -------------------------------------------------------------------------- 24 SECTION SEG_TEXT 25 BITS 64 26; 27; Load data into workspace, applying unsigned->signed conversion 28; 29; GLOBAL(void) 30; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col, 31; DCTELEM * workspace); 32; 33 34; r10 = JSAMPARRAY sample_data 35; r11 = JDIMENSION start_col 36; r12 = DCTELEM * workspace 37 38 align 16 39 global EXTN(jsimd_convsamp_sse2) 40 41EXTN(jsimd_convsamp_sse2): 42 push rbp 43 mov rax,rsp 44 mov rbp,rsp 45 collect_args 46 push rbx 47 48 pxor xmm6,xmm6 ; xmm6=(all 0's) 49 pcmpeqw xmm7,xmm7 50 psllw xmm7,7 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..} 51 52 mov rsi, r10 53 mov eax, r11d 54 mov rdi, r12 55 mov rcx, DCTSIZE/4 56.convloop: 57 mov rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *) 58 mov rdx, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *) 59 60 movq xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE] ; xmm0=(01234567) 61 movq xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE] ; xmm1=(89ABCDEF) 62 63 mov rbx, JSAMPROW [rsi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *) 64 mov rdx, JSAMPROW [rsi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *) 65 66 movq xmm2, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE] ; xmm2=(GHIJKLMN) 67 movq xmm3, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE] ; xmm3=(OPQRSTUV) 68 69 punpcklbw xmm0,xmm6 ; xmm0=(01234567) 70 punpcklbw xmm1,xmm6 ; xmm1=(89ABCDEF) 71 paddw xmm0,xmm7 72 paddw xmm1,xmm7 73 punpcklbw xmm2,xmm6 ; xmm2=(GHIJKLMN) 74 punpcklbw xmm3,xmm6 ; xmm3=(OPQRSTUV) 75 paddw xmm2,xmm7 76 paddw xmm3,xmm7 77 78 movdqa XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0 79 movdqa XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1 80 movdqa XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2 81 movdqa XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3 82 83 add rsi, byte 4*SIZEOF_JSAMPROW 84 add rdi, byte 4*DCTSIZE*SIZEOF_DCTELEM 85 dec rcx 86 jnz short .convloop 87 88 pop rbx 89 uncollect_args 90 pop rbp 91 ret 92 93; -------------------------------------------------------------------------- 94; 95; Quantize/descale the coefficients, and store into coef_block 96; 97; This implementation is based on an algorithm described in 98; "How to optimize for the Pentium family of microprocessors" 99; (http://www.agner.org/assem/). 100; 101; GLOBAL(void) 102; jsimd_quantize_sse2 (JCOEFPTR coef_block, DCTELEM * divisors, 103; DCTELEM * workspace); 104; 105 106%define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM) 107%define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM) 108%define SCALE(m,n,b) XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM) 109 110; r10 = JCOEFPTR coef_block 111; r11 = DCTELEM * divisors 112; r12 = DCTELEM * workspace 113 114 align 16 115 global EXTN(jsimd_quantize_sse2) 116 117EXTN(jsimd_quantize_sse2): 118 push rbp 119 mov rax,rsp 120 mov rbp,rsp 121 collect_args 122 123 mov rsi, r12 124 mov rdx, r11 125 mov rdi, r10 126 mov rax, DCTSIZE2/32 127.quantloop: 128 movdqa xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_DCTELEM)] 129 movdqa xmm5, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_DCTELEM)] 130 movdqa xmm6, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_DCTELEM)] 131 movdqa xmm7, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_DCTELEM)] 132 movdqa xmm0,xmm4 133 movdqa xmm1,xmm5 134 movdqa xmm2,xmm6 135 movdqa xmm3,xmm7 136 psraw xmm4,(WORD_BIT-1) 137 psraw xmm5,(WORD_BIT-1) 138 psraw xmm6,(WORD_BIT-1) 139 psraw xmm7,(WORD_BIT-1) 140 pxor xmm0,xmm4 141 pxor xmm1,xmm5 142 pxor xmm2,xmm6 143 pxor xmm3,xmm7 144 psubw xmm0,xmm4 ; if (xmm0 < 0) xmm0 = -xmm0; 145 psubw xmm1,xmm5 ; if (xmm1 < 0) xmm1 = -xmm1; 146 psubw xmm2,xmm6 ; if (xmm2 < 0) xmm2 = -xmm2; 147 psubw xmm3,xmm7 ; if (xmm3 < 0) xmm3 = -xmm3; 148 149 paddw xmm0, XMMWORD [CORRECTION(0,0,rdx)] ; correction + roundfactor 150 paddw xmm1, XMMWORD [CORRECTION(1,0,rdx)] 151 paddw xmm2, XMMWORD [CORRECTION(2,0,rdx)] 152 paddw xmm3, XMMWORD [CORRECTION(3,0,rdx)] 153 pmulhuw xmm0, XMMWORD [RECIPROCAL(0,0,rdx)] ; reciprocal 154 pmulhuw xmm1, XMMWORD [RECIPROCAL(1,0,rdx)] 155 pmulhuw xmm2, XMMWORD [RECIPROCAL(2,0,rdx)] 156 pmulhuw xmm3, XMMWORD [RECIPROCAL(3,0,rdx)] 157 pmulhuw xmm0, XMMWORD [SCALE(0,0,rdx)] ; scale 158 pmulhuw xmm1, XMMWORD [SCALE(1,0,rdx)] 159 pmulhuw xmm2, XMMWORD [SCALE(2,0,rdx)] 160 pmulhuw xmm3, XMMWORD [SCALE(3,0,rdx)] 161 162 pxor xmm0,xmm4 163 pxor xmm1,xmm5 164 pxor xmm2,xmm6 165 pxor xmm3,xmm7 166 psubw xmm0,xmm4 167 psubw xmm1,xmm5 168 psubw xmm2,xmm6 169 psubw xmm3,xmm7 170 movdqa XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0 171 movdqa XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1 172 movdqa XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2 173 movdqa XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3 174 175 add rsi, byte 32*SIZEOF_DCTELEM 176 add rdx, byte 32*SIZEOF_DCTELEM 177 add rdi, byte 32*SIZEOF_JCOEF 178 dec rax 179 jnz near .quantloop 180 181 uncollect_args 182 pop rbp 183 ret 184 185; For some reason, the OS X linker does not honor the request to align the 186; segment unless we do this. 187 align 16 188