1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jcqnts2i-64.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,187 @@
1.4 +;
1.5 +; jcqnts2i-64.asm - sample data conversion and quantization (64-bit SSE2)
1.6 +;
1.7 +; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
1.8 +; Copyright 2009 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 "jsimdext.inc"
1.24 +%include "jdct.inc"
1.25 +
1.26 +; --------------------------------------------------------------------------
1.27 + SECTION SEG_TEXT
1.28 + BITS 64
1.29 +;
1.30 +; Load data into workspace, applying unsigned->signed conversion
1.31 +;
1.32 +; GLOBAL(void)
1.33 +; jsimd_convsamp_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
1.34 +; DCTELEM * workspace);
1.35 +;
1.36 +
1.37 +; r10 = JSAMPARRAY sample_data
1.38 +; r11 = JDIMENSION start_col
1.39 +; r12 = DCTELEM * workspace
1.40 +
1.41 + align 16
1.42 + global EXTN(jsimd_convsamp_sse2)
1.43 +
1.44 +EXTN(jsimd_convsamp_sse2):
1.45 + push rbp
1.46 + mov rax,rsp
1.47 + mov rbp,rsp
1.48 + collect_args
1.49 + push rbx
1.50 +
1.51 + pxor xmm6,xmm6 ; xmm6=(all 0's)
1.52 + pcmpeqw xmm7,xmm7
1.53 + psllw xmm7,7 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
1.54 +
1.55 + mov rsi, r10
1.56 + mov rax, r11
1.57 + mov rdi, r12
1.58 + mov rcx, DCTSIZE/4
1.59 +.convloop:
1.60 + mov rbx, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
1.61 + mov rdx, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
1.62 +
1.63 + movq xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE] ; xmm0=(01234567)
1.64 + movq xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE] ; xmm1=(89ABCDEF)
1.65 +
1.66 + mov rbx, JSAMPROW [rsi+2*SIZEOF_JSAMPROW] ; (JSAMPLE *)
1.67 + mov rdx, JSAMPROW [rsi+3*SIZEOF_JSAMPROW] ; (JSAMPLE *)
1.68 +
1.69 + movq xmm2, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE] ; xmm2=(GHIJKLMN)
1.70 + movq xmm3, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE] ; xmm3=(OPQRSTUV)
1.71 +
1.72 + punpcklbw xmm0,xmm6 ; xmm0=(01234567)
1.73 + punpcklbw xmm1,xmm6 ; xmm1=(89ABCDEF)
1.74 + paddw xmm0,xmm7
1.75 + paddw xmm1,xmm7
1.76 + punpcklbw xmm2,xmm6 ; xmm2=(GHIJKLMN)
1.77 + punpcklbw xmm3,xmm6 ; xmm3=(OPQRSTUV)
1.78 + paddw xmm2,xmm7
1.79 + paddw xmm3,xmm7
1.80 +
1.81 + movdqa XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0
1.82 + movdqa XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1
1.83 + movdqa XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2
1.84 + movdqa XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3
1.85 +
1.86 + add rsi, byte 4*SIZEOF_JSAMPROW
1.87 + add rdi, byte 4*DCTSIZE*SIZEOF_DCTELEM
1.88 + dec rcx
1.89 + jnz short .convloop
1.90 +
1.91 + pop rbx
1.92 + uncollect_args
1.93 + pop rbp
1.94 + ret
1.95 +
1.96 +; --------------------------------------------------------------------------
1.97 +;
1.98 +; Quantize/descale the coefficients, and store into coef_block
1.99 +;
1.100 +; This implementation is based on an algorithm described in
1.101 +; "How to optimize for the Pentium family of microprocessors"
1.102 +; (http://www.agner.org/assem/).
1.103 +;
1.104 +; GLOBAL(void)
1.105 +; jsimd_quantize_sse2 (JCOEFPTR coef_block, DCTELEM * divisors,
1.106 +; DCTELEM * workspace);
1.107 +;
1.108 +
1.109 +%define RECIPROCAL(m,n,b) XMMBLOCK(DCTSIZE*0+(m),(n),(b),SIZEOF_DCTELEM)
1.110 +%define CORRECTION(m,n,b) XMMBLOCK(DCTSIZE*1+(m),(n),(b),SIZEOF_DCTELEM)
1.111 +%define SCALE(m,n,b) XMMBLOCK(DCTSIZE*2+(m),(n),(b),SIZEOF_DCTELEM)
1.112 +
1.113 +; r10 = JCOEFPTR coef_block
1.114 +; r11 = DCTELEM * divisors
1.115 +; r12 = DCTELEM * workspace
1.116 +
1.117 + align 16
1.118 + global EXTN(jsimd_quantize_sse2)
1.119 +
1.120 +EXTN(jsimd_quantize_sse2):
1.121 + push rbp
1.122 + mov rax,rsp
1.123 + mov rbp,rsp
1.124 + collect_args
1.125 +
1.126 + mov rsi, r12
1.127 + mov rdx, r11
1.128 + mov rdi, r10
1.129 + mov rax, DCTSIZE2/32
1.130 +.quantloop:
1.131 + movdqa xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_DCTELEM)]
1.132 + movdqa xmm5, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_DCTELEM)]
1.133 + movdqa xmm6, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_DCTELEM)]
1.134 + movdqa xmm7, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_DCTELEM)]
1.135 + movdqa xmm0,xmm4
1.136 + movdqa xmm1,xmm5
1.137 + movdqa xmm2,xmm6
1.138 + movdqa xmm3,xmm7
1.139 + psraw xmm4,(WORD_BIT-1)
1.140 + psraw xmm5,(WORD_BIT-1)
1.141 + psraw xmm6,(WORD_BIT-1)
1.142 + psraw xmm7,(WORD_BIT-1)
1.143 + pxor xmm0,xmm4
1.144 + pxor xmm1,xmm5
1.145 + pxor xmm2,xmm6
1.146 + pxor xmm3,xmm7
1.147 + psubw xmm0,xmm4 ; if (xmm0 < 0) xmm0 = -xmm0;
1.148 + psubw xmm1,xmm5 ; if (xmm1 < 0) xmm1 = -xmm1;
1.149 + psubw xmm2,xmm6 ; if (xmm2 < 0) xmm2 = -xmm2;
1.150 + psubw xmm3,xmm7 ; if (xmm3 < 0) xmm3 = -xmm3;
1.151 +
1.152 + paddw xmm0, XMMWORD [CORRECTION(0,0,rdx)] ; correction + roundfactor
1.153 + paddw xmm1, XMMWORD [CORRECTION(1,0,rdx)]
1.154 + paddw xmm2, XMMWORD [CORRECTION(2,0,rdx)]
1.155 + paddw xmm3, XMMWORD [CORRECTION(3,0,rdx)]
1.156 + pmulhuw xmm0, XMMWORD [RECIPROCAL(0,0,rdx)] ; reciprocal
1.157 + pmulhuw xmm1, XMMWORD [RECIPROCAL(1,0,rdx)]
1.158 + pmulhuw xmm2, XMMWORD [RECIPROCAL(2,0,rdx)]
1.159 + pmulhuw xmm3, XMMWORD [RECIPROCAL(3,0,rdx)]
1.160 + pmulhuw xmm0, XMMWORD [SCALE(0,0,rdx)] ; scale
1.161 + pmulhuw xmm1, XMMWORD [SCALE(1,0,rdx)]
1.162 + pmulhuw xmm2, XMMWORD [SCALE(2,0,rdx)]
1.163 + pmulhuw xmm3, XMMWORD [SCALE(3,0,rdx)]
1.164 +
1.165 + pxor xmm0,xmm4
1.166 + pxor xmm1,xmm5
1.167 + pxor xmm2,xmm6
1.168 + pxor xmm3,xmm7
1.169 + psubw xmm0,xmm4
1.170 + psubw xmm1,xmm5
1.171 + psubw xmm2,xmm6
1.172 + psubw xmm3,xmm7
1.173 + movdqa XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_DCTELEM)], xmm0
1.174 + movdqa XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_DCTELEM)], xmm1
1.175 + movdqa XMMWORD [XMMBLOCK(2,0,rdi,SIZEOF_DCTELEM)], xmm2
1.176 + movdqa XMMWORD [XMMBLOCK(3,0,rdi,SIZEOF_DCTELEM)], xmm3
1.177 +
1.178 + add rsi, byte 32*SIZEOF_DCTELEM
1.179 + add rdx, byte 32*SIZEOF_DCTELEM
1.180 + add rdi, byte 32*SIZEOF_JCOEF
1.181 + dec rax
1.182 + jnz near .quantloop
1.183 +
1.184 + uncollect_args
1.185 + pop rbp
1.186 + ret
1.187 +
1.188 +; For some reason, the OS X linker does not honor the request to align the
1.189 +; segment unless we do this.
1.190 + align 16
