1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jcqnts2f-64.asm Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,158 @@
1.4 +;
1.5 +; jcqnts2f-64.asm - sample data conversion and quantization (64-bit SSE & 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_float_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
1.34 +; FAST_FLOAT * workspace);
1.35 +;
1.36 +
1.37 +; r10 = JSAMPARRAY sample_data
1.38 +; r11 = JDIMENSION start_col
1.39 +; r12 = FAST_FLOAT * workspace
1.40 +
1.41 + align 16
1.42 + global EXTN(jsimd_convsamp_float_sse2)
1.43 +
1.44 +EXTN(jsimd_convsamp_float_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 + pcmpeqw xmm7,xmm7
1.52 + psllw xmm7,7
1.53 + packsswb xmm7,xmm7 ; xmm7 = PB_CENTERJSAMPLE (0x808080..)
1.54 +
1.55 + mov rsi, r10
1.56 + mov rax, r11
1.57 + mov rdi, r12
1.58 + mov rcx, DCTSIZE/2
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]
1.64 + movq xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]
1.65 +
1.66 + psubb xmm0,xmm7 ; xmm0=(01234567)
1.67 + psubb xmm1,xmm7 ; xmm1=(89ABCDEF)
1.68 +
1.69 + punpcklbw xmm0,xmm0 ; xmm0=(*0*1*2*3*4*5*6*7)
1.70 + punpcklbw xmm1,xmm1 ; xmm1=(*8*9*A*B*C*D*E*F)
1.71 +
1.72 + punpcklwd xmm2,xmm0 ; xmm2=(***0***1***2***3)
1.73 + punpckhwd xmm0,xmm0 ; xmm0=(***4***5***6***7)
1.74 + punpcklwd xmm3,xmm1 ; xmm3=(***8***9***A***B)
1.75 + punpckhwd xmm1,xmm1 ; xmm1=(***C***D***E***F)
1.76 +
1.77 + psrad xmm2,(DWORD_BIT-BYTE_BIT) ; xmm2=(0123)
1.78 + psrad xmm0,(DWORD_BIT-BYTE_BIT) ; xmm0=(4567)
1.79 + cvtdq2ps xmm2,xmm2 ; xmm2=(0123)
1.80 + cvtdq2ps xmm0,xmm0 ; xmm0=(4567)
1.81 + psrad xmm3,(DWORD_BIT-BYTE_BIT) ; xmm3=(89AB)
1.82 + psrad xmm1,(DWORD_BIT-BYTE_BIT) ; xmm1=(CDEF)
1.83 + cvtdq2ps xmm3,xmm3 ; xmm3=(89AB)
1.84 + cvtdq2ps xmm1,xmm1 ; xmm1=(CDEF)
1.85 +
1.86 + movaps XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm2
1.87 + movaps XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm0
1.88 + movaps XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
1.89 + movaps XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm1
1.90 +
1.91 + add rsi, byte 2*SIZEOF_JSAMPROW
1.92 + add rdi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
1.93 + dec rcx
1.94 + jnz short .convloop
1.95 +
1.96 + pop rbx
1.97 + uncollect_args
1.98 + pop rbp
1.99 + ret
1.100 +
1.101 +
1.102 +; --------------------------------------------------------------------------
1.103 +;
1.104 +; Quantize/descale the coefficients, and store into coef_block
1.105 +;
1.106 +; GLOBAL(void)
1.107 +; jsimd_quantize_float_sse2 (JCOEFPTR coef_block, FAST_FLOAT * divisors,
1.108 +; FAST_FLOAT * workspace);
1.109 +;
1.110 +
1.111 +; r10 = JCOEFPTR coef_block
1.112 +; r11 = FAST_FLOAT * divisors
1.113 +; r12 = FAST_FLOAT * workspace
1.114 +
1.115 + align 16
1.116 + global EXTN(jsimd_quantize_float_sse2)
1.117 +
1.118 +EXTN(jsimd_quantize_float_sse2):
1.119 + push rbp
1.120 + mov rax,rsp
1.121 + mov rbp,rsp
1.122 + collect_args
1.123 +
1.124 + mov rsi, r12
1.125 + mov rdx, r11
1.126 + mov rdi, r10
1.127 + mov rax, DCTSIZE2/16
1.128 +.quantloop:
1.129 + movaps xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_FAST_FLOAT)]
1.130 + movaps xmm1, XMMWORD [XMMBLOCK(0,1,rsi,SIZEOF_FAST_FLOAT)]
1.131 + mulps xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
1.132 + mulps xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
1.133 + movaps xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_FAST_FLOAT)]
1.134 + movaps xmm3, XMMWORD [XMMBLOCK(1,1,rsi,SIZEOF_FAST_FLOAT)]
1.135 + mulps xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
1.136 + mulps xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
1.137 +
1.138 + cvtps2dq xmm0,xmm0
1.139 + cvtps2dq xmm1,xmm1
1.140 + cvtps2dq xmm2,xmm2
1.141 + cvtps2dq xmm3,xmm3
1.142 +
1.143 + packssdw xmm0,xmm1
1.144 + packssdw xmm2,xmm3
1.145 +
1.146 + movdqa XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_JCOEF)], xmm0
1.147 + movdqa XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_JCOEF)], xmm2
1.148 +
1.149 + add rsi, byte 16*SIZEOF_FAST_FLOAT
1.150 + add rdx, byte 16*SIZEOF_FAST_FLOAT
1.151 + add rdi, byte 16*SIZEOF_JCOEF
1.152 + dec rax
1.153 + jnz short .quantloop
1.154 +
1.155 + uncollect_args
1.156 + pop rbp
1.157 + ret
1.158 +
1.159 +; For some reason, the OS X linker does not honor the request to align the
1.160 +; segment unless we do this.
1.161 + align 16
