2 ; jquantf.asm - sample data conversion and quantization (64-bit SSE & SSE2)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright (C) 2009, 2016, D. R. Commander.
6 ; Copyright (C) 2018, Matthias Räncker.
8 ; Based on the 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
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
18 %include "jsimdext.inc"
21 ; --------------------------------------------------------------------------
25 ; Load data into workspace, applying unsigned->signed conversion
28 ; jsimd_convsamp_float_sse2(JSAMPARRAY sample_data, JDIMENSION start_col,
29 ; FAST_FLOAT *workspace);
32 ; r10 = JSAMPARRAY sample_data
33 ; r11d = JDIMENSION start_col
34 ; r12 = FAST_FLOAT *workspace
37 GLOBAL_FUNCTION(jsimd_convsamp_float_sse2)
39 EXTN(jsimd_convsamp_float_sse2):
48 packsswb xmm7, xmm7 ; xmm7 = PB_CENTERJSAMPLE (0x808080..)
55 mov rbxp, JSAMPROW [rsi+0*SIZEOF_JSAMPROW] ; (JSAMPLE *)
56 mov rdxp, JSAMPROW [rsi+1*SIZEOF_JSAMPROW] ; (JSAMPLE *)
58 movq xmm0, XMM_MMWORD [rbx+rax*SIZEOF_JSAMPLE]
59 movq xmm1, XMM_MMWORD [rdx+rax*SIZEOF_JSAMPLE]
61 psubb xmm0, xmm7 ; xmm0=(01234567)
62 psubb xmm1, xmm7 ; xmm1=(89ABCDEF)
64 punpcklbw xmm0, xmm0 ; xmm0=(*0*1*2*3*4*5*6*7)
65 punpcklbw xmm1, xmm1 ; xmm1=(*8*9*A*B*C*D*E*F)
67 punpcklwd xmm2, xmm0 ; xmm2=(***0***1***2***3)
68 punpckhwd xmm0, xmm0 ; xmm0=(***4***5***6***7)
69 punpcklwd xmm3, xmm1 ; xmm3=(***8***9***A***B)
70 punpckhwd xmm1, xmm1 ; xmm1=(***C***D***E***F)
72 psrad xmm2, (DWORD_BIT-BYTE_BIT) ; xmm2=(0123)
73 psrad xmm0, (DWORD_BIT-BYTE_BIT) ; xmm0=(4567)
74 cvtdq2ps xmm2, xmm2 ; xmm2=(0123)
75 cvtdq2ps xmm0, xmm0 ; xmm0=(4567)
76 psrad xmm3, (DWORD_BIT-BYTE_BIT) ; xmm3=(89AB)
77 psrad xmm1, (DWORD_BIT-BYTE_BIT) ; xmm1=(CDEF)
78 cvtdq2ps xmm3, xmm3 ; xmm3=(89AB)
79 cvtdq2ps xmm1, xmm1 ; xmm1=(CDEF)
81 movaps XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_FAST_FLOAT)], xmm2
82 movaps XMMWORD [XMMBLOCK(0,1,rdi,SIZEOF_FAST_FLOAT)], xmm0
83 movaps XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_FAST_FLOAT)], xmm3
84 movaps XMMWORD [XMMBLOCK(1,1,rdi,SIZEOF_FAST_FLOAT)], xmm1
86 add rsi, byte 2*SIZEOF_JSAMPROW
87 add rdi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
96 ; --------------------------------------------------------------------------
98 ; Quantize/descale the coefficients, and store into coef_block
101 ; jsimd_quantize_float_sse2(JCOEFPTR coef_block, FAST_FLOAT *divisors,
102 ; FAST_FLOAT *workspace);
105 ; r10 = JCOEFPTR coef_block
106 ; r11 = FAST_FLOAT *divisors
107 ; r12 = FAST_FLOAT *workspace
110 GLOBAL_FUNCTION(jsimd_quantize_float_sse2)
112 EXTN(jsimd_quantize_float_sse2):
123 movaps xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_FAST_FLOAT)]
124 movaps xmm1, XMMWORD [XMMBLOCK(0,1,rsi,SIZEOF_FAST_FLOAT)]
125 mulps xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
126 mulps xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
127 movaps xmm2, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_FAST_FLOAT)]
128 movaps xmm3, XMMWORD [XMMBLOCK(1,1,rsi,SIZEOF_FAST_FLOAT)]
129 mulps xmm2, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
130 mulps xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
140 movdqa XMMWORD [XMMBLOCK(0,0,rdi,SIZEOF_JCOEF)], xmm0
141 movdqa XMMWORD [XMMBLOCK(1,0,rdi,SIZEOF_JCOEF)], xmm2
143 add rsi, byte 16*SIZEOF_FAST_FLOAT
144 add rdx, byte 16*SIZEOF_FAST_FLOAT
145 add rdi, byte 16*SIZEOF_JCOEF
153 ; For some reason, the OS X linker does not honor the request to align the
154 ; segment unless we do this.