1;
2; jquantf.asm - sample data conversion and quantization (SSE & SSE2)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5;
6; Based on the x86 SIMD extension for IJG JPEG library
7; Copyright (C) 1999-2006, MIYASAKA Masaru.
8; For conditions of distribution and use, see copyright notice in jsimdext.inc
9;
10; This file should be assembled with NASM (Netwide Assembler),
11; can *not* be assembled with Microsoft's MASM or any compatible
12; assembler (including Borland's Turbo Assembler).
13; NASM is available from http://nasm.sourceforge.net/ or
14; http://sourceforge.net/project/showfiles.php?group_id=6208
15;
16; [TAB8]
17
18%include "jsimdext.inc"
19%include "jdct.inc"
20
21; --------------------------------------------------------------------------
22        SECTION SEG_TEXT
23        BITS    32
24;
25; Load data into workspace, applying unsigned->signed conversion
26;
27; GLOBAL(void)
28; jsimd_convsamp_float_sse2 (JSAMPARRAY sample_data, JDIMENSION start_col,
29;                            FAST_FLOAT *workspace);
30;
31
32%define sample_data     ebp+8           ; JSAMPARRAY sample_data
33%define start_col       ebp+12          ; JDIMENSION start_col
34%define workspace       ebp+16          ; FAST_FLOAT *workspace
35
36        align   16
37        global  EXTN(jsimd_convsamp_float_sse2)
38
39EXTN(jsimd_convsamp_float_sse2):
40        push    ebp
41        mov     ebp,esp
42        push    ebx
43;       push    ecx             ; need not be preserved
44;       push    edx             ; need not be preserved
45        push    esi
46        push    edi
47
48        pcmpeqw  xmm7,xmm7
49        psllw    xmm7,7
50        packsswb xmm7,xmm7              ; xmm7 = PB_CENTERJSAMPLE (0x808080..)
51
52        mov     esi, JSAMPARRAY [sample_data]   ; (JSAMPROW *)
53        mov     eax, JDIMENSION [start_col]
54        mov     edi, POINTER [workspace]        ; (DCTELEM *)
55        mov     ecx, DCTSIZE/2
56        alignx  16,7
57.convloop:
58        mov     ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]   ; (JSAMPLE *)
59        mov     edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW]   ; (JSAMPLE *)
60
61        movq    xmm0, XMM_MMWORD [ebx+eax*SIZEOF_JSAMPLE]
62        movq    xmm1, XMM_MMWORD [edx+eax*SIZEOF_JSAMPLE]
63
64        psubb   xmm0,xmm7                       ; xmm0=(01234567)
65        psubb   xmm1,xmm7                       ; xmm1=(89ABCDEF)
66
67        punpcklbw xmm0,xmm0                     ; xmm0=(*0*1*2*3*4*5*6*7)
68        punpcklbw xmm1,xmm1                     ; xmm1=(*8*9*A*B*C*D*E*F)
69
70        punpcklwd xmm2,xmm0                     ; xmm2=(***0***1***2***3)
71        punpckhwd xmm0,xmm0                     ; xmm0=(***4***5***6***7)
72        punpcklwd xmm3,xmm1                     ; xmm3=(***8***9***A***B)
73        punpckhwd xmm1,xmm1                     ; xmm1=(***C***D***E***F)
74
75        psrad     xmm2,(DWORD_BIT-BYTE_BIT)     ; xmm2=(0123)
76        psrad     xmm0,(DWORD_BIT-BYTE_BIT)     ; xmm0=(4567)
77        cvtdq2ps  xmm2,xmm2                     ; xmm2=(0123)
78        cvtdq2ps  xmm0,xmm0                     ; xmm0=(4567)
79        psrad     xmm3,(DWORD_BIT-BYTE_BIT)     ; xmm3=(89AB)
80        psrad     xmm1,(DWORD_BIT-BYTE_BIT)     ; xmm1=(CDEF)
81        cvtdq2ps  xmm3,xmm3                     ; xmm3=(89AB)
82        cvtdq2ps  xmm1,xmm1                     ; xmm1=(CDEF)
83
84        movaps  XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_FAST_FLOAT)], xmm2
85        movaps  XMMWORD [XMMBLOCK(0,1,edi,SIZEOF_FAST_FLOAT)], xmm0
86        movaps  XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_FAST_FLOAT)], xmm3
87        movaps  XMMWORD [XMMBLOCK(1,1,edi,SIZEOF_FAST_FLOAT)], xmm1
88
89        add     esi, byte 2*SIZEOF_JSAMPROW
90        add     edi, byte 2*DCTSIZE*SIZEOF_FAST_FLOAT
91        dec     ecx
92        jnz     short .convloop
93
94        pop     edi
95        pop     esi
96;       pop     edx             ; need not be preserved
97;       pop     ecx             ; need not be preserved
98        pop     ebx
99        pop     ebp
100        ret
101
102
103; --------------------------------------------------------------------------
104;
105; Quantize/descale the coefficients, and store into coef_block
106;
107; GLOBAL(void)
108; jsimd_quantize_float_sse2 (JCOEFPTR coef_block, FAST_FLOAT *divisors,
109;                            FAST_FLOAT *workspace);
110;
111
112%define coef_block      ebp+8           ; JCOEFPTR coef_block
113%define divisors        ebp+12          ; FAST_FLOAT *divisors
114%define workspace       ebp+16          ; FAST_FLOAT *workspace
115
116        align   16
117        global  EXTN(jsimd_quantize_float_sse2)
118
119EXTN(jsimd_quantize_float_sse2):
120        push    ebp
121        mov     ebp,esp
122;       push    ebx             ; unused
123;       push    ecx             ; unused
124;       push    edx             ; need not be preserved
125        push    esi
126        push    edi
127
128        mov     esi, POINTER [workspace]
129        mov     edx, POINTER [divisors]
130        mov     edi, JCOEFPTR [coef_block]
131        mov     eax, DCTSIZE2/16
132        alignx  16,7
133.quantloop:
134        movaps  xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_FAST_FLOAT)]
135        movaps  xmm1, XMMWORD [XMMBLOCK(0,1,esi,SIZEOF_FAST_FLOAT)]
136        mulps   xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
137        mulps   xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
138        movaps  xmm2, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_FAST_FLOAT)]
139        movaps  xmm3, XMMWORD [XMMBLOCK(1,1,esi,SIZEOF_FAST_FLOAT)]
140        mulps   xmm2, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
141        mulps   xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
142
143        cvtps2dq xmm0,xmm0
144        cvtps2dq xmm1,xmm1
145        cvtps2dq xmm2,xmm2
146        cvtps2dq xmm3,xmm3
147
148        packssdw xmm0,xmm1
149        packssdw xmm2,xmm3
150
151        movdqa  XMMWORD [XMMBLOCK(0,0,edi,SIZEOF_JCOEF)], xmm0
152        movdqa  XMMWORD [XMMBLOCK(1,0,edi,SIZEOF_JCOEF)], xmm2
153
154        add     esi, byte 16*SIZEOF_FAST_FLOAT
155        add     edx, byte 16*SIZEOF_FAST_FLOAT
156        add     edi, byte 16*SIZEOF_JCOEF
157        dec     eax
158        jnz     short .quantloop
159
160        pop     edi
161        pop     esi
162;       pop     edx             ; need not be preserved
163;       pop     ecx             ; unused
164;       pop     ebx             ; unused
165        pop     ebp
166        ret
167
168; For some reason, the OS X linker does not honor the request to align the
169; segment unless we do this.
170        align   16
171