summaryrefslogtreecommitdiff
path: root/media/libjpeg/simd/i386/jquant-mmx.asm
blob: 61305c625de8559b7f57de4a73917afce335ba45 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
;
; jquant.asm - sample data conversion and quantization (MMX)
;
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
; Copyright (C) 2016, D. R. Commander.
;
; Based on the x86 SIMD extension for IJG JPEG library
; Copyright (C) 1999-2006, MIYASAKA Masaru.
; For conditions of distribution and use, see copyright notice in jsimdext.inc
;
; This file should be assembled with NASM (Netwide Assembler),
; can *not* be assembled with Microsoft's MASM or any compatible
; assembler (including Borland's Turbo Assembler).
; NASM is available from http://nasm.sourceforge.net/ or
; http://sourceforge.net/project/showfiles.php?group_id=6208

%include "jsimdext.inc"
%include "jdct.inc"

; --------------------------------------------------------------------------
    SECTION     SEG_TEXT
    BITS        32
;
; Load data into workspace, applying unsigned->signed conversion
;
; GLOBAL(void)
; jsimd_convsamp_mmx(JSAMPARRAY sample_data, JDIMENSION start_col,
;                    DCTELEM *workspace);
;

%define sample_data  ebp + 8            ; JSAMPARRAY sample_data
%define start_col    ebp + 12           ; JDIMENSION start_col
%define workspace    ebp + 16           ; DCTELEM *workspace

    align       32
    GLOBAL_FUNCTION(jsimd_convsamp_mmx)

EXTN(jsimd_convsamp_mmx):
    push        ebp
    mov         ebp, esp
    push        ebx
;   push        ecx                     ; need not be preserved
;   push        edx                     ; need not be preserved
    push        esi
    push        edi

    pxor        mm6, mm6                ; mm6=(all 0's)
    pcmpeqw     mm7, mm7
    psllw       mm7, 7                  ; mm7={0xFF80 0xFF80 0xFF80 0xFF80}

    mov         esi, JSAMPARRAY [sample_data]  ; (JSAMPROW *)
    mov         eax, JDIMENSION [start_col]
    mov         edi, POINTER [workspace]       ; (DCTELEM *)
    mov         ecx, DCTSIZE/4
    alignx      16, 7
.convloop:
    mov         ebx, JSAMPROW [esi+0*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
    mov         edx, JSAMPROW [esi+1*SIZEOF_JSAMPROW]  ; (JSAMPLE *)

    movq        mm0, MMWORD [ebx+eax*SIZEOF_JSAMPLE]   ; mm0=(01234567)
    movq        mm1, MMWORD [edx+eax*SIZEOF_JSAMPLE]   ; mm1=(89ABCDEF)

    mov         ebx, JSAMPROW [esi+2*SIZEOF_JSAMPROW]  ; (JSAMPLE *)
    mov         edx, JSAMPROW [esi+3*SIZEOF_JSAMPROW]  ; (JSAMPLE *)

    movq        mm2, MMWORD [ebx+eax*SIZEOF_JSAMPLE]   ; mm2=(GHIJKLMN)
    movq        mm3, MMWORD [edx+eax*SIZEOF_JSAMPLE]   ; mm3=(OPQRSTUV)

    movq        mm4, mm0
    punpcklbw   mm0, mm6                ; mm0=(0123)
    punpckhbw   mm4, mm6                ; mm4=(4567)
    movq        mm5, mm1
    punpcklbw   mm1, mm6                ; mm1=(89AB)
    punpckhbw   mm5, mm6                ; mm5=(CDEF)

    paddw       mm0, mm7
    paddw       mm4, mm7
    paddw       mm1, mm7
    paddw       mm5, mm7

    movq        MMWORD [MMBLOCK(0,0,edi,SIZEOF_DCTELEM)], mm0
    movq        MMWORD [MMBLOCK(0,1,edi,SIZEOF_DCTELEM)], mm4
    movq        MMWORD [MMBLOCK(1,0,edi,SIZEOF_DCTELEM)], mm1
    movq        MMWORD [MMBLOCK(1,1,edi,SIZEOF_DCTELEM)], mm5

    movq        mm0, mm2
    punpcklbw   mm2, mm6                ; mm2=(GHIJ)
    punpckhbw   mm0, mm6                ; mm0=(KLMN)
    movq        mm4, mm3
    punpcklbw   mm3, mm6                ; mm3=(OPQR)
    punpckhbw   mm4, mm6                ; mm4=(STUV)

    paddw       mm2, mm7
    paddw       mm0, mm7
    paddw       mm3, mm7
    paddw       mm4, mm7

    movq        MMWORD [MMBLOCK(2,0,edi,SIZEOF_DCTELEM)], mm2
    movq        MMWORD [MMBLOCK(2,1,edi,SIZEOF_DCTELEM)], mm0
    movq        MMWORD [MMBLOCK(3,0,edi,SIZEOF_DCTELEM)], mm3
    movq        MMWORD [MMBLOCK(3,1,edi,SIZEOF_DCTELEM)], mm4

    add         esi, byte 4*SIZEOF_JSAMPROW
    add         edi, byte 4*DCTSIZE*SIZEOF_DCTELEM
    dec         ecx
    jnz         short .convloop

    emms                                ; empty MMX state

    pop         edi
    pop         esi
;   pop         edx                     ; need not be preserved
;   pop         ecx                     ; need not be preserved
    pop         ebx
    pop         ebp
    ret

; --------------------------------------------------------------------------
;
; Quantize/descale the coefficients, and store into coef_block
;
; This implementation is based on an algorithm described in
;   "How to optimize for the Pentium family of microprocessors"
;   (http://www.agner.org/assem/).
;
; GLOBAL(void)
; jsimd_quantize_mmx(JCOEFPTR coef_block, DCTELEM *divisors,
;                    DCTELEM *workspace);
;

%define RECIPROCAL(m, n, b) \
  MMBLOCK(DCTSIZE * 0 + (m), (n), (b), SIZEOF_DCTELEM)
%define CORRECTION(m, n, b) \
  MMBLOCK(DCTSIZE * 1 + (m), (n), (b), SIZEOF_DCTELEM)
%define SCALE(m, n, b) \
  MMBLOCK(DCTSIZE * 2 + (m), (n), (b), SIZEOF_DCTELEM)
%define SHIFT(m, n, b) \
  MMBLOCK(DCTSIZE * 3 + (m), (n), (b), SIZEOF_DCTELEM)

%define coef_block  ebp + 8             ; JCOEFPTR coef_block
%define divisors    ebp + 12            ; DCTELEM *divisors
%define workspace   ebp + 16            ; DCTELEM *workspace

    align       32
    GLOBAL_FUNCTION(jsimd_quantize_mmx)

EXTN(jsimd_quantize_mmx):
    push        ebp
    mov         ebp, esp
;   push        ebx                     ; unused
;   push        ecx                     ; unused
;   push        edx                     ; need not be preserved
    push        esi
    push        edi

    mov         esi, POINTER [workspace]
    mov         edx, POINTER [divisors]
    mov         edi, JCOEFPTR [coef_block]
    mov         ah, 2
    alignx      16, 7
.quantloop1:
    mov         al, DCTSIZE2/8/2
    alignx      16, 7
.quantloop2:
    movq        mm2, MMWORD [MMBLOCK(0,0,esi,SIZEOF_DCTELEM)]
    movq        mm3, MMWORD [MMBLOCK(0,1,esi,SIZEOF_DCTELEM)]

    movq        mm0, mm2
    movq        mm1, mm3

    psraw       mm2, (WORD_BIT-1)       ; -1 if value < 0, 0 otherwise
    psraw       mm3, (WORD_BIT-1)

    pxor        mm0, mm2                ; val = -val
    pxor        mm1, mm3
    psubw       mm0, mm2
    psubw       mm1, mm3

    ;
    ; MMX is an annoyingly crappy instruction set. It has two
    ; misfeatures that are causing problems here:
    ;
    ; - All multiplications are signed.
    ;
    ; - The second operand for the shifts is not treated as packed.
    ;
    ;
    ; We work around the first problem by implementing this algorithm:
    ;
    ; unsigned long unsigned_multiply(unsigned short x, unsigned short y)
    ; {
    ;   enum { SHORT_BIT = 16 };
    ;   signed short sx = (signed short)x;
    ;   signed short sy = (signed short)y;
    ;   signed long sz;
    ;
    ;   sz = (long)sx * (long)sy;    /* signed multiply */
    ;
    ;   if (sx < 0) sz += (long)sy << SHORT_BIT;
    ;   if (sy < 0) sz += (long)sx << SHORT_BIT;
    ;
    ;   return (unsigned long)sz;
    ; }
    ;
    ; (note that a negative sx adds _sy_ and vice versa)
    ;
    ; For the second problem, we replace the shift by a multiplication.
    ; Unfortunately that means we have to deal with the signed issue again.
    ;

    paddw       mm0, MMWORD [CORRECTION(0,0,edx)]  ; correction + roundfactor
    paddw       mm1, MMWORD [CORRECTION(0,1,edx)]

    movq        mm4, mm0                ; store current value for later
    movq        mm5, mm1
    pmulhw      mm0, MMWORD [RECIPROCAL(0,0,edx)]  ; reciprocal
    pmulhw      mm1, MMWORD [RECIPROCAL(0,1,edx)]
    paddw       mm0, mm4  ; reciprocal is always negative (MSB=1),
    paddw       mm1, mm5  ; so we always need to add the initial value
                          ; (input value is never negative as we
                          ; inverted it at the start of this routine)

    ; here it gets a bit tricky as both scale
    ; and mm0/mm1 can be negative
    movq        mm6, MMWORD [SCALE(0,0,edx)]  ; scale
    movq        mm7, MMWORD [SCALE(0,1,edx)]
    movq        mm4, mm0
    movq        mm5, mm1
    pmulhw      mm0, mm6
    pmulhw      mm1, mm7

    psraw       mm6, (WORD_BIT-1)       ; determine if scale is negative
    psraw       mm7, (WORD_BIT-1)

    pand        mm6, mm4                ; and add input if it is
    pand        mm7, mm5
    paddw       mm0, mm6
    paddw       mm1, mm7

    psraw       mm4, (WORD_BIT-1)       ; then check if negative input
    psraw       mm5, (WORD_BIT-1)

    pand        mm4, MMWORD [SCALE(0,0,edx)]  ; and add scale if it is
    pand        mm5, MMWORD [SCALE(0,1,edx)]
    paddw       mm0, mm4
    paddw       mm1, mm5

    pxor        mm0, mm2                ; val = -val
    pxor        mm1, mm3
    psubw       mm0, mm2
    psubw       mm1, mm3

    movq        MMWORD [MMBLOCK(0,0,edi,SIZEOF_DCTELEM)], mm0
    movq        MMWORD [MMBLOCK(0,1,edi,SIZEOF_DCTELEM)], mm1

    add         esi, byte 8*SIZEOF_DCTELEM
    add         edx, byte 8*SIZEOF_DCTELEM
    add         edi, byte 8*SIZEOF_JCOEF
    dec         al
    jnz         near .quantloop2
    dec         ah
    jnz         near .quantloop1        ; to avoid branch misprediction

    emms                                ; empty MMX state

    pop         edi
    pop         esi
;   pop         edx                     ; need not be preserved
;   pop         ecx                     ; unused
;   pop         ebx                     ; unused
    pop         ebp
    ret

; For some reason, the OS X linker does not honor the request to align the
; segment unless we do this.
    align       32