1;
2; jidctred.asm - reduced-size IDCT (SSE2)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5;
6; Based on
7; x86 SIMD extension for IJG JPEG library
8; Copyright (C) 1999-2006, MIYASAKA Masaru.
9; For conditions of distribution and use, see copyright notice in jsimdext.inc
10;
11; This file should be assembled with NASM (Netwide Assembler),
12; can *not* be assembled with Microsoft's MASM or any compatible
13; assembler (including Borland's Turbo Assembler).
14; NASM is available from http://nasm.sourceforge.net/ or
15; http://sourceforge.net/project/showfiles.php?group_id=6208
16;
17; This file contains inverse-DCT routines that produce reduced-size
18; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
19; The following code is based directly on the IJG's original jidctred.c;
20; see the jidctred.c for more details.
21;
22; [TAB8]
23
24%include "jsimdext.inc"
25%include "jdct.inc"
26
27; --------------------------------------------------------------------------
28
29%define CONST_BITS      13
30%define PASS1_BITS      2
31
32%define DESCALE_P1_4    (CONST_BITS-PASS1_BITS+1)
33%define DESCALE_P2_4    (CONST_BITS+PASS1_BITS+3+1)
34%define DESCALE_P1_2    (CONST_BITS-PASS1_BITS+2)
35%define DESCALE_P2_2    (CONST_BITS+PASS1_BITS+3+2)
36
37%if CONST_BITS == 13
38F_0_211 equ      1730           ; FIX(0.211164243)
39F_0_509 equ      4176           ; FIX(0.509795579)
40F_0_601 equ      4926           ; FIX(0.601344887)
41F_0_720 equ      5906           ; FIX(0.720959822)
42F_0_765 equ      6270           ; FIX(0.765366865)
43F_0_850 equ      6967           ; FIX(0.850430095)
44F_0_899 equ      7373           ; FIX(0.899976223)
45F_1_061 equ      8697           ; FIX(1.061594337)
46F_1_272 equ     10426           ; FIX(1.272758580)
47F_1_451 equ     11893           ; FIX(1.451774981)
48F_1_847 equ     15137           ; FIX(1.847759065)
49F_2_172 equ     17799           ; FIX(2.172734803)
50F_2_562 equ     20995           ; FIX(2.562915447)
51F_3_624 equ     29692           ; FIX(3.624509785)
52%else
53; NASM cannot do compile-time arithmetic on floating-point constants.
54%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
55F_0_211 equ     DESCALE( 226735879,30-CONST_BITS)       ; FIX(0.211164243)
56F_0_509 equ     DESCALE( 547388834,30-CONST_BITS)       ; FIX(0.509795579)
57F_0_601 equ     DESCALE( 645689155,30-CONST_BITS)       ; FIX(0.601344887)
58F_0_720 equ     DESCALE( 774124714,30-CONST_BITS)       ; FIX(0.720959822)
59F_0_765 equ     DESCALE( 821806413,30-CONST_BITS)       ; FIX(0.765366865)
60F_0_850 equ     DESCALE( 913142361,30-CONST_BITS)       ; FIX(0.850430095)
61F_0_899 equ     DESCALE( 966342111,30-CONST_BITS)       ; FIX(0.899976223)
62F_1_061 equ     DESCALE(1139878239,30-CONST_BITS)       ; FIX(1.061594337)
63F_1_272 equ     DESCALE(1366614119,30-CONST_BITS)       ; FIX(1.272758580)
64F_1_451 equ     DESCALE(1558831516,30-CONST_BITS)       ; FIX(1.451774981)
65F_1_847 equ     DESCALE(1984016188,30-CONST_BITS)       ; FIX(1.847759065)
66F_2_172 equ     DESCALE(2332956230,30-CONST_BITS)       ; FIX(2.172734803)
67F_2_562 equ     DESCALE(2751909506,30-CONST_BITS)       ; FIX(2.562915447)
68F_3_624 equ     DESCALE(3891787747,30-CONST_BITS)       ; FIX(3.624509785)
69%endif
70
71; --------------------------------------------------------------------------
72        SECTION SEG_CONST
73
74        alignz  16
75        global  EXTN(jconst_idct_red_sse2)
76
77EXTN(jconst_idct_red_sse2):
78
79PW_F184_MF076   times 4 dw  F_1_847,-F_0_765
80PW_F256_F089    times 4 dw  F_2_562, F_0_899
81PW_F106_MF217   times 4 dw  F_1_061,-F_2_172
82PW_MF060_MF050  times 4 dw -F_0_601,-F_0_509
83PW_F145_MF021   times 4 dw  F_1_451,-F_0_211
84PW_F362_MF127   times 4 dw  F_3_624,-F_1_272
85PW_F085_MF072   times 4 dw  F_0_850,-F_0_720
86PD_DESCALE_P1_4 times 4 dd  1 << (DESCALE_P1_4-1)
87PD_DESCALE_P2_4 times 4 dd  1 << (DESCALE_P2_4-1)
88PD_DESCALE_P1_2 times 4 dd  1 << (DESCALE_P1_2-1)
89PD_DESCALE_P2_2 times 4 dd  1 << (DESCALE_P2_2-1)
90PB_CENTERJSAMP  times 16 db CENTERJSAMPLE
91
92        alignz  16
93
94; --------------------------------------------------------------------------
95        SECTION SEG_TEXT
96        BITS    32
97;
98; Perform dequantization and inverse DCT on one block of coefficients,
99; producing a reduced-size 4x4 output block.
100;
101; GLOBAL(void)
102; jsimd_idct_4x4_sse2 (void * dct_table, JCOEFPTR coef_block,
103;                      JSAMPARRAY output_buf, JDIMENSION output_col)
104;
105
106%define dct_table(b)    (b)+8           ; void * dct_table
107%define coef_block(b)   (b)+12          ; JCOEFPTR coef_block
108%define output_buf(b)   (b)+16          ; JSAMPARRAY output_buf
109%define output_col(b)   (b)+20          ; JDIMENSION output_col
110
111%define original_ebp    ebp+0
112%define wk(i)           ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
113%define WK_NUM          2
114
115        align   16
116        global  EXTN(jsimd_idct_4x4_sse2)
117
118EXTN(jsimd_idct_4x4_sse2):
119        push    ebp
120        mov     eax,esp                         ; eax = original ebp
121        sub     esp, byte 4
122        and     esp, byte (-SIZEOF_XMMWORD)     ; align to 128 bits
123        mov     [esp],eax
124        mov     ebp,esp                         ; ebp = aligned ebp
125        lea     esp, [wk(0)]
126        pushpic ebx
127;       push    ecx             ; unused
128;       push    edx             ; need not be preserved
129        push    esi
130        push    edi
131
132        get_GOT ebx             ; get GOT address
133
134        ; ---- Pass 1: process columns from input.
135
136;       mov     eax, [original_ebp]
137        mov     edx, POINTER [dct_table(eax)]           ; quantptr
138        mov     esi, JCOEFPTR [coef_block(eax)]         ; inptr
139
140%ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2
141        mov     eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
142        or      eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
143        jnz     short .columnDCT
144
145        movdqa  xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
146        movdqa  xmm1, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
147        por     xmm0, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
148        por     xmm1, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
149        por     xmm0, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
150        por     xmm1, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
151        por     xmm0,xmm1
152        packsswb xmm0,xmm0
153        packsswb xmm0,xmm0
154        movd    eax,xmm0
155        test    eax,eax
156        jnz     short .columnDCT
157
158        ; -- AC terms all zero
159
160        movdqa  xmm0, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
161        pmullw  xmm0, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
162
163        psllw   xmm0,PASS1_BITS
164
165        movdqa    xmm3,xmm0     ; xmm0=in0=(00 01 02 03 04 05 06 07)
166        punpcklwd xmm0,xmm0     ; xmm0=(00 00 01 01 02 02 03 03)
167        punpckhwd xmm3,xmm3     ; xmm3=(04 04 05 05 06 06 07 07)
168
169        pshufd  xmm1,xmm0,0x50  ; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01)
170        pshufd  xmm0,xmm0,0xFA  ; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03)
171        pshufd  xmm6,xmm3,0x50  ; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05)
172        pshufd  xmm3,xmm3,0xFA  ; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07)
173
174        jmp     near .column_end
175        alignx  16,7
176%endif
177.columnDCT:
178
179        ; -- Odd part
180
181        movdqa  xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
182        movdqa  xmm1, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
183        pmullw  xmm0, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
184        pmullw  xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
185        movdqa  xmm2, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
186        movdqa  xmm3, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
187        pmullw  xmm2, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
188        pmullw  xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
189
190        movdqa    xmm4,xmm0
191        movdqa    xmm5,xmm0
192        punpcklwd xmm4,xmm1
193        punpckhwd xmm5,xmm1
194        movdqa    xmm0,xmm4
195        movdqa    xmm1,xmm5
196        pmaddwd   xmm4,[GOTOFF(ebx,PW_F256_F089)]       ; xmm4=(tmp2L)
197        pmaddwd   xmm5,[GOTOFF(ebx,PW_F256_F089)]       ; xmm5=(tmp2H)
198        pmaddwd   xmm0,[GOTOFF(ebx,PW_F106_MF217)]      ; xmm0=(tmp0L)
199        pmaddwd   xmm1,[GOTOFF(ebx,PW_F106_MF217)]      ; xmm1=(tmp0H)
200
201        movdqa    xmm6,xmm2
202        movdqa    xmm7,xmm2
203        punpcklwd xmm6,xmm3
204        punpckhwd xmm7,xmm3
205        movdqa    xmm2,xmm6
206        movdqa    xmm3,xmm7
207        pmaddwd   xmm6,[GOTOFF(ebx,PW_MF060_MF050)]     ; xmm6=(tmp2L)
208        pmaddwd   xmm7,[GOTOFF(ebx,PW_MF060_MF050)]     ; xmm7=(tmp2H)
209        pmaddwd   xmm2,[GOTOFF(ebx,PW_F145_MF021)]      ; xmm2=(tmp0L)
210        pmaddwd   xmm3,[GOTOFF(ebx,PW_F145_MF021)]      ; xmm3=(tmp0H)
211
212        paddd   xmm6,xmm4               ; xmm6=tmp2L
213        paddd   xmm7,xmm5               ; xmm7=tmp2H
214        paddd   xmm2,xmm0               ; xmm2=tmp0L
215        paddd   xmm3,xmm1               ; xmm3=tmp0H
216
217        movdqa  XMMWORD [wk(0)], xmm2   ; wk(0)=tmp0L
218        movdqa  XMMWORD [wk(1)], xmm3   ; wk(1)=tmp0H
219
220        ; -- Even part
221
222        movdqa  xmm4, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
223        movdqa  xmm5, XMMWORD [XMMBLOCK(2,0,esi,SIZEOF_JCOEF)]
224        movdqa  xmm0, XMMWORD [XMMBLOCK(6,0,esi,SIZEOF_JCOEF)]
225        pmullw  xmm4, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
226        pmullw  xmm5, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
227        pmullw  xmm0, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
228
229        pxor      xmm1,xmm1
230        pxor      xmm2,xmm2
231        punpcklwd xmm1,xmm4             ; xmm1=tmp0L
232        punpckhwd xmm2,xmm4             ; xmm2=tmp0H
233        psrad     xmm1,(16-CONST_BITS-1) ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1
234        psrad     xmm2,(16-CONST_BITS-1) ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1
235
236        movdqa    xmm3,xmm5             ; xmm5=in2=z2
237        punpcklwd xmm5,xmm0             ; xmm0=in6=z3
238        punpckhwd xmm3,xmm0
239        pmaddwd   xmm5,[GOTOFF(ebx,PW_F184_MF076)]      ; xmm5=tmp2L
240        pmaddwd   xmm3,[GOTOFF(ebx,PW_F184_MF076)]      ; xmm3=tmp2H
241
242        movdqa  xmm4,xmm1
243        movdqa  xmm0,xmm2
244        paddd   xmm1,xmm5               ; xmm1=tmp10L
245        paddd   xmm2,xmm3               ; xmm2=tmp10H
246        psubd   xmm4,xmm5               ; xmm4=tmp12L
247        psubd   xmm0,xmm3               ; xmm0=tmp12H
248
249        ; -- Final output stage
250
251        movdqa  xmm5,xmm1
252        movdqa  xmm3,xmm2
253        paddd   xmm1,xmm6               ; xmm1=data0L
254        paddd   xmm2,xmm7               ; xmm2=data0H
255        psubd   xmm5,xmm6               ; xmm5=data3L
256        psubd   xmm3,xmm7               ; xmm3=data3H
257
258        movdqa  xmm6,[GOTOFF(ebx,PD_DESCALE_P1_4)]      ; xmm6=[PD_DESCALE_P1_4]
259
260        paddd   xmm1,xmm6
261        paddd   xmm2,xmm6
262        psrad   xmm1,DESCALE_P1_4
263        psrad   xmm2,DESCALE_P1_4
264        paddd   xmm5,xmm6
265        paddd   xmm3,xmm6
266        psrad   xmm5,DESCALE_P1_4
267        psrad   xmm3,DESCALE_P1_4
268
269        packssdw  xmm1,xmm2             ; xmm1=data0=(00 01 02 03 04 05 06 07)
270        packssdw  xmm5,xmm3             ; xmm5=data3=(30 31 32 33 34 35 36 37)
271
272        movdqa  xmm7, XMMWORD [wk(0)]   ; xmm7=tmp0L
273        movdqa  xmm6, XMMWORD [wk(1)]   ; xmm6=tmp0H
274
275        movdqa  xmm2,xmm4
276        movdqa  xmm3,xmm0
277        paddd   xmm4,xmm7               ; xmm4=data1L
278        paddd   xmm0,xmm6               ; xmm0=data1H
279        psubd   xmm2,xmm7               ; xmm2=data2L
280        psubd   xmm3,xmm6               ; xmm3=data2H
281
282        movdqa  xmm7,[GOTOFF(ebx,PD_DESCALE_P1_4)]      ; xmm7=[PD_DESCALE_P1_4]
283
284        paddd   xmm4,xmm7
285        paddd   xmm0,xmm7
286        psrad   xmm4,DESCALE_P1_4
287        psrad   xmm0,DESCALE_P1_4
288        paddd   xmm2,xmm7
289        paddd   xmm3,xmm7
290        psrad   xmm2,DESCALE_P1_4
291        psrad   xmm3,DESCALE_P1_4
292
293        packssdw  xmm4,xmm0             ; xmm4=data1=(10 11 12 13 14 15 16 17)
294        packssdw  xmm2,xmm3             ; xmm2=data2=(20 21 22 23 24 25 26 27)
295
296        movdqa    xmm6,xmm1     ; transpose coefficients(phase 1)
297        punpcklwd xmm1,xmm4     ; xmm1=(00 10 01 11 02 12 03 13)
298        punpckhwd xmm6,xmm4     ; xmm6=(04 14 05 15 06 16 07 17)
299        movdqa    xmm7,xmm2     ; transpose coefficients(phase 1)
300        punpcklwd xmm2,xmm5     ; xmm2=(20 30 21 31 22 32 23 33)
301        punpckhwd xmm7,xmm5     ; xmm7=(24 34 25 35 26 36 27 37)
302
303        movdqa    xmm0,xmm1     ; transpose coefficients(phase 2)
304        punpckldq xmm1,xmm2     ; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31)
305        punpckhdq xmm0,xmm2     ; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33)
306        movdqa    xmm3,xmm6     ; transpose coefficients(phase 2)
307        punpckldq xmm6,xmm7     ; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35)
308        punpckhdq xmm3,xmm7     ; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37)
309.column_end:
310
311        ; -- Prefetch the next coefficient block
312
313        prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
314        prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
315        prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
316        prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
317
318        ; ---- Pass 2: process rows, store into output array.
319
320        mov     eax, [original_ebp]
321        mov     edi, JSAMPARRAY [output_buf(eax)]       ; (JSAMPROW *)
322        mov     eax, JDIMENSION [output_col(eax)]
323
324        ; -- Even part
325
326        pxor      xmm4,xmm4
327        punpcklwd xmm4,xmm1             ; xmm4=tmp0
328        psrad     xmm4,(16-CONST_BITS-1) ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1
329
330        ; -- Odd part
331
332        punpckhwd xmm1,xmm0
333        punpckhwd xmm6,xmm3
334        movdqa    xmm5,xmm1
335        movdqa    xmm2,xmm6
336        pmaddwd   xmm1,[GOTOFF(ebx,PW_F256_F089)]       ; xmm1=(tmp2)
337        pmaddwd   xmm6,[GOTOFF(ebx,PW_MF060_MF050)]     ; xmm6=(tmp2)
338        pmaddwd   xmm5,[GOTOFF(ebx,PW_F106_MF217)]      ; xmm5=(tmp0)
339        pmaddwd   xmm2,[GOTOFF(ebx,PW_F145_MF021)]      ; xmm2=(tmp0)
340
341        paddd     xmm6,xmm1             ; xmm6=tmp2
342        paddd     xmm2,xmm5             ; xmm2=tmp0
343
344        ; -- Even part
345
346        punpcklwd xmm0,xmm3
347        pmaddwd   xmm0,[GOTOFF(ebx,PW_F184_MF076)]      ; xmm0=tmp2
348
349        movdqa    xmm7,xmm4
350        paddd     xmm4,xmm0             ; xmm4=tmp10
351        psubd     xmm7,xmm0             ; xmm7=tmp12
352
353        ; -- Final output stage
354
355        movdqa  xmm1,[GOTOFF(ebx,PD_DESCALE_P2_4)]      ; xmm1=[PD_DESCALE_P2_4]
356
357        movdqa  xmm5,xmm4
358        movdqa  xmm3,xmm7
359        paddd   xmm4,xmm6               ; xmm4=data0=(00 10 20 30)
360        paddd   xmm7,xmm2               ; xmm7=data1=(01 11 21 31)
361        psubd   xmm5,xmm6               ; xmm5=data3=(03 13 23 33)
362        psubd   xmm3,xmm2               ; xmm3=data2=(02 12 22 32)
363
364        paddd   xmm4,xmm1
365        paddd   xmm7,xmm1
366        psrad   xmm4,DESCALE_P2_4
367        psrad   xmm7,DESCALE_P2_4
368        paddd   xmm5,xmm1
369        paddd   xmm3,xmm1
370        psrad   xmm5,DESCALE_P2_4
371        psrad   xmm3,DESCALE_P2_4
372
373        packssdw  xmm4,xmm3             ; xmm4=(00 10 20 30 02 12 22 32)
374        packssdw  xmm7,xmm5             ; xmm7=(01 11 21 31 03 13 23 33)
375
376        movdqa    xmm0,xmm4             ; transpose coefficients(phase 1)
377        punpcklwd xmm4,xmm7             ; xmm4=(00 01 10 11 20 21 30 31)
378        punpckhwd xmm0,xmm7             ; xmm0=(02 03 12 13 22 23 32 33)
379
380        movdqa    xmm6,xmm4             ; transpose coefficients(phase 2)
381        punpckldq xmm4,xmm0             ; xmm4=(00 01 02 03 10 11 12 13)
382        punpckhdq xmm6,xmm0             ; xmm6=(20 21 22 23 30 31 32 33)
383
384        packsswb  xmm4,xmm6             ; xmm4=(00 01 02 03 10 11 12 13 20 ..)
385        paddb     xmm4,[GOTOFF(ebx,PB_CENTERJSAMP)]
386
387        pshufd    xmm2,xmm4,0x39        ; xmm2=(10 11 12 13 20 21 22 23 30 ..)
388        pshufd    xmm1,xmm4,0x4E        ; xmm1=(20 21 22 23 30 31 32 33 00 ..)
389        pshufd    xmm3,xmm4,0x93        ; xmm3=(30 31 32 33 00 01 02 03 10 ..)
390
391        mov     edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
392        mov     esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
393        movd    XMM_DWORD [edx+eax*SIZEOF_JSAMPLE], xmm4
394        movd    XMM_DWORD [esi+eax*SIZEOF_JSAMPLE], xmm2
395        mov     edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
396        mov     esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
397        movd    XMM_DWORD [edx+eax*SIZEOF_JSAMPLE], xmm1
398        movd    XMM_DWORD [esi+eax*SIZEOF_JSAMPLE], xmm3
399
400        pop     edi
401        pop     esi
402;       pop     edx             ; need not be preserved
403;       pop     ecx             ; unused
404        poppic  ebx
405        mov     esp,ebp         ; esp <- aligned ebp
406        pop     esp             ; esp <- original ebp
407        pop     ebp
408        ret
409
410
411; --------------------------------------------------------------------------
412;
413; Perform dequantization and inverse DCT on one block of coefficients,
414; producing a reduced-size 2x2 output block.
415;
416; GLOBAL(void)
417; jsimd_idct_2x2_sse2 (void * dct_table, JCOEFPTR coef_block,
418;                      JSAMPARRAY output_buf, JDIMENSION output_col)
419;
420
421%define dct_table(b)    (b)+8           ; void * dct_table
422%define coef_block(b)   (b)+12          ; JCOEFPTR coef_block
423%define output_buf(b)   (b)+16          ; JSAMPARRAY output_buf
424%define output_col(b)   (b)+20          ; JDIMENSION output_col
425
426        align   16
427        global  EXTN(jsimd_idct_2x2_sse2)
428
429EXTN(jsimd_idct_2x2_sse2):
430        push    ebp
431        mov     ebp,esp
432        push    ebx
433;       push    ecx             ; need not be preserved
434;       push    edx             ; need not be preserved
435        push    esi
436        push    edi
437
438        get_GOT ebx             ; get GOT address
439
440        ; ---- Pass 1: process columns from input.
441
442        mov     edx, POINTER [dct_table(ebp)]           ; quantptr
443        mov     esi, JCOEFPTR [coef_block(ebp)]         ; inptr
444
445        ; | input:                  | result:        |
446        ; | 00 01 ** 03 ** 05 ** 07 |                |
447        ; | 10 11 ** 13 ** 15 ** 17 |                |
448        ; | ** ** ** ** ** ** ** ** |                |
449        ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
450        ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
451        ; | 50 51 ** 53 ** 55 ** 57 |                |
452        ; | ** ** ** ** ** ** ** ** |                |
453        ; | 70 71 ** 73 ** 75 ** 77 |                |
454
455        ; -- Odd part
456
457        movdqa  xmm0, XMMWORD [XMMBLOCK(1,0,esi,SIZEOF_JCOEF)]
458        movdqa  xmm1, XMMWORD [XMMBLOCK(3,0,esi,SIZEOF_JCOEF)]
459        pmullw  xmm0, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
460        pmullw  xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
461        movdqa  xmm2, XMMWORD [XMMBLOCK(5,0,esi,SIZEOF_JCOEF)]
462        movdqa  xmm3, XMMWORD [XMMBLOCK(7,0,esi,SIZEOF_JCOEF)]
463        pmullw  xmm2, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
464        pmullw  xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
465
466        ; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37)
467        ; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77)
468
469        pcmpeqd   xmm7,xmm7
470        pslld     xmm7,WORD_BIT         ; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..}
471
472        movdqa    xmm4,xmm0             ; xmm4=(10 11 ** 13 ** 15 ** 17)
473        movdqa    xmm5,xmm2             ; xmm5=(50 51 ** 53 ** 55 ** 57)
474        punpcklwd xmm4,xmm1             ; xmm4=(10 30 11 31 ** ** 13 33)
475        punpcklwd xmm5,xmm3             ; xmm5=(50 70 51 71 ** ** 53 73)
476        pmaddwd   xmm4,[GOTOFF(ebx,PW_F362_MF127)]
477        pmaddwd   xmm5,[GOTOFF(ebx,PW_F085_MF072)]
478
479        psrld   xmm0,WORD_BIT           ; xmm0=(11 -- 13 -- 15 -- 17 --)
480        pand    xmm1,xmm7               ; xmm1=(-- 31 -- 33 -- 35 -- 37)
481        psrld   xmm2,WORD_BIT           ; xmm2=(51 -- 53 -- 55 -- 57 --)
482        pand    xmm3,xmm7               ; xmm3=(-- 71 -- 73 -- 75 -- 77)
483        por     xmm0,xmm1               ; xmm0=(11 31 13 33 15 35 17 37)
484        por     xmm2,xmm3               ; xmm2=(51 71 53 73 55 75 57 77)
485        pmaddwd xmm0,[GOTOFF(ebx,PW_F362_MF127)]
486        pmaddwd xmm2,[GOTOFF(ebx,PW_F085_MF072)]
487
488        paddd   xmm4,xmm5               ; xmm4=tmp0[col0 col1 **** col3]
489        paddd   xmm0,xmm2               ; xmm0=tmp0[col1 col3 col5 col7]
490
491        ; -- Even part
492
493        movdqa  xmm6, XMMWORD [XMMBLOCK(0,0,esi,SIZEOF_JCOEF)]
494        pmullw  xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
495
496        ; xmm6=(00 01 ** 03 ** 05 ** 07)
497
498        movdqa  xmm1,xmm6               ; xmm1=(00 01 ** 03 ** 05 ** 07)
499        pslld   xmm6,WORD_BIT           ; xmm6=(-- 00 -- ** -- ** -- **)
500        pand    xmm1,xmm7               ; xmm1=(-- 01 -- 03 -- 05 -- 07)
501        psrad   xmm6,(WORD_BIT-CONST_BITS-2) ; xmm6=tmp10[col0 **** **** ****]
502        psrad   xmm1,(WORD_BIT-CONST_BITS-2) ; xmm1=tmp10[col1 col3 col5 col7]
503
504        ; -- Final output stage
505
506        movdqa  xmm3,xmm6
507        movdqa  xmm5,xmm1
508        paddd   xmm6,xmm4       ; xmm6=data0[col0 **** **** ****]=(A0 ** ** **)
509        paddd   xmm1,xmm0       ; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7)
510        psubd   xmm3,xmm4       ; xmm3=data1[col0 **** **** ****]=(B0 ** ** **)
511        psubd   xmm5,xmm0       ; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7)
512
513        movdqa  xmm2,[GOTOFF(ebx,PD_DESCALE_P1_2)]      ; xmm2=[PD_DESCALE_P1_2]
514
515        punpckldq  xmm6,xmm3            ; xmm6=(A0 B0 ** **)
516
517        movdqa     xmm7,xmm1
518        punpcklqdq xmm1,xmm5            ; xmm1=(A1 A3 B1 B3)
519        punpckhqdq xmm7,xmm5            ; xmm7=(A5 A7 B5 B7)
520
521        paddd   xmm6,xmm2
522        psrad   xmm6,DESCALE_P1_2
523
524        paddd   xmm1,xmm2
525        paddd   xmm7,xmm2
526        psrad   xmm1,DESCALE_P1_2
527        psrad   xmm7,DESCALE_P1_2
528
529        ; -- Prefetch the next coefficient block
530
531        prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
532        prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
533        prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
534        prefetchnta [esi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
535
536        ; ---- Pass 2: process rows, store into output array.
537
538        mov     edi, JSAMPARRAY [output_buf(ebp)]       ; (JSAMPROW *)
539        mov     eax, JDIMENSION [output_col(ebp)]
540
541        ; | input:| result:|
542        ; | A0 B0 |        |
543        ; | A1 B1 | C0 C1  |
544        ; | A3 B3 | D0 D1  |
545        ; | A5 B5 |        |
546        ; | A7 B7 |        |
547
548        ; -- Odd part
549
550        packssdw  xmm1,xmm1             ; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3)
551        packssdw  xmm7,xmm7             ; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7)
552        pmaddwd   xmm1,[GOTOFF(ebx,PW_F362_MF127)]
553        pmaddwd   xmm7,[GOTOFF(ebx,PW_F085_MF072)]
554
555        paddd     xmm1,xmm7             ; xmm1=tmp0[row0 row1 row0 row1]
556
557        ; -- Even part
558
559        pslld     xmm6,(CONST_BITS+2)   ; xmm6=tmp10[row0 row1 **** ****]
560
561        ; -- Final output stage
562
563        movdqa    xmm4,xmm6
564        paddd     xmm6,xmm1     ; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **)
565        psubd     xmm4,xmm1     ; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **)
566
567        punpckldq xmm6,xmm4     ; xmm6=(C0 D0 C1 D1)
568
569        paddd     xmm6,[GOTOFF(ebx,PD_DESCALE_P2_2)]
570        psrad     xmm6,DESCALE_P2_2
571
572        packssdw  xmm6,xmm6             ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1)
573        packsswb  xmm6,xmm6             ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..)
574        paddb     xmm6,[GOTOFF(ebx,PB_CENTERJSAMP)]
575
576        pextrw  ebx,xmm6,0x00           ; ebx=(C0 D0 -- --)
577        pextrw  ecx,xmm6,0x01           ; ecx=(C1 D1 -- --)
578
579        mov     edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
580        mov     esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
581        mov     WORD [edx+eax*SIZEOF_JSAMPLE], bx
582        mov     WORD [esi+eax*SIZEOF_JSAMPLE], cx
583
584        pop     edi
585        pop     esi
586;       pop     edx             ; need not be preserved
587;       pop     ecx             ; need not be preserved
588        pop     ebx
589        pop     ebp
590        ret
591
592; For some reason, the OS X linker does not honor the request to align the
593; segment unless we do this.
594        align   16
595