• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jidctred.asm - reduced-size IDCT (MMX)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 2016, D. R. Commander.
6;
7; Based on the 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%include "jsimdext.inc"
23%include "jdct.inc"
24
25; --------------------------------------------------------------------------
26
27%define CONST_BITS    13
28%define PASS1_BITS    2
29
30%define DESCALE_P1_4  (CONST_BITS - PASS1_BITS + 1)
31%define DESCALE_P2_4  (CONST_BITS + PASS1_BITS + 3 + 1)
32%define DESCALE_P1_2  (CONST_BITS - PASS1_BITS + 2)
33%define DESCALE_P2_2  (CONST_BITS + PASS1_BITS + 3 + 2)
34
35%if CONST_BITS == 13
36F_0_211 equ  1730  ; FIX(0.211164243)
37F_0_509 equ  4176  ; FIX(0.509795579)
38F_0_601 equ  4926  ; FIX(0.601344887)
39F_0_720 equ  5906  ; FIX(0.720959822)
40F_0_765 equ  6270  ; FIX(0.765366865)
41F_0_850 equ  6967  ; FIX(0.850430095)
42F_0_899 equ  7373  ; FIX(0.899976223)
43F_1_061 equ  8697  ; FIX(1.061594337)
44F_1_272 equ 10426  ; FIX(1.272758580)
45F_1_451 equ 11893  ; FIX(1.451774981)
46F_1_847 equ 15137  ; FIX(1.847759065)
47F_2_172 equ 17799  ; FIX(2.172734803)
48F_2_562 equ 20995  ; FIX(2.562915447)
49F_3_624 equ 29692  ; FIX(3.624509785)
50%else
51; NASM cannot do compile-time arithmetic on floating-point constants.
52%define DESCALE(x, n)  (((x) + (1 << ((n) - 1))) >> (n))
53F_0_211 equ DESCALE( 226735879, 30 - CONST_BITS)  ; FIX(0.211164243)
54F_0_509 equ DESCALE( 547388834, 30 - CONST_BITS)  ; FIX(0.509795579)
55F_0_601 equ DESCALE( 645689155, 30 - CONST_BITS)  ; FIX(0.601344887)
56F_0_720 equ DESCALE( 774124714, 30 - CONST_BITS)  ; FIX(0.720959822)
57F_0_765 equ DESCALE( 821806413, 30 - CONST_BITS)  ; FIX(0.765366865)
58F_0_850 equ DESCALE( 913142361, 30 - CONST_BITS)  ; FIX(0.850430095)
59F_0_899 equ DESCALE( 966342111, 30 - CONST_BITS)  ; FIX(0.899976223)
60F_1_061 equ DESCALE(1139878239, 30 - CONST_BITS)  ; FIX(1.061594337)
61F_1_272 equ DESCALE(1366614119, 30 - CONST_BITS)  ; FIX(1.272758580)
62F_1_451 equ DESCALE(1558831516, 30 - CONST_BITS)  ; FIX(1.451774981)
63F_1_847 equ DESCALE(1984016188, 30 - CONST_BITS)  ; FIX(1.847759065)
64F_2_172 equ DESCALE(2332956230, 30 - CONST_BITS)  ; FIX(2.172734803)
65F_2_562 equ DESCALE(2751909506, 30 - CONST_BITS)  ; FIX(2.562915447)
66F_3_624 equ DESCALE(3891787747, 30 - CONST_BITS)  ; FIX(3.624509785)
67%endif
68
69; --------------------------------------------------------------------------
70    SECTION     SEG_CONST
71
72    alignz      32
73    GLOBAL_DATA(jconst_idct_red_mmx)
74
75EXTN(jconst_idct_red_mmx):
76
77PW_F184_MF076   times 2 dw  F_1_847, -F_0_765
78PW_F256_F089    times 2 dw  F_2_562,  F_0_899
79PW_F106_MF217   times 2 dw  F_1_061, -F_2_172
80PW_MF060_MF050  times 2 dw -F_0_601, -F_0_509
81PW_F145_MF021   times 2 dw  F_1_451, -F_0_211
82PW_F362_MF127   times 2 dw  F_3_624, -F_1_272
83PW_F085_MF072   times 2 dw  F_0_850, -F_0_720
84PD_DESCALE_P1_4 times 2 dd  1 << (DESCALE_P1_4 - 1)
85PD_DESCALE_P2_4 times 2 dd  1 << (DESCALE_P2_4 - 1)
86PD_DESCALE_P1_2 times 2 dd  1 << (DESCALE_P1_2 - 1)
87PD_DESCALE_P2_2 times 2 dd  1 << (DESCALE_P2_2 - 1)
88PB_CENTERJSAMP  times 8 db  CENTERJSAMPLE
89
90    alignz      32
91
92; --------------------------------------------------------------------------
93    SECTION     SEG_TEXT
94    BITS        32
95;
96; Perform dequantization and inverse DCT on one block of coefficients,
97; producing a reduced-size 4x4 output block.
98;
99; GLOBAL(void)
100; jsimd_idct_4x4_mmx(void *dct_table, JCOEFPTR coef_block,
101;                    JSAMPARRAY output_buf, JDIMENSION output_col)
102;
103
104%define dct_table(b)   (b) + 8          ; void *dct_table
105%define coef_block(b)  (b) + 12         ; JCOEFPTR coef_block
106%define output_buf(b)  (b) + 16         ; JSAMPARRAY output_buf
107%define output_col(b)  (b) + 20         ; JDIMENSION output_col
108
109%define original_ebp   ebp + 0
110%define wk(i)          ebp - (WK_NUM - (i)) * SIZEOF_MMWORD
111                                        ; mmword wk[WK_NUM]
112%define WK_NUM         2
113%define workspace      wk(0) - DCTSIZE2 * SIZEOF_JCOEF
114                                        ; JCOEF workspace[DCTSIZE2]
115
116    align       32
117    GLOBAL_FUNCTION(jsimd_idct_4x4_mmx)
118
119EXTN(jsimd_idct_4x4_mmx):
120    push        ebp
121    mov         eax, esp                    ; eax = original ebp
122    sub         esp, byte 4
123    and         esp, byte (-SIZEOF_MMWORD)  ; align to 64 bits
124    mov         [esp], eax
125    mov         ebp, esp                    ; ebp = aligned ebp
126    lea         esp, [workspace]
127    pushpic     ebx
128;   push        ecx                     ; need not be preserved
129;   push        edx                     ; need not be preserved
130    push        esi
131    push        edi
132
133    get_GOT     ebx                     ; get GOT address
134
135    ; ---- Pass 1: process columns from input, store into work array.
136
137;   mov         eax, [original_ebp]
138    mov         edx, POINTER [dct_table(eax)]    ; quantptr
139    mov         esi, JCOEFPTR [coef_block(eax)]  ; inptr
140    lea         edi, [workspace]                 ; JCOEF *wsptr
141    mov         ecx, DCTSIZE/4                   ; ctr
142    alignx      16, 7
143.columnloop:
144%ifndef NO_ZERO_COLUMN_TEST_4X4_MMX
145    mov         eax, dword [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
146    or          eax, dword [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
147    jnz         short .columnDCT
148
149    movq        mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
150    movq        mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
151    por         mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
152    por         mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
153    por         mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
154    por         mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
155    por         mm0, mm1
156    packsswb    mm0, mm0
157    movd        eax, mm0
158    test        eax, eax
159    jnz         short .columnDCT
160
161    ; -- AC terms all zero
162
163    movq        mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
164    pmullw      mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
165
166    psllw       mm0, PASS1_BITS
167
168    movq        mm2, mm0                ; mm0=in0=(00 01 02 03)
169    punpcklwd   mm0, mm0                ; mm0=(00 00 01 01)
170    punpckhwd   mm2, mm2                ; mm2=(02 02 03 03)
171
172    movq        mm1, mm0
173    punpckldq   mm0, mm0                ; mm0=(00 00 00 00)
174    punpckhdq   mm1, mm1                ; mm1=(01 01 01 01)
175    movq        mm3, mm2
176    punpckldq   mm2, mm2                ; mm2=(02 02 02 02)
177    punpckhdq   mm3, mm3                ; mm3=(03 03 03 03)
178
179    movq        MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
180    movq        MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
181    movq        MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
182    movq        MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
183    jmp         near .nextcolumn
184    alignx      16, 7
185%endif
186.columnDCT:
187
188    ; -- Odd part
189
190    movq        mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
191    movq        mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
192    pmullw      mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
193    pmullw      mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
194    movq        mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
195    movq        mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
196    pmullw      mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
197    pmullw      mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
198
199    movq        mm4, mm0
200    movq        mm5, mm0
201    punpcklwd   mm4, mm1
202    punpckhwd   mm5, mm1
203    movq        mm0, mm4
204    movq        mm1, mm5
205    pmaddwd     mm4, [GOTOFF(ebx,PW_F256_F089)]   ; mm4=(tmp2L)
206    pmaddwd     mm5, [GOTOFF(ebx,PW_F256_F089)]   ; mm5=(tmp2H)
207    pmaddwd     mm0, [GOTOFF(ebx,PW_F106_MF217)]  ; mm0=(tmp0L)
208    pmaddwd     mm1, [GOTOFF(ebx,PW_F106_MF217)]  ; mm1=(tmp0H)
209
210    movq        mm6, mm2
211    movq        mm7, mm2
212    punpcklwd   mm6, mm3
213    punpckhwd   mm7, mm3
214    movq        mm2, mm6
215    movq        mm3, mm7
216    pmaddwd     mm6, [GOTOFF(ebx,PW_MF060_MF050)]  ; mm6=(tmp2L)
217    pmaddwd     mm7, [GOTOFF(ebx,PW_MF060_MF050)]  ; mm7=(tmp2H)
218    pmaddwd     mm2, [GOTOFF(ebx,PW_F145_MF021)]   ; mm2=(tmp0L)
219    pmaddwd     mm3, [GOTOFF(ebx,PW_F145_MF021)]   ; mm3=(tmp0H)
220
221    paddd       mm6, mm4                ; mm6=tmp2L
222    paddd       mm7, mm5                ; mm7=tmp2H
223    paddd       mm2, mm0                ; mm2=tmp0L
224    paddd       mm3, mm1                ; mm3=tmp0H
225
226    movq        MMWORD [wk(0)], mm2     ; wk(0)=tmp0L
227    movq        MMWORD [wk(1)], mm3     ; wk(1)=tmp0H
228
229    ; -- Even part
230
231    movq        mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
232    movq        mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
233    movq        mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
234    pmullw      mm4, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
235    pmullw      mm5, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
236    pmullw      mm0, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
237
238    pxor        mm1, mm1
239    pxor        mm2, mm2
240    punpcklwd   mm1, mm4                ; mm1=tmp0L
241    punpckhwd   mm2, mm4                ; mm2=tmp0H
242    psrad       mm1, (16-CONST_BITS-1)  ; psrad mm1,16 & pslld mm1,CONST_BITS+1
243    psrad       mm2, (16-CONST_BITS-1)  ; psrad mm2,16 & pslld mm2,CONST_BITS+1
244
245    movq        mm3, mm5                ; mm5=in2=z2
246    punpcklwd   mm5, mm0                ; mm0=in6=z3
247    punpckhwd   mm3, mm0
248    pmaddwd     mm5, [GOTOFF(ebx,PW_F184_MF076)]  ; mm5=tmp2L
249    pmaddwd     mm3, [GOTOFF(ebx,PW_F184_MF076)]  ; mm3=tmp2H
250
251    movq        mm4, mm1
252    movq        mm0, mm2
253    paddd       mm1, mm5                ; mm1=tmp10L
254    paddd       mm2, mm3                ; mm2=tmp10H
255    psubd       mm4, mm5                ; mm4=tmp12L
256    psubd       mm0, mm3                ; mm0=tmp12H
257
258    ; -- Final output stage
259
260    movq        mm5, mm1
261    movq        mm3, mm2
262    paddd       mm1, mm6                ; mm1=data0L
263    paddd       mm2, mm7                ; mm2=data0H
264    psubd       mm5, mm6                ; mm5=data3L
265    psubd       mm3, mm7                ; mm3=data3H
266
267    movq        mm6, [GOTOFF(ebx,PD_DESCALE_P1_4)]  ; mm6=[PD_DESCALE_P1_4]
268
269    paddd       mm1, mm6
270    paddd       mm2, mm6
271    psrad       mm1, DESCALE_P1_4
272    psrad       mm2, DESCALE_P1_4
273    paddd       mm5, mm6
274    paddd       mm3, mm6
275    psrad       mm5, DESCALE_P1_4
276    psrad       mm3, DESCALE_P1_4
277
278    packssdw    mm1, mm2                ; mm1=data0=(00 01 02 03)
279    packssdw    mm5, mm3                ; mm5=data3=(30 31 32 33)
280
281    movq        mm7, MMWORD [wk(0)]     ; mm7=tmp0L
282    movq        mm6, MMWORD [wk(1)]     ; mm6=tmp0H
283
284    movq        mm2, mm4
285    movq        mm3, mm0
286    paddd       mm4, mm7                ; mm4=data1L
287    paddd       mm0, mm6                ; mm0=data1H
288    psubd       mm2, mm7                ; mm2=data2L
289    psubd       mm3, mm6                ; mm3=data2H
290
291    movq        mm7, [GOTOFF(ebx,PD_DESCALE_P1_4)]  ; mm7=[PD_DESCALE_P1_4]
292
293    paddd       mm4, mm7
294    paddd       mm0, mm7
295    psrad       mm4, DESCALE_P1_4
296    psrad       mm0, DESCALE_P1_4
297    paddd       mm2, mm7
298    paddd       mm3, mm7
299    psrad       mm2, DESCALE_P1_4
300    psrad       mm3, DESCALE_P1_4
301
302    packssdw    mm4, mm0                ; mm4=data1=(10 11 12 13)
303    packssdw    mm2, mm3                ; mm2=data2=(20 21 22 23)
304
305    movq        mm6, mm1                ; transpose coefficients(phase 1)
306    punpcklwd   mm1, mm4                ; mm1=(00 10 01 11)
307    punpckhwd   mm6, mm4                ; mm6=(02 12 03 13)
308    movq        mm7, mm2                ; transpose coefficients(phase 1)
309    punpcklwd   mm2, mm5                ; mm2=(20 30 21 31)
310    punpckhwd   mm7, mm5                ; mm7=(22 32 23 33)
311
312    movq        mm0, mm1                ; transpose coefficients(phase 2)
313    punpckldq   mm1, mm2                ; mm1=(00 10 20 30)
314    punpckhdq   mm0, mm2                ; mm0=(01 11 21 31)
315    movq        mm3, mm6                ; transpose coefficients(phase 2)
316    punpckldq   mm6, mm7                ; mm6=(02 12 22 32)
317    punpckhdq   mm3, mm7                ; mm3=(03 13 23 33)
318
319    movq        MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm1
320    movq        MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm0
321    movq        MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm6
322    movq        MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
323
324.nextcolumn:
325    add         esi, byte 4*SIZEOF_JCOEF            ; coef_block
326    add         edx, byte 4*SIZEOF_ISLOW_MULT_TYPE  ; quantptr
327    add         edi, byte 4*DCTSIZE*SIZEOF_JCOEF    ; wsptr
328    dec         ecx                                 ; ctr
329    jnz         near .columnloop
330
331    ; ---- Pass 2: process rows from work array, store into output array.
332
333    mov         eax, [original_ebp]
334    lea         esi, [workspace]                   ; JCOEF *wsptr
335    mov         edi, JSAMPARRAY [output_buf(eax)]  ; (JSAMPROW *)
336    mov         eax, JDIMENSION [output_col(eax)]
337
338    ; -- Odd part
339
340    movq        mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
341    movq        mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
342    movq        mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
343    movq        mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
344
345    movq        mm4, mm0
346    movq        mm5, mm0
347    punpcklwd   mm4, mm1
348    punpckhwd   mm5, mm1
349    movq        mm0, mm4
350    movq        mm1, mm5
351    pmaddwd     mm4, [GOTOFF(ebx,PW_F256_F089)]   ; mm4=(tmp2L)
352    pmaddwd     mm5, [GOTOFF(ebx,PW_F256_F089)]   ; mm5=(tmp2H)
353    pmaddwd     mm0, [GOTOFF(ebx,PW_F106_MF217)]  ; mm0=(tmp0L)
354    pmaddwd     mm1, [GOTOFF(ebx,PW_F106_MF217)]  ; mm1=(tmp0H)
355
356    movq        mm6, mm2
357    movq        mm7, mm2
358    punpcklwd   mm6, mm3
359    punpckhwd   mm7, mm3
360    movq        mm2, mm6
361    movq        mm3, mm7
362    pmaddwd     mm6, [GOTOFF(ebx,PW_MF060_MF050)]  ; mm6=(tmp2L)
363    pmaddwd     mm7, [GOTOFF(ebx,PW_MF060_MF050)]  ; mm7=(tmp2H)
364    pmaddwd     mm2, [GOTOFF(ebx,PW_F145_MF021)]   ; mm2=(tmp0L)
365    pmaddwd     mm3, [GOTOFF(ebx,PW_F145_MF021)]   ; mm3=(tmp0H)
366
367    paddd       mm6, mm4                ; mm6=tmp2L
368    paddd       mm7, mm5                ; mm7=tmp2H
369    paddd       mm2, mm0                ; mm2=tmp0L
370    paddd       mm3, mm1                ; mm3=tmp0H
371
372    movq        MMWORD [wk(0)], mm2     ; wk(0)=tmp0L
373    movq        MMWORD [wk(1)], mm3     ; wk(1)=tmp0H
374
375    ; -- Even part
376
377    movq        mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
378    movq        mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
379    movq        mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
380
381    pxor        mm1, mm1
382    pxor        mm2, mm2
383    punpcklwd   mm1, mm4                ; mm1=tmp0L
384    punpckhwd   mm2, mm4                ; mm2=tmp0H
385    psrad       mm1, (16-CONST_BITS-1)  ; psrad mm1,16 & pslld mm1,CONST_BITS+1
386    psrad       mm2, (16-CONST_BITS-1)  ; psrad mm2,16 & pslld mm2,CONST_BITS+1
387
388    movq        mm3, mm5                ; mm5=in2=z2
389    punpcklwd   mm5, mm0                ; mm0=in6=z3
390    punpckhwd   mm3, mm0
391    pmaddwd     mm5, [GOTOFF(ebx,PW_F184_MF076)]  ; mm5=tmp2L
392    pmaddwd     mm3, [GOTOFF(ebx,PW_F184_MF076)]  ; mm3=tmp2H
393
394    movq        mm4, mm1
395    movq        mm0, mm2
396    paddd       mm1, mm5                ; mm1=tmp10L
397    paddd       mm2, mm3                ; mm2=tmp10H
398    psubd       mm4, mm5                ; mm4=tmp12L
399    psubd       mm0, mm3                ; mm0=tmp12H
400
401    ; -- Final output stage
402
403    movq        mm5, mm1
404    movq        mm3, mm2
405    paddd       mm1, mm6                ; mm1=data0L
406    paddd       mm2, mm7                ; mm2=data0H
407    psubd       mm5, mm6                ; mm5=data3L
408    psubd       mm3, mm7                ; mm3=data3H
409
410    movq        mm6, [GOTOFF(ebx,PD_DESCALE_P2_4)]  ; mm6=[PD_DESCALE_P2_4]
411
412    paddd       mm1, mm6
413    paddd       mm2, mm6
414    psrad       mm1, DESCALE_P2_4
415    psrad       mm2, DESCALE_P2_4
416    paddd       mm5, mm6
417    paddd       mm3, mm6
418    psrad       mm5, DESCALE_P2_4
419    psrad       mm3, DESCALE_P2_4
420
421    packssdw    mm1, mm2                ; mm1=data0=(00 10 20 30)
422    packssdw    mm5, mm3                ; mm5=data3=(03 13 23 33)
423
424    movq        mm7, MMWORD [wk(0)]     ; mm7=tmp0L
425    movq        mm6, MMWORD [wk(1)]     ; mm6=tmp0H
426
427    movq        mm2, mm4
428    movq        mm3, mm0
429    paddd       mm4, mm7                ; mm4=data1L
430    paddd       mm0, mm6                ; mm0=data1H
431    psubd       mm2, mm7                ; mm2=data2L
432    psubd       mm3, mm6                ; mm3=data2H
433
434    movq        mm7, [GOTOFF(ebx,PD_DESCALE_P2_4)]  ; mm7=[PD_DESCALE_P2_4]
435
436    paddd       mm4, mm7
437    paddd       mm0, mm7
438    psrad       mm4, DESCALE_P2_4
439    psrad       mm0, DESCALE_P2_4
440    paddd       mm2, mm7
441    paddd       mm3, mm7
442    psrad       mm2, DESCALE_P2_4
443    psrad       mm3, DESCALE_P2_4
444
445    packssdw    mm4, mm0                ; mm4=data1=(01 11 21 31)
446    packssdw    mm2, mm3                ; mm2=data2=(02 12 22 32)
447
448    movq        mm6, [GOTOFF(ebx,PB_CENTERJSAMP)]  ; mm6=[PB_CENTERJSAMP]
449
450    packsswb    mm1, mm2                ; mm1=(00 10 20 30 02 12 22 32)
451    packsswb    mm4, mm5                ; mm4=(01 11 21 31 03 13 23 33)
452    paddb       mm1, mm6
453    paddb       mm4, mm6
454
455    movq        mm7, mm1                ; transpose coefficients(phase 1)
456    punpcklbw   mm1, mm4                ; mm1=(00 01 10 11 20 21 30 31)
457    punpckhbw   mm7, mm4                ; mm7=(02 03 12 13 22 23 32 33)
458
459    movq        mm0, mm1                ; transpose coefficients(phase 2)
460    punpcklwd   mm1, mm7                ; mm1=(00 01 02 03 10 11 12 13)
461    punpckhwd   mm0, mm7                ; mm0=(20 21 22 23 30 31 32 33)
462
463    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
464    mov         esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
465    movd        dword [edx+eax*SIZEOF_JSAMPLE], mm1
466    movd        dword [esi+eax*SIZEOF_JSAMPLE], mm0
467
468    psrlq       mm1, 4*BYTE_BIT
469    psrlq       mm0, 4*BYTE_BIT
470
471    mov         edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
472    mov         esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
473    movd        dword [edx+eax*SIZEOF_JSAMPLE], mm1
474    movd        dword [esi+eax*SIZEOF_JSAMPLE], mm0
475
476    emms                                ; empty MMX state
477
478    pop         edi
479    pop         esi
480;   pop         edx                     ; need not be preserved
481;   pop         ecx                     ; need not be preserved
482    poppic      ebx
483    mov         esp, ebp                ; esp <- aligned ebp
484    pop         esp                     ; esp <- original ebp
485    pop         ebp
486    ret
487
488; --------------------------------------------------------------------------
489;
490; Perform dequantization and inverse DCT on one block of coefficients,
491; producing a reduced-size 2x2 output block.
492;
493; GLOBAL(void)
494; jsimd_idct_2x2_mmx(void *dct_table, JCOEFPTR coef_block,
495;                    JSAMPARRAY output_buf, JDIMENSION output_col)
496;
497
498%define dct_table(b)   (b) + 8          ; void *dct_table
499%define coef_block(b)  (b) + 12         ; JCOEFPTR coef_block
500%define output_buf(b)  (b) + 16         ; JSAMPARRAY output_buf
501%define output_col(b)  (b) + 20         ; JDIMENSION output_col
502
503    align       32
504    GLOBAL_FUNCTION(jsimd_idct_2x2_mmx)
505
506EXTN(jsimd_idct_2x2_mmx):
507    push        ebp
508    mov         ebp, esp
509    push        ebx
510;   push        ecx                     ; need not be preserved
511;   push        edx                     ; need not be preserved
512    push        esi
513    push        edi
514
515    get_GOT     ebx                     ; get GOT address
516
517    ; ---- Pass 1: process columns from input.
518
519    mov         edx, POINTER [dct_table(ebp)]    ; quantptr
520    mov         esi, JCOEFPTR [coef_block(ebp)]  ; inptr
521
522    ; | input:                  | result:        |
523    ; | 00 01 ** 03 ** 05 ** 07 |                |
524    ; | 10 11 ** 13 ** 15 ** 17 |                |
525    ; | ** ** ** ** ** ** ** ** |                |
526    ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
527    ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
528    ; | 50 51 ** 53 ** 55 ** 57 |                |
529    ; | ** ** ** ** ** ** ** ** |                |
530    ; | 70 71 ** 73 ** 75 ** 77 |                |
531
532    ; -- Odd part
533
534    movq        mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
535    movq        mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
536    pmullw      mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
537    pmullw      mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
538    movq        mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
539    movq        mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
540    pmullw      mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
541    pmullw      mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
542
543    ; mm0=(10 11 ** 13), mm1=(30 31 ** 33)
544    ; mm2=(50 51 ** 53), mm3=(70 71 ** 73)
545
546    pcmpeqd     mm7, mm7
547    pslld       mm7, WORD_BIT           ; mm7={0x0000 0xFFFF 0x0000 0xFFFF}
548
549    movq        mm4, mm0                ; mm4=(10 11 ** 13)
550    movq        mm5, mm2                ; mm5=(50 51 ** 53)
551    punpcklwd   mm4, mm1                ; mm4=(10 30 11 31)
552    punpcklwd   mm5, mm3                ; mm5=(50 70 51 71)
553    pmaddwd     mm4, [GOTOFF(ebx,PW_F362_MF127)]
554    pmaddwd     mm5, [GOTOFF(ebx,PW_F085_MF072)]
555
556    psrld       mm0, WORD_BIT           ; mm0=(11 -- 13 --)
557    pand        mm1, mm7                ; mm1=(-- 31 -- 33)
558    psrld       mm2, WORD_BIT           ; mm2=(51 -- 53 --)
559    pand        mm3, mm7                ; mm3=(-- 71 -- 73)
560    por         mm0, mm1                ; mm0=(11 31 13 33)
561    por         mm2, mm3                ; mm2=(51 71 53 73)
562    pmaddwd     mm0, [GOTOFF(ebx,PW_F362_MF127)]
563    pmaddwd     mm2, [GOTOFF(ebx,PW_F085_MF072)]
564
565    paddd       mm4, mm5                ; mm4=tmp0[col0 col1]
566
567    movq        mm6, MMWORD [MMBLOCK(1,1,esi,SIZEOF_JCOEF)]
568    movq        mm1, MMWORD [MMBLOCK(3,1,esi,SIZEOF_JCOEF)]
569    pmullw      mm6, MMWORD [MMBLOCK(1,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
570    pmullw      mm1, MMWORD [MMBLOCK(3,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
571    movq        mm3, MMWORD [MMBLOCK(5,1,esi,SIZEOF_JCOEF)]
572    movq        mm5, MMWORD [MMBLOCK(7,1,esi,SIZEOF_JCOEF)]
573    pmullw      mm3, MMWORD [MMBLOCK(5,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
574    pmullw      mm5, MMWORD [MMBLOCK(7,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
575
576    ; mm6=(** 15 ** 17), mm1=(** 35 ** 37)
577    ; mm3=(** 55 ** 57), mm5=(** 75 ** 77)
578
579    psrld       mm6, WORD_BIT           ; mm6=(15 -- 17 --)
580    pand        mm1, mm7                ; mm1=(-- 35 -- 37)
581    psrld       mm3, WORD_BIT           ; mm3=(55 -- 57 --)
582    pand        mm5, mm7                ; mm5=(-- 75 -- 77)
583    por         mm6, mm1                ; mm6=(15 35 17 37)
584    por         mm3, mm5                ; mm3=(55 75 57 77)
585    pmaddwd     mm6, [GOTOFF(ebx,PW_F362_MF127)]
586    pmaddwd     mm3, [GOTOFF(ebx,PW_F085_MF072)]
587
588    paddd       mm0, mm2                ; mm0=tmp0[col1 col3]
589    paddd       mm6, mm3                ; mm6=tmp0[col5 col7]
590
591    ; -- Even part
592
593    movq        mm1, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
594    movq        mm5, MMWORD [MMBLOCK(0,1,esi,SIZEOF_JCOEF)]
595    pmullw      mm1, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
596    pmullw      mm5, MMWORD [MMBLOCK(0,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
597
598    ; mm1=(00 01 ** 03), mm5=(** 05 ** 07)
599
600    movq        mm2, mm1                      ; mm2=(00 01 ** 03)
601    pslld       mm1, WORD_BIT                 ; mm1=(-- 00 -- **)
602    psrad       mm1, (WORD_BIT-CONST_BITS-2)  ; mm1=tmp10[col0 ****]
603
604    pand        mm2, mm7                      ; mm2=(-- 01 -- 03)
605    pand        mm5, mm7                      ; mm5=(-- 05 -- 07)
606    psrad       mm2, (WORD_BIT-CONST_BITS-2)  ; mm2=tmp10[col1 col3]
607    psrad       mm5, (WORD_BIT-CONST_BITS-2)  ; mm5=tmp10[col5 col7]
608
609    ; -- Final output stage
610
611    movq        mm3, mm1
612    paddd       mm1, mm4                ; mm1=data0[col0 ****]=(A0 **)
613    psubd       mm3, mm4                ; mm3=data1[col0 ****]=(B0 **)
614    punpckldq   mm1, mm3                ; mm1=(A0 B0)
615
616    movq        mm7, [GOTOFF(ebx,PD_DESCALE_P1_2)]  ; mm7=[PD_DESCALE_P1_2]
617
618    movq        mm4, mm2
619    movq        mm3, mm5
620    paddd       mm2, mm0                ; mm2=data0[col1 col3]=(A1 A3)
621    paddd       mm5, mm6                ; mm5=data0[col5 col7]=(A5 A7)
622    psubd       mm4, mm0                ; mm4=data1[col1 col3]=(B1 B3)
623    psubd       mm3, mm6                ; mm3=data1[col5 col7]=(B5 B7)
624
625    paddd       mm1, mm7
626    psrad       mm1, DESCALE_P1_2
627
628    paddd       mm2, mm7
629    paddd       mm5, mm7
630    psrad       mm2, DESCALE_P1_2
631    psrad       mm5, DESCALE_P1_2
632    paddd       mm4, mm7
633    paddd       mm3, mm7
634    psrad       mm4, DESCALE_P1_2
635    psrad       mm3, DESCALE_P1_2
636
637    ; ---- Pass 2: process rows, store into output array.
638
639    mov         edi, JSAMPARRAY [output_buf(ebp)]  ; (JSAMPROW *)
640    mov         eax, JDIMENSION [output_col(ebp)]
641
642    ; | input:| result:|
643    ; | A0 B0 |        |
644    ; | A1 B1 | C0 C1  |
645    ; | A3 B3 | D0 D1  |
646    ; | A5 B5 |        |
647    ; | A7 B7 |        |
648
649    ; -- Odd part
650
651    packssdw    mm2, mm4                ; mm2=(A1 A3 B1 B3)
652    packssdw    mm5, mm3                ; mm5=(A5 A7 B5 B7)
653    pmaddwd     mm2, [GOTOFF(ebx,PW_F362_MF127)]
654    pmaddwd     mm5, [GOTOFF(ebx,PW_F085_MF072)]
655
656    paddd       mm2, mm5                ; mm2=tmp0[row0 row1]
657
658    ; -- Even part
659
660    pslld       mm1, (CONST_BITS+2)     ; mm1=tmp10[row0 row1]
661
662    ; -- Final output stage
663
664    movq        mm0, [GOTOFF(ebx,PD_DESCALE_P2_2)]  ; mm0=[PD_DESCALE_P2_2]
665
666    movq        mm6, mm1
667    paddd       mm1, mm2                ; mm1=data0[row0 row1]=(C0 C1)
668    psubd       mm6, mm2                ; mm6=data1[row0 row1]=(D0 D1)
669
670    paddd       mm1, mm0
671    paddd       mm6, mm0
672    psrad       mm1, DESCALE_P2_2
673    psrad       mm6, DESCALE_P2_2
674
675    movq        mm7, mm1                ; transpose coefficients
676    punpckldq   mm1, mm6                ; mm1=(C0 D0)
677    punpckhdq   mm7, mm6                ; mm7=(C1 D1)
678
679    packssdw    mm1, mm7                ; mm1=(C0 D0 C1 D1)
680    packsswb    mm1, mm1                ; mm1=(C0 D0 C1 D1 C0 D0 C1 D1)
681    paddb       mm1, [GOTOFF(ebx,PB_CENTERJSAMP)]
682
683    movd        ecx, mm1
684    movd        ebx, mm1                ; ebx=(C0 D0 C1 D1)
685    shr         ecx, 2*BYTE_BIT         ; ecx=(C1 D1 -- --)
686
687    mov         edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
688    mov         esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
689    mov         word [edx+eax*SIZEOF_JSAMPLE], bx
690    mov         word [esi+eax*SIZEOF_JSAMPLE], cx
691
692    emms                                ; empty MMX state
693
694    pop         edi
695    pop         esi
696;   pop         edx                     ; need not be preserved
697;   pop         ecx                     ; need not be preserved
698    pop         ebx
699    pop         ebp
700    ret
701
702; For some reason, the OS X linker does not honor the request to align the
703; segment unless we do this.
704    align       32
705