1;
2; jidctred.asm - reduced-size IDCT (64-bit SSE2)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 2009, 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; [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      32
75    GLOBAL_DATA(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      32
93
94; --------------------------------------------------------------------------
95    SECTION     SEG_TEXT
96    BITS        64
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; r10 = void *dct_table
107; r11 = JCOEFPTR coef_block
108; r12 = JSAMPARRAY output_buf
109; r13d = JDIMENSION output_col
110
111%define original_rbp  rbp + 0
112%define wk(i)         rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD
113                                        ; xmmword wk[WK_NUM]
114%define WK_NUM        2
115
116    align       32
117    GLOBAL_FUNCTION(jsimd_idct_4x4_sse2)
118
119EXTN(jsimd_idct_4x4_sse2):
120    push        rbp
121    mov         rax, rsp                     ; rax = original rbp
122    sub         rsp, byte 4
123    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
124    mov         [rsp], rax
125    mov         rbp, rsp                     ; rbp = aligned rbp
126    lea         rsp, [wk(0)]
127    collect_args 4
128
129    ; ---- Pass 1: process columns from input.
130
131    mov         rdx, r10                ; quantptr
132    mov         rsi, r11                ; inptr
133
134%ifndef NO_ZERO_COLUMN_TEST_4X4_SSE2
135    mov         eax, DWORD [DWBLOCK(1,0,rsi,SIZEOF_JCOEF)]
136    or          eax, DWORD [DWBLOCK(2,0,rsi,SIZEOF_JCOEF)]
137    jnz         short .columnDCT
138
139    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
140    movdqa      xmm1, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
141    por         xmm0, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
142    por         xmm1, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
143    por         xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
144    por         xmm1, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
145    por         xmm0, xmm1
146    packsswb    xmm0, xmm0
147    packsswb    xmm0, xmm0
148    movd        eax, xmm0
149    test        rax, rax
150    jnz         short .columnDCT
151
152    ; -- AC terms all zero
153
154    movdqa      xmm0, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
155    pmullw      xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
156
157    psllw       xmm0, PASS1_BITS
158
159    movdqa      xmm3, xmm0        ; xmm0=in0=(00 01 02 03 04 05 06 07)
160    punpcklwd   xmm0, xmm0        ; xmm0=(00 00 01 01 02 02 03 03)
161    punpckhwd   xmm3, xmm3        ; xmm3=(04 04 05 05 06 06 07 07)
162
163    pshufd      xmm1, xmm0, 0x50  ; xmm1=[col0 col1]=(00 00 00 00 01 01 01 01)
164    pshufd      xmm0, xmm0, 0xFA  ; xmm0=[col2 col3]=(02 02 02 02 03 03 03 03)
165    pshufd      xmm6, xmm3, 0x50  ; xmm6=[col4 col5]=(04 04 04 04 05 05 05 05)
166    pshufd      xmm3, xmm3, 0xFA  ; xmm3=[col6 col7]=(06 06 06 06 07 07 07 07)
167
168    jmp         near .column_end
169%endif
170.columnDCT:
171
172    ; -- Odd part
173
174    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
175    movdqa      xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
176    pmullw      xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
177    pmullw      xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
178    movdqa      xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
179    movdqa      xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
180    pmullw      xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
181    pmullw      xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
182
183    movdqa      xmm4, xmm0
184    movdqa      xmm5, xmm0
185    punpcklwd   xmm4, xmm1
186    punpckhwd   xmm5, xmm1
187    movdqa      xmm0, xmm4
188    movdqa      xmm1, xmm5
189    pmaddwd     xmm4, [rel PW_F256_F089]   ; xmm4=(tmp2L)
190    pmaddwd     xmm5, [rel PW_F256_F089]   ; xmm5=(tmp2H)
191    pmaddwd     xmm0, [rel PW_F106_MF217]  ; xmm0=(tmp0L)
192    pmaddwd     xmm1, [rel PW_F106_MF217]  ; xmm1=(tmp0H)
193
194    movdqa      xmm6, xmm2
195    movdqa      xmm7, xmm2
196    punpcklwd   xmm6, xmm3
197    punpckhwd   xmm7, xmm3
198    movdqa      xmm2, xmm6
199    movdqa      xmm3, xmm7
200    pmaddwd     xmm6, [rel PW_MF060_MF050]  ; xmm6=(tmp2L)
201    pmaddwd     xmm7, [rel PW_MF060_MF050]  ; xmm7=(tmp2H)
202    pmaddwd     xmm2, [rel PW_F145_MF021]   ; xmm2=(tmp0L)
203    pmaddwd     xmm3, [rel PW_F145_MF021]   ; xmm3=(tmp0H)
204
205    paddd       xmm6, xmm4              ; xmm6=tmp2L
206    paddd       xmm7, xmm5              ; xmm7=tmp2H
207    paddd       xmm2, xmm0              ; xmm2=tmp0L
208    paddd       xmm3, xmm1              ; xmm3=tmp0H
209
210    movdqa      XMMWORD [wk(0)], xmm2   ; wk(0)=tmp0L
211    movdqa      XMMWORD [wk(1)], xmm3   ; wk(1)=tmp0H
212
213    ; -- Even part
214
215    movdqa      xmm4, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
216    movdqa      xmm5, XMMWORD [XMMBLOCK(2,0,rsi,SIZEOF_JCOEF)]
217    movdqa      xmm0, XMMWORD [XMMBLOCK(6,0,rsi,SIZEOF_JCOEF)]
218    pmullw      xmm4, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
219    pmullw      xmm5, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
220    pmullw      xmm0, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
221
222    pxor        xmm1, xmm1
223    pxor        xmm2, xmm2
224    punpcklwd   xmm1, xmm4               ; xmm1=tmp0L
225    punpckhwd   xmm2, xmm4               ; xmm2=tmp0H
226    psrad       xmm1, (16-CONST_BITS-1)  ; psrad xmm1,16 & pslld xmm1,CONST_BITS+1
227    psrad       xmm2, (16-CONST_BITS-1)  ; psrad xmm2,16 & pslld xmm2,CONST_BITS+1
228
229    movdqa      xmm3, xmm5              ; xmm5=in2=z2
230    punpcklwd   xmm5, xmm0              ; xmm0=in6=z3
231    punpckhwd   xmm3, xmm0
232    pmaddwd     xmm5, [rel PW_F184_MF076]  ; xmm5=tmp2L
233    pmaddwd     xmm3, [rel PW_F184_MF076]  ; xmm3=tmp2H
234
235    movdqa      xmm4, xmm1
236    movdqa      xmm0, xmm2
237    paddd       xmm1, xmm5              ; xmm1=tmp10L
238    paddd       xmm2, xmm3              ; xmm2=tmp10H
239    psubd       xmm4, xmm5              ; xmm4=tmp12L
240    psubd       xmm0, xmm3              ; xmm0=tmp12H
241
242    ; -- Final output stage
243
244    movdqa      xmm5, xmm1
245    movdqa      xmm3, xmm2
246    paddd       xmm1, xmm6              ; xmm1=data0L
247    paddd       xmm2, xmm7              ; xmm2=data0H
248    psubd       xmm5, xmm6              ; xmm5=data3L
249    psubd       xmm3, xmm7              ; xmm3=data3H
250
251    movdqa      xmm6, [rel PD_DESCALE_P1_4]  ; xmm6=[rel PD_DESCALE_P1_4]
252
253    paddd       xmm1, xmm6
254    paddd       xmm2, xmm6
255    psrad       xmm1, DESCALE_P1_4
256    psrad       xmm2, DESCALE_P1_4
257    paddd       xmm5, xmm6
258    paddd       xmm3, xmm6
259    psrad       xmm5, DESCALE_P1_4
260    psrad       xmm3, DESCALE_P1_4
261
262    packssdw    xmm1, xmm2              ; xmm1=data0=(00 01 02 03 04 05 06 07)
263    packssdw    xmm5, xmm3              ; xmm5=data3=(30 31 32 33 34 35 36 37)
264
265    movdqa      xmm7, XMMWORD [wk(0)]   ; xmm7=tmp0L
266    movdqa      xmm6, XMMWORD [wk(1)]   ; xmm6=tmp0H
267
268    movdqa      xmm2, xmm4
269    movdqa      xmm3, xmm0
270    paddd       xmm4, xmm7              ; xmm4=data1L
271    paddd       xmm0, xmm6              ; xmm0=data1H
272    psubd       xmm2, xmm7              ; xmm2=data2L
273    psubd       xmm3, xmm6              ; xmm3=data2H
274
275    movdqa      xmm7, [rel PD_DESCALE_P1_4]  ; xmm7=[rel PD_DESCALE_P1_4]
276
277    paddd       xmm4, xmm7
278    paddd       xmm0, xmm7
279    psrad       xmm4, DESCALE_P1_4
280    psrad       xmm0, DESCALE_P1_4
281    paddd       xmm2, xmm7
282    paddd       xmm3, xmm7
283    psrad       xmm2, DESCALE_P1_4
284    psrad       xmm3, DESCALE_P1_4
285
286    packssdw    xmm4, xmm0        ; xmm4=data1=(10 11 12 13 14 15 16 17)
287    packssdw    xmm2, xmm3        ; xmm2=data2=(20 21 22 23 24 25 26 27)
288
289    movdqa      xmm6, xmm1        ; transpose coefficients(phase 1)
290    punpcklwd   xmm1, xmm4        ; xmm1=(00 10 01 11 02 12 03 13)
291    punpckhwd   xmm6, xmm4        ; xmm6=(04 14 05 15 06 16 07 17)
292    movdqa      xmm7, xmm2        ; transpose coefficients(phase 1)
293    punpcklwd   xmm2, xmm5        ; xmm2=(20 30 21 31 22 32 23 33)
294    punpckhwd   xmm7, xmm5        ; xmm7=(24 34 25 35 26 36 27 37)
295
296    movdqa      xmm0, xmm1        ; transpose coefficients(phase 2)
297    punpckldq   xmm1, xmm2        ; xmm1=[col0 col1]=(00 10 20 30 01 11 21 31)
298    punpckhdq   xmm0, xmm2        ; xmm0=[col2 col3]=(02 12 22 32 03 13 23 33)
299    movdqa      xmm3, xmm6        ; transpose coefficients(phase 2)
300    punpckldq   xmm6, xmm7        ; xmm6=[col4 col5]=(04 14 24 34 05 15 25 35)
301    punpckhdq   xmm3, xmm7        ; xmm3=[col6 col7]=(06 16 26 36 07 17 27 37)
302.column_end:
303
304    ; -- Prefetch the next coefficient block
305
306    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
307    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
308    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
309    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
310
311    ; ---- Pass 2: process rows, store into output array.
312
313    mov         rax, [original_rbp]
314    mov         rdi, r12                ; (JSAMPROW *)
315    mov         eax, r13d
316
317    ; -- Even part
318
319    pxor        xmm4, xmm4
320    punpcklwd   xmm4, xmm1               ; xmm4=tmp0
321    psrad       xmm4, (16-CONST_BITS-1)  ; psrad xmm4,16 & pslld xmm4,CONST_BITS+1
322
323    ; -- Odd part
324
325    punpckhwd   xmm1, xmm0
326    punpckhwd   xmm6, xmm3
327    movdqa      xmm5, xmm1
328    movdqa      xmm2, xmm6
329    pmaddwd     xmm1, [rel PW_F256_F089]    ; xmm1=(tmp2)
330    pmaddwd     xmm6, [rel PW_MF060_MF050]  ; xmm6=(tmp2)
331    pmaddwd     xmm5, [rel PW_F106_MF217]   ; xmm5=(tmp0)
332    pmaddwd     xmm2, [rel PW_F145_MF021]   ; xmm2=(tmp0)
333
334    paddd       xmm6, xmm1              ; xmm6=tmp2
335    paddd       xmm2, xmm5              ; xmm2=tmp0
336
337    ; -- Even part
338
339    punpcklwd   xmm0, xmm3
340    pmaddwd     xmm0, [rel PW_F184_MF076]  ; xmm0=tmp2
341
342    movdqa      xmm7, xmm4
343    paddd       xmm4, xmm0              ; xmm4=tmp10
344    psubd       xmm7, xmm0              ; xmm7=tmp12
345
346    ; -- Final output stage
347
348    movdqa      xmm1, [rel PD_DESCALE_P2_4]  ; xmm1=[rel PD_DESCALE_P2_4]
349
350    movdqa      xmm5, xmm4
351    movdqa      xmm3, xmm7
352    paddd       xmm4, xmm6              ; xmm4=data0=(00 10 20 30)
353    paddd       xmm7, xmm2              ; xmm7=data1=(01 11 21 31)
354    psubd       xmm5, xmm6              ; xmm5=data3=(03 13 23 33)
355    psubd       xmm3, xmm2              ; xmm3=data2=(02 12 22 32)
356
357    paddd       xmm4, xmm1
358    paddd       xmm7, xmm1
359    psrad       xmm4, DESCALE_P2_4
360    psrad       xmm7, DESCALE_P2_4
361    paddd       xmm5, xmm1
362    paddd       xmm3, xmm1
363    psrad       xmm5, DESCALE_P2_4
364    psrad       xmm3, DESCALE_P2_4
365
366    packssdw    xmm4, xmm3              ; xmm4=(00 10 20 30 02 12 22 32)
367    packssdw    xmm7, xmm5              ; xmm7=(01 11 21 31 03 13 23 33)
368
369    movdqa      xmm0, xmm4              ; transpose coefficients(phase 1)
370    punpcklwd   xmm4, xmm7              ; xmm4=(00 01 10 11 20 21 30 31)
371    punpckhwd   xmm0, xmm7              ; xmm0=(02 03 12 13 22 23 32 33)
372
373    movdqa      xmm6, xmm4              ; transpose coefficients(phase 2)
374    punpckldq   xmm4, xmm0              ; xmm4=(00 01 02 03 10 11 12 13)
375    punpckhdq   xmm6, xmm0              ; xmm6=(20 21 22 23 30 31 32 33)
376
377    packsswb    xmm4, xmm6              ; xmm4=(00 01 02 03 10 11 12 13 20 ..)
378    paddb       xmm4, [rel PB_CENTERJSAMP]
379
380    pshufd      xmm2, xmm4, 0x39        ; xmm2=(10 11 12 13 20 21 22 23 30 ..)
381    pshufd      xmm1, xmm4, 0x4E        ; xmm1=(20 21 22 23 30 31 32 33 00 ..)
382    pshufd      xmm3, xmm4, 0x93        ; xmm3=(30 31 32 33 00 01 02 03 10 ..)
383
384    mov         rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
385    mov         rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
386    movd        XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm4
387    movd        XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm2
388    mov         rdx, JSAMPROW [rdi+2*SIZEOF_JSAMPROW]
389    mov         rsi, JSAMPROW [rdi+3*SIZEOF_JSAMPROW]
390    movd        XMM_DWORD [rdx+rax*SIZEOF_JSAMPLE], xmm1
391    movd        XMM_DWORD [rsi+rax*SIZEOF_JSAMPLE], xmm3
392
393    uncollect_args 4
394    mov         rsp, rbp                ; rsp <- aligned rbp
395    pop         rsp                     ; rsp <- original rbp
396    pop         rbp
397    ret
398
399; --------------------------------------------------------------------------
400;
401; Perform dequantization and inverse DCT on one block of coefficients,
402; producing a reduced-size 2x2 output block.
403;
404; GLOBAL(void)
405; jsimd_idct_2x2_sse2(void *dct_table, JCOEFPTR coef_block,
406;                     JSAMPARRAY output_buf, JDIMENSION output_col)
407;
408
409; r10 = void *dct_table
410; r11 = JCOEFPTR coef_block
411; r12 = JSAMPARRAY output_buf
412; r13d = JDIMENSION output_col
413
414    align       32
415    GLOBAL_FUNCTION(jsimd_idct_2x2_sse2)
416
417EXTN(jsimd_idct_2x2_sse2):
418    push        rbp
419    mov         rax, rsp
420    mov         rbp, rsp
421    collect_args 4
422    push        rbx
423
424    ; ---- Pass 1: process columns from input.
425
426    mov         rdx, r10                ; quantptr
427    mov         rsi, r11                ; inptr
428
429    ; | input:                  | result:        |
430    ; | 00 01 ** 03 ** 05 ** 07 |                |
431    ; | 10 11 ** 13 ** 15 ** 17 |                |
432    ; | ** ** ** ** ** ** ** ** |                |
433    ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
434    ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
435    ; | 50 51 ** 53 ** 55 ** 57 |                |
436    ; | ** ** ** ** ** ** ** ** |                |
437    ; | 70 71 ** 73 ** 75 ** 77 |                |
438
439    ; -- Odd part
440
441    movdqa      xmm0, XMMWORD [XMMBLOCK(1,0,rsi,SIZEOF_JCOEF)]
442    movdqa      xmm1, XMMWORD [XMMBLOCK(3,0,rsi,SIZEOF_JCOEF)]
443    pmullw      xmm0, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
444    pmullw      xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
445    movdqa      xmm2, XMMWORD [XMMBLOCK(5,0,rsi,SIZEOF_JCOEF)]
446    movdqa      xmm3, XMMWORD [XMMBLOCK(7,0,rsi,SIZEOF_JCOEF)]
447    pmullw      xmm2, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
448    pmullw      xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
449
450    ; xmm0=(10 11 ** 13 ** 15 ** 17), xmm1=(30 31 ** 33 ** 35 ** 37)
451    ; xmm2=(50 51 ** 53 ** 55 ** 57), xmm3=(70 71 ** 73 ** 75 ** 77)
452
453    pcmpeqd     xmm7, xmm7
454    pslld       xmm7, WORD_BIT          ; xmm7={0x0000 0xFFFF 0x0000 0xFFFF ..}
455
456    movdqa      xmm4, xmm0              ; xmm4=(10 11 ** 13 ** 15 ** 17)
457    movdqa      xmm5, xmm2              ; xmm5=(50 51 ** 53 ** 55 ** 57)
458    punpcklwd   xmm4, xmm1              ; xmm4=(10 30 11 31 ** ** 13 33)
459    punpcklwd   xmm5, xmm3              ; xmm5=(50 70 51 71 ** ** 53 73)
460    pmaddwd     xmm4, [rel PW_F362_MF127]
461    pmaddwd     xmm5, [rel PW_F085_MF072]
462
463    psrld       xmm0, WORD_BIT          ; xmm0=(11 -- 13 -- 15 -- 17 --)
464    pand        xmm1, xmm7              ; xmm1=(-- 31 -- 33 -- 35 -- 37)
465    psrld       xmm2, WORD_BIT          ; xmm2=(51 -- 53 -- 55 -- 57 --)
466    pand        xmm3, xmm7              ; xmm3=(-- 71 -- 73 -- 75 -- 77)
467    por         xmm0, xmm1              ; xmm0=(11 31 13 33 15 35 17 37)
468    por         xmm2, xmm3              ; xmm2=(51 71 53 73 55 75 57 77)
469    pmaddwd     xmm0, [rel PW_F362_MF127]
470    pmaddwd     xmm2, [rel PW_F085_MF072]
471
472    paddd       xmm4, xmm5              ; xmm4=tmp0[col0 col1 **** col3]
473    paddd       xmm0, xmm2              ; xmm0=tmp0[col1 col3 col5 col7]
474
475    ; -- Even part
476
477    movdqa      xmm6, XMMWORD [XMMBLOCK(0,0,rsi,SIZEOF_JCOEF)]
478    pmullw      xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_ISLOW_MULT_TYPE)]
479
480    ; xmm6=(00 01 ** 03 ** 05 ** 07)
481
482    movdqa      xmm1, xmm6              ; xmm1=(00 01 ** 03 ** 05 ** 07)
483    pslld       xmm6, WORD_BIT          ; xmm6=(-- 00 -- ** -- ** -- **)
484    pand        xmm1, xmm7              ; xmm1=(-- 01 -- 03 -- 05 -- 07)
485    psrad       xmm6, (WORD_BIT-CONST_BITS-2)  ; xmm6=tmp10[col0 **** **** ****]
486    psrad       xmm1, (WORD_BIT-CONST_BITS-2)  ; xmm1=tmp10[col1 col3 col5 col7]
487
488    ; -- Final output stage
489
490    movdqa      xmm3, xmm6
491    movdqa      xmm5, xmm1
492    paddd       xmm6, xmm4      ; xmm6=data0[col0 **** **** ****]=(A0 ** ** **)
493    paddd       xmm1, xmm0      ; xmm1=data0[col1 col3 col5 col7]=(A1 A3 A5 A7)
494    psubd       xmm3, xmm4      ; xmm3=data1[col0 **** **** ****]=(B0 ** ** **)
495    psubd       xmm5, xmm0      ; xmm5=data1[col1 col3 col5 col7]=(B1 B3 B5 B7)
496
497    movdqa      xmm2, [rel PD_DESCALE_P1_2]  ; xmm2=[rel PD_DESCALE_P1_2]
498
499    punpckldq   xmm6, xmm3              ; xmm6=(A0 B0 ** **)
500
501    movdqa      xmm7, xmm1
502    punpcklqdq  xmm1, xmm5              ; xmm1=(A1 A3 B1 B3)
503    punpckhqdq  xmm7, xmm5              ; xmm7=(A5 A7 B5 B7)
504
505    paddd       xmm6, xmm2
506    psrad       xmm6, DESCALE_P1_2
507
508    paddd       xmm1, xmm2
509    paddd       xmm7, xmm2
510    psrad       xmm1, DESCALE_P1_2
511    psrad       xmm7, DESCALE_P1_2
512
513    ; -- Prefetch the next coefficient block
514
515    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 0*32]
516    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 1*32]
517    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 2*32]
518    prefetchnta [rsi + DCTSIZE2*SIZEOF_JCOEF + 3*32]
519
520    ; ---- Pass 2: process rows, store into output array.
521
522    mov         rdi, r12                ; (JSAMPROW *)
523    mov         eax, r13d
524
525    ; | input:| result:|
526    ; | A0 B0 |        |
527    ; | A1 B1 | C0 C1  |
528    ; | A3 B3 | D0 D1  |
529    ; | A5 B5 |        |
530    ; | A7 B7 |        |
531
532    ; -- Odd part
533
534    packssdw    xmm1, xmm1              ; xmm1=(A1 A3 B1 B3 A1 A3 B1 B3)
535    packssdw    xmm7, xmm7              ; xmm7=(A5 A7 B5 B7 A5 A7 B5 B7)
536    pmaddwd     xmm1, [rel PW_F362_MF127]
537    pmaddwd     xmm7, [rel PW_F085_MF072]
538
539    paddd       xmm1, xmm7              ; xmm1=tmp0[row0 row1 row0 row1]
540
541    ; -- Even part
542
543    pslld       xmm6, (CONST_BITS+2)    ; xmm6=tmp10[row0 row1 **** ****]
544
545    ; -- Final output stage
546
547    movdqa      xmm4, xmm6
548    paddd       xmm6, xmm1     ; xmm6=data0[row0 row1 **** ****]=(C0 C1 ** **)
549    psubd       xmm4, xmm1     ; xmm4=data1[row0 row1 **** ****]=(D0 D1 ** **)
550
551    punpckldq   xmm6, xmm4     ; xmm6=(C0 D0 C1 D1)
552
553    paddd       xmm6, [rel PD_DESCALE_P2_2]
554    psrad       xmm6, DESCALE_P2_2
555
556    packssdw    xmm6, xmm6              ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1)
557    packsswb    xmm6, xmm6              ; xmm6=(C0 D0 C1 D1 C0 D0 C1 D1 ..)
558    paddb       xmm6, [rel PB_CENTERJSAMP]
559
560    pextrw      ebx, xmm6, 0x00         ; ebx=(C0 D0 -- --)
561    pextrw      ecx, xmm6, 0x01         ; ecx=(C1 D1 -- --)
562
563    mov         rdx, JSAMPROW [rdi+0*SIZEOF_JSAMPROW]
564    mov         rsi, JSAMPROW [rdi+1*SIZEOF_JSAMPROW]
565    mov         WORD [rdx+rax*SIZEOF_JSAMPLE], bx
566    mov         WORD [rsi+rax*SIZEOF_JSAMPLE], cx
567
568    pop         rbx
569    uncollect_args 4
570    pop         rbp
571    ret
572
573; For some reason, the OS X linker does not honor the request to align the
574; segment unless we do this.
575    align       32
576