2 ; jfss2fst-64.asm - fast integer FDCT (64-bit SSE2)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright 2009 D. R. Commander
8 ; x86 SIMD extension for IJG JPEG library
9 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
10 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
12 ; This file should be assembled with NASM (Netwide Assembler),
13 ; can *not* be assembled with Microsoft's MASM or any compatible
14 ; assembler (including Borland's Turbo Assembler).
15 ; NASM is available from http://nasm.sourceforge.net/ or
16 ; http://sourceforge.net/project/showfiles.php?group_id=6208
18 ; This file contains a fast, not so accurate integer implementation of
19 ; the forward DCT (Discrete Cosine Transform). The following code is
20 ; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c
25 %include "jsimdext.inc"
28 ; --------------------------------------------------------------------------
30 %define CONST_BITS
8 ; 14 is also OK.
33 F_0_382
equ 98 ; FIX(0.382683433)
34 F_0_541
equ 139 ; FIX(0.541196100)
35 F_0_707
equ 181 ; FIX(0.707106781)
36 F_1_306
equ 334 ; FIX(1.306562965)
38 ; NASM cannot do compile-time arithmetic on floating-point constants.
39 %define DESCALE
(x
,n
) (((x
)+(1<<((n
)-1)))>>(n
))
40 F_0_382
equ DESCALE
( 410903207,30-CONST_BITS
) ; FIX(0.382683433)
41 F_0_541
equ DESCALE
( 581104887,30-CONST_BITS
) ; FIX(0.541196100)
42 F_0_707
equ DESCALE
( 759250124,30-CONST_BITS
) ; FIX(0.707106781)
43 F_1_306
equ DESCALE
(1402911301,30-CONST_BITS
) ; FIX(1.306562965)
46 ; --------------------------------------------------------------------------
49 ; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
50 ; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
52 %define PRE_MULTIPLY_SCALE_BITS
2
53 %define CONST_SHIFT
(16 - PRE_MULTIPLY_SCALE_BITS
- CONST_BITS
)
56 global EXTN
(jconst_fdct_ifast_sse2
)
58 EXTN
(jconst_fdct_ifast_sse2
):
60 PW_F0707 times
8 dw F_0_707
<< CONST_SHIFT
61 PW_F0382 times
8 dw F_0_382
<< CONST_SHIFT
62 PW_F0541 times
8 dw F_0_541
<< CONST_SHIFT
63 PW_F1306 times
8 dw F_1_306
<< CONST_SHIFT
67 ; --------------------------------------------------------------------------
71 ; Perform the forward DCT on one block of samples.
74 ; jsimd_fdct_ifast_sse2 (DCTELEM * data)
77 ; r10 = DCTELEM * data
79 %define wk
(i
) rbp
-(WK_NUM
-(i
))*SIZEOF_XMMWORD
; xmmword wk[WK_NUM]
83 global EXTN
(jsimd_fdct_ifast_sse2
)
85 EXTN
(jsimd_fdct_ifast_sse2
):
87 mov rax
,rsp
; rax = original rbp
89 and rsp
, byte (-SIZEOF_XMMWORD
) ; align to 128 bits
91 mov rbp
,rsp
; rbp = aligned rbp
95 ; ---- Pass 1: process rows.
97 mov rdx
, r10
; (DCTELEM *)
99 movdqa xmm0
, XMMWORD
[XMMBLOCK
(0,0,rdx
,SIZEOF_DCTELEM
)]
100 movdqa xmm1
, XMMWORD
[XMMBLOCK
(1,0,rdx
,SIZEOF_DCTELEM
)]
101 movdqa xmm2
, XMMWORD
[XMMBLOCK
(2,0,rdx
,SIZEOF_DCTELEM
)]
102 movdqa xmm3
, XMMWORD
[XMMBLOCK
(3,0,rdx
,SIZEOF_DCTELEM
)]
104 ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
105 ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
107 movdqa xmm4
,xmm0
; transpose coefficients(phase 1)
108 punpcklwd xmm0
,xmm1
; xmm0=(00 10 01 11 02 12 03 13)
109 punpckhwd xmm4
,xmm1
; xmm4=(04 14 05 15 06 16 07 17)
110 movdqa xmm5
,xmm2
; transpose coefficients(phase 1)
111 punpcklwd xmm2
,xmm3
; xmm2=(20 30 21 31 22 32 23 33)
112 punpckhwd xmm5
,xmm3
; xmm5=(24 34 25 35 26 36 27 37)
114 movdqa xmm6
, XMMWORD
[XMMBLOCK
(4,0,rdx
,SIZEOF_DCTELEM
)]
115 movdqa xmm7
, XMMWORD
[XMMBLOCK
(5,0,rdx
,SIZEOF_DCTELEM
)]
116 movdqa xmm1
, XMMWORD
[XMMBLOCK
(6,0,rdx
,SIZEOF_DCTELEM
)]
117 movdqa xmm3
, XMMWORD
[XMMBLOCK
(7,0,rdx
,SIZEOF_DCTELEM
)]
119 ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
120 ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
122 movdqa XMMWORD
[wk
(0)], xmm2
; wk(0)=(20 30 21 31 22 32 23 33)
123 movdqa XMMWORD
[wk
(1)], xmm5
; wk(1)=(24 34 25 35 26 36 27 37)
125 movdqa xmm2
,xmm6
; transpose coefficients(phase 1)
126 punpcklwd xmm6
,xmm7
; xmm6=(40 50 41 51 42 52 43 53)
127 punpckhwd xmm2
,xmm7
; xmm2=(44 54 45 55 46 56 47 57)
128 movdqa xmm5
,xmm1
; transpose coefficients(phase 1)
129 punpcklwd xmm1
,xmm3
; xmm1=(60 70 61 71 62 72 63 73)
130 punpckhwd xmm5
,xmm3
; xmm5=(64 74 65 75 66 76 67 77)
132 movdqa xmm7
,xmm6
; transpose coefficients(phase 2)
133 punpckldq xmm6
,xmm1
; xmm6=(40 50 60 70 41 51 61 71)
134 punpckhdq xmm7
,xmm1
; xmm7=(42 52 62 72 43 53 63 73)
135 movdqa xmm3
,xmm2
; transpose coefficients(phase 2)
136 punpckldq xmm2
,xmm5
; xmm2=(44 54 64 74 45 55 65 75)
137 punpckhdq xmm3
,xmm5
; xmm3=(46 56 66 76 47 57 67 77)
139 movdqa xmm1
, XMMWORD
[wk
(0)] ; xmm1=(20 30 21 31 22 32 23 33)
140 movdqa xmm5
, XMMWORD
[wk
(1)] ; xmm5=(24 34 25 35 26 36 27 37)
141 movdqa XMMWORD
[wk
(0)], xmm7
; wk(0)=(42 52 62 72 43 53 63 73)
142 movdqa XMMWORD
[wk
(1)], xmm2
; wk(1)=(44 54 64 74 45 55 65 75)
144 movdqa xmm7
,xmm0
; transpose coefficients(phase 2)
145 punpckldq xmm0
,xmm1
; xmm0=(00 10 20 30 01 11 21 31)
146 punpckhdq xmm7
,xmm1
; xmm7=(02 12 22 32 03 13 23 33)
147 movdqa xmm2
,xmm4
; transpose coefficients(phase 2)
148 punpckldq xmm4
,xmm5
; xmm4=(04 14 24 34 05 15 25 35)
149 punpckhdq xmm2
,xmm5
; xmm2=(06 16 26 36 07 17 27 37)
151 movdqa xmm1
,xmm0
; transpose coefficients(phase 3)
152 punpcklqdq xmm0
,xmm6
; xmm0=(00 10 20 30 40 50 60 70)=data0
153 punpckhqdq xmm1
,xmm6
; xmm1=(01 11 21 31 41 51 61 71)=data1
154 movdqa xmm5
,xmm2
; transpose coefficients(phase 3)
155 punpcklqdq xmm2
,xmm3
; xmm2=(06 16 26 36 46 56 66 76)=data6
156 punpckhqdq xmm5
,xmm3
; xmm5=(07 17 27 37 47 57 67 77)=data7
160 psubw xmm1
,xmm2
; xmm1=data1-data6=tmp6
161 psubw xmm0
,xmm5
; xmm0=data0-data7=tmp7
162 paddw xmm6
,xmm2
; xmm6=data1+data6=tmp1
163 paddw xmm3
,xmm5
; xmm3=data0+data7=tmp0
165 movdqa xmm2
, XMMWORD
[wk
(0)] ; xmm2=(42 52 62 72 43 53 63 73)
166 movdqa xmm5
, XMMWORD
[wk
(1)] ; xmm5=(44 54 64 74 45 55 65 75)
167 movdqa XMMWORD
[wk
(0)], xmm1
; wk(0)=tmp6
168 movdqa XMMWORD
[wk
(1)], xmm0
; wk(1)=tmp7
170 movdqa xmm1
,xmm7
; transpose coefficients(phase 3)
171 punpcklqdq xmm7
,xmm2
; xmm7=(02 12 22 32 42 52 62 72)=data2
172 punpckhqdq xmm1
,xmm2
; xmm1=(03 13 23 33 43 53 63 73)=data3
173 movdqa xmm0
,xmm4
; transpose coefficients(phase 3)
174 punpcklqdq xmm4
,xmm5
; xmm4=(04 14 24 34 44 54 64 74)=data4
175 punpckhqdq xmm0
,xmm5
; xmm0=(05 15 25 35 45 55 65 75)=data5
179 paddw xmm1
,xmm4
; xmm1=data3+data4=tmp3
180 paddw xmm7
,xmm0
; xmm7=data2+data5=tmp2
181 psubw xmm2
,xmm4
; xmm2=data3-data4=tmp4
182 psubw xmm5
,xmm0
; xmm5=data2-data5=tmp5
188 psubw xmm3
,xmm1
; xmm3=tmp13
189 psubw xmm6
,xmm7
; xmm6=tmp12
190 paddw xmm4
,xmm1
; xmm4=tmp10
191 paddw xmm0
,xmm7
; xmm0=tmp11
194 psllw xmm6
,PRE_MULTIPLY_SCALE_BITS
195 pmulhw xmm6
,[rel PW_F0707
] ; xmm6=z1
199 psubw xmm4
,xmm0
; xmm4=data4
200 psubw xmm3
,xmm6
; xmm3=data6
201 paddw xmm1
,xmm0
; xmm1=data0
202 paddw xmm7
,xmm6
; xmm7=data2
204 movdqa xmm0
, XMMWORD
[wk
(0)] ; xmm0=tmp6
205 movdqa xmm6
, XMMWORD
[wk
(1)] ; xmm6=tmp7
206 movdqa XMMWORD
[wk
(0)], xmm4
; wk(0)=data4
207 movdqa XMMWORD
[wk
(1)], xmm3
; wk(1)=data6
211 paddw xmm2
,xmm5
; xmm2=tmp10
212 paddw xmm5
,xmm0
; xmm5=tmp11
213 paddw xmm0
,xmm6
; xmm0=tmp12, xmm6=tmp7
215 psllw xmm2
,PRE_MULTIPLY_SCALE_BITS
216 psllw xmm0
,PRE_MULTIPLY_SCALE_BITS
218 psllw xmm5
,PRE_MULTIPLY_SCALE_BITS
219 pmulhw xmm5
,[rel PW_F0707
] ; xmm5=z3
221 movdqa xmm4
,xmm2
; xmm4=tmp10
223 pmulhw xmm2
,[rel PW_F0382
] ; xmm2=z5
224 pmulhw xmm4
,[rel PW_F0541
] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
225 pmulhw xmm0
,[rel PW_F1306
] ; xmm0=MULTIPLY(tmp12,FIX_1_306562)
226 paddw xmm4
,xmm2
; xmm4=z2
227 paddw xmm0
,xmm2
; xmm0=z4
230 psubw xmm6
,xmm5
; xmm6=z13
231 paddw xmm3
,xmm5
; xmm3=z11
235 psubw xmm6
,xmm4
; xmm6=data3
236 psubw xmm3
,xmm0
; xmm3=data7
237 paddw xmm2
,xmm4
; xmm2=data5
238 paddw xmm5
,xmm0
; xmm5=data1
240 ; ---- Pass 2: process columns.
242 ; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72)
243 ; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73)
245 movdqa xmm4
,xmm1
; transpose coefficients(phase 1)
246 punpcklwd xmm1
,xmm5
; xmm1=(00 01 10 11 20 21 30 31)
247 punpckhwd xmm4
,xmm5
; xmm4=(40 41 50 51 60 61 70 71)
248 movdqa xmm0
,xmm7
; transpose coefficients(phase 1)
249 punpcklwd xmm7
,xmm6
; xmm7=(02 03 12 13 22 23 32 33)
250 punpckhwd xmm0
,xmm6
; xmm0=(42 43 52 53 62 63 72 73)
252 movdqa xmm5
, XMMWORD
[wk
(0)] ; xmm5=col4
253 movdqa xmm6
, XMMWORD
[wk
(1)] ; xmm6=col6
255 ; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76)
256 ; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77)
258 movdqa XMMWORD
[wk
(0)], xmm7
; wk(0)=(02 03 12 13 22 23 32 33)
259 movdqa XMMWORD
[wk
(1)], xmm0
; wk(1)=(42 43 52 53 62 63 72 73)
261 movdqa xmm7
,xmm5
; transpose coefficients(phase 1)
262 punpcklwd xmm5
,xmm2
; xmm5=(04 05 14 15 24 25 34 35)
263 punpckhwd xmm7
,xmm2
; xmm7=(44 45 54 55 64 65 74 75)
264 movdqa xmm0
,xmm6
; transpose coefficients(phase 1)
265 punpcklwd xmm6
,xmm3
; xmm6=(06 07 16 17 26 27 36 37)
266 punpckhwd xmm0
,xmm3
; xmm0=(46 47 56 57 66 67 76 77)
268 movdqa xmm2
,xmm5
; transpose coefficients(phase 2)
269 punpckldq xmm5
,xmm6
; xmm5=(04 05 06 07 14 15 16 17)
270 punpckhdq xmm2
,xmm6
; xmm2=(24 25 26 27 34 35 36 37)
271 movdqa xmm3
,xmm7
; transpose coefficients(phase 2)
272 punpckldq xmm7
,xmm0
; xmm7=(44 45 46 47 54 55 56 57)
273 punpckhdq xmm3
,xmm0
; xmm3=(64 65 66 67 74 75 76 77)
275 movdqa xmm6
, XMMWORD
[wk
(0)] ; xmm6=(02 03 12 13 22 23 32 33)
276 movdqa xmm0
, XMMWORD
[wk
(1)] ; xmm0=(42 43 52 53 62 63 72 73)
277 movdqa XMMWORD
[wk
(0)], xmm2
; wk(0)=(24 25 26 27 34 35 36 37)
278 movdqa XMMWORD
[wk
(1)], xmm7
; wk(1)=(44 45 46 47 54 55 56 57)
280 movdqa xmm2
,xmm1
; transpose coefficients(phase 2)
281 punpckldq xmm1
,xmm6
; xmm1=(00 01 02 03 10 11 12 13)
282 punpckhdq xmm2
,xmm6
; xmm2=(20 21 22 23 30 31 32 33)
283 movdqa xmm7
,xmm4
; transpose coefficients(phase 2)
284 punpckldq xmm4
,xmm0
; xmm4=(40 41 42 43 50 51 52 53)
285 punpckhdq xmm7
,xmm0
; xmm7=(60 61 62 63 70 71 72 73)
287 movdqa xmm6
,xmm1
; transpose coefficients(phase 3)
288 punpcklqdq xmm1
,xmm5
; xmm1=(00 01 02 03 04 05 06 07)=data0
289 punpckhqdq xmm6
,xmm5
; xmm6=(10 11 12 13 14 15 16 17)=data1
290 movdqa xmm0
,xmm7
; transpose coefficients(phase 3)
291 punpcklqdq xmm7
,xmm3
; xmm7=(60 61 62 63 64 65 66 67)=data6
292 punpckhqdq xmm0
,xmm3
; xmm0=(70 71 72 73 74 75 76 77)=data7
296 psubw xmm6
,xmm7
; xmm6=data1-data6=tmp6
297 psubw xmm1
,xmm0
; xmm1=data0-data7=tmp7
298 paddw xmm5
,xmm7
; xmm5=data1+data6=tmp1
299 paddw xmm3
,xmm0
; xmm3=data0+data7=tmp0
301 movdqa xmm7
, XMMWORD
[wk
(0)] ; xmm7=(24 25 26 27 34 35 36 37)
302 movdqa xmm0
, XMMWORD
[wk
(1)] ; xmm0=(44 45 46 47 54 55 56 57)
303 movdqa XMMWORD
[wk
(0)], xmm6
; wk(0)=tmp6
304 movdqa XMMWORD
[wk
(1)], xmm1
; wk(1)=tmp7
306 movdqa xmm6
,xmm2
; transpose coefficients(phase 3)
307 punpcklqdq xmm2
,xmm7
; xmm2=(20 21 22 23 24 25 26 27)=data2
308 punpckhqdq xmm6
,xmm7
; xmm6=(30 31 32 33 34 35 36 37)=data3
309 movdqa xmm1
,xmm4
; transpose coefficients(phase 3)
310 punpcklqdq xmm4
,xmm0
; xmm4=(40 41 42 43 44 45 46 47)=data4
311 punpckhqdq xmm1
,xmm0
; xmm1=(50 51 52 53 54 55 56 57)=data5
315 paddw xmm6
,xmm4
; xmm6=data3+data4=tmp3
316 paddw xmm2
,xmm1
; xmm2=data2+data5=tmp2
317 psubw xmm7
,xmm4
; xmm7=data3-data4=tmp4
318 psubw xmm0
,xmm1
; xmm0=data2-data5=tmp5
324 psubw xmm3
,xmm6
; xmm3=tmp13
325 psubw xmm5
,xmm2
; xmm5=tmp12
326 paddw xmm4
,xmm6
; xmm4=tmp10
327 paddw xmm1
,xmm2
; xmm1=tmp11
330 psllw xmm5
,PRE_MULTIPLY_SCALE_BITS
331 pmulhw xmm5
,[rel PW_F0707
] ; xmm5=z1
335 psubw xmm4
,xmm1
; xmm4=data4
336 psubw xmm3
,xmm5
; xmm3=data6
337 paddw xmm6
,xmm1
; xmm6=data0
338 paddw xmm2
,xmm5
; xmm2=data2
340 movdqa XMMWORD
[XMMBLOCK
(4,0,rdx
,SIZEOF_DCTELEM
)], xmm4
341 movdqa XMMWORD
[XMMBLOCK
(6,0,rdx
,SIZEOF_DCTELEM
)], xmm3
342 movdqa XMMWORD
[XMMBLOCK
(0,0,rdx
,SIZEOF_DCTELEM
)], xmm6
343 movdqa XMMWORD
[XMMBLOCK
(2,0,rdx
,SIZEOF_DCTELEM
)], xmm2
347 movdqa xmm1
, XMMWORD
[wk
(0)] ; xmm1=tmp6
348 movdqa xmm5
, XMMWORD
[wk
(1)] ; xmm5=tmp7
350 paddw xmm7
,xmm0
; xmm7=tmp10
351 paddw xmm0
,xmm1
; xmm0=tmp11
352 paddw xmm1
,xmm5
; xmm1=tmp12, xmm5=tmp7
354 psllw xmm7
,PRE_MULTIPLY_SCALE_BITS
355 psllw xmm1
,PRE_MULTIPLY_SCALE_BITS
357 psllw xmm0
,PRE_MULTIPLY_SCALE_BITS
358 pmulhw xmm0
,[rel PW_F0707
] ; xmm0=z3
360 movdqa xmm4
,xmm7
; xmm4=tmp10
362 pmulhw xmm7
,[rel PW_F0382
] ; xmm7=z5
363 pmulhw xmm4
,[rel PW_F0541
] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
364 pmulhw xmm1
,[rel PW_F1306
] ; xmm1=MULTIPLY(tmp12,FIX_1_306562)
365 paddw xmm4
,xmm7
; xmm4=z2
366 paddw xmm1
,xmm7
; xmm1=z4
369 psubw xmm5
,xmm0
; xmm5=z13
370 paddw xmm3
,xmm0
; xmm3=z11
374 psubw xmm5
,xmm4
; xmm5=data3
375 psubw xmm3
,xmm1
; xmm3=data7
376 paddw xmm6
,xmm4
; xmm6=data5
377 paddw xmm2
,xmm1
; xmm2=data1
379 movdqa XMMWORD
[XMMBLOCK
(3,0,rdx
,SIZEOF_DCTELEM
)], xmm5
380 movdqa XMMWORD
[XMMBLOCK
(7,0,rdx
,SIZEOF_DCTELEM
)], xmm3
381 movdqa XMMWORD
[XMMBLOCK
(5,0,rdx
,SIZEOF_DCTELEM
)], xmm6
382 movdqa XMMWORD
[XMMBLOCK
(1,0,rdx
,SIZEOF_DCTELEM
)], xmm2
385 mov rsp
,rbp
; rsp <- aligned rbp
386 pop rsp
; rsp <- original rbp
390 ; For some reason, the OS X linker does not honor the request to align the
391 ; segment unless we do this.