2 ; jfss2fst.asm - fast integer FDCT (SSE2)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
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
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
17 ; This file contains a fast, not so accurate integer implementation of
18 ; the forward DCT (Discrete Cosine Transform). The following code is
19 ; based directly on the IJG's original jfdctfst.c; see the jfdctfst.c
24 %include "jsimdext.inc"
27 ; --------------------------------------------------------------------------
29 %define CONST_BITS
8 ; 14 is also OK.
32 F_0_382
equ 98 ; FIX(0.382683433)
33 F_0_541
equ 139 ; FIX(0.541196100)
34 F_0_707
equ 181 ; FIX(0.707106781)
35 F_1_306
equ 334 ; FIX(1.306562965)
37 ; NASM cannot do compile-time arithmetic on floating-point constants.
38 %define DESCALE
(x
,n
) (((x
)+(1<<((n
)-1)))>>(n
))
39 F_0_382
equ DESCALE
( 410903207,30-CONST_BITS
) ; FIX(0.382683433)
40 F_0_541
equ DESCALE
( 581104887,30-CONST_BITS
) ; FIX(0.541196100)
41 F_0_707
equ DESCALE
( 759250124,30-CONST_BITS
) ; FIX(0.707106781)
42 F_1_306
equ DESCALE
(1402911301,30-CONST_BITS
) ; FIX(1.306562965)
45 ; --------------------------------------------------------------------------
48 ; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
49 ; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
51 %define PRE_MULTIPLY_SCALE_BITS
2
52 %define CONST_SHIFT
(16 - PRE_MULTIPLY_SCALE_BITS
- CONST_BITS
)
55 global EXTN
(jconst_fdct_ifast_sse2
)
57 EXTN
(jconst_fdct_ifast_sse2
):
59 PW_F0707 times
8 dw F_0_707
<< CONST_SHIFT
60 PW_F0382 times
8 dw F_0_382
<< CONST_SHIFT
61 PW_F0541 times
8 dw F_0_541
<< CONST_SHIFT
62 PW_F1306 times
8 dw F_1_306
<< CONST_SHIFT
66 ; --------------------------------------------------------------------------
70 ; Perform the forward DCT on one block of samples.
73 ; jsimd_fdct_ifast_sse2 (DCTELEM * data)
76 %define data
(b
) (b
)+8 ; DCTELEM * data
78 %define original_ebp
ebp+0
79 %define wk
(i
) ebp-(WK_NUM
-(i
))*SIZEOF_XMMWORD
; xmmword wk[WK_NUM]
83 global EXTN
(jsimd_fdct_ifast_sse2
)
85 EXTN
(jsimd_fdct_ifast_sse2
):
87 mov eax,esp ; eax = original ebp
89 and esp, byte (-SIZEOF_XMMWORD
) ; align to 128 bits
91 mov ebp,esp ; ebp = aligned ebp
95 ; push edx ; need not be preserved
99 get_GOT
ebx ; get GOT address
101 ; ---- Pass 1: process rows.
103 mov edx, POINTER
[data
(eax)] ; (DCTELEM *)
105 movdqa xmm0
, XMMWORD
[XMMBLOCK
(0,0,edx,SIZEOF_DCTELEM
)]
106 movdqa xmm1
, XMMWORD
[XMMBLOCK
(1,0,edx,SIZEOF_DCTELEM
)]
107 movdqa xmm2
, XMMWORD
[XMMBLOCK
(2,0,edx,SIZEOF_DCTELEM
)]
108 movdqa xmm3
, XMMWORD
[XMMBLOCK
(3,0,edx,SIZEOF_DCTELEM
)]
110 ; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
111 ; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
113 movdqa xmm4
,xmm0
; transpose coefficients(phase 1)
114 punpcklwd xmm0
,xmm1
; xmm0=(00 10 01 11 02 12 03 13)
115 punpckhwd xmm4
,xmm1
; xmm4=(04 14 05 15 06 16 07 17)
116 movdqa xmm5
,xmm2
; transpose coefficients(phase 1)
117 punpcklwd xmm2
,xmm3
; xmm2=(20 30 21 31 22 32 23 33)
118 punpckhwd xmm5
,xmm3
; xmm5=(24 34 25 35 26 36 27 37)
120 movdqa xmm6
, XMMWORD
[XMMBLOCK
(4,0,edx,SIZEOF_DCTELEM
)]
121 movdqa xmm7
, XMMWORD
[XMMBLOCK
(5,0,edx,SIZEOF_DCTELEM
)]
122 movdqa xmm1
, XMMWORD
[XMMBLOCK
(6,0,edx,SIZEOF_DCTELEM
)]
123 movdqa xmm3
, XMMWORD
[XMMBLOCK
(7,0,edx,SIZEOF_DCTELEM
)]
125 ; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
126 ; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
128 movdqa XMMWORD
[wk
(0)], xmm2
; wk(0)=(20 30 21 31 22 32 23 33)
129 movdqa XMMWORD
[wk
(1)], xmm5
; wk(1)=(24 34 25 35 26 36 27 37)
131 movdqa xmm2
,xmm6
; transpose coefficients(phase 1)
132 punpcklwd xmm6
,xmm7
; xmm6=(40 50 41 51 42 52 43 53)
133 punpckhwd xmm2
,xmm7
; xmm2=(44 54 45 55 46 56 47 57)
134 movdqa xmm5
,xmm1
; transpose coefficients(phase 1)
135 punpcklwd xmm1
,xmm3
; xmm1=(60 70 61 71 62 72 63 73)
136 punpckhwd xmm5
,xmm3
; xmm5=(64 74 65 75 66 76 67 77)
138 movdqa xmm7
,xmm6
; transpose coefficients(phase 2)
139 punpckldq xmm6
,xmm1
; xmm6=(40 50 60 70 41 51 61 71)
140 punpckhdq xmm7
,xmm1
; xmm7=(42 52 62 72 43 53 63 73)
141 movdqa xmm3
,xmm2
; transpose coefficients(phase 2)
142 punpckldq xmm2
,xmm5
; xmm2=(44 54 64 74 45 55 65 75)
143 punpckhdq xmm3
,xmm5
; xmm3=(46 56 66 76 47 57 67 77)
145 movdqa xmm1
, XMMWORD
[wk
(0)] ; xmm1=(20 30 21 31 22 32 23 33)
146 movdqa xmm5
, XMMWORD
[wk
(1)] ; xmm5=(24 34 25 35 26 36 27 37)
147 movdqa XMMWORD
[wk
(0)], xmm7
; wk(0)=(42 52 62 72 43 53 63 73)
148 movdqa XMMWORD
[wk
(1)], xmm2
; wk(1)=(44 54 64 74 45 55 65 75)
150 movdqa xmm7
,xmm0
; transpose coefficients(phase 2)
151 punpckldq xmm0
,xmm1
; xmm0=(00 10 20 30 01 11 21 31)
152 punpckhdq xmm7
,xmm1
; xmm7=(02 12 22 32 03 13 23 33)
153 movdqa xmm2
,xmm4
; transpose coefficients(phase 2)
154 punpckldq xmm4
,xmm5
; xmm4=(04 14 24 34 05 15 25 35)
155 punpckhdq xmm2
,xmm5
; xmm2=(06 16 26 36 07 17 27 37)
157 movdqa xmm1
,xmm0
; transpose coefficients(phase 3)
158 punpcklqdq xmm0
,xmm6
; xmm0=(00 10 20 30 40 50 60 70)=data0
159 punpckhqdq xmm1
,xmm6
; xmm1=(01 11 21 31 41 51 61 71)=data1
160 movdqa xmm5
,xmm2
; transpose coefficients(phase 3)
161 punpcklqdq xmm2
,xmm3
; xmm2=(06 16 26 36 46 56 66 76)=data6
162 punpckhqdq xmm5
,xmm3
; xmm5=(07 17 27 37 47 57 67 77)=data7
166 psubw xmm1
,xmm2
; xmm1=data1-data6=tmp6
167 psubw xmm0
,xmm5
; xmm0=data0-data7=tmp7
168 paddw xmm6
,xmm2
; xmm6=data1+data6=tmp1
169 paddw xmm3
,xmm5
; xmm3=data0+data7=tmp0
171 movdqa xmm2
, XMMWORD
[wk
(0)] ; xmm2=(42 52 62 72 43 53 63 73)
172 movdqa xmm5
, XMMWORD
[wk
(1)] ; xmm5=(44 54 64 74 45 55 65 75)
173 movdqa XMMWORD
[wk
(0)], xmm1
; wk(0)=tmp6
174 movdqa XMMWORD
[wk
(1)], xmm0
; wk(1)=tmp7
176 movdqa xmm1
,xmm7
; transpose coefficients(phase 3)
177 punpcklqdq xmm7
,xmm2
; xmm7=(02 12 22 32 42 52 62 72)=data2
178 punpckhqdq xmm1
,xmm2
; xmm1=(03 13 23 33 43 53 63 73)=data3
179 movdqa xmm0
,xmm4
; transpose coefficients(phase 3)
180 punpcklqdq xmm4
,xmm5
; xmm4=(04 14 24 34 44 54 64 74)=data4
181 punpckhqdq xmm0
,xmm5
; xmm0=(05 15 25 35 45 55 65 75)=data5
185 paddw xmm1
,xmm4
; xmm1=data3+data4=tmp3
186 paddw xmm7
,xmm0
; xmm7=data2+data5=tmp2
187 psubw xmm2
,xmm4
; xmm2=data3-data4=tmp4
188 psubw xmm5
,xmm0
; xmm5=data2-data5=tmp5
194 psubw xmm3
,xmm1
; xmm3=tmp13
195 psubw xmm6
,xmm7
; xmm6=tmp12
196 paddw xmm4
,xmm1
; xmm4=tmp10
197 paddw xmm0
,xmm7
; xmm0=tmp11
200 psllw xmm6
,PRE_MULTIPLY_SCALE_BITS
201 pmulhw xmm6
,[GOTOFF
(ebx,PW_F0707
)] ; xmm6=z1
205 psubw xmm4
,xmm0
; xmm4=data4
206 psubw xmm3
,xmm6
; xmm3=data6
207 paddw xmm1
,xmm0
; xmm1=data0
208 paddw xmm7
,xmm6
; xmm7=data2
210 movdqa xmm0
, XMMWORD
[wk
(0)] ; xmm0=tmp6
211 movdqa xmm6
, XMMWORD
[wk
(1)] ; xmm6=tmp7
212 movdqa XMMWORD
[wk
(0)], xmm4
; wk(0)=data4
213 movdqa XMMWORD
[wk
(1)], xmm3
; wk(1)=data6
217 paddw xmm2
,xmm5
; xmm2=tmp10
218 paddw xmm5
,xmm0
; xmm5=tmp11
219 paddw xmm0
,xmm6
; xmm0=tmp12, xmm6=tmp7
221 psllw xmm2
,PRE_MULTIPLY_SCALE_BITS
222 psllw xmm0
,PRE_MULTIPLY_SCALE_BITS
224 psllw xmm5
,PRE_MULTIPLY_SCALE_BITS
225 pmulhw xmm5
,[GOTOFF
(ebx,PW_F0707
)] ; xmm5=z3
227 movdqa xmm4
,xmm2
; xmm4=tmp10
229 pmulhw xmm2
,[GOTOFF
(ebx,PW_F0382
)] ; xmm2=z5
230 pmulhw xmm4
,[GOTOFF
(ebx,PW_F0541
)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
231 pmulhw xmm0
,[GOTOFF
(ebx,PW_F1306
)] ; xmm0=MULTIPLY(tmp12,FIX_1_306562)
232 paddw xmm4
,xmm2
; xmm4=z2
233 paddw xmm0
,xmm2
; xmm0=z4
236 psubw xmm6
,xmm5
; xmm6=z13
237 paddw xmm3
,xmm5
; xmm3=z11
241 psubw xmm6
,xmm4
; xmm6=data3
242 psubw xmm3
,xmm0
; xmm3=data7
243 paddw xmm2
,xmm4
; xmm2=data5
244 paddw xmm5
,xmm0
; xmm5=data1
246 ; ---- Pass 2: process columns.
248 ; mov edx, POINTER [data(eax)] ; (DCTELEM *)
250 ; xmm1=(00 10 20 30 40 50 60 70), xmm7=(02 12 22 32 42 52 62 72)
251 ; xmm5=(01 11 21 31 41 51 61 71), xmm6=(03 13 23 33 43 53 63 73)
253 movdqa xmm4
,xmm1
; transpose coefficients(phase 1)
254 punpcklwd xmm1
,xmm5
; xmm1=(00 01 10 11 20 21 30 31)
255 punpckhwd xmm4
,xmm5
; xmm4=(40 41 50 51 60 61 70 71)
256 movdqa xmm0
,xmm7
; transpose coefficients(phase 1)
257 punpcklwd xmm7
,xmm6
; xmm7=(02 03 12 13 22 23 32 33)
258 punpckhwd xmm0
,xmm6
; xmm0=(42 43 52 53 62 63 72 73)
260 movdqa xmm5
, XMMWORD
[wk
(0)] ; xmm5=col4
261 movdqa xmm6
, XMMWORD
[wk
(1)] ; xmm6=col6
263 ; xmm5=(04 14 24 34 44 54 64 74), xmm6=(06 16 26 36 46 56 66 76)
264 ; xmm2=(05 15 25 35 45 55 65 75), xmm3=(07 17 27 37 47 57 67 77)
266 movdqa XMMWORD
[wk
(0)], xmm7
; wk(0)=(02 03 12 13 22 23 32 33)
267 movdqa XMMWORD
[wk
(1)], xmm0
; wk(1)=(42 43 52 53 62 63 72 73)
269 movdqa xmm7
,xmm5
; transpose coefficients(phase 1)
270 punpcklwd xmm5
,xmm2
; xmm5=(04 05 14 15 24 25 34 35)
271 punpckhwd xmm7
,xmm2
; xmm7=(44 45 54 55 64 65 74 75)
272 movdqa xmm0
,xmm6
; transpose coefficients(phase 1)
273 punpcklwd xmm6
,xmm3
; xmm6=(06 07 16 17 26 27 36 37)
274 punpckhwd xmm0
,xmm3
; xmm0=(46 47 56 57 66 67 76 77)
276 movdqa xmm2
,xmm5
; transpose coefficients(phase 2)
277 punpckldq xmm5
,xmm6
; xmm5=(04 05 06 07 14 15 16 17)
278 punpckhdq xmm2
,xmm6
; xmm2=(24 25 26 27 34 35 36 37)
279 movdqa xmm3
,xmm7
; transpose coefficients(phase 2)
280 punpckldq xmm7
,xmm0
; xmm7=(44 45 46 47 54 55 56 57)
281 punpckhdq xmm3
,xmm0
; xmm3=(64 65 66 67 74 75 76 77)
283 movdqa xmm6
, XMMWORD
[wk
(0)] ; xmm6=(02 03 12 13 22 23 32 33)
284 movdqa xmm0
, XMMWORD
[wk
(1)] ; xmm0=(42 43 52 53 62 63 72 73)
285 movdqa XMMWORD
[wk
(0)], xmm2
; wk(0)=(24 25 26 27 34 35 36 37)
286 movdqa XMMWORD
[wk
(1)], xmm7
; wk(1)=(44 45 46 47 54 55 56 57)
288 movdqa xmm2
,xmm1
; transpose coefficients(phase 2)
289 punpckldq xmm1
,xmm6
; xmm1=(00 01 02 03 10 11 12 13)
290 punpckhdq xmm2
,xmm6
; xmm2=(20 21 22 23 30 31 32 33)
291 movdqa xmm7
,xmm4
; transpose coefficients(phase 2)
292 punpckldq xmm4
,xmm0
; xmm4=(40 41 42 43 50 51 52 53)
293 punpckhdq xmm7
,xmm0
; xmm7=(60 61 62 63 70 71 72 73)
295 movdqa xmm6
,xmm1
; transpose coefficients(phase 3)
296 punpcklqdq xmm1
,xmm5
; xmm1=(00 01 02 03 04 05 06 07)=data0
297 punpckhqdq xmm6
,xmm5
; xmm6=(10 11 12 13 14 15 16 17)=data1
298 movdqa xmm0
,xmm7
; transpose coefficients(phase 3)
299 punpcklqdq xmm7
,xmm3
; xmm7=(60 61 62 63 64 65 66 67)=data6
300 punpckhqdq xmm0
,xmm3
; xmm0=(70 71 72 73 74 75 76 77)=data7
304 psubw xmm6
,xmm7
; xmm6=data1-data6=tmp6
305 psubw xmm1
,xmm0
; xmm1=data0-data7=tmp7
306 paddw xmm5
,xmm7
; xmm5=data1+data6=tmp1
307 paddw xmm3
,xmm0
; xmm3=data0+data7=tmp0
309 movdqa xmm7
, XMMWORD
[wk
(0)] ; xmm7=(24 25 26 27 34 35 36 37)
310 movdqa xmm0
, XMMWORD
[wk
(1)] ; xmm0=(44 45 46 47 54 55 56 57)
311 movdqa XMMWORD
[wk
(0)], xmm6
; wk(0)=tmp6
312 movdqa XMMWORD
[wk
(1)], xmm1
; wk(1)=tmp7
314 movdqa xmm6
,xmm2
; transpose coefficients(phase 3)
315 punpcklqdq xmm2
,xmm7
; xmm2=(20 21 22 23 24 25 26 27)=data2
316 punpckhqdq xmm6
,xmm7
; xmm6=(30 31 32 33 34 35 36 37)=data3
317 movdqa xmm1
,xmm4
; transpose coefficients(phase 3)
318 punpcklqdq xmm4
,xmm0
; xmm4=(40 41 42 43 44 45 46 47)=data4
319 punpckhqdq xmm1
,xmm0
; xmm1=(50 51 52 53 54 55 56 57)=data5
323 paddw xmm6
,xmm4
; xmm6=data3+data4=tmp3
324 paddw xmm2
,xmm1
; xmm2=data2+data5=tmp2
325 psubw xmm7
,xmm4
; xmm7=data3-data4=tmp4
326 psubw xmm0
,xmm1
; xmm0=data2-data5=tmp5
332 psubw xmm3
,xmm6
; xmm3=tmp13
333 psubw xmm5
,xmm2
; xmm5=tmp12
334 paddw xmm4
,xmm6
; xmm4=tmp10
335 paddw xmm1
,xmm2
; xmm1=tmp11
338 psllw xmm5
,PRE_MULTIPLY_SCALE_BITS
339 pmulhw xmm5
,[GOTOFF
(ebx,PW_F0707
)] ; xmm5=z1
343 psubw xmm4
,xmm1
; xmm4=data4
344 psubw xmm3
,xmm5
; xmm3=data6
345 paddw xmm6
,xmm1
; xmm6=data0
346 paddw xmm2
,xmm5
; xmm2=data2
348 movdqa XMMWORD
[XMMBLOCK
(4,0,edx,SIZEOF_DCTELEM
)], xmm4
349 movdqa XMMWORD
[XMMBLOCK
(6,0,edx,SIZEOF_DCTELEM
)], xmm3
350 movdqa XMMWORD
[XMMBLOCK
(0,0,edx,SIZEOF_DCTELEM
)], xmm6
351 movdqa XMMWORD
[XMMBLOCK
(2,0,edx,SIZEOF_DCTELEM
)], xmm2
355 movdqa xmm1
, XMMWORD
[wk
(0)] ; xmm1=tmp6
356 movdqa xmm5
, XMMWORD
[wk
(1)] ; xmm5=tmp7
358 paddw xmm7
,xmm0
; xmm7=tmp10
359 paddw xmm0
,xmm1
; xmm0=tmp11
360 paddw xmm1
,xmm5
; xmm1=tmp12, xmm5=tmp7
362 psllw xmm7
,PRE_MULTIPLY_SCALE_BITS
363 psllw xmm1
,PRE_MULTIPLY_SCALE_BITS
365 psllw xmm0
,PRE_MULTIPLY_SCALE_BITS
366 pmulhw xmm0
,[GOTOFF
(ebx,PW_F0707
)] ; xmm0=z3
368 movdqa xmm4
,xmm7
; xmm4=tmp10
370 pmulhw xmm7
,[GOTOFF
(ebx,PW_F0382
)] ; xmm7=z5
371 pmulhw xmm4
,[GOTOFF
(ebx,PW_F0541
)] ; xmm4=MULTIPLY(tmp10,FIX_0_541196)
372 pmulhw xmm1
,[GOTOFF
(ebx,PW_F1306
)] ; xmm1=MULTIPLY(tmp12,FIX_1_306562)
373 paddw xmm4
,xmm7
; xmm4=z2
374 paddw xmm1
,xmm7
; xmm1=z4
377 psubw xmm5
,xmm0
; xmm5=z13
378 paddw xmm3
,xmm0
; xmm3=z11
382 psubw xmm5
,xmm4
; xmm5=data3
383 psubw xmm3
,xmm1
; xmm3=data7
384 paddw xmm6
,xmm4
; xmm6=data5
385 paddw xmm2
,xmm1
; xmm2=data1
387 movdqa XMMWORD
[XMMBLOCK
(3,0,edx,SIZEOF_DCTELEM
)], xmm5
388 movdqa XMMWORD
[XMMBLOCK
(7,0,edx,SIZEOF_DCTELEM
)], xmm3
389 movdqa XMMWORD
[XMMBLOCK
(5,0,edx,SIZEOF_DCTELEM
)], xmm6
390 movdqa XMMWORD
[XMMBLOCK
(1,0,edx,SIZEOF_DCTELEM
)], xmm2
394 ; pop edx ; need not be preserved
397 mov esp,ebp ; esp <- aligned ebp
398 pop esp ; esp <- original ebp
402 ; For some reason, the OS X linker does not honor the request to align the
403 ; segment unless we do this.