2 ; jfmmxfst.asm - fast integer FDCT (MMX)
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_mmx)
57 EXTN(jconst_fdct_ifast_mmx):
59 PW_F0707 times 4 dw F_0_707 << CONST_SHIFT
60 PW_F0382 times 4 dw F_0_382 << CONST_SHIFT
61 PW_F0541 times 4 dw F_0_541 << CONST_SHIFT
62 PW_F1306 times 4 dw F_1_306 << CONST_SHIFT
66 ; --------------------------------------------------------------------------
70 ; Perform the forward DCT on one block of samples.
73 ; jsimd_fdct_ifast_mmx (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_MMWORD ; mmword wk[WK_NUM]
83 global EXTN(jsimd_fdct_ifast_mmx)
85 EXTN(jsimd_fdct_ifast_mmx):
87 mov eax,esp ; eax = original ebp
89 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
91 mov ebp,esp ; ebp = aligned ebp
94 ; push ecx ; need not be preserved
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 *)
108 movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
109 movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
110 movq mm2, MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)]
111 movq mm3, MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)]
113 ; mm0=(20 21 22 23), mm2=(24 25 26 27)
114 ; mm1=(30 31 32 33), mm3=(34 35 36 37)
116 movq mm4,mm0 ; transpose coefficients(phase 1)
117 punpcklwd mm0,mm1 ; mm0=(20 30 21 31)
118 punpckhwd mm4,mm1 ; mm4=(22 32 23 33)
119 movq mm5,mm2 ; transpose coefficients(phase 1)
120 punpcklwd mm2,mm3 ; mm2=(24 34 25 35)
121 punpckhwd mm5,mm3 ; mm5=(26 36 27 37)
123 movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
124 movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
125 movq mm1, MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)]
126 movq mm3, MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)]
128 ; mm6=(00 01 02 03), mm1=(04 05 06 07)
129 ; mm7=(10 11 12 13), mm3=(14 15 16 17)
131 movq MMWORD [wk(0)], mm4 ; wk(0)=(22 32 23 33)
132 movq MMWORD [wk(1)], mm2 ; wk(1)=(24 34 25 35)
134 movq mm4,mm6 ; transpose coefficients(phase 1)
135 punpcklwd mm6,mm7 ; mm6=(00 10 01 11)
136 punpckhwd mm4,mm7 ; mm4=(02 12 03 13)
137 movq mm2,mm1 ; transpose coefficients(phase 1)
138 punpcklwd mm1,mm3 ; mm1=(04 14 05 15)
139 punpckhwd mm2,mm3 ; mm2=(06 16 07 17)
141 movq mm7,mm6 ; transpose coefficients(phase 2)
142 punpckldq mm6,mm0 ; mm6=(00 10 20 30)=data0
143 punpckhdq mm7,mm0 ; mm7=(01 11 21 31)=data1
144 movq mm3,mm2 ; transpose coefficients(phase 2)
145 punpckldq mm2,mm5 ; mm2=(06 16 26 36)=data6
146 punpckhdq mm3,mm5 ; mm3=(07 17 27 37)=data7
150 psubw mm7,mm2 ; mm7=data1-data6=tmp6
151 psubw mm6,mm3 ; mm6=data0-data7=tmp7
152 paddw mm0,mm2 ; mm0=data1+data6=tmp1
153 paddw mm5,mm3 ; mm5=data0+data7=tmp0
155 movq mm2, MMWORD [wk(0)] ; mm2=(22 32 23 33)
156 movq mm3, MMWORD [wk(1)] ; mm3=(24 34 25 35)
157 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6
158 movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7
160 movq mm7,mm4 ; transpose coefficients(phase 2)
161 punpckldq mm4,mm2 ; mm4=(02 12 22 32)=data2
162 punpckhdq mm7,mm2 ; mm7=(03 13 23 33)=data3
163 movq mm6,mm1 ; transpose coefficients(phase 2)
164 punpckldq mm1,mm3 ; mm1=(04 14 24 34)=data4
165 punpckhdq mm6,mm3 ; mm6=(05 15 25 35)=data5
169 paddw mm7,mm1 ; mm7=data3+data4=tmp3
170 paddw mm4,mm6 ; mm4=data2+data5=tmp2
171 psubw mm2,mm1 ; mm2=data3-data4=tmp4
172 psubw mm3,mm6 ; mm3=data2-data5=tmp5
178 psubw mm5,mm7 ; mm5=tmp13
179 psubw mm0,mm4 ; mm0=tmp12
180 paddw mm1,mm7 ; mm1=tmp10
181 paddw mm6,mm4 ; mm6=tmp11
184 psllw mm0,PRE_MULTIPLY_SCALE_BITS
185 pmulhw mm0,[GOTOFF(ebx,PW_F0707)] ; mm0=z1
189 psubw mm1,mm6 ; mm1=data4
190 psubw mm5,mm0 ; mm5=data6
191 paddw mm7,mm6 ; mm7=data0
192 paddw mm4,mm0 ; mm4=data2
194 movq MMWORD [MMBLOCK(0,1,edx,SIZEOF_DCTELEM)], mm1
195 movq MMWORD [MMBLOCK(2,1,edx,SIZEOF_DCTELEM)], mm5
196 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm7
197 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
201 movq mm6, MMWORD [wk(0)] ; mm6=tmp6
202 movq mm0, MMWORD [wk(1)] ; mm0=tmp7
204 paddw mm2,mm3 ; mm2=tmp10
205 paddw mm3,mm6 ; mm3=tmp11
206 paddw mm6,mm0 ; mm6=tmp12, mm0=tmp7
208 psllw mm2,PRE_MULTIPLY_SCALE_BITS
209 psllw mm6,PRE_MULTIPLY_SCALE_BITS
211 psllw mm3,PRE_MULTIPLY_SCALE_BITS
212 pmulhw mm3,[GOTOFF(ebx,PW_F0707)] ; mm3=z3
214 movq mm1,mm2 ; mm1=tmp10
216 pmulhw mm2,[GOTOFF(ebx,PW_F0382)] ; mm2=z5
217 pmulhw mm1,[GOTOFF(ebx,PW_F0541)] ; mm1=MULTIPLY(tmp10,FIX_0_54119610)
218 pmulhw mm6,[GOTOFF(ebx,PW_F1306)] ; mm6=MULTIPLY(tmp12,FIX_1_30656296)
219 paddw mm1,mm2 ; mm1=z2
220 paddw mm6,mm2 ; mm6=z4
223 psubw mm0,mm3 ; mm0=z13
224 paddw mm5,mm3 ; mm5=z11
228 psubw mm0,mm1 ; mm0=data3
229 psubw mm5,mm6 ; mm5=data7
230 paddw mm7,mm1 ; mm7=data5
231 paddw mm4,mm6 ; mm4=data1
233 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm0
234 movq MMWORD [MMBLOCK(3,1,edx,SIZEOF_DCTELEM)], mm5
235 movq MMWORD [MMBLOCK(1,1,edx,SIZEOF_DCTELEM)], mm7
236 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm4
238 add edx, byte 4*DCTSIZE*SIZEOF_DCTELEM
242 ; ---- Pass 2: process columns.
244 mov edx, POINTER [data(eax)] ; (DCTELEM *)
249 movq mm0, MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)]
250 movq mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)]
251 movq mm2, MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)]
252 movq mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)]
254 ; mm0=(02 12 22 32), mm2=(42 52 62 72)
255 ; mm1=(03 13 23 33), mm3=(43 53 63 73)
257 movq mm4,mm0 ; transpose coefficients(phase 1)
258 punpcklwd mm0,mm1 ; mm0=(02 03 12 13)
259 punpckhwd mm4,mm1 ; mm4=(22 23 32 33)
260 movq mm5,mm2 ; transpose coefficients(phase 1)
261 punpcklwd mm2,mm3 ; mm2=(42 43 52 53)
262 punpckhwd mm5,mm3 ; mm5=(62 63 72 73)
264 movq mm6, MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)]
265 movq mm7, MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)]
266 movq mm1, MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)]
267 movq mm3, MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)]
269 ; mm6=(00 10 20 30), mm1=(40 50 60 70)
270 ; mm7=(01 11 21 31), mm3=(41 51 61 71)
272 movq MMWORD [wk(0)], mm4 ; wk(0)=(22 23 32 33)
273 movq MMWORD [wk(1)], mm2 ; wk(1)=(42 43 52 53)
275 movq mm4,mm6 ; transpose coefficients(phase 1)
276 punpcklwd mm6,mm7 ; mm6=(00 01 10 11)
277 punpckhwd mm4,mm7 ; mm4=(20 21 30 31)
278 movq mm2,mm1 ; transpose coefficients(phase 1)
279 punpcklwd mm1,mm3 ; mm1=(40 41 50 51)
280 punpckhwd mm2,mm3 ; mm2=(60 61 70 71)
282 movq mm7,mm6 ; transpose coefficients(phase 2)
283 punpckldq mm6,mm0 ; mm6=(00 01 02 03)=data0
284 punpckhdq mm7,mm0 ; mm7=(10 11 12 13)=data1
285 movq mm3,mm2 ; transpose coefficients(phase 2)
286 punpckldq mm2,mm5 ; mm2=(60 61 62 63)=data6
287 punpckhdq mm3,mm5 ; mm3=(70 71 72 73)=data7
291 psubw mm7,mm2 ; mm7=data1-data6=tmp6
292 psubw mm6,mm3 ; mm6=data0-data7=tmp7
293 paddw mm0,mm2 ; mm0=data1+data6=tmp1
294 paddw mm5,mm3 ; mm5=data0+data7=tmp0
296 movq mm2, MMWORD [wk(0)] ; mm2=(22 23 32 33)
297 movq mm3, MMWORD [wk(1)] ; mm3=(42 43 52 53)
298 movq MMWORD [wk(0)], mm7 ; wk(0)=tmp6
299 movq MMWORD [wk(1)], mm6 ; wk(1)=tmp7
301 movq mm7,mm4 ; transpose coefficients(phase 2)
302 punpckldq mm4,mm2 ; mm4=(20 21 22 23)=data2
303 punpckhdq mm7,mm2 ; mm7=(30 31 32 33)=data3
304 movq mm6,mm1 ; transpose coefficients(phase 2)
305 punpckldq mm1,mm3 ; mm1=(40 41 42 43)=data4
306 punpckhdq mm6,mm3 ; mm6=(50 51 52 53)=data5
310 paddw mm7,mm1 ; mm7=data3+data4=tmp3
311 paddw mm4,mm6 ; mm4=data2+data5=tmp2
312 psubw mm2,mm1 ; mm2=data3-data4=tmp4
313 psubw mm3,mm6 ; mm3=data2-data5=tmp5
319 psubw mm5,mm7 ; mm5=tmp13
320 psubw mm0,mm4 ; mm0=tmp12
321 paddw mm1,mm7 ; mm1=tmp10
322 paddw mm6,mm4 ; mm6=tmp11
325 psllw mm0,PRE_MULTIPLY_SCALE_BITS
326 pmulhw mm0,[GOTOFF(ebx,PW_F0707)] ; mm0=z1
330 psubw mm1,mm6 ; mm1=data4
331 psubw mm5,mm0 ; mm5=data6
332 paddw mm7,mm6 ; mm7=data0
333 paddw mm4,mm0 ; mm4=data2
335 movq MMWORD [MMBLOCK(4,0,edx,SIZEOF_DCTELEM)], mm1
336 movq MMWORD [MMBLOCK(6,0,edx,SIZEOF_DCTELEM)], mm5
337 movq MMWORD [MMBLOCK(0,0,edx,SIZEOF_DCTELEM)], mm7
338 movq MMWORD [MMBLOCK(2,0,edx,SIZEOF_DCTELEM)], mm4
342 movq mm6, MMWORD [wk(0)] ; mm6=tmp6
343 movq mm0, MMWORD [wk(1)] ; mm0=tmp7
345 paddw mm2,mm3 ; mm2=tmp10
346 paddw mm3,mm6 ; mm3=tmp11
347 paddw mm6,mm0 ; mm6=tmp12, mm0=tmp7
349 psllw mm2,PRE_MULTIPLY_SCALE_BITS
350 psllw mm6,PRE_MULTIPLY_SCALE_BITS
352 psllw mm3,PRE_MULTIPLY_SCALE_BITS
353 pmulhw mm3,[GOTOFF(ebx,PW_F0707)] ; mm3=z3
355 movq mm1,mm2 ; mm1=tmp10
357 pmulhw mm2,[GOTOFF(ebx,PW_F0382)] ; mm2=z5
358 pmulhw mm1,[GOTOFF(ebx,PW_F0541)] ; mm1=MULTIPLY(tmp10,FIX_0_54119610)
359 pmulhw mm6,[GOTOFF(ebx,PW_F1306)] ; mm6=MULTIPLY(tmp12,FIX_1_30656296)
360 paddw mm1,mm2 ; mm1=z2
361 paddw mm6,mm2 ; mm6=z4
364 psubw mm0,mm3 ; mm0=z13
365 paddw mm5,mm3 ; mm5=z11
369 psubw mm0,mm1 ; mm0=data3
370 psubw mm5,mm6 ; mm5=data7
371 paddw mm7,mm1 ; mm7=data5
372 paddw mm4,mm6 ; mm4=data1
374 movq MMWORD [MMBLOCK(3,0,edx,SIZEOF_DCTELEM)], mm0
375 movq MMWORD [MMBLOCK(7,0,edx,SIZEOF_DCTELEM)], mm5
376 movq MMWORD [MMBLOCK(5,0,edx,SIZEOF_DCTELEM)], mm7
377 movq MMWORD [MMBLOCK(1,0,edx,SIZEOF_DCTELEM)], mm4
379 add edx, byte 4*SIZEOF_DCTELEM
383 emms ; empty MMX state
387 ; pop edx ; need not be preserved
388 ; pop ecx ; need not be preserved
390 mov esp,ebp ; esp <- aligned ebp
391 pop esp ; esp <- original ebp
395 ; For some reason, the OS X linker does not honor the request to align the
396 ; segment unless we do this.