PageRenderTime 77ms CodeModel.GetById 15ms app.highlight 55ms RepoModel.GetById 2ms app.codeStats 0ms

/media/libjpeg/simd/jfss2int-64.asm

http://github.com/zpao/v8monkey
Assembly | 622 lines | 408 code | 102 blank | 112 comment | 1 complexity | 5a2264f06e07bdc6f4c6d5897a384e2d MD5 | raw file
  1;
  2; jfss2int-64.asm - accurate integer FDCT (64-bit SSE2)
  3;
  4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
  5; Copyright 2009 D. R. Commander
  6;
  7; Based on
  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
 11;
 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
 17;
 18; This file contains a slow-but-accurate integer implementation of the
 19; forward DCT (Discrete Cosine Transform). The following code is based
 20; directly on the IJG's original jfdctint.c; see the jfdctint.c for
 21; more details.
 22;
 23; [TAB8]
 24
 25%include "jsimdext.inc"
 26%include "jdct.inc"
 27
 28; --------------------------------------------------------------------------
 29
 30%define CONST_BITS	13
 31%define PASS1_BITS	2
 32
 33%define DESCALE_P1	(CONST_BITS-PASS1_BITS)
 34%define DESCALE_P2	(CONST_BITS+PASS1_BITS)
 35
 36%if CONST_BITS == 13
 37F_0_298	equ	 2446		; FIX(0.298631336)
 38F_0_390	equ	 3196		; FIX(0.390180644)
 39F_0_541	equ	 4433		; FIX(0.541196100)
 40F_0_765	equ	 6270		; FIX(0.765366865)
 41F_0_899	equ	 7373		; FIX(0.899976223)
 42F_1_175	equ	 9633		; FIX(1.175875602)
 43F_1_501	equ	12299		; FIX(1.501321110)
 44F_1_847	equ	15137		; FIX(1.847759065)
 45F_1_961	equ	16069		; FIX(1.961570560)
 46F_2_053	equ	16819		; FIX(2.053119869)
 47F_2_562	equ	20995		; FIX(2.562915447)
 48F_3_072	equ	25172		; FIX(3.072711026)
 49%else
 50; NASM cannot do compile-time arithmetic on floating-point constants.
 51%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
 52F_0_298	equ	DESCALE( 320652955,30-CONST_BITS)	; FIX(0.298631336)
 53F_0_390	equ	DESCALE( 418953276,30-CONST_BITS)	; FIX(0.390180644)
 54F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
 55F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
 56F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
 57F_1_175	equ	DESCALE(1262586813,30-CONST_BITS)	; FIX(1.175875602)
 58F_1_501	equ	DESCALE(1612031267,30-CONST_BITS)	; FIX(1.501321110)
 59F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
 60F_1_961	equ	DESCALE(2106220350,30-CONST_BITS)	; FIX(1.961570560)
 61F_2_053	equ	DESCALE(2204520673,30-CONST_BITS)	; FIX(2.053119869)
 62F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
 63F_3_072	equ	DESCALE(3299298341,30-CONST_BITS)	; FIX(3.072711026)
 64%endif
 65
 66; --------------------------------------------------------------------------
 67	SECTION	SEG_CONST
 68
 69	alignz	16
 70	global	EXTN(jconst_fdct_islow_sse2)
 71
 72EXTN(jconst_fdct_islow_sse2):
 73
 74PW_F130_F054	times 4 dw  (F_0_541+F_0_765), F_0_541
 75PW_F054_MF130	times 4 dw  F_0_541, (F_0_541-F_1_847)
 76PW_MF078_F117	times 4 dw  (F_1_175-F_1_961), F_1_175
 77PW_F117_F078	times 4 dw  F_1_175, (F_1_175-F_0_390)
 78PW_MF060_MF089	times 4 dw  (F_0_298-F_0_899),-F_0_899
 79PW_MF089_F060	times 4 dw -F_0_899, (F_1_501-F_0_899)
 80PW_MF050_MF256	times 4 dw  (F_2_053-F_2_562),-F_2_562
 81PW_MF256_F050	times 4 dw -F_2_562, (F_3_072-F_2_562)
 82PD_DESCALE_P1	times 4 dd  1 << (DESCALE_P1-1)
 83PD_DESCALE_P2	times 4 dd  1 << (DESCALE_P2-1)
 84PW_DESCALE_P2X	times 8 dw  1 << (PASS1_BITS-1)
 85
 86	alignz	16
 87
 88; --------------------------------------------------------------------------
 89	SECTION	SEG_TEXT
 90	BITS	64
 91;
 92; Perform the forward DCT on one block of samples.
 93;
 94; GLOBAL(void)
 95; jsimd_fdct_islow_sse2 (DCTELEM * data)
 96;
 97
 98; r10 = DCTELEM * data
 99
100%define wk(i)		rbp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
101%define WK_NUM		6
102
103	align	16
104	global	EXTN(jsimd_fdct_islow_sse2)
105
106EXTN(jsimd_fdct_islow_sse2):
107	push	rbp
108	mov	rax,rsp				; rax = original rbp
109	sub	rsp, byte 4
110	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
111	mov	[rsp],rax
112	mov	rbp,rsp				; rbp = aligned rbp
113	lea	rsp, [wk(0)]
114	collect_args
115
116	; ---- Pass 1: process rows.
117
118	mov	rdx, r10	; (DCTELEM *)
119
120	movdqa	xmm0, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)]
121	movdqa	xmm1, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)]
122	movdqa	xmm2, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)]
123	movdqa	xmm3, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)]
124
125	; xmm0=(00 01 02 03 04 05 06 07), xmm2=(20 21 22 23 24 25 26 27)
126	; xmm1=(10 11 12 13 14 15 16 17), xmm3=(30 31 32 33 34 35 36 37)
127
128	movdqa    xmm4,xmm0		; transpose coefficients(phase 1)
129	punpcklwd xmm0,xmm1		; xmm0=(00 10 01 11 02 12 03 13)
130	punpckhwd xmm4,xmm1		; xmm4=(04 14 05 15 06 16 07 17)
131	movdqa    xmm5,xmm2		; transpose coefficients(phase 1)
132	punpcklwd xmm2,xmm3		; xmm2=(20 30 21 31 22 32 23 33)
133	punpckhwd xmm5,xmm3		; xmm5=(24 34 25 35 26 36 27 37)
134
135	movdqa	xmm6, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)]
136	movdqa	xmm7, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)]
137	movdqa	xmm1, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)]
138	movdqa	xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)]
139
140	; xmm6=( 4 12 20 28 36 44 52 60), xmm1=( 6 14 22 30 38 46 54 62)
141	; xmm7=( 5 13 21 29 37 45 53 61), xmm3=( 7 15 23 31 39 47 55 63)
142
143	movdqa	XMMWORD [wk(0)], xmm2	; wk(0)=(20 30 21 31 22 32 23 33)
144	movdqa	XMMWORD [wk(1)], xmm5	; wk(1)=(24 34 25 35 26 36 27 37)
145
146	movdqa    xmm2,xmm6		; transpose coefficients(phase 1)
147	punpcklwd xmm6,xmm7		; xmm6=(40 50 41 51 42 52 43 53)
148	punpckhwd xmm2,xmm7		; xmm2=(44 54 45 55 46 56 47 57)
149	movdqa    xmm5,xmm1		; transpose coefficients(phase 1)
150	punpcklwd xmm1,xmm3		; xmm1=(60 70 61 71 62 72 63 73)
151	punpckhwd xmm5,xmm3		; xmm5=(64 74 65 75 66 76 67 77)
152
153	movdqa    xmm7,xmm6		; transpose coefficients(phase 2)
154	punpckldq xmm6,xmm1		; xmm6=(40 50 60 70 41 51 61 71)
155	punpckhdq xmm7,xmm1		; xmm7=(42 52 62 72 43 53 63 73)
156	movdqa    xmm3,xmm2		; transpose coefficients(phase 2)
157	punpckldq xmm2,xmm5		; xmm2=(44 54 64 74 45 55 65 75)
158	punpckhdq xmm3,xmm5		; xmm3=(46 56 66 76 47 57 67 77)
159
160	movdqa	xmm1, XMMWORD [wk(0)]	; xmm1=(20 30 21 31 22 32 23 33)
161	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=(24 34 25 35 26 36 27 37)
162	movdqa	XMMWORD [wk(2)], xmm7	; wk(2)=(42 52 62 72 43 53 63 73)
163	movdqa	XMMWORD [wk(3)], xmm2	; wk(3)=(44 54 64 74 45 55 65 75)
164
165	movdqa    xmm7,xmm0		; transpose coefficients(phase 2)
166	punpckldq xmm0,xmm1		; xmm0=(00 10 20 30 01 11 21 31)
167	punpckhdq xmm7,xmm1		; xmm7=(02 12 22 32 03 13 23 33)
168	movdqa    xmm2,xmm4		; transpose coefficients(phase 2)
169	punpckldq xmm4,xmm5		; xmm4=(04 14 24 34 05 15 25 35)
170	punpckhdq xmm2,xmm5		; xmm2=(06 16 26 36 07 17 27 37)
171
172	movdqa     xmm1,xmm0		; transpose coefficients(phase 3)
173	punpcklqdq xmm0,xmm6		; xmm0=(00 10 20 30 40 50 60 70)=data0
174	punpckhqdq xmm1,xmm6		; xmm1=(01 11 21 31 41 51 61 71)=data1
175	movdqa     xmm5,xmm2		; transpose coefficients(phase 3)
176	punpcklqdq xmm2,xmm3		; xmm2=(06 16 26 36 46 56 66 76)=data6
177	punpckhqdq xmm5,xmm3		; xmm5=(07 17 27 37 47 57 67 77)=data7
178
179	movdqa	xmm6,xmm1
180	movdqa	xmm3,xmm0
181	psubw	xmm1,xmm2		; xmm1=data1-data6=tmp6
182	psubw	xmm0,xmm5		; xmm0=data0-data7=tmp7
183	paddw	xmm6,xmm2		; xmm6=data1+data6=tmp1
184	paddw	xmm3,xmm5		; xmm3=data0+data7=tmp0
185
186	movdqa	xmm2, XMMWORD [wk(2)]	; xmm2=(42 52 62 72 43 53 63 73)
187	movdqa	xmm5, XMMWORD [wk(3)]	; xmm5=(44 54 64 74 45 55 65 75)
188	movdqa	XMMWORD [wk(0)], xmm1	; wk(0)=tmp6
189	movdqa	XMMWORD [wk(1)], xmm0	; wk(1)=tmp7
190
191	movdqa     xmm1,xmm7		; transpose coefficients(phase 3)
192	punpcklqdq xmm7,xmm2		; xmm7=(02 12 22 32 42 52 62 72)=data2
193	punpckhqdq xmm1,xmm2		; xmm1=(03 13 23 33 43 53 63 73)=data3
194	movdqa     xmm0,xmm4		; transpose coefficients(phase 3)
195	punpcklqdq xmm4,xmm5		; xmm4=(04 14 24 34 44 54 64 74)=data4
196	punpckhqdq xmm0,xmm5		; xmm0=(05 15 25 35 45 55 65 75)=data5
197
198	movdqa	xmm2,xmm1
199	movdqa	xmm5,xmm7
200	paddw	xmm1,xmm4		; xmm1=data3+data4=tmp3
201	paddw	xmm7,xmm0		; xmm7=data2+data5=tmp2
202	psubw	xmm2,xmm4		; xmm2=data3-data4=tmp4
203	psubw	xmm5,xmm0		; xmm5=data2-data5=tmp5
204
205	; -- Even part
206
207	movdqa	xmm4,xmm3
208	movdqa	xmm0,xmm6
209	paddw	xmm3,xmm1		; xmm3=tmp10
210	paddw	xmm6,xmm7		; xmm6=tmp11
211	psubw	xmm4,xmm1		; xmm4=tmp13
212	psubw	xmm0,xmm7		; xmm0=tmp12
213
214	movdqa	xmm1,xmm3
215	paddw	xmm3,xmm6		; xmm3=tmp10+tmp11
216	psubw	xmm1,xmm6		; xmm1=tmp10-tmp11
217
218	psllw	xmm3,PASS1_BITS		; xmm3=data0
219	psllw	xmm1,PASS1_BITS		; xmm1=data4
220
221	movdqa	XMMWORD [wk(2)], xmm3	; wk(2)=data0
222	movdqa	XMMWORD [wk(3)], xmm1	; wk(3)=data4
223
224	; (Original)
225	; z1 = (tmp12 + tmp13) * 0.541196100;
226	; data2 = z1 + tmp13 * 0.765366865;
227	; data6 = z1 + tmp12 * -1.847759065;
228	;
229	; (This implementation)
230	; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
231	; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
232
233	movdqa    xmm7,xmm4		; xmm4=tmp13
234	movdqa    xmm6,xmm4
235	punpcklwd xmm7,xmm0		; xmm0=tmp12
236	punpckhwd xmm6,xmm0
237	movdqa    xmm4,xmm7
238	movdqa    xmm0,xmm6
239	pmaddwd   xmm7,[rel PW_F130_F054]	; xmm7=data2L
240	pmaddwd   xmm6,[rel PW_F130_F054]	; xmm6=data2H
241	pmaddwd   xmm4,[rel PW_F054_MF130]	; xmm4=data6L
242	pmaddwd   xmm0,[rel PW_F054_MF130]	; xmm0=data6H
243
244	paddd	xmm7,[rel PD_DESCALE_P1]
245	paddd	xmm6,[rel PD_DESCALE_P1]
246	psrad	xmm7,DESCALE_P1
247	psrad	xmm6,DESCALE_P1
248	paddd	xmm4,[rel PD_DESCALE_P1]
249	paddd	xmm0,[rel PD_DESCALE_P1]
250	psrad	xmm4,DESCALE_P1
251	psrad	xmm0,DESCALE_P1
252
253	packssdw  xmm7,xmm6		; xmm7=data2
254	packssdw  xmm4,xmm0		; xmm4=data6
255
256	movdqa	XMMWORD [wk(4)], xmm7	; wk(4)=data2
257	movdqa	XMMWORD [wk(5)], xmm4	; wk(5)=data6
258
259	; -- Odd part
260
261	movdqa	xmm3, XMMWORD [wk(0)]	; xmm3=tmp6
262	movdqa	xmm1, XMMWORD [wk(1)]	; xmm1=tmp7
263
264	movdqa	xmm6,xmm2		; xmm2=tmp4
265	movdqa	xmm0,xmm5		; xmm5=tmp5
266	paddw	xmm6,xmm3		; xmm6=z3
267	paddw	xmm0,xmm1		; xmm0=z4
268
269	; (Original)
270	; z5 = (z3 + z4) * 1.175875602;
271	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
272	; z3 += z5;  z4 += z5;
273	;
274	; (This implementation)
275	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
276	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
277
278	movdqa    xmm7,xmm6
279	movdqa    xmm4,xmm6
280	punpcklwd xmm7,xmm0
281	punpckhwd xmm4,xmm0
282	movdqa    xmm6,xmm7
283	movdqa    xmm0,xmm4
284	pmaddwd   xmm7,[rel PW_MF078_F117]	; xmm7=z3L
285	pmaddwd   xmm4,[rel PW_MF078_F117]	; xmm4=z3H
286	pmaddwd   xmm6,[rel PW_F117_F078]	; xmm6=z4L
287	pmaddwd   xmm0,[rel PW_F117_F078]	; xmm0=z4H
288
289	movdqa	XMMWORD [wk(0)], xmm7	; wk(0)=z3L
290	movdqa	XMMWORD [wk(1)], xmm4	; wk(1)=z3H
291
292	; (Original)
293	; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
294	; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
295	; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
296	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
297	; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
298	; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
299	;
300	; (This implementation)
301	; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
302	; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
303	; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
304	; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
305	; data7 = tmp4 + z3;  data5 = tmp5 + z4;
306	; data3 = tmp6 + z3;  data1 = tmp7 + z4;
307
308	movdqa    xmm7,xmm2
309	movdqa    xmm4,xmm2
310	punpcklwd xmm7,xmm1
311	punpckhwd xmm4,xmm1
312	movdqa    xmm2,xmm7
313	movdqa    xmm1,xmm4
314	pmaddwd   xmm7,[rel PW_MF060_MF089]	; xmm7=tmp4L
315	pmaddwd   xmm4,[rel PW_MF060_MF089]	; xmm4=tmp4H
316	pmaddwd   xmm2,[rel PW_MF089_F060]	; xmm2=tmp7L
317	pmaddwd   xmm1,[rel PW_MF089_F060]	; xmm1=tmp7H
318
319	paddd	xmm7, XMMWORD [wk(0)]	; xmm7=data7L
320	paddd	xmm4, XMMWORD [wk(1)]	; xmm4=data7H
321	paddd	xmm2,xmm6		; xmm2=data1L
322	paddd	xmm1,xmm0		; xmm1=data1H
323
324	paddd	xmm7,[rel PD_DESCALE_P1]
325	paddd	xmm4,[rel PD_DESCALE_P1]
326	psrad	xmm7,DESCALE_P1
327	psrad	xmm4,DESCALE_P1
328	paddd	xmm2,[rel PD_DESCALE_P1]
329	paddd	xmm1,[rel PD_DESCALE_P1]
330	psrad	xmm2,DESCALE_P1
331	psrad	xmm1,DESCALE_P1
332
333	packssdw  xmm7,xmm4		; xmm7=data7
334	packssdw  xmm2,xmm1		; xmm2=data1
335
336	movdqa    xmm4,xmm5
337	movdqa    xmm1,xmm5
338	punpcklwd xmm4,xmm3
339	punpckhwd xmm1,xmm3
340	movdqa    xmm5,xmm4
341	movdqa    xmm3,xmm1
342	pmaddwd   xmm4,[rel PW_MF050_MF256]	; xmm4=tmp5L
343	pmaddwd   xmm1,[rel PW_MF050_MF256]	; xmm1=tmp5H
344	pmaddwd   xmm5,[rel PW_MF256_F050]	; xmm5=tmp6L
345	pmaddwd   xmm3,[rel PW_MF256_F050]	; xmm3=tmp6H
346
347	paddd	xmm4,xmm6		; xmm4=data5L
348	paddd	xmm1,xmm0		; xmm1=data5H
349	paddd	xmm5, XMMWORD [wk(0)]	; xmm5=data3L
350	paddd	xmm3, XMMWORD [wk(1)]	; xmm3=data3H
351
352	paddd	xmm4,[rel PD_DESCALE_P1]
353	paddd	xmm1,[rel PD_DESCALE_P1]
354	psrad	xmm4,DESCALE_P1
355	psrad	xmm1,DESCALE_P1
356	paddd	xmm5,[rel PD_DESCALE_P1]
357	paddd	xmm3,[rel PD_DESCALE_P1]
358	psrad	xmm5,DESCALE_P1
359	psrad	xmm3,DESCALE_P1
360
361	packssdw  xmm4,xmm1		; xmm4=data5
362	packssdw  xmm5,xmm3		; xmm5=data3
363
364	; ---- Pass 2: process columns.
365
366	movdqa	xmm6, XMMWORD [wk(2)]	; xmm6=col0
367	movdqa	xmm0, XMMWORD [wk(4)]	; xmm0=col2
368
369	; xmm6=(00 10 20 30 40 50 60 70), xmm0=(02 12 22 32 42 52 62 72)
370	; xmm2=(01 11 21 31 41 51 61 71), xmm5=(03 13 23 33 43 53 63 73)
371
372	movdqa    xmm1,xmm6		; transpose coefficients(phase 1)
373	punpcklwd xmm6,xmm2		; xmm6=(00 01 10 11 20 21 30 31)
374	punpckhwd xmm1,xmm2		; xmm1=(40 41 50 51 60 61 70 71)
375	movdqa    xmm3,xmm0		; transpose coefficients(phase 1)
376	punpcklwd xmm0,xmm5		; xmm0=(02 03 12 13 22 23 32 33)
377	punpckhwd xmm3,xmm5		; xmm3=(42 43 52 53 62 63 72 73)
378
379	movdqa	xmm2, XMMWORD [wk(3)]	; xmm2=col4
380	movdqa	xmm5, XMMWORD [wk(5)]	; xmm5=col6
381
382	; xmm2=(04 14 24 34 44 54 64 74), xmm5=(06 16 26 36 46 56 66 76)
383	; xmm4=(05 15 25 35 45 55 65 75), xmm7=(07 17 27 37 47 57 67 77)
384
385	movdqa	XMMWORD [wk(0)], xmm0	; wk(0)=(02 03 12 13 22 23 32 33)
386	movdqa	XMMWORD [wk(1)], xmm3	; wk(1)=(42 43 52 53 62 63 72 73)
387
388	movdqa    xmm0,xmm2		; transpose coefficients(phase 1)
389	punpcklwd xmm2,xmm4		; xmm2=(04 05 14 15 24 25 34 35)
390	punpckhwd xmm0,xmm4		; xmm0=(44 45 54 55 64 65 74 75)
391	movdqa    xmm3,xmm5		; transpose coefficients(phase 1)
392	punpcklwd xmm5,xmm7		; xmm5=(06 07 16 17 26 27 36 37)
393	punpckhwd xmm3,xmm7		; xmm3=(46 47 56 57 66 67 76 77)
394
395	movdqa    xmm4,xmm2		; transpose coefficients(phase 2)
396	punpckldq xmm2,xmm5		; xmm2=(04 05 06 07 14 15 16 17)
397	punpckhdq xmm4,xmm5		; xmm4=(24 25 26 27 34 35 36 37)
398	movdqa    xmm7,xmm0		; transpose coefficients(phase 2)
399	punpckldq xmm0,xmm3		; xmm0=(44 45 46 47 54 55 56 57)
400	punpckhdq xmm7,xmm3		; xmm7=(64 65 66 67 74 75 76 77)
401
402	movdqa	xmm5, XMMWORD [wk(0)]	; xmm5=(02 03 12 13 22 23 32 33)
403	movdqa	xmm3, XMMWORD [wk(1)]	; xmm3=(42 43 52 53 62 63 72 73)
404	movdqa	XMMWORD [wk(2)], xmm4	; wk(2)=(24 25 26 27 34 35 36 37)
405	movdqa	XMMWORD [wk(3)], xmm0	; wk(3)=(44 45 46 47 54 55 56 57)
406
407	movdqa    xmm4,xmm6		; transpose coefficients(phase 2)
408	punpckldq xmm6,xmm5		; xmm6=(00 01 02 03 10 11 12 13)
409	punpckhdq xmm4,xmm5		; xmm4=(20 21 22 23 30 31 32 33)
410	movdqa    xmm0,xmm1		; transpose coefficients(phase 2)
411	punpckldq xmm1,xmm3		; xmm1=(40 41 42 43 50 51 52 53)
412	punpckhdq xmm0,xmm3		; xmm0=(60 61 62 63 70 71 72 73)
413
414	movdqa     xmm5,xmm6		; transpose coefficients(phase 3)
415	punpcklqdq xmm6,xmm2		; xmm6=(00 01 02 03 04 05 06 07)=data0
416	punpckhqdq xmm5,xmm2		; xmm5=(10 11 12 13 14 15 16 17)=data1
417	movdqa     xmm3,xmm0		; transpose coefficients(phase 3)
418	punpcklqdq xmm0,xmm7		; xmm0=(60 61 62 63 64 65 66 67)=data6
419	punpckhqdq xmm3,xmm7		; xmm3=(70 71 72 73 74 75 76 77)=data7
420
421	movdqa	xmm2,xmm5
422	movdqa	xmm7,xmm6
423	psubw	xmm5,xmm0		; xmm5=data1-data6=tmp6
424	psubw	xmm6,xmm3		; xmm6=data0-data7=tmp7
425	paddw	xmm2,xmm0		; xmm2=data1+data6=tmp1
426	paddw	xmm7,xmm3		; xmm7=data0+data7=tmp0
427
428	movdqa	xmm0, XMMWORD [wk(2)]	; xmm0=(24 25 26 27 34 35 36 37)
429	movdqa	xmm3, XMMWORD [wk(3)]	; xmm3=(44 45 46 47 54 55 56 57)
430	movdqa	XMMWORD [wk(0)], xmm5	; wk(0)=tmp6
431	movdqa	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
432
433	movdqa     xmm5,xmm4		; transpose coefficients(phase 3)
434	punpcklqdq xmm4,xmm0		; xmm4=(20 21 22 23 24 25 26 27)=data2
435	punpckhqdq xmm5,xmm0		; xmm5=(30 31 32 33 34 35 36 37)=data3
436	movdqa     xmm6,xmm1		; transpose coefficients(phase 3)
437	punpcklqdq xmm1,xmm3		; xmm1=(40 41 42 43 44 45 46 47)=data4
438	punpckhqdq xmm6,xmm3		; xmm6=(50 51 52 53 54 55 56 57)=data5
439
440	movdqa	xmm0,xmm5
441	movdqa	xmm3,xmm4
442	paddw	xmm5,xmm1		; xmm5=data3+data4=tmp3
443	paddw	xmm4,xmm6		; xmm4=data2+data5=tmp2
444	psubw	xmm0,xmm1		; xmm0=data3-data4=tmp4
445	psubw	xmm3,xmm6		; xmm3=data2-data5=tmp5
446
447	; -- Even part
448
449	movdqa	xmm1,xmm7
450	movdqa	xmm6,xmm2
451	paddw	xmm7,xmm5		; xmm7=tmp10
452	paddw	xmm2,xmm4		; xmm2=tmp11
453	psubw	xmm1,xmm5		; xmm1=tmp13
454	psubw	xmm6,xmm4		; xmm6=tmp12
455
456	movdqa	xmm5,xmm7
457	paddw	xmm7,xmm2		; xmm7=tmp10+tmp11
458	psubw	xmm5,xmm2		; xmm5=tmp10-tmp11
459
460	paddw	xmm7,[rel PW_DESCALE_P2X]
461	paddw	xmm5,[rel PW_DESCALE_P2X]
462	psraw	xmm7,PASS1_BITS		; xmm7=data0
463	psraw	xmm5,PASS1_BITS		; xmm5=data4
464
465	movdqa	XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_DCTELEM)], xmm7
466	movdqa	XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_DCTELEM)], xmm5
467
468	; (Original)
469	; z1 = (tmp12 + tmp13) * 0.541196100;
470	; data2 = z1 + tmp13 * 0.765366865;
471	; data6 = z1 + tmp12 * -1.847759065;
472	;
473	; (This implementation)
474	; data2 = tmp13 * (0.541196100 + 0.765366865) + tmp12 * 0.541196100;
475	; data6 = tmp13 * 0.541196100 + tmp12 * (0.541196100 - 1.847759065);
476
477	movdqa    xmm4,xmm1		; xmm1=tmp13
478	movdqa    xmm2,xmm1
479	punpcklwd xmm4,xmm6		; xmm6=tmp12
480	punpckhwd xmm2,xmm6
481	movdqa    xmm1,xmm4
482	movdqa    xmm6,xmm2
483	pmaddwd   xmm4,[rel PW_F130_F054]	; xmm4=data2L
484	pmaddwd   xmm2,[rel PW_F130_F054]	; xmm2=data2H
485	pmaddwd   xmm1,[rel PW_F054_MF130]	; xmm1=data6L
486	pmaddwd   xmm6,[rel PW_F054_MF130]	; xmm6=data6H
487
488	paddd	xmm4,[rel PD_DESCALE_P2]
489	paddd	xmm2,[rel PD_DESCALE_P2]
490	psrad	xmm4,DESCALE_P2
491	psrad	xmm2,DESCALE_P2
492	paddd	xmm1,[rel PD_DESCALE_P2]
493	paddd	xmm6,[rel PD_DESCALE_P2]
494	psrad	xmm1,DESCALE_P2
495	psrad	xmm6,DESCALE_P2
496
497	packssdw  xmm4,xmm2		; xmm4=data2
498	packssdw  xmm1,xmm6		; xmm1=data6
499
500	movdqa	XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_DCTELEM)], xmm4
501	movdqa	XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_DCTELEM)], xmm1
502
503	; -- Odd part
504
505	movdqa	xmm7, XMMWORD [wk(0)]	; xmm7=tmp6
506	movdqa	xmm5, XMMWORD [wk(1)]	; xmm5=tmp7
507
508	movdqa	xmm2,xmm0		; xmm0=tmp4
509	movdqa	xmm6,xmm3		; xmm3=tmp5
510	paddw	xmm2,xmm7		; xmm2=z3
511	paddw	xmm6,xmm5		; xmm6=z4
512
513	; (Original)
514	; z5 = (z3 + z4) * 1.175875602;
515	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
516	; z3 += z5;  z4 += z5;
517	;
518	; (This implementation)
519	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
520	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
521
522	movdqa    xmm4,xmm2
523	movdqa    xmm1,xmm2
524	punpcklwd xmm4,xmm6
525	punpckhwd xmm1,xmm6
526	movdqa    xmm2,xmm4
527	movdqa    xmm6,xmm1
528	pmaddwd   xmm4,[rel PW_MF078_F117]	; xmm4=z3L
529	pmaddwd   xmm1,[rel PW_MF078_F117]	; xmm1=z3H
530	pmaddwd   xmm2,[rel PW_F117_F078]	; xmm2=z4L
531	pmaddwd   xmm6,[rel PW_F117_F078]	; xmm6=z4H
532
533	movdqa	XMMWORD [wk(0)], xmm4	; wk(0)=z3L
534	movdqa	XMMWORD [wk(1)], xmm1	; wk(1)=z3H
535
536	; (Original)
537	; z1 = tmp4 + tmp7;  z2 = tmp5 + tmp6;
538	; tmp4 = tmp4 * 0.298631336;  tmp5 = tmp5 * 2.053119869;
539	; tmp6 = tmp6 * 3.072711026;  tmp7 = tmp7 * 1.501321110;
540	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
541	; data7 = tmp4 + z1 + z3;  data5 = tmp5 + z2 + z4;
542	; data3 = tmp6 + z2 + z3;  data1 = tmp7 + z1 + z4;
543	;
544	; (This implementation)
545	; tmp4 = tmp4 * (0.298631336 - 0.899976223) + tmp7 * -0.899976223;
546	; tmp5 = tmp5 * (2.053119869 - 2.562915447) + tmp6 * -2.562915447;
547	; tmp6 = tmp5 * -2.562915447 + tmp6 * (3.072711026 - 2.562915447);
548	; tmp7 = tmp4 * -0.899976223 + tmp7 * (1.501321110 - 0.899976223);
549	; data7 = tmp4 + z3;  data5 = tmp5 + z4;
550	; data3 = tmp6 + z3;  data1 = tmp7 + z4;
551
552	movdqa    xmm4,xmm0
553	movdqa    xmm1,xmm0
554	punpcklwd xmm4,xmm5
555	punpckhwd xmm1,xmm5
556	movdqa    xmm0,xmm4
557	movdqa    xmm5,xmm1
558	pmaddwd   xmm4,[rel PW_MF060_MF089]	; xmm4=tmp4L
559	pmaddwd   xmm1,[rel PW_MF060_MF089]	; xmm1=tmp4H
560	pmaddwd   xmm0,[rel PW_MF089_F060]	; xmm0=tmp7L
561	pmaddwd   xmm5,[rel PW_MF089_F060]	; xmm5=tmp7H
562
563	paddd	xmm4, XMMWORD [wk(0)]	; xmm4=data7L
564	paddd	xmm1, XMMWORD [wk(1)]	; xmm1=data7H
565	paddd	xmm0,xmm2		; xmm0=data1L
566	paddd	xmm5,xmm6		; xmm5=data1H
567
568	paddd	xmm4,[rel PD_DESCALE_P2]
569	paddd	xmm1,[rel PD_DESCALE_P2]
570	psrad	xmm4,DESCALE_P2
571	psrad	xmm1,DESCALE_P2
572	paddd	xmm0,[rel PD_DESCALE_P2]
573	paddd	xmm5,[rel PD_DESCALE_P2]
574	psrad	xmm0,DESCALE_P2
575	psrad	xmm5,DESCALE_P2
576
577	packssdw  xmm4,xmm1		; xmm4=data7
578	packssdw  xmm0,xmm5		; xmm0=data1
579
580	movdqa	XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_DCTELEM)], xmm4
581	movdqa	XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_DCTELEM)], xmm0
582
583	movdqa    xmm1,xmm3
584	movdqa    xmm5,xmm3
585	punpcklwd xmm1,xmm7
586	punpckhwd xmm5,xmm7
587	movdqa    xmm3,xmm1
588	movdqa    xmm7,xmm5
589	pmaddwd   xmm1,[rel PW_MF050_MF256]	; xmm1=tmp5L
590	pmaddwd   xmm5,[rel PW_MF050_MF256]	; xmm5=tmp5H
591	pmaddwd   xmm3,[rel PW_MF256_F050]	; xmm3=tmp6L
592	pmaddwd   xmm7,[rel PW_MF256_F050]	; xmm7=tmp6H
593
594	paddd	xmm1,xmm2		; xmm1=data5L
595	paddd	xmm5,xmm6		; xmm5=data5H
596	paddd	xmm3, XMMWORD [wk(0)]	; xmm3=data3L
597	paddd	xmm7, XMMWORD [wk(1)]	; xmm7=data3H
598
599	paddd	xmm1,[rel PD_DESCALE_P2]
600	paddd	xmm5,[rel PD_DESCALE_P2]
601	psrad	xmm1,DESCALE_P2
602	psrad	xmm5,DESCALE_P2
603	paddd	xmm3,[rel PD_DESCALE_P2]
604	paddd	xmm7,[rel PD_DESCALE_P2]
605	psrad	xmm3,DESCALE_P2
606	psrad	xmm7,DESCALE_P2
607
608	packssdw  xmm1,xmm5		; xmm1=data5
609	packssdw  xmm3,xmm7		; xmm3=data3
610
611	movdqa	XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_DCTELEM)], xmm1
612	movdqa	XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_DCTELEM)], xmm3
613
614	uncollect_args
615	mov	rsp,rbp		; rsp <- aligned rbp
616	pop	rsp		; rsp <- original rbp
617	pop	rbp
618	ret
619
620; For some reason, the OS X linker does not honor the request to align the
621; segment unless we do this.
622	align	16