PageRenderTime 39ms CodeModel.GetById 16ms app.highlight 17ms RepoModel.GetById 1ms app.codeStats 0ms

/media/libjpeg/simd/jimmxfst.asm

http://github.com/zpao/v8monkey
Assembly | 500 lines | 345 code | 87 blank | 68 comment | 2 complexity | 00f86e2bd3ebfb7fe17210e27bc3c852 MD5 | raw file
  1;
  2; jimmxfst.asm - fast integer IDCT (MMX)
  3;
  4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
  5;
  6; Based on
  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
 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 a fast, not so accurate integer implementation of
 18; the inverse DCT (Discrete Cosine Transform). The following code is
 19; based directly on the IJG's original jidctfst.c; see the jidctfst.c
 20; for more details.
 21;
 22; [TAB8]
 23
 24%include "jsimdext.inc"
 25%include "jdct.inc"
 26
 27; --------------------------------------------------------------------------
 28
 29%define CONST_BITS	8	; 14 is also OK.
 30%define PASS1_BITS	2
 31
 32%if IFAST_SCALE_BITS != PASS1_BITS
 33%error "'IFAST_SCALE_BITS' must be equal to 'PASS1_BITS'."
 34%endif
 35
 36%if CONST_BITS == 8
 37F_1_082	equ	277		; FIX(1.082392200)
 38F_1_414	equ	362		; FIX(1.414213562)
 39F_1_847	equ	473		; FIX(1.847759065)
 40F_2_613	equ	669		; FIX(2.613125930)
 41F_1_613	equ	(F_2_613 - 256)	; FIX(2.613125930) - FIX(1)
 42%else
 43; NASM cannot do compile-time arithmetic on floating-point constants.
 44%define	DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
 45F_1_082	equ	DESCALE(1162209775,30-CONST_BITS)	; FIX(1.082392200)
 46F_1_414	equ	DESCALE(1518500249,30-CONST_BITS)	; FIX(1.414213562)
 47F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
 48F_2_613	equ	DESCALE(2805822602,30-CONST_BITS)	; FIX(2.613125930)
 49F_1_613	equ	(F_2_613 - (1 << CONST_BITS))	; FIX(2.613125930) - FIX(1)
 50%endif
 51
 52; --------------------------------------------------------------------------
 53	SECTION	SEG_CONST
 54
 55; PRE_MULTIPLY_SCALE_BITS <= 2 (to avoid overflow)
 56; CONST_BITS + CONST_SHIFT + PRE_MULTIPLY_SCALE_BITS == 16 (for pmulhw)
 57
 58%define PRE_MULTIPLY_SCALE_BITS   2
 59%define CONST_SHIFT     (16 - PRE_MULTIPLY_SCALE_BITS - CONST_BITS)
 60
 61	alignz	16
 62	global	EXTN(jconst_idct_ifast_mmx)
 63
 64EXTN(jconst_idct_ifast_mmx):
 65
 66PW_F1414	times 4 dw  F_1_414 << CONST_SHIFT
 67PW_F1847	times 4 dw  F_1_847 << CONST_SHIFT
 68PW_MF1613	times 4 dw -F_1_613 << CONST_SHIFT
 69PW_F1082	times 4 dw  F_1_082 << CONST_SHIFT
 70PB_CENTERJSAMP	times 8 db  CENTERJSAMPLE
 71
 72	alignz	16
 73
 74; --------------------------------------------------------------------------
 75	SECTION	SEG_TEXT
 76	BITS	32
 77;
 78; Perform dequantization and inverse DCT on one block of coefficients.
 79;
 80; GLOBAL(void)
 81; jsimd_idct_ifast_mmx (void * dct_table, JCOEFPTR coef_block,
 82;                       JSAMPARRAY output_buf, JDIMENSION output_col)
 83;
 84
 85%define dct_table(b)	(b)+8			; jpeg_component_info * compptr
 86%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
 87%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
 88%define output_col(b)	(b)+20		; JDIMENSION output_col
 89
 90%define original_ebp	ebp+0
 91%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
 92%define WK_NUM		2
 93%define workspace	wk(0)-DCTSIZE2*SIZEOF_JCOEF
 94					; JCOEF workspace[DCTSIZE2]
 95
 96	align	16
 97	global	EXTN(jsimd_idct_ifast_mmx)
 98
 99EXTN(jsimd_idct_ifast_mmx):
100	push	ebp
101	mov	eax,esp				; eax = original ebp
102	sub	esp, byte 4
103	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
104	mov	[esp],eax
105	mov	ebp,esp				; ebp = aligned ebp
106	lea	esp, [workspace]
107	push	ebx
108;	push	ecx		; need not be preserved
109;	push	edx		; need not be preserved
110	push	esi
111	push	edi
112
113	get_GOT	ebx		; get GOT address
114
115	; ---- Pass 1: process columns from input, store into work array.
116
117;	mov	eax, [original_ebp]
118	mov	edx, POINTER [dct_table(eax)]	; quantptr
119	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
120	lea	edi, [workspace]			; JCOEF * wsptr
121	mov	ecx, DCTSIZE/4				; ctr
122	alignx	16,7
123.columnloop:
124%ifndef NO_ZERO_COLUMN_TEST_IFAST_MMX
125	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
126	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
127	jnz	short .columnDCT
128
129	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
130	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
131	por	mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
132	por	mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
133	por	mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
134	por	mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
135	por	mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
136	por	mm1,mm0
137	packsswb mm1,mm1
138	movd	eax,mm1
139	test	eax,eax
140	jnz	short .columnDCT
141
142	; -- AC terms all zero
143
144	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
145	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
146
147	movq      mm2,mm0		; mm0=in0=(00 01 02 03)
148	punpcklwd mm0,mm0		; mm0=(00 00 01 01)
149	punpckhwd mm2,mm2		; mm2=(02 02 03 03)
150
151	movq      mm1,mm0
152	punpckldq mm0,mm0		; mm0=(00 00 00 00)
153	punpckhdq mm1,mm1		; mm1=(01 01 01 01)
154	movq      mm3,mm2
155	punpckldq mm2,mm2		; mm2=(02 02 02 02)
156	punpckhdq mm3,mm3		; mm3=(03 03 03 03)
157
158	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
159	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
160	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
161	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm1
162	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
163	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
164	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
165	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm3
166	jmp	near .nextcolumn
167	alignx	16,7
168%endif
169.columnDCT:
170
171	; -- Even part
172
173	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
174	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
175	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_IFAST_MULT_TYPE)]
176	pmullw	mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_IFAST_MULT_TYPE)]
177	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
178	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
179	pmullw	mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_IFAST_MULT_TYPE)]
180	pmullw	mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_IFAST_MULT_TYPE)]
181
182	movq	mm4,mm0
183	movq	mm5,mm1
184	psubw	mm0,mm2			; mm0=tmp11
185	psubw	mm1,mm3
186	paddw	mm4,mm2			; mm4=tmp10
187	paddw	mm5,mm3			; mm5=tmp13
188
189	psllw	mm1,PRE_MULTIPLY_SCALE_BITS
190	pmulhw	mm1,[GOTOFF(ebx,PW_F1414)]
191	psubw	mm1,mm5			; mm1=tmp12
192
193	movq	mm6,mm4
194	movq	mm7,mm0
195	psubw	mm4,mm5			; mm4=tmp3
196	psubw	mm0,mm1			; mm0=tmp2
197	paddw	mm6,mm5			; mm6=tmp0
198	paddw	mm7,mm1			; mm7=tmp1
199
200	movq	MMWORD [wk(1)], mm4	; wk(1)=tmp3
201	movq	MMWORD [wk(0)], mm0	; wk(0)=tmp2
202
203	; -- Odd part
204
205	movq	mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
206	movq	mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
207	pmullw	mm2, MMWORD [MMBLOCK(1,0,edx,SIZEOF_IFAST_MULT_TYPE)]
208	pmullw	mm3, MMWORD [MMBLOCK(3,0,edx,SIZEOF_IFAST_MULT_TYPE)]
209	movq	mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
210	movq	mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
211	pmullw	mm5, MMWORD [MMBLOCK(5,0,edx,SIZEOF_IFAST_MULT_TYPE)]
212	pmullw	mm1, MMWORD [MMBLOCK(7,0,edx,SIZEOF_IFAST_MULT_TYPE)]
213
214	movq	mm4,mm2
215	movq	mm0,mm5
216	psubw	mm2,mm1			; mm2=z12
217	psubw	mm5,mm3			; mm5=z10
218	paddw	mm4,mm1			; mm4=z11
219	paddw	mm0,mm3			; mm0=z13
220
221	movq	mm1,mm5			; mm1=z10(unscaled)
222	psllw	mm2,PRE_MULTIPLY_SCALE_BITS
223	psllw	mm5,PRE_MULTIPLY_SCALE_BITS
224
225	movq	mm3,mm4
226	psubw	mm4,mm0
227	paddw	mm3,mm0			; mm3=tmp7
228
229	psllw	mm4,PRE_MULTIPLY_SCALE_BITS
230	pmulhw	mm4,[GOTOFF(ebx,PW_F1414)]	; mm4=tmp11
231
232	; To avoid overflow...
233	;
234	; (Original)
235	; tmp12 = -2.613125930 * z10 + z5;
236	;
237	; (This implementation)
238	; tmp12 = (-1.613125930 - 1) * z10 + z5;
239	;       = -1.613125930 * z10 - z10 + z5;
240
241	movq	mm0,mm5
242	paddw	mm5,mm2
243	pmulhw	mm5,[GOTOFF(ebx,PW_F1847)]	; mm5=z5
244	pmulhw	mm0,[GOTOFF(ebx,PW_MF1613)]
245	pmulhw	mm2,[GOTOFF(ebx,PW_F1082)]
246	psubw	mm0,mm1
247	psubw	mm2,mm5			; mm2=tmp10
248	paddw	mm0,mm5			; mm0=tmp12
249
250	; -- Final output stage
251
252	psubw	mm0,mm3			; mm0=tmp6
253	movq	mm1,mm6
254	movq	mm5,mm7
255	paddw	mm6,mm3			; mm6=data0=(00 01 02 03)
256	paddw	mm7,mm0			; mm7=data1=(10 11 12 13)
257	psubw	mm1,mm3			; mm1=data7=(70 71 72 73)
258	psubw	mm5,mm0			; mm5=data6=(60 61 62 63)
259	psubw	mm4,mm0			; mm4=tmp5
260
261	movq      mm3,mm6		; transpose coefficients(phase 1)
262	punpcklwd mm6,mm7		; mm6=(00 10 01 11)
263	punpckhwd mm3,mm7		; mm3=(02 12 03 13)
264	movq      mm0,mm5		; transpose coefficients(phase 1)
265	punpcklwd mm5,mm1		; mm5=(60 70 61 71)
266	punpckhwd mm0,mm1		; mm0=(62 72 63 73)
267
268	movq	mm7, MMWORD [wk(0)]	; mm7=tmp2
269	movq	mm1, MMWORD [wk(1)]	; mm1=tmp3
270
271	movq	MMWORD [wk(0)], mm5	; wk(0)=(60 70 61 71)
272	movq	MMWORD [wk(1)], mm0	; wk(1)=(62 72 63 73)
273
274	paddw	mm2,mm4			; mm2=tmp4
275	movq	mm5,mm7
276	movq	mm0,mm1
277	paddw	mm7,mm4			; mm7=data2=(20 21 22 23)
278	paddw	mm1,mm2			; mm1=data4=(40 41 42 43)
279	psubw	mm5,mm4			; mm5=data5=(50 51 52 53)
280	psubw	mm0,mm2			; mm0=data3=(30 31 32 33)
281
282	movq      mm4,mm7		; transpose coefficients(phase 1)
283	punpcklwd mm7,mm0		; mm7=(20 30 21 31)
284	punpckhwd mm4,mm0		; mm4=(22 32 23 33)
285	movq      mm2,mm1		; transpose coefficients(phase 1)
286	punpcklwd mm1,mm5		; mm1=(40 50 41 51)
287	punpckhwd mm2,mm5		; mm2=(42 52 43 53)
288
289	movq      mm0,mm6		; transpose coefficients(phase 2)
290	punpckldq mm6,mm7		; mm6=(00 10 20 30)
291	punpckhdq mm0,mm7		; mm0=(01 11 21 31)
292	movq      mm5,mm3		; transpose coefficients(phase 2)
293	punpckldq mm3,mm4		; mm3=(02 12 22 32)
294	punpckhdq mm5,mm4		; mm5=(03 13 23 33)
295
296	movq	mm7, MMWORD [wk(0)]	; mm7=(60 70 61 71)
297	movq	mm4, MMWORD [wk(1)]	; mm4=(62 72 63 73)
298
299	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm6
300	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm0
301	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm3
302	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm5
303
304	movq      mm6,mm1		; transpose coefficients(phase 2)
305	punpckldq mm1,mm7		; mm1=(40 50 60 70)
306	punpckhdq mm6,mm7		; mm6=(41 51 61 71)
307	movq      mm0,mm2		; transpose coefficients(phase 2)
308	punpckldq mm2,mm4		; mm2=(42 52 62 72)
309	punpckhdq mm0,mm4		; mm0=(43 53 63 73)
310
311	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm1
312	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm6
313	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
314	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm0
315
316.nextcolumn:
317	add	esi, byte 4*SIZEOF_JCOEF		; coef_block
318	add	edx, byte 4*SIZEOF_IFAST_MULT_TYPE	; quantptr
319	add	edi, byte 4*DCTSIZE*SIZEOF_JCOEF	; wsptr
320	dec	ecx					; ctr
321	jnz	near .columnloop
322
323	; ---- Pass 2: process rows from work array, store into output array.
324
325	mov	eax, [original_ebp]
326	lea	esi, [workspace]			; JCOEF * wsptr
327	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
328	mov	eax, JDIMENSION [output_col(eax)]
329	mov	ecx, DCTSIZE/4				; ctr
330	alignx	16,7
331.rowloop:
332
333	; -- Even part
334
335	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
336	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
337	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
338	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
339
340	movq	mm4,mm0
341	movq	mm5,mm1
342	psubw	mm0,mm2			; mm0=tmp11
343	psubw	mm1,mm3
344	paddw	mm4,mm2			; mm4=tmp10
345	paddw	mm5,mm3			; mm5=tmp13
346
347	psllw	mm1,PRE_MULTIPLY_SCALE_BITS
348	pmulhw	mm1,[GOTOFF(ebx,PW_F1414)]
349	psubw	mm1,mm5			; mm1=tmp12
350
351	movq	mm6,mm4
352	movq	mm7,mm0
353	psubw	mm4,mm5			; mm4=tmp3
354	psubw	mm0,mm1			; mm0=tmp2
355	paddw	mm6,mm5			; mm6=tmp0
356	paddw	mm7,mm1			; mm7=tmp1
357
358	movq	MMWORD [wk(1)], mm4	; wk(1)=tmp3
359	movq	MMWORD [wk(0)], mm0	; wk(0)=tmp2
360
361	; -- Odd part
362
363	movq	mm2, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
364	movq	mm3, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
365	movq	mm5, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
366	movq	mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
367
368	movq	mm4,mm2
369	movq	mm0,mm5
370	psubw	mm2,mm1			; mm2=z12
371	psubw	mm5,mm3			; mm5=z10
372	paddw	mm4,mm1			; mm4=z11
373	paddw	mm0,mm3			; mm0=z13
374
375	movq	mm1,mm5			; mm1=z10(unscaled)
376	psllw	mm2,PRE_MULTIPLY_SCALE_BITS
377	psllw	mm5,PRE_MULTIPLY_SCALE_BITS
378
379	movq	mm3,mm4
380	psubw	mm4,mm0
381	paddw	mm3,mm0			; mm3=tmp7
382
383	psllw	mm4,PRE_MULTIPLY_SCALE_BITS
384	pmulhw	mm4,[GOTOFF(ebx,PW_F1414)]	; mm4=tmp11
385
386	; To avoid overflow...
387	;
388	; (Original)
389	; tmp12 = -2.613125930 * z10 + z5;
390	;
391	; (This implementation)
392	; tmp12 = (-1.613125930 - 1) * z10 + z5;
393	;       = -1.613125930 * z10 - z10 + z5;
394
395	movq	mm0,mm5
396	paddw	mm5,mm2
397	pmulhw	mm5,[GOTOFF(ebx,PW_F1847)]	; mm5=z5
398	pmulhw	mm0,[GOTOFF(ebx,PW_MF1613)]
399	pmulhw	mm2,[GOTOFF(ebx,PW_F1082)]
400	psubw	mm0,mm1
401	psubw	mm2,mm5			; mm2=tmp10
402	paddw	mm0,mm5			; mm0=tmp12
403
404	; -- Final output stage
405
406	psubw	mm0,mm3			; mm0=tmp6
407	movq	mm1,mm6
408	movq	mm5,mm7
409	paddw	mm6,mm3			; mm6=data0=(00 10 20 30)
410	paddw	mm7,mm0			; mm7=data1=(01 11 21 31)
411	psraw	mm6,(PASS1_BITS+3)	; descale
412	psraw	mm7,(PASS1_BITS+3)	; descale
413	psubw	mm1,mm3			; mm1=data7=(07 17 27 37)
414	psubw	mm5,mm0			; mm5=data6=(06 16 26 36)
415	psraw	mm1,(PASS1_BITS+3)	; descale
416	psraw	mm5,(PASS1_BITS+3)	; descale
417	psubw	mm4,mm0			; mm4=tmp5
418
419	packsswb  mm6,mm5		; mm6=(00 10 20 30 06 16 26 36)
420	packsswb  mm7,mm1		; mm7=(01 11 21 31 07 17 27 37)
421
422	movq	mm3, MMWORD [wk(0)]	; mm3=tmp2
423	movq	mm0, MMWORD [wk(1)]	; mm0=tmp3
424
425	paddw	mm2,mm4			; mm2=tmp4
426	movq	mm5,mm3
427	movq	mm1,mm0
428	paddw	mm3,mm4			; mm3=data2=(02 12 22 32)
429	paddw	mm0,mm2			; mm0=data4=(04 14 24 34)
430	psraw	mm3,(PASS1_BITS+3)	; descale
431	psraw	mm0,(PASS1_BITS+3)	; descale
432	psubw	mm5,mm4			; mm5=data5=(05 15 25 35)
433	psubw	mm1,mm2			; mm1=data3=(03 13 23 33)
434	psraw	mm5,(PASS1_BITS+3)	; descale
435	psraw	mm1,(PASS1_BITS+3)	; descale
436
437	movq      mm4,[GOTOFF(ebx,PB_CENTERJSAMP)]	; mm4=[PB_CENTERJSAMP]
438
439	packsswb  mm3,mm0		; mm3=(02 12 22 32 04 14 24 34)
440	packsswb  mm1,mm5		; mm1=(03 13 23 33 05 15 25 35)
441
442	paddb     mm6,mm4
443	paddb     mm7,mm4
444	paddb     mm3,mm4
445	paddb     mm1,mm4
446
447	movq      mm2,mm6		; transpose coefficients(phase 1)
448	punpcklbw mm6,mm7		; mm6=(00 01 10 11 20 21 30 31)
449	punpckhbw mm2,mm7		; mm2=(06 07 16 17 26 27 36 37)
450	movq      mm0,mm3		; transpose coefficients(phase 1)
451	punpcklbw mm3,mm1		; mm3=(02 03 12 13 22 23 32 33)
452	punpckhbw mm0,mm1		; mm0=(04 05 14 15 24 25 34 35)
453
454	movq      mm5,mm6		; transpose coefficients(phase 2)
455	punpcklwd mm6,mm3		; mm6=(00 01 02 03 10 11 12 13)
456	punpckhwd mm5,mm3		; mm5=(20 21 22 23 30 31 32 33)
457	movq      mm4,mm0		; transpose coefficients(phase 2)
458	punpcklwd mm0,mm2		; mm0=(04 05 06 07 14 15 16 17)
459	punpckhwd mm4,mm2		; mm4=(24 25 26 27 34 35 36 37)
460
461	movq      mm7,mm6		; transpose coefficients(phase 3)
462	punpckldq mm6,mm0		; mm6=(00 01 02 03 04 05 06 07)
463	punpckhdq mm7,mm0		; mm7=(10 11 12 13 14 15 16 17)
464	movq      mm1,mm5		; transpose coefficients(phase 3)
465	punpckldq mm5,mm4		; mm5=(20 21 22 23 24 25 26 27)
466	punpckhdq mm1,mm4		; mm1=(30 31 32 33 34 35 36 37)
467
468	pushpic	ebx			; save GOT address
469
470	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
471	mov	ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
472	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm6
473	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm7
474	mov	edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
475	mov	ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
476	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm5
477	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm1
478
479	poppic	ebx			; restore GOT address
480
481	add	esi, byte 4*SIZEOF_JCOEF	; wsptr
482	add	edi, byte 4*SIZEOF_JSAMPROW
483	dec	ecx				; ctr
484	jnz	near .rowloop
485
486	emms		; empty MMX state
487
488	pop	edi
489	pop	esi
490;	pop	edx		; need not be preserved
491;	pop	ecx		; need not be preserved
492	pop	ebx
493	mov	esp,ebp		; esp <- aligned ebp
494	pop	esp		; esp <- original ebp
495	pop	ebp
496	ret
497
498; For some reason, the OS X linker does not honor the request to align the
499; segment unless we do this.
500	align	16