PageRenderTime 64ms CodeModel.GetById 15ms app.highlight 43ms RepoModel.GetById 1ms app.codeStats 1ms

/media/libjpeg/simd/jimmxint.asm

http://github.com/zpao/v8monkey
Assembly | 852 lines | 597 code | 143 blank | 112 comment | 1 complexity | 0eaf46a641f04a8c2c93ac45b71bb56a MD5 | raw file
  1;
  2; jimmxint.asm - accurate 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 slow-but-accurate integer implementation of the
 18; inverse DCT (Discrete Cosine Transform). The following code is based
 19; directly on the IJG's original jidctint.c; see the jidctint.c for
 20; more details.
 21;
 22; [TAB8]
 23
 24%include "jsimdext.inc"
 25%include "jdct.inc"
 26
 27; --------------------------------------------------------------------------
 28
 29%define CONST_BITS	13
 30%define PASS1_BITS	2
 31
 32%define DESCALE_P1	(CONST_BITS-PASS1_BITS)
 33%define DESCALE_P2	(CONST_BITS+PASS1_BITS+3)
 34
 35%if CONST_BITS == 13
 36F_0_298	equ	 2446		; FIX(0.298631336)
 37F_0_390	equ	 3196		; FIX(0.390180644)
 38F_0_541	equ	 4433		; FIX(0.541196100)
 39F_0_765	equ	 6270		; FIX(0.765366865)
 40F_0_899	equ	 7373		; FIX(0.899976223)
 41F_1_175	equ	 9633		; FIX(1.175875602)
 42F_1_501	equ	12299		; FIX(1.501321110)
 43F_1_847	equ	15137		; FIX(1.847759065)
 44F_1_961	equ	16069		; FIX(1.961570560)
 45F_2_053	equ	16819		; FIX(2.053119869)
 46F_2_562	equ	20995		; FIX(2.562915447)
 47F_3_072	equ	25172		; FIX(3.072711026)
 48%else
 49; NASM cannot do compile-time arithmetic on floating-point constants.
 50%define DESCALE(x,n)  (((x)+(1<<((n)-1)))>>(n))
 51F_0_298	equ	DESCALE( 320652955,30-CONST_BITS)	; FIX(0.298631336)
 52F_0_390	equ	DESCALE( 418953276,30-CONST_BITS)	; FIX(0.390180644)
 53F_0_541	equ	DESCALE( 581104887,30-CONST_BITS)	; FIX(0.541196100)
 54F_0_765	equ	DESCALE( 821806413,30-CONST_BITS)	; FIX(0.765366865)
 55F_0_899	equ	DESCALE( 966342111,30-CONST_BITS)	; FIX(0.899976223)
 56F_1_175	equ	DESCALE(1262586813,30-CONST_BITS)	; FIX(1.175875602)
 57F_1_501	equ	DESCALE(1612031267,30-CONST_BITS)	; FIX(1.501321110)
 58F_1_847	equ	DESCALE(1984016188,30-CONST_BITS)	; FIX(1.847759065)
 59F_1_961	equ	DESCALE(2106220350,30-CONST_BITS)	; FIX(1.961570560)
 60F_2_053	equ	DESCALE(2204520673,30-CONST_BITS)	; FIX(2.053119869)
 61F_2_562	equ	DESCALE(2751909506,30-CONST_BITS)	; FIX(2.562915447)
 62F_3_072	equ	DESCALE(3299298341,30-CONST_BITS)	; FIX(3.072711026)
 63%endif
 64
 65; --------------------------------------------------------------------------
 66	SECTION	SEG_CONST
 67
 68	alignz	16
 69	global	EXTN(jconst_idct_islow_mmx)
 70
 71EXTN(jconst_idct_islow_mmx):
 72
 73PW_F130_F054	times 2 dw  (F_0_541+F_0_765), F_0_541
 74PW_F054_MF130	times 2 dw  F_0_541, (F_0_541-F_1_847)
 75PW_MF078_F117	times 2 dw  (F_1_175-F_1_961), F_1_175
 76PW_F117_F078	times 2 dw  F_1_175, (F_1_175-F_0_390)
 77PW_MF060_MF089	times 2 dw  (F_0_298-F_0_899),-F_0_899
 78PW_MF089_F060	times 2 dw -F_0_899, (F_1_501-F_0_899)
 79PW_MF050_MF256	times 2 dw  (F_2_053-F_2_562),-F_2_562
 80PW_MF256_F050	times 2 dw -F_2_562, (F_3_072-F_2_562)
 81PD_DESCALE_P1	times 2 dd  1 << (DESCALE_P1-1)
 82PD_DESCALE_P2	times 2 dd  1 << (DESCALE_P2-1)
 83PB_CENTERJSAMP	times 8 db  CENTERJSAMPLE
 84
 85	alignz	16
 86
 87; --------------------------------------------------------------------------
 88	SECTION	SEG_TEXT
 89	BITS	32
 90;
 91; Perform dequantization and inverse DCT on one block of coefficients.
 92;
 93; GLOBAL(void)
 94; jsimd_idct_islow_mmx (void * dct_table, JCOEFPTR coef_block,
 95;                       JSAMPARRAY output_buf, JDIMENSION output_col)
 96;
 97
 98%define dct_table(b)	(b)+8			; jpeg_component_info * compptr
 99%define coef_block(b)	(b)+12		; JCOEFPTR coef_block
100%define output_buf(b)	(b)+16		; JSAMPARRAY output_buf
101%define output_col(b)	(b)+20		; JDIMENSION output_col
102
103%define original_ebp	ebp+0
104%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_MMWORD	; mmword wk[WK_NUM]
105%define WK_NUM		12
106%define workspace	wk(0)-DCTSIZE2*SIZEOF_JCOEF
107					; JCOEF workspace[DCTSIZE2]
108
109	align	16
110	global	EXTN(jsimd_idct_islow_mmx)
111
112EXTN(jsimd_idct_islow_mmx):
113	push	ebp
114	mov	eax,esp				; eax = original ebp
115	sub	esp, byte 4
116	and	esp, byte (-SIZEOF_MMWORD)	; align to 64 bits
117	mov	[esp],eax
118	mov	ebp,esp				; ebp = aligned ebp
119	lea	esp, [workspace]
120	push	ebx
121;	push	ecx		; need not be preserved
122;	push	edx		; need not be preserved
123	push	esi
124	push	edi
125
126	get_GOT	ebx		; get GOT address
127
128	; ---- Pass 1: process columns from input, store into work array.
129
130;	mov	eax, [original_ebp]
131	mov	edx, POINTER [dct_table(eax)]	; quantptr
132	mov	esi, JCOEFPTR [coef_block(eax)]		; inptr
133	lea	edi, [workspace]			; JCOEF * wsptr
134	mov	ecx, DCTSIZE/4				; ctr
135	alignx	16,7
136.columnloop:
137%ifndef NO_ZERO_COLUMN_TEST_ISLOW_MMX
138	mov	eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
139	or	eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
140	jnz	short .columnDCT
141
142	movq	mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
143	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
144	por	mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
145	por	mm1, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
146	por	mm0, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
147	por	mm1, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
148	por	mm0, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
149	por	mm1,mm0
150	packsswb mm1,mm1
151	movd	eax,mm1
152	test	eax,eax
153	jnz	short .columnDCT
154
155	; -- AC terms all zero
156
157	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
158	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
159
160	psllw	mm0,PASS1_BITS
161
162	movq      mm2,mm0		; mm0=in0=(00 01 02 03)
163	punpcklwd mm0,mm0		; mm0=(00 00 01 01)
164	punpckhwd mm2,mm2		; mm2=(02 02 03 03)
165
166	movq      mm1,mm0
167	punpckldq mm0,mm0		; mm0=(00 00 00 00)
168	punpckhdq mm1,mm1		; mm1=(01 01 01 01)
169	movq      mm3,mm2
170	punpckldq mm2,mm2		; mm2=(02 02 02 02)
171	punpckhdq mm3,mm3		; mm3=(03 03 03 03)
172
173	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
174	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
175	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
176	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm1
177	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
178	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm2
179	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
180	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm3
181	jmp	near .nextcolumn
182	alignx	16,7
183%endif
184.columnDCT:
185
186	; -- Even part
187
188	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
189	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
190	pmullw	mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
191	pmullw	mm1, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
192	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
193	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
194	pmullw	mm2, MMWORD [MMBLOCK(4,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
195	pmullw	mm3, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
196
197	; (Original)
198	; z1 = (z2 + z3) * 0.541196100;
199	; tmp2 = z1 + z3 * -1.847759065;
200	; tmp3 = z1 + z2 * 0.765366865;
201	;
202	; (This implementation)
203	; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
204	; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
205
206	movq      mm4,mm1		; mm1=in2=z2
207	movq      mm5,mm1
208	punpcklwd mm4,mm3		; mm3=in6=z3
209	punpckhwd mm5,mm3
210	movq      mm1,mm4
211	movq      mm3,mm5
212	pmaddwd   mm4,[GOTOFF(ebx,PW_F130_F054)]	; mm4=tmp3L
213	pmaddwd   mm5,[GOTOFF(ebx,PW_F130_F054)]	; mm5=tmp3H
214	pmaddwd   mm1,[GOTOFF(ebx,PW_F054_MF130)]	; mm1=tmp2L
215	pmaddwd   mm3,[GOTOFF(ebx,PW_F054_MF130)]	; mm3=tmp2H
216
217	movq      mm6,mm0
218	paddw     mm0,mm2		; mm0=in0+in4
219	psubw     mm6,mm2		; mm6=in0-in4
220
221	pxor      mm7,mm7
222	pxor      mm2,mm2
223	punpcklwd mm7,mm0		; mm7=tmp0L
224	punpckhwd mm2,mm0		; mm2=tmp0H
225	psrad     mm7,(16-CONST_BITS)	; psrad mm7,16 & pslld mm7,CONST_BITS
226	psrad     mm2,(16-CONST_BITS)	; psrad mm2,16 & pslld mm2,CONST_BITS
227
228	movq	mm0,mm7
229	paddd	mm7,mm4			; mm7=tmp10L
230	psubd	mm0,mm4			; mm0=tmp13L
231	movq	mm4,mm2
232	paddd	mm2,mm5			; mm2=tmp10H
233	psubd	mm4,mm5			; mm4=tmp13H
234
235	movq	MMWORD [wk(0)], mm7	; wk(0)=tmp10L
236	movq	MMWORD [wk(1)], mm2	; wk(1)=tmp10H
237	movq	MMWORD [wk(2)], mm0	; wk(2)=tmp13L
238	movq	MMWORD [wk(3)], mm4	; wk(3)=tmp13H
239
240	pxor      mm5,mm5
241	pxor      mm7,mm7
242	punpcklwd mm5,mm6		; mm5=tmp1L
243	punpckhwd mm7,mm6		; mm7=tmp1H
244	psrad     mm5,(16-CONST_BITS)	; psrad mm5,16 & pslld mm5,CONST_BITS
245	psrad     mm7,(16-CONST_BITS)	; psrad mm7,16 & pslld mm7,CONST_BITS
246
247	movq	mm2,mm5
248	paddd	mm5,mm1			; mm5=tmp11L
249	psubd	mm2,mm1			; mm2=tmp12L
250	movq	mm0,mm7
251	paddd	mm7,mm3			; mm7=tmp11H
252	psubd	mm0,mm3			; mm0=tmp12H
253
254	movq	MMWORD [wk(4)], mm5	; wk(4)=tmp11L
255	movq	MMWORD [wk(5)], mm7	; wk(5)=tmp11H
256	movq	MMWORD [wk(6)], mm2	; wk(6)=tmp12L
257	movq	MMWORD [wk(7)], mm0	; wk(7)=tmp12H
258
259	; -- Odd part
260
261	movq	mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
262	movq	mm6, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
263	pmullw	mm4, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
264	pmullw	mm6, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
265	movq	mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
266	movq	mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
267	pmullw	mm1, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
268	pmullw	mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
269
270	movq	mm5,mm6
271	movq	mm7,mm4
272	paddw	mm5,mm3			; mm5=z3
273	paddw	mm7,mm1			; mm7=z4
274
275	; (Original)
276	; z5 = (z3 + z4) * 1.175875602;
277	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
278	; z3 += z5;  z4 += z5;
279	;
280	; (This implementation)
281	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
282	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
283
284	movq      mm2,mm5
285	movq      mm0,mm5
286	punpcklwd mm2,mm7
287	punpckhwd mm0,mm7
288	movq      mm5,mm2
289	movq      mm7,mm0
290	pmaddwd   mm2,[GOTOFF(ebx,PW_MF078_F117)]	; mm2=z3L
291	pmaddwd   mm0,[GOTOFF(ebx,PW_MF078_F117)]	; mm0=z3H
292	pmaddwd   mm5,[GOTOFF(ebx,PW_F117_F078)]	; mm5=z4L
293	pmaddwd   mm7,[GOTOFF(ebx,PW_F117_F078)]	; mm7=z4H
294
295	movq	MMWORD [wk(10)], mm2	; wk(10)=z3L
296	movq	MMWORD [wk(11)], mm0	; wk(11)=z3H
297
298	; (Original)
299	; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
300	; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
301	; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
302	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
303	; tmp0 += z1 + z3;  tmp1 += z2 + z4;
304	; tmp2 += z2 + z3;  tmp3 += z1 + z4;
305	;
306	; (This implementation)
307	; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
308	; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
309	; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
310	; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
311	; tmp0 += z3;  tmp1 += z4;
312	; tmp2 += z3;  tmp3 += z4;
313
314	movq      mm2,mm3
315	movq      mm0,mm3
316	punpcklwd mm2,mm4
317	punpckhwd mm0,mm4
318	movq      mm3,mm2
319	movq      mm4,mm0
320	pmaddwd   mm2,[GOTOFF(ebx,PW_MF060_MF089)]	; mm2=tmp0L
321	pmaddwd   mm0,[GOTOFF(ebx,PW_MF060_MF089)]	; mm0=tmp0H
322	pmaddwd   mm3,[GOTOFF(ebx,PW_MF089_F060)]	; mm3=tmp3L
323	pmaddwd   mm4,[GOTOFF(ebx,PW_MF089_F060)]	; mm4=tmp3H
324
325	paddd	mm2, MMWORD [wk(10)]	; mm2=tmp0L
326	paddd	mm0, MMWORD [wk(11)]	; mm0=tmp0H
327	paddd	mm3,mm5			; mm3=tmp3L
328	paddd	mm4,mm7			; mm4=tmp3H
329
330	movq	MMWORD [wk(8)], mm2	; wk(8)=tmp0L
331	movq	MMWORD [wk(9)], mm0	; wk(9)=tmp0H
332
333	movq      mm2,mm1
334	movq      mm0,mm1
335	punpcklwd mm2,mm6
336	punpckhwd mm0,mm6
337	movq      mm1,mm2
338	movq      mm6,mm0
339	pmaddwd   mm2,[GOTOFF(ebx,PW_MF050_MF256)]	; mm2=tmp1L
340	pmaddwd   mm0,[GOTOFF(ebx,PW_MF050_MF256)]	; mm0=tmp1H
341	pmaddwd   mm1,[GOTOFF(ebx,PW_MF256_F050)]	; mm1=tmp2L
342	pmaddwd   mm6,[GOTOFF(ebx,PW_MF256_F050)]	; mm6=tmp2H
343
344	paddd	mm2,mm5			; mm2=tmp1L
345	paddd	mm0,mm7			; mm0=tmp1H
346	paddd	mm1, MMWORD [wk(10)]	; mm1=tmp2L
347	paddd	mm6, MMWORD [wk(11)]	; mm6=tmp2H
348
349	movq	MMWORD [wk(10)], mm2	; wk(10)=tmp1L
350	movq	MMWORD [wk(11)], mm0	; wk(11)=tmp1H
351
352	; -- Final output stage
353
354	movq	mm5, MMWORD [wk(0)]	; mm5=tmp10L
355	movq	mm7, MMWORD [wk(1)]	; mm7=tmp10H
356
357	movq	mm2,mm5
358	movq	mm0,mm7
359	paddd	mm5,mm3			; mm5=data0L
360	paddd	mm7,mm4			; mm7=data0H
361	psubd	mm2,mm3			; mm2=data7L
362	psubd	mm0,mm4			; mm0=data7H
363
364	movq	mm3,[GOTOFF(ebx,PD_DESCALE_P1)]	; mm3=[PD_DESCALE_P1]
365
366	paddd	mm5,mm3
367	paddd	mm7,mm3
368	psrad	mm5,DESCALE_P1
369	psrad	mm7,DESCALE_P1
370	paddd	mm2,mm3
371	paddd	mm0,mm3
372	psrad	mm2,DESCALE_P1
373	psrad	mm0,DESCALE_P1
374
375	packssdw  mm5,mm7		; mm5=data0=(00 01 02 03)
376	packssdw  mm2,mm0		; mm2=data7=(70 71 72 73)
377
378	movq	mm4, MMWORD [wk(4)]	; mm4=tmp11L
379	movq	mm3, MMWORD [wk(5)]	; mm3=tmp11H
380
381	movq	mm7,mm4
382	movq	mm0,mm3
383	paddd	mm4,mm1			; mm4=data1L
384	paddd	mm3,mm6			; mm3=data1H
385	psubd	mm7,mm1			; mm7=data6L
386	psubd	mm0,mm6			; mm0=data6H
387
388	movq	mm1,[GOTOFF(ebx,PD_DESCALE_P1)]	; mm1=[PD_DESCALE_P1]
389
390	paddd	mm4,mm1
391	paddd	mm3,mm1
392	psrad	mm4,DESCALE_P1
393	psrad	mm3,DESCALE_P1
394	paddd	mm7,mm1
395	paddd	mm0,mm1
396	psrad	mm7,DESCALE_P1
397	psrad	mm0,DESCALE_P1
398
399	packssdw  mm4,mm3		; mm4=data1=(10 11 12 13)
400	packssdw  mm7,mm0		; mm7=data6=(60 61 62 63)
401
402	movq      mm6,mm5		; transpose coefficients(phase 1)
403	punpcklwd mm5,mm4		; mm5=(00 10 01 11)
404	punpckhwd mm6,mm4		; mm6=(02 12 03 13)
405	movq      mm1,mm7		; transpose coefficients(phase 1)
406	punpcklwd mm7,mm2		; mm7=(60 70 61 71)
407	punpckhwd mm1,mm2		; mm1=(62 72 63 73)
408
409	movq	mm3, MMWORD [wk(6)]	; mm3=tmp12L
410	movq	mm0, MMWORD [wk(7)]	; mm0=tmp12H
411	movq	mm4, MMWORD [wk(10)]	; mm4=tmp1L
412	movq	mm2, MMWORD [wk(11)]	; mm2=tmp1H
413
414	movq	MMWORD [wk(0)], mm5	; wk(0)=(00 10 01 11)
415	movq	MMWORD [wk(1)], mm6	; wk(1)=(02 12 03 13)
416	movq	MMWORD [wk(4)], mm7	; wk(4)=(60 70 61 71)
417	movq	MMWORD [wk(5)], mm1	; wk(5)=(62 72 63 73)
418
419	movq	mm5,mm3
420	movq	mm6,mm0
421	paddd	mm3,mm4			; mm3=data2L
422	paddd	mm0,mm2			; mm0=data2H
423	psubd	mm5,mm4			; mm5=data5L
424	psubd	mm6,mm2			; mm6=data5H
425
426	movq	mm7,[GOTOFF(ebx,PD_DESCALE_P1)]	; mm7=[PD_DESCALE_P1]
427
428	paddd	mm3,mm7
429	paddd	mm0,mm7
430	psrad	mm3,DESCALE_P1
431	psrad	mm0,DESCALE_P1
432	paddd	mm5,mm7
433	paddd	mm6,mm7
434	psrad	mm5,DESCALE_P1
435	psrad	mm6,DESCALE_P1
436
437	packssdw  mm3,mm0		; mm3=data2=(20 21 22 23)
438	packssdw  mm5,mm6		; mm5=data5=(50 51 52 53)
439
440	movq	mm1, MMWORD [wk(2)]	; mm1=tmp13L
441	movq	mm4, MMWORD [wk(3)]	; mm4=tmp13H
442	movq	mm2, MMWORD [wk(8)]	; mm2=tmp0L
443	movq	mm7, MMWORD [wk(9)]	; mm7=tmp0H
444
445	movq	mm0,mm1
446	movq	mm6,mm4
447	paddd	mm1,mm2			; mm1=data3L
448	paddd	mm4,mm7			; mm4=data3H
449	psubd	mm0,mm2			; mm0=data4L
450	psubd	mm6,mm7			; mm6=data4H
451
452	movq	mm2,[GOTOFF(ebx,PD_DESCALE_P1)]	; mm2=[PD_DESCALE_P1]
453
454	paddd	mm1,mm2
455	paddd	mm4,mm2
456	psrad	mm1,DESCALE_P1
457	psrad	mm4,DESCALE_P1
458	paddd	mm0,mm2
459	paddd	mm6,mm2
460	psrad	mm0,DESCALE_P1
461	psrad	mm6,DESCALE_P1
462
463	packssdw  mm1,mm4		; mm1=data3=(30 31 32 33)
464	packssdw  mm0,mm6		; mm0=data4=(40 41 42 43)
465
466	movq	mm7, MMWORD [wk(0)]	; mm7=(00 10 01 11)
467	movq	mm2, MMWORD [wk(1)]	; mm2=(02 12 03 13)
468
469	movq      mm4,mm3		; transpose coefficients(phase 1)
470	punpcklwd mm3,mm1		; mm3=(20 30 21 31)
471	punpckhwd mm4,mm1		; mm4=(22 32 23 33)
472	movq      mm6,mm0		; transpose coefficients(phase 1)
473	punpcklwd mm0,mm5		; mm0=(40 50 41 51)
474	punpckhwd mm6,mm5		; mm6=(42 52 43 53)
475
476	movq      mm1,mm7		; transpose coefficients(phase 2)
477	punpckldq mm7,mm3		; mm7=(00 10 20 30)
478	punpckhdq mm1,mm3		; mm1=(01 11 21 31)
479	movq      mm5,mm2		; transpose coefficients(phase 2)
480	punpckldq mm2,mm4		; mm2=(02 12 22 32)
481	punpckhdq mm5,mm4		; mm5=(03 13 23 33)
482
483	movq	mm3, MMWORD [wk(4)]	; mm3=(60 70 61 71)
484	movq	mm4, MMWORD [wk(5)]	; mm4=(62 72 63 73)
485
486	movq	MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm7
487	movq	MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
488	movq	MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
489	movq	MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm5
490
491	movq      mm7,mm0		; transpose coefficients(phase 2)
492	punpckldq mm0,mm3		; mm0=(40 50 60 70)
493	punpckhdq mm7,mm3		; mm7=(41 51 61 71)
494	movq      mm1,mm6		; transpose coefficients(phase 2)
495	punpckldq mm6,mm4		; mm6=(42 52 62 72)
496	punpckhdq mm1,mm4		; mm1=(43 53 63 73)
497
498	movq	MMWORD [MMBLOCK(0,1,edi,SIZEOF_JCOEF)], mm0
499	movq	MMWORD [MMBLOCK(1,1,edi,SIZEOF_JCOEF)], mm7
500	movq	MMWORD [MMBLOCK(2,1,edi,SIZEOF_JCOEF)], mm6
501	movq	MMWORD [MMBLOCK(3,1,edi,SIZEOF_JCOEF)], mm1
502
503.nextcolumn:
504	add	esi, byte 4*SIZEOF_JCOEF		; coef_block
505	add	edx, byte 4*SIZEOF_ISLOW_MULT_TYPE	; quantptr
506	add	edi, byte 4*DCTSIZE*SIZEOF_JCOEF	; wsptr
507	dec	ecx					; ctr
508	jnz	near .columnloop
509
510	; ---- Pass 2: process rows from work array, store into output array.
511
512	mov	eax, [original_ebp]
513	lea	esi, [workspace]			; JCOEF * wsptr
514	mov	edi, JSAMPARRAY [output_buf(eax)]	; (JSAMPROW *)
515	mov	eax, JDIMENSION [output_col(eax)]
516	mov	ecx, DCTSIZE/4				; ctr
517	alignx	16,7
518.rowloop:
519
520	; -- Even part
521
522	movq	mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
523	movq	mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
524	movq	mm2, MMWORD [MMBLOCK(4,0,esi,SIZEOF_JCOEF)]
525	movq	mm3, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
526
527	; (Original)
528	; z1 = (z2 + z3) * 0.541196100;
529	; tmp2 = z1 + z3 * -1.847759065;
530	; tmp3 = z1 + z2 * 0.765366865;
531	;
532	; (This implementation)
533	; tmp2 = z2 * 0.541196100 + z3 * (0.541196100 - 1.847759065);
534	; tmp3 = z2 * (0.541196100 + 0.765366865) + z3 * 0.541196100;
535
536	movq      mm4,mm1		; mm1=in2=z2
537	movq      mm5,mm1
538	punpcklwd mm4,mm3		; mm3=in6=z3
539	punpckhwd mm5,mm3
540	movq      mm1,mm4
541	movq      mm3,mm5
542	pmaddwd   mm4,[GOTOFF(ebx,PW_F130_F054)]	; mm4=tmp3L
543	pmaddwd   mm5,[GOTOFF(ebx,PW_F130_F054)]	; mm5=tmp3H
544	pmaddwd   mm1,[GOTOFF(ebx,PW_F054_MF130)]	; mm1=tmp2L
545	pmaddwd   mm3,[GOTOFF(ebx,PW_F054_MF130)]	; mm3=tmp2H
546
547	movq      mm6,mm0
548	paddw     mm0,mm2		; mm0=in0+in4
549	psubw     mm6,mm2		; mm6=in0-in4
550
551	pxor      mm7,mm7
552	pxor      mm2,mm2
553	punpcklwd mm7,mm0		; mm7=tmp0L
554	punpckhwd mm2,mm0		; mm2=tmp0H
555	psrad     mm7,(16-CONST_BITS)	; psrad mm7,16 & pslld mm7,CONST_BITS
556	psrad     mm2,(16-CONST_BITS)	; psrad mm2,16 & pslld mm2,CONST_BITS
557
558	movq	mm0,mm7
559	paddd	mm7,mm4			; mm7=tmp10L
560	psubd	mm0,mm4			; mm0=tmp13L
561	movq	mm4,mm2
562	paddd	mm2,mm5			; mm2=tmp10H
563	psubd	mm4,mm5			; mm4=tmp13H
564
565	movq	MMWORD [wk(0)], mm7	; wk(0)=tmp10L
566	movq	MMWORD [wk(1)], mm2	; wk(1)=tmp10H
567	movq	MMWORD [wk(2)], mm0	; wk(2)=tmp13L
568	movq	MMWORD [wk(3)], mm4	; wk(3)=tmp13H
569
570	pxor      mm5,mm5
571	pxor      mm7,mm7
572	punpcklwd mm5,mm6		; mm5=tmp1L
573	punpckhwd mm7,mm6		; mm7=tmp1H
574	psrad     mm5,(16-CONST_BITS)	; psrad mm5,16 & pslld mm5,CONST_BITS
575	psrad     mm7,(16-CONST_BITS)	; psrad mm7,16 & pslld mm7,CONST_BITS
576
577	movq	mm2,mm5
578	paddd	mm5,mm1			; mm5=tmp11L
579	psubd	mm2,mm1			; mm2=tmp12L
580	movq	mm0,mm7
581	paddd	mm7,mm3			; mm7=tmp11H
582	psubd	mm0,mm3			; mm0=tmp12H
583
584	movq	MMWORD [wk(4)], mm5	; wk(4)=tmp11L
585	movq	MMWORD [wk(5)], mm7	; wk(5)=tmp11H
586	movq	MMWORD [wk(6)], mm2	; wk(6)=tmp12L
587	movq	MMWORD [wk(7)], mm0	; wk(7)=tmp12H
588
589	; -- Odd part
590
591	movq	mm4, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
592	movq	mm6, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
593	movq	mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
594	movq	mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
595
596	movq	mm5,mm6
597	movq	mm7,mm4
598	paddw	mm5,mm3			; mm5=z3
599	paddw	mm7,mm1			; mm7=z4
600
601	; (Original)
602	; z5 = (z3 + z4) * 1.175875602;
603	; z3 = z3 * -1.961570560;  z4 = z4 * -0.390180644;
604	; z3 += z5;  z4 += z5;
605	;
606	; (This implementation)
607	; z3 = z3 * (1.175875602 - 1.961570560) + z4 * 1.175875602;
608	; z4 = z3 * 1.175875602 + z4 * (1.175875602 - 0.390180644);
609
610	movq      mm2,mm5
611	movq      mm0,mm5
612	punpcklwd mm2,mm7
613	punpckhwd mm0,mm7
614	movq      mm5,mm2
615	movq      mm7,mm0
616	pmaddwd   mm2,[GOTOFF(ebx,PW_MF078_F117)]	; mm2=z3L
617	pmaddwd   mm0,[GOTOFF(ebx,PW_MF078_F117)]	; mm0=z3H
618	pmaddwd   mm5,[GOTOFF(ebx,PW_F117_F078)]	; mm5=z4L
619	pmaddwd   mm7,[GOTOFF(ebx,PW_F117_F078)]	; mm7=z4H
620
621	movq	MMWORD [wk(10)], mm2	; wk(10)=z3L
622	movq	MMWORD [wk(11)], mm0	; wk(11)=z3H
623
624	; (Original)
625	; z1 = tmp0 + tmp3;  z2 = tmp1 + tmp2;
626	; tmp0 = tmp0 * 0.298631336;  tmp1 = tmp1 * 2.053119869;
627	; tmp2 = tmp2 * 3.072711026;  tmp3 = tmp3 * 1.501321110;
628	; z1 = z1 * -0.899976223;  z2 = z2 * -2.562915447;
629	; tmp0 += z1 + z3;  tmp1 += z2 + z4;
630	; tmp2 += z2 + z3;  tmp3 += z1 + z4;
631	;
632	; (This implementation)
633	; tmp0 = tmp0 * (0.298631336 - 0.899976223) + tmp3 * -0.899976223;
634	; tmp1 = tmp1 * (2.053119869 - 2.562915447) + tmp2 * -2.562915447;
635	; tmp2 = tmp1 * -2.562915447 + tmp2 * (3.072711026 - 2.562915447);
636	; tmp3 = tmp0 * -0.899976223 + tmp3 * (1.501321110 - 0.899976223);
637	; tmp0 += z3;  tmp1 += z4;
638	; tmp2 += z3;  tmp3 += z4;
639
640	movq      mm2,mm3
641	movq      mm0,mm3
642	punpcklwd mm2,mm4
643	punpckhwd mm0,mm4
644	movq      mm3,mm2
645	movq      mm4,mm0
646	pmaddwd   mm2,[GOTOFF(ebx,PW_MF060_MF089)]	; mm2=tmp0L
647	pmaddwd   mm0,[GOTOFF(ebx,PW_MF060_MF089)]	; mm0=tmp0H
648	pmaddwd   mm3,[GOTOFF(ebx,PW_MF089_F060)]	; mm3=tmp3L
649	pmaddwd   mm4,[GOTOFF(ebx,PW_MF089_F060)]	; mm4=tmp3H
650
651	paddd	mm2, MMWORD [wk(10)]	; mm2=tmp0L
652	paddd	mm0, MMWORD [wk(11)]	; mm0=tmp0H
653	paddd	mm3,mm5			; mm3=tmp3L
654	paddd	mm4,mm7			; mm4=tmp3H
655
656	movq	MMWORD [wk(8)], mm2	; wk(8)=tmp0L
657	movq	MMWORD [wk(9)], mm0	; wk(9)=tmp0H
658
659	movq      mm2,mm1
660	movq      mm0,mm1
661	punpcklwd mm2,mm6
662	punpckhwd mm0,mm6
663	movq      mm1,mm2
664	movq      mm6,mm0
665	pmaddwd   mm2,[GOTOFF(ebx,PW_MF050_MF256)]	; mm2=tmp1L
666	pmaddwd   mm0,[GOTOFF(ebx,PW_MF050_MF256)]	; mm0=tmp1H
667	pmaddwd   mm1,[GOTOFF(ebx,PW_MF256_F050)]	; mm1=tmp2L
668	pmaddwd   mm6,[GOTOFF(ebx,PW_MF256_F050)]	; mm6=tmp2H
669
670	paddd	mm2,mm5			; mm2=tmp1L
671	paddd	mm0,mm7			; mm0=tmp1H
672	paddd	mm1, MMWORD [wk(10)]	; mm1=tmp2L
673	paddd	mm6, MMWORD [wk(11)]	; mm6=tmp2H
674
675	movq	MMWORD [wk(10)], mm2	; wk(10)=tmp1L
676	movq	MMWORD [wk(11)], mm0	; wk(11)=tmp1H
677
678	; -- Final output stage
679
680	movq	mm5, MMWORD [wk(0)]	; mm5=tmp10L
681	movq	mm7, MMWORD [wk(1)]	; mm7=tmp10H
682
683	movq	mm2,mm5
684	movq	mm0,mm7
685	paddd	mm5,mm3			; mm5=data0L
686	paddd	mm7,mm4			; mm7=data0H
687	psubd	mm2,mm3			; mm2=data7L
688	psubd	mm0,mm4			; mm0=data7H
689
690	movq	mm3,[GOTOFF(ebx,PD_DESCALE_P2)]	; mm3=[PD_DESCALE_P2]
691
692	paddd	mm5,mm3
693	paddd	mm7,mm3
694	psrad	mm5,DESCALE_P2
695	psrad	mm7,DESCALE_P2
696	paddd	mm2,mm3
697	paddd	mm0,mm3
698	psrad	mm2,DESCALE_P2
699	psrad	mm0,DESCALE_P2
700
701	packssdw  mm5,mm7		; mm5=data0=(00 10 20 30)
702	packssdw  mm2,mm0		; mm2=data7=(07 17 27 37)
703
704	movq	mm4, MMWORD [wk(4)]	; mm4=tmp11L
705	movq	mm3, MMWORD [wk(5)]	; mm3=tmp11H
706
707	movq	mm7,mm4
708	movq	mm0,mm3
709	paddd	mm4,mm1			; mm4=data1L
710	paddd	mm3,mm6			; mm3=data1H
711	psubd	mm7,mm1			; mm7=data6L
712	psubd	mm0,mm6			; mm0=data6H
713
714	movq	mm1,[GOTOFF(ebx,PD_DESCALE_P2)]	; mm1=[PD_DESCALE_P2]
715
716	paddd	mm4,mm1
717	paddd	mm3,mm1
718	psrad	mm4,DESCALE_P2
719	psrad	mm3,DESCALE_P2
720	paddd	mm7,mm1
721	paddd	mm0,mm1
722	psrad	mm7,DESCALE_P2
723	psrad	mm0,DESCALE_P2
724
725	packssdw  mm4,mm3		; mm4=data1=(01 11 21 31)
726	packssdw  mm7,mm0		; mm7=data6=(06 16 26 36)
727
728	packsswb  mm5,mm7		; mm5=(00 10 20 30 06 16 26 36)
729	packsswb  mm4,mm2		; mm4=(01 11 21 31 07 17 27 37)
730
731	movq	mm6, MMWORD [wk(6)]	; mm6=tmp12L
732	movq	mm1, MMWORD [wk(7)]	; mm1=tmp12H
733	movq	mm3, MMWORD [wk(10)]	; mm3=tmp1L
734	movq	mm0, MMWORD [wk(11)]	; mm0=tmp1H
735
736	movq	MMWORD [wk(0)], mm5	; wk(0)=(00 10 20 30 06 16 26 36)
737	movq	MMWORD [wk(1)], mm4	; wk(1)=(01 11 21 31 07 17 27 37)
738
739	movq	mm7,mm6
740	movq	mm2,mm1
741	paddd	mm6,mm3			; mm6=data2L
742	paddd	mm1,mm0			; mm1=data2H
743	psubd	mm7,mm3			; mm7=data5L
744	psubd	mm2,mm0			; mm2=data5H
745
746	movq	mm5,[GOTOFF(ebx,PD_DESCALE_P2)]	; mm5=[PD_DESCALE_P2]
747
748	paddd	mm6,mm5
749	paddd	mm1,mm5
750	psrad	mm6,DESCALE_P2
751	psrad	mm1,DESCALE_P2
752	paddd	mm7,mm5
753	paddd	mm2,mm5
754	psrad	mm7,DESCALE_P2
755	psrad	mm2,DESCALE_P2
756
757	packssdw  mm6,mm1		; mm6=data2=(02 12 22 32)
758	packssdw  mm7,mm2		; mm7=data5=(05 15 25 35)
759
760	movq	mm4, MMWORD [wk(2)]	; mm4=tmp13L
761	movq	mm3, MMWORD [wk(3)]	; mm3=tmp13H
762	movq	mm0, MMWORD [wk(8)]	; mm0=tmp0L
763	movq	mm5, MMWORD [wk(9)]	; mm5=tmp0H
764
765	movq	mm1,mm4
766	movq	mm2,mm3
767	paddd	mm4,mm0			; mm4=data3L
768	paddd	mm3,mm5			; mm3=data3H
769	psubd	mm1,mm0			; mm1=data4L
770	psubd	mm2,mm5			; mm2=data4H
771
772	movq	mm0,[GOTOFF(ebx,PD_DESCALE_P2)]	; mm0=[PD_DESCALE_P2]
773
774	paddd	mm4,mm0
775	paddd	mm3,mm0
776	psrad	mm4,DESCALE_P2
777	psrad	mm3,DESCALE_P2
778	paddd	mm1,mm0
779	paddd	mm2,mm0
780	psrad	mm1,DESCALE_P2
781	psrad	mm2,DESCALE_P2
782
783	movq      mm5,[GOTOFF(ebx,PB_CENTERJSAMP)]	; mm5=[PB_CENTERJSAMP]
784
785	packssdw  mm4,mm3		; mm4=data3=(03 13 23 33)
786	packssdw  mm1,mm2		; mm1=data4=(04 14 24 34)
787
788	movq      mm0, MMWORD [wk(0)]	; mm0=(00 10 20 30 06 16 26 36)
789	movq      mm3, MMWORD [wk(1)]	; mm3=(01 11 21 31 07 17 27 37)
790
791	packsswb  mm6,mm1		; mm6=(02 12 22 32 04 14 24 34)
792	packsswb  mm4,mm7		; mm4=(03 13 23 33 05 15 25 35)
793
794	paddb     mm0,mm5
795	paddb     mm3,mm5
796	paddb     mm6,mm5
797	paddb     mm4,mm5
798
799	movq      mm2,mm0		; transpose coefficients(phase 1)
800	punpcklbw mm0,mm3		; mm0=(00 01 10 11 20 21 30 31)
801	punpckhbw mm2,mm3		; mm2=(06 07 16 17 26 27 36 37)
802	movq      mm1,mm6		; transpose coefficients(phase 1)
803	punpcklbw mm6,mm4		; mm6=(02 03 12 13 22 23 32 33)
804	punpckhbw mm1,mm4		; mm1=(04 05 14 15 24 25 34 35)
805
806	movq      mm7,mm0		; transpose coefficients(phase 2)
807	punpcklwd mm0,mm6		; mm0=(00 01 02 03 10 11 12 13)
808	punpckhwd mm7,mm6		; mm7=(20 21 22 23 30 31 32 33)
809	movq      mm5,mm1		; transpose coefficients(phase 2)
810	punpcklwd mm1,mm2		; mm1=(04 05 06 07 14 15 16 17)
811	punpckhwd mm5,mm2		; mm5=(24 25 26 27 34 35 36 37)
812
813	movq      mm3,mm0		; transpose coefficients(phase 3)
814	punpckldq mm0,mm1		; mm0=(00 01 02 03 04 05 06 07)
815	punpckhdq mm3,mm1		; mm3=(10 11 12 13 14 15 16 17)
816	movq      mm4,mm7		; transpose coefficients(phase 3)
817	punpckldq mm7,mm5		; mm7=(20 21 22 23 24 25 26 27)
818	punpckhdq mm4,mm5		; mm4=(30 31 32 33 34 35 36 37)
819
820	pushpic	ebx			; save GOT address
821
822	mov	edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
823	mov	ebx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
824	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm0
825	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm3
826	mov	edx, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
827	mov	ebx, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
828	movq	MMWORD [edx+eax*SIZEOF_JSAMPLE], mm7
829	movq	MMWORD [ebx+eax*SIZEOF_JSAMPLE], mm4
830
831	poppic	ebx			; restore GOT address
832
833	add	esi, byte 4*SIZEOF_JCOEF	; wsptr
834	add	edi, byte 4*SIZEOF_JSAMPROW
835	dec	ecx				; ctr
836	jnz	near .rowloop
837
838	emms		; empty MMX state
839
840	pop	edi
841	pop	esi
842;	pop	edx		; need not be preserved
843;	pop	ecx		; need not be preserved
844	pop	ebx
845	mov	esp,ebp		; esp <- aligned ebp
846	pop	esp		; esp <- original ebp
847	pop	ebp
848	ret
849
850; For some reason, the OS X linker does not honor the request to align the
851; segment unless we do this.
852	align	16