1;
2; jfsseflt.asm - floating-point FDCT (SSE)
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 floating-point implementation of the forward DCT
18; (Discrete Cosine Transform). The following code is based directly on
19; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
20;
21; [TAB8]
22
23%include "jsimdext.inc"
24%include "jdct.inc"
25
26; --------------------------------------------------------------------------
27
28%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
29	shufps	%1,%2,0x44
30%endmacro
31
32%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
33	shufps	%1,%2,0xEE
34%endmacro
35
36; --------------------------------------------------------------------------
37	SECTION	SEG_CONST
38
39	alignz	16
40	global	EXTN(jconst_fdct_float_sse) PRIVATE
41
42EXTN(jconst_fdct_float_sse):
43
44PD_0_382	times 4 dd  0.382683432365089771728460
45PD_0_707	times 4 dd  0.707106781186547524400844
46PD_0_541	times 4 dd  0.541196100146196984399723
47PD_1_306	times 4 dd  1.306562964876376527856643
48
49	alignz	16
50
51; --------------------------------------------------------------------------
52	SECTION	SEG_TEXT
53	BITS	32
54;
55; Perform the forward DCT on one block of samples.
56;
57; GLOBAL(void)
58; jsimd_fdct_float_sse (FAST_FLOAT * data)
59;
60
61%define data(b)		(b)+8		; FAST_FLOAT * data
62
63%define original_ebp	ebp+0
64%define wk(i)		ebp-(WK_NUM-(i))*SIZEOF_XMMWORD	; xmmword wk[WK_NUM]
65%define WK_NUM		2
66
67	align	16
68	global	EXTN(jsimd_fdct_float_sse) PRIVATE
69
70EXTN(jsimd_fdct_float_sse):
71	push	ebp
72	mov	eax,esp				; eax = original ebp
73	sub	esp, byte 4
74	and	esp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
75	mov	[esp],eax
76	mov	ebp,esp				; ebp = aligned ebp
77	lea	esp, [wk(0)]
78	pushpic	ebx
79;	push	ecx		; need not be preserved
80;	push	edx		; need not be preserved
81;	push	esi		; unused
82;	push	edi		; unused
83
84	get_GOT	ebx		; get GOT address
85
86	; ---- Pass 1: process rows.
87
88	mov	edx, POINTER [data(eax)]	; (FAST_FLOAT *)
89	mov	ecx, DCTSIZE/4
90	alignx	16,7
91.rowloop:
92
93	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
94	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
95	movaps	xmm2, XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)]
96	movaps	xmm3, XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)]
97
98	; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
99	; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
100
101	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
102	unpcklps xmm0,xmm1		; xmm0=(20 30 21 31)
103	unpckhps xmm4,xmm1		; xmm4=(22 32 23 33)
104	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
105	unpcklps xmm2,xmm3		; xmm2=(24 34 25 35)
106	unpckhps xmm5,xmm3		; xmm5=(26 36 27 37)
107
108	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
109	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
110	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)]
111	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)]
112
113	; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
114	; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
115
116	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 32 23 33)
117	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(24 34 25 35)
118
119	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
120	unpcklps xmm6,xmm7		; xmm6=(00 10 01 11)
121	unpckhps xmm4,xmm7		; xmm4=(02 12 03 13)
122	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
123	unpcklps xmm1,xmm3		; xmm1=(04 14 05 15)
124	unpckhps xmm2,xmm3		; xmm2=(06 16 07 17)
125
126	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
127	unpcklps2 xmm6,xmm0		; xmm6=(00 10 20 30)=data0
128	unpckhps2 xmm7,xmm0		; xmm7=(01 11 21 31)=data1
129	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
130	unpcklps2 xmm2,xmm5		; xmm2=(06 16 26 36)=data6
131	unpckhps2 xmm3,xmm5		; xmm3=(07 17 27 37)=data7
132
133	movaps	xmm0,xmm7
134	movaps	xmm5,xmm6
135	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
136	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
137	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
138	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
139
140	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 32 23 33)
141	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(24 34 25 35)
142	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
143	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
144
145	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
146	unpcklps2 xmm4,xmm2		; xmm4=(02 12 22 32)=data2
147	unpckhps2 xmm7,xmm2		; xmm7=(03 13 23 33)=data3
148	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
149	unpcklps2 xmm1,xmm3		; xmm1=(04 14 24 34)=data4
150	unpckhps2 xmm6,xmm3		; xmm6=(05 15 25 35)=data5
151
152	movaps	xmm2,xmm7
153	movaps	xmm3,xmm4
154	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
155	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
156	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
157	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
158
159	; -- Even part
160
161	movaps	xmm1,xmm5
162	movaps	xmm6,xmm0
163	subps	xmm5,xmm7		; xmm5=tmp13
164	subps	xmm0,xmm4		; xmm0=tmp12
165	addps	xmm1,xmm7		; xmm1=tmp10
166	addps	xmm6,xmm4		; xmm6=tmp11
167
168	addps	xmm0,xmm5
169	mulps	xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1
170
171	movaps	xmm7,xmm1
172	movaps	xmm4,xmm5
173	subps	xmm1,xmm6		; xmm1=data4
174	subps	xmm5,xmm0		; xmm5=data6
175	addps	xmm7,xmm6		; xmm7=data0
176	addps	xmm4,xmm0		; xmm4=data2
177
178	movaps	XMMWORD [XMMBLOCK(0,1,edx,SIZEOF_FAST_FLOAT)], xmm1
179	movaps	XMMWORD [XMMBLOCK(2,1,edx,SIZEOF_FAST_FLOAT)], xmm5
180	movaps	XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7
181	movaps	XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4
182
183	; -- Odd part
184
185	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
186	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
187
188	addps	xmm2,xmm3		; xmm2=tmp10
189	addps	xmm3,xmm6		; xmm3=tmp11
190	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
191
192	mulps	xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3
193
194	movaps	xmm1,xmm2		; xmm1=tmp10
195	subps	xmm2,xmm6
196	mulps	xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5
197	mulps	xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
198	mulps	xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
199	addps	xmm1,xmm2		; xmm1=z2
200	addps	xmm6,xmm2		; xmm6=z4
201
202	movaps	xmm5,xmm0
203	subps	xmm0,xmm3		; xmm0=z13
204	addps	xmm5,xmm3		; xmm5=z11
205
206	movaps	xmm7,xmm0
207	movaps	xmm4,xmm5
208	subps	xmm0,xmm1		; xmm0=data3
209	subps	xmm5,xmm6		; xmm5=data7
210	addps	xmm7,xmm1		; xmm7=data5
211	addps	xmm4,xmm6		; xmm4=data1
212
213	movaps	XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0
214	movaps	XMMWORD [XMMBLOCK(3,1,edx,SIZEOF_FAST_FLOAT)], xmm5
215	movaps	XMMWORD [XMMBLOCK(1,1,edx,SIZEOF_FAST_FLOAT)], xmm7
216	movaps	XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4
217
218	add	edx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
219	dec	ecx
220	jnz	near .rowloop
221
222	; ---- Pass 2: process columns.
223
224	mov	edx, POINTER [data(eax)]	; (FAST_FLOAT *)
225	mov	ecx, DCTSIZE/4
226	alignx	16,7
227.columnloop:
228
229	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)]
230	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)]
231	movaps	xmm2, XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)]
232	movaps	xmm3, XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)]
233
234	; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
235	; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
236
237	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
238	unpcklps xmm0,xmm1		; xmm0=(02 03 12 13)
239	unpckhps xmm4,xmm1		; xmm4=(22 23 32 33)
240	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
241	unpcklps xmm2,xmm3		; xmm2=(42 43 52 53)
242	unpckhps xmm5,xmm3		; xmm5=(62 63 72 73)
243
244	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)]
245	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)]
246	movaps	xmm1, XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)]
247	movaps	xmm3, XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)]
248
249	; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
250	; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
251
252	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 23 32 33)
253	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(42 43 52 53)
254
255	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
256	unpcklps xmm6,xmm7		; xmm6=(00 01 10 11)
257	unpckhps xmm4,xmm7		; xmm4=(20 21 30 31)
258	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
259	unpcklps xmm1,xmm3		; xmm1=(40 41 50 51)
260	unpckhps xmm2,xmm3		; xmm2=(60 61 70 71)
261
262	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
263	unpcklps2 xmm6,xmm0		; xmm6=(00 01 02 03)=data0
264	unpckhps2 xmm7,xmm0		; xmm7=(10 11 12 13)=data1
265	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
266	unpcklps2 xmm2,xmm5		; xmm2=(60 61 62 63)=data6
267	unpckhps2 xmm3,xmm5		; xmm3=(70 71 72 73)=data7
268
269	movaps	xmm0,xmm7
270	movaps	xmm5,xmm6
271	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
272	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
273	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
274	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
275
276	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 23 32 33)
277	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(42 43 52 53)
278	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
279	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
280
281	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
282	unpcklps2 xmm4,xmm2		; xmm4=(20 21 22 23)=data2
283	unpckhps2 xmm7,xmm2		; xmm7=(30 31 32 33)=data3
284	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
285	unpcklps2 xmm1,xmm3		; xmm1=(40 41 42 43)=data4
286	unpckhps2 xmm6,xmm3		; xmm6=(50 51 52 53)=data5
287
288	movaps	xmm2,xmm7
289	movaps	xmm3,xmm4
290	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
291	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
292	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
293	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
294
295	; -- Even part
296
297	movaps	xmm1,xmm5
298	movaps	xmm6,xmm0
299	subps	xmm5,xmm7		; xmm5=tmp13
300	subps	xmm0,xmm4		; xmm0=tmp12
301	addps	xmm1,xmm7		; xmm1=tmp10
302	addps	xmm6,xmm4		; xmm6=tmp11
303
304	addps	xmm0,xmm5
305	mulps	xmm0,[GOTOFF(ebx,PD_0_707)] ; xmm0=z1
306
307	movaps	xmm7,xmm1
308	movaps	xmm4,xmm5
309	subps	xmm1,xmm6		; xmm1=data4
310	subps	xmm5,xmm0		; xmm5=data6
311	addps	xmm7,xmm6		; xmm7=data0
312	addps	xmm4,xmm0		; xmm4=data2
313
314	movaps	XMMWORD [XMMBLOCK(4,0,edx,SIZEOF_FAST_FLOAT)], xmm1
315	movaps	XMMWORD [XMMBLOCK(6,0,edx,SIZEOF_FAST_FLOAT)], xmm5
316	movaps	XMMWORD [XMMBLOCK(0,0,edx,SIZEOF_FAST_FLOAT)], xmm7
317	movaps	XMMWORD [XMMBLOCK(2,0,edx,SIZEOF_FAST_FLOAT)], xmm4
318
319	; -- Odd part
320
321	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
322	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
323
324	addps	xmm2,xmm3		; xmm2=tmp10
325	addps	xmm3,xmm6		; xmm3=tmp11
326	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
327
328	mulps	xmm3,[GOTOFF(ebx,PD_0_707)] ; xmm3=z3
329
330	movaps	xmm1,xmm2		; xmm1=tmp10
331	subps	xmm2,xmm6
332	mulps	xmm2,[GOTOFF(ebx,PD_0_382)] ; xmm2=z5
333	mulps	xmm1,[GOTOFF(ebx,PD_0_541)] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
334	mulps	xmm6,[GOTOFF(ebx,PD_1_306)] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
335	addps	xmm1,xmm2		; xmm1=z2
336	addps	xmm6,xmm2		; xmm6=z4
337
338	movaps	xmm5,xmm0
339	subps	xmm0,xmm3		; xmm0=z13
340	addps	xmm5,xmm3		; xmm5=z11
341
342	movaps	xmm7,xmm0
343	movaps	xmm4,xmm5
344	subps	xmm0,xmm1		; xmm0=data3
345	subps	xmm5,xmm6		; xmm5=data7
346	addps	xmm7,xmm1		; xmm7=data5
347	addps	xmm4,xmm6		; xmm4=data1
348
349	movaps	XMMWORD [XMMBLOCK(3,0,edx,SIZEOF_FAST_FLOAT)], xmm0
350	movaps	XMMWORD [XMMBLOCK(7,0,edx,SIZEOF_FAST_FLOAT)], xmm5
351	movaps	XMMWORD [XMMBLOCK(5,0,edx,SIZEOF_FAST_FLOAT)], xmm7
352	movaps	XMMWORD [XMMBLOCK(1,0,edx,SIZEOF_FAST_FLOAT)], xmm4
353
354	add	edx, byte 4*SIZEOF_FAST_FLOAT
355	dec	ecx
356	jnz	near .columnloop
357
358;	pop	edi		; unused
359;	pop	esi		; unused
360;	pop	edx		; need not be preserved
361;	pop	ecx		; need not be preserved
362	poppic	ebx
363	mov	esp,ebp		; esp <- aligned ebp
364	pop	esp		; esp <- original ebp
365	pop	ebp
366	ret
367
368; For some reason, the OS X linker does not honor the request to align the
369; segment unless we do this.
370	align	16
371