1;
2; jfsseflt-64.asm - floating-point FDCT (64-bit SSE)
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 floating-point implementation of the forward DCT
19; (Discrete Cosine Transform). The following code is based directly on
20; the IJG's original jfdctflt.c; see the jfdctflt.c for more details.
21;
22; [TAB8]
23
24%include "jsimdext.inc"
25%include "jdct.inc"
26
27; --------------------------------------------------------------------------
28
29%macro	unpcklps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(0 1 4 5)
30	shufps	%1,%2,0x44
31%endmacro
32
33%macro	unpckhps2 2	; %1=(0 1 2 3) / %2=(4 5 6 7) => %1=(2 3 6 7)
34	shufps	%1,%2,0xEE
35%endmacro
36
37; --------------------------------------------------------------------------
38	SECTION	SEG_CONST
39
40	alignz	16
41	global	EXTN(jconst_fdct_float_sse) PRIVATE
42
43EXTN(jconst_fdct_float_sse):
44
45PD_0_382	times 4 dd  0.382683432365089771728460
46PD_0_707	times 4 dd  0.707106781186547524400844
47PD_0_541	times 4 dd  0.541196100146196984399723
48PD_1_306	times 4 dd  1.306562964876376527856643
49
50	alignz	16
51
52; --------------------------------------------------------------------------
53	SECTION	SEG_TEXT
54	BITS	64
55;
56; Perform the forward DCT on one block of samples.
57;
58; GLOBAL(void)
59; jsimd_fdct_float_sse (FAST_FLOAT * data)
60;
61
62; r10 = FAST_FLOAT * data
63
64%define wk(i)		rbp-(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	rbp
72	mov	rax,rsp				; rax = original rbp
73	sub	rsp, byte 4
74	and	rsp, byte (-SIZEOF_XMMWORD)	; align to 128 bits
75	mov	[rsp],rax
76	mov	rbp,rsp				; rbp = aligned rbp
77	lea	rsp, [wk(0)]
78	collect_args
79
80	; ---- Pass 1: process rows.
81
82	mov	rdx, r10	; (FAST_FLOAT *)
83	mov	rcx, DCTSIZE/4
84.rowloop:
85
86	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
87	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
88	movaps	xmm2, XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)]
89	movaps	xmm3, XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)]
90
91	; xmm0=(20 21 22 23), xmm2=(24 25 26 27)
92	; xmm1=(30 31 32 33), xmm3=(34 35 36 37)
93
94	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
95	unpcklps xmm0,xmm1		; xmm0=(20 30 21 31)
96	unpckhps xmm4,xmm1		; xmm4=(22 32 23 33)
97	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
98	unpcklps xmm2,xmm3		; xmm2=(24 34 25 35)
99	unpckhps xmm5,xmm3		; xmm5=(26 36 27 37)
100
101	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
102	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
103	movaps	xmm1, XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)]
104	movaps	xmm3, XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)]
105
106	; xmm6=(00 01 02 03), xmm1=(04 05 06 07)
107	; xmm7=(10 11 12 13), xmm3=(14 15 16 17)
108
109	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 32 23 33)
110	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(24 34 25 35)
111
112	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
113	unpcklps xmm6,xmm7		; xmm6=(00 10 01 11)
114	unpckhps xmm4,xmm7		; xmm4=(02 12 03 13)
115	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
116	unpcklps xmm1,xmm3		; xmm1=(04 14 05 15)
117	unpckhps xmm2,xmm3		; xmm2=(06 16 07 17)
118
119	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
120	unpcklps2 xmm6,xmm0		; xmm6=(00 10 20 30)=data0
121	unpckhps2 xmm7,xmm0		; xmm7=(01 11 21 31)=data1
122	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
123	unpcklps2 xmm2,xmm5		; xmm2=(06 16 26 36)=data6
124	unpckhps2 xmm3,xmm5		; xmm3=(07 17 27 37)=data7
125
126	movaps	xmm0,xmm7
127	movaps	xmm5,xmm6
128	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
129	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
130	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
131	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
132
133	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 32 23 33)
134	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(24 34 25 35)
135	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
136	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
137
138	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
139	unpcklps2 xmm4,xmm2		; xmm4=(02 12 22 32)=data2
140	unpckhps2 xmm7,xmm2		; xmm7=(03 13 23 33)=data3
141	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
142	unpcklps2 xmm1,xmm3		; xmm1=(04 14 24 34)=data4
143	unpckhps2 xmm6,xmm3		; xmm6=(05 15 25 35)=data5
144
145	movaps	xmm2,xmm7
146	movaps	xmm3,xmm4
147	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
148	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
149	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
150	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
151
152	; -- Even part
153
154	movaps	xmm1,xmm5
155	movaps	xmm6,xmm0
156	subps	xmm5,xmm7		; xmm5=tmp13
157	subps	xmm0,xmm4		; xmm0=tmp12
158	addps	xmm1,xmm7		; xmm1=tmp10
159	addps	xmm6,xmm4		; xmm6=tmp11
160
161	addps	xmm0,xmm5
162	mulps	xmm0,[rel PD_0_707] ; xmm0=z1
163
164	movaps	xmm7,xmm1
165	movaps	xmm4,xmm5
166	subps	xmm1,xmm6		; xmm1=data4
167	subps	xmm5,xmm0		; xmm5=data6
168	addps	xmm7,xmm6		; xmm7=data0
169	addps	xmm4,xmm0		; xmm4=data2
170
171	movaps	XMMWORD [XMMBLOCK(0,1,rdx,SIZEOF_FAST_FLOAT)], xmm1
172	movaps	XMMWORD [XMMBLOCK(2,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
173	movaps	XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
174	movaps	XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
175
176	; -- Odd part
177
178	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
179	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
180
181	addps	xmm2,xmm3		; xmm2=tmp10
182	addps	xmm3,xmm6		; xmm3=tmp11
183	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
184
185	mulps	xmm3,[rel PD_0_707] ; xmm3=z3
186
187	movaps	xmm1,xmm2		; xmm1=tmp10
188	subps	xmm2,xmm6
189	mulps	xmm2,[rel PD_0_382] ; xmm2=z5
190	mulps	xmm1,[rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
191	mulps	xmm6,[rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
192	addps	xmm1,xmm2		; xmm1=z2
193	addps	xmm6,xmm2		; xmm6=z4
194
195	movaps	xmm5,xmm0
196	subps	xmm0,xmm3		; xmm0=z13
197	addps	xmm5,xmm3		; xmm5=z11
198
199	movaps	xmm7,xmm0
200	movaps	xmm4,xmm5
201	subps	xmm0,xmm1		; xmm0=data3
202	subps	xmm5,xmm6		; xmm5=data7
203	addps	xmm7,xmm1		; xmm7=data5
204	addps	xmm4,xmm6		; xmm4=data1
205
206	movaps	XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
207	movaps	XMMWORD [XMMBLOCK(3,1,rdx,SIZEOF_FAST_FLOAT)], xmm5
208	movaps	XMMWORD [XMMBLOCK(1,1,rdx,SIZEOF_FAST_FLOAT)], xmm7
209	movaps	XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
210
211	add	rdx, 4*DCTSIZE*SIZEOF_FAST_FLOAT
212	dec	rcx
213	jnz	near .rowloop
214
215	; ---- Pass 2: process columns.
216
217	mov	rdx, r10	; (FAST_FLOAT *)
218	mov	rcx, DCTSIZE/4
219.columnloop:
220
221	movaps	xmm0, XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)]
222	movaps	xmm1, XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)]
223	movaps	xmm2, XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)]
224	movaps	xmm3, XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)]
225
226	; xmm0=(02 12 22 32), xmm2=(42 52 62 72)
227	; xmm1=(03 13 23 33), xmm3=(43 53 63 73)
228
229	movaps   xmm4,xmm0		; transpose coefficients(phase 1)
230	unpcklps xmm0,xmm1		; xmm0=(02 03 12 13)
231	unpckhps xmm4,xmm1		; xmm4=(22 23 32 33)
232	movaps   xmm5,xmm2		; transpose coefficients(phase 1)
233	unpcklps xmm2,xmm3		; xmm2=(42 43 52 53)
234	unpckhps xmm5,xmm3		; xmm5=(62 63 72 73)
235
236	movaps	xmm6, XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)]
237	movaps	xmm7, XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)]
238	movaps	xmm1, XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)]
239	movaps	xmm3, XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)]
240
241	; xmm6=(00 10 20 30), xmm1=(40 50 60 70)
242	; xmm7=(01 11 21 31), xmm3=(41 51 61 71)
243
244	movaps	XMMWORD [wk(0)], xmm4	; wk(0)=(22 23 32 33)
245	movaps	XMMWORD [wk(1)], xmm2	; wk(1)=(42 43 52 53)
246
247	movaps   xmm4,xmm6		; transpose coefficients(phase 1)
248	unpcklps xmm6,xmm7		; xmm6=(00 01 10 11)
249	unpckhps xmm4,xmm7		; xmm4=(20 21 30 31)
250	movaps   xmm2,xmm1		; transpose coefficients(phase 1)
251	unpcklps xmm1,xmm3		; xmm1=(40 41 50 51)
252	unpckhps xmm2,xmm3		; xmm2=(60 61 70 71)
253
254	movaps    xmm7,xmm6		; transpose coefficients(phase 2)
255	unpcklps2 xmm6,xmm0		; xmm6=(00 01 02 03)=data0
256	unpckhps2 xmm7,xmm0		; xmm7=(10 11 12 13)=data1
257	movaps    xmm3,xmm2		; transpose coefficients(phase 2)
258	unpcklps2 xmm2,xmm5		; xmm2=(60 61 62 63)=data6
259	unpckhps2 xmm3,xmm5		; xmm3=(70 71 72 73)=data7
260
261	movaps	xmm0,xmm7
262	movaps	xmm5,xmm6
263	subps	xmm7,xmm2		; xmm7=data1-data6=tmp6
264	subps	xmm6,xmm3		; xmm6=data0-data7=tmp7
265	addps	xmm0,xmm2		; xmm0=data1+data6=tmp1
266	addps	xmm5,xmm3		; xmm5=data0+data7=tmp0
267
268	movaps	xmm2, XMMWORD [wk(0)]	; xmm2=(22 23 32 33)
269	movaps	xmm3, XMMWORD [wk(1)]	; xmm3=(42 43 52 53)
270	movaps	XMMWORD [wk(0)], xmm7	; wk(0)=tmp6
271	movaps	XMMWORD [wk(1)], xmm6	; wk(1)=tmp7
272
273	movaps    xmm7,xmm4		; transpose coefficients(phase 2)
274	unpcklps2 xmm4,xmm2		; xmm4=(20 21 22 23)=data2
275	unpckhps2 xmm7,xmm2		; xmm7=(30 31 32 33)=data3
276	movaps    xmm6,xmm1		; transpose coefficients(phase 2)
277	unpcklps2 xmm1,xmm3		; xmm1=(40 41 42 43)=data4
278	unpckhps2 xmm6,xmm3		; xmm6=(50 51 52 53)=data5
279
280	movaps	xmm2,xmm7
281	movaps	xmm3,xmm4
282	addps	xmm7,xmm1		; xmm7=data3+data4=tmp3
283	addps	xmm4,xmm6		; xmm4=data2+data5=tmp2
284	subps	xmm2,xmm1		; xmm2=data3-data4=tmp4
285	subps	xmm3,xmm6		; xmm3=data2-data5=tmp5
286
287	; -- Even part
288
289	movaps	xmm1,xmm5
290	movaps	xmm6,xmm0
291	subps	xmm5,xmm7		; xmm5=tmp13
292	subps	xmm0,xmm4		; xmm0=tmp12
293	addps	xmm1,xmm7		; xmm1=tmp10
294	addps	xmm6,xmm4		; xmm6=tmp11
295
296	addps	xmm0,xmm5
297	mulps	xmm0,[rel PD_0_707] ; xmm0=z1
298
299	movaps	xmm7,xmm1
300	movaps	xmm4,xmm5
301	subps	xmm1,xmm6		; xmm1=data4
302	subps	xmm5,xmm0		; xmm5=data6
303	addps	xmm7,xmm6		; xmm7=data0
304	addps	xmm4,xmm0		; xmm4=data2
305
306	movaps	XMMWORD [XMMBLOCK(4,0,rdx,SIZEOF_FAST_FLOAT)], xmm1
307	movaps	XMMWORD [XMMBLOCK(6,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
308	movaps	XMMWORD [XMMBLOCK(0,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
309	movaps	XMMWORD [XMMBLOCK(2,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
310
311	; -- Odd part
312
313	movaps	xmm6, XMMWORD [wk(0)]	; xmm6=tmp6
314	movaps	xmm0, XMMWORD [wk(1)]	; xmm0=tmp7
315
316	addps	xmm2,xmm3		; xmm2=tmp10
317	addps	xmm3,xmm6		; xmm3=tmp11
318	addps	xmm6,xmm0		; xmm6=tmp12, xmm0=tmp7
319
320	mulps	xmm3,[rel PD_0_707] ; xmm3=z3
321
322	movaps	xmm1,xmm2		; xmm1=tmp10
323	subps	xmm2,xmm6
324	mulps	xmm2,[rel PD_0_382] ; xmm2=z5
325	mulps	xmm1,[rel PD_0_541] ; xmm1=MULTIPLY(tmp10,FIX_0_541196)
326	mulps	xmm6,[rel PD_1_306] ; xmm6=MULTIPLY(tmp12,FIX_1_306562)
327	addps	xmm1,xmm2		; xmm1=z2
328	addps	xmm6,xmm2		; xmm6=z4
329
330	movaps	xmm5,xmm0
331	subps	xmm0,xmm3		; xmm0=z13
332	addps	xmm5,xmm3		; xmm5=z11
333
334	movaps	xmm7,xmm0
335	movaps	xmm4,xmm5
336	subps	xmm0,xmm1		; xmm0=data3
337	subps	xmm5,xmm6		; xmm5=data7
338	addps	xmm7,xmm1		; xmm7=data5
339	addps	xmm4,xmm6		; xmm4=data1
340
341	movaps	XMMWORD [XMMBLOCK(3,0,rdx,SIZEOF_FAST_FLOAT)], xmm0
342	movaps	XMMWORD [XMMBLOCK(7,0,rdx,SIZEOF_FAST_FLOAT)], xmm5
343	movaps	XMMWORD [XMMBLOCK(5,0,rdx,SIZEOF_FAST_FLOAT)], xmm7
344	movaps	XMMWORD [XMMBLOCK(1,0,rdx,SIZEOF_FAST_FLOAT)], xmm4
345
346	add	rdx, byte 4*SIZEOF_FAST_FLOAT
347	dec	rcx
348	jnz	near .columnloop
349
350	uncollect_args
351	mov	rsp,rbp		; rsp <- aligned rbp
352	pop	rsp		; rsp <- original rbp
353	pop	rbp
354	ret
355
356; For some reason, the OS X linker does not honor the request to align the
357; segment unless we do this.
358	align	16
359