2 * FFT transform with Altivec optimizations
3 * Copyright (c) 2009 Loren Merritt
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 * These functions are not individually interchangeable with the C versions.
24 * While C takes arrays of FFTComplex, Altivec leaves intermediate results
25 * in blocks as convenient to the vector size.
26 * i.e. {4x real, 4x imaginary, 4x real, ...}
28 * I ignore standard calling convention.
29 * Instead, the following registers are treated as global constants:
32 * v19..v29: permutations
35 * and the rest are free for local use.
43 /* Apple gas doesn't support this shorthand */
48 .macro addi2 ra, imm // add 32-bit immediate
53 addis \ra, \ra, \imm@ha
59 .macro LOAD_PTR ra, rbase, offset
60 ld \ra,(\offset)*8(\rbase)
62 .macro STORE_PTR ra, rbase, offset
63 std \ra,(\offset)*8(\rbase)
67 .macro LOAD_PTR ra, rbase, offset
68 lwz \ra,(\offset)*4(\rbase)
70 .macro STORE_PTR ra, rbase, offset
71 stw \ra,(\offset)*4(\rbase)
75 .macro FFT4 a0, a1, a2, a3 // in:0-1 out:2-3
76 vperm \a2,\a0,\a1,v20 // vcprm(0,1,s2,s1) // {r0,i0,r3,i2}
77 vperm \a3,\a0,\a1,v21 // vcprm(2,3,s0,s3) // {r1,i1,r2,i3}
78 vaddfp \a0,\a2,\a3 // {t1,t2,t6,t5}
79 vsubfp \a1,\a2,\a3 // {t3,t4,t8,t7}
80 vmrghw \a2,\a0,\a1 // vcprm(0,s0,1,s1) // {t1,t3,t2,t4}
81 vperm \a3,\a0,\a1,v22 // vcprm(2,s3,3,s2) // {t6,t7,t5,t8}
82 vaddfp \a0,\a2,\a3 // {r0,r1,i0,i1}
83 vsubfp \a1,\a2,\a3 // {r2,r3,i2,i3}
84 vperm \a2,\a0,\a1,v23 // vcprm(0,1,s0,s1) // {r0,r1,r2,r3}
85 vperm \a3,\a0,\a1,v24 // vcprm(2,3,s2,s3) // {i0,i1,i2,i3}
88 .macro FFT4x2 a0, a1, b0, b1, a2, a3, b2, b3
89 vperm \a2,\a0,\a1,v20 // vcprm(0,1,s2,s1) // {r0,i0,r3,i2}
90 vperm \a3,\a0,\a1,v21 // vcprm(2,3,s0,s3) // {r1,i1,r2,i3}
93 vaddfp \a0,\a2,\a3 // {t1,t2,t6,t5}
94 vsubfp \a1,\a2,\a3 // {t3,t4,t8,t7}
97 vmrghw \a2,\a0,\a1 // vcprm(0,s0,1,s1) // {t1,t3,t2,t4}
98 vperm \a3,\a0,\a1,v22 // vcprm(2,s3,3,s2) // {t6,t7,t5,t8}
100 vperm \b3,\b0,\b1,v22
101 vaddfp \a0,\a2,\a3 // {r0,r1,i0,i1}
102 vsubfp \a1,\a2,\a3 // {r2,r3,i2,i3}
105 vperm \a2,\a0,\a1,v23 // vcprm(0,1,s0,s1) // {r0,r1,r2,r3}
106 vperm \a3,\a0,\a1,v24 // vcprm(2,3,s2,s3) // {i0,i1,i2,i3}
107 vperm \b2,\b0,\b1,v23
108 vperm \b3,\b0,\b1,v24
111 .macro FFT8 a0, a1, b0, b1, a2, a3, b2, b3, b4 // in,out:a0-b1
112 vmrghw \b2,\b0,\b1 // vcprm(0,s0,1,s1) // {r4,r6,i4,i6}
113 vmrglw \b3,\b0,\b1 // vcprm(2,s2,3,s3) // {r5,r7,i5,i7}
114 vperm \a2,\a0,\a1,v20 // FFT4 ...
115 vperm \a3,\a0,\a1,v21
116 vaddfp \b0,\b2,\b3 // {t1,t3,t2,t4}
117 vsubfp \b1,\b2,\b3 // {r5,r7,i5,i7}
118 vperm \b4,\b1,\b1,v25 // vcprm(2,3,0,1) // {i5,i7,r5,r7}
121 vmaddfp \b1,\b1,v17,v14 // * {-1,1,1,-1}/sqrt(2)
122 vmaddfp \b1,\b4,v18,\b1 // * { 1,1,1,1 }/sqrt(2) // {t8,ta,t7,t9}
124 vperm \a3,\a0,\a1,v22
125 vperm \b2,\b0,\b1,v26 // vcprm(1,2,s3,s0) // {t3,t2,t9,t8}
126 vperm \b3,\b0,\b1,v27 // vcprm(0,3,s2,s1) // {t1,t4,t7,ta}
129 vaddfp \b0,\b2,\b3 // {t1,t2,t9,ta}
130 vsubfp \b1,\b2,\b3 // {t6,t5,tc,tb}
131 vperm \a2,\a0,\a1,v23
132 vperm \a3,\a0,\a1,v24
133 vperm \b2,\b0,\b1,v28 // vcprm(0,2,s1,s3) // {t1,t9,t5,tb}
134 vperm \b3,\b0,\b1,v29 // vcprm(1,3,s0,s2) // {t2,ta,t6,tc}
135 vsubfp \b0,\a2,\b2 // {r4,r5,r6,r7}
136 vsubfp \b1,\a3,\b3 // {i4,i5,i6,i7}
137 vaddfp \a0,\a2,\b2 // {r0,r1,r2,r3}
138 vaddfp \a1,\a3,\b3 // {i0,i1,i2,i3}
141 .macro BF d0,d1,s0,s1
146 .macro zip d0,d1,s0,s1
151 .macro def_fft4 interleave
152 fft4\interleave\()_altivec:
167 .macro def_fft8 interleave
168 fft8\interleave\()_altivec:
174 FFT8 v0,v1,v2,v3,v4,v5,v6,v7,v8
191 .macro def_fft16 interleave
192 fft16\interleave\()_altivec:
200 FFT4x2 v0,v1,v2,v3,v4,v5,v6,v7
205 FFT8 v0,v1,v2,v3,v8,v9,v10,v11,v12
206 vmaddfp v8,v4,v15,v14 // r2*wre
207 vmaddfp v9,v5,v15,v14 // i2*wre
208 vmaddfp v10,v6,v15,v14 // r3*wre
209 vmaddfp v11,v7,v15,v14 // i3*wre
210 vmaddfp v8,v5,v16,v8 // i2*wim
211 vnmsubfp v9,v4,v16,v9 // r2*wim
212 vnmsubfp v10,v7,v16,v10 // i3*wim
213 vmaddfp v11,v6,v16,v11 // r3*wim
246 // void pass(float *z, float *wre, int n)
247 .macro PASS interleave, suffix
248 fft_pass\suffix\()_altivec:
255 addi r6,r5,16 // o1+16
256 addi r8,r7,16 // o2+16
257 addi r11,r10,16 // o3+16
263 vperm v9,v9,v10,v19 // vcprm(s0,3,2,1) => wim[0 .. -3]
264 lvx v4,r3,r7 // r2 = z[o2]
265 lvx v5,r3,r8 // i2 = z[o2+16]
266 lvx v6,r3,r10 // r3 = z[o3]
267 lvx v7,r3,r11 // i3 = z[o3+16]
268 vmaddfp v10,v4,v8,v14 // r2*wre
269 vmaddfp v11,v5,v8,v14 // i2*wre
270 vmaddfp v12,v6,v8,v14 // r3*wre
271 vmaddfp v13,v7,v8,v14 // i3*wre
272 lvx v0, 0,r3 // r0 = z[0]
273 lvx v3,r3,r6 // i1 = z[o1+16]
274 vmaddfp v10,v5,v9,v10 // i2*wim
275 vnmsubfp v11,v4,v9,v11 // r2*wim
276 vnmsubfp v12,v7,v9,v12 // i3*wim
277 vmaddfp v13,v6,v9,v13 // r3*wim
278 lvx v1,r3,r9 // i0 = z[16]
279 lvx v2,r3,r5 // r1 = z[o1]
322 .macro DECL_FFT suffix, bits, n, n2, n4
323 fft\n\suffix\()_altivec:
325 STORE_PTR r0,r1,\bits-5
332 LOAD_PTR r0,r1,\bits-5
333 LOAD_PTR r4,r12,\bits
336 b fft_pass\suffix\()_altivec
339 .macro DECL_FFTS interleave, suffix
344 PASS \interleave, \suffix
345 DECL_FFT \suffix, 5, 32, 16, 8
346 DECL_FFT \suffix, 6, 64, 32, 16
347 DECL_FFT \suffix, 7, 128, 64, 32
348 DECL_FFT \suffix, 8, 256, 128, 64
349 DECL_FFT \suffix, 9, 512, 256, 128
350 DECL_FFT \suffix,10, 1024, 512, 256
351 DECL_FFT \suffix,11, 2048, 1024, 512
352 DECL_FFT \suffix,12, 4096, 2048, 1024
353 DECL_FFT \suffix,13, 8192, 4096, 2048
354 DECL_FFT \suffix,14,16384, 8192, 4096
355 DECL_FFT \suffix,15,32768,16384, 8192
356 DECL_FFT \suffix,16,65536,32768,16384
359 .global EXTERN_ASM\()ff_fft_dispatch\suffix\()_altivec
360 EXTERN_ASM\()ff_fft_dispatch\suffix\()_altivec:
361 PTR fft4\suffix\()_altivec
362 PTR fft8\suffix\()_altivec
363 PTR fft16\suffix\()_altivec
364 PTR fft32\suffix\()_altivec
365 PTR fft64\suffix\()_altivec
366 PTR fft128\suffix\()_altivec
367 PTR fft256\suffix\()_altivec
368 PTR fft512\suffix\()_altivec
369 PTR fft1024\suffix\()_altivec
370 PTR fft2048\suffix\()_altivec
371 PTR fft4096\suffix\()_altivec
372 PTR fft8192\suffix\()_altivec
373 PTR fft16384\suffix\()_altivec
374 PTR fft32768\suffix\()_altivec
375 PTR fft65536\suffix\()_altivec
379 DECL_FFTS 1, _interleave