commit b90e85469bc89e51fc22d38a6fe70b83885fb9b1 Author: Loren Merritt Date: Sat Aug 22 11:22:09 2009 +0100 altivec split-radix FFT 1.85x faster than altivec radix-2 on a G4 8% faster vorbis decoding diff --git a/libavcodec/Makefile b/libavcodec/Makefile index 24e6af4..5a9113b 100644 --- a/libavcodec/Makefile +++ b/libavcodec/Makefile @@ -523,6 +523,7 @@ OBJS-$(HAVE_ALTIVEC) += ppc/check_altivec.o \ ppc/dsputil_altivec.o \ ppc/fdct_altivec.o \ ppc/fft_altivec.o \ + ppc/fft_altivec_s.o \ ppc/float_altivec.o \ ppc/gmc_altivec.o \ ppc/idct_altivec.o \ diff --git a/libavcodec/fft.c b/libavcodec/fft.c index a3f1151..d864f82 100644 --- a/libavcodec/fft.c +++ b/libavcodec/fft.c @@ -110,7 +110,6 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse) has_vectors = mm_support(); if (has_vectors & FF_MM_ALTIVEC) { s->fft_calc = ff_fft_calc_altivec; - split_radix = 0; } #endif diff --git a/libavcodec/ppc/fft_altivec.c b/libavcodec/ppc/fft_altivec.c index 7391131..0d47f1a 100644 --- a/libavcodec/ppc/fft_altivec.c +++ b/libavcodec/ppc/fft_altivec.c @@ -1,8 +1,7 @@ /* * FFT/IFFT transforms * AltiVec-enabled - * Copyright (c) 2003 Romain Dolbeau - * Based on code Copyright (c) 2002 Fabrice Bellard + * Copyright (c) 2009 Loren Merritt * * This file is part of FFmpeg. * @@ -23,6 +22,7 @@ #include "libavcodec/dsputil.h" #include "dsputil_ppc.h" #include "util_altivec.h" +#include "types_altivec.h" /** * Do a complex FFT with the parameters defined in ff_fft_init(). The * input data must be permuted before with s->revtab table. No @@ -30,106 +30,59 @@ * AltiVec-enabled * This code assumes that the 'z' pointer is 16 bytes-aligned * It also assumes all FFTComplex are 8 bytes-aligned pair of float - * The code is exactly the same as the SSE version, except - * that successive MUL + ADD/SUB have been merged into - * fused multiply-add ('vec_madd' in altivec) */ -void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) -{ -POWERPC_PERF_DECLARE(altivec_fft_num, s->nbits >= 6); - register const vector float vczero = (const vector float)vec_splat_u32(0.); - - int ln = s->nbits; - int j, np, np2; - int nblocks, nloops; - register FFTComplex *p, *q; - FFTComplex *cptr, *cptr1; - int k; - -POWERPC_PERF_START_COUNT(altivec_fft_num, s->nbits >= 6); - - np = 1 << ln; - - { - vector float *r, a, b, a1, c1, c2; - - r = (vector float *)&z[0]; - - c1 = vcii(p,p,n,n); - - if (s->inverse) { - c2 = vcii(p,p,n,p); - } else { - c2 = vcii(p,p,p,n); - } - - j = (np >> 2); - do { - a = vec_ld(0, r); - a1 = vec_ld(sizeof(vector float), r); - - b = vec_perm(a,a,vcprmle(1,0,3,2)); - a = vec_madd(a,c1,b); - /* do the pass 0 butterfly */ - - b = vec_perm(a1,a1,vcprmle(1,0,3,2)); - b = vec_madd(a1,c1,b); - /* do the pass 0 butterfly */ - - /* multiply third by -i */ - b = vec_perm(b,b,vcprmle(2,3,1,0)); - /* do the pass 1 butterfly */ - vec_st(vec_madd(b,c2,a), 0, r); - vec_st(vec_nmsub(b,c2,a), sizeof(vector float), r); +extern FFTSample ff_cos_16[]; +// pointers to functions. but unlike function pointers on some PPC ABIs, these aren't function descriptors. +extern void *ff_fft_dispatch_altivec[2][15]; - r += 2; - } while (--j != 0); +// convert from simd order to C order +static void swizzle(vec_f *z, int n) +{ + int i; + n >>= 1; + for(i=0; i> 3; - nloops = 1 << 2; - np2 = np >> 1; - - cptr1 = s->exptab1; - do { - p = z; - q = z + nloops; - j = nblocks; - do { - cptr = cptr1; - k = nloops >> 1; - do { - vector float a,b,c,t1; - - a = vec_ld(0, (float*)p); - b = vec_ld(0, (float*)q); - - /* complex mul */ - c = vec_ld(0, (float*)cptr); - /* cre*re cim*re */ - t1 = vec_madd(c, vec_perm(b,b,vcprmle(2,2,0,0)),vczero); - c = vec_ld(sizeof(vector float), (float*)cptr); - /* -cim*im cre*im */ - b = vec_madd(c, vec_perm(b,b,vcprmle(3,3,1,1)),t1); - - /* butterfly */ - vec_st(vec_add(a,b), 0, (float*)p); - vec_st(vec_sub(a,b), 0, (float*)q); - - p += 2; - q += 2; - cptr += 4; - } while (--k); - - p += nloops; - q += nloops; - } while (--j); - cptr1 += nloops * 2; - nblocks = nblocks >> 1; - nloops = nloops << 1; - } while (nblocks != 0); +} -POWERPC_PERF_STOP_COUNT(altivec_fft_num, s->nbits >= 6); +void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) +{ + register vec_f v14 __asm__("v14") = {0,0,0,0}; + register vec_f v15 __asm__("v15") = *(const vec_f*)ff_cos_16; + register vec_f v16 __asm__("v16") = {0, 0.38268343, M_SQRT1_2, 0.92387953}; + register vec_f v17 __asm__("v17") = {-M_SQRT1_2, M_SQRT1_2, M_SQRT1_2,-M_SQRT1_2}; + register vec_f v18 __asm__("v18") = { M_SQRT1_2, M_SQRT1_2, M_SQRT1_2, M_SQRT1_2}; + register vec_u8 v19 __asm__("v19") = vcprm(s0,3,2,1); + register vec_u8 v20 __asm__("v20") = vcprm(0,1,s2,s1); + register vec_u8 v21 __asm__("v21") = vcprm(2,3,s0,s3); + register vec_u8 v22 __asm__("v22") = vcprm(2,s3,3,s2); + register vec_u8 v23 __asm__("v23") = vcprm(0,1,s0,s1); + register vec_u8 v24 __asm__("v24") = vcprm(2,3,s2,s3); + register vec_u8 v25 __asm__("v25") = vcprm(2,3,0,1); + register vec_u8 v26 __asm__("v26") = vcprm(1,2,s3,s0); + register vec_u8 v27 __asm__("v27") = vcprm(0,3,s2,s1); + register vec_u8 v28 __asm__("v28") = vcprm(0,2,s1,s3); + register vec_u8 v29 __asm__("v29") = vcprm(1,3,s0,s2); + register FFTSample **cos_tabs __asm__("r12") = ff_cos_tabs; + register FFTComplex *zarg __asm__("r3") = z; + __asm__( + "mtctr %0 \n" + "li 11,16 \n" + "subi 1,1,%1 \n" + "bctrl \n" + "addi 1,1,%1 \n" + ::"r"(ff_fft_dispatch_altivec[1][s->nbits-2]), "i"(12*sizeof(void*)), + "r"(zarg), "r"(cos_tabs), + "v"(v14),"v"(v15),"v"(v16),"v"(v17),"v"(v18),"v"(v19),"v"(v20),"v"(v21), + "v"(v22),"v"(v23),"v"(v24),"v"(v25),"v"(v26),"v"(v27),"v"(v28),"v"(v29) + : "lr","ctr","r0","r1","r4","r5","r6","r7","r8","r9","r10","r11", + "v0","v1","v2","v3","v4","v5","v6","v7","v8","v9","v10","v11","v12","v13" + ); + if(s->nbits <= 4) + swizzle((vec_f*)z, 1<nbits); } + diff --git a/libavcodec/ppc/fft_altivec_s.S b/libavcodec/ppc/fft_altivec_s.S new file mode 100644 index 0000000..692ccf0 --- /dev/null +++ b/libavcodec/ppc/fft_altivec_s.S @@ -0,0 +1,318 @@ +/* + * FFT transform with Altivec optimizations + * Copyright (c) 2009 Loren Merritt + * + * This file is part of FFmpeg. + * + * FFmpeg is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * FFmpeg is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with FFmpeg; if not, write to the Free Software + * 51, Inc., Foundation Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA + */ + +/* + * These functions are not individually interchangeable with the C versions. + * While C takes arrays of FFTComplex, Altivec leaves intermediate results + * in blocks as convenient to the vector size. + * i.e. {4x real, 4x imaginary, 4x real, ...} + * + * I ignore standard calling convention. + * Instead, the following registers are treated as global constants: + * v14: zero + * v15..v18: cosines + * v19..v29: permutations + * r11: 16 + * r12: ff_cos_tabs + * and the rest are free for local use. + */ + +.text + +.macro slwi ra, rb, imm + rlwinm \ra, \rb, \imm, 0, 31-\imm +.endm + +.macro addi2 ra, imm // add 32-bit immediate +.if \imm & 0xffff + addi \ra, \ra, \imm@l +.endif +.if (\imm+0x8000)>>16 + addis \ra, \ra, ((\imm+0x8000)>>16) + (\imm&0x800000000000)*0x1fffe +.endif +.endm + +#if ARCH_PPC64 +.macro LOAD_PTR ra, rbase, offset + ld \ra,(\offset)*8(\rbase) +.endm +.macro STORE_PTR ra, rbase, offset + std \ra,(\offset)*8(\rbase) +.endm +#else +.macro LOAD_PTR ra, rbase, offset + lwz \ra,(\offset)*4(\rbase) +.endm +.macro STORE_PTR ra, rbase, offset + stw \ra,(\offset)*4(\rbase) +.endm +#endif + +.macro FFT4 a0, a1, a2, a3 // in:0-1 out:2-3 + vperm \a2,\a0,\a1,20 // vcprm(0,1,s2,s1) // {r0,i0,r3,i2} + vperm \a3,\a0,\a1,21 // vcprm(2,3,s0,s3) // {r1,i1,r2,i3} + vaddfp \a0,\a2,\a3 // {t1,t2,t6,t5} + vsubfp \a1,\a2,\a3 // {t3,t4,t8,t7} + vmrghw \a2,\a0,\a1 // vcprm(0,s0,1,s1) // {t1,t3,t2,t4} + vperm \a3,\a0,\a1,22 // vcprm(2,s3,3,s2) // {t6,t7,t5,t8} + vaddfp \a0,\a2,\a3 // {r0,r1,i0,i1} + vsubfp \a1,\a2,\a3 // {r2,r3,i2,i3} + vperm \a2,\a0,\a1,23 // vcprm(0,1,s0,s1) // {r0,r1,r2,r3} + vperm \a3,\a0,\a1,24 // vcprm(2,3,s2,s3) // {i0,i1,i2,i3} +.endm + +.macro FFT4x2 a0, a1, b0, b1, a2, a3, b2, b3 + vperm \a2,\a0,\a1,20 // vcprm(0,1,s2,s1) // {r0,i0,r3,i2} + vperm \a3,\a0,\a1,21 // vcprm(2,3,s0,s3) // {r1,i1,r2,i3} + vperm \b2,\b0,\b1,20 + vperm \b3,\b0,\b1,21 + vaddfp \a0,\a2,\a3 // {t1,t2,t6,t5} + vsubfp \a1,\a2,\a3 // {t3,t4,t8,t7} + vaddfp \b0,\b2,\b3 + vsubfp \b1,\b2,\b3 + vmrghw \a2,\a0,\a1 // vcprm(0,s0,1,s1) // {t1,t3,t2,t4} + vperm \a3,\a0,\a1,22 // vcprm(2,s3,3,s2) // {t6,t7,t5,t8} + vmrghw \b2,\b0,\b1 + vperm \b3,\b0,\b1,22 + vaddfp \a0,\a2,\a3 // {r0,r1,i0,i1} + vsubfp \a1,\a2,\a3 // {r2,r3,i2,i3} + vaddfp \b0,\b2,\b3 + vsubfp \b1,\b2,\b3 + vperm \a2,\a0,\a1,23 // vcprm(0,1,s0,s1) // {r0,r1,r2,r3} + vperm \a3,\a0,\a1,24 // vcprm(2,3,s2,s3) // {i0,i1,i2,i3} + vperm \b2,\b0,\b1,23 + vperm \b3,\b0,\b1,24 +.endm + +.macro FFT8 a0, a1, b0, b1, a2, a3, b2, b3, b4 // in,out:a0-b1 + vmrghw \b2,\b0,\b1 // vcprm(0,s0,1,s1) // {r4,r6,i4,i6} + vmrglw \b3,\b0,\b1 // vcprm(2,s2,3,s3) // {r5,r7,i5,i7} + vperm \a2,\a0,\a1,20 // FFT4 ... + vperm \a3,\a0,\a1,21 + vaddfp \b0,\b2,\b3 // {t1,t3,t2,t4} + vsubfp \b1,\b2,\b3 // {r5,r7,i5,i7} + vperm \b4,\b1,\b1,25 // vcprm(2,3,0,1) // {i5,i7,r5,r7} + vaddfp \a0,\a2,\a3 + vsubfp \a1,\a2,\a3 + vmaddfp \b1,\b1,17,14 // * {-1,1,1,-1}/sqrt(2) + vmaddfp \b1,\b4,18,\b1 // * { 1,1,1,1 }/sqrt(2) // {t8,ta,t7,t9} + vmrghw \a2,\a0,\a1 + vperm \a3,\a0,\a1,22 + vperm \b2,\b0,\b1,26 // vcprm(1,2,s3,s0) // {t3,t2,t9,t8} + vperm \b3,\b0,\b1,27 // vcprm(0,3,s2,s1) // {t1,t4,t7,ta} + vaddfp \a0,\a2,\a3 + vsubfp \a1,\a2,\a3 + vaddfp \b0,\b2,\b3 // {t1,t2,t9,ta} + vsubfp \b1,\b2,\b3 // {t6,t5,tc,tb} + vperm \a2,\a0,\a1,23 + vperm \a3,\a0,\a1,24 + vperm \b2,\b0,\b1,28 // vcprm(0,2,s1,s3) // {t1,t9,t5,tb} + vperm \b3,\b0,\b1,29 // vcprm(1,3,s0,s2) // {t2,ta,t6,tc} + vsubfp \b0,\a2,\b2 // {r4,r5,r6,r7} + vsubfp \b1,\a3,\b3 // {i4,i5,i6,i7} + vaddfp \a0,\a2,\b2 // {r0,r1,r2,r3} + vaddfp \a1,\a3,\b3 // {i0,i1,i2,i3} +.endm + +.macro BF d0,d1,s0,s1 + vsubfp \d1,\s0,\s1 + vaddfp \d0,\s0,\s1 +.endm + +fft4_altivec: + lvx 0,0,3 + lvx 1,11,3 + FFT4 0,1,2,3 + stvx 2,0,3 + stvx 3,11,3 + blr + +fft8_altivec: + addi 4,3,32 + lvx 0,0,3 + lvx 1,11,3 + lvx 2,0,4 + lvx 3,11,4 + FFT8 0,1,2,3,4,5,6,7,8 + stvx 0,0,3 + stvx 1,11,3 + stvx 2,0,4 + stvx 3,11,4 + blr + +fft16_altivec: + addi 5,3,64 + addi 6,3,96 + addi 4,3,32 + lvx 0,0,5 + lvx 1,11,5 + lvx 2,0,6 + lvx 3,11,6 + FFT4x2 0,1,2,3,4,5,6,7 + lvx 0,0,3 + lvx 1,11,3 + lvx 2,0,4 + lvx 3,11,4 + FFT8 0,1,2,3,8,9,10,11,12 + vmaddfp 8,4,15,14 // r2*wre + vmaddfp 9,5,15,14 // i2*wre + vmaddfp 10,6,15,14 // r3*wre + vmaddfp 11,7,15,14 // i3*wre + vmaddfp 8,5,16,8 // i2*wim + vnmsubfp 9,4,16,9 // r2*wim + vnmsubfp 10,7,16,10 // i3*wim + vmaddfp 11,6,16,11 // r3*wim + BF 10,12,10,8 + BF 11,13,9,11 + BF 0,4,0,10 + BF 3,7,3,12 + stvx 0,0,3 + stvx 4,0,5 + stvx 3,11,4 + stvx 7,11,6 + BF 1,5,1,11 + BF 2,6,2,13 + stvx 1,11,3 + stvx 5,11,5 + stvx 2,0,4 + stvx 6,0,6 + blr + +// void pass(float *z, float *wre, int n) +.macro PASS interleave +fft_pass\interleave\()_altivec: + mtctr 5 + slwi 10,5,4 + slwi 7,5,6 // o2 + slwi 5,5,5 // o1 + add 9,5,7 // o3 + add 10,4,10 // wim + addi 6,5,16 // o1+16 + addi 8,7,16 // o2+16 + addi 0,9,16 // o3+16 +1: + lvx 8,0,4 // wre + lvx 10,0,10 + subi 10,10,16 + lvx 9,0,10 + vperm 9,9,10,19 // vcprm(s0,3,2,1) => wim[0 .. -3] + lvx 4,3,7 // r2 = z[o2] + lvx 5,3,8 // i2 = z[o2+16] + lvx 6,3,9 // r3 = z[o3] + lvx 7,3,0 // i3 = z[o3+16] + vmaddfp 10,4,8,14 // r2*wre + vmaddfp 11,5,8,14 // i2*wre + vmaddfp 12,6,8,14 // r3*wre + vmaddfp 13,7,8,14 // i3*wre + lvx 0,0,3 // r0 = z[0] + lvx 3,3,6 // i1 = z[o1+16] + vmaddfp 10,5,9,10 // i2*wim + vnmsubfp 11,4,9,11 // r2*wim + vnmsubfp 12,7,9,12 // i3*wim + vmaddfp 13,6,9,13 // r3*wim + lvx 1,3,11 // i0 = z[16] + lvx 2,3,5 // r1 = z[o1] + BF 12,8,12,10 + BF 13,9,11,13 + BF 0,4,0,12 + BF 3,7,3,8 +.ifb \interleave + stvx 0,0,3 + stvx 4,3,7 + stvx 3,3,6 + stvx 7,3,0 +.endif + BF 1,5,1,13 + BF 2,6,2,9 +.ifb \interleave + stvx 1,3,11 + stvx 2,3,5 + stvx 5,3,8 + stvx 6,3,9 +.else + vmrghw 8,0,1 + vmrglw 9,0,1 + stvx 8,0,3 + stvx 9,3,11 + vmrghw 8,2,3 + vmrglw 9,2,3 + stvx 8,3,5 + stvx 9,3,6 + vmrghw 8,4,5 + vmrglw 9,4,5 + stvx 8,3,7 + stvx 9,3,8 + vmrghw 8,6,7 + vmrglw 9,6,7 + stvx 8,3,9 + stvx 9,3,0 +.endif + addi 3,3,32 + addi 4,4,16 + bdnz 1b + sub 3,3,5 + blr +.endm + +.macro DECL_FFT ver, bits, n, n2, n4 +fft\n\ver\()_altivec: + mflr 0 + STORE_PTR 0,1,\bits-5 + bl fft\n2\()_altivec + addi2 3,\n*4 + bl fft\n4\()_altivec + addi2 3,\n*2 + bl fft\n4\()_altivec + addi2 3,\n*-6 + LOAD_PTR 0,1,\bits-5 + LOAD_PTR 4,12,\bits-4 + mtlr 0 + li 5,\n/16 + b fft_pass\ver\()_altivec +.endm + +.macro DECL_FFTS ver +PASS \ver +DECL_FFT \ver,5,32,16,8 +DECL_FFT \ver,6,64,32,16 +DECL_FFT \ver,7,128,64,32 +DECL_FFT \ver,8,256,128,64 +DECL_FFT \ver,9,512,256,128 +DECL_FFT \ver,10,1024,512,256 +DECL_FFT \ver,11,2048,1024,512 +DECL_FFT \ver,12,4096,2048,1024 +DECL_FFT \ver,13,8192,4096,2048 +DECL_FFT \ver,14,16384,8192,4096 +DECL_FFT \ver,15,32768,16384,8192 +DECL_FFT \ver,16,65536,32768,16384 +.rodata +.global ff_fft_dispatch\ver\()_altivec +ff_fft_dispatch\ver\()_altivec: +.long fft4_altivec, fft8_altivec, fft16_altivec, fft32\ver\()_altivec,\ + fft64\ver\()_altivec, fft128\ver\()_altivec, fft256\ver\()_altivec,\ + fft512\ver\()_altivec, fft1024\ver\()_altivec, fft2048\ver\()_altivec,\ + fft4096\ver\()_altivec, fft8192\ver\()_altivec, fft16384\ver\()_altivec,\ + fft32768\ver\()_altivec, fft65536\ver\()_altivec +.text +.endm + +DECL_FFTS +DECL_FFTS _interleave diff --git a/libavcodec/ppc/types_altivec.h b/libavcodec/ppc/types_altivec.h index 2870e83..36b6e1f 100644 --- a/libavcodec/ppc/types_altivec.h +++ b/libavcodec/ppc/types_altivec.h @@ -30,6 +30,7 @@ #define vec_s16 vector signed short #define vec_u32 vector unsigned int #define vec_s32 vector signed int +#define vec_f vector float /*********************************************************************** * Null vector