Loren Merritt <lorenm at u.washington.edu> writes:
Now without Apple workarounds.
--Loren Merritt
From 100bf86add9537c20f4db2c5d2eeb2880c9c705e Mon Sep 17 00:00:00 2001 From: Loren Merritt <pengvado at akuvian.org> Date: Sat, 22 Aug 2009 11:22:09 +0100 Subject: [PATCH 1/3] altivec split-radix FFT 1.8x faster than altivec radix-2 on a G4 8% faster vorbis decoding
--- configure | 2 + libavcodec/Makefile | 4 +- libavcodec/fft.c | 2 +- libavcodec/ppc/asm.S | 59 ++++++++ libavcodec/ppc/fft_altivec.c | 152 +++++++------------- libavcodec/ppc/fft_altivec_s.S | 322 ++++++++++++++++++++++++++++++++++++++++ libavcodec/ppc/types_altivec.h | 1 + 7 files changed, 440 insertions(+), 102 deletions(-) create mode 100644 libavcodec/ppc/asm.S create mode 100644 libavcodec/ppc/fft_altivec_s.S
diff --git a/configure b/configure index 6aa6177..fdb6ade 100755 --- a/configure +++ b/configure @@ -933,6 +933,7 @@ HAVE_LIST=" fast_cmov fast_unaligned fork + fsf_gas gethrtime GetProcessTimes getrusage @@ -2108,6 +2109,7 @@ elif enabled mips; then elif enabled ppc; then
check_asm dcbzl '"dcbzl 0, 1"' + check_asm fsf_gas '"add 0, 0, 0"' check_asm ppc4xx '"maclhw r10, r11, r12"' check_asm xform_asm '"lwzx 0, %y0" :: "Z"(*(int*)0)'
This is the syntax the IBM PPC documentation uses, so branding it fsf seems a bit wrong.
diff --git a/libavcodec/fft.c b/libavcodec/fft.c index c827139..52bc73a 100644 --- a/libavcodec/fft.c +++ b/libavcodec/fft.c @@ -89,7 +89,7 @@ av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse) s->split_radix = 1;
if (ARCH_ARM) ff_fft_init_arm(s); - if (HAVE_ALTIVEC) ff_fft_init_altivec(s); + if (HAVE_ALTIVEC && HAVE_FSF_GAS) ff_fft_init_altivec(s); if (HAVE_MMX) ff_fft_init_mmx(s);
if (s->split_radix) {
I'd rather hide this condition inside ff_fft_init_altivec().
diff --git a/libavcodec/ppc/fft_altivec.c b/libavcodec/ppc/fft_altivec.c index 83b2b7f..235f205 100644 --- a/libavcodec/ppc/fft_altivec.c +++ b/libavcodec/ppc/fft_altivec.c [...] +extern FFTSample ff_cos_16[];
Isn't this declared in some header file? If not, it should be.
+// pointers to functions. but unlike function pointers on some PPC ABIs, these aren't function descriptors.
What does that mean? Not saying it's wrong, I just don't understand it.
From fc014fba0a5c39e47d2cb651856d1b1e83c4edf9 Mon Sep 17 00:00:00 2001 From: Loren Merritt <pengvado at akuvian.org> Date: Wed, 26 Aug 2009 09:44:32 +0100 Subject: [PATCH 2/3] remove vestiges of radix-2 FFT
--- libavcodec/dsputil.h | 3 - libavcodec/fft.c | 83 +++--------------------------------------------------- libavcodec/ppc/fft_altivec.c | 1 3 files changed, 5 insertions(+), 82 deletions(-)
diff --git a/libavcodec/dsputil.h b/libavcodec/dsputil.h index d100d5f..901f8a7 100644 --- a/libavcodec/dsputil.h +++ b/libavcodec/dsputil.h @@ -675,15 +675,12 @@ typedef struct FFTContext { int nbits; int inverse; uint16_t *revtab; - FFTComplex *exptab; - FFTComplex *exptab1; /* only used by SSE code */
I need to update the offsets in the NEON code for this, so don't commit without telling me first.
diff --git a/libavcodec/fft.c b/libavcodec/fft.c index 52bc73a..a7c7b96 100644 --- a/libavcodec/fft.c +++ b/libavcodec/fft.c @@ -60,39 +60,30 @@ static int split_radix_permutation(int i, int n, int inverse)
av_cold int ff_fft_init(FFTContext *s, int nbits, int inverse) { - int i, j, m, n; - float alpha, c1, s1, s2; - int av_unused has_vectors; + int i, j, n;
if (nbits < 2 || nbits > 16) goto fail; s->nbits = nbits; n = 1 << nbits;
- s->tmp_buf = NULL; - s->exptab = av_malloc((n / 2) * sizeof(FFTComplex)); - if (!s->exptab) - goto fail; s->revtab = av_malloc(n * sizeof(uint16_t)); if (!s->revtab) goto fail; + s->tmp_buf = av_malloc(n * sizeof(FFTComplex)); + if (!s->tmp_buf) + goto fail; s->inverse = inverse;
- s2 = inverse ? 1.0 : -1.0; - s->fft_permute = ff_fft_permute_c; s->fft_calc = ff_fft_calc_c; s->imdct_calc = ff_imdct_calc_c; s->imdct_half = ff_imdct_half_c; - s->mdct_calc = ff_mdct_calc_c;
I think that line should stay.
- s->exptab1 = NULL; - s->split_radix = 1;
if (ARCH_ARM) ff_fft_init_arm(s); if (HAVE_ALTIVEC && HAVE_FSF_GAS) ff_fft_init_altivec(s); if (HAVE_MMX) ff_fft_init_mmx(s);
diff --git a/libavcodec/ppc/fft_altivec.c b/libavcodec/ppc/fft_altivec.c index 235f205..9f01992 100644 --- a/libavcodec/ppc/fft_altivec.c +++ b/libavcodec/ppc/fft_altivec.c @@ -89,5 +89,4 @@ void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) av_cold void ff_fft_init_altivec(FFTContext *s) { s->fft_calc = ff_fft_calc_altivec; - s->split_radix = 1; }
From e2e5f93792d62becaa666723ce26f81472e2eddc Mon Sep 17 00:00:00 2001 From: Loren Merritt <pengvado at akuvian.org> Date: Wed, 26 Aug 2009 10:08:06 +0100 Subject: [PATCH 3/3] altivec imdct 1.8x faster than C iMDCT (excluding the FFT part) on a G4 10% faster vorbis decoding
--- libavcodec/ppc/fft_altivec.c | 113 ++++++++++++++++++++++++++++++++++++++++- 1 files changed, 110 insertions(+), 3 deletions(-)
diff --git a/libavcodec/ppc/fft_altivec.c b/libavcodec/ppc/fft_altivec.c index 9f01992..0d4ced3 100644 --- a/libavcodec/ppc/fft_altivec.c +++ b/libavcodec/ppc/fft_altivec.c @@ -49,7 +49,7 @@ static void swizzle(vec_f *z, int n) } }
-void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) +static av_always_inline void fft_dispatch(FFTContext *s, FFTComplex *z, int do_swizzle) { register vec_f v14 __asm__("v14") = {0,0,0,0}; register vec_f v15 __asm__("v15") = *(const vec_f*)ff_cos_16; @@ -75,18 +75,125 @@ void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) "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"(ff_fft_dispatch_altivec[do_swizzle][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","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) + if (do_swizzle && s->nbits <= 4) swizzle((vec_f*)z, 1<<s->nbits); }
+static void ff_fft_calc_altivec(FFTContext *s, FFTComplex *z) +{ + fft_dispatch(s, z, 1); +} + +static void ff_imdct_half_altivec(MDCTContext *s, FFTSample *output, const FFTSample *input) +{ + int j, k; + int n = 1 << s->nbits; + int n4 = n >> 2; + int n8 = n >> 3; + int n32 = n >> 5; + const uint16_t *revtabj = s->fft.revtab; + const uint16_t *revtabk = s->fft.revtab+n4; + const vec_f *tcos = (const vec_f*)(s->tcos+n8); + const vec_f *tsin = (const vec_f*)(s->tsin+n8); + const vec_f *pin = (const vec_f*)(input+n4); + vec_f *pout = (vec_f*)(output+n4);
Why the intrinsics here? -- M?ns Rullg?rd mans at mansr.com