[FFmpeg-devel] [PATCH 1/2] Optimization of AC3 floating point decoder for MIPS
Babic, Nedeljko
nbabic at mips.com
Mon Jun 25 10:59:37 CEST 2012
Hi,
>> libavcodec/dsputil.c | 1 +
>> libavcodec/dsputil.h | 1 +
>> libavcodec/fft.c | 1 +
>> libavcodec/fft.h | 11 +
>> libavcodec/fmtconvert.c | 1 +
>> libavcodec/fmtconvert.h | 1 +
>> libavcodec/mips/Makefile | 4 +
>> libavcodec/mips/dsputil_mips.c | 168 +++++++++
>> libavcodec/mips/fft_mips.c | 689 +++++++++++++++++++++++++++++++++++++
>> libavcodec/mips/fft_table.h | 482 ++++++++++++++++++++++++++
>> libavcodec/mips/fmtconvert_mips.c | 336 ++++++++++++++++++
>> 11 files changed, 1695 insertions(+), 0 deletions(-)
>> create mode 100644 libavcodec/mips/dsputil_mips.c
>> create mode 100644 libavcodec/mips/fft_mips.c
>> create mode 100644 libavcodec/mips/fft_table.h
>> create mode 100644 libavcodec/mips/fmtconvert_mips.c
>>
>> diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
>> index 442b900..b7d928f 100644
>> --- a/libavcodec/dsputil.c
>> +++ b/libavcodec/dsputil.c
>> @@ -3161,6 +3161,7 @@ av_cold void ff_dsputil_init(DSPContext* c, AVCodecContext *avctx)
>> if (HAVE_MMI) ff_dsputil_init_mmi (c, avctx);
>> if (ARCH_SH4) ff_dsputil_init_sh4 (c, avctx);
>> if (ARCH_BFIN) ff_dsputil_init_bfin (c, avctx);
>> + if (HAVE_MIPSFPU) ff_dsputil_init_mips (c, avctx);
>
>> --- a/libavcodec/fft.h
>> +++ b/libavcodec/fft.h
>> @@ -38,6 +38,16 @@
>>
>> typedef float FFTDouble;
>>
>> +#if ARCH_MIPS
>> +enum _fftConsts{
>> + MIN_LOG2_NFFT = 5, //!< Specifies miniumum allowed fft size
>> + MAX_LOG2_NFFT = 12 //!< Specifies maxiumum allowed fft size
>> +};
>> +
>> +#define MAX_FFT_SIZE (1<< MAX_LOG2_NFFT)
>> +#define MIN_FFT_SIZE (1<< MAX_LOG2_NFFT)
>> +
>> +#endif
>
>MIPS-specific code should not be in common code.
I will place this in appropriate MIPS file.
>
>> diff --git a/libavcodec/mips/fft_mips.c b/libavcodec/mips/fft_mips.c
>> new file mode 100644
>
>Nice, can you post the benchmarks results of "fft-
>test -s"?
Posted below
>
>> index 0000000..286c67f
>> --- /dev/null
>> +++ b/libavcodec/mips/fft_mips.c
>> @@ -0,0 +1,689 @@
>> +/*
>> + * Copyright (c) 2012
>> + * MIPS Technologies, Inc., California.
>> + *
>> + * Redistribution and use in source and binary forms, with or without
>> + * modification, are permitted provided that the following conditions
>> + * are met:
>> + * 1. Redistributions of source code must retain the above copyright
>> + * notice, this list of conditions and the following disclaimer.
>> + * 2. Redistributions in binary form must reproduce the above copyright
>> + * notice, this list of conditions and the following disclaimer in the
>> + * documentation and/or other materials provided with the distribution.
>> + * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
>> + * contributors may be used to endorse or promote products derived from
>> + * this software without specific prior written permission.
>> + *
>> + * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
>> + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
>> + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
>> + * ARE DISCLAIMED. IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
>> + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
>> + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
>> + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
>> + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
>> + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
>> + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
>> + * SUCH DAMAGE.
>> + *
>> + * Author: Stanisalv Ocovaj (socovaj at mips.com)
>> + * Author: Zoran Lukic (zoranl at mips.com)
>> + *
>> + * Optimized MDCT/IMDCT and FFT transforms
>> + *
>> + * 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
>> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
>> + */
>> +#include "config.h"
>> +#include "libavcodec/fft.h"
>> +#include "fft_table.h"
>> +
>> +/**
>> + * FFT transform
>> + */
>> +
>> +#if HAVE_INLINE_ASM
>> +static void ff_fft_calc_mips(FFTContext *s, FFTComplex *z) {
>> +
>> + int nbits, i, n, num_transforms, offset, step;
>> + int n4, n2, n34;
>> + FFTSample tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
>> + FFTComplex *tmpz;
>> + float w_re, w_im;
>> + float *w_re_ptr;
>> + const int fft_size = (1<< s->nbits);
>> + int s_n = s->nbits;
>> + int tem1, tem2;
>> + float pom, pom1, pom2, pom3;
>> + float temp, temp1, temp3, temp4;
>> + FFTComplex * tmpz_n2, * tmpz_n34, * tmpz_n4;
>> + FFTComplex * tmpz_n2_i, * tmpz_n34_i, * tmpz_n4_i, * tmpz_i;
>> +
>> + /**
>> + *num_transforms = (0x2aab>> (16 - s->nbits)) | 1;
>> + */
>> + __asm__ __volatile__ (
>> + "li %[tem1], 16 \n\t"
>> + "sub %[s_n], %[tem1], %[s_n] \n\t"
>> + "li %[tem2], 10923 \n\t"
>> + "srav %[tem2], %[tem2], %[s_n] \n\t"
>> + "ori %[num_t],%[tem2], 1 \n\t"
>> + : [num_t]"=r"(num_transforms), [s_n]"+r"(s_n),
>> + [tem1]"=&r"(tem1), [tem2]"=&r"(tem2)
>> + );
>> +
>> +
>> + for (n=0; n<num_transforms; n++)
>> + {
>> + offset = fft_offsets_lut[n]<< 2;
>> + tmpz = z + offset;
>
>What is the point of this LUT? If you want your input permutated in some
>particular order, you can just use revtab struct.
We use this LUT because of the somewhat irregular (when compared to radix-2 FFT) structure
of split-radix FFT. It tells us at which starting offsets the sub-transforms of particular
size have to be performed. It has nothing to do with input permutation, which is the same
as in the original algorithm (we use the same revtab for that).
>
>> +
>> + tmp1 = tmpz[0].re + tmpz[1].re;
>> + tmp5 = tmpz[2].re + tmpz[3].re;
>> + tmp2 = tmpz[0].im + tmpz[1].im;
>> + tmp6 = tmpz[2].im + tmpz[3].im;
>> + tmp3 = tmpz[0].re - tmpz[1].re;
>> + tmp8 = tmpz[2].im - tmpz[3].im;
>> + tmp4 = tmpz[0].im - tmpz[1].im;
>> + tmp7 = tmpz[2].re - tmpz[3].re;
>> +
>> + tmpz[0].re = tmp1 + tmp5;
>> + tmpz[2].re = tmp1 - tmp5;
>> + tmpz[0].im = tmp2 + tmp6;
>> + tmpz[2].im = tmp2 - tmp6;
>> + tmpz[1].re = tmp3 + tmp8;
>> + tmpz[3].re = tmp3 - tmp8;
>> + tmpz[1].im = tmp4 - tmp7;
>> + tmpz[3].im = tmp4 + tmp7;
>> +
>> +}
>> +
>> + if (fft_size< 8)
>> + return;
>> +
>> + num_transforms = (num_transforms>> 1) | 1;
>> + for (n=0; n<num_transforms; n++)
>> + {
>> + offset = fft_offsets_lut[n]<< 3;
>> + tmpz = z + offset;
>> +
>> + __asm__ __volatile__ (
>> + "lwc1 %[tmp1], 32(%[tmpz]) \n\t"
>> + "lwc1 %[pom], 40(%[tmpz]) \n\t"
>> + "lwc1 %[tmp3], 48(%[tmpz]) \n\t"
>> + "lwc1 %[pom1], 56(%[tmpz]) \n\t"
>> + "lwc1 %[tmp2], 36(%[tmpz]) \n\t"
>> + "lwc1 %[pom2], 44(%[tmpz]) \n\t"
>> + "lwc1 %[pom3], 60(%[tmpz]) \n\t"
>> + "lwc1 %[tmp4], 52(%[tmpz]) \n\t"
>> + "add.s %[tmp1], %[tmp1], %[pom] \n\t" // tmp1 = tmpz[4].re + tmpz[5].re;
>> + "add.s %[tmp3], %[tmp3], %[pom1] \n\t" // tmp3 = tmpz[6].re + tmpz[7].re;
>> + "add.s %[tmp2], %[tmp2], %[pom2] \n\t" // tmp2 = tmpz[4].im + tmpz[5].im;
>> + "lwc1 %[pom], 40(%[tmpz]) \n\t"
>> + "add.s %[tmp4], %[tmp4], %[pom3] \n\t" // tmp4 = tmpz[6].im + tmpz[7].im;
>> + "add.s %[tmp5], %[tmp1], %[tmp3] \n\t" // tmp5 = tmp1 + tmp3;
>> + "sub.s %[tmp7], %[tmp1], %[tmp3] \n\t" // tmp7 = tmp1 - tmp3;
>> + "lwc1 %[tmp1], 32(%[tmpz]) \n\t"
>> + "lwc1 %[pom1], 44(%[tmpz]) \n\t"
>> + "add.s %[tmp6], %[tmp2], %[tmp4] \n\t" // tmp6 = tmp2 + tmp4;
>> + "sub.s %[tmp8], %[tmp2], %[tmp4] \n\t" // tmp8 = tmp2 - tmp4;
>> + "lwc1 %[tmp2], 36(%[tmpz]) \n\t"
>> + "lwc1 %[pom2], 56(%[tmpz]) \n\t"
>> + "lwc1 %[pom3], 60(%[tmpz]) \n\t"
>> + "lwc1 %[tmp3], 48(%[tmpz]) \n\t"
>> + "lwc1 %[tmp4], 52(%[tmpz]) \n\t"
>> + "sub.s %[tmp1], %[tmp1], %[pom] \n\t" // tmp1 = tmpz[4].re - tmpz[5].re;
>> + "lwc1 %[pom], 0(%[tmpz]) \n\t"
>> + "sub.s %[tmp2], %[tmp2], %[pom1] \n\t" // tmp2 = tmpz[4].im - tmpz[5].im;
>> + "sub.s %[tmp3], %[tmp3], %[pom2] \n\t" // tmp3 = tmpz[6].re - tmpz[7].re;
>> + "lwc1 %[pom2], 4(%[tmpz]) \n\t"
>> + "sub.s %[pom1], %[pom], %[tmp5] \n\t"
>> + "sub.s %[tmp4], %[tmp4], %[pom3] \n\t" // tmp4 = tmpz[6].im - tmpz[7].im;
>> + "add.s %[pom3], %[pom], %[tmp5] \n\t"
>> + "sub.s %[pom], %[pom2], %[tmp6] \n\t"
>> + "add.s %[pom2], %[pom2], %[tmp6] \n\t"
>> + "swc1 %[pom1], 32(%[tmpz]) \n\t" // tmpz[4].re = tmpz[0].re - tmp5;
>> + "swc1 %[pom3], 0(%[tmpz]) \n\t" // tmpz[0].re = tmpz[0].re + tmp5;
>> + "swc1 %[pom], 36(%[tmpz]) \n\t" // tmpz[4].im = tmpz[0].im - tmp6;
>> + "swc1 %[pom2], 4(%[tmpz]) \n\t" // tmpz[0].im = tmpz[0].im + tmp6;
>> + "lwc1 %[pom1], 16(%[tmpz]) \n\t"
>> + "lwc1 %[pom3], 20(%[tmpz]) \n\t"
>> + "li.s %[pom], 0.7071067812 \n\t" // float pom = 0.7071067812f;
>> + "add.s %[temp1],%[tmp1], %[tmp2] \n\t"
>> + "sub.s %[temp], %[pom1], %[tmp8] \n\t"
>> + "add.s %[pom2], %[pom3], %[tmp7] \n\t"
>> + "sub.s %[temp3],%[tmp3], %[tmp4] \n\t"
>> + "sub.s %[temp4],%[tmp2], %[tmp1] \n\t"
>> + "swc1 %[temp], 48(%[tmpz]) \n\t" // tmpz[6].re = tmpz[2].re - tmp8;
>> + "swc1 %[pom2], 52(%[tmpz]) \n\t" // tmpz[6].im = tmpz[2].im + tmp7;
>> + "add.s %[pom1], %[pom1], %[tmp8] \n\t"
>> + "sub.s %[pom3], %[pom3], %[tmp7] \n\t"
>> + "add.s %[tmp3], %[tmp3], %[tmp4] \n\t"
>> + "mul.s %[tmp5], %[pom], %[temp1] \n\t" // tmp5 = pom * (tmp1 + tmp2);
>> + "mul.s %[tmp7], %[pom], %[temp3] \n\t" // tmp7 = pom * (tmp3 - tmp4);
>> + "mul.s %[tmp6], %[pom], %[temp4] \n\t" // tmp6 = pom * (tmp2 - tmp1);
>> + "mul.s %[tmp8], %[pom], %[tmp3] \n\t" // tmp8 = pom * (tmp3 + tmp4);
>> + "swc1 %[pom1], 16(%[tmpz]) \n\t" // tmpz[2].re = tmpz[2].re + tmp8;
>> + "swc1 %[pom3], 20(%[tmpz]) \n\t" // tmpz[2].im = tmpz[2].im - tmp7;
>> + "add.s %[tmp1], %[tmp5], %[tmp7] \n\t" // tmp1 = tmp5 + tmp7;
>> + "sub.s %[tmp3], %[tmp5], %[tmp7] \n\t" // tmp3 = tmp5 - tmp7;
>> + "add.s %[tmp2], %[tmp6], %[tmp8] \n\t" // tmp2 = tmp6 + tmp8;
>> + "sub.s %[tmp4], %[tmp6], %[tmp8] \n\t" // tmp4 = tmp6 - tmp8;
>> + "lwc1 %[temp], 8(%[tmpz]) \n\t"
>> + "lwc1 %[temp1],12(%[tmpz]) \n\t"
>> + "lwc1 %[pom], 24(%[tmpz]) \n\t"
>> + "lwc1 %[pom2], 28(%[tmpz]) \n\t"
>> + "sub.s %[temp4],%[temp], %[tmp1] \n\t"
>> + "sub.s %[temp3],%[temp1], %[tmp2] \n\t"
>> + "add.s %[temp], %[temp], %[tmp1] \n\t"
>> + "add.s %[temp1],%[temp1], %[tmp2] \n\t"
>> + "sub.s %[pom1], %[pom], %[tmp4] \n\t"
>> + "add.s %[pom3], %[pom2], %[tmp3] \n\t"
>> + "add.s %[pom], %[pom], %[tmp4] \n\t"
>> + "sub.s %[pom2], %[pom2], %[tmp3] \n\t"
>> + "swc1 %[temp4],40(%[tmpz]) \n\t" // tmpz[5].re = tmpz[1].re - tmp1;
>> + "swc1 %[temp3],44(%[tmpz]) \n\t" // tmpz[5].im = tmpz[1].im - tmp2;
>> + "swc1 %[temp], 8(%[tmpz]) \n\t" // tmpz[1].re = tmpz[1].re + tmp1;
>> + "swc1 %[temp1],12(%[tmpz]) \n\t" // tmpz[1].im = tmpz[1].im + tmp2;
>> + "swc1 %[pom1], 56(%[tmpz]) \n\t" // tmpz[7].re = tmpz[3].re - tmp4;
>> + "swc1 %[pom3], 60(%[tmpz]) \n\t" // tmpz[7].im = tmpz[3].im + tmp3;
>> + "swc1 %[pom], 24(%[tmpz]) \n\t" // tmpz[3].re = tmpz[3].re + tmp4;
>> + "swc1 %[pom2], 28(%[tmpz]) \n\t" // tmpz[3].im = tmpz[3].im - tmp3;
>> + : [tmpz]"+r"(tmpz), [tmp1]"=f"(tmp1), [pom]"=f"(pom), [pom1]"=&f"(pom1), [pom2]"=&f"(pom2),
>> + [tmp3]"=f"(tmp3), [tmp2]"=f"(tmp2), [tmp4]"=f"(tmp4), [tmp5]"=f"(tmp5), [tmp7]"=f"(tmp7),
>> + [tmp6]"=f"(tmp6), [tmp8]"=f"(tmp8), [pom3]"=&f"(pom3),[temp]"=&f"(temp), [temp1]"=&f"(temp1),
>> + [temp3]"=&f"(temp3), [temp4]"=&f"(temp4)
>> + :
>> + : "memory"
>> + );
>> + }
>> +
>> + step = 1<< (MAX_LOG2_NFFT - 4);
>> + n4 = 4;
>> + for (nbits=4; nbits<=s->nbits; nbits++)
>> + {
>> + /*
>> + * num_transforms = (num_transforms>> 1) | 1;
>> + */
>> + __asm__ __volatile__ (
>> + "sra %[num_t], %[num_t], 1 \n\t"
>> + "ori %[num_t], %[num_t], 1 \n\t"
>> +
>> + : [num_t] "+r" (num_transforms)
>> + );
>> + n2 = 2 * n4;
>> + n34 = 3 * n4;
>> +
>> + for (n=0; n<num_transforms; n++)
>> + {
>> + offset = fft_offsets_lut[n]<< nbits;
>> + tmpz = z + offset;
>> +
>> + tmpz_n2 = tmpz + n2;
>> + tmpz_n4 = tmpz + n4;
>> + tmpz_n34 = tmpz + n34;
>> +
>> + __asm__ __volatile__ (
>> + "lwc1 %[pom1], 0(%[tmpz_n2]) \n\t"
>> + "lwc1 %[pom], 0(%[tmpz_n34]) \n\t"
>> + "lwc1 %[pom2], 4(%[tmpz_n2]) \n\t"
>> + "lwc1 %[pom3], 4(%[tmpz_n34]) \n\t"
>> + "lwc1 %[temp1],0(%[tmpz]) \n\t"
>> + "lwc1 %[temp3],4(%[tmpz]) \n\t"
>> + "add.s %[tmp5], %[pom1], %[pom] \n\t" // tmp5 = tmpz[ n2].re + tmpz[n34].re;
>> + "sub.s %[tmp1], %[pom1], %[pom] \n\t" // tmp1 = tmpz[ n2].re - tmpz[n34].re;
>> + "add.s %[tmp6], %[pom2], %[pom3] \n\t" // tmp6 = tmpz[ n2].im + tmpz[n34].im;
>> + "sub.s %[tmp2], %[pom2], %[pom3] \n\t" // tmp2 = tmpz[ n2].im - tmpz[n34].im;
>> + "sub.s %[temp], %[temp1], %[tmp5] \n\t"
>> + "add.s %[temp1],%[temp1], %[tmp5] \n\t"
>> + "sub.s %[temp4],%[temp3], %[tmp6] \n\t"
>> + "add.s %[temp3],%[temp3], %[tmp6] \n\t"
>> + "swc1 %[temp], 0(%[tmpz_n2]) \n\t" // tmpz[ n2].re = tmpz[ 0].re - tmp5;
>> + "swc1 %[temp1],0(%[tmpz]) \n\t" // tmpz[ 0].re = tmpz[ 0].re + tmp5;
>> + "lwc1 %[pom1], 0(%[tmpz_n4]) \n\t"
>> + "swc1 %[temp4],4(%[tmpz_n2]) \n\t" // tmpz[ n2].im = tmpz[ 0].im - tmp6;
>> + "lwc1 %[temp], 4(%[tmpz_n4]) \n\t"
>> + "swc1 %[temp3],4(%[tmpz]) \n\t" // tmpz[ 0].im = tmpz[ 0].im + tmp6;
>> + "sub.s %[pom], %[pom1], %[tmp2] \n\t"
>> + "add.s %[pom1], %[pom1], %[tmp2] \n\t"
>> + "add.s %[temp1],%[temp], %[tmp1] \n\t"
>> + "sub.s %[temp], %[temp], %[tmp1] \n\t"
>> + "swc1 %[pom], 0(%[tmpz_n34]) \n\t" // tmpz[n34].re = tmpz[n4].re - tmp2;
>> + "swc1 %[pom1], 0(%[tmpz_n4]) \n\t" // tmpz[ n4].re = tmpz[n4].re + tmp2;
>> + "swc1 %[temp1],4(%[tmpz_n34]) \n\t" // tmpz[n34].im = tmpz[n4].im + tmp1;
>> + "swc1 %[temp], 4(%[tmpz_n4]) \n\t" // tmpz[ n4].im = tmpz[n4].im - tmp1;
>> + : [tmpz]"+r"(tmpz), [tmpz_n2]"+r"(tmpz_n2), [tmpz_n34]"+r"(tmpz_n34), [tmp5]"=f"(tmp5),
>> + [tmp1]"=f"(tmp1), [pom]"=&f"(pom), [pom1]"=&f"(pom1), [pom2]"=&f"(pom2),
>> + [tmp2]"=f"(tmp2), [tmp6]"=f"(tmp6), [tmpz_n4]"+r"(tmpz_n4), [pom3]"=&f"(pom3),
>> + [temp]"=f"(temp), [temp1]"=f"(temp1), [temp3]"=f"(temp3), [temp4]"=f"(temp4)
>> + :
>> + : "memory"
>> + );
>> +
>> + w_re_ptr = w_tab + step;
>> +
>> + for (i=1; i<n4; i++)
>> + {
>> + w_re = w_re_ptr[0];
>> + w_im = w_re_ptr[MAX_FFT_SIZE/4];
>
>Can you explain why you cannot use the same cos/sin tab that the C
>version uses?
We can use them and I will rewrite this part of code to use them.
========================
fft-test -s results on MIPS 74Kf board:
========================
original fft:
FFT 512 test
Checking...
max:0.000008 e:3.92148e-08
Speed test...
time: 136.9 us/transform [total time=1.12 s its=8192]
========================================================
optimized fft:
FFT 512 test
Checking...
max:0.000005 e:3.86258e-08
Speed test...
time: 89.7 us/transform [total time=1.47 s its=16384]
-Nedeljko
________________________________________
From: Vitor Sessak [vitor1001 at gmail.com]
Sent: Thursday, June 21, 2012 14:26
To: FFmpeg development discussions and patches
Cc: Babic, Nedeljko; Lukac, Zeljko
Subject: Re: [FFmpeg-devel] [PATCH 1/2] Optimization of AC3 floating point decoder for MIPS
Hi
On 06/21/2012 12:04 PM, Nedeljko Babic wrote:
> FFT in MIPS implementation is working iteratively instead
> of "recursively" calling functions for smaller FFT sizes.
> Some of DSP and format convert utils functions are also optimized.
Some comments
> libavcodec/dsputil.c | 1 +
> libavcodec/dsputil.h | 1 +
> libavcodec/fft.c | 1 +
> libavcodec/fft.h | 11 +
> libavcodec/fmtconvert.c | 1 +
> libavcodec/fmtconvert.h | 1 +
> libavcodec/mips/Makefile | 4 +
> libavcodec/mips/dsputil_mips.c | 168 +++++++++
> libavcodec/mips/fft_mips.c | 689 +++++++++++++++++++++++++++++++++++++
> libavcodec/mips/fft_table.h | 482 ++++++++++++++++++++++++++
> libavcodec/mips/fmtconvert_mips.c | 336 ++++++++++++++++++
> 11 files changed, 1695 insertions(+), 0 deletions(-)
> create mode 100644 libavcodec/mips/dsputil_mips.c
> create mode 100644 libavcodec/mips/fft_mips.c
> create mode 100644 libavcodec/mips/fft_table.h
> create mode 100644 libavcodec/mips/fmtconvert_mips.c
>
> diff --git a/libavcodec/dsputil.c b/libavcodec/dsputil.c
> index 442b900..b7d928f 100644
> --- a/libavcodec/dsputil.c
> +++ b/libavcodec/dsputil.c
> @@ -3161,6 +3161,7 @@ av_cold void ff_dsputil_init(DSPContext* c, AVCodecContext *avctx)
> if (HAVE_MMI) ff_dsputil_init_mmi (c, avctx);
> if (ARCH_SH4) ff_dsputil_init_sh4 (c, avctx);
> if (ARCH_BFIN) ff_dsputil_init_bfin (c, avctx);
> + if (HAVE_MIPSFPU) ff_dsputil_init_mips (c, avctx);
> --- a/libavcodec/fft.h
> +++ b/libavcodec/fft.h
> @@ -38,6 +38,16 @@
>
> typedef float FFTDouble;
>
> +#if ARCH_MIPS
> +enum _fftConsts{
> + MIN_LOG2_NFFT = 5, //!< Specifies miniumum allowed fft size
> + MAX_LOG2_NFFT = 12 //!< Specifies maxiumum allowed fft size
> +};
> +
> +#define MAX_FFT_SIZE (1<< MAX_LOG2_NFFT)
> +#define MIN_FFT_SIZE (1<< MAX_LOG2_NFFT)
> +
> +#endif
MIPS-specific code should not be in common code.
> diff --git a/libavcodec/mips/fft_mips.c b/libavcodec/mips/fft_mips.c
> new file mode 100644
Nice, can you post the benchmarks results of "fft-
test -s"?
> index 0000000..286c67f
> --- /dev/null
> +++ b/libavcodec/mips/fft_mips.c
> @@ -0,0 +1,689 @@
> +/*
> + * Copyright (c) 2012
> + * MIPS Technologies, Inc., California.
> + *
> + * Redistribution and use in source and binary forms, with or without
> + * modification, are permitted provided that the following conditions
> + * are met:
> + * 1. Redistributions of source code must retain the above copyright
> + * notice, this list of conditions and the following disclaimer.
> + * 2. Redistributions in binary form must reproduce the above copyright
> + * notice, this list of conditions and the following disclaimer in the
> + * documentation and/or other materials provided with the distribution.
> + * 3. Neither the name of the MIPS Technologies, Inc., nor the names of its
> + * contributors may be used to endorse or promote products derived from
> + * this software without specific prior written permission.
> + *
> + * THIS SOFTWARE IS PROVIDED BY THE MIPS TECHNOLOGIES, INC. ``AS IS'' AND
> + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
> + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
> + * ARE DISCLAIMED. IN NO EVENT SHALL THE MIPS TECHNOLOGIES, INC. BE LIABLE
> + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
> + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
> + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
> + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
> + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
> + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
> + * SUCH DAMAGE.
> + *
> + * Author: Stanisalv Ocovaj (socovaj at mips.com)
> + * Author: Zoran Lukic (zoranl at mips.com)
> + *
> + * Optimized MDCT/IMDCT and FFT transforms
> + *
> + * 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
> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
> + */
> +#include "config.h"
> +#include "libavcodec/fft.h"
> +#include "fft_table.h"
> +
> +/**
> + * FFT transform
> + */
> +
> +#if HAVE_INLINE_ASM
> +static void ff_fft_calc_mips(FFTContext *s, FFTComplex *z) {
> +
> + int nbits, i, n, num_transforms, offset, step;
> + int n4, n2, n34;
> + FFTSample tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
> + FFTComplex *tmpz;
> + float w_re, w_im;
> + float *w_re_ptr;
> + const int fft_size = (1<< s->nbits);
> + int s_n = s->nbits;
> + int tem1, tem2;
> + float pom, pom1, pom2, pom3;
> + float temp, temp1, temp3, temp4;
> + FFTComplex * tmpz_n2, * tmpz_n34, * tmpz_n4;
> + FFTComplex * tmpz_n2_i, * tmpz_n34_i, * tmpz_n4_i, * tmpz_i;
> +
> + /**
> + *num_transforms = (0x2aab>> (16 - s->nbits)) | 1;
> + */
> + __asm__ __volatile__ (
> + "li %[tem1], 16 \n\t"
> + "sub %[s_n], %[tem1], %[s_n] \n\t"
> + "li %[tem2], 10923 \n\t"
> + "srav %[tem2], %[tem2], %[s_n] \n\t"
> + "ori %[num_t],%[tem2], 1 \n\t"
> + : [num_t]"=r"(num_transforms), [s_n]"+r"(s_n),
> + [tem1]"=&r"(tem1), [tem2]"=&r"(tem2)
> + );
> +
> +
> + for (n=0; n<num_transforms; n++)
> + {
> + offset = fft_offsets_lut[n]<< 2;
> + tmpz = z + offset;
What is the point of this LUT? If you want your input permutated in some
particular order, you can just use revtab struct.
> +
> + tmp1 = tmpz[0].re + tmpz[1].re;
> + tmp5 = tmpz[2].re + tmpz[3].re;
> + tmp2 = tmpz[0].im + tmpz[1].im;
> + tmp6 = tmpz[2].im + tmpz[3].im;
> + tmp3 = tmpz[0].re - tmpz[1].re;
> + tmp8 = tmpz[2].im - tmpz[3].im;
> + tmp4 = tmpz[0].im - tmpz[1].im;
> + tmp7 = tmpz[2].re - tmpz[3].re;
> +
> + tmpz[0].re = tmp1 + tmp5;
> + tmpz[2].re = tmp1 - tmp5;
> + tmpz[0].im = tmp2 + tmp6;
> + tmpz[2].im = tmp2 - tmp6;
> + tmpz[1].re = tmp3 + tmp8;
> + tmpz[3].re = tmp3 - tmp8;
> + tmpz[1].im = tmp4 - tmp7;
> + tmpz[3].im = tmp4 + tmp7;
> +
> +}
> +
> + if (fft_size< 8)
> + return;
> +
> + num_transforms = (num_transforms>> 1) | 1;
> + for (n=0; n<num_transforms; n++)
> + {
> + offset = fft_offsets_lut[n]<< 3;
> + tmpz = z + offset;
> +
> + __asm__ __volatile__ (
> + "lwc1 %[tmp1], 32(%[tmpz]) \n\t"
> + "lwc1 %[pom], 40(%[tmpz]) \n\t"
> + "lwc1 %[tmp3], 48(%[tmpz]) \n\t"
> + "lwc1 %[pom1], 56(%[tmpz]) \n\t"
> + "lwc1 %[tmp2], 36(%[tmpz]) \n\t"
> + "lwc1 %[pom2], 44(%[tmpz]) \n\t"
> + "lwc1 %[pom3], 60(%[tmpz]) \n\t"
> + "lwc1 %[tmp4], 52(%[tmpz]) \n\t"
> + "add.s %[tmp1], %[tmp1], %[pom] \n\t" // tmp1 = tmpz[4].re + tmpz[5].re;
> + "add.s %[tmp3], %[tmp3], %[pom1] \n\t" // tmp3 = tmpz[6].re + tmpz[7].re;
> + "add.s %[tmp2], %[tmp2], %[pom2] \n\t" // tmp2 = tmpz[4].im + tmpz[5].im;
> + "lwc1 %[pom], 40(%[tmpz]) \n\t"
> + "add.s %[tmp4], %[tmp4], %[pom3] \n\t" // tmp4 = tmpz[6].im + tmpz[7].im;
> + "add.s %[tmp5], %[tmp1], %[tmp3] \n\t" // tmp5 = tmp1 + tmp3;
> + "sub.s %[tmp7], %[tmp1], %[tmp3] \n\t" // tmp7 = tmp1 - tmp3;
> + "lwc1 %[tmp1], 32(%[tmpz]) \n\t"
> + "lwc1 %[pom1], 44(%[tmpz]) \n\t"
> + "add.s %[tmp6], %[tmp2], %[tmp4] \n\t" // tmp6 = tmp2 + tmp4;
> + "sub.s %[tmp8], %[tmp2], %[tmp4] \n\t" // tmp8 = tmp2 - tmp4;
> + "lwc1 %[tmp2], 36(%[tmpz]) \n\t"
> + "lwc1 %[pom2], 56(%[tmpz]) \n\t"
> + "lwc1 %[pom3], 60(%[tmpz]) \n\t"
> + "lwc1 %[tmp3], 48(%[tmpz]) \n\t"
> + "lwc1 %[tmp4], 52(%[tmpz]) \n\t"
> + "sub.s %[tmp1], %[tmp1], %[pom] \n\t" // tmp1 = tmpz[4].re - tmpz[5].re;
> + "lwc1 %[pom], 0(%[tmpz]) \n\t"
> + "sub.s %[tmp2], %[tmp2], %[pom1] \n\t" // tmp2 = tmpz[4].im - tmpz[5].im;
> + "sub.s %[tmp3], %[tmp3], %[pom2] \n\t" // tmp3 = tmpz[6].re - tmpz[7].re;
> + "lwc1 %[pom2], 4(%[tmpz]) \n\t"
> + "sub.s %[pom1], %[pom], %[tmp5] \n\t"
> + "sub.s %[tmp4], %[tmp4], %[pom3] \n\t" // tmp4 = tmpz[6].im - tmpz[7].im;
> + "add.s %[pom3], %[pom], %[tmp5] \n\t"
> + "sub.s %[pom], %[pom2], %[tmp6] \n\t"
> + "add.s %[pom2], %[pom2], %[tmp6] \n\t"
> + "swc1 %[pom1], 32(%[tmpz]) \n\t" // tmpz[4].re = tmpz[0].re - tmp5;
> + "swc1 %[pom3], 0(%[tmpz]) \n\t" // tmpz[0].re = tmpz[0].re + tmp5;
> + "swc1 %[pom], 36(%[tmpz]) \n\t" // tmpz[4].im = tmpz[0].im - tmp6;
> + "swc1 %[pom2], 4(%[tmpz]) \n\t" // tmpz[0].im = tmpz[0].im + tmp6;
> + "lwc1 %[pom1], 16(%[tmpz]) \n\t"
> + "lwc1 %[pom3], 20(%[tmpz]) \n\t"
> + "li.s %[pom], 0.7071067812 \n\t" // float pom = 0.7071067812f;
> + "add.s %[temp1],%[tmp1], %[tmp2] \n\t"
> + "sub.s %[temp], %[pom1], %[tmp8] \n\t"
> + "add.s %[pom2], %[pom3], %[tmp7] \n\t"
> + "sub.s %[temp3],%[tmp3], %[tmp4] \n\t"
> + "sub.s %[temp4],%[tmp2], %[tmp1] \n\t"
> + "swc1 %[temp], 48(%[tmpz]) \n\t" // tmpz[6].re = tmpz[2].re - tmp8;
> + "swc1 %[pom2], 52(%[tmpz]) \n\t" // tmpz[6].im = tmpz[2].im + tmp7;
> + "add.s %[pom1], %[pom1], %[tmp8] \n\t"
> + "sub.s %[pom3], %[pom3], %[tmp7] \n\t"
> + "add.s %[tmp3], %[tmp3], %[tmp4] \n\t"
> + "mul.s %[tmp5], %[pom], %[temp1] \n\t" // tmp5 = pom * (tmp1 + tmp2);
> + "mul.s %[tmp7], %[pom], %[temp3] \n\t" // tmp7 = pom * (tmp3 - tmp4);
> + "mul.s %[tmp6], %[pom], %[temp4] \n\t" // tmp6 = pom * (tmp2 - tmp1);
> + "mul.s %[tmp8], %[pom], %[tmp3] \n\t" // tmp8 = pom * (tmp3 + tmp4);
> + "swc1 %[pom1], 16(%[tmpz]) \n\t" // tmpz[2].re = tmpz[2].re + tmp8;
> + "swc1 %[pom3], 20(%[tmpz]) \n\t" // tmpz[2].im = tmpz[2].im - tmp7;
> + "add.s %[tmp1], %[tmp5], %[tmp7] \n\t" // tmp1 = tmp5 + tmp7;
> + "sub.s %[tmp3], %[tmp5], %[tmp7] \n\t" // tmp3 = tmp5 - tmp7;
> + "add.s %[tmp2], %[tmp6], %[tmp8] \n\t" // tmp2 = tmp6 + tmp8;
> + "sub.s %[tmp4], %[tmp6], %[tmp8] \n\t" // tmp4 = tmp6 - tmp8;
> + "lwc1 %[temp], 8(%[tmpz]) \n\t"
> + "lwc1 %[temp1],12(%[tmpz]) \n\t"
> + "lwc1 %[pom], 24(%[tmpz]) \n\t"
> + "lwc1 %[pom2], 28(%[tmpz]) \n\t"
> + "sub.s %[temp4],%[temp], %[tmp1] \n\t"
> + "sub.s %[temp3],%[temp1], %[tmp2] \n\t"
> + "add.s %[temp], %[temp], %[tmp1] \n\t"
> + "add.s %[temp1],%[temp1], %[tmp2] \n\t"
> + "sub.s %[pom1], %[pom], %[tmp4] \n\t"
> + "add.s %[pom3], %[pom2], %[tmp3] \n\t"
> + "add.s %[pom], %[pom], %[tmp4] \n\t"
> + "sub.s %[pom2], %[pom2], %[tmp3] \n\t"
> + "swc1 %[temp4],40(%[tmpz]) \n\t" // tmpz[5].re = tmpz[1].re - tmp1;
> + "swc1 %[temp3],44(%[tmpz]) \n\t" // tmpz[5].im = tmpz[1].im - tmp2;
> + "swc1 %[temp], 8(%[tmpz]) \n\t" // tmpz[1].re = tmpz[1].re + tmp1;
> + "swc1 %[temp1],12(%[tmpz]) \n\t" // tmpz[1].im = tmpz[1].im + tmp2;
> + "swc1 %[pom1], 56(%[tmpz]) \n\t" // tmpz[7].re = tmpz[3].re - tmp4;
> + "swc1 %[pom3], 60(%[tmpz]) \n\t" // tmpz[7].im = tmpz[3].im + tmp3;
> + "swc1 %[pom], 24(%[tmpz]) \n\t" // tmpz[3].re = tmpz[3].re + tmp4;
> + "swc1 %[pom2], 28(%[tmpz]) \n\t" // tmpz[3].im = tmpz[3].im - tmp3;
> + : [tmpz]"+r"(tmpz), [tmp1]"=f"(tmp1), [pom]"=f"(pom), [pom1]"=&f"(pom1), [pom2]"=&f"(pom2),
> + [tmp3]"=f"(tmp3), [tmp2]"=f"(tmp2), [tmp4]"=f"(tmp4), [tmp5]"=f"(tmp5), [tmp7]"=f"(tmp7),
> + [tmp6]"=f"(tmp6), [tmp8]"=f"(tmp8), [pom3]"=&f"(pom3),[temp]"=&f"(temp), [temp1]"=&f"(temp1),
> + [temp3]"=&f"(temp3), [temp4]"=&f"(temp4)
> + :
> + : "memory"
> + );
> + }
> +
> + step = 1<< (MAX_LOG2_NFFT - 4);
> + n4 = 4;
> + for (nbits=4; nbits<=s->nbits; nbits++)
> + {
> + /*
> + * num_transforms = (num_transforms>> 1) | 1;
> + */
> + __asm__ __volatile__ (
> + "sra %[num_t], %[num_t], 1 \n\t"
> + "ori %[num_t], %[num_t], 1 \n\t"
> +
> + : [num_t] "+r" (num_transforms)
> + );
> + n2 = 2 * n4;
> + n34 = 3 * n4;
> +
> + for (n=0; n<num_transforms; n++)
> + {
> + offset = fft_offsets_lut[n]<< nbits;
> + tmpz = z + offset;
> +
> + tmpz_n2 = tmpz + n2;
> + tmpz_n4 = tmpz + n4;
> + tmpz_n34 = tmpz + n34;
> +
> + __asm__ __volatile__ (
> + "lwc1 %[pom1], 0(%[tmpz_n2]) \n\t"
> + "lwc1 %[pom], 0(%[tmpz_n34]) \n\t"
> + "lwc1 %[pom2], 4(%[tmpz_n2]) \n\t"
> + "lwc1 %[pom3], 4(%[tmpz_n34]) \n\t"
> + "lwc1 %[temp1],0(%[tmpz]) \n\t"
> + "lwc1 %[temp3],4(%[tmpz]) \n\t"
> + "add.s %[tmp5], %[pom1], %[pom] \n\t" // tmp5 = tmpz[ n2].re + tmpz[n34].re;
> + "sub.s %[tmp1], %[pom1], %[pom] \n\t" // tmp1 = tmpz[ n2].re - tmpz[n34].re;
> + "add.s %[tmp6], %[pom2], %[pom3] \n\t" // tmp6 = tmpz[ n2].im + tmpz[n34].im;
> + "sub.s %[tmp2], %[pom2], %[pom3] \n\t" // tmp2 = tmpz[ n2].im - tmpz[n34].im;
> + "sub.s %[temp], %[temp1], %[tmp5] \n\t"
> + "add.s %[temp1],%[temp1], %[tmp5] \n\t"
> + "sub.s %[temp4],%[temp3], %[tmp6] \n\t"
> + "add.s %[temp3],%[temp3], %[tmp6] \n\t"
> + "swc1 %[temp], 0(%[tmpz_n2]) \n\t" // tmpz[ n2].re = tmpz[ 0].re - tmp5;
> + "swc1 %[temp1],0(%[tmpz]) \n\t" // tmpz[ 0].re = tmpz[ 0].re + tmp5;
> + "lwc1 %[pom1], 0(%[tmpz_n4]) \n\t"
> + "swc1 %[temp4],4(%[tmpz_n2]) \n\t" // tmpz[ n2].im = tmpz[ 0].im - tmp6;
> + "lwc1 %[temp], 4(%[tmpz_n4]) \n\t"
> + "swc1 %[temp3],4(%[tmpz]) \n\t" // tmpz[ 0].im = tmpz[ 0].im + tmp6;
> + "sub.s %[pom], %[pom1], %[tmp2] \n\t"
> + "add.s %[pom1], %[pom1], %[tmp2] \n\t"
> + "add.s %[temp1],%[temp], %[tmp1] \n\t"
> + "sub.s %[temp], %[temp], %[tmp1] \n\t"
> + "swc1 %[pom], 0(%[tmpz_n34]) \n\t" // tmpz[n34].re = tmpz[n4].re - tmp2;
> + "swc1 %[pom1], 0(%[tmpz_n4]) \n\t" // tmpz[ n4].re = tmpz[n4].re + tmp2;
> + "swc1 %[temp1],4(%[tmpz_n34]) \n\t" // tmpz[n34].im = tmpz[n4].im + tmp1;
> + "swc1 %[temp], 4(%[tmpz_n4]) \n\t" // tmpz[ n4].im = tmpz[n4].im - tmp1;
> + : [tmpz]"+r"(tmpz), [tmpz_n2]"+r"(tmpz_n2), [tmpz_n34]"+r"(tmpz_n34), [tmp5]"=f"(tmp5),
> + [tmp1]"=f"(tmp1), [pom]"=&f"(pom), [pom1]"=&f"(pom1), [pom2]"=&f"(pom2),
> + [tmp2]"=f"(tmp2), [tmp6]"=f"(tmp6), [tmpz_n4]"+r"(tmpz_n4), [pom3]"=&f"(pom3),
> + [temp]"=f"(temp), [temp1]"=f"(temp1), [temp3]"=f"(temp3), [temp4]"=f"(temp4)
> + :
> + : "memory"
> + );
> +
> + w_re_ptr = w_tab + step;
> +
> + for (i=1; i<n4; i++)
> + {
> + w_re = w_re_ptr[0];
> + w_im = w_re_ptr[MAX_FFT_SIZE/4];
Can you explain why you cannot use the same cos/sin tab that the C
version uses?
-Vitor
More information about the ffmpeg-devel
mailing list