RepoMirrors
/
ffmpeg
mirror of https://git.ffmpeg.org/ffmpeg.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity
ffmpeg/libavutil/tx_priv.h

314 lines
14 KiB
C
Raw Normal View History

lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
/*
* 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
*/
#ifndef AVUTIL_TX_PRIV_H
#define AVUTIL_TX_PRIV_H
#include "tx.h"
#include "thread.h"
lavu/mem: move the DECLARE_ALIGNED macro family to mem_internal on next+1 bump They are not properly namespaced and not intended for public use.
2020-05-27 12:54:38 +00:00
#include "mem_internal.h"
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
#include "attributes.h"
#ifdef TX_FLOAT
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define TX_TAB(x) x ## _float
#define TX_NAME(x) x ## _float_c
lavu/tx: disable debugging information when CONFIG_SMALL
2022-02-07 02:42:19 +00:00
#define TX_NAME_STR(x) NULL_IF_CONFIG_SMALL(x "_float_c")
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define TX_TYPE(x) AV_TX_FLOAT_ ## x
lavu/tx: refactor assembly codelet definition This commit does some refactoring to make defining assembly codelets smaller, and fixes compiler redefinition warnings. It also allows for other assembly versions to reuse the same boilerplate code as x86. Finally, it also adds the out_of_place flag to all assembly codelets. This changes nothing, as out-of-place operation was assumed to be available anyway, but this makes it more explicit.
2022-02-07 02:30:27 +00:00
#define TX_FN_NAME(fn, suffix) ff_tx_ ## fn ## _float_ ## suffix
lavu/tx: disable debugging information when CONFIG_SMALL
2022-02-07 02:42:19 +00:00
#define TX_FN_NAME_STR(fn, suffix) NULL_IF_CONFIG_SMALL(#fn "_float_" #suffix)
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define MULT(x, m) ((x) * (m))
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
#define SCALE_TYPE float
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
typedef float TXSample;
typedef AVComplexFloat TXComplex;
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
#elif defined(TX_DOUBLE)
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define TX_TAB(x) x ## _double
#define TX_NAME(x) x ## _double_c
lavu/tx: disable debugging information when CONFIG_SMALL
2022-02-07 02:42:19 +00:00
#define TX_NAME_STR(x) NULL_IF_CONFIG_SMALL(x "_double_c")
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define TX_TYPE(x) AV_TX_DOUBLE_ ## x
lavu/tx: refactor assembly codelet definition This commit does some refactoring to make defining assembly codelets smaller, and fixes compiler redefinition warnings. It also allows for other assembly versions to reuse the same boilerplate code as x86. Finally, it also adds the out_of_place flag to all assembly codelets. This changes nothing, as out-of-place operation was assumed to be available anyway, but this makes it more explicit.
2022-02-07 02:30:27 +00:00
#define TX_FN_NAME(fn, suffix) ff_tx_ ## fn ## _double_ ## suffix
lavu/tx: disable debugging information when CONFIG_SMALL
2022-02-07 02:42:19 +00:00
#define TX_FN_NAME_STR(fn, suffix) NULL_IF_CONFIG_SMALL(#fn "_double_" #suffix)
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define MULT(x, m) ((x) * (m))
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
#define SCALE_TYPE double
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
typedef double TXSample;
typedef AVComplexDouble TXComplex;
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
#elif defined(TX_INT32)
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define TX_TAB(x) x ## _int32
#define TX_NAME(x) x ## _int32_c
lavu/tx: disable debugging information when CONFIG_SMALL
2022-02-07 02:42:19 +00:00
#define TX_NAME_STR(x) NULL_IF_CONFIG_SMALL(x "_int32_c")
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define TX_TYPE(x) AV_TX_INT32_ ## x
lavu/tx: refactor assembly codelet definition This commit does some refactoring to make defining assembly codelets smaller, and fixes compiler redefinition warnings. It also allows for other assembly versions to reuse the same boilerplate code as x86. Finally, it also adds the out_of_place flag to all assembly codelets. This changes nothing, as out-of-place operation was assumed to be available anyway, but this makes it more explicit.
2022-02-07 02:30:27 +00:00
#define TX_FN_NAME(fn, suffix) ff_tx_ ## fn ## _int32_ ## suffix
lavu/tx: disable debugging information when CONFIG_SMALL
2022-02-07 02:42:19 +00:00
#define TX_FN_NAME_STR(fn, suffix) NULL_IF_CONFIG_SMALL(#fn "_int32_" #suffix)
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define MULT(x, m) (((((int64_t)(x)) * (int64_t)(m)) + 0x40000000) >> 31)
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
#define SCALE_TYPE float
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
typedef int32_t TXSample;
typedef AVComplexInt32 TXComplex;
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
#else
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
typedef void TXComplex;
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
#endif
lavu/tx: refactor assembly codelet definition This commit does some refactoring to make defining assembly codelets smaller, and fixes compiler redefinition warnings. It also allows for other assembly versions to reuse the same boilerplate code as x86. Finally, it also adds the out_of_place flag to all assembly codelets. This changes nothing, as out-of-place operation was assumed to be available anyway, but this makes it more explicit.
2022-02-07 02:30:27 +00:00
#define TX_DECL_FN(fn, suffix) \
void TX_FN_NAME(fn, suffix)(AVTXContext *s, void *o, void *i, ptrdiff_t st);
#define TX_DEF(fn, tx_type, len_min, len_max, f1, f2, \
p, init_fn, suffix, cf, cd_flags, cf2) \
&(const FFTXCodelet){ \
.name = TX_FN_NAME_STR(fn, suffix), \
.function = TX_FN_NAME(fn, suffix), \
.type = TX_TYPE(tx_type), \
.flags = FF_TX_ALIGNED | FF_TX_OUT_OF_PLACE | cd_flags, \
.factors = { f1, f2 }, \
.min_len = len_min, \
.max_len = len_max, \
.init = init_fn, \
.cpu_flags = cf2 | AV_CPU_FLAG_ ## cf, \
.prio = p, \
}
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
#if defined(TX_FLOAT) || defined(TX_DOUBLE)
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define CMUL(dre, dim, are, aim, bre, bim) \
do { \
(dre) = (are) * (bre) - (aim) * (bim); \
(dim) = (are) * (bim) + (aim) * (bre); \
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
} while (0)
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define SMUL(dre, dim, are, aim, bre, bim) \
do { \
(dre) = (are) * (bre) - (aim) * (bim); \
(dim) = (are) * (bim) - (aim) * (bre); \
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
} while (0)
lavu: support arbitrary-point FFTs and all even (i)MDCT transforms This patch adds support for arbitrary-point FFTs and all even MDCT transforms. Odd MDCTs are not supported yet as they're based on the DCT-II and DCT-III and they're very niche. With this we can now write tests.
2021-01-12 07:11:47 +00:00
#define UNSCALE(x) (x)
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
#define RESCALE(x) (x)
#define FOLD(a, b) ((a) + (b))
#elif defined(TX_INT32)
/* Properly rounds the result */
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define CMUL(dre, dim, are, aim, bre, bim) \
do { \
int64_t accu; \
(accu) = (int64_t)(bre) * (are); \
(accu) -= (int64_t)(bim) * (aim); \
(dre) = (int)(((accu) + 0x40000000) >> 31); \
(accu) = (int64_t)(bim) * (are); \
(accu) += (int64_t)(bre) * (aim); \
(dim) = (int)(((accu) + 0x40000000) >> 31); \
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
} while (0)
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define SMUL(dre, dim, are, aim, bre, bim) \
do { \
int64_t accu; \
(accu) = (int64_t)(bre) * (are); \
(accu) -= (int64_t)(bim) * (aim); \
(dre) = (int)(((accu) + 0x40000000) >> 31); \
(accu) = (int64_t)(bim) * (are); \
(accu) -= (int64_t)(bre) * (aim); \
(dim) = (int)(((accu) + 0x40000000) >> 31); \
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
} while (0)
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define UNSCALE(x) ((double)(x)/2147483648.0)
lavu/tx: clip when converting table values to fixed-point INT32_MAX (2147483647) isn't exactly representable by a floating point value, with the closest being 2147483648.0. So when rescaling a value of 1.0, this could overflow when casting the 64-bit value returned from lrintf() into 32 bits. Unfortunately the properties of integer overflows don't match up well with how a Fourier Transform operates. So clip the value before casting to a 32-bit int. Should be noted we don't have overflows with the table values we're currently using. However, converting a Kaiser-Bessel window function with a length of 256 and a parameter of 5.0 to fixed point did create overflows. So this is more of insurance to save debugging time in case something changes in the future. The macro is only used during init, so it being a little slower is not a problem.
2021-01-09 19:41:25 +00:00
#define RESCALE(x) (av_clip64(lrintf((x) * 2147483648.0), INT32_MIN, INT32_MAX))
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define FOLD(x, y) ((int32_t)((x) + (unsigned)(y) + 32) >> 6)
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#endif /* TX_INT32 */
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define BF(x, y, a, b) \
do { \
x = (a) - (b); \
y = (a) + (b); \
lavu/tx: implement 32 bit fixed point FFT and MDCT Required minimal changes to the code so made sense to implement. FFT and MDCT tested, the output of both was properly rounded. Fun fact: the non-power-of-two fixed-point FFT and MDCT are the fastest ever non-power-of-two fixed-point FFT and MDCT written. This can replace the power of two integer MDCTs in aac and ac3 if the MIPS optimizations are ported across. Unfortunately the ac3 encoder uses a 16-bit fixed point forward transform, unlike the encoder which uses a 32bit inverse transform, so some modifications might be required there. The 3-point FFT is somewhat less accurate than it otherwise could be, having minor rounding errors with bigger transforms. However, this could be improved later, and the way its currently written is the way one would write assembly for it. Similar rounding errors can also be found throughout the power of two FFTs as well, though those are more difficult to correct. Despite this, the integer transforms are more than accurate enough.
2020-02-08 23:13:28 +00:00
} while (0)
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
#define CMUL3(c, a, b) CMUL((c).re, (c).im, (a).re, (a).im, (b).re, (b).im)
/* Codelet flags, used to pick codelets. Must be a superset of enum AVTXFlags,
* but if it runs out of bits, it can be made separate. */
lavu/tx: avoid using 64-bit enums MSVC (2016, but possibly more) still force enums to be basic ints.
2022-01-27 09:19:45 +00:00
#define FF_TX_OUT_OF_PLACE (1ULL << 63) /* Can be OR'd with AV_TX_INPLACE */
#define FF_TX_ALIGNED (1ULL << 62) /* Cannot be OR'd with AV_TX_UNALIGNED */
#define FF_TX_PRESHUFFLE (1ULL << 61) /* Codelet expects permuted coeffs */
#define FF_TX_INVERSE_ONLY (1ULL << 60) /* For non-orthogonal inverse-only transforms */
#define FF_TX_FORWARD_ONLY (1ULL << 59) /* For non-orthogonal forward-only transforms */
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
typedef enum FFTXCodeletPriority {
FF_TX_PRIO_BASE = 0, /* Baseline priority */
/* For SIMD, set base prio to the register size in bits and increment in
* steps of 64 depending on faster/slower features, like FMA. */
FF_TX_PRIO_MIN = -131072, /* For naive implementations */
FF_TX_PRIO_MAX = 32768, /* For custom implementations/ASICs */
} FFTXCodeletPriority;
/* Codelet options */
typedef struct FFTXCodeletOptions {
int invert_lookup; /* If codelet is flagged as FF_TX_CODELET_PRESHUFFLE,
invert the lookup direction for the map generated */
} FFTXCodeletOptions;
/* Maximum amount of subtransform functions, subtransforms and factors. Arbitrary. */
#define TX_MAX_SUB 4
typedef struct FFTXCodelet {
const char *name; /* Codelet name, for debugging */
av_tx_fn function; /* Codelet function, != NULL */
enum AVTXType type; /* Type of codelet transform */
#define TX_TYPE_ANY INT32_MAX /* Special type to allow all types */
lavu/tx: avoid using 64-bit enums MSVC (2016, but possibly more) still force enums to be basic ints.
2022-01-27 09:19:45 +00:00
uint64_t flags; /* A combination of AVTXFlags and codelet
* flags that describe its properties. */
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
int factors[TX_MAX_SUB]; /* Length factors */
#define TX_FACTOR_ANY -1 /* When used alone, signals that the codelet
* supports all factors. Otherwise, if other
* factors are present, it signals that whatever
* remains will be supported, as long as the
* other factors are a component of the length */
int min_len; /* Minimum length of transform, must be >= 1 */
int max_len; /* Maximum length of transform */
#define TX_LEN_UNLIMITED -1 /* Special length value to permit all lengths */
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
int (*init)(AVTXContext *s, /* Optional callback for current context initialization. */
const struct FFTXCodelet *cd,
uint64_t flags,
FFTXCodeletOptions *opts,
int len, int inv,
const void *scale);
int (*uninit)(AVTXContext *s); /* Optional callback for uninitialization. */
int cpu_flags; /* CPU flags. If any negative flags like
* SLOW are present, will avoid picking.
* 0x0 to signal it's a C codelet */
#define FF_TX_CPU_FLAGS_ALL 0x0 /* Special CPU flag for C */
int prio; /* < 0 = least, 0 = no pref, > 0 = prefer */
} FFTXCodelet;
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
struct AVTXContext {
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
/* Fields the root transform and subtransforms use or may use.
* NOTE: This section is used by assembly, do not reorder or change */
int len; /* Length of the transform */
int inv; /* If transform is inverse */
int *map; /* Lookup table(s) */
TXComplex *exp; /* Any non-pre-baked multiplication factors needed */
TXComplex *tmp; /* Temporary buffer, if needed */
AVTXContext *sub; /* Subtransform context(s), if needed */
av_tx_fn fn[TX_MAX_SUB]; /* Function(s) for the subtransforms */
int nb_sub; /* Number of subtransforms.
* The reason all of these are set here
* rather than in each separate context
* is to eliminate extra pointer
* dereferences. */
/* Fields mainly useul/applicable for the root transform or initialization.
* Fields below are not used by assembly code. */
const FFTXCodelet *cd[TX_MAX_SUB]; /* Subtransform codelets */
const FFTXCodelet *cd_self; /* Codelet for the current context */
enum AVTXType type; /* Type of transform */
lavu/tx: avoid using 64-bit enums MSVC (2016, but possibly more) still force enums to be basic ints.
2022-01-27 09:19:45 +00:00
uint64_t flags; /* A combination of AVTXFlags and
* codelet flags used when creating */
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
float scale_f;
double scale_d;
void *opaque; /* Free to use by implementations */
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
};
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
/* Create a subtransform in the current context with the given parameters.
* The flags parameter from FFTXCodelet.init() should be preserved as much
* as that's possible.
* MUST be called during the sub() callback of each codelet. */
int ff_tx_init_subtx(AVTXContext *s, enum AVTXType type,
uint64_t flags, FFTXCodeletOptions *opts,
int len, int inv, const void *scale);
lavu/tx: minor code style improvements and additional comments
2021-04-10 01:45:03 +00:00
/*
* Generates the PFA permutation table into AVTXContext->pfatab. The end table
* is appended to the start table.
*/
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
int ff_tx_gen_compound_mapping(AVTXContext *s, int n, int m);
lavu/tx: minor code style improvements and additional comments
2021-04-10 01:45:03 +00:00
/*
* Generates a standard-ish (slightly modified) Split-Radix revtab into
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
* AVTXContext->map. Invert lookup changes how the mapping needs to be applied.
* If it's set to 0, it has to be applied like out[map[i]] = in[i], otherwise
* if it's set to 1, has to be applied as out[i] = in[map[i]]
lavu/tx: minor code style improvements and additional comments
2021-04-10 01:45:03 +00:00
*/
lavu/tx: invert permutation lookups out[lut[i]] = in[i] lookups were 4.04 times(!) slower than out[i] = in[lut[i]] lookups for an out-of-place FFT of length 4096. The permutes remain unchanged for anything but out-of-place monolithic FFT, as those benefit quite a lot from the current order (it means there's only 1 lookup necessary to add to an offset, rather than a full gather). The code was based around non-power-of-two FFTs, so this wasn't benchmarked early on.
2021-02-27 03:11:04 +00:00
int ff_tx_gen_ptwo_revtab(AVTXContext *s, int invert_lookup);
lavu/tx: minor code style improvements and additional comments
2021-04-10 01:45:03 +00:00
/*
* Generates an index into AVTXContext->inplace_idx that if followed in the
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
* specific order, allows the revtab to be done in-place. The sub-transform
* and its map should already be initialized.
lavu/tx: minor code style improvements and additional comments
2021-04-10 01:45:03 +00:00
*/
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
int ff_tx_gen_ptwo_inplace_revtab_idx(AVTXContext *s);
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
lavu/tx: add parity revtab generator version This will be used for SIMD support.
2021-04-10 01:52:31 +00:00
/*
* This generates a parity-based revtab of length len and direction inv.
*
* Parity means even and odd complex numbers will be split, e.g. the even
* coefficients will come first, after which the odd coefficients will be
* placed. For example, a 4-point transform's coefficients after reordering:
* z[0].re, z[0].im, z[2].re, z[2].im, z[1].re, z[1].im, z[3].re, z[3].im
*
* The basis argument is the length of the largest non-composite transform
* supported, and also implies that the basis/2 transform is supported as well,
* as the split-radix algorithm requires it to be.
*
* The dual_stride argument indicates that both the basis, as well as the
* basis/2 transforms support doing two transforms at once, and the coefficients
* will be interleaved between each pair in a split-radix like so (stride == 2):
* tx1[0], tx1[2], tx2[0], tx2[2], tx1[1], tx1[3], tx2[1], tx2[3]
* A non-zero number switches this on, with the value indicating the stride
* (how many values of 1 transform to put first before switching to the other).
* Must be a power of two or 0. Must be less than the basis.
* Value will be clipped to the transform size, so for a basis of 16 and a
* dual_stride of 8, dual 8-point transforms will be laid out as if dual_stride
* was set to 4.
* Usually you'll set this to half the complex numbers that fit in a single
* register or 0. This allows to reuse SSE functions as dual-transform
* functions in AVX mode.
*
* If length is smaller than basis/2 this function will not do anything.
*/
lavu/tx: rewrite internal code as a tree-based codelet constructor This commit rewrites the internal transform code into a constructor that stitches transforms (codelets). This allows for transforms to reuse arbitrary parts of other transforms, and allows transforms to be stacked onto one another (such as a full iMDCT using a half-iMDCT which in turn uses an FFT). It also permits for each step to be individually replaced by assembly or a custom implementation (such as an ASIC).
2022-01-20 06:14:46 +00:00
int ff_tx_gen_split_radix_parity_revtab(AVTXContext *s, int invert_lookup,
int basis, int dual_stride);
/* Typed init function to initialize shared tables. Will initialize all tables
* for all factors of a length. */
void ff_tx_init_tabs_float (int len);
void ff_tx_init_tabs_double(int len);
void ff_tx_init_tabs_int32 (int len);
/* Typed init function to initialize an MDCT exptab in a context. */
int ff_tx_mdct_gen_exp_float (AVTXContext *s);
int ff_tx_mdct_gen_exp_double(AVTXContext *s);
int ff_tx_mdct_gen_exp_int32 (AVTXContext *s);
/* Lists of codelets */
extern const FFTXCodelet * const ff_tx_codelet_list_float_c [];
extern const FFTXCodelet * const ff_tx_codelet_list_float_x86 [];
extern const FFTXCodelet * const ff_tx_codelet_list_double_c [];
extern const FFTXCodelet * const ff_tx_codelet_list_int32_c [];
lavu/x86: add FFT assembly This commit adds a pure x86 assembly SIMD version of the FFT in libavutil/tx. The design of this pure assembly FFT is pretty unconventional. On the lowest level, instead of splitting the complex numbers into real and imaginary parts, we keep complex numbers together but split them in terms of parity. This saves a number of shuffles in each transform, but more importantly, it splits each transform into two independent paths, which we process using separate registers in parallel. This allows us to keep all units saturated and lets us use all available registers to avoid dependencies. Moreover, it allows us to double the granularity of our per-load permutation, skipping many expensive lookups and allowing us to use just 4 loads per register, rather than 8, or in case FMA3 (and by extension, AVX2), use the vgatherdpd instruction, which is at least as fast as 4 separate loads on old hardware, and quite a bit faster on modern CPUs). Higher up, we go for a bottom-up construction of large transforms, foregoing the traditional per-transform call-return recursion chains. Instead, we always start at the bottom-most basis transform (in this case, a 32-point transform), and continue constructing larger and larger transforms until we return to the top-most transform. This way, we only touch the stack 3 times per a complete target transform: once for the 1/2 length transform and two times for the 1/4 length transform. The combination algorithm we use is a standard Split-Radix algorithm, as used in our C code. Although a version with less operations exists (Steven G. Johnson and Matteo Frigo's "A modified split-radix FFT with fewer arithmetic operations", IEEE Trans. Signal Process. 55 (1), 111–119 (2007), which is the one FFTW uses), it only has 2% less operations and requires at least 4x the binary code (due to it needing 4 different paths to do a single transform). That version also has other issues which prevent it from being implemented with SIMD code as efficiently, which makes it lose the marginal gains it offered, and cannot be performed bottom-up, requiring many recursive call-return chains, whose overhead adds up. We go through a lot of effort to minimize load/stores by keeping as much in registers in between construcring transforms. This saves us around 32 cycles, on paper, but in reality a lot more due to load/store aliasing (a load from a memory location cannot be issued while there's a store pending, and there are only so many (2 for Zen 3) load/store units in a CPU). Also, we interleave coefficients during the last stage to save on a store+load per register. Each of the smallest, basis transforms (4, 8 and 16-point in our case) has been extremely optimized. Our 8-point transform is barely 20 instructions in total, beating our old implementation 8-point transform by 1 instruction. Our 2x8-point transform is 23 instructions, beating our old implementation by 6 instruction and needing 50% less cycles. Our 16-point transform's combination code takes slightly more instructions than our old implementation, but makes up for it by requiring a lot less arithmetic operations. Overall, the transform was optimized for the timings of Zen 3, which at the time of writing has the most IPC from all documented CPUs. Shuffles were preferred over arithmetic operations due to their 1/0.5 latency/throughput. On average, this code is 30% faster than our old libavcodec implementation. It's able to trade blows with the previously-untouchable FFTW on small transforms, and due to its tiny size and better prediction, outdoes FFTW on larger transforms by 11% on the largest currently supported size.
2021-04-10 01:54:22 +00:00
lavu/tx: add support for double precision FFT and MDCT Simply moves and templates the actual transforms to support an additional data type. Unlike the float version, which is equal or better than libfftw3f, double precision output is bit identical with libfftw3.
2019-07-27 17:54:20 +00:00
#endif /* AVUTIL_TX_PRIV_H */