Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
|
|
|
|
|
/* starch generated code. Do not edit. */
|
|
|
|
|
|
|
|
|
|
#include "dsp-types.h"
|
|
|
|
|
#include "cpu.h"
|
|
|
|
|
|
|
|
|
|
/* mixes */
|
|
|
|
|
|
|
|
|
|
/* Generic build, compiler defaults only */
|
|
|
|
|
#ifdef STARCH_MIX_GENERIC
|
|
|
|
|
#define STARCH_FLAVOR_GENERIC
|
|
|
|
|
#define STARCH_MIX_ALIGNMENT 1
|
|
|
|
|
#endif /* STARCH_MIX_GENERIC */
|
|
|
|
|
|
|
|
|
|
/* ARM */
|
|
|
|
|
#ifdef STARCH_MIX_ARM
|
|
|
|
|
#define STARCH_FLAVOR_ARMV7A_NEON_VFPV4
|
|
|
|
|
#define STARCH_FLAVOR_GENERIC
|
|
|
|
|
#define STARCH_MIX_ALIGNMENT 16
|
|
|
|
|
#endif /* STARCH_MIX_ARM */
|
|
|
|
|
|
2021-02-08 00:07:17 +00:00
|
|
|
/* AARCH64 */
|
|
|
|
|
#ifdef STARCH_MIX_AARCH64
|
|
|
|
|
#define STARCH_FLAVOR_ARMV8_A
|
|
|
|
|
#define STARCH_FLAVOR_GENERIC
|
|
|
|
|
#define STARCH_MIX_ALIGNMENT 32
|
|
|
|
|
#endif /* STARCH_MIX_AARCH64 */
|
|
|
|
|
|
Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
/* x64 */
|
|
|
|
|
#ifdef STARCH_MIX_X86
|
|
|
|
|
#define STARCH_FLAVOR_X86_AVX2
|
|
|
|
|
#define STARCH_FLAVOR_GENERIC
|
|
|
|
|
#define STARCH_MIX_ALIGNMENT 32
|
|
|
|
|
#endif /* STARCH_MIX_X86 */
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
#ifdef STARCH_MIX_ALIGNMENT
|
|
|
|
|
#define STARCH_ALIGNMENT STARCH_MIX_ALIGNMENT
|
|
|
|
|
#define STARCH_IS_ALIGNED(_ptr) (((uintptr_t)(_ptr) & (STARCH_MIX_ALIGNMENT-1)) == 0)
|
|
|
|
|
#else
|
|
|
|
|
/* mix not defined, alignment is unknown, treat everything as unaligned */
|
|
|
|
|
#define STARCH_IS_ALIGNED(_ptr) (0)
|
|
|
|
|
#endif
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/* entry points and registries */
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_magnitude_uc8_ptr) ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
extern starch_magnitude_uc8_ptr starch_magnitude_uc8;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_magnitude_uc8_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_magnitude_uc8_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_magnitude_uc8_regentry starch_magnitude_uc8_registry[];
|
|
|
|
|
starch_magnitude_uc8_regentry * starch_magnitude_uc8_select();
|
|
|
|
|
void starch_magnitude_uc8_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_magnitude_uc8_aligned_ptr) ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
extern starch_magnitude_uc8_aligned_ptr starch_magnitude_uc8_aligned;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_magnitude_uc8_aligned_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_magnitude_uc8_aligned_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_magnitude_uc8_aligned_regentry starch_magnitude_uc8_aligned_registry[];
|
|
|
|
|
starch_magnitude_uc8_aligned_regentry * starch_magnitude_uc8_aligned_select();
|
|
|
|
|
void starch_magnitude_uc8_aligned_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_magnitude_power_uc8_ptr) ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
extern starch_magnitude_power_uc8_ptr starch_magnitude_power_uc8;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_magnitude_power_uc8_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_magnitude_power_uc8_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_magnitude_power_uc8_regentry starch_magnitude_power_uc8_registry[];
|
|
|
|
|
starch_magnitude_power_uc8_regentry * starch_magnitude_power_uc8_select();
|
|
|
|
|
void starch_magnitude_power_uc8_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_magnitude_power_uc8_aligned_ptr) ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
extern starch_magnitude_power_uc8_aligned_ptr starch_magnitude_power_uc8_aligned;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_magnitude_power_uc8_aligned_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_magnitude_power_uc8_aligned_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_magnitude_power_uc8_aligned_regentry starch_magnitude_power_uc8_aligned_registry[];
|
|
|
|
|
starch_magnitude_power_uc8_aligned_regentry * starch_magnitude_power_uc8_aligned_select();
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_magnitude_sc16_ptr) ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
extern starch_magnitude_sc16_ptr starch_magnitude_sc16;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_magnitude_sc16_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_magnitude_sc16_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_magnitude_sc16_regentry starch_magnitude_sc16_registry[];
|
|
|
|
|
starch_magnitude_sc16_regentry * starch_magnitude_sc16_select();
|
|
|
|
|
void starch_magnitude_sc16_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_magnitude_sc16_aligned_ptr) ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
extern starch_magnitude_sc16_aligned_ptr starch_magnitude_sc16_aligned;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_magnitude_sc16_aligned_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_magnitude_sc16_aligned_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_magnitude_sc16_aligned_regentry starch_magnitude_sc16_aligned_registry[];
|
|
|
|
|
starch_magnitude_sc16_aligned_regentry * starch_magnitude_sc16_aligned_select();
|
|
|
|
|
void starch_magnitude_sc16_aligned_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_magnitude_sc16q11_ptr) ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
extern starch_magnitude_sc16q11_ptr starch_magnitude_sc16q11;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_magnitude_sc16q11_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_magnitude_sc16q11_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_magnitude_sc16q11_regentry starch_magnitude_sc16q11_registry[];
|
|
|
|
|
starch_magnitude_sc16q11_regentry * starch_magnitude_sc16q11_select();
|
|
|
|
|
void starch_magnitude_sc16q11_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_magnitude_sc16q11_aligned_ptr) ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
extern starch_magnitude_sc16q11_aligned_ptr starch_magnitude_sc16q11_aligned;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_magnitude_sc16q11_aligned_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_magnitude_sc16q11_aligned_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_magnitude_sc16q11_aligned_regentry starch_magnitude_sc16q11_aligned_registry[];
|
|
|
|
|
starch_magnitude_sc16q11_aligned_regentry * starch_magnitude_sc16q11_aligned_select();
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_mean_power_u16_ptr) ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
extern starch_mean_power_u16_ptr starch_mean_power_u16;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_mean_power_u16_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_mean_power_u16_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_mean_power_u16_regentry starch_mean_power_u16_registry[];
|
|
|
|
|
starch_mean_power_u16_regentry * starch_mean_power_u16_select();
|
|
|
|
|
void starch_mean_power_u16_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
typedef void (* starch_mean_power_u16_aligned_ptr) ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
extern starch_mean_power_u16_aligned_ptr starch_mean_power_u16_aligned;
|
|
|
|
|
|
|
|
|
|
typedef struct {
|
|
|
|
|
int rank;
|
|
|
|
|
const char *name;
|
|
|
|
|
const char *flavor;
|
|
|
|
|
starch_mean_power_u16_aligned_ptr callable;
|
|
|
|
|
int (*flavor_supported)();
|
|
|
|
|
} starch_mean_power_u16_aligned_regentry;
|
|
|
|
|
|
|
|
|
|
extern starch_mean_power_u16_aligned_regentry starch_mean_power_u16_aligned_registry[];
|
|
|
|
|
starch_mean_power_u16_aligned_regentry * starch_mean_power_u16_aligned_select();
|
|
|
|
|
void starch_mean_power_u16_aligned_set_wisdom( const char * const * received_wisdom );
|
|
|
|
|
|
|
|
|
|
/* flavors and prototypes */
|
|
|
|
|
|
|
|
|
|
#ifdef STARCH_FLAVOR_GENERIC
|
|
|
|
|
void starch_magnitude_power_uc8_twopass_generic ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_lookup_generic ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_lookup_unroll_4_generic ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
2021-02-08 00:07:17 +00:00
|
|
|
void starch_magnitude_sc16_exact_u32_generic ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_exact_float_generic ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
void starch_magnitude_sc16q11_exact_u32_generic ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_exact_float_generic ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_11bit_table_generic ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_12bit_table_generic ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
2021-02-08 00:07:17 +00:00
|
|
|
void starch_magnitude_uc8_lookup_generic ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_lookup_unroll_4_generic ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_exact_generic ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_mean_power_u16_float_generic ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_u32_generic ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_u64_generic ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
#endif /* STARCH_FLAVOR_GENERIC */
|
|
|
|
|
|
|
|
|
|
int starch_read_wisdom (const char * path);
|
|
|
|
|
|
|
|
|
|
#ifdef STARCH_FLAVOR_ARMV7A_NEON_VFPV4
|
|
|
|
|
int cpu_supports_armv7_neon_vfpv4 (void);
|
|
|
|
|
void starch_magnitude_power_uc8_twopass_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_twopass_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_lookup_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_lookup_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_lookup_unroll_4_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_lookup_unroll_4_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_neon_vrsqrte_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_neon_vrsqrte_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
2021-02-08 00:07:17 +00:00
|
|
|
void starch_magnitude_sc16_exact_u32_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_aligned_exact_u32_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_exact_float_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_aligned_exact_float_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_neon_vrsqrte_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_aligned_neon_vrsqrte_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
void starch_magnitude_sc16q11_exact_u32_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_exact_u32_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_exact_float_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_exact_float_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_11bit_table_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_11bit_table_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_12bit_table_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_12bit_table_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_neon_vrsqrte_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_neon_vrsqrte_armv7a_neon_vfpv4 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
2021-02-08 00:07:17 +00:00
|
|
|
void starch_magnitude_uc8_lookup_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_lookup_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_lookup_unroll_4_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_lookup_unroll_4_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_exact_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_exact_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_neon_vrsqrte_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_neon_vrsqrte_armv7a_neon_vfpv4 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_mean_power_u16_float_armv7a_neon_vfpv4 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_float_armv7a_neon_vfpv4 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_u32_armv7a_neon_vfpv4 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_u32_armv7a_neon_vfpv4 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_u64_armv7a_neon_vfpv4 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_u64_armv7a_neon_vfpv4 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_neon_float_armv7a_neon_vfpv4 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_neon_float_armv7a_neon_vfpv4 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
#endif /* STARCH_FLAVOR_ARMV7A_NEON_VFPV4 */
|
|
|
|
|
|
|
|
|
|
int starch_read_wisdom (const char * path);
|
|
|
|
|
|
2021-02-08 00:07:17 +00:00
|
|
|
#ifdef STARCH_FLAVOR_ARMV8_A
|
|
|
|
|
void starch_magnitude_power_uc8_twopass_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_twopass_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_lookup_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_lookup_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_lookup_unroll_4_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_lookup_unroll_4_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_sc16_exact_u32_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_aligned_exact_u32_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_exact_float_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_aligned_exact_float_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_exact_u32_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_exact_u32_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_exact_float_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_exact_float_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_11bit_table_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_11bit_table_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_12bit_table_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_12bit_table_armv8_a ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_lookup_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_lookup_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_lookup_unroll_4_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_lookup_unroll_4_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_exact_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_exact_armv8_a ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_mean_power_u16_float_armv8_a ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_float_armv8_a ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_u32_armv8_a ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_u32_armv8_a ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_u64_armv8_a ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_u64_armv8_a ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
#endif /* STARCH_FLAVOR_ARMV8_A */
|
|
|
|
|
|
|
|
|
|
int starch_read_wisdom (const char * path);
|
|
|
|
|
|
Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
#ifdef STARCH_FLAVOR_X86_AVX2
|
|
|
|
|
int cpu_supports_avx2 (void);
|
|
|
|
|
void starch_magnitude_power_uc8_twopass_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_twopass_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_lookup_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_lookup_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_lookup_unroll_4_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
|
|
|
|
void starch_magnitude_power_uc8_aligned_lookup_unroll_4_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2, double * arg3, double * arg4 );
|
2021-02-08 00:07:17 +00:00
|
|
|
void starch_magnitude_sc16_exact_u32_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_aligned_exact_u32_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_exact_float_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16_aligned_exact_float_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
void starch_magnitude_sc16q11_exact_u32_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_exact_u32_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_exact_float_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_exact_float_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_11bit_table_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_11bit_table_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_12bit_table_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_sc16q11_aligned_12bit_table_x86_avx2 ( const sc16_t * arg0, uint16_t * arg1, unsigned arg2 );
|
2021-02-08 00:07:17 +00:00
|
|
|
void starch_magnitude_uc8_lookup_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_lookup_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_lookup_unroll_4_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_lookup_unroll_4_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_exact_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_magnitude_uc8_aligned_exact_x86_avx2 ( const uc8_t * arg0, uint16_t * arg1, unsigned arg2 );
|
|
|
|
|
void starch_mean_power_u16_float_x86_avx2 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_float_x86_avx2 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_u32_x86_avx2 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_u32_x86_avx2 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_u64_x86_avx2 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
|
|
|
|
void starch_mean_power_u16_aligned_u64_x86_avx2 ( const uint16_t * arg0, unsigned arg1, double * arg2, double * arg3 );
|
Move all converters to starch-based implementations (#97)
* Switch all conversion routines to use starch.
main user-visible changes:
* ensure you check out submodules ('git clone --recurse-submodules")
* --version shows the CPU features and DSP implementations in use
* --wisdom allows overriding of the built-in architecture wisdom
* --dcfilter no longer supported
* "starch-benchmark" binary will benchmark all options on the
current machine and can produce a wisdom file to feed to
the --wisdom option
If you have a usecase for --dcfilter, please get in touch and
let me know - it's an edge case and for now there's no starch/DSP
support for it, but support can be written if needed.
In almost all cases the new conversion routines are slightly or
substantially faster than the old conversion routines. The only case
that is slower is SC16/SC16Q11 on a Pi 0, which is around 10% slower
due to changing from heavily approximated lookup tables to higher
quality results (but SC16 is probably already out of reach of a Pi 0)
* No need to build with SC16Q11_TABLE_BITS any more
* Add oneoff/uc8_capture_stats
(reads a UC8 capture; measures min/max/mean I and Q)
* Switch UC8 conversion to 127.4 center, 128 range.
Looking at actual UC8 captures from a RTL2832, the mean I and Q
are actually at 127.4, so use that as the zero point.
This means that the resulting I/Q maximum values could be as large as
127.6. Switch to 128 for simplicity.
* Switch to the new UC8 zero offset in benchmarks, fix some bugs
* Fix some bugs in SC16/SC16Q11 validation, tighten the max error requirements
* Ditch UC8 approximation path, add a NEON VRQSQRTE path.
* Tweak the SC16 exact path, add a new impl that uses a mix of
u32 & floats.
* SC16Q11 impl tweaks:
* add a u32->float exact path
* ditch the approximation path
* add a NEON VRSQRTE path
* add a 12-bit table path (using the full signed I/Q value, not absolute value)
* Ditch SC16 approximation path, add NEON vrsqrte path
* Add oneoff/dsp_error_measurement
This runs sample input through the DSP functions that are
allowed to be inexact and dumps the results as a TSV suitable for
feeding to gnuplot to look at the actual errors.
* Update make clean, make wisdom targets
* Update wisdom based on benchmarking
* Preserve the raw wisdom benchmark data
* Update to latest starch
* Update .gitignore for new wisdom files
* Update starch generated code
* Build starch-benchmark as part of the 'all' target
* Use wisdom from /etc/dump1090-fa/wisdom.local if present
* Package starch-benchmark and a helper script to generate local wisdom data
* Remove submodules in preparation for importing them directly
* Import cpu_features v0.6.0 from https://github.com/google/cpu_features/releases/tag/v0.6.0
* Import starch at commit a725c8491dc33a321565d451b385131e589d8490
from https://github.com/flightaware/starch
2021-01-21 11:45:00 +00:00
|
|
|
#endif /* STARCH_FLAVOR_X86_AVX2 */
|
|
|
|
|
|
|
|
|
|
int starch_read_wisdom (const char * path);
|
|
|
|
|
|
