mirror of
https://gitlab.freedesktop.org/gstreamer/gstreamer.git
synced 2024-11-18 07:47:17 +00:00
d793a2b560
This adds ARM NEON accelerated code paths for 16-bit integer and 32-bit floating point samples. It is a modified combination of patches #3 and #5 from Jyri Sarha ( http://lists.xiph.org/pipermail/speex-dev/2011-September/008240.html & http://lists.xiph.org/pipermail/speex-dev/2011-September/008238.html ) Signed-off-by: Carlos Rafael Giani <dv@pseudoterminal.org> |
||
---|---|---|
.. | ||
arch.h | ||
fixed_arm4.h | ||
fixed_arm5e.h | ||
fixed_bfin.h | ||
fixed_debug.h | ||
fixed_generic.h | ||
gstaudioresample.c | ||
gstaudioresample.h | ||
Makefile.am | ||
README | ||
resample.c | ||
resample_neon.h | ||
resample_sse.h | ||
speex_resampler.h | ||
speex_resampler_double.c | ||
speex_resampler_float.c | ||
speex_resampler_int.c | ||
speex_resampler_wrapper.h |
arch.h fixed_arm4.h fixed_arm5e.h fixed_bfin.h fixed_debug.h fixed_generic.h resample.c speex_resampler.h are taken from http://git.xiph.org/speex.git/ as of 2009-11-10. The only changes are: diff -Naur old/arch.h new/arch.h --- old/arch.h 2009-11-10 12:18:29.000000000 +0100 +++ new/arch.h 2009-11-10 12:19:09.000000000 +0100 @@ -78,7 +78,10 @@ #include "../include/speex/speex_types.h" #endif +#ifndef ABS #define ABS(x) ((x) < 0 ? (-(x)) : (x)) /**< Absolute integer value. */ +#endif + #define ABS16(x) ((x) < 0 ? (-(x)) : (x)) /**< Absolute 16-bit value. */ #define MIN16(a,b) ((a) < (b) ? (a) : (b)) /**< Maximum 16-bit value. */ #define MAX16(a,b) ((a) > (b) ? (a) : (b)) /**< Maximum 16-bit value. */ @@ -134,6 +137,28 @@ #else +#ifdef DOUBLE_PRECISION +typedef double spx_mem_t; +typedef double spx_coef_t; +typedef double spx_lsp_t; +typedef double spx_sig_t; +typedef double spx_word16_t; +typedef double spx_word32_t; + +#define Q15ONE 1.0 +#define LPC_SCALING 1. +#define SIG_SCALING 1. +#define LSP_SCALING 1. +#define GAMMA_SCALING 1. +#define GAIN_SCALING 1. +#define GAIN_SCALING_1 1. + + +#define VERY_SMALL 1e-20 +#define VERY_LARGE32 1e20 +#define VERY_LARGE16 1e20 +#define Q15_ONE ((spx_word16_t)1.) +#else /* !DOUBLE_PRECISION */ typedef float spx_mem_t; typedef float spx_coef_t; typedef float spx_lsp_t; @@ -154,6 +179,7 @@ #define VERY_LARGE32 1e15f #define VERY_LARGE16 1e15f #define Q15_ONE ((spx_word16_t)1.f) +#endif /* DOUBLE_PRECISION */ #define QCONST16(x,bits) (x) #define QCONST32(x,bits) (x) diff -Naur old/resample.c new/resample.c --- old/resample.c 2009-11-10 12:18:51.000000000 +0100 +++ new/resample.c 2009-11-10 12:19:09.000000000 +0100 @@ -63,22 +63,27 @@ #ifdef OUTSIDE_SPEEX #include <stdlib.h> -static void * + +#include <glib.h> + +#define EXPORT G_GNUC_INTERNAL + +static inline void * speex_alloc (int size) { - return calloc (size, 1); + return g_malloc0 (size); } -static void * +static inline void * speex_realloc (void *ptr, int size) { - return realloc (ptr, size); + return g_realloc (ptr, size); } -static void +static inline void speex_free (void *ptr) { - free (ptr); + g_free (ptr); } #include "speex_resampler.h" @@ -90,7 +95,6 @@ #include "os_support.h" #endif /* OUTSIDE_SPEEX */ -#include "stack_alloc.h" #include <math.h> #ifndef M_PI @@ -263,10 +267,17 @@ }; /*8,24,40,56,80,104,128,160,200,256,320*/ +#ifdef DOUBLE_PRECISION +static double +compute_func (double x, struct FuncDef *func) +{ + double y, frac; +#else static double compute_func (float x, struct FuncDef *func) { float y, frac; +#endif double interp[4]; int ind; y = x * func->oversample; @@ -317,11 +328,19 @@ } #else /* The slow way of computing a sinc for the table. Should improve that some day */ +#ifdef DOUBLE_PRECISION +static spx_word16_t +sinc (double cutoff, double x, int N, struct FuncDef *window_func) +{ + /*fprintf (stderr, "%f ", x); */ + double xx = x * cutoff; +#else static spx_word16_t sinc (float cutoff, float x, int N, struct FuncDef *window_func) { /*fprintf (stderr, "%f ", x); */ float xx = x * cutoff; +#endif if (fabs (x) < 1e-6) return cutoff; else if (fabs (x) > .5 * N) @@ -372,6 +391,7 @@ } #endif +#ifndef DOUBLE_PRECISION static int resampler_basic_direct_single (SpeexResamplerState * st, spx_uint32_t channel_index, const spx_word16_t * in, spx_uint32_t * in_len, @@ -428,6 +448,7 @@ st->samp_frac_num[channel_index] = samp_frac_num; return out_sample; } +#endif #ifdef FIXED_POINT #else @@ -483,6 +504,7 @@ } #endif +#ifndef DOUBLE_PRECISION static int resampler_basic_interpolate_single (SpeexResamplerState * st, spx_uint32_t channel_index, const spx_word16_t * in, spx_uint32_t * in_len, @@ -562,6 +584,7 @@ st->samp_frac_num[channel_index] = samp_frac_num; return out_sample; } +#endif #ifdef FIXED_POINT #else @@ -592,10 +615,16 @@ PDIV32 (SHL32 ((samp_frac_num * st->oversample) % st->den_rate, 15), st->den_rate); #else +#ifdef DOUBLE_PRECISION + const spx_word16_t frac = + ((double) ((samp_frac_num * st->oversample) % st->den_rate)) / + st->den_rate; +#else const spx_word16_t frac = ((float) ((samp_frac_num * st->oversample) % st->den_rate)) / st->den_rate; #endif +#endif spx_word16_t interp[4]; @@ -696,20 +725,27 @@ spx_int32_t j; for (j = 0; j < st->filt_len; j++) { st->sinc_table[i * st->filt_len + j] = - sinc (st->cutoff, - ((j - (spx_int32_t) st->filt_len / 2 + 1) - + sinc (st->cutoff, ((j - (spx_int32_t) st->filt_len / 2 + 1) - +#ifdef DOUBLE_PRECISION + ((double) i) / st->den_rate), st->filt_len, +#else ((float) i) / st->den_rate), st->filt_len, +#endif quality_map[st->quality].window_func); } } #ifdef FIXED_POINT st->resampler_ptr = resampler_basic_direct_single; #else +#ifdef DOUBLE_PRECISION + st->resampler_ptr = resampler_basic_direct_double; +#else if (st->quality > 8) st->resampler_ptr = resampler_basic_direct_double; else st->resampler_ptr = resampler_basic_direct_single; #endif +#endif /*fprintf (stderr, "resampler uses direct sinc table and normalised cutoff %f\n", cutoff); */ } else { spx_int32_t i; @@ -725,16 +761,24 @@ } for (i = -4; i < (spx_int32_t) (st->oversample * st->filt_len + 4); i++) st->sinc_table[i + 4] = +#ifdef DOUBLE_PRECISION + sinc (st->cutoff, (i / (double) st->oversample - st->filt_len / 2), +#else sinc (st->cutoff, (i / (float) st->oversample - st->filt_len / 2), +#endif st->filt_len, quality_map[st->quality].window_func); #ifdef FIXED_POINT st->resampler_ptr = resampler_basic_interpolate_single; #else +#ifdef DOUBLE_PRECISION + st->resampler_ptr = resampler_basic_interpolate_double; +#else if (st->quality > 8) st->resampler_ptr = resampler_basic_interpolate_double; else st->resampler_ptr = resampler_basic_interpolate_single; #endif +#endif /*fprintf (stderr, "resampler uses interpolated sinc table and normalised cutoff %f\n", cutoff); */ } st->int_advance = st->num_rate / st->den_rate; @@ -964,11 +1008,18 @@ spx_uint32_t channel_index, const spx_int16_t * in, spx_uint32_t * in_len, spx_int16_t * out, spx_uint32_t * out_len) #else +#ifdef DOUBLE_PRECISION +EXPORT int +speex_resampler_process_float (SpeexResamplerState * st, + spx_uint32_t channel_index, const double *in, spx_uint32_t * in_len, + double *out, spx_uint32_t * out_len) +#else EXPORT int speex_resampler_process_float (SpeexResamplerState * st, spx_uint32_t channel_index, const float *in, spx_uint32_t * in_len, float *out, spx_uint32_t * out_len) #endif +#endif { int j; spx_uint32_t ilen = *in_len; @@ -1086,9 +1137,16 @@ return RESAMPLER_ERR_SUCCESS; } +#ifdef DOUBLE_PRECISION +EXPORT int +speex_resampler_process_interleaved_float (SpeexResamplerState * st, + const double *in, spx_uint32_t * in_len, double *out, + spx_uint32_t * out_len) +#else EXPORT int speex_resampler_process_interleaved_float (SpeexResamplerState * st, const float *in, spx_uint32_t * in_len, float *out, spx_uint32_t * out_len) +#endif { spx_uint32_t i; int istride_save, ostride_save; diff -Naur old/speex_resampler.h new/speex_resampler.h --- old/speex_resampler.h 2009-11-10 12:18:09.000000000 +0100 +++ new/speex_resampler.h 2009-11-10 12:19:09.000000000 +0100 @@ -77,10 +77,10 @@ #define speex_resampler_reset_mem CAT_PREFIX(RANDOM_PREFIX,_resampler_reset_mem) #define speex_resampler_strerror CAT_PREFIX(RANDOM_PREFIX,_resampler_strerror) -#define spx_int16_t short -#define spx_int32_t int -#define spx_uint16_t unsigned short -#define spx_uint32_t unsigned int +#define spx_int16_t gint16 +#define spx_int32_t gint32 +#define spx_uint16_t guint16 +#define spx_uint32_t guint32 #else /* OUTSIDE_SPEEX */ @@ -166,12 +166,21 @@ * @param out Output buffer * @param out_len Size of the output buffer. Returns the number of samples written */ +#ifdef DOUBLE_PRECISION +int speex_resampler_process_float(SpeexResamplerState *st, + spx_uint32_t channel_index, + const double *in, + spx_uint32_t *in_len, + double *out, + spx_uint32_t *out_len); +#else int speex_resampler_process_float(SpeexResamplerState *st, spx_uint32_t channel_index, const float *in, spx_uint32_t *in_len, float *out, spx_uint32_t *out_len); +#endif /** Resample an int array. The input and output buffers must *not* overlap. * @param st Resampler state @@ -199,11 +208,19 @@ * @param out_len Size of the output buffer. Returns the number of samples written. * This is all per-channel. */ +#ifdef DOUBLE_PRECISION +int speex_resampler_process_interleaved_float(SpeexResamplerState *st, + const double *in, + spx_uint32_t *in_len, + double *out, + spx_uint32_t *out_len); +#else int speex_resampler_process_interleaved_float(SpeexResamplerState *st, const float *in, spx_uint32_t *in_len, float *out, spx_uint32_t *out_len); +#endif /** Resample an interleaved int array. The input and output buffers must *not* overlap. * @param st Resampler state