On 2024/05/01 11:21, Jan Stary wrote: > Hi, > > On Apr 26 20:46:51, b...@comstyle.com wrote: > > Implement SSE2 lrint() and lrintf() on amd64. > > I don't think this is worth the added complexity: > seven more patches to have a different lrint()? > Does it make the resampling noticably better/faster?
Playing with the benchmark mentioned in https://github.com/libsndfile/libsamplerate/issues/187 suggests that it's going to be *hugely* faster with clang (and a bit faster with gcc). > Also, the patch changes the CONFIGURE_STYLE from gnu to autoreconf > and hardwires the autoconf and automake version (without explicitly > depending on them) - presumably because configure.ac is patched > so ./configure must be recreated. > This seems to basicaly replicate the SSE2 github commit(s) in the port. Exactly, that is the best way to do it. For code changes use a patch direct from an upstream commit, or as close as possible. For autoconf, patch the input files and regenerate. Patching the output files is much more likely to result in breakage at update time. > I would wait for a release that will already contain this. The last release was 3 years ago so that could be quite some wait. An alternative would be to bump the port to a hash-based git archive download, though I prefer making this sort of change as a targetted patch like this. ok sthen FWIW > > Index: Makefile > > =================================================================== > > RCS file: /cvs/ports/audio/libsamplerate/Makefile,v > > retrieving revision 1.27 > > diff -u -p -u -p -r1.27 Makefile > > --- Makefile 5 Sep 2023 16:13:38 -0000 1.27 > > +++ Makefile 27 Apr 2024 00:26:05 -0000 > > @@ -2,7 +2,7 @@ COMMENT= audio sample rate conversion li > > > > VER= 0.2.2 > > DISTNAME= libsamplerate-${VER} > > -REVISION= 0 > > +REVISION= 1 > > CATEGORIES= audio > > EXTRACT_SUFX= .tar.xz > > > > @@ -18,7 +18,9 @@ SITES= https://github.com/libsndfile/lib > > > > WANTLIB= m > > > > -CONFIGURE_STYLE=gnu > > +AUTOCONF_VERSION= 2.71 > > +AUTOMAKE_VERSION= 1.16 > > +CONFIGURE_STYLE=autoreconf > > CONFIGURE_ARGS= --disable-cpu-clip \ > > --disable-fftw \ > > --disable-sndfile > > Index: patches/patch-configure_ac > > =================================================================== > > RCS file: patches/patch-configure_ac > > diff -N patches/patch-configure_ac > > --- /dev/null 1 Jan 1970 00:00:00 -0000 > > +++ patches/patch-configure_ac 27 Apr 2024 00:26:05 -0000 > > @@ -0,0 +1,43 @@ > > +- Implement SSE2 lrint() and lrintf() > > + 7a81766b14fa03e97822cf1e0b1651648df13116 > > +- use sse2 intrinsics for lrint/lrintf only on windows x64 > > + c01e2405612ad3561bf93e8e6dddb9ba0dffe4d9 > > +- sse2 lrint/lrintf updates > > + c164eaa25ffdeedc7d25e731172cc45a25f483d4 > > + > > +Index: configure.ac > > +--- configure.ac.orig > > ++++ configure.ac > > +@@ -89,7 +89,7 @@ m4_define([abi_version_patch], [lt_revision]) > > + > > + dnl > > ==================================================================================== > > + > > +-AC_CHECK_HEADERS([stdbool.h stdint.h sys/times.h unistd.h]) > > ++AC_CHECK_HEADERS([stdbool.h stdint.h sys/times.h unistd.h immintrin.h]) > > + > > + dnl > > ==================================================================================== > > + dnl Couple of initializations here. Fill in real values later. > > +@@ -105,6 +105,9 @@ AC_ARG_ENABLE([werror], > > + AC_ARG_ENABLE([cpu-clip], > > + [AS_HELP_STRING([--disable-cpu-clip], [disable tricky cpu specific > > clipper])]) > > + > > ++AC_ARG_ENABLE([sse2-lrint], > > ++ [AS_HELP_STRING([--enable-sse2-lrint], [implement lrintf using SSE2 on > > x86 CPUs if possible])]) > > ++ > > + AC_ARG_ENABLE([sndfile], > > + [AS_HELP_STRING([--disable-sndfile], [disable support for sndfile > > (default=autodetect)])], [], [enable_sndfile=auto]) > > + > > +@@ -178,6 +181,13 @@ AS_IF([test "x$enable_cpu_clip" != "xno"], [ > > + > > + AC_DEFINE_UNQUOTED([CPU_CLIPS_POSITIVE], [${ac_cv_c_clip_positive}], > > [Host processor clips on positive float to int conversion.]) > > + AC_DEFINE_UNQUOTED([CPU_CLIPS_NEGATIVE], [${ac_cv_c_clip_negative}], > > [Host processor clips on negative float to int conversion.]) > > ++ > > ++dnl > > ==================================================================================== > > ++dnl Determine if the user enabled lrint implementations using SSE2. > > ++ > > ++AS_IF([test "x$enable_sse2_lrint" = "xyes"], [ > > ++ CFLAGS="$CFLAGS -DENABLE_SSE2_LRINT" > > ++ ]) > > + > > + dnl > > ==================================================================================== > > + dnl Check for libsndfile which is required for the test and example > > programs. > > Index: patches/patch-examples_audio_out_c > > =================================================================== > > RCS file: patches/patch-examples_audio_out_c > > diff -N patches/patch-examples_audio_out_c > > --- /dev/null 1 Jan 1970 00:00:00 -0000 > > +++ patches/patch-examples_audio_out_c 27 Apr 2024 00:26:05 -0000 > > @@ -0,0 +1,19 @@ > > +- Implement SSE2 lrint() and lrintf() > > + 7a81766b14fa03e97822cf1e0b1651648df13116 > > +- use sse2 intrinsics for lrint/lrintf only on windows x64 > > + c01e2405612ad3561bf93e8e6dddb9ba0dffe4d9 > > +- sse2 lrint/lrintf updates > > + c164eaa25ffdeedc7d25e731172cc45a25f483d4 > > + > > +Index: examples/audio_out.c > > +--- examples/audio_out.c.orig > > ++++ examples/audio_out.c > > +@@ -960,7 +960,7 @@ solaris_play (get_audio_callback_t callback, AUDIO_OUT > > + > > + while ((read_frames = callback (callback_data, float_buffer, BUFFER_LEN > > / solaris_out->channels))) > > + { for (k = 0 ; k < read_frames * solaris_out->channels ; k++) > > +- buffer [k] = lrint (32767.0 * float_buffer [k]) ; > > ++ buffer [k] = psf_lrint (32767.0 * float_buffer [k]) ; > > + write (solaris_out->fd, buffer, read_frames * > > solaris_out->channels * sizeof (short)) ; > > + } ; > > + > > Index: patches/patch-src_common_h > > =================================================================== > > RCS file: patches/patch-src_common_h > > diff -N patches/patch-src_common_h > > --- /dev/null 1 Jan 1970 00:00:00 -0000 > > +++ patches/patch-src_common_h 27 Apr 2024 00:26:05 -0000 > > @@ -0,0 +1,98 @@ > > +- Implement SSE2 lrint() and lrintf() > > + 7a81766b14fa03e97822cf1e0b1651648df13116 > > +- use sse2 intrinsics for lrint/lrintf only on windows x64 > > + c01e2405612ad3561bf93e8e6dddb9ba0dffe4d9 > > +- sse2 lrint/lrintf updates > > + c164eaa25ffdeedc7d25e731172cc45a25f483d4 > > + > > +Index: src/common.h > > +--- src/common.h.orig > > ++++ src/common.h > > +@@ -14,6 +14,36 @@ > > + #include <stdbool.h> > > + #endif > > + > > ++#if defined(__x86_64__) || defined(_M_X64) > > ++# define HAVE_SSE2_INTRINSICS > > ++#elif defined(ENABLE_SSE2_LRINT) && (defined(_M_IX86) || > > defined(__i386__)) > > ++# if defined(_MSC_VER) > > ++# define HAVE_SSE2_INTRINSICS > > ++# elif defined(__clang__) > > ++# ifdef __SSE2__ > > ++# define HAVE_SSE2_INTRINSICS > > ++# elif (__has_attribute(target)) > > ++# define HAVE_SSE2_INTRINSICS > > ++# define USE_TARGET_ATTRIBUTE > > ++# endif > > ++# elif defined(__GNUC__) > > ++# ifdef __SSE2__ > > ++# define HAVE_SSE2_INTRINSICS > > ++# elif (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 9)) > > ++# define HAVE_SSE2_INTRINSICS > > ++# define USE_TARGET_ATTRIBUTE > > ++# endif > > ++# endif > > ++#endif > > ++ > > ++#ifdef HAVE_SSE2_INTRINSICS > > ++#ifdef HAVE_IMMINTRIN_H > > ++#include <immintrin.h> > > ++#else > > ++#include <emmintrin.h> > > ++#endif > > ++#endif /* HAVE_SSE2_INTRINSICS */ > > ++ > > + #include <math.h> > > + > > + #ifdef HAVE_VISIBILITY > > +@@ -163,6 +193,41 @@ const char* zoh_get_description (int src_enum) ; > > + SRC_STATE *zoh_state_new (int channels, SRC_ERROR *error) ; > > + > > + /*---------------------------------------------------------- > > ++** SIMD optimized math functions. > > ++*/ > > ++ > > ++#ifdef HAVE_SSE2_INTRINSICS > > ++static inline int > > ++#ifdef USE_TARGET_ATTRIBUTE > > ++__attribute__((target("sse2"))) > > ++#endif > > ++psf_lrintf (float x) > > ++{ > > ++ return _mm_cvtss_si32 (_mm_load_ss (&x)) ; > > ++} > > ++static inline int > > ++#ifdef USE_TARGET_ATTRIBUTE > > ++__attribute__((target("sse2"))) > > ++#endif > > ++psf_lrint (double x) > > ++{ > > ++ return _mm_cvtsd_si32 (_mm_load_sd (&x)) ; > > ++} > > ++ > > ++#else > > ++ > > ++static inline int psf_lrintf (float x) > > ++{ > > ++ return lrintf (x) ; > > ++} /* psf_lrintf */ > > ++ > > ++static inline int psf_lrint (double x) > > ++{ > > ++ return lrint (x) ; > > ++} /* psf_lrint */ > > ++#endif > > ++ > > ++/*---------------------------------------------------------- > > + ** Common static inline functions. > > + */ > > + > > +@@ -170,7 +235,7 @@ static inline double > > + fmod_one (double x) > > + { double res ; > > + > > +- res = x - lrint (x) ; > > ++ res = x - psf_lrint (x) ; > > + if (res < 0.0) > > + return res + 1.0 ; > > + > > Index: patches/patch-src_samplerate_c > > =================================================================== > > RCS file: patches/patch-src_samplerate_c > > diff -N patches/patch-src_samplerate_c > > --- /dev/null 1 Jan 1970 00:00:00 -0000 > > +++ patches/patch-src_samplerate_c 27 Apr 2024 00:26:05 -0000 > > @@ -0,0 +1,28 @@ > > +- Implement SSE2 lrint() and lrintf() > > + 7a81766b14fa03e97822cf1e0b1651648df13116 > > +- use sse2 intrinsics for lrint/lrintf only on windows x64 > > + c01e2405612ad3561bf93e8e6dddb9ba0dffe4d9 > > +- sse2 lrint/lrintf updates > > + c164eaa25ffdeedc7d25e731172cc45a25f483d4 > > + > > +Index: src/samplerate.c > > +--- src/samplerate.c.orig > > ++++ src/samplerate.c > > +@@ -445,7 +445,7 @@ src_float_to_short_array (const float *in, short *out, > > + else if (scaled_value <= -32768.f) > > + out [i] = -32768 ; > > + else > > +- out [i] = (short) (lrintf (scaled_value)) ; > > ++ out [i] = (short) (psf_lrintf (scaled_value)) ; > > + } > > + } /* src_float_to_short_array */ > > + > > +@@ -477,7 +477,7 @@ src_float_to_int_array (const float *in, int *out, int > > + continue ; > > + } ; > > + #endif > > +- out [i] = (int) lrint (scaled_value) ; > > ++ out [i] = (int) psf_lrint (scaled_value) ; > > + } ; > > + > > + } /* src_float_to_int_array */ > > Index: patches/patch-src_src_linear_c > > =================================================================== > > RCS file: patches/patch-src_src_linear_c > > diff -N patches/patch-src_src_linear_c > > --- /dev/null 1 Jan 1970 00:00:00 -0000 > > +++ patches/patch-src_src_linear_c 27 Apr 2024 00:26:05 -0000 > > @@ -0,0 +1,28 @@ > > +- Implement SSE2 lrint() and lrintf() > > + 7a81766b14fa03e97822cf1e0b1651648df13116 > > +- use sse2 intrinsics for lrint/lrintf only on windows x64 > > + c01e2405612ad3561bf93e8e6dddb9ba0dffe4d9 > > +- sse2 lrint/lrintf updates > > + c164eaa25ffdeedc7d25e731172cc45a25f483d4 > > + > > +Index: src/src_linear.c > > +--- src/src_linear.c.orig > > ++++ src/src_linear.c > > +@@ -102,7 +102,7 @@ linear_vari_process (SRC_STATE *state, SRC_DATA *data) > > + } ; > > + > > + rem = fmod_one (input_index) ; > > +- priv->in_used += state->channels * lrint (input_index - rem) ; > > ++ priv->in_used += state->channels * psf_lrint (input_index - rem) ; > > + input_index = rem ; > > + > > + /* Main processing loop. */ > > +@@ -128,7 +128,7 @@ linear_vari_process (SRC_STATE *state, SRC_DATA *data) > > + input_index += 1.0 / src_ratio ; > > + rem = fmod_one (input_index) ; > > + > > +- priv->in_used += state->channels * lrint (input_index - rem) ; > > ++ priv->in_used += state->channels * psf_lrint (input_index - > > rem) ; > > + input_index = rem ; > > + } ; > > + > > Index: patches/patch-src_src_sinc_c > > =================================================================== > > RCS file: patches/patch-src_src_sinc_c > > diff -N patches/patch-src_src_sinc_c > > --- /dev/null 1 Jan 1970 00:00:00 -0000 > > +++ patches/patch-src_src_sinc_c 27 Apr 2024 00:26:05 -0000 > > @@ -0,0 +1,148 @@ > > +- Implement SSE2 lrint() and lrintf() > > + 7a81766b14fa03e97822cf1e0b1651648df13116 > > +- use sse2 intrinsics for lrint/lrintf only on windows x64 > > + c01e2405612ad3561bf93e8e6dddb9ba0dffe4d9 > > +- sse2 lrint/lrintf updates > > + c164eaa25ffdeedc7d25e731172cc45a25f483d4 > > + > > +Index: src/src_sinc.c > > +--- src/src_sinc.c.orig > > ++++ src/src_sinc.c > > +@@ -132,7 +132,7 @@ static SRC_STATE_VT sinc_mono_state_vt = > > + > > + static inline increment_t > > + double_to_fp (double x) > > +-{ return (increment_t) (lrint ((x) * FP_ONE)) ; > > ++{ return (increment_t) (psf_lrint ((x) * FP_ONE)) ; > > + } /* double_to_fp */ > > + > > + static inline increment_t > > +@@ -240,7 +240,7 @@ sinc_filter_new (int converter_type, int channels) > > + #endif > > + } > > + > > +- priv->b_len = 3 * (int) lrint ((priv->coeff_half_len + 2.0) / > > priv->index_inc * SRC_MAX_RATIO + 1) ; > > ++ priv->b_len = 3 * (int) psf_lrint ((priv->coeff_half_len + 2.0) > > / priv->index_inc * SRC_MAX_RATIO + 1) ; > > + priv->b_len = MAX (priv->b_len, 4096) ; > > + priv->b_len *= channels ; > > + priv->b_len += 1 ; // There is a <= check against > > samples_in_hand requiring a buffer bigger than the calculation above > > +@@ -458,12 +458,12 @@ sinc_mono_vari_process (SRC_STATE *state, SRC_DATA > > *da > > + count /= MIN (state->last_ratio, data->src_ratio) ; > > + > > + /* Maximum coefficientson either side of center point. */ > > +- half_filter_chan_len = state->channels * (int) (lrint (count) + 1) ; > > ++ half_filter_chan_len = state->channels * (int) (psf_lrint (count) + 1) ; > > + > > + input_index = state->last_position ; > > + > > + rem = fmod_one (input_index) ; > > +- filter->b_current = (filter->b_current + state->channels * lrint > > (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * psf_lrint > > (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + > > + terminate = 1.0 / src_ratio + 1e-20 ; > > +@@ -505,7 +505,7 @@ sinc_mono_vari_process (SRC_STATE *state, SRC_DATA *da > > + input_index += 1.0 / src_ratio ; > > + rem = fmod_one (input_index) ; > > + > > +- filter->b_current = (filter->b_current + state->channels * > > lrint (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * > > psf_lrint (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + } ; > > + > > +@@ -614,12 +614,12 @@ sinc_stereo_vari_process (SRC_STATE *state, SRC_DATA > > * > > + count /= MIN (state->last_ratio, data->src_ratio) ; > > + > > + /* Maximum coefficientson either side of center point. */ > > +- half_filter_chan_len = state->channels * (int) (lrint (count) + 1) ; > > ++ half_filter_chan_len = state->channels * (int) (psf_lrint (count) + 1) ; > > + > > + input_index = state->last_position ; > > + > > + rem = fmod_one (input_index) ; > > +- filter->b_current = (filter->b_current + state->channels * lrint > > (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * psf_lrint > > (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + > > + terminate = 1.0 / src_ratio + 1e-20 ; > > +@@ -660,7 +660,7 @@ sinc_stereo_vari_process (SRC_STATE *state, SRC_DATA * > > + input_index += 1.0 / src_ratio ; > > + rem = fmod_one (input_index) ; > > + > > +- filter->b_current = (filter->b_current + state->channels * > > lrint (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * > > psf_lrint (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + } ; > > + > > +@@ -770,12 +770,12 @@ sinc_quad_vari_process (SRC_STATE *state, SRC_DATA > > *da > > + count /= MIN (state->last_ratio, data->src_ratio) ; > > + > > + /* Maximum coefficientson either side of center point. */ > > +- half_filter_chan_len = state->channels * (int) (lrint (count) + 1) ; > > ++ half_filter_chan_len = state->channels * (int) (psf_lrint (count) + 1) ; > > + > > + input_index = state->last_position ; > > + > > + rem = fmod_one (input_index) ; > > +- filter->b_current = (filter->b_current + state->channels * lrint > > (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * psf_lrint > > (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + > > + terminate = 1.0 / src_ratio + 1e-20 ; > > +@@ -816,7 +816,7 @@ sinc_quad_vari_process (SRC_STATE *state, SRC_DATA *da > > + input_index += 1.0 / src_ratio ; > > + rem = fmod_one (input_index) ; > > + > > +- filter->b_current = (filter->b_current + state->channels * > > lrint (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * > > psf_lrint (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + } ; > > + > > +@@ -925,12 +925,12 @@ sinc_hex_vari_process (SRC_STATE *state, SRC_DATA > > *dat > > + count /= MIN (state->last_ratio, data->src_ratio) ; > > + > > + /* Maximum coefficientson either side of center point. */ > > +- half_filter_chan_len = state->channels * (int) (lrint (count) + 1) ; > > ++ half_filter_chan_len = state->channels * (int) (psf_lrint (count) + 1) ; > > + > > + input_index = state->last_position ; > > + > > + rem = fmod_one (input_index) ; > > +- filter->b_current = (filter->b_current + state->channels * lrint > > (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * psf_lrint > > (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + > > + terminate = 1.0 / src_ratio + 1e-20 ; > > +@@ -971,7 +971,7 @@ sinc_hex_vari_process (SRC_STATE *state, SRC_DATA *dat > > + input_index += 1.0 / src_ratio ; > > + rem = fmod_one (input_index) ; > > + > > +- filter->b_current = (filter->b_current + state->channels * > > lrint (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * > > psf_lrint (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + } ; > > + > > +@@ -1090,12 +1090,12 @@ sinc_multichan_vari_process (SRC_STATE *state, > > SRC_DAT > > + count /= MIN (state->last_ratio, data->src_ratio) ; > > + > > + /* Maximum coefficientson either side of center point. */ > > +- half_filter_chan_len = state->channels * (int) (lrint (count) + 1) ; > > ++ half_filter_chan_len = state->channels * (int) (psf_lrint (count) + 1) ; > > + > > + input_index = state->last_position ; > > + > > + rem = fmod_one (input_index) ; > > +- filter->b_current = (filter->b_current + state->channels * lrint > > (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * psf_lrint > > (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + > > + terminate = 1.0 / src_ratio + 1e-20 ; > > +@@ -1136,7 +1136,7 @@ sinc_multichan_vari_process (SRC_STATE *state, > > SRC_DAT > > + input_index += 1.0 / src_ratio ; > > + rem = fmod_one (input_index) ; > > + > > +- filter->b_current = (filter->b_current + state->channels * > > lrint (input_index - rem)) % filter->b_len ; > > ++ filter->b_current = (filter->b_current + state->channels * > > psf_lrint (input_index - rem)) % filter->b_len ; > > + input_index = rem ; > > + } ; > > + > > Index: patches/patch-src_src_zoh_c > > =================================================================== > > RCS file: patches/patch-src_src_zoh_c > > diff -N patches/patch-src_src_zoh_c > > --- /dev/null 1 Jan 1970 00:00:00 -0000 > > +++ patches/patch-src_src_zoh_c 27 Apr 2024 00:26:05 -0000 > > @@ -0,0 +1,28 @@ > > +- Implement SSE2 lrint() and lrintf() > > + 7a81766b14fa03e97822cf1e0b1651648df13116 > > +- use sse2 intrinsics for lrint/lrintf only on windows x64 > > + c01e2405612ad3561bf93e8e6dddb9ba0dffe4d9 > > +- sse2 lrint/lrintf updates > > + c164eaa25ffdeedc7d25e731172cc45a25f483d4 > > + > > +Index: src/src_zoh.c > > +--- src/src_zoh.c.orig > > ++++ src/src_zoh.c > > +@@ -99,7 +99,7 @@ zoh_vari_process (SRC_STATE *state, SRC_DATA *data) > > + } ; > > + > > + rem = fmod_one (input_index) ; > > +- priv->in_used += state->channels * lrint (input_index - rem) ; > > ++ priv->in_used += state->channels * psf_lrint (input_index - rem) ; > > + input_index = rem ; > > + > > + /* Main processing loop. */ > > +@@ -117,7 +117,7 @@ zoh_vari_process (SRC_STATE *state, SRC_DATA *data) > > + input_index += 1.0 / src_ratio ; > > + rem = fmod_one (input_index) ; > > + > > +- priv->in_used += state->channels * lrint (input_index - rem) ; > > ++ priv->in_used += state->channels * psf_lrint (input_index - > > rem) ; > > + input_index = rem ; > > + } ; > > + > > > > >
