Viswanath Puttagunta
2014-Nov-09 21:34 UTC
[opus] [RFC PATCH v1] arm: kf_bfly4: Introduce ARM neon intrinsics
Hello, This patch introduces ARM NEON Intrinsics to optimize kf_bfly4 routine in celt part of libopus. Using NEON optimized kf_bfly4(_neon) routine helped improve performance of opus_fft_impl function by about 21.4%. The end use case was decoding a music opus ogg file. The end use case saw performance improvement of about 4.47%. This patch has 2 components i. Actual neon code to improve kf_bfly4 ii. Infrastructure to include neon intrinsics into this project I am reasonably confident about part "i" above. However, I need some direction with "ii". With this patch, users can explicitly enable neon intrinsics for SoCs that have ARMv7 NEON VFP support using --enable-armv7-neon-float I enabled this feature with minimal invasion of existing configure.ac and Makefile.am code base. I suspect having runtime function detection and seamless enablement of neon intrinsics without using --enable-armv7-neon-float will take more work and collaboration. Can we in the mean time take this patch as a starting point (ofcourse after due review). My idea is that once this patch gets accepted, work can go forward in two fronts *independently*. i. Optimizing more functions using neon intrinsics ii. Proper way to enable neon intrinsics in configure.ac,Makefile.am etc. More details on how I verified this patch and performance measurements is available at [1] Please let me know your thoughts. [1]: https://docs.google.com/document/d/1l_VWknKMdR_6nn1zIjaawxP2u7p4F3OAt7jBeuAyqe0/edit?usp=sharing Viswanath Puttagunta (1): arm: kf_bfly4: Introduce ARM neon intrinsics Makefile.am | 16 ++++ celt/_kiss_fft_guts.h | 13 +++ celt/arm/kiss_fft_neon.c | 211 ++++++++++++++++++++++++++++++++++++++++++++++ celt/arm/kiss_fft_neon.h | 37 ++++++++ celt/kiss_fft.c | 2 +- celt_headers.mk | 1 + celt_sources.mk | 3 + configure.ac | 14 +++ 8 files changed, 296 insertions(+), 1 deletion(-) create mode 100644 celt/arm/kiss_fft_neon.c create mode 100644 celt/arm/kiss_fft_neon.h -- 1.7.9.5
Viswanath Puttagunta
2014-Nov-09 21:34 UTC
[opus] [RFC PATCH v1] arm: kf_bfly4: Introduce ARM neon intrinsics
Optimize kf_bfly4 function using ARM NEON intrinsics for SoCs that have NEON VFP unit As initial step, only targetting ARMv7-VFP based SoCs. To enable this optimization, use --enable-armv7-neon-float when running configure command. This is disabled by default. --- Makefile.am | 16 ++++ celt/_kiss_fft_guts.h | 13 +++ celt/arm/kiss_fft_neon.c | 211 ++++++++++++++++++++++++++++++++++++++++++++++ celt/arm/kiss_fft_neon.h | 37 ++++++++ celt/kiss_fft.c | 2 +- celt_headers.mk | 1 + celt_sources.mk | 3 + configure.ac | 14 +++ 8 files changed, 296 insertions(+), 1 deletion(-) create mode 100644 celt/arm/kiss_fft_neon.c create mode 100644 celt/arm/kiss_fft_neon.h diff --git a/Makefile.am b/Makefile.am index e20f7b4..94e2419 100644 --- a/Makefile.am +++ b/Makefile.am @@ -38,6 +38,12 @@ CELT_SOURCES += $(CELT_SOURCES_SSE) endif endif +if ARMv7_NEON_INTRINSICS_FLOAT +noinst_LTLIBRARIES = libneon.la +libneon_la_SOURCES = $(CELT_SOURCES_ARM_NEON) +libneon_la_CPPFLAGS = -mfpu=neon-vfpv4 -O3 -I$(top_srcdir)/include +endif + if CPU_ARM CELT_SOURCES += $(CELT_SOURCES_ARM) SILK_SOURCES += $(SILK_SOURCES_ARM) @@ -60,6 +66,10 @@ libopus_la_SOURCES = $(CELT_SOURCES) $(SILK_SOURCES) $(OPUS_SOURCES) libopus_la_LDFLAGS = -no-undefined -version-info @OPUS_LT_CURRENT@:@OPUS_LT_REVISION@:@OPUS_LT_AGE@ libopus_la_LIBADD = $(LIBM) +if ARMv7_NEON_INTRINSICS_FLOAT +libopus_la_LIBADD += ./libneon.la +endif + pkginclude_HEADERS = include/opus.h include/opus_multistream.h include/opus_types.h include/opus_defines.h noinst_HEADERS = $(OPUS_HEAD) $(SILK_HEAD) $(CELT_HEAD) @@ -97,6 +107,9 @@ celt_tests_test_unit_cwrs32_LDADD = $(LIBM) celt_tests_test_unit_dft_SOURCES = celt/tests/test_unit_dft.c celt_tests_test_unit_dft_LDADD = $(LIBM) +if ARMv7_NEON_INTRINSICS_FLOAT +celt_tests_test_unit_dft_LDADD += ./libneon.la +endif celt_tests_test_unit_entropy_SOURCES = celt/tests/test_unit_entropy.c celt_tests_test_unit_entropy_LDADD = $(LIBM) @@ -114,6 +127,9 @@ endif celt_tests_test_unit_mdct_SOURCES = celt/tests/test_unit_mdct.c celt_tests_test_unit_mdct_LDADD = $(LIBM) +if ARMv7_NEON_INTRINSICS_FLOAT +celt_tests_test_unit_mdct_LDADD += ./libneon.la +endif celt_tests_test_unit_rotation_SOURCES = celt/tests/test_unit_rotation.c celt_tests_test_unit_rotation_LDADD = $(LIBM) diff --git a/celt/_kiss_fft_guts.h b/celt/_kiss_fft_guts.h index 5e3d58f..219b431 100644 --- a/celt/_kiss_fft_guts.h +++ b/celt/_kiss_fft_guts.h @@ -34,6 +34,19 @@ and defines typedef struct { kiss_fft_scalar r; kiss_fft_scalar i; }kiss_fft_cpx; */ #include "kiss_fft.h" +#include "arch.h" +/* +void kf_bfly4_c( kiss_fft_cpx * Fout, const size_t fstride, + const kiss_fft_state *st, + int m, int N, int mm); +*/ + +#if defined (ARMv7_NEON_INTRINSICS_FLOAT) +#include "arm/kiss_fft_neon.h" +#define kf_bfly4 kf_bfly4_neon +#else +#define kf_bfly4 kf_bfly4_c +#endif /* Explanation of macros dealing with complex math: diff --git a/celt/arm/kiss_fft_neon.c b/celt/arm/kiss_fft_neon.c new file mode 100644 index 0000000..a2ae42c --- /dev/null +++ b/celt/arm/kiss_fft_neon.c @@ -0,0 +1,211 @@ +/* Copyright (c) 2014 Linaro Ltd + Written by Viswanath Puttagunta */ +/* + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions + are met: + + - Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + + - Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER + OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +#include "../kiss_fft.h" +#include <arm_neon.h> + +#define C_MUL_NEON(m, a, b, t, ones, tv) \ + do{ \ + t = vrev64q_f32(b); \ + m = vmulq_f32(a, b); \ + m = vmulq_f32(m, ones); \ + t = vmulq_f32(a, t); \ + tv = vtrnq_f32(m, t); \ + m = vaddq_f32(tv.val[0], tv.val[1]); \ + }while(0) + +#define ONES_MINUS_ONE 0xbf8000003f800000 //{-1.0, 1.0} +#define MINUS_ONE 0xbf800000bf800000 // {-1.0, -1.0} + +static void kf_bfly4_neon_m1(kiss_fft_cpx *Fout, int N) { + float32x4_t Fout_4[2]; + float32x2_t Fout_2[4]; + float32x2_t scratch_2[2]; + float32x2_t ones_2 = vcreate_f32(ONES_MINUS_ONE); + float32x2_t minusones_2 = vcreate_f32(MINUS_ONE); + float *ai = (float *)Fout; + float *bi = (float *)Fout; + int i; + + /* Consume/update 4 complex Fout values per cycle + * just like normal C code, except each neon + * instruction consumes 1 complex number (2 floats) + * In theory, one could use Q regs instead of + * D regs, but you need to consider case when N is odd + * One can do that if it justifies performance improment + */ + + for (i = 0; i < N; i++) { + Fout_4[0] = vld1q_f32(ai); + ai += 4; + Fout_4[1] = vld1q_f32(ai); + ai += 4; + Fout_2[0] = vget_low_f32(Fout_4[0]); + Fout_2[1] = vget_high_f32(Fout_4[0]); + Fout_2[2] = vget_low_f32(Fout_4[1]); + Fout_2[3] = vget_high_f32(Fout_4[1]); + + scratch_2[0] = vsub_f32(Fout_2[0], Fout_2[2]); + Fout_2[0] = vadd_f32(Fout_2[0], Fout_2[2]); + scratch_2[1] = vadd_f32(Fout_2[1], Fout_2[3]); + Fout_2[2] = vsub_f32(Fout_2[0], scratch_2[1]); + Fout_2[0] = vadd_f32(Fout_2[0], scratch_2[1]); + scratch_2[1] = vsub_f32(Fout_2[1], Fout_2[3]); + + scratch_2[1] = vrev64_f32(scratch_2[1]); + /* scratch_2[1] *= (1, -1) */ + scratch_2[1] = vmul_f32(scratch_2[1], ones_2); + Fout_2[1] = vadd_f32(scratch_2[0], scratch_2[1]); + + /* scratch_2[1] *= (-1, -1) */ + scratch_2[1] = vmul_f32(scratch_2[1], minusones_2); + Fout_2[3] = vadd_f32(scratch_2[0], scratch_2[1]); + + Fout_4[0] = vcombine_f32(Fout_2[0], Fout_2[1]); + Fout_4[1] = vcombine_f32(Fout_2[2], Fout_2[3]); + + vst1q_f32(bi, Fout_4[0]); + bi += 4; + vst1q_f32(bi, Fout_4[1]); + bi += 4; + } +} + +static void kf_bfly4_neon_m8(kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_state *st, + int m, + int N, + int mm) { + int i, j; + float32x4_t scratch[6]; + float32x4_t Fout_4[4]; + float32x2_t ones_2 = vcreate_f32(ONES_MINUS_ONE); + float32x2_t minusones_2 = vcreate_f32(MINUS_ONE); + float32x4_t ones = vcombine_f32(ones_2, ones_2); + float32x4_t minusones = vcombine_f32(minusones_2, minusones_2); + float32x4_t t; + float32x4x2_t tv; + float *tw1, *tw2, *tw3; + float *tw1_2, *tw2_2, *tw3_2; + int fstride_2 = 2*fstride; + int fs_tw1 = 2*fstride_2; + int fs_tw2 = 4*fstride_2; + int fs_tw3 = 6*fstride_2; + int fs_x = 3*fstride_2; + const int m1 = 2*m; + const int m2 = 4*m; // 2*(2*m) + const int m3 = 6*m; // 3*(2*m) + kiss_fft_cpx *Fout_beg = Fout; + float32x4_t tw[3]; + float32x2_t tw_2[6]; + float *ai; + + /* m is guaranteed to be a multiple of 4 + * however, this function will function properly + * so long as m is a multiple of 2 + */ + celt_assert((m%2 == 0)); + + for (i = 0; i < N; i++) { + Fout = Fout_beg + i*mm; + ai = (float *) Fout; + tw1 = tw2 = tw3 = (float *)st->twiddles; + tw1_2 = tw1 + fstride_2; + tw2_2 = tw1 + fs_tw1; //fstride_2*2; + tw3_2 = tw1 + fs_x; //fstride_2*3; + + /* In each cycle, left 2 lanes of q regs have data + * corresponding to m even and right 2 lanes have + * data corresponding to m odd. + * So, instead of consuming/updating 4 complex + * values of Fout per cycle, we consume/update 8 + * complex values of Fout + */ + for (j = 0; j < m/2; j++) { + Fout_4[0] = vld1q_f32(ai); + Fout_4[1] = vld1q_f32(ai+m1); + Fout_4[2] = vld1q_f32(ai+m2); + Fout_4[3] = vld1q_f32(ai+m3); + + tw_2[0] = vld1_f32(tw1); tw_2[1] = vld1_f32(tw1_2); + tw[0] = vcombine_f32(tw_2[0], tw_2[1]); + tw1 += fs_tw1; + tw1_2 += fs_tw1; + + tw_2[2] = vld1_f32(tw2); tw_2[3] = vld1_f32(tw2_2); + tw[1] = vcombine_f32(tw_2[2], tw_2[3]); + tw2 += fs_tw2; + tw2_2 += fs_tw2; + + tw_2[4] = vld1_f32(tw3); tw_2[5] = vld1_f32(tw3_2); + tw[2] = vcombine_f32(tw_2[4], tw_2[5]); + tw3 += fs_tw3; + tw3_2 += fs_tw3; + + C_MUL_NEON(scratch[0], Fout_4[1], tw[0], t, ones, tv); + C_MUL_NEON(scratch[1], Fout_4[2], tw[1], t, ones, tv); + C_MUL_NEON(scratch[2], Fout_4[3], tw[2], t, ones, tv); + + scratch[5] = vsubq_f32(Fout_4[0], scratch[1]); + Fout_4[0] = vaddq_f32(Fout_4[0], scratch[1]); + scratch[3] = vaddq_f32(scratch[0], scratch[2]); + scratch[4] = vsubq_f32(scratch[0], scratch[2]); + Fout_4[2] = vsubq_f32(Fout_4[0], scratch[3]); + Fout_4[0] = vaddq_f32(Fout_4[0], scratch[3]); + + scratch[4] = vrev64q_f32(scratch[4]); + scratch[4] = vmulq_f32(scratch[4], ones); + Fout_4[1] = vaddq_f32(scratch[5], scratch[4]); + scratch[4] = vmulq_f32(scratch[4], minusones); + Fout_4[3] = vaddq_f32(scratch[5], scratch[4]); + + vst1q_f32(ai, Fout_4[0]); + vst1q_f32(ai+m1, Fout_4[1]); + vst1q_f32(ai+m2, Fout_4[2]); + vst1q_f32(ai+m3, Fout_4[3]); + + ai += 4; + } + } +} + +/* NEON Implementation of kf_bfly_c */ +void kf_bfly4_neon(kiss_fft_cpx * Fout, + const size_t fstride, + const kiss_fft_state *st, + int m, + int N, + int mm) { + if (m == 1) { + /* Degenerate case whre all twiddles are 1 */ + kf_bfly4_neon_m1(Fout, N); + } else { + /* m is guaranteed to be a multiple of 4 */ + kf_bfly4_neon_m8(Fout, fstride, st, m, N, mm); + } +} diff --git a/celt/arm/kiss_fft_neon.h b/celt/arm/kiss_fft_neon.h new file mode 100644 index 0000000..b332e87 --- /dev/null +++ b/celt/arm/kiss_fft_neon.h @@ -0,0 +1,37 @@ +/* Copyright (c) 2014 Linaro Ltd + Written by Viswanath Puttagunta */ +/* + Redistribution and use in source and binary forms, with or without + modification, are permitted provided that the following conditions + are met: + + - Redistributions of source code must retain the above copyright + notice, this list of conditions and the following disclaimer. + + - Redistributions in binary form must reproduce the above copyright + notice, this list of conditions and the following disclaimer in the + documentation and/or other materials provided with the distribution. + + THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER + OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. +*/ + +#ifndef KISS_FFT_NEON_H +#define KISS_FFT_NEON_H + +#include "../kiss_fft.h" + +/* NEON Implementation of kf_bfly_c */ +void kf_bfly4_neon( kiss_fft_cpx * Fout, const size_t fstride, + const kiss_fft_state *st, + int m, int N, int mm); +#endif diff --git a/celt/kiss_fft.c b/celt/kiss_fft.c index cc487fc..5c81f78 100644 --- a/celt/kiss_fft.c +++ b/celt/kiss_fft.c @@ -101,7 +101,7 @@ static void kf_bfly2( } } -static void kf_bfly4( +void kf_bfly4_c( kiss_fft_cpx * Fout, const size_t fstride, const kiss_fft_state *st, diff --git a/celt_headers.mk b/celt_headers.mk index 5bb193e..db4c2f3 100644 --- a/celt_headers.mk +++ b/celt_headers.mk @@ -37,5 +37,6 @@ celt/arm/fixed_armv5e.h \ celt/arm/kiss_fft_armv4.h \ celt/arm/kiss_fft_armv5e.h \ celt/arm/pitch_arm.h \ +celt/arm/kiss_fft_neon.h \ celt/x86/pitch_sse.h \ celt/x86/x86cpu.h diff --git a/celt_sources.mk b/celt_sources.mk index 20b1b1b..b27bf3e 100644 --- a/celt_sources.mk +++ b/celt_sources.mk @@ -32,3 +32,6 @@ celt/arm/celt_pitch_xcorr_arm.s CELT_AM_SOURCES_ARM_ASM = \ celt/arm/armopts.s.in + +CELT_SOURCES_ARM_NEON = \ +celt/arm/kiss_fft_neon.c diff --git a/configure.ac b/configure.ac index 9b2f51f..488295f 100644 --- a/configure.ac +++ b/configure.ac @@ -193,6 +193,20 @@ AC_ARG_ENABLE([intrinsics], [AS_HELP_STRING([--enable-intrinsics], [Enable intrinsics optimizations (only for fixed point x86)])],, [enable_intrinsics=no]) +AC_ARG_ENABLE([armv7-neon-intrinsics], + [AS_HELP_STRING([--enable-armv7-neon-intrinsics], [Enable intrinsics optimizations for ARMv7 NEON VFP])],, + [enable_armv7_neon_intrinsics=no]) + +AS_IF([test "${enable_armv7_neon_intrinsics}" = "yes"], + [ + AC_CHECK_HEADERS([arm_neon.h]) + AC_DEFINE([ARMv7_NEON_INTRINSICS_FLOAT], [1], [Use ARMv7 NEON Intrinsics]) + AM_CONDITIONAL([ARMv7_NEON_INTRINSICS_FLOAT],[true]) + ], + [ + AM_CONDITIONAL([ARMv7_NEON_INTRINSICS_FLOAT],[false]) + ]) + rtcd_support=no cpu_arm=no -- 1.7.9.5
Viswanath Puttagunta
2014-Nov-14 20:53 UTC
[opus] [RFC PATCH v1] arm: kf_bfly4: Introduce ARM neon intrinsics
Hello, I see from [1] Jean-Marc comment that opus project is open to adopting ARM focussed fft optimizations in NE10 project/library at [2]. In light of this information, if and when fft implementations for 60, 120, 240, 480 become available in NE10 library, a. Will the approach of enabling this optimization using --enable-armv7-neon-float or similar flag be acceptable to begin with? or b. Should the "AS_IF([test x"${enable_asm}" = x"yes"]".. section in configure.ac be re-written? Please advise. [1]: https://code.google.com/p/webrtc/issues/detail?id=3350&can=1&q=ne10&colspec=ID%20Pri%20Mstone%20ReleaseBlock%20Area%20Status%20Owner%20Summary [2]: https://github.com/projectNe10/Ne10 Regards, Vish On 9 November 2014 15:34, Viswanath Puttagunta <viswanath.puttagunta at linaro.org> wrote:> > Hello, > > This patch introduces ARM NEON Intrinsics to optimize > kf_bfly4 routine in celt part of libopus. > > Using NEON optimized kf_bfly4(_neon) routine helped improve > performance of opus_fft_impl function by about 21.4%. The > end use case was decoding a music opus ogg file. The end > use case saw performance improvement of about 4.47%. > > This patch has 2 components > i. Actual neon code to improve kf_bfly4 > ii. Infrastructure to include neon intrinsics into this project > > I am reasonably confident about part "i" above. > However, I need some direction with "ii". > > With this patch, users can explicitly enable neon intrinsics for > SoCs that have ARMv7 NEON VFP support using --enable-armv7-neon-float > > I enabled this feature with minimal invasion of existing configure.ac > and Makefile.am code base. > > I suspect having runtime function detection and seamless enablement > of neon intrinsics without using --enable-armv7-neon-float will take > more work and collaboration. > > Can we in the mean time take this patch as a starting point (ofcourse > after due review). My idea is that once this patch gets accepted, work > can go forward in two fronts *independently*. > > i. Optimizing more functions using neon intrinsics > ii. Proper way to enable neon intrinsics in configure.ac,Makefile.am etc. > > More details on how I verified this patch and performance measurements > is available at [1] > > Please let me know your thoughts. > > [1]: https://docs.google.com/document/d/1l_VWknKMdR_6nn1zIjaawxP2u7p4F3OAt7jBeuAyqe0/edit?usp=sharing > > Viswanath Puttagunta (1): > arm: kf_bfly4: Introduce ARM neon intrinsics > > Makefile.am | 16 ++++ > celt/_kiss_fft_guts.h | 13 +++ > celt/arm/kiss_fft_neon.c | 211 ++++++++++++++++++++++++++++++++++++++++++++++ > celt/arm/kiss_fft_neon.h | 37 ++++++++ > celt/kiss_fft.c | 2 +- > celt_headers.mk | 1 + > celt_sources.mk | 3 + > configure.ac | 14 +++ > 8 files changed, 296 insertions(+), 1 deletion(-) > create mode 100644 celt/arm/kiss_fft_neon.c > create mode 100644 celt/arm/kiss_fft_neon.h > > -- > 1.7.9.5 >
Timothy B. Terriberry
2014-Nov-14 21:00 UTC
[opus] [RFC PATCH v1] arm: kf_bfly4: Introduce ARM neon intrinsics
Viswanath Puttagunta wrote:> a. Will the approach of enabling this optimization using > --enable-armv7-neon-float or similar flag be acceptable to begin with? > > or > > b. Should the "AS_IF([test x"${enable_asm}" = x"yes"]".. section in > configure.ac be re-written?Automatic detection of compiler support in configure (in addition to run-time checks) would be preferable. We used a manual switches for, e.g., the SSE 4.2 intrinsics, just because the submitted auto-detection code was not very reliable, and no one has written anything better yet.
Possibly Parallel Threads
- [RFC PATCH v1] arm: kf_bfly4: Introduce ARM neon intrinsics
- [RFC PATCHv1] cover: celt_pitch_xcorr: Introduce ARM neon intrinsics
- [RFC PATCH v2] cover: armv7: celt_pitch_xcorr: Introduce ARM neon intrinsics
- [RFC PATCH v1] arm: kf_bfly4: Introduce ARM neon intrinsics
- [PATCH v1] cover: armv7: celt_pitch_xcorr: Introduce ARM neon intrinsics