first commit

This commit is contained in:
Uladzimir Karpenka
2026-06-01 15:58:45 +03:00
commit 89c8a0e2b5
313 changed files with 945364 additions and 0 deletions
+85
View File
@@ -0,0 +1,85 @@
/* Copyright (c) 2011-2019 Mozilla
2023 Amazon */
/*
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 FOUNDATION 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 DNN_X86_H
#define DNN_X86_H
#include "cpu_support.h"
#include "opus_types.h"
void compute_linear_sse4_1(const LinearLayer *linear, float *out, const float *in);
void compute_activation_sse4_1(float *output, const float *input, int N, int activation);
void compute_conv2d_sse4_1(const Conv2dLayer *conv, float *out, float *mem, const float *in, int height, int hstride, int activation);
void compute_linear_avx2(const LinearLayer *linear, float *out, const float *in);
void compute_activation_avx2(float *output, const float *input, int N, int activation);
void compute_conv2d_avx2(const Conv2dLayer *conv, float *out, float *mem, const float *in, int height, int hstride, int activation);
#ifdef RNN_ENABLE_X86_RTCD
extern void (*const RNN_COMPUTE_LINEAR_IMPL[OPUS_ARCHMASK + 1])(
const LinearLayer *linear,
float *out,
const float *in
);
#define OVERRIDE_COMPUTE_LINEAR
#define compute_linear(linear, out, in, arch) \
((*RNN_COMPUTE_LINEAR_IMPL[(arch) & OPUS_ARCHMASK])(linear, out, in))
extern void (*const RNN_COMPUTE_ACTIVATION_IMPL[OPUS_ARCHMASK + 1])(
float *output,
const float *input,
int N,
int activation
);
#define OVERRIDE_COMPUTE_ACTIVATION
#define compute_activation(output, input, N, activation, arch) \
((*RNN_COMPUTE_ACTIVATION_IMPL[(arch) & OPUS_ARCHMASK])(output, input, N, activation))
extern void (*const RNN_COMPUTE_CONV2D_IMPL[OPUS_ARCHMASK + 1])(
const Conv2dLayer *conv,
float *out,
float *mem,
const float *in,
int height,
int hstride,
int activation
);
#define OVERRIDE_COMPUTE_CONV2D
#define compute_conv2d(conv, out, mem, in, height, hstride, activation, arch) \
((*RNN_COMPUTE_CONV2D_IMPL[(arch) & OPUS_ARCHMASK])(conv, out, mem, in, height, hstride, activation))
#endif
#endif /* DNN_X86_H */
+40
View File
@@ -0,0 +1,40 @@
/* Copyright (c) 2018-2019 Mozilla
2023 Amazon */
/*
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 FOUNDATION 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.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "x86/x86_arch_macros.h"
#ifndef __AVX2__
#error nnet_avx2.c is being compiled without AVX2 enabled
#endif
#define RTCD_ARCH avx2
#include "nnet_arch.h"
+40
View File
@@ -0,0 +1,40 @@
/* Copyright (c) 2018-2019 Mozilla
2023 Amazon */
/*
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 FOUNDATION 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.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "x86/x86_arch_macros.h"
#ifndef __SSE4_1__
#error nnet_sse4_1.c is being compiled without SSE4.1 enabled
#endif
#define RTCD_ARCH sse4_1
#include "nnet_arch.h"
+47
View File
@@ -0,0 +1,47 @@
/* Copyright (c) 2023 Amazon */
/*
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 FOUNDATION 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.
*/
#ifdef _MSC_VER
# ifdef OPUS_X86_MAY_HAVE_SSE
# ifndef __SSE__
# define __SSE__
# endif
# endif
# ifdef OPUS_X86_MAY_HAVE_SSE2
# ifndef __SSE2__
# define __SSE2__
# endif
# endif
# ifdef OPUS_X86_MAY_HAVE_SSE4_1
# ifndef __SSE4_1__
# define __SSE4_1__
# endif
# endif
#endif
+74
View File
@@ -0,0 +1,74 @@
/* Copyright (c) 2018-2019 Mozilla
2023 Amazon */
/*
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 FOUNDATION 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.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "x86/x86cpu.h"
#include "nnet.h"
#ifdef RNN_ENABLE_X86_RTCD
void (*const RNN_COMPUTE_LINEAR_IMPL[OPUS_ARCHMASK + 1])(
const LinearLayer *linear,
float *out,
const float *in
) = {
compute_linear_c, /* non-sse */
MAY_HAVE_SSE4_1(compute_linear), /* sse4.1 */
MAY_HAVE_AVX2(compute_linear) /* avx */
};
void (*const RNN_COMPUTE_ACTIVATION_IMPL[OPUS_ARCHMASK + 1])(
float *output,
const float *input,
int N,
int activation
) = {
compute_activation_c, /* non-sse */
MAY_HAVE_SSE4_1(compute_activation), /* sse4.1 */
MAY_HAVE_AVX2(compute_activation) /* avx */
};
void (*const RNN_COMPUTE_CONV2D_IMPL[OPUS_ARCHMASK + 1])(
const Conv2dLayer *conv,
float *out,
float *mem,
const float *in,
int height,
int hstride,
int activation
) = {
compute_conv2d_c, /* non-sse */
MAY_HAVE_SSE4_1(compute_conv2d), /* sse4.1 */
MAY_HAVE_AVX2(compute_conv2d) /* avx */
};
#endif
+166
View File
@@ -0,0 +1,166 @@
/* Copyright (c) 2014, Cisco Systems, INC
Written by XiangMingZhu WeiZhou MinPeng YanWang
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.
*/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "cpu_support.h"
#include "pitch.h"
#include "x86cpu.h"
#ifdef RNN_ENABLE_X86_RTCD
#if defined(_MSC_VER)
#include <intrin.h>
static _inline void cpuid(unsigned int CPUInfo[4], unsigned int InfoType)
{
__cpuid((int*)CPUInfo, InfoType);
}
#else
#if defined(CPU_INFO_BY_C)
#include <cpuid.h>
#endif
static void cpuid(unsigned int CPUInfo[4], unsigned int InfoType)
{
#if defined(CPU_INFO_BY_ASM)
#if defined(__i386__) && defined(__PIC__)
/* %ebx is PIC register in 32-bit, so mustn't clobber it. */
__asm__ __volatile__ (
"xchg %%ebx, %1\n"
"cpuid\n"
"xchg %%ebx, %1\n":
"=a" (CPUInfo[0]),
"=r" (CPUInfo[1]),
"=c" (CPUInfo[2]),
"=d" (CPUInfo[3]) :
/* We clear ECX to avoid a valgrind false-positive prior to v3.17.0. */
"0" (InfoType), "2" (0)
);
#else
__asm__ __volatile__ (
"cpuid":
"=a" (CPUInfo[0]),
"=b" (CPUInfo[1]),
"=c" (CPUInfo[2]),
"=d" (CPUInfo[3]) :
/* We clear ECX to avoid a valgrind false-positive prior to v3.17.0. */
"0" (InfoType), "2" (0)
);
#endif
#elif defined(CPU_INFO_BY_C)
/* We use __get_cpuid_count to clear ECX to avoid a valgrind false-positive
prior to v3.17.0.*/
if (!__get_cpuid_count(InfoType, 0, &(CPUInfo[0]), &(CPUInfo[1]), &(CPUInfo[2]), &(CPUInfo[3]))) {
/* Our function cannot fail, but __get_cpuid{_count} can.
Returning all zeroes will effectively disable all SIMD, which is
what we want on CPUs that don't support CPUID. */
CPUInfo[3] = CPUInfo[2] = CPUInfo[1] = CPUInfo[0] = 0;
}
#else
# error "Configured to use x86 RTCD, but no CPU detection method available. " \
"Reconfigure with --disable-rtcd (or send patches)."
#endif
}
#endif
typedef struct CPU_Feature{
/* SIMD: 128-bit */
int HW_SSE;
int HW_SSE2;
int HW_SSE41;
/* SIMD: 256-bit */
int HW_AVX2;
} CPU_Feature;
static void rnn_cpu_feature_check(CPU_Feature *cpu_feature)
{
unsigned int info[4];
unsigned int nIds = 0;
cpuid(info, 0);
nIds = info[0];
if (nIds >= 1){
cpuid(info, 1);
cpu_feature->HW_SSE = (info[3] & (1 << 25)) != 0;
cpu_feature->HW_SSE2 = (info[3] & (1 << 26)) != 0;
cpu_feature->HW_SSE41 = (info[2] & (1 << 19)) != 0;
cpu_feature->HW_AVX2 = (info[2] & (1 << 28)) != 0 && (info[2] & (1 << 12)) != 0;
if (cpu_feature->HW_AVX2 && nIds >= 7) {
cpuid(info, 7);
cpu_feature->HW_AVX2 = cpu_feature->HW_AVX2 && (info[1] & (1 << 5)) != 0;
} else {
cpu_feature->HW_AVX2 = 0;
}
}
else {
cpu_feature->HW_SSE = 0;
cpu_feature->HW_SSE2 = 0;
cpu_feature->HW_SSE41 = 0;
cpu_feature->HW_AVX2 = 0;
}
}
static int rnn_select_arch_impl(void)
{
CPU_Feature cpu_feature;
int arch;
rnn_cpu_feature_check(&cpu_feature);
arch = 0;
if (!cpu_feature.HW_SSE41)
{
return arch;
}
arch++;
if (!cpu_feature.HW_AVX2)
{
return arch;
}
arch++;
return arch;
}
int rnn_select_arch(void) {
int arch = rnn_select_arch_impl();
#ifdef FUZZING
/* Randomly downgrade the architecture. */
arch = rand()%(arch+1);
#endif
return arch;
}
#endif
+88
View File
@@ -0,0 +1,88 @@
/* Copyright (c) 2014, Cisco Systems, INC
Written by XiangMingZhu WeiZhou MinPeng YanWang
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.
*/
#if !defined(X86CPU_H)
# define X86CPU_H
# define MAY_HAVE_SSE4_1(name) name ## _sse4_1
# define MAY_HAVE_AVX2(name) name ## _avx2
# ifdef RNN_ENABLE_X86_RTCD
int opus_select_arch(void);
# endif
# if defined(__SSE2__)
# include "common.h"
/*MOVD should not impose any alignment restrictions, but the C standard does,
and UBSan will report errors if we actually make unaligned accesses.
Use this to work around those restrictions (which should hopefully all get
optimized to a single MOVD instruction).
GCC implemented _mm_loadu_si32() since GCC 11; HOWEVER, there is a bug!
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99754
LLVM implemented _mm_loadu_si32() since Clang 8.0, however the
__clang_major__ version number macro is unreliable, as vendors
(specifically, Apple) will use different numbering schemes than upstream.
Clang's advice is "use feature detection", but they do not provide feature
detection support for specific SIMD functions.
We follow the approach from the SIMDe project and instead detect unrelated
features that should be available in the version we want (see
<https://github.com/simd-everywhere/simde/blob/master/simde/simde-detect-clang.h>).*/
# if defined(__clang__)
# if __has_warning("-Wextra-semi-stmt") || \
__has_builtin(__builtin_rotateleft32)
# define OPUS_CLANG_8 (1)
# endif
# endif
# if !defined(_MSC_VER) && !OPUS_GNUC_PREREQ(11,3) && !defined(OPUS_CLANG_8)
# include <string.h>
# include <emmintrin.h>
# ifdef _mm_loadu_si32
# undef _mm_loadu_si32
# endif
# define _mm_loadu_si32 WORKAROUND_mm_loadu_si32
static inline __m128i WORKAROUND_mm_loadu_si32(void const* mem_addr) {
int val;
memcpy(&val, mem_addr, sizeof(val));
return _mm_cvtsi32_si128(val);
}
# elif defined(_MSC_VER)
/* MSVC needs this for _mm_loadu_si32 */
# include <immintrin.h>
# endif
# define OP_CVTEPI8_EPI32_M32(x) \
(_mm_cvtepi8_epi32(_mm_loadu_si32(x)))
# define OP_CVTEPI16_EPI32_M64(x) \
(_mm_cvtepi16_epi32(_mm_loadl_epi64((__m128i *)(void*)(x))))
# endif
#endif