T-Gander/Sunshine
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

perf(deps): switch to upstream nanors runtime dispatching (#5369)

Browse Source
This commit is contained in:
Cameron Gutman
2026-07-03 22:31:23 -04:00
committed by GitHub
parent fbafc49768
commit c42d20eb1a
9 changed files with 9 additions and 352 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
cmake/compile_definitions/common.cmake
View File

@@ -62,6 +62,9 @@ set(SUNSHINE_TARGET_FILES
"${CMAKE_SOURCE_DIR}/third-party/moonlight-common-c/src/Rtsp.h" "${CMAKE_SOURCE_DIR}/third-party/moonlight-common-c/src/Rtsp.h"
"${CMAKE_SOURCE_DIR}/third-party/moonlight-common-c/src/RtspParser.c" "${CMAKE_SOURCE_DIR}/third-party/moonlight-common-c/src/RtspParser.c"
"${CMAKE_SOURCE_DIR}/third-party/moonlight-common-c/src/Video.h" "${CMAKE_SOURCE_DIR}/third-party/moonlight-common-c/src/Video.h"
"${CMAKE_SOURCE_DIR}/third-party/nanors/deps/obl/oblas_common.c"
"${CMAKE_SOURCE_DIR}/third-party/nanors/deps/obl/oblas_lite.c"
"${CMAKE_SOURCE_DIR}/third-party/nanors/rs.c"
"${CMAKE_SOURCE_DIR}/src/upnp.cpp" "${CMAKE_SOURCE_DIR}/src/upnp.cpp"
"${CMAKE_SOURCE_DIR}/src/upnp.h" "${CMAKE_SOURCE_DIR}/src/upnp.h"
"${CMAKE_SOURCE_DIR}/src/cbs.cpp" "${CMAKE_SOURCE_DIR}/src/cbs.cpp"
@@ -116,8 +119,6 @@ set(SUNSHINE_TARGET_FILES
"${CMAKE_SOURCE_DIR}/src/round_robin.h" "${CMAKE_SOURCE_DIR}/src/round_robin.h"
"${CMAKE_SOURCE_DIR}/src/stat_trackers.h" "${CMAKE_SOURCE_DIR}/src/stat_trackers.h"
"${CMAKE_SOURCE_DIR}/src/stat_trackers.cpp" "${CMAKE_SOURCE_DIR}/src/stat_trackers.cpp"
"${CMAKE_SOURCE_DIR}/src/rswrapper.h"
"${CMAKE_SOURCE_DIR}/src/rswrapper.c"
${PLATFORM_TARGET_FILES}) ${PLATFORM_TARGET_FILES})
if(NOT SUNSHINE_ASSETS_DIR_DEF) if(NOT SUNSHINE_ASSETS_DIR_DEF)

5
cmake/targets/common.cmake
View File

@@ -85,11 +85,6 @@ set_source_files_properties("${CMAKE_SOURCE_DIR}/src/upnp.cpp"
DIRECTORY "${CMAKE_SOURCE_DIR}" "${TEST_DIR}" DIRECTORY "${CMAKE_SOURCE_DIR}" "${TEST_DIR}"
PROPERTIES COMPILE_FLAGS -Wno-pedantic) PROPERTIES COMPILE_FLAGS -Wno-pedantic)
# third-party/nanors
set_source_files_properties("${CMAKE_SOURCE_DIR}/src/rswrapper.c"
DIRECTORY "${CMAKE_SOURCE_DIR}" "${TEST_DIR}"
PROPERTIES COMPILE_FLAGS "-ftree-vectorize -funroll-loops")
# third-party/ViGEmClient # third-party/ViGEmClient
set(VIGEM_COMPILE_FLAGS "") set(VIGEM_COMPILE_FLAGS "")
string(APPEND VIGEM_COMPILE_FLAGS "-Wno-unknown-pragmas ") string(APPEND VIGEM_COMPILE_FLAGS "-Wno-unknown-pragmas ")

7
src/main.cpp
View File

@@ -13,6 +13,9 @@
#include <mach-o/dyld.h> #include <mach-o/dyld.h>
#endif #endif
// lib includes
#include <rs.h>
// local includes // local includes
#include "confighttp.h" #include "confighttp.h"
#include "display_device.h" #include "display_device.h"
@@ -27,10 +30,6 @@
#include "upnp.h" #include "upnp.h"
#include "video.h" #include "video.h"
extern "C" {
#include "rswrapper.h"
}
using namespace std::literals; using namespace std::literals;
std::map<int, std::function<void()>> signal_handlers; ///< Signal handlers. std::map<int, std::function<void()>> signal_handlers; ///< Signal handlers.

233
src/rswrapper.c
View File

@@ -1,233 +0,0 @@
/**
* @file src/rswrapper.c
* @brief Wrappers for nanors vectorization with different ISA options
*/
// _FORTIY_SOURCE can cause some versions of GCC to try to inline
// memset() with incompatible target options when compiling rs.c
#ifdef _FORTIFY_SOURCE
#undef _FORTIFY_SOURCE
#endif
// The assert() function is decorated with __cold on macOS which
// is incompatible with Clang's target multiversioning feature
#ifndef NDEBUG
/**
* @def NDEBUG
* @brief Macro for NDEBUG.
*/
#define NDEBUG
#endif
/**
* @def DECORATE_FUNC_I(a, b)
* @brief Macro for DECORATE FUNC i.
*/
#define DECORATE_FUNC_I(a, b) a##b
/**
* @def DECORATE_FUNC(a, b)
* @brief Macro for DECORATE FUNC.
*/
#define DECORATE_FUNC(a, b) DECORATE_FUNC_I(a, b)
// Append an ISA suffix to the public RS API
#define reed_solomon_init DECORATE_FUNC(reed_solomon_init, ISA_SUFFIX)
/**
* @def reed_solomon_new
* @brief Macro for reed solomon new.
*/
#define reed_solomon_new DECORATE_FUNC(reed_solomon_new, ISA_SUFFIX)
/**
* @def reed_solomon_new_static
* @brief Macro for reed solomon new static.
*/
#define reed_solomon_new_static DECORATE_FUNC(reed_solomon_new_static, ISA_SUFFIX)
/**
* @def reed_solomon_release
* @brief Macro for reed solomon release.
*/
#define reed_solomon_release DECORATE_FUNC(reed_solomon_release, ISA_SUFFIX)
/**
* @def reed_solomon_decode
* @brief Macro for reed solomon decode.
*/
#define reed_solomon_decode DECORATE_FUNC(reed_solomon_decode, ISA_SUFFIX)
/**
* @def reed_solomon_encode
* @brief Macro for reed solomon encode.
*/
#define reed_solomon_encode DECORATE_FUNC(reed_solomon_encode, ISA_SUFFIX)
// Append an ISA suffix to internal functions to prevent multiple definition errors
/**
* @def obl_axpy_ref
* @brief Macro for obl axpy ref.
*/
#define obl_axpy_ref DECORATE_FUNC(obl_axpy_ref, ISA_SUFFIX)
/**
* @def obl_scal_ref
* @brief Macro for obl scal ref.
*/
#define obl_scal_ref DECORATE_FUNC(obl_scal_ref, ISA_SUFFIX)
/**
* @def obl_axpyb32_ref
* @brief Macro for obl axpyb32 ref.
*/
#define obl_axpyb32_ref DECORATE_FUNC(obl_axpyb32_ref, ISA_SUFFIX)
/**
* @def obl_axpy
* @brief Macro for obl axpy.
*/
#define obl_axpy DECORATE_FUNC(obl_axpy, ISA_SUFFIX)
/**
* @def obl_scal
* @brief Macro for obl scal.
*/
#define obl_scal DECORATE_FUNC(obl_scal, ISA_SUFFIX)
/**
* @def obl_swap
* @brief Macro for obl swap.
*/
#define obl_swap DECORATE_FUNC(obl_swap, ISA_SUFFIX)
/**
* @def obl_axpyb32
* @brief Macro for obl axpyb32.
*/
#define obl_axpyb32 DECORATE_FUNC(obl_axpyb32, ISA_SUFFIX)
/**
* @def axpy
* @brief Macro for axpy.
*/
#define axpy DECORATE_FUNC(axpy, ISA_SUFFIX)
/**
* @def scal
* @brief Macro for scal.
*/
#define scal DECORATE_FUNC(scal, ISA_SUFFIX)
/**
* @def gemm
* @brief Macro for gemm.
*/
#define gemm DECORATE_FUNC(gemm, ISA_SUFFIX)
/**
* @def invert_mat
* @brief Macro for invert mat.
*/
#define invert_mat DECORATE_FUNC(invert_mat, ISA_SUFFIX)
#if defined(__x86_64) || defined(__x86_64__) || defined(__amd64) || defined(__amd64__) || defined(_M_AMD64)
// Compile a variant for SSSE3
#if defined(__clang__)
#pragma clang attribute push(__attribute__((target("ssse3"))), apply_to = function)
#else
#pragma GCC push_options
#pragma GCC target("ssse3")
#endif
#define ISA_SUFFIX _ssse3
#define OBLAS_SSE3
#include "../third-party/nanors/rs.c"
#undef OBLAS_SSE3
#undef ISA_SUFFIX
#if defined(__clang__)
#pragma clang attribute pop
#else
#pragma GCC pop_options
#endif
// Compile a variant for AVX2
#if defined(__clang__)
#pragma clang attribute push(__attribute__((target("avx2"))), apply_to = function)
#else
#pragma GCC push_options
#pragma GCC target("avx2")
#endif
#define ISA_SUFFIX _avx2
#define OBLAS_AVX2
#include "../third-party/nanors/rs.c"
#undef OBLAS_AVX2
#undef ISA_SUFFIX
#if defined(__clang__)
#pragma clang attribute pop
#else
#pragma GCC pop_options
#endif
// Compile a variant for AVX512BW
#if defined(__clang__)
#pragma clang attribute push(__attribute__((target("avx512f,avx512bw"))), apply_to = function)
#else
#pragma GCC push_options
#pragma GCC target("avx512f,avx512bw")
#endif
#define ISA_SUFFIX _avx512
#define OBLAS_AVX512
#include "../third-party/nanors/rs.c"
#undef OBLAS_AVX512
#undef ISA_SUFFIX
#if defined(__clang__)
#pragma clang attribute pop
#else
#pragma GCC pop_options
#endif
#endif
// Compile a default variant
/**
* @def ISA_SUFFIX
* @brief Macro for ISA SUFFIX.
*/
#define ISA_SUFFIX _def
#include "../third-party/nanors/deps/obl/autoshim.h"
#include "../third-party/nanors/rs.c"
#undef ISA_SUFFIX
#undef reed_solomon_init
#undef reed_solomon_new
#undef reed_solomon_new_static
#undef reed_solomon_release
#undef reed_solomon_decode
#undef reed_solomon_encode
#include "rswrapper.h"
reed_solomon_new_t reed_solomon_new_fn; ///< Reed solomon new.
reed_solomon_release_t reed_solomon_release_fn; ///< Reed solomon release.
reed_solomon_encode_t reed_solomon_encode_fn; ///< Reed solomon encode.
reed_solomon_decode_t reed_solomon_decode_fn; ///< Reed solomon decode.
/**
* @brief This initializes the RS function pointers to the best vectorized version available.
* @details The streaming code will directly invoke these function pointers during encoding.
*/
void reed_solomon_init(void) {
#if defined(__x86_64) || defined(__x86_64__) || defined(__amd64) || defined(__amd64__) || defined(_M_AMD64)
if (__builtin_cpu_supports("avx512f") && __builtin_cpu_supports("avx512bw")) {
reed_solomon_new_fn = reed_solomon_new_avx512;
reed_solomon_release_fn = reed_solomon_release_avx512;
reed_solomon_encode_fn = reed_solomon_encode_avx512;
reed_solomon_decode_fn = reed_solomon_decode_avx512;
reed_solomon_init_avx512();
} else if (__builtin_cpu_supports("avx2")) {
reed_solomon_new_fn = reed_solomon_new_avx2;
reed_solomon_release_fn = reed_solomon_release_avx2;
reed_solomon_encode_fn = reed_solomon_encode_avx2;
reed_solomon_decode_fn = reed_solomon_decode_avx2;
reed_solomon_init_avx2();
} else if (__builtin_cpu_supports("ssse3")) {
reed_solomon_new_fn = reed_solomon_new_ssse3;
reed_solomon_release_fn = reed_solomon_release_ssse3;
reed_solomon_encode_fn = reed_solomon_encode_ssse3;
reed_solomon_decode_fn = reed_solomon_decode_ssse3;
reed_solomon_init_ssse3();
} else
#endif
{
reed_solomon_new_fn = reed_solomon_new_def;
reed_solomon_release_fn = reed_solomon_release_def;
reed_solomon_encode_fn = reed_solomon_encode_def;
reed_solomon_decode_fn = reed_solomon_decode_def;
reed_solomon_init_def();
}
}

69
src/rswrapper.h
View File

@@ -1,69 +0,0 @@
/**
* @file src/rswrapper.h
* @brief Wrappers for nanors vectorization
* @details This is a drop-in replacement for nanors rs.h
*/
#pragma once
// standard includes
#include <stdint.h>
/**
* @def DATA_SHARDS_MAX
* @brief Macro for DATA SHARDS MAX.
*/
#define DATA_SHARDS_MAX 255
/**
* @brief Opaque Reed-Solomon encoder/decoder context.
*/
typedef struct _reed_solomon reed_solomon;
/**
* @brief Function pointer used to create a Reed-Solomon context.
*/
typedef reed_solomon *(*reed_solomon_new_t)(int data_shards, int parity_shards);
/**
* @brief Function pointer used to release a Reed-Solomon context.
*/
typedef void (*reed_solomon_release_t)(reed_solomon *rs);
/**
* @brief Function pointer used to encode Reed-Solomon recovery shards.
*/
typedef int (*reed_solomon_encode_t)(reed_solomon *rs, uint8_t **shards, int nr_shards, int bs);
/**
* @brief Function pointer used to decode Reed-Solomon recovery shards.
*/
typedef int (*reed_solomon_decode_t)(reed_solomon *rs, uint8_t **shards, uint8_t *marks, int nr_shards, int bs);
extern reed_solomon_new_t reed_solomon_new_fn; ///< Reed solomon new.
extern reed_solomon_release_t reed_solomon_release_fn; ///< Reed solomon release.
extern reed_solomon_encode_t reed_solomon_encode_fn; ///< Reed solomon encode.
extern reed_solomon_decode_t reed_solomon_decode_fn; ///< Reed solomon decode.
/**
* @def reed_solomon_new
* @brief Macro for reed solomon new.
*/
#define reed_solomon_new reed_solomon_new_fn
/**
* @def reed_solomon_release
* @brief Macro for reed solomon release.
*/
#define reed_solomon_release reed_solomon_release_fn
/**
* @def reed_solomon_encode
* @brief Macro for reed solomon encode.
*/
#define reed_solomon_encode reed_solomon_encode_fn
/**
* @def reed_solomon_decode
* @brief Macro for reed solomon decode.
*/
#define reed_solomon_decode reed_solomon_decode_fn
/**
* @brief This initializes the RS function pointers to the best vectorized version available.
* @details The streaming code will directly invoke these function pointers during encoding.
*/
void reed_solomon_init(void);

2
src/stream.cpp
View File

@@ -11,11 +11,11 @@
// lib includes // lib includes
#include <boost/endian/arithmetic.hpp> #include <boost/endian/arithmetic.hpp>
#include <openssl/err.h> #include <openssl/err.h>
#include <rs.h>
extern "C" { extern "C" {
// clang-format off // clang-format off
#include <moonlight-common-c/src/Limelight-internal.h> #include <moonlight-common-c/src/Limelight-internal.h>
#include "rswrapper.h"
// clang-format on // clang-format on
} }

36
tests/unit/test_rswrapper.cpp
View File

@@ -1,36 +0,0 @@
/**
* @file tests/unit/test_rswrapper.cpp
* @brief Test src/rswrapper.*
*/
extern "C" {
#include <src/rswrapper.h>
}
#include "../tests_common.h"
TEST(ReedSolomonWrapperTests, InitTest) {
reed_solomon_init();
// Ensure all function pointers were populated
ASSERT_NE(reed_solomon_new, nullptr);
ASSERT_NE(reed_solomon_release, nullptr);
ASSERT_NE(reed_solomon_encode, nullptr);
ASSERT_NE(reed_solomon_decode, nullptr);
}
TEST(ReedSolomonWrapperTests, EncodeTest) {
reed_solomon_init();
auto rs = reed_solomon_new(1, 1);
ASSERT_NE(rs, nullptr);
uint8_t dataShard[16] = {};
uint8_t fecShard[16] = {};
// If we picked the incorrect ISA in our wrapper, we should crash here
uint8_t *shardPtrs[2] = {dataShard, fecShard};
auto ret = reed_solomon_encode(rs, shardPtrs, 2, sizeof(dataShard));
ASSERT_EQ(ret, 0);
reed_solomon_release(rs);
}

2
third-party/moonlight-common-c vendored

Submodule third-party/moonlight-common-c updated: 47b4d338b1...99c45d35ad

2
third-party/nanors vendored

Submodule third-party/nanors updated: 19f07b513e...57ee5e921b