loongarch: add lsx simd opt

This commit is contained in:
yuanhecai
2025-08-26 14:21:45 +08:00
parent 9feadb98a9
commit c94458d8f2
11 changed files with 135 additions and 7 deletions

View File

@@ -109,6 +109,21 @@ if(CMAKE_SYSTEM_NAME MATCHES "^(Linux|kFreeBSD|GNU)$" AND NOT CMAKE_CROSSCOMPILI
include(GNUInstallDirs) include(GNUInstallDirs)
endif() endif()
include(CheckCXXSourceCompiles)
if (CMAKE_SYSTEM_PROCESSOR MATCHES "loongarch32|loongarch64")
set (CMAKE_REQUIRED_FLAGS -mlsx)
check_cxx_source_compiles(
"int main() {
#if !defined(__loongarch_sx)
static_assert(false, \"__loongarch_sx is not defined\");
#endif
return 0;
}"
COMPILER_SUPPORT_LSX_FLAG
)
endif()
# #
# Declare PKGINCLUDEDIR for kissfft include path # Declare PKGINCLUDEDIR for kissfft include path
# #
@@ -167,7 +182,10 @@ else()
if(KISSFFT_DATATYPE MATCHES "^simd$") if(KISSFFT_DATATYPE MATCHES "^simd$")
list(APPEND KISSFFT_COMPILE_DEFINITIONS USE_SIMD) list(APPEND KISSFFT_COMPILE_DEFINITIONS USE_SIMD)
if (NOT MSVC) if(COMPILER_SUPPORT_LSX_FLAG)
list(APPEND KISSFFT_COMPILE_DEFINITIONS HAVE_LSX)
target_compile_options(kissfft PRIVATE -mlsx)
elseif(NOT MSVC)
target_compile_options(kissfft PRIVATE -msse) target_compile_options(kissfft PRIVATE -msse)
else() else()
target_compile_options(kissfft PRIVATE "/arch:SSE") target_compile_options(kissfft PRIVATE "/arch:SSE")
@@ -262,6 +280,10 @@ function(add_kissfft_executable NAME)
target_link_libraries(${NAME} PRIVATE m) target_link_libraries(${NAME} PRIVATE m)
endif() endif()
if(COMPILER_SUPPORT_LSX_FLAG)
target_compile_options(${NAME} PRIVATE -mlsx)
endif()
if (NOT KISSFFT_OPENMP) if (NOT KISSFFT_OPENMP)
set_target_properties(${NAME} PROPERTIES set_target_properties(${NAME} PROPERTIES
OUTPUT_NAME "${NAME}-${KISSFFT_DATATYPE}") OUTPUT_NAME "${NAME}-${KISSFFT_DATATYPE}")
@@ -329,7 +351,9 @@ if (KISSFFT_PKGCONFIG)
join_paths(PKGCONFIG_KISSFFT_INCLUDEDIR "\${prefix}" "${CMAKE_INSTALL_INCLUDEDIR}") join_paths(PKGCONFIG_KISSFFT_INCLUDEDIR "\${prefix}" "${CMAKE_INSTALL_INCLUDEDIR}")
if(KISSFFT_DATATYPE MATCHES "^simd$") if(KISSFFT_DATATYPE MATCHES "^simd$")
list(APPEND KISSFFT_COMPILE_DEFINITIONS USE_SIMD) list(APPEND KISSFFT_COMPILE_DEFINITIONS USE_SIMD)
if (NOT MSVC) if(COMPILER_SUPPORT_LSX_FLAG)
set(PKG_KISSFFT_DEFS "${PKG_KISSFFT_DEFS} -mlsx")
elseif(NOT MSVC)
set(PKG_KISSFFT_DEFS "${PKG_KISSFFT_DEFS} -msse") set(PKG_KISSFFT_DEFS "${PKG_KISSFFT_DEFS} -msse")
else() else()
set(PKG_KISSFFT_DEFS "${PKG_KISSFFT_DEFS} /ARCH:SSE") set(PKG_KISSFFT_DEFS "${PKG_KISSFFT_DEFS} /ARCH:SSE")

View File

@@ -72,6 +72,11 @@ ifneq ($(MAKECMDGOALS),clean)
ifeq ($(_UNAME_ARCH),x86_64) ifeq ($(_UNAME_ARCH),x86_64)
CANDIDATE_LIBDIR_NAME = lib64 CANDIDATE_LIBDIR_NAME = lib64
endif endif
ifeq ($(_UNAME_ARCH),loongarch64)
CANDIDATE_LIBDIR_NAME = lib64
HAVE_LSX=lsx
export HAVE_LSX
endif
endif endif
endif endif
@@ -127,7 +132,11 @@ ifeq "$(KISSFFT_DATATYPE)" "int32_t"
else ifeq "$(KISSFFT_DATATYPE)" "int16_t" else ifeq "$(KISSFFT_DATATYPE)" "int16_t"
TYPEFLAGS += -DFIXED_POINT=16 TYPEFLAGS += -DFIXED_POINT=16
else ifeq "$(KISSFFT_DATATYPE)" "simd" else ifeq "$(KISSFFT_DATATYPE)" "simd"
ifeq "$(HAVE_LSX)" "lsx"
TYPEFLAGS += -DUSE_SIMD=1 -DHAVE_LSX=1 -mlsx
else
TYPEFLAGS += -DUSE_SIMD=1 -msse TYPEFLAGS += -DUSE_SIMD=1 -msse
endif
else ifeq "$(KISSFFT_DATATYPE)" "float" else ifeq "$(KISSFFT_DATATYPE)" "float"
TYPEFLAGS += -Dkiss_fft_scalar=$(KISSFFT_DATATYPE) TYPEFLAGS += -Dkiss_fft_scalar=$(KISSFFT_DATATYPE)
else ifeq "$(KISSFFT_DATATYPE)" "double" else ifeq "$(KISSFFT_DATATYPE)" "double"
@@ -273,21 +282,33 @@ testall:
$(MAKE) KISSFFT_DATATYPE=int16_t testsingle $(MAKE) KISSFFT_DATATYPE=int16_t testsingle
# The simd and int32_t types may or may not work on your machine # The simd and int32_t types may or may not work on your machine
$(MAKE) KISSFFT_DATATYPE=int32_t testsingle $(MAKE) KISSFFT_DATATYPE=int32_t testsingle
$(MAKE) KISSFFT_DATATYPE=simd testsingle @if [ "$(HAVE_LSX)" = "lsx" ]; then \
$(MAKE) KISSFFT_DATATYPE=simd HAVE_LSX=lsx testsingle; \
else \
$(MAKE) KISSFFT_DATATYPE=simd testsingle; \
fi
# Static libraries # Static libraries
$(MAKE) KISSFFT_DATATYPE=double KISSFFT_STATIC=1 testsingle $(MAKE) KISSFFT_DATATYPE=double KISSFFT_STATIC=1 testsingle
$(MAKE) KISSFFT_DATATYPE=float KISSFFT_STATIC=1 testsingle $(MAKE) KISSFFT_DATATYPE=float KISSFFT_STATIC=1 testsingle
$(MAKE) KISSFFT_DATATYPE=int16_t KISSFFT_STATIC=1 testsingle $(MAKE) KISSFFT_DATATYPE=int16_t KISSFFT_STATIC=1 testsingle
# The simd and int32_t types may or may not work on your machine # The simd and int32_t types may or may not work on your machine
$(MAKE) KISSFFT_DATATYPE=int32_t KISSFFT_STATIC=1 testsingle $(MAKE) KISSFFT_DATATYPE=int32_t KISSFFT_STATIC=1 testsingle
$(MAKE) KISSFFT_DATATYPE=simd KISSFFT_STATIC=1 testsingle @if [ "$(HAVE_LSX)" = "lsx" ]; then \
$(MAKE) KISSFFT_DATATYPE=simd HAVE_LSX=lsx KISSFFT_STATIC=1 testsingle; \
else \
$(MAKE) KISSFFT_DATATYPE=simd KISSFFT_STATIC=1 testsingle; \
fi
# OpenMP libraries # OpenMP libraries
$(MAKE) KISSFFT_DATATYPE=double KISSFFT_OPENMP=1 testsingle $(MAKE) KISSFFT_DATATYPE=double KISSFFT_OPENMP=1 testsingle
$(MAKE) KISSFFT_DATATYPE=float KISSFFT_OPENMP=1 testsingle $(MAKE) KISSFFT_DATATYPE=float KISSFFT_OPENMP=1 testsingle
$(MAKE) KISSFFT_DATATYPE=int16_t KISSFFT_OPENMP=1 testsingle $(MAKE) KISSFFT_DATATYPE=int16_t KISSFFT_OPENMP=1 testsingle
# The simd and int32_t types may or may not work on your machine # The simd and int32_t types may or may not work on your machine
$(MAKE) KISSFFT_DATATYPE=int32_t KISSFFT_OPENMP=1 testsingle $(MAKE) KISSFFT_DATATYPE=int32_t KISSFFT_OPENMP=1 testsingle
$(MAKE) KISSFFT_DATATYPE=simd KISSFFT_OPENMP=1 testsingle @if [ "$(HAVE_LSX)" = "lsx" ]; then \
$(MAKE) KISSFFT_DATATYPE=simd HAVE_LSX=lsx KISSFFT_OPENMP=1 testsingle; \
else \
$(MAKE) KISSFFT_DATATYPE=simd KISSFFT_OPENMP=1 testsingle; \
fi
$(warning All tests passed!) $(warning All tests passed!)
# #

View File

@@ -128,6 +128,16 @@ struct kiss_fft_state{
# define KISS_FFT_COS(phase) floor(.5+SAMP_MAX * cos (phase)) # define KISS_FFT_COS(phase) floor(.5+SAMP_MAX * cos (phase))
# define KISS_FFT_SIN(phase) floor(.5+SAMP_MAX * sin (phase)) # define KISS_FFT_SIN(phase) floor(.5+SAMP_MAX * sin (phase))
# define HALF_OF(x) ((x)>>1) # define HALF_OF(x) ((x)>>1)
#elif defined(USE_SIMD) && defined(HAVE_LSX)
#define KISS_FFT_COS(phase) ({ \
float __cos_val = cosf(phase); \
(__m128)(__lsx_vldrepl_w(&__cos_val, 0)); \
})
#define KISS_FFT_SIN(phase) ({ \
float __sin_val = sinf(phase); \
(__m128)(__lsx_vldrepl_w(&__sin_val, 0)); \
})
#define HALF_OF(x) ((x) * (__m128)(__lsx_vreplgr2vr_w(0x3F000000))) // 0.5f
#elif defined(USE_SIMD) #elif defined(USE_SIMD)
# define KISS_FFT_COS(phase) _mm_set1_ps( cos(phase) ) # define KISS_FFT_COS(phase) _mm_set1_ps( cos(phase) )
# define KISS_FFT_SIN(phase) _mm_set1_ps( sin(phase) ) # define KISS_FFT_SIN(phase) _mm_set1_ps( sin(phase) )

View File

@@ -48,6 +48,18 @@ extern "C" {
/* User may override KISS_FFT_MALLOC and/or KISS_FFT_FREE. */ /* User may override KISS_FFT_MALLOC and/or KISS_FFT_FREE. */
#ifdef USE_SIMD #ifdef USE_SIMD
#ifdef HAVE_LSX
# include <lsxintrin.h>
# define kiss_fft_scalar __m128
# ifndef KISS_FFT_MALLOC
# define KISS_FFT_MALLOC(nbytes) aligned_alloc(16, KISS_FFT_ALIGN_SIZE_UP(nbytes))
# define KISS_FFT_ALIGN_CHECK(ptr)
# define KISS_FFT_ALIGN_SIZE_UP(size) ((size + 15UL) & ~0xFUL)
# endif
# ifndef KISS_FFT_FREE
# define KISS_FFT_FREE free
# endif
#else
# include <xmmintrin.h> # include <xmmintrin.h>
# define kiss_fft_scalar __m128 # define kiss_fft_scalar __m128
# ifndef KISS_FFT_MALLOC # ifndef KISS_FFT_MALLOC
@@ -58,6 +70,7 @@ extern "C" {
# ifndef KISS_FFT_FREE # ifndef KISS_FFT_FREE
# define KISS_FFT_FREE _mm_free # define KISS_FFT_FREE _mm_free
# endif # endif
#endif
#else #else
# define KISS_FFT_ALIGN_CHECK(ptr) # define KISS_FFT_ALIGN_CHECK(ptr)
# define KISS_FFT_ALIGN_SIZE_UP(size) (size) # define KISS_FFT_ALIGN_SIZE_UP(size) (size)

View File

@@ -93,7 +93,12 @@ void kiss_fftr(kiss_fftr_cfg st,const kiss_fft_scalar *timedata,kiss_fft_cpx *fr
freqdata[0].r = tdc.r + tdc.i; freqdata[0].r = tdc.r + tdc.i;
freqdata[ncfft].r = tdc.r - tdc.i; freqdata[ncfft].r = tdc.r - tdc.i;
#ifdef USE_SIMD #ifdef USE_SIMD
#ifdef HAVE_LSX
freqdata[0].i = (__m128)(__lsx_vreplgr2vr_w(0));
freqdata[ncfft].i = freqdata[0].i;
#else
freqdata[ncfft].i = freqdata[0].i = _mm_set1_ps(0); freqdata[ncfft].i = freqdata[0].i = _mm_set1_ps(0);
#endif
#else #else
freqdata[ncfft].i = freqdata[0].i = 0; freqdata[ncfft].i = freqdata[0].i = 0;
#endif #endif
@@ -146,7 +151,12 @@ void kiss_fftri(kiss_fftr_cfg st,const kiss_fft_cpx *freqdata,kiss_fft_scalar *t
C_ADD (st->tmpbuf[k], fek, fok); C_ADD (st->tmpbuf[k], fek, fok);
C_SUB (st->tmpbuf[ncfft - k], fek, fok); C_SUB (st->tmpbuf[ncfft - k], fek, fok);
#ifdef USE_SIMD #ifdef USE_SIMD
#ifdef HAVE_LSX
__m128 neg_one = (__m128)__lsx_vreplgr2vr_w(0xBF800000); // -1.0f
st->tmpbuf[ncfft - k].i = __lsx_vfmul_s(st->tmpbuf[ncfft - k].i, neg_one);
#else
st->tmpbuf[ncfft - k].i *= _mm_set1_ps(-1.0); st->tmpbuf[ncfft - k].i *= _mm_set1_ps(-1.0);
#endif
#else #else
st->tmpbuf[ncfft - k].i *= -1; st->tmpbuf[ncfft - k].i *= -1;
#endif #endif

View File

@@ -2,6 +2,10 @@ function(add_kissfft_test_executable NAME)
add_kissfft_executable(${NAME} ${ARGN}) add_kissfft_executable(${NAME} ${ARGN})
target_include_directories(${NAME} PRIVATE ..) target_include_directories(${NAME} PRIVATE ..)
if(COMPILER_SUPPORT_LSX_FLAG)
target_compile_options(${NAME} PRIVATE -mlsx)
endif()
add_test(NAME ${NAME} COMMAND ${NAME}) add_test(NAME ${NAME} COMMAND ${NAME})
set_tests_properties(${NAME} PROPERTIES TIMEOUT 3600) set_tests_properties(${NAME} PROPERTIES TIMEOUT 3600)
endfunction() endfunction()
@@ -46,7 +50,9 @@ add_kissfft_test_executable(testcpp testcpp.cc)
if(KISSFFT_DATATYPE MATCHES "^simd$") if(KISSFFT_DATATYPE MATCHES "^simd$")
add_kissfft_test_executable(tsimd test_simd.c) add_kissfft_test_executable(tsimd test_simd.c)
target_compile_definitions(tsimd PRIVATE USE_SIMD) target_compile_definitions(tsimd PRIVATE USE_SIMD)
if (NOT MSVC) if(COMPILER_SUPPORT_LSX_FLAG)
target_compile_options(kissfft PRIVATE -mlsx)
elseif(NOT MSVC)
target_compile_options(kissfft PRIVATE -msse) target_compile_options(kissfft PRIVATE -msse)
else() else()
target_compile_options(kissfft PRIVATE "/arch:SSE") target_compile_options(kissfft PRIVATE "/arch:SSE")

View File

@@ -143,7 +143,11 @@ endif
$(TESTSIMD): test_simd.c $(TESTSIMD): test_simd.c
ifeq "$(KISSFFT_DATATYPE)" "simd" ifeq "$(KISSFFT_DATATYPE)" "simd"
ifeq "$(HAVE_LSX)" "lsx"
$(CC) -o $@ -g $(CFLAGS) -DUSE_SIMD=1 -DHAVE_LSX=1 -mlsx $< -L.. -l$(KISSFFTLIB_SHORTNAME) -lm
else
$(CC) -o $@ -g $(CFLAGS) -DUSE_SIMD=1 -msse $< -L.. -l$(KISSFFTLIB_SHORTNAME) -lm $(CC) -o $@ -g $(CFLAGS) -DUSE_SIMD=1 -msse $< -L.. -l$(KISSFFTLIB_SHORTNAME) -lm
endif
else else
$(error ERROR: This test makes sense only with KISSFFT_DATATYPE=simd) $(error ERROR: This test makes sense only with KISSFFT_DATATYPE=simd)
endif endif

View File

@@ -22,7 +22,12 @@ static
kiss_fft_scalar rand_scalar(void) kiss_fft_scalar rand_scalar(void)
{ {
#ifdef USE_SIMD #ifdef USE_SIMD
#ifdef HAVE_LSX
float tmp = rand()-RAND_MAX/2;
return (__m128)(__lsx_vldrepl_w(&tmp, 0));
#else
return _mm_set1_ps(rand()-RAND_MAX/2); return _mm_set1_ps(rand()-RAND_MAX/2);
#endif
#else #else
kiss_fft_scalar s = (kiss_fft_scalar)(rand() -RAND_MAX/2); kiss_fft_scalar s = (kiss_fft_scalar)(rand() -RAND_MAX/2);
return s/2; return s/2;

View File

@@ -6,16 +6,30 @@ static void test1(void)
int n[2] = {256,256}; int n[2] = {256,256};
size_t nbytes = sizeof(kiss_fft_cpx)*n[0]*n[1]; size_t nbytes = sizeof(kiss_fft_cpx)*n[0]*n[1];
#ifdef HAVE_LSX
kiss_fft_cpx * inbuf = NULL;
kiss_fft_cpx * outbuf = NULL;
if (posix_memalign((void**)&inbuf, 16, nbytes) ||
posix_memalign((void**)&outbuf, 16, nbytes))
return;
#else
kiss_fft_cpx * inbuf = _mm_malloc(nbytes,16); kiss_fft_cpx * inbuf = _mm_malloc(nbytes,16);
kiss_fft_cpx * outbuf = _mm_malloc(nbytes,16); kiss_fft_cpx * outbuf = _mm_malloc(nbytes,16);
#endif
memset(inbuf,0,nbytes); memset(inbuf,0,nbytes);
memset(outbuf,0,nbytes); memset(outbuf,0,nbytes);
kiss_fftnd_cfg cfg = kiss_fftnd_alloc(n,2,is_inverse,0,0); kiss_fftnd_cfg cfg = kiss_fftnd_alloc(n,2,is_inverse,0,0);
kiss_fftnd(cfg,inbuf,outbuf); kiss_fftnd(cfg,inbuf,outbuf);
kiss_fft_free(cfg); kiss_fft_free(cfg);
#ifdef HAVE_LSX
free(inbuf);
free(outbuf);
#else
_mm_free(inbuf); _mm_free(inbuf);
_mm_free(outbuf); _mm_free(outbuf);
#endif
} }
int main(void) int main(void)

View File

@@ -38,8 +38,14 @@ double two_tone_test( int nfft, int bin1,int bin2)
/* generate a signal with two tones*/ /* generate a signal with two tones*/
for (i = 0; i < nfft; i++) { for (i = 0; i < nfft; i++) {
#ifdef USE_SIMD #ifdef USE_SIMD
#ifdef HAVE_LSX
float tmp = (maxrange>>1)*cos(f1*i) + (maxrange>>1)*cos(f2*i);
tbuf[i] = (__m128)__lsx_vldrepl_w(&tmp, 0);
#else
tbuf[i] = _mm_set1_ps( (maxrange>>1)*cos(f1*i) tbuf[i] = _mm_set1_ps( (maxrange>>1)*cos(f1*i)
+ (maxrange>>1)*cos(f2*i) ); + (maxrange>>1)*cos(f2*i) );
#endif
#else #else
tbuf[i] = (maxrange>>1)*cos(f1*i) tbuf[i] = (maxrange>>1)*cos(f1*i)
+ (maxrange>>1)*cos(f2*i); + (maxrange>>1)*cos(f2*i);

View File

@@ -154,8 +154,14 @@ kiss_fastfir_cfg kiss_fastfir_alloc(
for ( i=0; i < st->n_freq_bins; ++i ) { for ( i=0; i < st->n_freq_bins; ++i ) {
#ifdef USE_SIMD #ifdef USE_SIMD
#ifdef HAVE_LSX
__m128 tmp = (__m128)__lsx_vldrepl_w(&scale, 0);
st->fir_freq_resp[i].r = __lsx_vfmul_s(tmp, st->fir_freq_resp[i].r);
st->fir_freq_resp[i].i = __lsx_vfmul_s(tmp, st->fir_freq_resp[i].i);
#else
st->fir_freq_resp[i].r *= _mm_set1_ps(scale); st->fir_freq_resp[i].r *= _mm_set1_ps(scale);
st->fir_freq_resp[i].i *= _mm_set1_ps(scale); st->fir_freq_resp[i].i *= _mm_set1_ps(scale);
#endif
#else #else
st->fir_freq_resp[i].r *= scale; st->fir_freq_resp[i].r *= scale;
st->fir_freq_resp[i].i *= scale; st->fir_freq_resp[i].i *= scale;
@@ -286,7 +292,11 @@ void direct_file_filter(
tmph = imp_resp+nlag; tmph = imp_resp+nlag;
#ifdef REAL_FASTFIR #ifdef REAL_FASTFIR
# ifdef USE_SIMD # ifdef USE_SIMD
# ifdef HAVE_LSX
outval = (__m128)(__lsx_vreplgr2vr_w(0));
#else
outval = _mm_set1_ps(0); outval = _mm_set1_ps(0);
#endif
#else #else
outval = 0; outval = 0;
#endif #endif
@@ -297,7 +307,12 @@ void direct_file_filter(
outval += buf[k] * *tmph; outval += buf[k] * *tmph;
#else #else
# ifdef USE_SIMD # ifdef USE_SIMD
# ifdef HAVE_LSX
outval.i = (__m128)(__lsx_vreplgr2vr_w(0));
outval.r = outval.i;
#else
outval.r = outval.i = _mm_set1_ps(0); outval.r = outval.i = _mm_set1_ps(0);
#endif
#else #else
outval.r = outval.i = 0; outval.r = outval.i = 0;
#endif #endif