Merge pull request #111 from HecaiYuan/master

loongarch: add loongarch simd opt

test/CMakeLists.txt

@@ -2,6 +2,14 @@ function(add_kissfft_test_executable NAME)
     add_kissfft_executable(${NAME} ${ARGN})
     target_include_directories(${NAME} PRIVATE ..)
 
+    if(HAVE_LASX)
+        target_compile_options(${NAME} PRIVATE -mlasx)
+    endif()
+
+    if(HAVE_LSX)
+        target_compile_options(${NAME} PRIVATE -mlsx)
+    endif()
+
     add_test(NAME ${NAME} COMMAND ${NAME})
     set_tests_properties(${NAME} PROPERTIES TIMEOUT 3600)
 endfunction()
@@ -46,7 +54,11 @@ add_kissfft_test_executable(testcpp testcpp.cc)
 if(KISSFFT_DATATYPE MATCHES "^simd$")
     add_kissfft_test_executable(tsimd test_simd.c)
     target_compile_definitions(tsimd PRIVATE USE_SIMD)
-    if (NOT MSVC)
+    if(HAVE_LASX)
+        target_compile_options(kissfft PRIVATE -mlasx)
+    elseif(HAVE_LSX)
+        target_compile_options(kissfft PRIVATE -mlsx)
+    elseif(NOT MSVC)
         target_compile_options(kissfft PRIVATE -msse)
     else()
         target_compile_options(kissfft PRIVATE "/arch:SSE")

test/Makefile

@@ -138,7 +138,13 @@ endif
 
 $(TESTSIMD): test_simd.c
 ifeq "$(KISSFFT_DATATYPE)" "simd"
+ifeq "$(HAVE_LASX)" "lasx"
+	$(CC) -o $@ -g $(CFLAGS) -DUSE_SIMD=1 -DHAVE_LASX=1 -mlasx $< -L.. -l$(KISSFFTLIB_SHORTNAME) -lm
+else 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
+endif
 else
 	$(error ERROR: This test makes sense only with KISSFFT_DATATYPE=simd)
 endif

test/benchkiss.c

@@ -70,8 +70,13 @@ int main(int argc,char ** argv)
         nbytes *= nfft[k];
 
 #ifdef USE_SIMD        
+#ifdef HAVE_LASX
+    numffts /= 8;
+    fprintf(stderr,"since SIMD implementation does 8 ffts at a time, numffts is being reduced to %d\n",numffts);
+#else
     numffts /= 4;
     fprintf(stderr,"since SIMD implementation does 4 ffts at a time, numffts is being reduced to %d\n",numffts);
+#endif
 #endif
 
     buf=(kiss_fft_cpx*)KISS_FFT_MALLOC(nbytes);

test/test_real.c

@@ -22,7 +22,15 @@ static
 kiss_fft_scalar rand_scalar(void) 
 {
 #ifdef USE_SIMD
+#ifdef HAVE_LASX
+    float tmp = rand()-RAND_MAX/2;
+    return (__m256)(__lasx_xvldrepl_w(&tmp, 0));
+#elif defined(HAVE_LSX)
+    float tmp = rand()-RAND_MAX/2;
+    return (__m128)(__lsx_vldrepl_w(&tmp, 0));
+#else
     return _mm_set1_ps(rand()-RAND_MAX/2);
+#endif
 #else
     kiss_fft_scalar s = (kiss_fft_scalar)(rand() -RAND_MAX/2);
     return s/2;

test/test_simd.c

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

test/twotonetest.c

@@ -38,8 +38,16 @@ double two_tone_test( int nfft, int bin1,int bin2)
     /* generate a signal with two tones*/
     for (i = 0; i < nfft; i++) {
 #ifdef USE_SIMD
+#ifdef HAVE_LASX
+        float tmp = (maxrange>>1)*cos(f1*i) + (maxrange>>1)*cos(f2*i);
+        tbuf[i] = (__m256)__lasx_xvldrepl_w(&tmp, 0);
+#elif defined(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)
                              + (maxrange>>1)*cos(f2*i) );
+#endif
 #else        
         tbuf[i] =  (maxrange>>1)*cos(f1*i)
                  + (maxrange>>1)*cos(f2*i);