performance improved

kernel/simd/intrin.h

@@ -1,6 +1,26 @@
 #ifndef _INTRIN_H_
 #define _INTRIN_H_
 
+#if defined(_MSC_VER)
+#define BLAS_INLINE __inline
+#elif defined(__GNUC__)
+#if defined(__STRICT_ANSI__)
+#define BLAS_INLINE __inline__
+#else
+#define BLAS_INLINE inline
+#endif
+#else
+#define BLAS_INLINE
+#endif
+
+#ifdef _MSC_VER
+#define BLAS_FINLINE static __forceinline
+#elif defined(__GNUC__)
+#define BLAS_FINLINE static BLAS_INLINE __attribute__((always_inline))
+#else
+#define BLAS_FINLINE static
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif

kernel/simd/intrin_avx.h

@@ -10,6 +10,16 @@ arithmetic
 */
 #define v_add_f32 _mm256_add_ps
 #define v_mul_f32 _mm256_mul_ps
+
+#ifdef HAVE_FMA3
+    // multiply and add, a*b + c
+    #define v_muladd_f32 _mm256_fmadd_ps
+#else
+    // multiply and add, a*b + c
+    BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
+    { return v_add_f32(v_mul_f32(a, b), c); }
+#endif // !HAVE_FMA3
+
 /*
 memory
 */

kernel/simd/intrin_avx512.h

@@ -10,10 +10,12 @@ arithmetic
 */
 #define v_add_f32 _mm512_add_ps
 #define v_mul_f32 _mm512_mul_ps
+// multiply and add, a*b + c
+#define v_muladd_f32 _mm512_fmadd_ps
 /*
 memory
 */
 // unaligned load
 #define v_loadu_f32(PTR) _mm512_loadu_ps((const __m512*)(PTR))
-#define v_storeu_f32(PTR) _mm512_storeu_ps((const __m512*)(PTR))
+#define v_storeu_f32 _mm512_storeu_ps
 #define v_setall_f32(VAL) _mm512_set1_ps(VAL)

kernel/simd/intrin_sse.h

@@ -10,6 +10,17 @@ arithmetic
 */
 #define v_add_f32 _mm_add_ps
 #define v_mul_f32 _mm_mul_ps
+#ifdef HAVE_FMA3
+    // multiply and add, a*b + c
+    #define v_muladd_f32 _mm_fmadd_ps
+#elif defined(HAVE_FMA4)
+    // multiply and add, a*b + c
+    #define v_muladd_f32 _mm_macc_ps
+#else
+    // multiply and add, a*b + c
+    BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
+    { return v_add_f32(v_mul_f32(a, b), c); }
+#endif // HAVE_FMA3
 /*
 memory
 */

kernel/x86_64/daxpy.c

@@ -48,7 +48,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #ifndef HAVE_KERNEL_8
 #include"../simd/intrin.h"
 
-void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
+static void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 {
 	BLASLONG register i = 0;
 	FLOAT a = *alpha;
@@ -57,7 +57,7 @@ void daxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	__alpha =  v_setall_f32(*alpha);
 	const int vstep = v_nlanes_f32;
 	for (; i < n; i += vstep) {
-		tmp = v_add_f32(v_loadu_f32(y + i), v_mul_f32(__alpha, v_loadu_f32( x + i )));
+		tmp = v_muladd_f32(__alpha, v_loadu_f32( x + i ), v_loadu_f32(y + i));
 		v_storeu_f32(y + i, tmp);
 	}
 #else