Use f32 [skip ci]

src/halfutils.c

@@ -1,5 +1,7 @@
 #include "postgres.h"
 
+#include <arm_neon.h>
+
 #include "halfutils.h"
 #include "halfvec.h"
 
@@ -28,9 +30,27 @@ static float
 HalfvecL2SquaredDistanceDefault(int dim, half * ax, half * bx)
 {
 	float		distance = 0.0;
+	int			i = 0;
+
+/* TODO Improve */
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+	int			count = (dim / 4) * 4;
+	float32x4_t dist = vmovq_n_f32(0);
+
+	for (; i < count; i += 4)
+	{
+		float16x4_t axs = vld1_f16((const __fp16 *) (ax + i));
+		float16x4_t bxs = vld1_f16((const __fp16 *) (bx + i));
+		float32x4_t diff = vsubq_f32(vcvt_f32_f16(axs), vcvt_f32_f16(bxs));
+
+		dist = vfmaq_f32(dist, diff, diff);
+	}
+
+	distance += vaddvq_f32(dist);
+#endif
 
 	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
+	for (; i < dim; i++)
 	{
 		float		diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);