ggml-cpu: add 4 element loops for fp32->fp16

Signed-off-by: Aaron Teo <[email protected]>

Author: taronaeo

ggml/src/ggml-cpu/ggml-cpu.c

@@ -3146,8 +3146,15 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
         float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0));
         float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4));
         uint16x8_t v_xd = vec_round_from_fp32(v_xh, v_xl, 0);
-        uint16x8_t v_xfp16 = vec_convert_to_fp16(v_xd, 0);
-        vec_xst(v_xfp16, 0, (ggml_fp16_t *)(y + i));
+        uint16x8_t v_y = vec_convert_to_fp16(v_xd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
+    for (; i + 3 < n; i += 4) {
+        float32x4_t v_x = vec_xl(0, (const float *)(x + i));
+        float32x4_t v_zero = vec_splats(0.0f);
+        uint16x8_t v_xd = vec_round_from_fp32(v_x, v_zero, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_xd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
     }
 #endif
     for (; i < n; ++i) {