diff --git a/src/layer/arm/prelu_arm.cpp b/src/layer/arm/prelu_arm.cpp
index 845a8c217..d1959e0a9 100644
--- a/src/layer/arm/prelu_arm.cpp
+++ b/src/layer/arm/prelu_arm.cpp
@@ -25,6 +25,112 @@ DEFINE_LAYER_CREATOR(PReLU_arm)
 int PReLU_arm::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 {
     int dims = bottom_top_blob.dims;
+    int elempack = bottom_top_blob.elempack;
+
+    if (opt.use_packing_layout)
+    {
+
+    if (elempack == 4)
+    {
+        float32x4_t _zero = vdupq_n_f32(0.f);
+
+        if (dims == 1)
+        {
+            int w = bottom_top_blob.w;
+
+            float* ptr = bottom_top_blob;
+
+            if (num_slope > 1)
+            {
+                const float* slope = slope_data;
+
+                #pragma omp parallel for num_threads(opt.num_threads)
+                for (int i=0; i<w; i++)
+                {
+                    float32x4_t _p = vld1q_f32(ptr);
+                    float32x4_t _slope = vld1q_f32(slope);
+                    uint32x4_t _lemask = vcleq_f32(_p, _zero);
+                    float32x4_t _ps = vmulq_f32(_p, _slope);
+                    _p = vbslq_f32(_lemask, _ps, _p);
+                    vst1q_f32(ptr, _p);
+
+                    ptr += 4;
+                    slope += 4;
+                }
+            }
+            else
+            {
+                float32x4_t _slope = vdupq_n_f32(slope_data[0]);
+
+                #pragma omp parallel for num_threads(opt.num_threads)
+                for (int i=0; i<w; i++)
+                {
+                    float32x4_t _p = vld1q_f32(ptr);
+                    uint32x4_t _lemask = vcleq_f32(_p, _zero);
+                    float32x4_t _ps = vmulq_f32(_p, _slope);
+                    _p = vbslq_f32(_lemask, _ps, _p);
+                    vst1q_f32(ptr, _p);
+
+                    ptr += 4;
+                }
+            }
+        }
+
+        if (dims == 2)
+        {
+            int w = bottom_top_blob.w;
+            int h = bottom_top_blob.h;
+
+            #pragma omp parallel for num_threads(opt.num_threads)
+            for (int i=0; i<h; i++)
+            {
+                float* ptr = bottom_top_blob.row(i);
+                float32x4_t _slope = num_slope > 1 ? vld1q_f32((const float*)slope_data + i * 4) : vdupq_n_f32(slope_data[0]);
+
+                for (int j=0; j<w; j++)
+                {
+                    float32x4_t _p = vld1q_f32(ptr);
+                    uint32x4_t _lemask = vcleq_f32(_p, _zero);
+                    float32x4_t _ps = vmulq_f32(_p, _slope);
+                    _p = vbslq_f32(_lemask, _ps, _p);
+                    vst1q_f32(ptr, _p);
+
+                    ptr += 4;
+                }
+            }
+        }
+
+        if (dims == 3)
+        {
+            int w = bottom_top_blob.w;
+            int h = bottom_top_blob.h;
+            int channels = bottom_top_blob.c;
+            int size = w * h;
+
+            #pragma omp parallel for num_threads(opt.num_threads)
+            for (int q=0; q<channels; q++)
+            {
+                float* ptr = bottom_top_blob.channel(q);
+                float32x4_t _slope = num_slope > 1 ? vld1q_f32((const float*)slope_data + q * 4) : vdupq_n_f32(slope_data[0]);
+
+                for (int i=0; i<size; i++)
+                {
+                    float32x4_t _p = vld1q_f32(ptr);
+                    uint32x4_t _lemask = vcleq_f32(_p, _zero);
+                    float32x4_t _ps = vmulq_f32(_p, _slope);
+                    _p = vbslq_f32(_lemask, _ps, _p);
+                    vst1q_f32(ptr, _p);
+
+                    ptr += 4;
+                }
+            }
+        }
+
+        return 0;
+    }
+
+    } // opt.use_packing_layout
+
     if (dims != 3)
         return PReLU::forward_inplace(bottom_top_blob, opt);