diff --git a/src/layer/arm/neon_mathfun_fp16s.h b/src/layer/arm/neon_mathfun_fp16s.h
index ff6f91d47..c26ddb673 100644
--- a/src/layer/arm/neon_mathfun_fp16s.h
+++ b/src/layer/arm/neon_mathfun_fp16s.h
@@ -844,7 +844,7 @@ static inline float16x4_t gelu_ps_f16(float16x4_t x)
     float16x4_t _half = vdup_n_f16(0.5f);
     float16x4_t _blob = vmul_f16(vdup_n_f16(-0.70710678f), x);
     _blob = erfc_ps_f16(_blob);
-    return vmul_f16(_half, vmul_f16(_blob, x));
+    return vmul_f16(vmul_f16(_half, _blob), x);
 }
 
 static inline float16x8_t gelu_ps_f16(float16x8_t x)
@@ -852,7 +852,7 @@ static inline float16x8_t gelu_ps_f16(float16x8_t x)
     float16x8_t _half = vdupq_n_f16(0.5f);
     float16x8_t _blob = vmulq_f16(vdupq_n_f16(-0.70710678f), x);
     _blob = erfc_ps_f16(_blob);
-    return vmulq_f16(_half, vmulq_f16(_blob, x));
+    return vmulq_f16(vmulq_f16(_half, _blob), x);
 }
 
 static inline float16x4_t fast_gelu_ps_f16(float16x4_t x)
@@ -865,7 +865,7 @@ static inline float16x4_t fast_gelu_ps_f16(float16x4_t x)
     _blob = vfma_f16(_blob, vdup_n_f16(0.79788452f), x);
     _blob = tanh_ps_f16(_blob);
     _blob = vadd_f16(_one, _blob);
-    return vmul_f16(_half, vmul_f16(_blob, x));
+    return vmul_f16(vmul_f16(_half, _blob), x);
 }
 
 static inline float16x8_t fast_gelu_ps_f16(float16x8_t x)
@@ -878,7 +878,7 @@ static inline float16x8_t fast_gelu_ps_f16(float16x8_t x)
     _blob = vfmaq_f16(_blob, vdupq_n_f16(0.79788452f), x);
     _blob = tanh_ps_f16(_blob);
     _blob = vaddq_f16(_one, _blob);
-    return vmulq_f16(_half, vmulq_f16(_blob, x));
+    return vmulq_f16(vmulq_f16(_half, _blob), x);
 }
 
 static inline float16x4_t selu_ps_f16(float16x4_t x, float16x4_t alphaxlambda, float16x4_t lambda)
diff --git a/src/layer/riscv/gelu_riscv_zfh.cpp b/src/layer/riscv/gelu_riscv_zfh.cpp
index ce819111a..64ba66820 100644
--- a/src/layer/riscv/gelu_riscv_zfh.cpp
+++ b/src/layer/riscv/gelu_riscv_zfh.cpp
@@ -49,8 +49,8 @@ int GELU_riscv::forward_inplace_fp16s(Mat& bottom_top_blob, const Option& opt) c
 
                 vfloat16m4_t _one = __riscv_vfmv_v_f_f16m4((__fp16)1.f, vl);
                 _tanharg = __riscv_vfadd_vv_f16m4(_tanharg, _one, vl);
+                _tanharg = __riscv_vfmul_vf_f16m4(_tanharg, (__fp16)0.5f, vl);
                 _p = __riscv_vfmul_vv_f16m4(_p, _tanharg, vl);
-                _p = __riscv_vfmul_vf_f16m4(_p, (__fp16)0.5f, vl);
 
                 __riscv_vse16_v_f16m4(ptr, _p, vl);
                 n -= vl;
diff --git a/tests/test_gelu.cpp b/tests/test_gelu.cpp
index d330f5e2a..71d2d9bf5 100644
--- a/tests/test_gelu.cpp
+++ b/tests/test_gelu.cpp
@@ -89,6 +89,24 @@ static int test_gelu_4()
            || test_gelu(RandomMat(128, -2000.f, 2000.f), true);
 }
 
+static int test_gelu_5()
+{
+    // finite fp16 inputs must not overflow when multiplying by 1 + tanh(...)
+    const float values[] = {-65504.f, -40000.f, -32768.f, -32752.f, -1.f, -0.f, 0.f, 1.f, 32752.f, 32768.f, 40000.f, 65504.f};
+    ncnn::Mat a(128);
+    ncnn::Mat b(127);
+    for (int i = 0; i < a.w; i++)
+        a[i] = values[i % 12];
+    for (int i = 0; i < b.w; i++)
+        b[i] = values[i % 12];
+
+    return 0
+           || test_gelu(a, false)
+           || test_gelu(a, true)
+           || test_gelu(b, false)
+           || test_gelu(b, true);
+}
+
 int main()
 {
     SRAND(7767517);
@@ -98,5 +116,6 @@ int main()
            || test_gelu_1()
            || test_gelu_2()
            || test_gelu_3()
-           || test_gelu_4();
+           || test_gelu_4()
+           || test_gelu_5();
 }