fix padding arm pack4 per_channel_pad

src/layer/arm/padding_arm.cpp

@@ -30,7 +30,7 @@ Padding_arm::Padding_arm()
 }
 
 #if __ARM_NEON
-static void padding_constant_pack4_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right, float v)
+static void padding_constant_pack4_neon(const Mat& src, Mat& dst, int top, int bottom, int left, int right, float32x4_t v)
 {
     const float* ptr = src;
     float* outptr = dst;
@@ -43,10 +43,10 @@ static void padding_constant_pack4_neon(const Mat& src, Mat& dst, int top, int b
 
 #if __aarch64__
     asm volatile(
-        "dup    v0.4s, %w10             \n"
-        "dup    v1.4s, %w10             \n"
-        "dup    v2.4s, %w10             \n"
-        "dup    v3.4s, %w10             \n"
+        "mov    v0.4s, %10.4s           \n"
+        "mov    v1.4s, %10.4s           \n"
+        "mov    v2.4s, %10.4s           \n"
+        "mov    v3.4s, %10.4s           \n"
 
         // fill top
         "lsr    w4, %w8, #3             \n"// w4 = nn = top_size >> 3
@@ -198,15 +198,15 @@ static void padding_constant_pack4_neon(const Mat& src, Mat& dst, int top, int b
           "r"(right),       // %7
           "r"(top_size),    // %8
           "r"(bottom_size), // %9
-          "r"(v)            // %10
+          "w"(v)            // %10
         : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23"
     );
 #else // __aarch64__
     asm volatile(
-        "vdup.f32   q0, %10             \n"
-        "vdup.f32   q1, %10             \n"
-        "vdup.f32   q2, %10             \n"
-        "vdup.f32   q3, %10             \n"
+        "vmov.f32   q0, %q10            \n"
+        "vmov.f32   q1, %q10            \n"
+        "vmov.f32   q2, %q10            \n"
+        "vmov.f32   q3, %q10            \n"
 
         // fill top
         "lsr        r4, %8, #3          \n"// r4 = nn = top_size >> 3
@@ -358,7 +358,7 @@ static void padding_constant_pack4_neon(const Mat& src, Mat& dst, int top, int b
           "r"(right),       // %7
           "r"(top_size),    // %8
           "r"(bottom_size), // %9
-          "r"(v)            // %10
+          "r"(w)            // %10
         : "cc", "memory", "r4", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15"
     );
 #endif // __aarch64__
@@ -553,7 +553,7 @@ int Padding_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
                 return -100;
 
             if (type == 0)
-                padding_constant_pack4_neon(bottom_blob, top_blob, 0, 0, left, right, value);
+                padding_constant_pack4_neon(bottom_blob, top_blob, 0, 0, left, right, vdupq_n_f32(value));
             else if (type == 1)
                 padding_replicate_pack4_neon(bottom_blob, top_blob, 0, 0, left, right);
             else // if (type == 2)
@@ -571,7 +571,7 @@ int Padding_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
                 return -100;
 
             if (type == 0)
-                padding_constant_pack4_neon(bottom_blob, top_blob, top, bottom, left, right, value);
+                padding_constant_pack4_neon(bottom_blob, top_blob, top, bottom, left, right, vdupq_n_f32(value));
             else if (type == 1)
                 padding_replicate_pack4_neon(bottom_blob, top_blob, top, bottom, left, right);
             else // if (type == 2)
@@ -592,7 +592,7 @@ int Padding_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& op
                 const Mat m = bottom_blob.channel(q);
                 Mat borderm = top_blob.channel(q);
 
-                float pad_value = per_channel_pad_data_size ? per_channel_pad_data[q] : value;
+                float32x4_t pad_value = per_channel_pad_data_size ? vld1q_f32((const float*)per_channel_pad_data + q * 4) : vdupq_n_f32(value);
 
                 if (type == 0)
                     padding_constant_pack4_neon(m, borderm, top, bottom, left, right, pad_value);