wip

Tencent · Nov 17, 2023 · 23ecdc3 · 23ecdc3
1 parent 98ccc3e
commit 23ecdc3
Show file tree

Hide file tree

Showing 4 changed files with 18 additions and 18 deletions.
diff --git a/src/layer/arm/padding_pack4.h b/src/layer/arm/padding_pack4.h
@@ -372,7 +372,7 @@ static void padding_constant_pack4_neon(const Mat& src, Mat& dst, int top, int b
         }
         for (int x = 0; x < src.w; x++)
         {
-            _p = vld1q_f32(ptr);
+            float32x4_t _p = vld1q_f32(ptr);
             vst1q_f32(outptr, _p);
             ptr += 4;
             outptr += 4;

diff --git a/src/layer/arm/padding_pack4_bf16s_fp16s.h b/src/layer/arm/padding_pack4_bf16s_fp16s.h
@@ -360,7 +360,7 @@ static void padding_constant_pack4_bf16_fp16s_neon(const Mat& src, Mat& dst, int
         }
         for (int x = 0; x < src.w; x++)
         {
-            _p = vld1_u16(ptr);
+            uint16x4_t _p = vld1_u16(ptr);
             vst1_u16(outptr, _p);
             ptr += 4;
             outptr += 4;

diff --git a/src/layer/arm/padding_pack8_fp16s.h b/src/layer/arm/padding_pack8_fp16s.h
@@ -155,7 +155,7 @@ static void padding_constant_pack8_fp16s_neon(const Mat& src, Mat& dst, int top,
         "r"(right),       // %7
         "r"(top_size),    // %8
         "r"(bottom_size), // %9
-        "w"(v)           // %10
+        "w"(v)            // %10
         : "cc", "memory", "x4", "v0", "v1", "v2", "v3", "v16", "v17", "v18", "v19");
 #else  // NCNN_GNU_INLINE_ASM
 
@@ -186,7 +186,7 @@ static void padding_constant_pack8_fp16s_neon(const Mat& src, Mat& dst, int top,
         }
         for (int x = 0; x < src.w; x++)
         {
-            _p = vld1q_u16(ptr);
+            uint16x8_t _p = vld1q_u16(ptr);
             vst1q_u16(outptr, _p);
             ptr += 8;
             outptr += 8;

diff --git a/src/layer/arm/padding_pack8_int8.h b/src/layer/arm/padding_pack8_int8.h
@@ -342,15 +342,15 @@ static void padding_constant_pack8_int8_neon(const Mat& src, Mat& dst, int top,
         int x = 0;
         for (; x + 3 < top_size; x += 4)
         {
-            vst1q_s8(outptr, v);
-            vst1q_s8(outptr + 8, v);
-            vst1q_s8(outptr + 16, v);
-            vst1q_s8(outptr + 24, v);
+            vst1_s8(outptr, v);
+            vst1_s8(outptr + 8, v);
+            vst1_s8(outptr + 16, v);
+            vst1_s8(outptr + 24, v);
             outptr += 32;
         }
         for (; x < top_size; x++)
         {
-            vst1q_s8(outptr, v);
+            vst1_s8(outptr, v);
             outptr += 8;
         }
     }
@@ -359,19 +359,19 @@ static void padding_constant_pack8_int8_neon(const Mat& src, Mat& dst, int top,
     {
         for (int x = 0; x < left; x++)
         {
-            vst1q_s8(outptr, v);
+            vst1_s8(outptr, v);
             outptr += 8;
         }
         for (int x = 0; x < src.w; x++)
         {
-            _p = vld1q_s8(ptr);
-            vst1q_s8(outptr, _p);
+            int8x8_t _p = vld1_s8(ptr);
+            vst1_s8(outptr, _p);
             ptr += 8;
             outptr += 8;
         }
         for (int x = 0; x < right; x++)
         {
-            vst1q_s8(outptr, v);
+            vst1_s8(outptr, v);
             outptr += 8;
         }
     }
@@ -380,15 +380,15 @@ static void padding_constant_pack8_int8_neon(const Mat& src, Mat& dst, int top,
         int x = 0;
         for (; x + 3 < bottom_size; x += 4)
         {
-            vst1q_s8(outptr, v);
-            vst1q_s8(outptr + 8, v);
-            vst1q_s8(outptr + 16, v);
-            vst1q_s8(outptr + 24, v);
+            vst1_s8(outptr, v);
+            vst1_s8(outptr + 8, v);
+            vst1_s8(outptr + 16, v);
+            vst1_s8(outptr + 24, v);
             outptr += 32;
         }
         for (; x < bottom_size; x++)
         {
-            vst1q_s8(outptr, v);
+            vst1_s8(outptr, v);
             outptr += 8;
         }
     }