x86 packed convolution transform kernel avx2/avx512 optimization (#4819)

* fix non-sse non-neon weight pack

Author: nihui

src/layer/arm/convolution1d_packed.h

@@ -60,10 +60,11 @@ static void convolution1d_transform_kernel_packed(const Mat& kernel, Mat& kernel
 #endif // __aarch64__
         if (inh >= 4)
             kernel_tm.create(2 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2);
-        else if (inh >= 2)
-            kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2);
         else
 #endif // __ARM_NEON
+        if (inh >= 2)
+            kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2);
+        else
             kernel_tm.create(2 * kernel_w, inh, outh / 2 + outh % 2);
     }
     else
@@ -76,10 +77,11 @@ static void convolution1d_transform_kernel_packed(const Mat& kernel, Mat& kernel
 #endif // __aarch64__
         if (inh >= 4)
             kernel_tm.create(4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh);
-        else if (inh >= 2)
-            kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh);
         else
 #endif // __ARM_NEON
+        if (inh >= 2)
+            kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh);
+        else
             kernel_tm.create(kernel_w, inh, outh);
     }
     // *INDENT-ON*

src/layer/arm/convolution1d_packed_bf16s.h

@@ -60,10 +60,11 @@ static void convolution1d_transform_kernel_packed_bf16s(const Mat& kernel, Mat&
 #endif // __aarch64__
         if (inh >= 4)
             kernel_tm.create(2 * 4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
-        else if (inh >= 2)
-            kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
         else
 #endif // __ARM_NEON
+        if (inh >= 2)
+            kernel_tm.create(2 * 2 * kernel_w, inh / 2 + inh % 2, outh / 2 + outh % 2, (size_t)2u);
+        else
             kernel_tm.create(2 * kernel_w, inh, outh / 2 + outh % 2, (size_t)2u);
     }
     else
@@ -76,10 +77,11 @@ static void convolution1d_transform_kernel_packed_bf16s(const Mat& kernel, Mat&
 #endif // __aarch64__
         if (inh >= 4)
             kernel_tm.create(4 * kernel_w, inh / 4 + (inh % 4) / 2 + inh % 2, outh, (size_t)2u);
-        else if (inh >= 2)
-            kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh, (size_t)2u);
         else
 #endif // __ARM_NEON
+        if (inh >= 2)
+            kernel_tm.create(2 * kernel_w, inh / 2 + inh % 2, outh, (size_t)2u);
+        else
             kernel_tm.create(kernel_w, inh, outh, (size_t)2u);
     }
     // *INDENT-ON*

src/layer/arm/convolution_packed.h

@@ -62,10 +62,11 @@ static void convolution_transform_kernel_packed(const Mat& kernel, Mat& kernel_t
 #endif // __aarch64__
         if (inch >= 4)
             kernel_tm.create(2 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2);
-        else if (inch >= 2)
-            kernel_tm.create(2 * 2 * maxk, inch / 2 + inch % 2, outch / 2 + outch % 2);
         else
 #endif // __ARM_NEON
+        if (inch >= 2)
+            kernel_tm.create(2 * 2 * maxk, inch / 2 + inch % 2, outch / 2 + outch % 2);
+        else
             kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2);
     }
     else
@@ -78,10 +79,11 @@ static void convolution_transform_kernel_packed(const Mat& kernel, Mat& kernel_t
 #endif // __aarch64__
         if (inch >= 4)
             kernel_tm.create(4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch);
-        else if (inch >= 2)
-            kernel_tm.create(2 * maxk, inch / 2 + inch % 2, outch);
         else
 #endif // __ARM_NEON
+        if (inch >= 2)
+            kernel_tm.create(2 * maxk, inch / 2 + inch % 2, outch);
+        else
             kernel_tm.create(maxk, inch, outch);
     }
     // *INDENT-ON*

src/layer/arm/convolution_packed_bf16s.h

@@ -62,10 +62,11 @@ static void convolution_transform_kernel_packed_bf16s(const Mat& kernel, Mat& ke
 #endif // __aarch64__
         if (inch >= 4)
             kernel_tm.create(2 * 4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
-        else if (inch >= 2)
-            kernel_tm.create(2 * 2 * maxk, inch / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
         else
 #endif // __ARM_NEON
+        if (inch >= 2)
+            kernel_tm.create(2 * 2 * maxk, inch / 2 + inch % 2, outch / 2 + outch % 2, (size_t)2u);
+        else
             kernel_tm.create(2 * maxk, inch, outch / 2 + outch % 2, (size_t)2u);
     }
     else
@@ -78,10 +79,11 @@ static void convolution_transform_kernel_packed_bf16s(const Mat& kernel, Mat& ke
 #endif // __aarch64__
         if (inch >= 4)
             kernel_tm.create(4 * maxk, inch / 4 + (inch % 4) / 2 + inch % 2, outch, (size_t)2u);
-        else if (inch >= 2)
-            kernel_tm.create(2 * maxk, inch / 2 + inch % 2, outch, (size_t)2u);
         else
 #endif // __ARM_NEON
+        if (inch >= 2)
+            kernel_tm.create(2 * maxk, inch / 2 + inch % 2, outch, (size_t)2u);
+        else
             kernel_tm.create(maxk, inch, outch, (size_t)2u);
     }
     // *INDENT-ON*