fix gemm transpose B wrong result when tile N is not a multiple of 4, optimize load C (#4430)

src/layer/arm/gemm_arm.cpp

@@ -707,7 +707,7 @@ static void transpose_pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int
     {
         if (elempack == 4)
         {
-            const float* p0 = (const float*)B + k / 4 * 4 * B_hstep + (j + jj) * 4;
+            const float* p0 = (const float*)B + k * B_hstep + (j + jj) * 4;
 
             int kk = 0;
             for (; kk + 3 < max_kk; kk += 4)
@@ -750,7 +750,7 @@ static void transpose_pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int
     {
         if (elempack == 4)
         {
-            const float* p0 = (const float*)B + k / 4 * 4 * B_hstep + (j + jj) * 4;
+            const float* p0 = (const float*)B + k * B_hstep + (j + jj) * 4;
 
             int kk = 0;
             for (; kk + 3 < max_kk; kk += 4)
@@ -787,7 +787,7 @@ static void transpose_pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int
     {
         if (elempack == 4)
         {
-            const float* p0 = (const float*)B + k / 4 * 4 * B_hstep + (j + jj) * 4;
+            const float* p0 = (const float*)B + k * B_hstep + (j + jj) * 4;
 
             int kk = 0;
             for (; kk + 3 < max_kk; kk += 4)
@@ -820,7 +820,7 @@ static void transpose_pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int
 #if __ARM_NEON
         if (elempack == 4)
         {
-            const float* p0 = (const float*)B + k / 4 * 4 * B_hstep + (j + jj) * 4;
+            const float* p0 = (const float*)B + k * B_hstep + (j + jj) * 4;
 
             int kk = 0;
             for (; kk + 3 < max_kk; kk += 4)
@@ -853,7 +853,7 @@ static void transpose_pack_B_tile(const Mat& B, Mat& BT, int j, int max_jj, int
 #if __ARM_NEON
         if (elempack == 4)
         {
-            const float* p0 = (const float*)B + k / 4 * 4 * B_hstep + (j + jj) * 4;
+            const float* p0 = (const float*)B + k * B_hstep + (j + jj) * 4;
 
             int kk = 0;
             for (; kk + 3 < max_kk; kk += 4)
@@ -3789,11 +3789,11 @@ static void get_optimal_tile_mnk(int M, int N, int K, int& TILE_M, int& TILE_N,
 
 #if __aarch64__
     TILE_M = tile_size / 8 * 8;
-    TILE_N = tile_size;
+    TILE_N = tile_size / 4 * 4;
     TILE_K = tile_size / 8 * 8;
 #elif __ARM_NEON
     TILE_M = tile_size / 4 * 4;
-    TILE_N = tile_size;
+    TILE_N = tile_size / 4 * 4;
     TILE_K = tile_size / 4 * 4;
 #else
     TILE_M = tile_size / 2 * 2;
@@ -3818,10 +3818,10 @@ static void get_optimal_tile_mnk(int M, int N, int K, int& TILE_M, int& TILE_N,
 
 #if __aarch64__
             TILE_M = tile_size / 8 * 8;
-            TILE_N = tile_size;
+            TILE_N = tile_size / 4 * 4;
 #elif __ARM_NEON
             TILE_M = tile_size / 4 * 4;
-            TILE_N = tile_size;
+            TILE_N = tile_size / 4 * 4;
 #else
             TILE_M = tile_size / 2 * 2;
             TILE_N = tile_size;
@@ -3846,7 +3846,13 @@ static void get_optimal_tile_mnk(int M, int N, int K, int& TILE_M, int& TILE_N,
     if (N > 0)
     {
         int nn_N = (N + TILE_N - 1) / TILE_N;
+#if __aarch64__
+        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
+#elif __ARM_NEON
+        TILE_N = std::min(TILE_N, ((N + nn_N - 1) / nn_N + 3) / 4 * 4);
+#else
         TILE_N = std::min(TILE_N, (N + nn_N - 1) / nn_N);
+#endif
     }
 
     if (nT > 1)