riscv: add int8 convolution winograd (#7093)

Author: nihui

src/layer/arm/convolution_3x3_winograd_int8.h

@@ -3432,7 +3432,7 @@ static void get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TIL
 #if __aarch64__
         TILE_K = std::max(8, tile_size / 8 * 8);
 #elif __ARM_NEON
-        TILE_K = std::max(4, tile_size / 4 * 4);
+        TILE_K = std::max(8, tile_size / 8 * 8);
 #else
         TILE_K = std::max(2, tile_size / 2 * 2);
 #endif
@@ -3441,7 +3441,7 @@ static void get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TIL
 #if __aarch64__
         TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
 #elif __ARM_NEON
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
+        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
 #else
         TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
 #endif

src/layer/loongarch/convolution_3x3_winograd_int8.h

@@ -2551,7 +2551,7 @@ static void get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TIL
 #if __loongarch_asx
         TILE_K = std::max(8, tile_size / 8 * 8);
 #elif __loongarch_sx
-        TILE_K = std::max(4, tile_size / 4 * 4);
+        TILE_K = std::max(8, tile_size / 8 * 8);
 #else
         TILE_K = std::max(2, tile_size / 2 * 2);
 #endif
@@ -2560,7 +2560,7 @@ static void get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TIL
 #if __loongarch_asx
         TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
 #elif __loongarch_sx
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
+        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
 #else
         TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
 #endif

src/layer/mips/convolution_3x3_winograd_int8.h

@@ -1901,14 +1901,14 @@ static void get_optimal_tile_mnk_int8(int M, int N, int K, int& TILE_M, int& TIL
         int tile_size = (int)(sqrt((float)l2_cache_size_int8) - TILE_M);
 
 #if __mips_msa
-        TILE_K = std::max(4, tile_size / 4 * 4);
+        TILE_K = std::max(8, tile_size / 8 * 8);
 #else
         TILE_K = std::max(2, tile_size / 2 * 2);
 #endif
 
         int nn_K = (K + TILE_K - 1) / TILE_K;
 #if __mips_msa
-        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 3) / 4 * 4);
+        TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 7) / 8 * 8);
 #else
         TILE_K = std::min(TILE_K, ((K + nn_K - 1) / nn_K + 1) / 2 * 2);
 #endif