[ARM] fix 5x5s2_int8 compute bug, test=develop (#3079)

lite/backends/arm/math/conv5x5s2_depthwise_int8.cc

@@ -102,7 +102,7 @@ void conv_depthwise_5x5s2_int8(Dtype* dout,
       if (h + hout_r_block > hout) {
         h_kernel = hout - h;
       }
-      int hs = h - padh;
+      int hs = h * 2 - padh;
       int he = hs + h_kernel * 2 + 3;
 
 #pragma omp parallel for num_threads(threads)

lite/tests/math/conv_int8_compute_test.cc

@@ -502,7 +502,7 @@ TEST(TestConv5x5DWInt8, test_conv5x5_depthwise) {
               std::vector<DDim> dims;
               DDim weights_dim({c, 1, 5, 5});
               for (auto& batch : {1, 2}) {
-                for (auto& h : {1, 3, 15, 33}) {
+                for (auto& h : {1, 3, 15, 33, 112, 224}) {
                   dims.push_back(DDim({batch, c, h, h}));
                 }
               }
@@ -514,7 +514,7 @@ TEST(TestConv5x5DWInt8, test_conv5x5_depthwise) {
                              {1, 1},
                              flag_bias,
                              flag_relu,
-                             {4},
+                             {1, 4},
                              {FLAGS_power_mode});
             }
           }