fix winograd reinitwhenneed (#2511)

* add winograd c4 implement (#2494)
test=develop
fix: fix conv_block prepack_input_nxwc4 bug
* fix: optimize sgemm_c4 in armv7
     change condition of choose winograd kernel
* fix: change conv choose kernel condition
test=develop

lite/kernels/arm/conv_compute.cc

@@ -68,10 +68,6 @@ void ConvCompute<PRECISION(kFloat), PRECISION(kFloat)>::PrepareForRun() {
     VLOG(3) << "invoking dw conv";
   } else if (param.groups == 1 && kw == 3 && stride == 1 && kps_equal &&
              no_dilation) {
-    int tile_block = 8;
-#ifdef __aarch64__
-    tile_block = 16;
-#endif
     bool use_winograd =
         (threads == 1 && oc >= 4 && ic >= 4 && hout >= 6 && wout >= 6 &&
          pads_equal) ||

lite/kernels/arm/conv_winograd.cc

@@ -49,6 +49,10 @@ void WinogradConv<PRECISION(kFloat), PRECISION(kFloat)>::ReInitWhenNeeded() {
   int parallel_threads =
       (((ow + 5) / 6) * ((oh + 5) / 6) + tile_block - 1) / tile_block;
   if (threads <= 2 && parallel_threads >= threads) {
+    if (last_kernel_is_c4_ == 1) {
+      return;
+    }
+    last_kernel_is_c4_ = 1;
     auto pad = *(param.paddings);
     int pad_h = pad[0];
     int pad_w = pad[2];
@@ -68,6 +72,10 @@ void WinogradConv<PRECISION(kFloat), PRECISION(kFloat)>::ReInitWhenNeeded() {
         weights_data_, param.filter->data<float>(), ic, oc, trans_tmp_ptr);
     free(trans_tmp_ptr);
   } else {
+    if (last_kernel_is_c4_ == 0) {
+      return;
+    }
+    last_kernel_is_c4_ = 0;
     int tile_w = (ow + 5) / 6;
     int tile_h = (oh + 5) / 6;
 

lite/kernels/arm/conv_winograd.h

@@ -40,6 +40,7 @@ class WinogradConv : public KernelLite<TARGET(kARM), Ptype> {
   Tensor weights_;
   DDim last_shape_;
   int workspace_size_{0};
+  int last_kernel_is_c4_{-1};
 };
 
 }  // namespace arm