fix memory reuse

mace/core/BUILD

@@ -15,6 +15,7 @@ load(
     "if_production_mode",
     "if_not_production_mode",
     "if_openmp_enabled",
+    "if_neon_enabled",
 )
 
 cc_library(
@@ -47,7 +48,11 @@ cc_library(
         "-DMACE_ENABLE_OPENMP",
     ]) + if_android([
         "-DMACE_ENABLE_OPENCL",
-    ]) + if_hexagon_enabled(["-DMACE_ENABLE_HEXAGON"]),
+    ]) + if_hexagon_enabled([
+        "-DMACE_ENABLE_HEXAGON",
+    ]) + if_neon_enabled([
+        "-DMACE_ENABLE_NEON",
+    ]),
     linkopts = ["-ldl"] + if_android([
         "-pie",
         "-lm",

mace/core/tensor.h

@@ -222,6 +222,8 @@ class Tensor {
     if (buffer_ != nullptr) {
       MACE_CHECK(!has_opencl_image(), "Cannot resize image, use ResizeImage.");
       if (raw_size() + EXTRA_BUFFER_PAD_SIZE > buffer_->size()) {
+        LOG(WARNING) << "Resize buffer from size " << buffer_->size() << " to "
+                     << raw_size() + EXTRA_BUFFER_PAD_SIZE;
         return buffer_->Resize(raw_size() + EXTRA_BUFFER_PAD_SIZE);
       }
       return MaceStatus::MACE_SUCCESS;

mace/core/workspace.cc

@@ -177,7 +177,7 @@ MaceStatus Workspace::CreateOutputTensorBuffer(const NetDef &net_def,
         std::unique_ptr<BufferBase> tensor_buf(
             new Buffer(GetDeviceAllocator(device_type)));
         MaceStatus status = tensor_buf->Allocate(
-            mem_block.x() * GetEnumTypeSize(dtype));
+            mem_block.x() * GetEnumTypeSize(dtype) + EXTRA_BUFFER_PAD_SIZE);
         if (status != MaceStatus::MACE_SUCCESS) {
           return status;
         }