diff --git a/WORKSPACE b/WORKSPACE
index daa855de4784d5b968dbc877bc3cc0031f3e455e..d7472c7b270b1d22e9db58718ea7a8434ce10a03 100644
--- a/WORKSPACE
+++ b/WORKSPACE
@@ -79,19 +79,19 @@ new_http_archive(
 
 http_archive(
     name = "gemmlowp",
-    sha256 = "afbea037aee2d21b625985238486b4219396f9c2550b0fde3157fab4d2580205",
-    strip_prefix = "gemmlowp-master-1f6d8d442805a400c74e63a4a017390733df2e28",
+    sha256 = "f340384e7728cea605e83597593699dfe8d13ff333b834d24c256935e3dc1758",
+    strip_prefix = "gemmlowp-master-48c0547a046d49b466aa01e3a82a18028f288924",
     urls = [
-        "http://cnbj1.fds.api.xiaomi.com/mace/third-party/gemmlowp/gemmlowp-master-1f6d8d442805a400c74e63a4a017390733df2e28.zip",
+        "http://cnbj1.fds.api.xiaomi.com/mace/third-party/gemmlowp/gemmlowp-master-48c0547a046d49b466aa01e3a82a18028f288924.zip",
     ],
 )
 
 http_archive(
     name = "tflite",
-    sha256 = "8b4c1b2ad2d31da9859e17b0ad551b12e1db7ff2faf7e83218901ab48d9fa91a",
-    strip_prefix = "tensorflow-mace-dfabaf85145e4d5ad39f34a0cea57b44c32dbe43",
+    sha256 = "6f2671a02fe635a82c289c8c40a6e5bc24670ff1d4c3c2ab4a7aa9b825256a18",
+    strip_prefix = "tensorflow-mace-d73e88fc830320d3818ac24e57cd441820a85cc9",
     urls = [
-        "http://cnbj1.fds.api.xiaomi.com/mace/third-party/tflite/tensorflow-mace-dfabaf85145e4d5ad39f34a0cea57b44c32dbe43.zip",
+        "http://cnbj1.fds.api.xiaomi.com/mace/third-party/tflite/tensorflow-mace-d73e88fc830320d3818ac24e57cd441820a85cc9.zip",
     ],
 )
 
diff --git a/mace/core/runtime/cpu/cpu_runtime.h b/mace/core/runtime/cpu/cpu_runtime.h
index f8cb2111cedd5fa125a6dae264c2210359900c8d..d97c90ce7cb1e06390582f3dc25e3bc94e2e0680 100644
--- a/mace/core/runtime/cpu/cpu_runtime.h
+++ b/mace/core/runtime/cpu/cpu_runtime.h
@@ -57,7 +57,7 @@ class CPURuntime {
 
   gemmlowp::GemmContext *GetGemmlowpContext() {
     if (gemm_context_ == nullptr) {
-      gemm_context_ = new gemmlowp::GemmContext();
+      gemm_context_ = new gemmlowp::GemmContext(thread_pool_);
     }
     return static_cast<gemmlowp::GemmContext*>(gemm_context_);
   }
diff --git a/mace/ops/depthwise_conv2d.cc b/mace/ops/depthwise_conv2d.cc
index 522a3b357ed24f5804a9bb2d4af41f8605e644a2..67339ef9c5a78ef37369e4b6c197781dea5690db 100644
--- a/mace/ops/depthwise_conv2d.cc
+++ b/mace/ops/depthwise_conv2d.cc
@@ -250,6 +250,7 @@ class DepthwiseConv2dOp<DeviceType::CPU, uint8_t>
       std::vector<index_t> bias_shape{out_channels};
 
       tflite::optimized_ops::DepthwiseConv(
+          &context->device()->cpu_runtime()->thread_pool(),
           input_data, ShapeToTfliteDims(input->shape()), -input->zero_point(),
           filter_data, ShapeToTfliteDims(filter_shape), -filter->zero_point(),
           bias_data, ShapeToTfliteDims(bias_shape), stride_w, stride_h,
diff --git a/tools/bazel.rc b/tools/bazel.rc
index 0a49d2cebd18add2bab3b7eae341b730acf771dd..629129cf4cc38015ac49dee5b9f3a97ed77b09da 100644
--- a/tools/bazel.rc
+++ b/tools/bazel.rc
@@ -8,6 +8,8 @@ build --copt=-std=c++11
 build --copt=-fPIC
 build --copt=-D_GLIBCXX_USE_C99_MATH_TR1
 build --copt=-DMACE_OBFUSCATE_LITERALS
+build --copt=-DGEMMLOWP_USE_MACE_THREAD_POOL
+build --copt=-DMACE_DEPTHWISE_U8_USE_MULTI_THREAD
 
 # Usage example: bazel build --config android
 build:android --linkopt=-pie