Add enable_neon rule to leave neon code alone when compiling opencl.

mace/BUILD

@@ -32,6 +32,14 @@ config_setting(
     visibility = ["//visibility:public"],
 )
 
+config_setting(
+    name = "enable_neon",
+    define_values = {
+        "neon": "true",
+    },
+    visibility = ["//visibility:public"],
+)
+
 config_setting(
     name = "embed_binary_program",
     define_values = {

mace/examples/BUILD

 # Examples
-load("//mace:mace.bzl", "if_android", "if_embed_binary_program")
+load("//mace:mace.bzl", "if_android", "if_enable_neon", "if_embed_binary_program")
 
 cc_binary(
     name = "helloworld",
@@ -7,7 +7,7 @@ cc_binary(
         "helloworld.cc",
     ],
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
-    linkopts = ["-fopenmp"],
+    linkopts = if_enable_neon(["-fopenmp"]),
     deps = [
         "//mace/core",
         "//mace/ops",
@@ -20,7 +20,7 @@ cc_test(
     testonly = 1,
     srcs = ["benchmark_example.cc"],
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
-    linkopts = ["-fopenmp"],
+    linkopts = if_enable_neon(["-fopenmp"]),
     linkstatic = 1,
     deps = [
         "//mace/core",
@@ -32,7 +32,7 @@ cc_binary(
     name = "mace_run",
     srcs = ["mace_run.cc"],
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
-    linkopts = ["-fopenmp"],
+    linkopts = if_enable_neon(["-fopenmp"]),
     linkstatic = 1,
     deps = [
         "//mace/codegen:generated_models_lib",

mace/kernels/BUILD

@@ -7,22 +7,22 @@ package(
 
 licenses(["notice"])  # Apache 2.0
 
-load("//mace:mace.bzl", "if_android", "if_android_arm64", "if_embed_binary_program")
+load("//mace:mace.bzl", "if_android", "if_enable_neon", "if_embed_binary_program")
 
 cc_library(
     name = "kernels",
-    srcs = glob(["*.cc", "opencl/*.cc"]) + if_android_arm64(glob([
+    srcs = glob(["*.cc", "opencl/*.cc"]) + if_enable_neon(glob([
         "neon/*.cc",
     ])),
-    hdrs = glob(["*.h", "opencl/*.h"]) + if_android_arm64(glob([
+    hdrs = glob(["*.h", "opencl/*.h"]) + if_enable_neon(glob([
         "neon/*.h",
     ])),
     copts = [
         "-std=c++11",
-        "-fopenmp",
         "-D_GLIBCXX_USE_C99_MATH_TR1",
     ] +
-    if_embed_binary_program(["-DMACE_EMBED_BINARY_PROGRAM"]),
+    if_embed_binary_program(["-DMACE_EMBED_BINARY_PROGRAM"]) +
+    if_enable_neon(["-fopenmp"]),
     linkopts = if_android(["-lm"]),
     deps = [
         "//mace/core",

mace/mace.bzl

@@ -35,3 +35,9 @@ def if_embed_binary_program(a):
       "//mace:embed_binary_program": a,
       "//conditions:default": [],
   })
+
+def if_enable_neon(a):
+  return select({
+      "//mace:enable_neon": a,
+      "//conditions:default": [],
+  })

mace/ops/BUILD

@@ -7,7 +7,7 @@ package(
 
 licenses(["notice"])  # Apache 2.0
 
-load("//mace:mace.bzl", "if_android")
+load("//mace:mace.bzl", "if_android", "if_enable_neon")
 
 cc_library(
     name = "test",
@@ -34,7 +34,8 @@ cc_library(
         ["*.h"],
         exclude = ["ops_test_util.h"],
     ),
-    copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
+    copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"] +
+            if_enable_neon(["-DMACE_ENABLE_NEON"]),
     deps = [
         "//mace/kernels",
     ],

mace/ops/addn.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("AddN")
                              .Build(),
                       AddNOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("AddN")
                              .TypeConstraint<float>("T")
                              .Build(),
                        AddNOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("AddN")
                              .TypeConstraint<float>("T")

mace/ops/batch_norm.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("BatchNorm")
                              .Build(),
                       BatchNormOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("BatchNorm")
                              .TypeConstraint<float>("T")
                              .Build(),
                        BatchNormOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("BatchNorm")
                              .TypeConstraint<float>("T")

mace/ops/conv_2d.cc

@@ -16,12 +16,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("Conv2D")
                           .Build(),
                       Conv2dOp<DeviceType::CPU, half>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("Conv2D")
                              .TypeConstraint<float>("T")
                              .Build(),
                        Conv2dOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Conv2D")
                              .TypeConstraint<float>("T")

mace/ops/depthwise_conv2d.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
                              .Build(),
                       DepthwiseConv2dOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
                              .TypeConstraint<float>("T")
                              .Build(),
                        DepthwiseConv2dOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("DepthwiseConv2d")
                              .TypeConstraint<float>("T")

mace/ops/global_avg_pooling.cc

@@ -11,11 +11,11 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("GlobalAvgPooling")
                              .Build(),
                       GlobalAvgPoolingOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("GlobalAvgPooling")
                              .TypeConstraint<float>("T")
                              .Build(),
                        GlobalAvgPoolingOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 }  //  namespace mace

mace/ops/global_avg_pooling.h

-//
+//DMACE_ENABLE_NEON
 // Copyright (c) 2017 XiaoMi All rights reserved.
 //
 

mace/ops/ops_test_util.h

@@ -11,6 +11,7 @@
 #include "mace/core/common.h"
 #include "mace/core/net.h"
 #include "mace/core/tensor.h"
+#include "mace/core/workspace.h"
 #include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/kernels/opencl/helper.h"
 #include "mace/utils/utils.h"

mace/ops/pooling.cc

@@ -15,12 +15,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("Pooling")
                           .Build(),
                       PoolingOp<DeviceType::CPU, half>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("Pooling")
                              .TypeConstraint<float>("T")
                              .Build(),
                        PoolingOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Pooling")
                              .TypeConstraint<float>("T")

mace/ops/relu.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("Relu")
                           .Build(),
                       ReluOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("Relu")
                              .TypeConstraint<float>("T")
                              .Build(),
                        ReluOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("Relu")
                              .TypeConstraint<float>("T")

mace/ops/resize_bilinear.cc

@@ -11,12 +11,12 @@ REGISTER_CPU_OPERATOR(OpKeyBuilder("ResizeBilinear")
                           .Build(),
                       ResizeBilinearOp<DeviceType::CPU, float>);
 
-#if __ARM_NEON
+#if MACE_ENABLE_NEON
 REGISTER_NEON_OPERATOR(OpKeyBuilder("ResizeBilinear")
                              .TypeConstraint<float>("T")
                              .Build(),
                        ResizeBilinearOp<DeviceType::NEON, float>);
-#endif  // __ARM_NEON
+#endif  // MACE_ENABLE_NEON
 
 REGISTER_OPENCL_OPERATOR(OpKeyBuilder("ResizeBilinear")
                              .TypeConstraint<float>("T")

tools/bazel-adb-run.sh

@@ -27,7 +27,11 @@ PROFILINE="--define profiling=true"
 BRANCH=$(git symbolic-ref --short HEAD)
 COMMIT_ID=$(git rev-parse --short HEAD)
 
-bazel build -c opt $STRIP --verbose_failures $BAZEL_TARGET --crosstool_top=//external:android/crosstool --host_crosstool_top=@bazel_tools//tools/cpp:toolchain --cpu=$ANDROID_ABI
+bazel build -c opt $STRIP --verbose_failures $BAZEL_TARGET \
+   --crosstool_top=//external:android/crosstool \
+   --host_crosstool_top=@bazel_tools//tools/cpp:toolchain  \
+   --cpu=$ANDROID_ABI \
+   --define neon=true
 
 if [ $? -ne 0 ]; then
   exit 1

tools/validate_gcn.sh

@@ -51,7 +51,6 @@ build_and_run()
   adb push ${MODEL_DIR}/${INPUT_FILE_NAME} ${PHONE_DATA_DIR}
   adb push bazel-bin/mace/examples/mace_run ${PHONE_DATA_DIR}
 
-  num_threads=${1:-4}
   if [[ "${TUNING_OR_NOT}" != "0" && "$EMBED_OPENCL_BINARY" != true ]];then
     tuning_flag=1
     round=0 # only warm up
@@ -64,7 +63,6 @@ build_and_run()
     MACE_CPP_MIN_VLOG_LEVEL=0 \
     MACE_RUN_PARAMETER_PATH=${PHONE_DATA_DIR}/mace_run.config \
     MACE_KERNEL_PATH=$KERNEL_DIR \
-    OMP_NUM_THREADS=$num_threads \
     ${PHONE_DATA_DIR}/mace_run \
     --model=${PHONE_DATA_DIR}/${MACE_MODEL_NAME} \
     --input=mace_input_node \