From 4dd645b8f0e38450d97cbf43bc4a61dc8ba57df8 Mon Sep 17 00:00:00 2001
From: liuqi <liuqi10@xiaomi.com>
Date: Tue, 31 Oct 2017 15:25:04 +0800
Subject: [PATCH] Vectorization batch norm opencl kernel.

---
 mace/kernels/opencl/cl/batch_norm.cl | 10 +++++-----
 mace/utils/BUILD                     |  1 +
 mace/utils/tuner.h                   | 10 +++++-----
 mace/utils/tuner_test.cc             |  3 ++-
 4 files changed, 13 insertions(+), 11 deletions(-)

diff --git a/mace/kernels/opencl/cl/batch_norm.cl b/mace/kernels/opencl/cl/batch_norm.cl
index 3b7dcf08..3fc449ce 100644
--- a/mace/kernels/opencl/cl/batch_norm.cl
+++ b/mace/kernels/opencl/cl/batch_norm.cl
@@ -22,12 +22,12 @@ void kernel batch_norm(global const float *input,
 
   barrier(CLK_LOCAL_MEM_FENCE);
 
-  const int sample_offset = (batch * channels + channel) * pixels + pixel_offset*4;
-  const float *input_ptr = input + sample_offset;
-  float *output_ptr = output + sample_offset;
+  const int image_offset = (batch * channels + channel) * pixels + pixel_offset*4;
+  const float *input_ptr = input + image_offset;
+  float *output_ptr = output + image_offset;
   const int end = (batch * channels + channel + 1) * pixels;
-  if ((sample_offset+4) > end) {
-    for (int i = sample_offset; i < end; ++i) {
+  if ((image_offset+4) > end) {
+    for (int i = image_offset; i < end; ++i) {
       *output_ptr = new_scale[local_channel].x * *input_ptr + new_offset[local_channel].x;
       ++input_ptr;
       ++output_ptr;
diff --git a/mace/utils/BUILD b/mace/utils/BUILD
index 50f65f4e..cd5583df 100644
--- a/mace/utils/BUILD
+++ b/mace/utils/BUILD
@@ -39,6 +39,7 @@ cc_library(
     copts = ["-std=c++11"],
     deps = [
         "//mace/core",
+        "//mace/core:opencl_runtime",
     ],
 )
 
diff --git a/mace/utils/tuner.h b/mace/utils/tuner.h
index de96f87e..1d36f7f5 100644
--- a/mace/utils/tuner.h
+++ b/mace/utils/tuner.h
@@ -33,7 +33,7 @@ class Tuner {
               const std::function<std::vector<std::vector<param_type>>()> &param_generator,
               const std::function<RetType(const std::vector<param_type> &)> &func) {
 
-    if (IsTuning()) {
+    if (IsTuning() && param_generator != nullptr) {
       // tune
       std::vector<param_type> opt_param = default_param;
       RetType res = Tune<RetType>(param_generator, func, opt_param);
@@ -68,7 +68,7 @@ class Tuner {
   }
 
   inline void WriteRunParameters() {
-    VLOG(0) << path_;
+    VLOG(1) << path_;
     if (path_ != nullptr) {
       std::ofstream ofs(path_, std::ios::binary | std::ios::out);
       if (ofs.is_open()) {
@@ -78,14 +78,14 @@ class Tuner {
           int32_t key_size = kp.first.size();
           ofs.write(reinterpret_cast<char *>(&key_size), sizeof(key_size));
           ofs.write(kp.first.c_str(), key_size);
-          VLOG(0) << kp.first.c_str();
+          VLOG(1) << kp.first.c_str();
 
           auto &params = kp.second;
           int32_t params_size = params.size() * sizeof(param_type);
           ofs.write(reinterpret_cast<char*>(&params_size), sizeof(params_size));
           for (auto &param : params) {
             ofs.write(reinterpret_cast<char *>(&param), sizeof(params_size));
-            VLOG(0) << param;
+            VLOG(1) << param;
           }
         }
         ofs.close();
@@ -144,7 +144,7 @@ class Tuner {
   }
 
   template <typename RetType>
-  inline RetType Tune(std::function<std::vector<std::vector<param_type>>()> param_generator,
+  inline RetType Tune(const std::function<std::vector<std::vector<param_type>>()> &param_generator,
                    const std::function<RetType(const std::vector<param_type> &)> &func,
                    std::vector<param_type> &opt_params) {
     RetType res;
diff --git a/mace/utils/tuner_test.cc b/mace/utils/tuner_test.cc
index bcb5c620..ea80dd4d 100644
--- a/mace/utils/tuner_test.cc
+++ b/mace/utils/tuner_test.cc
@@ -13,7 +13,8 @@ class TunerTest: public ::testing::Test {
  protected:
   virtual void SetUp() {
     remove( "/data/local/tmp/mace.config" );
-    setenv("MACE_RUN_PARAMTER_PATH", "/data/local/tmp/mace.config", 1);
+    setenv("MACE_RUN_PARAMETER_PATH", "/data/local/tmp/mace.config", 1);
+    setenv("MACE_TUNING", "1", 1);
   }
 };
 
-- 
GitLab