Merge branch 'tuning-binary' into 'master'

Finish convert tuning params file to source file.

See merge request !170

mace/codegen/BUILD

@@ -15,7 +15,10 @@ cc_library(
     deps = [
         "//mace/core",
         "//mace/ops",
-    ] + if_embed_binary_program(['//mace/codegen:generated_opencl_lib']),
+    ] + if_embed_binary_program([
+        '//mace/codegen:generated_opencl_lib',
+        '//mace/codegen:generated_tuning_lib',
+    ]),
 )
 
 cc_library(
@@ -24,3 +27,10 @@ cc_library(
     copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
     linkstatic = 1,
 )
+
+cc_library(
+    name = "generated_tuning_lib",
+    srcs = glob(["tuning/*.cc"]),
+    copts = ["-std=c++11", "-D_GLIBCXX_USE_C99_MATH_TR1"],
+    linkstatic = 1,
+)

mace/kernels/BUILD

@@ -7,7 +7,7 @@ package(
 
 licenses(["notice"])  # Apache 2.0
 
-load("//mace:mace.bzl", "if_android", "if_android_arm64")
+load("//mace:mace.bzl", "if_android", "if_android_arm64", "if_embed_binary_program")
 
 cc_library(
     name = "kernels",
@@ -21,7 +21,8 @@ cc_library(
         "-std=c++11",
         "-fopenmp",
         "-D_GLIBCXX_USE_C99_MATH_TR1",
-    ],
+    ] +
+    if_embed_binary_program(["-DMACE_EMBED_BINARY_PROGRAM"]),
     linkopts = if_android(["-lm"]),
     deps = [
         "//mace/core",

mace/python/tools/opencl_compiled_program.cc.tmpl → mace/python/tools/binary.cc.tmpl

@@ -10,21 +10,17 @@
 
 namespace mace {
 
-{% for map in binary_maps %}
-//  {{map.name}}
-{% endfor %}
-
-extern const std::map<std::string, std::vector<unsigned char>> kCompiledProgramMap =
+extern const std::map<std::string, std::vector<{{data_type}}>> {{variable_name}}=
 {
-  {% for map in binary_maps %}
+  {% for key, value in maps.iteritems() %}
   {
-    "{{map.name}}",
+    "{{key}}",
     {
-      {%- for ele in map.content -%}
+      {%- for ele in value -%}
       {{ele}},
       {%- endfor -%}
     }
-  },  // {{map.name}}
+  },  // {{key}}
 {% endfor %}
 };
 

mace/python/tools/binary_codegen.py

+import argparse
+import os
+import sys
+import struct
+
+import jinja2
+
+import numpy as np
+
+# python mace/python/tools/binary_codegen.py \
+#     --binary_file=${BIN_FILE} --output_path=${CODE_GEN_PATH} --variable_name=kTuningParamsData
+
+FLAGS = None
+
+
+def generate_cpp_source():
+  with open(FLAGS.binary_file, "rb") as binary_file:
+    binary_array = np.fromfile(binary_file, dtype=np.uint8)
+
+  data_map = {}
+
+  idx = 0
+  size, = struct.unpack("Q", binary_array[idx:idx+8])
+  print size
+  idx += 8
+  for _ in xrange(size):
+    key_size, = struct.unpack("i", binary_array[idx:idx+4])
+    idx += 4
+    key, = struct.unpack(str(key_size) + "s", binary_array[idx:idx+key_size])
+    idx += key_size
+    params_size, = struct.unpack("i", binary_array[idx:idx+4])
+    idx += 4
+    data_map[key] = []
+    count = params_size / 4
+    params = struct.unpack(str(count) + "i", binary_array[idx:idx+params_size])
+    for i in params:
+      data_map[key].append(i)
+    idx += params_size
+
+  env = jinja2.Environment(loader=jinja2.FileSystemLoader(sys.path[0]))
+  return env.get_template('binary.cc.tmpl').render(
+    maps = data_map,
+    data_type = 'int',
+    variable_name = FLAGS.variable_name
+  )
+
+def main(unused_args):
+
+  cpp_binary_source = generate_cpp_source()
+  if os.path.isfile(FLAGS.output_path):
+    os.remove(FLAGS.output_path)
+  w_file = open(FLAGS.output_path, "w")
+  w_file.write(cpp_binary_source)
+  w_file.close()
+
+def parse_args():
+  """Parses command line arguments."""
+  parser = argparse.ArgumentParser()
+  parser.add_argument(
+      "--binary_file",
+      type=str,
+      default="",
+      help="The binaries file path.")
+  parser.add_argument(
+      "--output_path",
+      type=str,
+      default="",
+      help="The path of generated C++ source file which contains the binary.")
+  parser.add_argument(
+    "--variable_name",
+    type=str,
+    default="kTuningParamsData",
+    help="global variable name.")
+  return parser.parse_known_args()
+
+
+if __name__ == '__main__':
+  FLAGS, unparsed = parse_args()
+  main(unused_args=[sys.argv[0]] + unparsed)

mace/python/tools/opencl_codegen.py

@@ -14,7 +14,7 @@ FLAGS = None
 
 
 def generate_cpp_source():
-  maps = {"binary_maps": []}
+  maps = {}
   for file_name in os.listdir(FLAGS.cl_binary_dir):
     file_path = os.path.join(FLAGS.cl_binary_dir, file_name)
     if file_path[-4:] == ".bin":
@@ -23,13 +23,16 @@ def generate_cpp_source():
       binary_array = np.fromfile(f, dtype=np.uint8)
       f.close()
 
-      binary_dict = {"name": file_name[:-4], "content": []}
+      maps[file_name[:-4]] = []
       for ele in binary_array:
-        binary_dict["content"].append(hex(ele))
-      maps["binary_maps"].append(binary_dict)
+        maps[file_name[:-4]].append(hex(ele))
 
   env = jinja2.Environment(loader=jinja2.FileSystemLoader(sys.path[0]))
-  return env.get_template('opencl_compiled_program.cc.tmpl').render(maps)
+  return env.get_template('binary.cc.tmpl').render(
+    maps = maps,
+    data_type = 'unsigned char',
+    variable_name = 'kCompiledProgramMap'
+  )
 
 
 def main(unused_args):

mace/utils/tuner.h

@@ -11,6 +11,7 @@
 #include <string>
 #include <unordered_map>
 #include <vector>
+#include <map>
 
 #include "mace/utils/logging.h"
 #include "mace/utils/timer.h"
@@ -72,7 +73,7 @@ class Tuner {
     if (path_ != nullptr) {
       std::ofstream ofs(path_, std::ios::binary | std::ios::out);
       if (ofs.is_open()) {
-        size_t num_pramas = param_table_.size();
+        int64_t num_pramas = param_table_.size();
         ofs.write(reinterpret_cast<char *>(&num_pramas), sizeof(num_pramas));
         for (auto &kp : param_table_) {
           int32_t key_size = kp.first.size();
@@ -97,13 +98,20 @@ class Tuner {
   }
 
   inline void ReadRunParamters() {
+#ifdef MACE_EMBED_BINARY_PROGRAM
+    extern const std::map<std::string, std::vector<int>> kTuningParamsData;
+    VLOG(1) << "Read tuning parameters from source";
+    for (auto it = kTuningParamsData.begin(); it != kTuningParamsData.end(); ++it) {
+      param_table_.emplace(it->first, std::vector<param_type>(it->second.begin(), it->second.end()));
+    }
+#else
     if (path_ != nullptr) {
       std::ifstream ifs(path_, std::ios::binary | std::ios::in);
       if (ifs.is_open()) {
         int32_t key_size = 0;
         int32_t params_size = 0;
         int32_t params_count = 0;
-        size_t num_pramas = 0;
+        int64_t num_pramas = 0;
         ifs.read(reinterpret_cast<char *>(&num_pramas), sizeof(num_pramas));
         while (num_pramas--) {
           ifs.read(reinterpret_cast<char *>(&key_size), sizeof(key_size));
@@ -123,6 +131,7 @@ class Tuner {
         LOG(WARNING) << "Read run parameter file failed.";
       }
     }
+#endif
   }
 
   template <typename RetType>

tools/validate_gcn.sh

@@ -2,10 +2,10 @@
 # Must run at root dir of mace project.
 set +x
 Usage() {
-  echo 'Usage: bash tools/validate_gcn.sh tf_model_path image_size'
+  echo 'Usage: bash tools/validate_gcn.sh tf_model_path image_size [tuning]'
 }
 
-if [ $# != 2 ];then
+if [ $# -lt 2 ];then
   Usage
   exit -1
 fi
@@ -25,6 +25,8 @@ CODEGEN_DIR=${MACE_SOURCE_DIR}/mace/codegen
 MODEL_CODEGEN_DIR=${CODEGEN_DIR}/models/gcn-$IMAGE_SIZE
 CL_CODEGEN_DIR=${CODEGEN_DIR}/opencl
 CL_BIN_DIR=${CODEGEN_DIR}/opencl_bin
+TUNING_CODEGEN_DIR=${CODEGEN_DIR}/tuning
+TUNING_OR_NOT=${3:-0}
 
 build_and_run()
 {
@@ -40,18 +42,24 @@ build_and_run()
     $EMBED_OPENCL_BINARY_BUILD_FLAGS \
     --copt=-DMACE_MODEL_FUNCTION=Create${MODEL_TAG}
 
-  adb shell "rm -rf ${PHONE_DATA_DIR}"
   adb shell "mkdir -p ${PHONE_DATA_DIR}"
   if [ "$EMBED_OPENCL_BINARY" = false ]; then
+    adb shell "rm -rf ${KERNEL_DIR}"
     adb shell "mkdir -p ${KERNEL_DIR}"
-    adb push mace/kernels/opencl/cl/ ${KERNEL_DIR}
+    adb push mace/kernels/opencl/cl/. ${KERNEL_DIR}
   fi
   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
+  else
+    tuning_flag=0
+  fi
 
-  adb </dev/null shell MACE_CPP_MIN_VLOG_LEVEL=0 \
+  adb </dev/null shell MACE_TUNING=${tuning_flag} \
+    MACE_CPP_MIN_VLOG_LEVEL=0 \
     MACE_RUN_PARAMETER_PATH=${PHONE_DATA_DIR}/mace_run.config \
     MACE_KERNEL_PATH=$KERNEL_DIR \
     OMP_NUM_THREADS=$num_threads \
@@ -73,7 +81,7 @@ python tools/validate.py --generate_data true --random_seed 1 \
 
 echo "Step 2: Convert tf model to mace model and optimize memory"
 bazel build //mace/python/tools:tf_converter
-rm -rf ${MODEL_CODEGEN_DIR}
+rm -rf ${CODEGEN_DIR}/models
 mkdir -p ${MODEL_CODEGEN_DIR}
 bazel-bin/mace/python/tools/tf_converter --input=${TF_MODEL_FILE_PATH} \
                                          --output=${MODEL_CODEGEN_DIR}/mace_gcn${IMAGE_SIZE}.cc \
@@ -86,25 +94,31 @@ bazel-bin/mace/python/tools/tf_converter --input=${TF_MODEL_FILE_PATH} \
                                          --model_tag=${MODEL_TAG} \
                                          --confuse=True
 
-echo "Step 3: Run model on the phone"
+echo "Step 3: Run model on the phone with files"
 build_and_run false
 
 echo "Step 4: Generate OpenCL binary program and config code"
-rm -rf ${CL_BIN_DIR}/*
+rm -rf ${CL_BIN_DIR}
 adb pull ${KERNEL_DIR} ${CL_BIN_DIR}
 rm -rf ${CL_CODEGEN_DIR}
 mkdir -p ${CL_CODEGEN_DIR}
 python mace/python/tools/opencl_codegen.py \
   --cl_binary_dir=${CL_BIN_DIR} --output_path=${CL_CODEGEN_DIR}/opencl_compiled_program.cc
 
-echo "Step 5: Run model on the phone"
+echo "Step 5: Generate tuning source file"
+adb pull ${PHONE_DATA_DIR}/mace_run.config ${CL_BIN_DIR}
+mkdir -p ${TUNING_CODEGEN_DIR}
+python mace/python/tools/binary_codegen.py \
+  --binary_file=${CL_BIN_DIR}/mace_run.config --output_path=${TUNING_CODEGEN_DIR}/tuning_params.cc
+
+echo "Step 6: Run model on the phone using binary"
 build_and_run true
 
-echo "Step 6: Pull the mace run result."
+echo "Step 7: Pull the mace run result."
 rm -rf ${MODEL_DIR}/${OUTPUT_FILE_NAME}
 adb </dev/null pull ${PHONE_DATA_DIR}/${OUTPUT_FILE_NAME} ${MODEL_DIR}
 
-echo "Step 7: Validate the result"
+echo "Step 8: Validate the result"
 python tools/validate.py --model_file ${TF_MODEL_FILE_PATH} \
     --input_file ${MODEL_DIR}/${INPUT_FILE_NAME} \
     --mace_out_file ${MODEL_DIR}/${OUTPUT_FILE_NAME} \