Use static functions/variables if possible (pytorch#4432)

Summary:
Pull Request resolved: pytorch#4432

X-link: facebookresearch/FBGEMM#1497

There are two changes:
1. Marks in-file templates and other inner functions as static; this provides more opportunities to optimise code, i.e. followed by enabling link time optimization. An unused function in test code is removed.
2. Enables `misc-use-internal-linkage` check.

Pull Request resolved: pytorch#4423

Reviewed By: cthi

Differential Revision: D77635949

Pulled By: q10

fbshipit-source-id: 3ce5cfd435be216ccc48c5be9eb30e91dcfcfbb5

Author: cyyever

.clang-tidy

@@ -4,8 +4,10 @@
 # Get options for config files in parent directories,
 # but override them if there's a conflict.
 InheritParentConfig: true
+# @nolint
 Checks: '
 bugprone-argument-comment,
+misc-use-internal-linkage,
 '
 CheckOptions:
  - key: facebook-cuda-safe-api-call-check.HandlerName

bench/BenchUtils.cc

@@ -19,7 +19,7 @@
 
 namespace fbgemm {
 
-std::default_random_engine eng;
+static std::default_random_engine eng;
 
 template <typename T>
 void randFill(aligned_vector<T>& vec, T low, T high, std::true_type) {

bench/ConvUnifiedBenchmark.cc

@@ -28,7 +28,7 @@ using namespace fbgemm;
 
 // clang-format off
 // 1D conv shapes
-vector<conv_param_t<1>> shapes_1d = {
+static vector<conv_param_t<1>> shapes_1d = {
   // MB, IC, OC, IW, G, KW, stride_w, pad_w_left, pad_w_right,
   // (dilation, output_padding_w, tranpose)
   // regular
@@ -46,7 +46,7 @@ vector<conv_param_t<1>> shapes_1d = {
 };
 
 // 2D conv shapes
-vector<conv_param_t<2>> shapes_2d = {
+static vector<conv_param_t<2>> shapes_2d = {
   // MB, IC, OC, IH, IW, G, KH, KW, stride_h, stride_w,
   // pad_h_top, pad_w_left, pad_h_bottom, pad_w_right,
   // (dilation_h, dilation_w, output_padding_h, output_padding_w, tranpose)
@@ -84,7 +84,7 @@ vector<conv_param_t<2>> shapes_2d = {
       {1, 1}, {0, 0, 0, 0})
 };
 
-vector<conv_param_t<2>> shapes_2d_resnext_101 = {
+static vector<conv_param_t<2>> shapes_2d_resnext_101 = {
   // ResNext-101 (unique shapes only)
   // conv_param_t<>(N, C, M, H, W, groups, /* kern */ {KH, KW}, /* stride */
   //   {stride_h, stride_w}, /* padding pad_l = pad_h */ {pad_l, pad_l, pad_l, pad_l}, /* dialation */
@@ -143,7 +143,7 @@ vector<conv_param_t<2>> shapes_2d_resnext_101 = {
 };
 
 // 3D conv shapes
-vector<conv_param_t<3>> shapes_3d = {
+static vector<conv_param_t<3>> shapes_3d = {
   // MB, IC, OC, {IT, IH, IW}, G, {KT, KH, KW}, {stride_t, stride_h,
   // stride_w},
   // {pad_prev, pad_h_top, pad_w_left, pad_next, pad_h_bottom, pad_w_right},
@@ -216,7 +216,7 @@ vector<conv_param_t<3>> shapes_3d = {
 // clang-format on
 
 template <int SPATIAL_DIM, typename Acc_t>
-void performance_test(
+static void performance_test(
     const vector<conv_param_t<SPATIAL_DIM>>& shapes,
     bool flush,
     int repetitions) {

bench/ConvertBenchmark.cc

@@ -21,7 +21,7 @@
 using namespace std;
 using namespace fbgemm;
 
-void performance_test() {
+static void performance_test() {
   constexpr int NWARMUP = 4;
   constexpr int NITER = 256;
 

bench/EmbeddingIndexRemappingBenchmark.cc

@@ -37,7 +37,7 @@ static vector<vector<int>> GetInputs_() {
   return input_dims;
 }
 
-int run_benchmark(
+static int run_benchmark(
     int batch_size,
     int num_rows,
     int average_len,

bench/EmbeddingQuantizeBenchmark.cc

@@ -25,7 +25,7 @@ using namespace fbgemm;
 
 // T is the type of scale and bias
 template <typename T>
-void performance_test() {
+static void performance_test() {
   constexpr int NWARMUP = 4;
   constexpr int NITER = 256;
 

bench/EmbeddingQuantizeFloatToFloatOrHalfBenchmark.cc

@@ -25,7 +25,7 @@ using namespace fbgemm;
 
 // T is the type of scale and bias
 template <typename T>
-void performance_test() {
+static void performance_test() {
   constexpr int NWARMUP = 4;
   constexpr int NITER = 256;
 

bench/EmbeddingSpMDM8BitBenchmark.cc

@@ -30,15 +30,16 @@
 using namespace std;
 using namespace fbgemm;
 
-void print_fused_table(int rows, int embedding_dim, const uint8_t* table) {
-  for (int i = 0; i < rows; i++) {
-    cout << "row: " << i << " : " << endl;
-    for (int ii = 0; ii < embedding_dim; ii++) {
-      cout << (int)table[i * (embedding_dim + 2 * sizeof(float)) + ii] << ",";
+/*
+static void print_fused_table(int rows, int embedding_dim, const uint8_t* table)
+{ for (int i = 0; i < rows; i++) { cout << "row: " << i << " : " << endl; for
+(int ii = 0; ii < embedding_dim; ii++) { cout << (int)table[i * (embedding_dim +
+2 * sizeof(float)) + ii] << ",";
     }
     cout << endl;
   }
 }
+*/
 
 static vector<vector<int>> GetInputs_() {
   vector<vector<int>> input_dims = {
@@ -58,10 +59,10 @@ static vector<vector<int>> GetInputs_() {
   return input_dims;
 }
 
-vector<double> benchmarkTimes;
+static vector<double> benchmarkTimes;
 
 template <typename OutType>
-int run_benchmark(
+static int run_benchmark(
     int batch_size,
     int num_rows,
     int embedding_dim,

bench/EmbeddingSpMDMBenchmark.cc

@@ -49,7 +49,7 @@ static vector<vector<int>> GetInputs_() {
   return input_dims;
 }
 
-void run_benchmark(
+static void run_benchmark(
     int batch_size,
     int num_rows,
     int embedding_dim,

bench/EmbeddingSpMDMNBit2Benchmark.cc

@@ -171,17 +171,6 @@ static void print_benchmark_results() {
   }
 }
 
-void print_fused_table(int rows, int embedding_dim, const uint8_t* table) {
-  for (int i = 0; i < rows; i++) {
-    std::cout << "row: " << i << " : " << std::endl;
-    for (int ii = 0; ii < embedding_dim; ii++) {
-      std::cout << (int)table[i * (embedding_dim + 2 * sizeof(float)) + ii]
-                << ",";
-    }
-    std::cout << std::endl;
-  }
-}
-
 static vector<vector<int>> GetInputs_() {
   vector<vector<int>> input_dims = {
       // batch size, number of rows of table, emb dim , avg lengthl
@@ -200,7 +189,7 @@ static vector<vector<int>> GetInputs_() {
   return input_dims;
 }
 
-int run_benchmark(
+static int run_benchmark(
     int bit_rate,
     int batch_size,
     int num_rows,
@@ -488,7 +477,7 @@ int run_benchmark(
   return 0;
 }
 
-void sweep_benchmark(KernelType kern_type) {
+static void sweep_benchmark(KernelType kern_type) {
   int batch_size;
   int num_rows;
   int embedding_dim;