[Performance] Remove phmap dependency. (#7658)

Author: mfbalin

.gitmodules

@@ -10,9 +10,6 @@
 [submodule "third_party/METIS"]
 	path = third_party/METIS
 	url = https://github.com/KarypisLab/METIS.git
-[submodule "third_party/phmap"]
-	path = third_party/phmap
-	url = https://github.com/greg7mdp/parallel-hashmap.git
 [submodule "third_party/nanoflann"]
 	path = third_party/nanoflann
 	url = https://github.com/jlblancoc/nanoflann

CMakeLists.txt

@@ -340,15 +340,9 @@ else(EXTERNAL_DMLC_PATH)
   set(GOOGLE_TEST 0) # Turn off dmlc-core test
 endif(EXTERNAL_DMLC_PATH)
 
-if(EXTERNAL_PHMAP_PATH)
-  include_directories(SYSTEM ${EXTERNAL_PHMAP_PATH})
-else(EXTERNAL_PHMAP_PATH)
-  target_include_directories(dgl PRIVATE "third_party/phmap")
-endif(EXTERNAL_PHMAP_PATH)
-
-
 target_include_directories(dgl PRIVATE "tensoradapter/include")
 target_include_directories(dgl PRIVATE "third_party/pcg/include")
+target_include_directories(dgl PRIVATE "third_party/tsl_robin_map/include")
 
 if(EXTERNAL_NANOFLANN_PATH)
   include_directories(SYSTEM ${EXTERNAL_NANOFLANN_PATH})
@@ -473,7 +467,7 @@ if(BUILD_CPP_TEST)
   include_directories("include")
   include_directories("third_party/dlpack/include")
   include_directories("third_party/dmlc-core/include")
-  include_directories("third_party/phmap")
+  include_directories("third_party/tsl_robin_map/include")
   include_directories("third_party/libxsmm/include")
   include_directories("third_party/pcg/include")
   file(GLOB_RECURSE TEST_SRC_FILES ${PROJECT_SOURCE_DIR}/tests/cpp/*.cc)

src/array/cpu/array_utils.h

@@ -7,7 +7,7 @@
 #define DGL_ARRAY_CPU_ARRAY_UTILS_H_
 
 #include <dgl/aten/types.h>
-#include <parallel_hashmap/phmap.h>
+#include <tsl/robin_map.h>
 
 #include <unordered_map>
 #include <utility>
@@ -53,8 +53,7 @@ class IdHashMap {
     const int64_t len = ids->shape[0];
     for (int64_t i = 0; i < len; ++i) {
       const IdType id = ids_data[i];
-      // phmap::flat_hash_map::insert assures that an insertion will not happen
-      // if the key already exists.
+      // Insertion will not happen if the key already exists.
       oldv2newv_.insert({id, oldv2newv_.size()});
       filter_[id & kFilterMask] = true;
     }
@@ -106,7 +105,7 @@ class IdHashMap {
   // lookups.
   std::vector<bool> filter_;
   // The hashmap from old vid to new vid
-  phmap::flat_hash_map<IdType, IdType> oldv2newv_;
+  tsl::robin_map<IdType, IdType> oldv2newv_;
 };
 
 /**

src/array/cpu/csr_mm.cc

@@ -6,7 +6,8 @@
 
 #include <dgl/array.h>
 #include <dgl/runtime/parallel_for.h>
-#include <parallel_hashmap/phmap.h>
+#include <tsl/robin_map.h>
+#include <tsl/robin_set.h>
 
 #include <vector>
 
@@ -28,7 +29,7 @@ void CountNNZPerRow(
     const IdType* B_indices, IdType* C_indptr_data, int64_t M) {
   parallel_for(0, M, [=](size_t b, size_t e) {
     for (auto i = b; i < e; ++i) {
-      phmap::flat_hash_set<IdType> set;
+      tsl::robin_set<IdType> set;
       for (IdType u = A_indptr[i]; u < A_indptr[i + 1]; ++u) {
         IdType w = A_indices[u];
         for (IdType v = B_indptr[w]; v < B_indptr[w + 1]; ++v)
@@ -60,7 +61,7 @@ void ComputeIndicesAndData(
     IdType* C_indices_data, DType* C_weights_data, int64_t M) {
   parallel_for(0, M, [=](size_t b, size_t e) {
     for (auto i = b; i < e; ++i) {
-      phmap::flat_hash_map<IdType, DType> map;
+      tsl::robin_map<IdType, DType> map;
       for (IdType u = A_indptr[i]; u < A_indptr[i + 1]; ++u) {
         IdType w = A_indices[u];
         DType vA = A_data[A_eids ? A_eids[u] : u];

src/array/cpu/csr_sum.cc

@@ -6,7 +6,8 @@
 
 #include <dgl/array.h>
 #include <dgl/runtime/parallel_for.h>
-#include <parallel_hashmap/phmap.h>
+#include <tsl/robin_map.h>
+#include <tsl/robin_set.h>
 
 #include <vector>
 
@@ -30,7 +31,7 @@ void CountNNZPerRow(
 
   runtime::parallel_for(0, M, [=](size_t b, size_t e) {
     for (size_t i = b; i < e; ++i) {
-      phmap::flat_hash_set<IdType> set;
+      tsl::robin_set<IdType> set;
       for (int64_t k = 0; k < n; ++k) {
         for (IdType u = A_indptr[k][i]; u < A_indptr[k][i + 1]; ++u)
           set.insert(A_indices[k][u]);
@@ -63,7 +64,7 @@ void ComputeIndicesAndData(
   int64_t n = A_indptr.size();
   runtime::parallel_for(0, M, [=](size_t b, size_t e) {
     for (auto i = b; i < e; ++i) {
-      phmap::flat_hash_map<IdType, DType> map;
+      tsl::robin_map<IdType, DType> map;
       for (int64_t k = 0; k < n; ++k) {
         for (IdType u = A_indptr[k][i]; u < A_indptr[k][i + 1]; ++u) {
           IdType kA = A_indices[k][u];

src/array/cpu/labor_pick.h

@@ -26,7 +26,7 @@
 #include <dgl/random.h>
 #include <dgl/runtime/parallel_for.h>
 #include <dmlc/omp.h>
-#include <parallel_hashmap/phmap.h>
+#include <tsl/robin_map.h>
 
 #include <algorithm>
 #include <cmath>
@@ -45,6 +45,13 @@ namespace impl {
 
 using dgl::random::continuous_seed;
 
+template <typename K, typename V>
+using map_t = tsl::robin_map<K, V>;
+template <typename iterator>
+auto& mutable_value_ref(iterator it) {
+  return it.value();
+}
+
 constexpr double eps = 0.0001;
 
 template <typename IdxType, typename FloatType>
@@ -61,7 +68,7 @@ auto compute_importance_sampling_probabilities(
 
   double prev_ex_nodes = max_degree * num_rows;
 
-  phmap::flat_hash_map<IdxType, FloatType> hop_map, hop_map2;
+  map_t<IdxType, FloatType> hop_map, hop_map2;
   for (int iters = 0; iters < importance_sampling || importance_sampling < 0;
        iters++) {
     // NOTE(mfbalin) When the graph is unweighted, the first c values in
@@ -83,7 +90,9 @@ auto compute_importance_sampling_probabilities(
         for (auto j = indptr[rid]; j < indptr[rid + 1]; j++) {
           const auto ct = c * (weighted && iters == 1 ? A[j] : 1);
           auto itb = hop_map2.emplace(indices[j], ct);
-          if (!itb.second) itb.first->second = std::max(ct, itb.first->second);
+          if (!itb.second) {
+            mutable_value_ref(itb.first) = std::max(ct, itb.first->second);
+          }
         }
       }
       if (hop_map.empty())
@@ -203,7 +212,7 @@ std::pair<COOMatrix, FloatArray> CSRLaborPick(
     hop_size += act_degree;
   }
 
-  phmap::flat_hash_map<IdxType, FloatType> hop_map;
+  map_t<IdxType, FloatType> hop_map;
 
   if (importance_sampling)
     hop_map = compute_importance_sampling_probabilities<IdxType, FloatType>(