Revert parallel code generation (incomplete prerequisite handling)

Removes LIBINT2_NUM_WORKERS, WorkerConfig, build_libint_parallel.sh, and
all worker partitioning logic. The process-level parallelism produced
incomplete output (missing CR header files) because generate_rr_code needs
external symbols from ALL quartets but workers only discover their subset.

Retains: type dispatch cache, CSE disable, braket tiebreaker, progress bar.

Author: kshitij-05

CMakeLists.txt

@@ -270,8 +270,6 @@ option_with_print(LIBINT2_SINGLE_EVALTYPE
         "Generate single evaluator type (i.e. all tasks use the same evaluator). OFF is NYI" ON)
 option_with_default(LIBINT2_ENABLE_UNROLLING
         "Unroll shell sets into integrals (will unroll shell sets larger than N) (0 for never, N for N, 1000000000 for always)" 100)
-option_with_default(LIBINT2_NUM_WORKERS
-        "Number of parallel build_libint processes for code generation (1 for serial)" 1)
 option_with_default(LIBINT2_ALIGN_SIZE
         "(EXPERT) if posix_memalign is available, this will specify alignment of Libint data, in units of
    sizeof(LIBINT2_REALTYPE). Default is to use built-in heuristics: system-determined for vectorization off (default) or veclen * sizeof(LIBINT2_REALTYPE) for vectorization on." 0)

bin/build_libint_parallel.sh

@@ -32,7 +32,6 @@
 
 #include <boost/preprocessor.hpp>
 #include <chrono>
-#include <cstdlib>
 #include <fstream>
 #include <iomanip>
 #include <iostream>
@@ -71,55 +70,6 @@ using namespace libint2;
 
 CodeGenProgress g_progress;
 
-/// Process-level parallelism: each worker generates code for a subset of
-/// quartets. Configured via LIBINT_NUM_WORKERS and LIBINT_WORKER_ID env vars.
-struct WorkerConfig {
-  unsigned int num_workers = 1;
-  unsigned int worker_id = 0;
-  std::unique_ptr<std::ofstream> si_frag;  // static_init fragment
-  std::unique_ptr<std::ofstream> ii_frag;  // int_iface fragment
-
-  void init(const std::string& source_dir = "") {
-    const char* nw = std::getenv("LIBINT_NUM_WORKERS");
-    const char* wi = std::getenv("LIBINT_WORKER_ID");
-    if (nw) num_workers = std::max(1u, static_cast<unsigned>(std::atoi(nw)));
-    if (wi) worker_id = static_cast<unsigned>(std::atoi(wi));
-    if (worker_id >= num_workers) worker_id = 0;
-  }
-
-  void open_fragments(const std::string& source_dir) {
-    if (!is_parallel()) return;
-    si_frag.reset(new std::ofstream(source_dir + "libint2_static_init_" +
-                                    std::to_string(worker_id) + ".inc"));
-    ii_frag.reset(new std::ofstream(source_dir + "libint2_iface_internal_" +
-                                    std::to_string(worker_id) + ".inc"));
-  }
-
-  /// Write to static_init: either iface or fragment file
-  void write_static_init(Libint2Iface* iface, const std::string& s) {
-    if (si_frag)
-      *si_frag << s << std::endl;
-    else
-      iface->to_static_init(s);
-  }
-
-  /// Write to int_iface: either iface or fragment file
-  void write_int_iface(Libint2Iface* iface, const std::string& s) {
-    if (ii_frag)
-      *ii_frag << s << std::endl;
-    else
-      iface->to_int_iface(s);
-  }
-
-  /// Returns true if this worker should process the given global quartet index
-  bool should_process(unsigned int global_idx) const {
-    return (global_idx % num_workers) == worker_id;
-  }
-
-  bool is_parallel() const { return num_workers > 1; }
-};
-static WorkerConfig g_worker;
-
 enum ShellSetType {
   ShellSetType_Standard = LIBINT_SHELL_SET_STANDARD,
   ShellSetType_ORCA = LIBINT_SHELL_SET_ORCA
@@ -597,14 +547,14 @@ void build_onebody_1b_1k(std::ostream& os, std::string label,
           << la << "][" << lb
           << "] = " << context->label_to_function_name(eval_label)
           << context->end_of_stat() << endl;
-      g_worker.write_static_init(iface.get(), oss.str());
+      iface->to_static_init(oss.str());
 
       // need to declare this function internally
       for (std::deque<std::string>::const_iterator i = decl_filenames.begin();
            i != decl_filenames.end(); ++i) {
         oss.str("");
         oss << "#include <" << *i << ">" << endl;
-        g_worker.write_int_iface(iface.get(), oss.str());
+        iface->to_int_iface(oss.str());
       }
 
 #if DEBUG
@@ -621,12 +571,6 @@ void build_onebody_1b_1k(std::ostream& os, std::string label,
 void try_main(int argc, char* argv[]) {
   std::ostream& os = cout;
 
-  g_worker.init();
-  if (g_worker.is_parallel()) {
-    std::cerr << "  Worker " << g_worker.worker_id << " of "
-              << g_worker.num_workers << std::endl;
-  }
-
   // First must declare the tasks
   LibraryTaskManager& taskmgr = LibraryTaskManager::Instance();
   taskmgr.add("default");
@@ -1025,7 +969,6 @@ void try_main(int argc, char* argv[]) {
   cparams->print(os);
 
   g_progress.start();
-  g_worker.open_fragments(cparams->source_directory());
 
 #ifdef LIBINT_INCLUDE_ONEBODY
   for (unsigned int d = 0; d <= LIBINT_INCLUDE_ONEBODY; ++d) {
@@ -1087,41 +1030,6 @@ void try_main(int argc, char* argv[]) {
 
   g_progress.finish();
 
-  // Close fragment files if in parallel mode
-  g_worker.si_frag.reset();
-  g_worker.ii_frag.reset();
-
-  // In parallel mode, non-zero workers exit here after generating source
-  // files and iface fragments. Worker 0 continues to merge and finalize.
-  if (g_worker.is_parallel() && g_worker.worker_id != 0) {
-    os << "Worker " << g_worker.worker_id << " finished." << endl;
-    return;
-  }
-
-  // Merge worker fragment files into the iface (worker 0 only)
-  if (g_worker.is_parallel()) {
-    for (unsigned int w = 0; w < g_worker.num_workers; ++w) {
-      auto read_and_append =
-          [&](const std::string& frag_name,
-              void (Libint2Iface::*writer)(const std::string&)) {
-            std::string fpath = cparams->source_directory() + frag_name;
-            std::ifstream fin(fpath);
-            if (fin.is_open()) {
-              std::string line;
-              while (std::getline(fin, line)) {
-                if (!line.empty()) (iface.get()->*writer)(line);
-              }
-              fin.close();
-              std::remove(fpath.c_str());
-            }
-          };
-      read_and_append("libint2_static_init_" + std::to_string(w) + ".inc",
-                      &Libint2Iface::to_static_init);
-      read_and_append("libint2_iface_internal_" + std::to_string(w) + ".inc",
-                      &Libint2Iface::to_int_iface);
-    }
-  }
-
   // Generate code for the set-level RRs
   std::deque<std::string> decl_filenames, def_filenames;
   generate_rr_code(os, cparams, decl_filenames, def_filenames);
@@ -1245,7 +1153,6 @@ static void build_TwoPRep_2b_2k(
 
   // Note: la, lb, lc, ld generate code for chemist notation (ab|O|cd), where O
   // is a two-body operator.
-  unsigned int quartet_idx = 0;
   for (unsigned int la = 0; la <= lmax; la++) {
     for (unsigned int lb = 0; lb <= lmax; lb++) {
       for (unsigned int lc = 0; lc <= lmax; lc++) {
@@ -1255,9 +1162,6 @@ static void build_TwoPRep_2b_2k(
                                             bra_ket_coswappable))
             continue;
 
-          // Worker filter: skip quartets not assigned to this worker
-          if (!g_worker.should_process(quartet_idx++)) continue;
-
           // std::shared_ptr<Tactic> tactic(new ParticleDirectionTactic(la+lb >
           // lc+ld ? false : true));
           std::shared_ptr<Tactic> tactic(
@@ -1424,13 +1328,13 @@ static void build_TwoPRep_2b_2k(
               << la << "][" << lb << "][" << lc << "][" << ld
               << "] = " << context->label_to_function_name(eval_label)
               << context->end_of_stat() << endl;
-          g_worker.write_static_init(iface.get(), oss.str());
+          iface->to_static_init(oss.str());
 
           // need to declare this function internally
           for (auto& decl_filename : decl_filenames) {
             oss.str("");
             oss << "#include <" << decl_filename << ">" << endl;
-            g_worker.write_int_iface(iface.get(), oss.str());
+            iface->to_int_iface(oss.str());
           }
 
 #if DEBUG
@@ -1634,13 +1538,13 @@ void build_TwoPRep_1b_2k(std::ostream& os,
             << lbra << "][" << lc << "][" << ld
             << "] = " << context->label_to_function_name(label)
             << context->end_of_stat() << endl;
-        g_worker.write_static_init(iface.get(), oss.str());
+        iface->to_static_init(oss.str());
 
         // need to declare this function internally
         for (auto& decl_filename : decl_filenames) {
           oss.str("");
           oss << "#include <" << decl_filename << ">" << endl;
-          g_worker.write_int_iface(iface.get(), oss.str());
+          iface->to_int_iface(oss.str());
         }
 
 #if DEBUG
@@ -1833,13 +1737,13 @@ void build_TwoPRep_1b_1k(std::ostream& os,
           << lbra << "][" << lket
           << "] = " << context->label_to_function_name(label)
           << context->end_of_stat() << endl;
-      g_worker.write_static_init(iface.get(), oss.str());
+      iface->to_static_init(oss.str());
 
       // need to declare this function internally
       for (auto& decl_filename : decl_filenames) {
         oss.str("");
         oss << "#include <" << decl_filename << ">" << endl;
-        g_worker.write_int_iface(iface.get(), oss.str());
+        iface->to_int_iface(oss.str());
       }
 
 #if DEBUG
@@ -2031,15 +1935,15 @@ void build_R12kG12_2b_2k(std::ostream& os,
               << la << "][" << lb << "][" << lc << "][" << ld
               << "] = " << context->label_to_function_name(label)
               << context->end_of_stat() << endl;
-          g_worker.write_static_init(iface.get(), oss.str());
+          iface->to_static_init(oss.str());
 
           // need to declare this function internally
           for (std::deque<std::string>::const_iterator i =
                    decl_filenames.begin();
                i != decl_filenames.end(); ++i) {
             oss.str("");
             oss << "#include <" << *i << ">" << endl;
-            g_worker.write_int_iface(iface.get(), oss.str());
+            iface->to_int_iface(oss.str());
           }
 
 #if DEBUG
@@ -2176,15 +2080,15 @@ void build_R12kG12_2b_2k_separate(
                 << "[" << la << "][" << lb << "][" << lc << "][" << ld
                 << "] = " << context->label_to_function_name(label)
                 << context->end_of_stat() << endl;
-            g_worker.write_static_init(iface.get(), oss.str());
+            iface->to_static_init(oss.str());
 
             // need to declare this function internally
             for (std::deque<std::string>::const_iterator i =
                      decl_filenames.begin();
                  i != decl_filenames.end(); ++i) {
               oss.str("");
               oss << "#include <" << *i << ">" << endl;
-              g_worker.write_int_iface(iface.get(), oss.str());
+              iface->to_int_iface(oss.str());
             }
 
 #if DEBUG
@@ -2362,11 +2266,11 @@ void build_G12DKH_2b_2k(std::ostream& os,
               << la << "][" << lb << "][" << lc << "][" << ld
               << "] = " << context->label_to_function_name(label)
               << context->end_of_stat() << endl;
-          g_worker.write_static_init(iface.get(), oss.str());
+          iface->to_static_init(oss.str());
 
           oss.str("");
           oss << "#include <" << decl_filename << ">" << endl;
-          g_worker.write_int_iface(iface.get(), oss.str());
+          iface->to_int_iface(oss.str());
 
           // For the most expensive (i.e. presumably complete) graph extract
           // all precomputed quantities -- these will be members of the

src/bin/libint/build_libint.cc

@@ -6,37 +6,17 @@ include(int_computed)  # for macros.tex and features
 # <<<  Generate Library Source  >>>
 
 file(MAKE_DIRECTORY "${PROJECT_BINARY_DIR}/generated")
-if (LIBINT2_NUM_WORKERS GREATER 1)
-  # Parallel code generation requires bash for background process management.
-  # The script launches N-1 workers in parallel, then worker 0 merges results.
-  find_program(BASH_EXECUTABLE bash REQUIRED)
-  add_custom_command(
-    OUTPUT
-      "${PROJECT_BINARY_DIR}/generated/libint2_params.h"
-      "${PROJECT_BINARY_DIR}/generated/HRRPart0bra0ket0pp.h"
-      "${PROJECT_BINARY_DIR}/generated/HRRPart0bra0ket0pp.cc"
-    COMMAND ${BASH_EXECUTABLE}
-      "${PROJECT_SOURCE_DIR}/bin/build_libint_parallel.sh"
-      $<TARGET_FILE:build_libint> ${LIBINT2_NUM_WORKERS}
-    DEPENDS
-      build_libint
-      "${PROJECT_SOURCE_DIR}/bin/build_libint_parallel.sh"
-    WORKING_DIRECTORY "${PROJECT_BINARY_DIR}/generated"
-    COMMENT "Generating Libint2 library source (${LIBINT2_NUM_WORKERS} workers)"
-    )
-else()
-  add_custom_command(
-    OUTPUT
-      "${PROJECT_BINARY_DIR}/generated/libint2_params.h"
-      "${PROJECT_BINARY_DIR}/generated/HRRPart0bra0ket0pp.h"
-      "${PROJECT_BINARY_DIR}/generated/HRRPart0bra0ket0pp.cc"
-    COMMAND $<TARGET_FILE:build_libint>
-    DEPENDS
-      build_libint
-    WORKING_DIRECTORY "${PROJECT_BINARY_DIR}/generated"
-    COMMENT "Generating Libint2 library source"
-    )
-endif()
+add_custom_command(
+  OUTPUT
+    "${PROJECT_BINARY_DIR}/generated/libint2_params.h"
+    "${PROJECT_BINARY_DIR}/generated/HRRPart0bra0ket0pp.h"
+    "${PROJECT_BINARY_DIR}/generated/HRRPart0bra0ket0pp.cc"
+  COMMAND $<TARGET_FILE:build_libint>
+  DEPENDS
+    build_libint
+  WORKING_DIRECTORY "${PROJECT_BINARY_DIR}/generated"
+  COMMENT "Generating Libint2 library source"
+  )
 add_custom_target(libint-library-generate DEPENDS "${PROJECT_BINARY_DIR}/generated/libint2_params.h")