Merge remote-tracking branch 'private/master' into gpl-fix-hpwl-print

Author: gudeh

BUILD.bazel

@@ -8,7 +8,7 @@ load("//bazel:tcl_wrap_cc.bzl", "tcl_wrap_cc")
 package(
     features = [
         "-parse_headers",
-        "-layering_check",
+        "layering_check",
         # TODO(b/299593765): Fix strict ordering.
         "-libcxx_assertions",
     ],
@@ -84,6 +84,8 @@ OPENROAD_LIBRARY_DEPS = [
     "//src/psm",
     "//src/rcx",
     "//src/rcx:ui",
+    "//src/rmp",
+    "//src/rmp:ui",
     "//src/rsz",
     "//src/rsz:ui",
     "//src/stt",
@@ -129,24 +131,20 @@ cc_binary(
     name = "openroad",
     srcs = [
         "src/Main.cc",
-        "src/OpenRoad.cc",
-        ":openroad_swig",
-        ":openroad_tcl",
-        ":rmp_swig",
-        ":rmp_tcl",
         "//bazel:runfiles",
     ],
     copts = OPENROAD_COPTS,
     features = ["-use_header_modules"],
     malloc = "@tcmalloc//tcmalloc",
     visibility = ["//visibility:public"],
     deps = [
-        ":openroad_lib_private",
+        ":openroad_lib",
         ":openroad_version",
-        "//src/odb",
-        "//src/rsz",
+        "//:ord",
+        "//src/gui",
         "//src/sta:opensta_lib",
         "//src/utl",
+        "@boost.stacktrace",
         "@rules_cc//cc/runfiles",
         "@tk_tcl//:tcl",
     ],
@@ -165,21 +163,17 @@ cc_library(
         "src/Design.cc",
         "src/Tech.cc",
         "src/Timing.cc",
-    ] + glob([
-        "src/rmp/src/*.h",
-        "src/rmp/src/*.cpp",
-    ]),
-    hdrs = glob([
-        "src/rmp/include/rmp/*.h",
-    ]),
+    ],
     copts = OPENROAD_COPTS,
     defines = OPENROAD_DEFINES + GUI_BUILD_FLAGS,
     features = ["-use_header_modules"],
     includes = [
         "include",
-        "src/rmp/include",
     ],
-    deps = OPENROAD_LIBRARY_DEPS,
+    deps = OPENROAD_LIBRARY_DEPS + [
+        "//src/sta:opensta_lib",
+        "@tk_tcl//:tcl",
+    ]
 )
 
 cc_library(
@@ -191,22 +185,19 @@ cc_library(
         "src/Timing.cc",
         ":openroad_swig",
         ":openroad_tcl",
-        ":rmp_swig",
-        ":rmp_tcl",
-    ] + glob([
-        "src/rmp/src/*.cpp",
-        "src/rmp/src/*.h",
-    ]),
-    hdrs = glob(["src/rmp/include/rmp/*.h"]),
+    ],
     copts = OPENROAD_COPTS,
     defines = OPENROAD_DEFINES + GUI_BUILD_FLAGS,
     features = ["-use_header_modules"],
     includes = [
         "include",
-        "src/rmp/include",
     ],
     visibility = ["//visibility:public"],
-    deps = OPENROAD_LIBRARY_DEPS,
+    deps = OPENROAD_LIBRARY_DEPS + [
+        "//src/sta:opensta_lib",
+        "@boost.stacktrace",
+        "@tk_tcl//:tcl",
+    ]
 )
 
 cc_library(
@@ -254,15 +245,6 @@ tcl_encode(
     namespace = "ord",
 )
 
-tcl_encode(
-    name = "rmp_tcl",
-    srcs = [
-        "src/rmp/src/rmp.tcl",
-    ],
-    char_array_name = "rmp_tcl_inits",
-    namespace = "rmp",
-)
-
 tcl_wrap_cc(
     name = "openroad_swig",
     srcs = [
@@ -277,22 +259,6 @@ tcl_wrap_cc(
     ],
 )
 
-tcl_wrap_cc(
-    name = "rmp_swig",
-    srcs = [
-        "src/rmp/src/rmp.i",
-        ":error_swig",
-        "//src/sta:sta_swig_files",
-    ],
-    module = "rmp",
-    namespace_prefix = "rmp",
-    root_swig_src = "src/rmp/src/rmp.i",
-    swig_includes = [
-        "src/rmp/src",
-        "src/sta",
-    ],
-)
-
 filegroup(
     name = "error_swig",
     srcs = [

WORKSPACE

@@ -8,9 +8,9 @@ workspace(name = "openroad")
 
 load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
 
-rules_hdl_git_hash = "56da46a87e8e5a4dbe84c0bbe5d00e92b936494f"
+rules_hdl_git_hash = "cf2bd95334741db74b39b47fa1d4622b0d45ce6c"
 
-rules_hdl_git_sha256 = "dc184ad0fe92f315eb5600fb3293c94ce1fce3fc1d0fd79400107038ed917d70"
+rules_hdl_git_sha256 = "137e1fbde970a41f295ca348f9105bc7eedd6640374ec7f4a8abaee9a2cfc2d0"
 
 http_archive(
     name = "rules_hdl",

etc/DependencyInstaller.sh

@@ -91,7 +91,7 @@ _installCommonDev() {
     cmakeBin=${cmakePrefix}/bin/cmake
     if [[ ! -f ${cmakeBin} || -z $(${cmakeBin} --version | grep ${cmakeVersionBig}) ]]; then
         cd "${baseDir}"
-        eval wget https://cmake.org/files/v${cmakeVersionBig}/cmake-${cmakeVersionSmall}-${osName}-${arch}.sh
+        eval wget https://github.com/Kitware/CMake/releases/download/v${cmakeVersionSmall}/cmake-${cmakeVersionSmall}-linux-${arch}.sh
         md5sum -c <(echo "${cmakeChecksum} cmake-${cmakeVersionSmall}-${osName}-${arch}.sh") || exit 1
         chmod +x cmake-${cmakeVersionSmall}-${osName}-${arch}.sh
         ./cmake-${cmakeVersionSmall}-${osName}-${arch}.sh --skip-license --prefix=${cmakePrefix}
@@ -861,10 +861,10 @@ while [ "$#" -gt 0 ]; do
             export GIT_SSL_NO_VERIFY=true
             ;;
         -save-deps-prefixes=*)
-            saveDepsPrefixes=$(realpath ${1#-save-deps-prefixes=})
+            saveDepsPrefixes=$(realpath ${1#*=})
             ;;
         -threads=*)
-            numThreads=${1}
+            numThreads=${1#*=}
             ;;
         *)
             echo "unknown option: ${1}" >&2
@@ -933,11 +933,13 @@ case "${os}" in
             _installOrTools "ubuntu" "${ubuntuVersion}" "amd64"
         fi
         ;;
-    "Red Hat Enterprise Linux" | "Rocky Linux")
+    "Red Hat Enterprise Linux" | "Rocky Linux" | "AlmaLinux")
     if [[ "${os}" == "Red Hat Enterprise Linux" ]]; then
         rhelVersion=$(rpm -q --queryformat '%{VERSION}' redhat-release | cut -d. -f1)
     elif  [[ "${os}" == "Rocky Linux" ]]; then
         rhelVersion=$(rpm -q --queryformat '%{VERSION}' rocky-release | cut -d. -f1)
+    elif [[ "${os}" == "AlmaLinux" ]]; then
+        rhelVersion=$(rpm -q --queryformat '%{VERSION}' almalinux-release | cut -d. -f1)
     fi
         if [[ "${rhelVersion}" != "8" ]] && [[ "${rhelVersion}" != "9" ]]; then
             echo "ERROR: Unsupported ${rhelVersion} version. Versions '8' and '9' are supported."
@@ -1018,4 +1020,8 @@ esac
 if [[ ! -z ${saveDepsPrefixes} ]]; then
     mkdir -p "$(dirname $saveDepsPrefixes)"
     echo "$CMAKE_PACKAGE_ROOT_ARGS" > $saveDepsPrefixes
+    # Fix permissions if running as root to allow user access
+    if [[ $(id -u) == 0 && ! -z "${SUDO_USER+x}" ]]; then
+        chown "$SUDO_USER:$(id -gn "$SUDO_USER")" "$saveDepsPrefixes" 2>/dev/null || true
+    fi
 fi

src/cts/README.md

@@ -61,7 +61,7 @@ clock_tree_synthesis
     [-num_static_layers]
     [-sink_clustering_buffer]
     [-obstruction_aware]
-    [-apply_ndr]
+    [-apply_ndr strategy]
     [-insertion_delay]
     [-dont_use_dummy_load]
     [-sink_buffer_max_cap_derate derate_value]
@@ -92,7 +92,7 @@ clock_tree_synthesis
 | `-num_static_layers` | Set the number of static layers. The default value is `0`, and the allowed values are integers `[0, MAX_INT]`. |
 | `-sink_clustering_buffer` | Set the sink clustering buffer(s) to be used. |
 | `-obstruction_aware` | Enables obstruction-aware buffering such that clock buffers are not placed on top of blockages or hard macros. This option may reduce legalizer displacement, leading to better latency, skew or timing QoR.  The default value is `False`, and the allowed values are bool. |
-| `-apply_ndr` | Applies 2X spacing non-default rule to all clock nets except leaf-level nets. The default value is `False`. |
+| `-apply_ndr` | Applies 2X spacing non-default rule to clock nets except leaf-level nets following some strategy. There are four strategy options: `none, root_only, half, full`. If this is not specified, the default value is `none`. |
 | `-dont_use_dummy_load` | Don't apply dummy buffer or inverter cells at clock tree leaves to balance loads. The default values is `False`. |
 | `-sink_buffer_max_cap_derate` | Use this option to control automatic buffer selection. To favor strong(weak) drive strength buffers use a small(large) value.  The default value is `0.01`, meaning that buffers are selected by derating max cap limit by 0.01. The value of 1.0 means no derating of max cap limit.  |
 | `-delay_buffer_derate` | This option balances latencies between macro cells and registers by inserting delay buffers.  The default value is `1.0`, meaning all needed delay buffers are inserted.  A value of 0.5 means only half of necessary delay buffers are inserted.  A value of 0.0 means no insertion of delay buffers. |

src/cts/include/cts/TritonCTS.h

@@ -98,12 +98,25 @@ class TritonCTS
   void buildClockTrees();
   void writeDataToDb();
 
+  // NDR functions
+  std::vector<int> getAllClockTreeLevels(Clock& clockNet);
+  int applyNDRToClockLevels(Clock& clockNet,
+                            odb::dbTechNonDefaultRule* clockNDR,
+                            const std::vector<int>& targetLevels);
+
+  int applyNDRToClockLevelRange(Clock& clockNet,
+                                odb::dbTechNonDefaultRule* clockNDR,
+                                int minLevel,
+                                int maxLevel);
+  int applyNDRToFirstHalfLevels(Clock& clockNet,
+                                odb::dbTechNonDefaultRule* clockNDR);
+
   // db functions
   bool masterExists(const std::string& master) const;
   void populateTritonCTS();
   void writeClockNetsToDb(TreeBuilder* builder,
                           std::set<odb::dbNet*>& clkLeafNets);
-  void writeClockNDRsToDb(const std::set<odb::dbNet*>& clkLeafNets);
+  void writeClockNDRsToDb(TreeBuilder* builder);
   void incrementNumClocks() { ++numberOfClocks_; }
   void clearNumClocks() { numberOfClocks_ = 0; }
   unsigned getNumClocks() const { return numberOfClocks_; }

src/cts/src/Clock.h

@@ -94,10 +94,16 @@ class ClockSubNet
   std::deque<ClockInst*> instances_;
   std::unordered_map<ClockInst*, unsigned> mapInstToIdx_;
   bool leafLevel_ = false;
+  int level_ = -1;
+  odb::dbNet* netObj_ = nullptr;
 
  public:
   explicit ClockSubNet(const std::string& name) : name_(name) {}
 
+  void setNetObj(odb::dbNet* net) { netObj_ = net; }
+  odb::dbNet* getNetObj() { return netObj_; }
+  void setTreeLevel(int level) { level_ = level; }
+  int getTreeLevel() { return level_; }
   void setLeafLevel(bool isLeaf) { leafLevel_ = isLeaf; }
   bool isLeafLevel() const { return leafLevel_; }
 

src/cts/src/CtsOptions.h

@@ -27,6 +27,14 @@ namespace cts {
 class CtsOptions : public odb::dbBlockCallBackObj
 {
  public:
+  enum class NdrStrategy
+  {
+    NONE,
+    ROOT_ONLY,
+    HALF,
+    FULL
+  };
+
   enum class MasterType
   {
     DUMMY,
@@ -222,8 +230,6 @@ class CtsOptions : public odb::dbBlockCallBackObj
   stt::SteinerTreeBuilder* getSttBuilder() const { return sttBuilder_; }
   void setObstructionAware(bool obs) { obsAware_ = obs; }
   bool getObstructionAware() const { return obsAware_; }
-  void setApplyNDR(bool ndr) { applyNDR_ = ndr; }
-  bool applyNDR() const { return applyNDR_; }
   void enableInsertionDelay(bool insDelay) { insertionDelay_ = insDelay; }
   bool insertionDelayEnabled() const { return insertionDelay_; }
   void setBufferListInferred(bool inferred) { bufferListInferred_ = inferred; }
@@ -264,6 +270,10 @@ class CtsOptions : public odb::dbBlockCallBackObj
   void setRepairClockNets(bool value) { repairClockNets_ = value; }
   bool getRepairClockNets() { return repairClockNets_; }
 
+  // NDR strategies
+  void setApplyNDR(NdrStrategy strategy) { ndrStrategy_ = strategy; }
+  NdrStrategy getApplyNdr() const { return ndrStrategy_; }
+
  private:
   std::string clockNets_;
   std::string rootBuffer_;
@@ -315,7 +325,6 @@ class CtsOptions : public odb::dbBlockCallBackObj
   utl::Logger* logger_ = nullptr;
   stt::SteinerTreeBuilder* sttBuilder_ = nullptr;
   bool obsAware_ = true;
-  bool applyNDR_ = false;
   bool insertionDelay_ = true;
   bool bufferListInferred_ = false;
   bool sinkBufferInferred_ = false;
@@ -330,6 +339,7 @@ class CtsOptions : public odb::dbBlockCallBackObj
   std::string dummyload_prefix_ = "clkload";
   MasterCount dummy_count_;
   bool repairClockNets_ = false;
+  NdrStrategy ndrStrategy_ = NdrStrategy::NONE;
 };
 
 }  // namespace cts

src/cts/src/HTreeBuilder.cpp

@@ -1849,6 +1849,7 @@ void HTreeBuilder::createClockSubNets()
 
   addTreeLevelBuffer(&rootBuffer);
   ClockSubNet& rootClockSubNet = clock_.addSubNet("clknet_0");
+  rootClockSubNet.setTreeLevel(0);
   rootClockSubNet.addInst(rootBuffer);
   treeBufLevels_++;
 
@@ -1950,6 +1951,11 @@ void HTreeBuilder::createClockSubNets()
                              wireSegmentUnit_,
                              this);
 
+      // Set clock tree level the first time only.
+      if (builder.getDrivingSubNet()->getTreeLevel() < 0) {
+        builder.getDrivingSubNet()->setTreeLevel(levelIdx);
+      }
+
       if (!options_->getTreeBuffer().empty()) {
         builder.build(options_->getTreeBuffer());
       } else {
@@ -2017,6 +2023,7 @@ void HTreeBuilder::createSingleBufferClockNet()
 
   addTreeLevelBuffer(&rootBuffer);
   ClockSubNet& clockSubNet = clock_.addSubNet("clknet_0");
+  clockSubNet.setTreeLevel(0);
   clockSubNet.addInst(rootBuffer);
 
   clock_.forEachSink([&](ClockInst& inst) { clockSubNet.addInst(inst); });