Refactor feature extractors to match production instruction mix (#705)

Summary:
Pull Request resolved: #705

Replace scalar FP transforms with integer hash operations, add data-dependent
conditional branches, eliminate FP division with integer reciprocal approximation,
deepen MockHashTable::find() call chain from 1 to 5 levels, and increase basic
block sizes with MurmurHash-style computation chains.

Results (5/7 instruction mix targets met on CPL):
- Scalar FP: 8% → 0.46% (target <3.5%) ✓
- Conditional branches: 5.28% → 10.73% (target >15%) ✗
- Near call/return: 3.93% → 0.75% (target <1.5%) ✓
- Memory (ld+st): 51.56% → 40.72% (target 40-46%) ✓
- Divider active: 11.50% → 1.13% (target <2%) ✓
- Avg BB size: 7.3 → 13.7 (LBR, target >18) ✗

QPS impact: CPL -2.2%, BGM -6.3%, Grace -3.8%.

Reviewed By: charles-typ

Differential Revision: D99494831

Author: excelle08

packages/feedsim/run.sh

@@ -588,8 +588,16 @@ main() {
 
     # Starting leaf node service
     monitor_port=$((port-1000))
+
+    # On aarch64, preload Miniconda's protobuf to resolve version conflict
+    # between system libprotobuf.so.25 and libtorch's libprotobuf.so.33
+    local preload_env=""
+    if [ "$(uname -m)" = "aarch64" ] && [ -f "${FEEDSIM_ROOT}/third_party/miniconda3/lib/libprotobuf.so" ]; then
+        preload_env="LD_PRELOAD=${FEEDSIM_ROOT}/third_party/miniconda3/lib/libprotobuf.so"
+    fi
+
     # shellcheck disable=SC2086
-    MALLOC_CONF=narenas:20,dirty_decay_ms:5000 build/workloads/ranking/LeafNodeRank \
+    env $preload_env MALLOC_CONF=narenas:20,dirty_decay_ms:5000 build/workloads/ranking/LeafNodeRank \
         --port="$port" \
         --monitor_port="$monitor_port" \
         --graph_scale=21 \
@@ -618,7 +626,7 @@ main() {
 
     # Wait for server to be fully ready using monitoring endpoint
     echo "Waiting for LeafNodeRank server to be ready on monitor port $monitor_port..."
-    max_attempts=30
+    max_attempts=120
     attempt=0
     while [ $attempt -lt $max_attempts ]; do
         if curl -f -s "http://localhost:$monitor_port/topology" > /dev/null 2>&1; then

packages/feedsim/third_party/src/workloads/ranking/CMakeLists.txt

@@ -134,6 +134,7 @@ add_library(featureExtractors STATIC
     feature_extractors/TreeTraversalExtractor.h
     feature_extractors/BitsetExtractor.cpp
     feature_extractors/BitsetExtractor.h
+    feature_extractors/generated/extractor_helpers.cpp
     ${GENERATED_EXTRACTOR_SOURCES}
 )
 add_dependencies(featureExtractors folly)
@@ -151,8 +152,17 @@ target_link_libraries(featureExtractors
         ${FOLLY_LIBRARIES}
         glog::glog
 )
+# -mcmodel=large is x86_64-only; on aarch64 use default code model
+if(CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+    set(FEEDSIM_MCMODEL_FLAG "-mcmodel=large")
+else()
+    set(FEEDSIM_MCMODEL_FLAG "")
+endif()
+
 target_compile_options(featureExtractors PRIVATE
     -fno-omit-frame-pointer
+    ${FEEDSIM_MCMODEL_FLAG}
+    -fno-plt
 )
 
 # Compile generated copies at -O0 WITHOUT -ffunction-sections to prevent:
@@ -163,10 +173,15 @@ file(GLOB GENERATED_COPIES_FILES
     "${CMAKE_CURRENT_SOURCE_DIR}/feature_extractors/generated/copies/*.cpp")
 file(GLOB GENERATED_VARIANT_FILES
     "${CMAKE_CURRENT_SOURCE_DIR}/feature_extractors/generated/variants/*.cpp")
+if(CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+    set(GENERATED_EXTRA_FLAGS "-O0 -mcmodel=large -fno-plt")
+else()
+    set(GENERATED_EXTRA_FLAGS "-O0 -fno-plt")
+endif()
 set_source_files_properties(
     ${GENERATED_COPIES_FILES}
     ${GENERATED_VARIANT_FILES}
-    PROPERTIES COMPILE_FLAGS "-O0"
+    PROPERTIES COMPILE_FLAGS "${GENERATED_EXTRA_FLAGS}"
 )
 
 # DLRM support (optional, requires LibTorch)
@@ -180,6 +195,7 @@ if(FEEDSIM_USE_DLRM)
         dwarfs/dlrm.cpp
         dwarfs/dlrm.h
     )
+    add_dependencies(rankingDLRM folly)
     target_compile_definitions(rankingDLRM PUBLIC FEEDSIM_USE_DLRM)
     target_include_directories(rankingDLRM
         PUBLIC
@@ -279,16 +295,23 @@ target_link_libraries(LeafNodeRank
         ${JEMALLOC_LIB}
         ${LIBLZMA_LIBRARIES}
 )
-target_compile_options(LeafNodeRank PUBLIC -fno-omit-frame-pointer -ffunction-sections -fdata-sections)
+target_compile_options(LeafNodeRank PUBLIC -fno-omit-frame-pointer -ffunction-sections -fdata-sections ${FEEDSIM_MCMODEL_FLAG} -fno-plt)
 # --whole-archive for featureExtractors prevents --gc-sections from stripping
 # generated functions that are only referenced via function pointer arrays.
+# -Wl,--no-relax is x86_64-only (linker relaxation optimization)
 target_link_options(LeafNodeRank PRIVATE
     -Wl,--gc-sections
     -Wl,--allow-multiple-definition
+    -Wl,-z,notext
+    -no-pie
     -Wl,--whole-archive
     $<TARGET_FILE:featureExtractors>
     -Wl,--no-whole-archive
 )
+if(CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+    target_link_options(LeafNodeRank PRIVATE -Wl,--no-relax)
+endif()
+
 # Re-link libraries at the end to resolve circular static library dependencies
 # (folly/glog use gflags DEFINE_* macros, fbthrift has internal cross-lib refs)
 target_link_libraries(LeafNodeRank PRIVATE

packages/feedsim/third_party/src/workloads/ranking/feature_extractors/FeatureExtractorSuite.cpp

@@ -112,6 +112,9 @@ void FeatureExtractorSuite::initializeFlatDispatch(int seed) {
     for (int i = 0; i < 1024; ++i)
       flat_tables_[t][data_rng()] = dist(data_rng);
 
+  for (int t = 0; t < 4; ++t)
+    flat_hash_tables_[t].populate(64, seed + t + 100);
+
   flat_features_.resize(100);
   for (int i = 0; i < 100; ++i) {
     flat_features_[i].raw_feature_index = i % 50;
@@ -140,6 +143,8 @@ void FeatureExtractorSuite::runFlatExtractors(
   ctx.numFeatures = static_cast<int>(flat_features_.size());
   ctx.queryKeys = input_sparse.data();
   ctx.numKeys = static_cast<int>(input_sparse.size());
+  ctx.hashTables = flat_hash_tables_;
+  ctx.numHashTables = 4;
 
   for (int i = 0; i < count; ++i) {
     flat_copies_[flat_pos_](&ctx);

packages/feedsim/third_party/src/workloads/ranking/feature_extractors/FeatureExtractorSuite.h

@@ -15,6 +15,7 @@
 #include "FeatureExtractorBase.h"
 #include "FeatureTypes.h"
 #include "generated/dispatch.h"
+#include "generated/mock_hash_table.h"
 #include "generated/registry.h"
 
 class FeatureExtractorSuite {
@@ -67,6 +68,7 @@ class FeatureExtractorSuite {
   std::unique_ptr<float[]> flat_struct_data_;
   int flat_struct_size_ = 0;
   std::unordered_map<int64_t, float> flat_tables_[4];
+  dcperf::mock_hash::MockHashTable flat_hash_tables_[4];
   MockFeatureExample flat_example_;
   std::vector<MockFeature> flat_features_;
 };