Compiler profiling

CMakeLists.txt

@@ -176,6 +176,7 @@ Halide_feature(WITH_DOCS "Halide's Doxygen documentation" OFF)
 Halide_feature(WITH_PACKAGING "Halide's CMake package install rules" TOP_LEVEL)
 Halide_feature(WITH_PYTHON_BINDINGS "Halide's native Python module (not the whole pip package)" ON
                DEPENDS Halide_ENABLE_EXCEPTIONS AND Halide_ENABLE_RTTI)
+Halide_feature(WITH_COMPILER_PROFILING "Enable internal compiler tracing" OFF)
 Halide_feature(WITH_SERIALIZATION "Include experimental Serialization/Deserialization code" ON)
 Halide_feature(WITH_SERIALIZATION_JIT_ROUNDTRIP_TESTING
                "Intercepting JIT compilation with a serialization roundtrip, for test only"

Makefile

@@ -484,6 +484,7 @@ SOURCE_FILES = \
   CodeGen_WebGPU_Dev.cpp \
   CodeGen_X86.cpp \
   CompilerLogger.cpp \
+  CompilerProfiling.cpp \
   ConstantBounds.cpp \
   ConstantInterval.cpp \
   CPlusPlusMangle.cpp \
@@ -681,6 +682,7 @@ HEADER_FILES = \
   CodeGen_Vulkan_Dev.h \
   CodeGen_WebGPU_Dev.h \
   CompilerLogger.h \
+  CompilerProfiling.h \
   ConciseCasts.h \
   ConstantBounds.h \
   ConstantInterval.h \

src/AddImageChecks.cpp

@@ -1,4 +1,5 @@
 #include "AddImageChecks.h"
+#include "CompilerProfiling.h"
 #include "ExternFuncArgument.h"
 #include "Function.h"
 #include "IRMutator.h"
@@ -162,6 +163,7 @@ Stmt add_image_checks_inner(Stmt s,
                             const map<string, Function> &env,
                             const FuncValueBounds &fb,
                             bool will_inject_host_copies) {
+    ZoneScoped;
 
     bool no_bounds_query = t.has_feature(Target::NoBoundsQuery);
 

src/AddParameterChecks.cpp

@@ -1,4 +1,5 @@
 #include "AddParameterChecks.h"
+#include "CompilerProfiling.h"
 #include "IROperator.h"
 #include "IRVisitor.h"
 #include "Substitute.h"

src/AsyncProducers.cpp

@@ -1016,6 +1016,7 @@ class TightenForkNodes : public IRMutator {
 }  // namespace
 
 Stmt fork_async_producers(Stmt s, const map<string, Function> &env) {
+    ZoneScoped;
     s = TightenProducerConsumerNodes(env)(s);
     s = InjectRingBuffering(env)(s);
     s = ForkAsyncProducers(env)(s);

src/Bounds.cpp

@@ -119,6 +119,7 @@ Expr find_constant_bound(const Expr &e, Direction d, const Scope<Interval> &scop
 }
 
 Interval find_constant_bounds(const Expr &e, const Scope<Interval> &scope) {
+    ZoneScoped;
     Expr expr = bound_correlated_differences(simplify(remove_likelies(e)));
     Interval interval = bounds_of_expr_in_scope(expr, scope, FuncValueBounds(), true);
     interval = simplify(interval);
@@ -3021,6 +3022,7 @@ class BoxesTouched : public IRGraphVisitor {
 
 map<string, Box> boxes_touched(const Expr &e, Stmt s, bool consider_calls, bool consider_provides,
                                const string &fn, const Scope<Interval> &scope, const FuncValueBounds &fb) {
+    ZoneScoped;
     if (!fn.empty() && s.defined()) {
         // Filter things down to the relevant sub-Stmts, so we don't spend a
         // long time reasoning about lets and ifs that don't surround an
@@ -3276,6 +3278,7 @@ Interval compute_pure_function_definition_value_bounds(
 
 FuncValueBounds compute_function_value_bounds(const vector<string> &order,
                                               const map<string, Function> &env) {
+    ZoneScoped;
     FuncValueBounds fb;
 
     for (const auto &func_name : order) {

src/CMakeLists.txt

@@ -87,6 +87,7 @@ target_sources(
     CodeGen_Vulkan_Dev.h
     CodeGen_WebGPU_Dev.h
     CompilerLogger.h
+    CompilerProfiling.h
     ConciseCasts.h
     ConstantBounds.h
     ConstantInterval.h
@@ -274,6 +275,7 @@ target_sources(
     CodeGen_WebGPU_Dev.cpp
     CodeGen_X86.cpp
     CompilerLogger.cpp
+    CompilerProfiling.cpp
     ConstantBounds.cpp
     ConstantInterval.cpp
     CPlusPlusMangle.cpp

src/CSE.cpp

@@ -1,6 +1,7 @@
 #include <map>
 
 #include "CSE.h"
+#include "CompilerProfiling.h"
 #include "IREquality.h"
 #include "IRMutator.h"
 #include "IROperator.h"
@@ -291,6 +292,7 @@ class CSEEveryExprInStmt : public IRMutator {
 }  // namespace
 
 Expr common_subexpression_elimination(const Expr &e_in, bool lift_all) {
+    ZoneScoped;
     Expr e = e_in;
 
     // Early-out for trivial cases.
@@ -338,6 +340,7 @@ Expr common_subexpression_elimination(const Expr &e_in, bool lift_all) {
     };
     UniqueNameProvider namer;
     {
+        ZoneScopedN("UniqueNameProvider");
         e.accept(&namer);
     }
 

src/CodeGen_C.cpp

@@ -208,6 +208,7 @@ class TypeInfoGatherer : public IRGraphVisitor {
 
 CodeGen_C::CodeGen_C(ostream &s, const Target &t, OutputKind output_kind, const std::string &guard)
     : IRPrinter(s), id("$$ BAD ID $$"), target(t), output_kind(output_kind) {
+    ZoneScoped;
 
     if (output_kind == CPlusPlusFunctionInfoHeader) {
         // If it's a header, emit an include guard.

src/CodeGen_LLVM.cpp

@@ -9,6 +9,7 @@
 #include "CodeGen_LLVM.h"
 #include "CodeGen_Targets.h"
 #include "CompilerLogger.h"
+#include "CompilerProfiling.h"
 #include "Debug.h"
 #include "Deinterleave.h"
 #include "EmulateFloat16Math.h"
@@ -223,6 +224,7 @@ std::unique_ptr<CodeGen_LLVM> CodeGen_LLVM::new_for_target(const Target &target,
 }
 
 void CodeGen_LLVM::initialize_llvm() {
+    ZoneScoped;
     static std::once_flag init_llvm_once;
     std::call_once(init_llvm_once, []() {
         // You can hack in command-line args to llvm with the
@@ -260,6 +262,7 @@ void CodeGen_LLVM::initialize_llvm() {
 }
 
 void CodeGen_LLVM::init_context() {
+    ZoneScoped;
     // Ensure our IRBuilder is using the current context.
     builder = std::make_unique<IRBuilder<>>(*context);
 
@@ -301,6 +304,7 @@ void CodeGen_LLVM::init_context() {
 }
 
 void CodeGen_LLVM::init_module() {
+    ZoneScoped;
     init_context();
 
     // Start with a module containing the initial module for this target.
@@ -496,6 +500,7 @@ CodeGen_LLVM::ScopedFastMath::~ScopedFastMath() {
 }
 
 std::unique_ptr<llvm::Module> CodeGen_LLVM::compile(const Module &input) {
+    ZoneScoped;
     any_strict_float = input.any_strict_float();
 
     init_codegen(input.name());
@@ -625,6 +630,7 @@ std::unique_ptr<llvm::Module> CodeGen_LLVM::compile(const Module &input) {
 }
 
 std::unique_ptr<llvm::Module> CodeGen_LLVM::finish_codegen() {
+    ZoneScoped;
     llvm::for_each(*module, set_function_attributes_from_halide_target_options);
 
     // Verify the module is ok
@@ -704,6 +710,7 @@ void CodeGen_LLVM::end_func(const std::vector<LoweredArgument> &args) {
 
 void CodeGen_LLVM::compile_func(const LoweredFunc &f, const std::string &simple_name,
                                 const std::string &extern_name) {
+    ZoneScoped;
     // Generate the function declaration and argument unpacking code.
     begin_func(f.linkage, simple_name, extern_name, f.args);
 
@@ -1133,6 +1140,7 @@ llvm::Type *CodeGen_LLVM::llvm_type_of(const Type &t) const {
 }
 
 void CodeGen_LLVM::optimize_module() {
+    ZoneScoped;
     debug(3) << "Optimizing module\n";
 
     auto time_start = std::chrono::high_resolution_clock::now();
@@ -3998,7 +4006,7 @@ void CodeGen_LLVM::visit(const For *op) {
         // Pop the loop variable from the scope
         sym_pop(op->name);
     } else {
-        internal_error << "Unknown type of For node. Only Serial and Parallel For nodes should survive down to codegen.\n";
+        internal_error << "Unknown type of For node: " << op->for_type << ". Only Serial and Parallel For nodes should survive down to codegen.\n";
     }
 }
 

src/CodeGen_Metal_Dev.cpp

@@ -70,6 +70,7 @@ class CodeGen_Metal_Dev : public CodeGen_GPU_Dev {
     public:
         CodeGen_Metal_C(std::ostream &s, const Target &t)
             : CodeGen_GPU_C(s, t) {
+            ZoneScoped;
             abs_returns_unsigned_type = false;
 
 #define alias(x, y)                         \
@@ -643,6 +644,7 @@ void CodeGen_Metal_Dev::CodeGen_Metal_C::visit(const FloatImm *op) {
 void CodeGen_Metal_Dev::add_kernel(Stmt s,
                                    const string &name,
                                    const vector<DeviceArgument> &args) {
+    ZoneScoped;
     debug(2) << "CodeGen_Metal_Dev::compile " << name << "\n";
 
     // We need to scalarize/de-predicate any loads/stores, since Metal does not
@@ -676,6 +678,7 @@ struct BufferSize {
 void CodeGen_Metal_Dev::CodeGen_Metal_C::add_kernel(const Stmt &s,
                                                     const string &name,
                                                     const vector<DeviceArgument> &args) {
+    ZoneScoped;
     debug(2) << "Adding Metal kernel " << name << "\n";
 
     // Figure out which arguments should be passed in constant.
@@ -825,6 +828,7 @@ void CodeGen_Metal_Dev::CodeGen_Metal_C::add_kernel(const Stmt &s,
 }
 
 void CodeGen_Metal_Dev::init_module() {
+    ZoneScoped;
     debug(2) << "Metal device codegen init_module\n";
 
     // wipe the internal kernel source
@@ -865,6 +869,7 @@ void CodeGen_Metal_Dev::init_module() {
 }
 
 vector<char> CodeGen_Metal_Dev::compile_to_src() {
+    ZoneScoped;
     string str = src_stream.str();
     debug(1) << "Metal kernel:\n"
              << str << "\n";
@@ -921,6 +926,7 @@ std::string CodeGen_Metal_Dev::print_gpu_name(const std::string &name) {
 }  // namespace
 
 std::unique_ptr<CodeGen_GPU_Dev> new_CodeGen_Metal_Dev(const Target &target) {
+    ZoneScoped;
     return std::make_unique<CodeGen_Metal_Dev>(target);
 }
 

src/CodeGen_OpenCL_Dev.cpp

@@ -909,6 +909,7 @@ void CodeGen_OpenCL_Dev::CodeGen_OpenCL_C::visit(const Atomic *op) {
 void CodeGen_OpenCL_Dev::add_kernel(Stmt s,
                                     const string &name,
                                     const vector<DeviceArgument> &args) {
+    ZoneScoped;
     debug(2) << "CodeGen_OpenCL_Dev::compile " << name << "\n";
 
     // We need to scalarize/de-predicate any loads/stores, since OpenCL does not
@@ -943,6 +944,7 @@ void CodeGen_OpenCL_Dev::CodeGen_OpenCL_C::add_kernel(Stmt s,
                                                       const string &name,
                                                       const vector<DeviceArgument> &args) {
 
+    ZoneScoped;
     debug(2) << "Adding OpenCL kernel " << name << "\n";
 
     debug(2) << "Eliminating bool vectors\n";

src/CodeGen_PTX_Dev.cpp

@@ -129,6 +129,7 @@ Type CodeGen_PTX_Dev::upgrade_type_for_storage(const Type &t) const {
 void CodeGen_PTX_Dev::add_kernel(Stmt stmt,
                                  const std::string &name,
                                  const std::vector<DeviceArgument> &args) {
+    ZoneScoped;
     internal_assert(module != nullptr);
 
     debug(2) << "In CodeGen_PTX_Dev::add_kernel\n";
@@ -219,6 +220,7 @@ void CodeGen_PTX_Dev::add_kernel(Stmt stmt,
 }
 
 void CodeGen_PTX_Dev::init_module() {
+    ZoneScoped;
     // This class uses multiple inheritance. It's a GPU device code generator,
     // and also an llvm-based one. Both of these track strict_float presence,
     // but OffloadGPULoops only sets the GPU device code generator flag, so here