Compiler profiling

CMakeLists.txt

@@ -232,6 +232,7 @@ 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

@@ -557,6 +557,7 @@ SOURCE_FILES = \
   Pipeline.cpp \
   Prefetch.cpp \
   PrintLoopNest.cpp \
+  ProfiledIRVisitor.cpp \
   Profiling.cpp \
   PurifyIndexMath.cpp \
   PythonExtensionGen.cpp \

src/AddImageChecks.cpp

@@ -1,4 +1,5 @@
 #include "AddImageChecks.h"
+#include "CompilerProfiling.h"
 #include "ExternFuncArgument.h"
 #include "Function.h"
 #include "IRMutator.h"
@@ -103,6 +104,7 @@ class TrimStmtToPartsThatAccessBuffers : public IRMutator {
     bool touches_buffer = false;
     const map<string, FindBuffers::Result> &buffers;
 
+protected:
     using IRMutator::visit;
 
     Expr visit(const Call *op) override {
@@ -165,7 +167,7 @@ Stmt add_image_checks_inner(Stmt s,
     bool no_bounds_query = t.has_feature(Target::NoBoundsQuery);
 
     // First hunt for all the referenced buffers
-    FindBuffers finder;
+    Profiled<FindBuffers> finder;
     map<string, FindBuffers::Result> &bufs = finder.buffers;
 
     // Add the output buffer(s).
@@ -188,7 +190,7 @@ Stmt add_image_checks_inner(Stmt s,
     s.accept(&finder);
 
     Scope<Interval> empty_scope;
-    Stmt sub_stmt = TrimStmtToPartsThatAccessBuffers(bufs).mutate(s);
+    Stmt sub_stmt = Profiled<TrimStmtToPartsThatAccessBuffers>(bufs).mutate(s);
     map<string, Box> boxes = boxes_touched(sub_stmt, empty_scope, fb);
 
     // Now iterate through all the buffers, creating a list of lets
@@ -737,6 +739,7 @@ Stmt add_image_checks(const Stmt &s,
     // Checks for images go at the marker deposited by computation
     // bounds inference.
     class Injector : public IRMutator {
+    protected:
         using IRMutator::visit;
 
         Expr visit(const Variable *op) override {
@@ -794,7 +797,8 @@ Stmt add_image_checks(const Stmt &s,
                  bool will_inject_host_copies)
             : outputs(outputs), t(t), order(order), env(env), fb(fb), will_inject_host_copies(will_inject_host_copies) {
         }
-    } injector(outputs, t, order, env, fb, will_inject_host_copies);
+    };
+    Profiled<Injector> injector(outputs, t, order, env, fb, will_inject_host_copies);
 
     return injector.mutate(s);
 }

src/AddParameterChecks.cpp

@@ -3,6 +3,7 @@
 #include "IRVisitor.h"
 #include "Substitute.h"
 #include "Target.h"
+#include "CompilerProfiling.h"
 
 namespace Halide {
 namespace Internal {
@@ -35,7 +36,7 @@ class FindParameters : public IRGraphVisitor {
 Stmt add_parameter_checks(const vector<Stmt> &preconditions, Stmt s, const Target &t) {
 
     // First, find all the parameters
-    FindParameters finder;
+    Profiled<FindParameters> finder;
     s.accept(&finder);
 
     map<string, Expr> replace_with_constrained;

src/AsyncProducers.cpp

@@ -116,6 +116,7 @@ class NoOpCollapsingMutator : public IRMutator {
 };
 
 class GenerateProducerBody : public NoOpCollapsingMutator {
+protected:
     const string &func;
     vector<Expr> sema;
     std::set<string> producers_dropped;
@@ -285,6 +286,7 @@ class GenerateProducerBody : public NoOpCollapsingMutator {
 };
 
 class GenerateConsumerBody : public NoOpCollapsingMutator {
+protected:
     const string &func;
     vector<Expr> sema;
 
@@ -342,6 +344,7 @@ class GenerateConsumerBody : public NoOpCollapsingMutator {
 };
 
 class CloneAcquire : public IRMutator {
+protected:
     using IRMutator::visit;
 
     const string &old_name;
@@ -390,6 +393,7 @@ class CountConsumeNodes : public IRVisitor {
 };
 
 class ForkAsyncProducers : public IRMutator {
+protected:
     using IRMutator::visit;
 
     const map<string, Function> &env;
@@ -493,6 +497,7 @@ class ForkAsyncProducers : public IRMutator {
 // simple failure case, error_async_require_fail. One has not been
 // written for the complex nested case yet.)
 class InitializeSemaphores : public IRMutator {
+protected:
     using IRMutator::visit;
 
     const Type sema_type = type_of<halide_semaphore_t *>();
@@ -558,6 +563,7 @@ class InitializeSemaphores : public IRMutator {
 // A class to support stmt_uses_vars queries that repeatedly hit the same
 // sub-stmts. Used to support TightenProducerConsumerNodes below.
 class CachingStmtUsesVars : public IRMutator {
+protected:
     const Scope<> &query;
     bool found_use = false;
     std::map<Stmt, bool> cache;
@@ -613,6 +619,7 @@ class CachingStmtUsesVars : public IRMutator {
 
 // Tighten the scope of consume nodes as much as possible to avoid needless synchronization.
 class TightenProducerConsumerNodes : public IRMutator {
+protected:
     using IRMutator::visit;
 
     Stmt make_producer_consumer(const string &name, bool is_producer, Stmt body, const Scope<> &scope, CachingStmtUsesVars &uses_vars) {
@@ -703,6 +710,7 @@ class TightenProducerConsumerNodes : public IRMutator {
 
 // Update indices to add ring buffer.
 class UpdateIndices : public IRMutator {
+protected:
     using IRMutator::visit;
 
     Stmt visit(const Provide *op) override {
@@ -734,6 +742,7 @@ class UpdateIndices : public IRMutator {
 
 // Inject ring buffering.
 class InjectRingBuffering : public IRMutator {
+protected:
     using IRMutator::visit;
 
     struct Loop {
@@ -816,6 +825,7 @@ class InjectRingBuffering : public IRMutator {
 // Broaden the scope of acquire nodes to pack trailing work into the
 // same task and to potentially reduce the nesting depth of tasks.
 class ExpandAcquireNodes : public IRMutator {
+protected:
     using IRMutator::visit;
 
     Stmt visit(const Block *op) override {
@@ -918,6 +928,7 @@ class ExpandAcquireNodes : public IRMutator {
 };
 
 class TightenForkNodes : public IRMutator {
+protected:
     using IRMutator::visit;
 
     Stmt make_fork(const Stmt &first, const Stmt &rest) {
@@ -1005,12 +1016,13 @@ class TightenForkNodes : public IRMutator {
 }  // namespace
 
 Stmt fork_async_producers(Stmt s, const map<string, Function> &env) {
-    s = TightenProducerConsumerNodes(env).mutate(s);
-    s = InjectRingBuffering(env).mutate(s);
-    s = ForkAsyncProducers(env).mutate(s);
-    s = ExpandAcquireNodes().mutate(s);
-    s = TightenForkNodes().mutate(s);
-    s = InitializeSemaphores().mutate(s);
+    ZoneScoped;
+    s = Profiled<TightenProducerConsumerNodes>(env).profiled_mutate(s);
+    s = Profiled<InjectRingBuffering>(env).profiled_mutate(s);
+    s = Profiled<ForkAsyncProducers>(env).profiled_mutate(s);
+    s = Profiled<ExpandAcquireNodes>().profiled_mutate(s);
+    s = Profiled<TightenForkNodes>().profiled_mutate(s);
+    s = Profiled<InitializeSemaphores>().profiled_mutate(s);
     return s;
 }
 

src/BoundSmallAllocations.cpp

@@ -156,6 +156,7 @@ class BoundSmallAllocations : public IRMutator {
 }  // namespace
 
 Stmt bound_small_allocations(const Stmt &s) {
+    ZoneScoped;
     return BoundSmallAllocations().mutate(s);
 }
 

src/CMakeLists.txt

@@ -86,6 +86,7 @@ target_sources(
     CodeGen_Vulkan_Dev.h
     CodeGen_WebGPU_Dev.h
     CompilerLogger.h
+    CompilerProfiling.h
     ConciseCasts.h
     CPlusPlusMangle.h
     ConstantBounds.h
@@ -166,6 +167,7 @@ target_sources(
     Param.h
     Parameter.h
     PartitionLoops.h
+    PerformanceCounter.h
     Pipeline.h
     Prefetch.h
     PrefetchDirective.h
@@ -269,6 +271,7 @@ target_sources(
     CodeGen_WebGPU_Dev.cpp
     CodeGen_X86.cpp
     CompilerLogger.cpp
+    CompilerProfiling.cpp
     CPlusPlusMangle.cpp
     ConstantBounds.cpp
     ConstantInterval.cpp
@@ -515,6 +518,11 @@ target_compile_definitions(Halide PRIVATE WITH_SPIRV)
 target_compile_definitions(Halide PRIVATE WITH_VULKAN)
 target_compile_definitions(Halide PRIVATE WITH_WEBGPU)
 
+if (WITH_COMPILER_PROFILING)
+    target_compile_definitions(Halide PRIVATE WITH_COMPILER_PROFILING)
+endif()
+
+
 ##
 # Flatbuffers and Serialization dependencies.
 ##

src/CSE.cpp

@@ -1,6 +1,7 @@
 #include <map>
 
 #include "CSE.h"
+#include "CompilerProfiling.h"
 #include "IREquality.h"
 #include "IRMutator.h"
 #include "IROperator.h"
@@ -186,6 +187,7 @@ class Replacer : public IRGraphMutator {
 };
 
 class RemoveLets : public IRGraphMutator {
+protected:
     using IRGraphMutator::visit;
 
     Scope<Expr> scope;
@@ -218,6 +220,7 @@ class RemoveLets : public IRGraphMutator {
 };
 
 class CSEEveryExprInStmt : public IRMutator {
+protected:
     bool lift_all;
     using IRMutator::visit;
 
@@ -260,6 +263,7 @@ 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.
@@ -269,14 +273,15 @@ Expr common_subexpression_elimination(const Expr &e_in, bool lift_all) {
 
     debug(4) << "\n\n\nInput to CSE " << e << "\n";
 
-    e = RemoveLets().mutate(e);
+    e = Profiled<RemoveLets>().profiled_mutate(e);
 
     debug(4) << "After removing lets: " << e << "\n";
 
     // CSE is run on unsanitized Exprs from the user, and may contain Vars with
     // the same name as the temporaries we intend to introduce. Find any such
     // Vars so that we know not to use those names.
     class UniqueNameProvider : public IRGraphVisitor {
+    protected:
         using IRGraphVisitor::visit;
 
         const char prefix = 't';  // Annoyingly, this can't be static because this is a local class.
@@ -303,14 +308,18 @@ Expr common_subexpression_elimination(const Expr &e_in, bool lift_all) {
             } while (vars.count(name));
             return name;
         }
-    } namer;
-    e.accept(&namer);
+    };
+    Profiled<UniqueNameProvider> namer;
+    {
+        ZoneScopedN("UniqueNameProvider");
+        e.accept(&namer);
+    }
 
-    GVN gvn;
-    e = gvn.mutate(e);
+    Profiled<GVN> gvn;
+    e = gvn.profiled_mutate(e);
 
-    ComputeUseCounts count_uses(gvn, lift_all);
-    count_uses.include(e);
+    Profiled<ComputeUseCounts> count_uses(gvn, lift_all);
+    count_uses.profiled_include(e);
 
     debug(4) << "Canonical form without lets " << e << "\n";
 
@@ -330,8 +339,8 @@ Expr common_subexpression_elimination(const Expr &e_in, bool lift_all) {
     }
 
     // Rebuild the expr to include references to the variables:
-    Replacer replacer(replacements);
-    e = replacer.mutate(e);
+    Profiled<Replacer> replacer(replacements);
+    e = replacer.profiled_mutate(e);
 
     debug(4) << "With variables " << e << "\n";
 
@@ -340,7 +349,7 @@ Expr common_subexpression_elimination(const Expr &e_in, bool lift_all) {
         // Drop this variable as an acceptable replacement for this expr.
         replacer.erase(value);
         // Use containing lets in the value.
-        e = Let::make(var, replacer.mutate(value), e);
+        e = Let::make(var, replacer.profiled_mutate(value), e);
     }
 
     debug(4) << "With lets: " << e << "\n";
@@ -349,7 +358,8 @@ Expr common_subexpression_elimination(const Expr &e_in, bool lift_all) {
 }
 
 Stmt common_subexpression_elimination(const Stmt &s, bool lift_all) {
-    return CSEEveryExprInStmt(lift_all).mutate(s);
+    ZoneScoped;
+    return Profiled<CSEEveryExprInStmt>(lift_all).profiled_mutate(s);
 }
 
 // Testing code.

src/CodeGen_ARM.cpp

@@ -2551,7 +2551,7 @@ bool CodeGen_ARM::supports_call_as_float16(const Call *op) const {
 }  // namespace
 
 std::unique_ptr<CodeGen_Posix> new_CodeGen_ARM(const Target &target) {
-    return std::make_unique<CodeGen_ARM>(target);
+    return std::make_unique<Profiled<CodeGen_ARM>>(target);
 }
 
 #else  // WITH_ARM || WITH_AARCH64

src/CodeGen_C.cpp

@@ -189,6 +189,7 @@ 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_Hexagon.cpp

@@ -2295,7 +2295,7 @@ void CodeGen_Hexagon::visit(const Allocate *alloc) {
 }  // namespace
 
 std::unique_ptr<CodeGen_Posix> new_CodeGen_Hexagon(const Target &target) {
-    return std::make_unique<CodeGen_Hexagon>(target);
+    return std::make_unique<Profiled<CodeGen_Hexagon>>(target);
 }
 
 #else  // WITH_HEXAGON