[refactor] [ir] IR system refactorings (taichi-dev#1058)

* [refactor] [ir] IR system refactorings part 1 * format * fix segfault * refactor typecheck signature and misc * fix * fix build * use utility function as * fix init * fix segfault to pass CI * format again * retrigger CI * modify test with fake kernel * format * remove CompileConfig from some transform passes * Update taichi/transforms/alg_simp.cpp Co-authored-by: xumingkuan <[email protected]> Co-authored-by: xumingkuan <[email protected]>
firedtoad · May 26, 2020 · 2bf0bee · 2bf0bee
1 parent 387c823
commit 2bf0bee
Show file tree

Hide file tree

Showing 19 changed files with 112 additions and 58 deletions.
diff --git a/taichi/analysis/clone.cpp b/taichi/analysis/clone.cpp
@@ -112,15 +112,15 @@ class IRCloner : public IRVisitor {
 
   static std::unique_ptr<IRNode> run(IRNode *root, Kernel *kernel) {
     if (kernel == nullptr) {
-      kernel = &get_current_program().get_current_kernel();
+      kernel = root->get_kernel();
     }
     std::unique_ptr<IRNode> new_root = root->clone();
     IRCloner cloner(new_root.get());
     cloner.phase = IRCloner::register_operand_map;
     root->accept(&cloner);
     cloner.phase = IRCloner::replace_operand;
     root->accept(&cloner);
-    irpass::typecheck(new_root.get(), kernel);
+    irpass::typecheck(new_root.get());
     irpass::fix_block_parents(new_root.get());
     return new_root;
   }

diff --git a/taichi/backends/metal/codegen_metal.cpp b/taichi/backends/metal/codegen_metal.cpp
@@ -1031,7 +1031,7 @@ CodeGen::CodeGen(Kernel *kernel,
 FunctionType CodeGen::compile() {
   auto &config = kernel_->program.config;
   config.demote_dense_struct_fors = true;
-  irpass::compile_to_offloads(kernel_->ir, config,
+  irpass::compile_to_offloads(kernel_->ir.get(), config,
                               /*vectorize=*/false, kernel_->grad,
                               /*ad_use_stack=*/false, config.print_ir);
 

diff --git a/taichi/backends/opengl/codegen_opengl.cpp b/taichi/backends/opengl/codegen_opengl.cpp
@@ -702,7 +702,7 @@ FunctionType OpenglCodeGen::gen(void) {
 }
 
 void OpenglCodeGen::lower() {
-  auto ir = kernel_->ir;
+  auto ir = kernel_->ir.get();
   auto &config = kernel_->program.config;
   config.demote_dense_struct_fors = true;
   irpass::compile_to_offloads(ir, config,

diff --git a/taichi/codegen/codegen.cpp b/taichi/codegen/codegen.cpp
@@ -15,7 +15,7 @@ TLANG_NAMESPACE_BEGIN
 KernelCodeGen::KernelCodeGen(Kernel *kernel, IRNode *ir)
     : prog(&kernel->program), kernel(kernel), ir(ir) {
   if (ir == nullptr)
-    this->ir = kernel->ir;
+    this->ir = kernel->ir.get();
 
   auto num_stmts = irpass::analysis::count_statements(this->ir);
   if (kernel->is_evaluator)

diff --git a/taichi/codegen/codegen_llvm.cpp b/taichi/codegen/codegen_llvm.cpp
@@ -272,7 +272,7 @@ CodeGenLLVM::CodeGenLLVM(Kernel *kernel, IRNode *ir)
       ir(ir),
       prog(&kernel->program) {
   if (ir == nullptr)
-    this->ir = kernel->ir;
+    this->ir = kernel->ir.get();
   initialize_context();
 
   context_ty = get_runtime_type("Context");

diff --git a/taichi/ir/ir.cpp b/taichi/ir/ir.cpp
@@ -297,6 +297,14 @@ IRNode *Stmt::get_ir_root() {
   return dynamic_cast<IRNode *>(block);
 }
 
+Kernel *Stmt::get_kernel() const {
+  if (parent) {
+    return parent->get_kernel();
+  } else {
+    return nullptr;
+  }
+}
+
 std::vector<Stmt *> Stmt::get_operands() const {
   std::vector<Stmt *> ret;
   for (int i = 0; i < num_operands(); i++) {
@@ -706,6 +714,17 @@ Stmt *Block::mask() {
   }
 }
 
+Kernel *Block::get_kernel() const {
+  Block *parent = this->parent;
+  if (parent == nullptr) {
+    return kernel;
+  }
+  while (parent->parent) {
+    parent = parent->parent;
+  }
+  return parent->kernel;
+}
+
 void Block::set_statements(VecStatement &&stmts) {
   statements.clear();
   for (int i = 0; i < (int)stmts.size(); i++) {

diff --git a/taichi/ir/ir.h b/taichi/ir/ir.h
@@ -245,6 +245,9 @@ class IRNode {
   virtual void accept(IRVisitor *visitor) {
     TI_NOT_IMPLEMENTED
   }
+  virtual Kernel *get_kernel() const {
+    return nullptr;
+  }
   virtual ~IRNode() = default;
 
   template <typename T>
@@ -553,6 +556,8 @@ class Stmt : public IRNode {
 
   IRNode *get_ir_root();
 
+  Kernel *get_kernel() const override;
+
   virtual void repeat(int factor) {
     ret_type.width *= factor;
   }
@@ -809,6 +814,7 @@ class Block : public IRNode {
   std::vector<std::unique_ptr<Stmt>> statements, trash_bin;
   Stmt *mask_var;
   std::vector<SNode *> stop_gradients;
+  Kernel *kernel;
 
   // Only used in frontend. Stores LoopIndexStmt or BinaryOpStmt for loop
   // variables, and AllocaStmt for other variables.
@@ -817,6 +823,7 @@ class Block : public IRNode {
   Block() {
     mask_var = nullptr;
     parent = nullptr;
+    kernel = nullptr;
   }
 
   bool has_container_statements();
@@ -838,6 +845,7 @@ class Block : public IRNode {
                     bool replace_usages = true);
   Stmt *lookup_var(const Identifier &ident) const;
   Stmt *mask();
+  Kernel *get_kernel() const override;
 
   Stmt *back() const {
     return statements.back().get();

diff --git a/taichi/ir/transforms.h b/taichi/ir/transforms.h
@@ -14,16 +14,15 @@ void re_id(IRNode *root);
 void flag_access(IRNode *root);
 void die(IRNode *root);
 void simplify(IRNode *root, Kernel *kernel = nullptr);
-bool alg_simp(IRNode *root, const CompileConfig &config);
+
+bool alg_simp(IRNode *root);
 void whole_kernel_cse(IRNode *root);
 void variable_optimization(IRNode *root, bool after_lower_access);
 void extract_constant(IRNode *root);
-void full_simplify(IRNode *root,
-                   const CompileConfig &config,
-                   Kernel *kernel = nullptr);
+void full_simplify(IRNode *root, Kernel *kernel = nullptr);
 void print(IRNode *root, std::string *output = nullptr);
 void lower(IRNode *root);
-void typecheck(IRNode *root, Kernel *kernel = nullptr);
+void typecheck(IRNode *root);
 void loop_vectorize(IRNode *root);
 void slp_vectorize(IRNode *root);
 void vector_split(IRNode *root, int max_width, bool serial_schedule);

diff --git a/taichi/program/async_engine.cpp b/taichi/program/async_engine.cpp
@@ -51,7 +51,7 @@ void ExecutionQueue::enqueue(KernelLaunchRecord &&ker) {
         flag_access(stmt);
         lower_access(stmt, true, kernel);
         flag_access(stmt);
-        full_simplify(stmt, kernel->program.config, kernel);
+        full_simplify(stmt, kernel);
         // analysis::verify(stmt);
       }
       auto func = CodeGenCPU(kernel, stmt).codegen();
@@ -108,7 +108,7 @@ ExecutionQueue::ExecutionQueue()
 void AsyncEngine::launch(Kernel *kernel) {
   if (!kernel->lowered)
     kernel->lower(false);
-  auto block = dynamic_cast<Block *>(kernel->ir);
+  auto block = dynamic_cast<Block *>(kernel->ir.get());
   TI_ASSERT(block);
   auto &offloads = block->statements;
   for (std::size_t i = 0; i < offloads.size(); i++) {
@@ -266,7 +266,7 @@ bool AsyncEngine::fuse() {
       irpass::fix_block_parents(task_a);
 
       auto kernel = task_queue[i].kernel;
-      irpass::full_simplify(task_a, kernel->program.config, kernel);
+      irpass::full_simplify(task_a, kernel);
       task_queue[i].h = hash(task_a);
 
       modified = true;

diff --git a/taichi/program/kernel.cpp b/taichi/program/kernel.cpp
@@ -36,14 +36,14 @@ Kernel::Kernel(Program &program,
   is_evaluator = false;
   compiled = nullptr;
   taichi::lang::context = std::make_unique<FrontendContext>();
-  ir_holder = taichi::lang::context->get_root();
-  ir = ir_holder.get();
+  ir = taichi::lang::context->get_root();
 
   {
     CurrentKernelGuard _(program, this);
     program.start_function_definition(this);
     func();
     program.end_function_definition();
+    ir->as<Block>()->kernel = this;
   }
 
   arch = program.config.arch;
@@ -74,7 +74,7 @@ void Kernel::lower(bool lower_access) {  // TODO: is a "Lowerer" class necessary
     if (is_accessor && !config.print_accessor_ir)
       verbose = false;
     irpass::compile_to_offloads(
-        ir, config, /*vectorize*/ arch_is_cpu(arch), grad,
+        ir.get(), config, /*vectorize*/ arch_is_cpu(arch), grad,
         /*ad_use_stack*/ true, verbose, /*lower_global_access*/ lower_access);
   } else {
     TI_NOT_IMPLEMENTED

diff --git a/taichi/program/kernel.h b/taichi/program/kernel.h
@@ -10,8 +10,7 @@ class Program;
 
 class Kernel {
  public:
-  std::unique_ptr<IRNode> ir_holder;
-  IRNode *ir;
+  std::unique_ptr<IRNode> ir;
   Program &program;
   FunctionType compiled;
   std::string name;

diff --git a/taichi/transforms/alg_simp.cpp b/taichi/transforms/alg_simp.cpp
@@ -178,7 +178,8 @@ class AlgSimp : public BasicStmtVisitor {
 
 namespace irpass {
 
-bool alg_simp(IRNode *root, const CompileConfig &config) {
+bool alg_simp(IRNode *root) {
+  const auto &config = root->get_kernel()->program.config;
   return AlgSimp::run(root, config.fast_math);
 }
 

diff --git a/taichi/transforms/compile_to_offloads.cpp b/taichi/transforms/compile_to_offloads.cpp
@@ -53,9 +53,9 @@ void compile_to_offloads(IRNode *ir,
 
   if (grad) {
     irpass::demote_atomics(ir);
-    irpass::full_simplify(ir, config);
+    irpass::full_simplify(ir);
     irpass::make_adjoint(ir, ad_use_stack);
-    irpass::full_simplify(ir, config);
+    irpass::full_simplify(ir);
     print("Adjoint");
     irpass::analysis::verify(ir);
   }
@@ -91,7 +91,7 @@ void compile_to_offloads(IRNode *ir,
     irpass::analysis::verify(ir);
   }
 
-  irpass::full_simplify(ir, config);
+  irpass::full_simplify(ir);
   print("Simplified II");
   irpass::analysis::verify(ir);
 
@@ -122,7 +122,7 @@ void compile_to_offloads(IRNode *ir,
   irpass::variable_optimization(ir, true);
   print("Store forwarded II");
 
-  irpass::full_simplify(ir, config);
+  irpass::full_simplify(ir);
   print("Simplified III");
 
   // Final field registration correctness & type checking

diff --git a/taichi/transforms/constant_fold.cpp b/taichi/transforms/constant_fold.cpp
@@ -113,13 +113,14 @@ class ConstantFold : public BasicStmtVisitor {
                       rhs.dt,
                       true};
     auto *ker = get_jit_evaluator_kernel(id);
-    auto &ctx = get_current_program().get_context();
+    auto &current_program = stmt->get_kernel()->program;
+    auto &ctx = current_program.get_context();
     ContextArgSaveGuard _(
         ctx);  // save input args, prevent override current kernel
     ctx.set_arg<int64_t>(0, lhs.val_i64);
     ctx.set_arg<int64_t>(1, rhs.val_i64);
     (*ker)();
-    ret.val_i64 = get_current_program().fetch_result<int64_t>(0);
+    ret.val_i64 = current_program.fetch_result<int64_t>(0);
     return true;
   }
 
@@ -135,12 +136,13 @@ class ConstantFold : public BasicStmtVisitor {
                       stmt->cast_type,
                       false};
     auto *ker = get_jit_evaluator_kernel(id);
-    auto &ctx = get_current_program().get_context();
+    auto &current_program = stmt->get_kernel()->program;
+    auto &ctx = current_program.get_context();
     ContextArgSaveGuard _(
         ctx);  // save input args, prevent override current kernel
     ctx.set_arg<int64_t>(0, operand.val_i64);
     (*ker)();
-    ret.val_i64 = get_current_program().fetch_result<int64_t>(0);
+    ret.val_i64 = current_program.fetch_result<int64_t>(0);
     return true;
   }
 
@@ -204,7 +206,8 @@ void constant_fold(IRNode *root) {
   // disable constant_fold when config.debug is turned on.
   // Discussion:
   // https://github.com/taichi-dev/taichi/pull/839#issuecomment-626107010
-  if (get_current_program().config.debug) {
+  auto kernel = root->get_kernel();
+  if (kernel && kernel->program.config.debug) {
     TI_TRACE("config.debug enabled, ignoring constant fold");
     return;
   }

diff --git a/taichi/transforms/lower_access.cpp b/taichi/transforms/lower_access.cpp
@@ -246,7 +246,7 @@ namespace irpass {
 
 void lower_access(IRNode *root, bool lower_atomic, Kernel *kernel) {
   LowerAccess::run(root, lower_atomic);
-  typecheck(root, kernel);
+  typecheck(root);
 }
 
 }  // namespace irpass

diff --git a/taichi/transforms/simplify.cpp b/taichi/transforms/simplify.cpp
@@ -807,7 +807,7 @@ class BasicBlockSimplify : public IRVisitor {
     stmt->insert_before_me(std::move(sum));
     stmt->parent->erase(stmt);
     // get types of adds and muls
-    irpass::typecheck(stmt->parent, kernel);
+    irpass::typecheck(stmt->parent);
     throw IRModified();
   }
 
@@ -1160,10 +1160,10 @@ void simplify(IRNode *root, Kernel *kernel) {
   }
 }
 
-void full_simplify(IRNode *root, const CompileConfig &config, Kernel *kernel) {
+void full_simplify(IRNode *root, Kernel *kernel) {
   constant_fold(root);
   if (advanced_optimization) {
-    alg_simp(root, config);
+    alg_simp(root);
     die(root);
     whole_kernel_cse(root);
   }

diff --git a/taichi/transforms/type_check.cpp b/taichi/transforms/type_check.cpp
@@ -17,10 +17,11 @@ class TypeCheck : public IRVisitor {
   CompileConfig config;
 
  public:
-  TypeCheck(Kernel *kernel) : kernel(kernel) {
-    // TODO: remove dependency on get_current_program here
-    if (current_program != nullptr)
-      config = get_current_program().config;
+  TypeCheck(IRNode *root) {
+    kernel = root->get_kernel();
+    if (kernel != nullptr) {
+      config = kernel->program.config;
+    }
     allow_undefined_visitor = true;
   }
 
@@ -316,7 +317,7 @@ class TypeCheck : public IRVisitor {
   void visit(ArgLoadStmt *stmt) {
     Kernel *current_kernel = kernel;
     if (current_kernel == nullptr) {
-      current_kernel = &get_current_program().get_current_kernel();
+      current_kernel = stmt->get_kernel();
     }
     auto &args = current_kernel->args;
     TI_ASSERT(0 <= stmt->arg_id && stmt->arg_id < args.size());
@@ -326,7 +327,7 @@ class TypeCheck : public IRVisitor {
   void visit(KernelReturnStmt *stmt) {
     Kernel *current_kernel = kernel;
     if (current_kernel == nullptr) {
-      current_kernel = &get_current_program().get_current_kernel();
+      current_kernel = stmt->get_kernel();
     }
     auto &rets = current_kernel->rets;
     TI_ASSERT(rets.size() >= 1);
@@ -416,9 +417,9 @@ class TypeCheck : public IRVisitor {
 
 namespace irpass {
 
-void typecheck(IRNode *root, Kernel *kernel) {
+void typecheck(IRNode *root) {
   analysis::check_fields_registered(root);
-  TypeCheck inst(kernel);
+  TypeCheck inst(root);
   root->accept(&inst);
 }