[IR][refactor] Convert loop_var into LoopIndexStmt (#953)

* [skip ci] Convert loop_var into LoopIndexStmt (stage 1) * [skip ci] Fix irpass::offload * Should work on CPU now * Move convert_into_loop_index to right after lower_ast * Merge master * Fix tests * Fix CUDA compilation * [skip ci] Remove comments * Metal backend * [skip ci] Metal backend (continued) * opengl backend * [skip ci] enforce code format * Revert "[skip ci] enforce code format" This reverts commit f284688. * [skip ci] Remove LoopIndexStmt::is_struct_for * Finally move the pass into `lower` * [skip ci] add comment * Fix opengl backend * Use const Stmt* * [skip ci] enforce code format Co-authored-by: Taichi Gardener <taichigardener@gmail.com>
taichi-dev · May 14, 2020 · c6086e0 · c6086e0
1 parent 0cef342
commit c6086e0
Show file tree

Hide file tree

Showing 20 changed files with 243 additions and 161 deletions.
diff --git a/docs/hello.rst b/docs/hello.rst
@@ -55,9 +55,9 @@ Let's dive into this simple Taichi program.
 
 import taichi as ti
 -------------------
-Taichi is a domain-specific language (DSL) embedded in Python. To make Taichi as easy to use as a Python package, 
-we have done heavy engineering with this goal in mind - letting every Python programmer write Taichi codes with 
-minimal learning effort. You can even use your favorite Python package management system, Python IDEs and other 
+Taichi is a domain-specific language (DSL) embedded in Python. To make Taichi as easy to use as a Python package,
+we have done heavy engineering with this goal in mind - letting every Python programmer write Taichi codes with
+minimal learning effort. You can even use your favorite Python package management system, Python IDEs and other
 Python packages in conjunction with Taichi.
 
 Portability

diff --git a/taichi/analysis/verify.cpp b/taichi/analysis/verify.cpp
@@ -73,28 +73,46 @@ class IRVerifier : public BasicStmtVisitor {
     TI_ASSERT(stmt->ptr->is<AllocaStmt>());
   }
 
+  void visit(LoopIndexStmt *stmt) override {
+    basic_verify(stmt);
+    TI_ASSERT(stmt->loop);
+    if (stmt->loop->is<OffloadedStmt>()) {
+      TI_ASSERT(stmt->loop->as<OffloadedStmt>()->task_type ==
+                    OffloadedStmt::TaskType::struct_for ||
+                stmt->loop->as<OffloadedStmt>()->task_type ==
+                    OffloadedStmt::TaskType::range_for);
+    } else {
+      TI_ASSERT(stmt->loop->is<StructForStmt>() ||
+                stmt->loop->is<RangeForStmt>());
+    }
+  }
+
   void visit(RangeForStmt *for_stmt) override {
     basic_verify(for_stmt);
-    TI_ASSERT(for_stmt->loop_var->is<AllocaStmt>());
-    TI_ASSERT_INFO(irpass::analysis::gather_statements(
-                       for_stmt->loop_var->parent,
-                       [&](Stmt *s) {
-                         if (auto store = s->cast<LocalStoreStmt>())
-                           return store->ptr == for_stmt->loop_var;
-                         else if (auto atomic = s->cast<AtomicOpStmt>()) {
-                           return atomic->dest == for_stmt->loop_var;
-                         } else {
-                           return false;
-                         }
-                       })
-                       .empty(),
-                   "loop_var of {} modified", for_stmt->id);
+    if (for_stmt->loop_var) {
+      TI_ASSERT(for_stmt->loop_var->is<AllocaStmt>());
+      TI_ASSERT_INFO(irpass::analysis::gather_statements(
+                         for_stmt->loop_var->parent,
+                         [&](Stmt *s) {
+                           if (auto store = s->cast<LocalStoreStmt>())
+                             return store->ptr == for_stmt->loop_var;
+                           else if (auto atomic = s->cast<AtomicOpStmt>()) {
+                             return atomic->dest == for_stmt->loop_var;
+                           } else {
+                             return false;
+                           }
+                         })
+                         .empty(),
+                     "loop_var of {} modified", for_stmt->id);
+    }
     for_stmt->body->accept(this);
   }
 
   void visit(StructForStmt *for_stmt) override {
     basic_verify(for_stmt);
     for (auto &loop_var : for_stmt->loop_vars) {
+      if (!loop_var)
+        continue;
       TI_ASSERT(loop_var->is<AllocaStmt>());
       TI_ASSERT_INFO(irpass::analysis::gather_statements(
                          loop_var->parent,

diff --git a/taichi/backends/cpu/codegen_cpu.cpp b/taichi/backends/cpu/codegen_cpu.cpp
@@ -32,7 +32,7 @@ class CodeGenLLVMCPU : public CodeGenLLVM {
            tlctx->get_data_type<int>()});
 
       auto loop_var = create_entry_block_alloca(DataType::i32);
-      offloaded_loop_vars_llvm[stmt].push_back(loop_var);
+      loop_vars_llvm[stmt].push_back(loop_var);
       builder->CreateStore(get_arg(1), loop_var);
       stmt->body->accept(this);
 

diff --git a/taichi/backends/cuda/codegen_cuda.cpp b/taichi/backends/cuda/codegen_cuda.cpp
@@ -346,7 +346,7 @@ class CodeGenLLVMCUDA : public CodeGenLLVM {
            tlctx->get_data_type<int>()});
 
       auto loop_var = create_entry_block_alloca(DataType::i32);
-      offloaded_loop_vars_llvm[stmt].push_back(loop_var);
+      loop_vars_llvm[stmt].push_back(loop_var);
       builder->CreateStore(get_arg(1), loop_var);
       stmt->body->accept(this);
 

diff --git a/taichi/backends/metal/codegen_metal.cpp b/taichi/backends/metal/codegen_metal.cpp
@@ -309,15 +309,22 @@ class KernelCodegen : public IRVisitor {
   }
 
   void visit(LoopIndexStmt *stmt) override {
-    using TaskType = OffloadedStmt::TaskType;
-    const auto type = current_kernel_attribs_->task_type;
     const auto stmt_name = stmt->raw_name();
-    if (type == TaskType::range_for) {
+    if (stmt->loop->is<OffloadedStmt>()) {
+      using TaskType = OffloadedStmt::TaskType;
+      const auto type = stmt->loop->as<OffloadedStmt>()->task_type;
+      if (type == TaskType::range_for) {
+        TI_ASSERT(stmt->index == 0);
+        emit("const int {} = {};", stmt_name, kLinearLoopIndexName);
+      } else if (type == TaskType::struct_for) {
+        emit("const int {} = {}.coords[{}];", stmt_name, kListgenElemVarName,
+             stmt->index);
+      } else {
+        TI_NOT_IMPLEMENTED;
+      }
+    } else if (stmt->loop->is<RangeForStmt>()) {
       TI_ASSERT(stmt->index == 0);
-      emit("const int {} = {};", stmt_name, kLinearLoopIndexName);
-    } else if (type == TaskType::struct_for) {
-      emit("const int {} = {}.coords[{}];", stmt_name, kListgenElemVarName,
-           stmt->index);
+      emit("const int {} = {};", stmt_name, stmt->loop->raw_name());
     } else {
       TI_NOT_IMPLEMENTED;
     }
@@ -445,29 +452,18 @@ class KernelCodegen : public IRVisitor {
 
   void visit(RangeForStmt *for_stmt) override {
     TI_ASSERT(for_stmt->width() == 1);
-    auto *loop_var = for_stmt->loop_var;
-    if (loop_var->ret_type.data_type == DataType::i32) {
-      if (!for_stmt->reversed) {
-        emit("for (int {}_ = {}; {}_ < {}; {}_ = {}_ + {}) {{",
-             loop_var->raw_name(), for_stmt->begin->raw_name(),
-             loop_var->raw_name(), for_stmt->end->raw_name(),
-             loop_var->raw_name(), loop_var->raw_name(), 1);
-        emit("  int {} = {}_;", loop_var->raw_name(), loop_var->raw_name());
-      } else {
-        // reversed for loop
-        emit("for (int {}_ = {} - 1; {}_ >= {}; {}_ = {}_ - {}) {{",
-             loop_var->raw_name(), for_stmt->end->raw_name(),
-             loop_var->raw_name(), for_stmt->begin->raw_name(),
-             loop_var->raw_name(), loop_var->raw_name(), 1);
-        emit("  int {} = {}_;", loop_var->raw_name(), loop_var->raw_name());
-      }
+    auto loop_var_name = for_stmt->raw_name();
+    if (!for_stmt->reversed) {
+      emit("for (int {}_ = {}; {}_ < {}; {}_ = {}_ + {}) {{", loop_var_name,
+           for_stmt->begin->raw_name(), loop_var_name,
+           for_stmt->end->raw_name(), loop_var_name, loop_var_name, 1);
+      emit("  int {} = {}_;", loop_var_name, loop_var_name);
     } else {
-      TI_ASSERT(!for_stmt->reversed);
-      const auto type_name = metal_data_type_name(loop_var->element_type());
-      emit("for ({} {} = {}; {} < {}; {} = {} + ({})1) {{", type_name,
-           loop_var->raw_name(), for_stmt->begin->raw_name(),
-           loop_var->raw_name(), for_stmt->end->raw_name(),
-           loop_var->raw_name(), loop_var->raw_name(), type_name);
+      // reversed for loop
+      emit("for (int {}_ = {} - 1; {}_ >= {}; {}_ = {}_ - {}) {{",
+           loop_var_name, for_stmt->end->raw_name(), loop_var_name,
+           for_stmt->begin->raw_name(), loop_var_name, loop_var_name, 1);
+      emit("  int {} = {}_;", loop_var_name, loop_var_name);
     }
     for_stmt->body->accept(this);
     emit("}}");

diff --git a/taichi/backends/metal/shaders/helpers.metal.h b/taichi/backends/metal/shaders/helpers.metal.h
@@ -33,8 +33,9 @@ STR(
 
     inline int ifloordiv(int lhs, int rhs) {
       const int intm = (lhs / rhs);
-      return (((lhs < 0) != (rhs < 0) && lhs &&
-            (rhs * intm != lhs)) ? (intm - 1) : intm);
+      return (((lhs < 0) != (rhs < 0) && lhs && (rhs * intm != lhs))
+                  ? (intm - 1)
+                  : intm);
     }
 
     int32_t pow_i32(int32_t x, int32_t n) {

diff --git a/taichi/backends/opengl/codegen_opengl.cpp b/taichi/backends/opengl/codegen_opengl.cpp
@@ -414,14 +414,13 @@ class KernelGen : public IRVisitor {
         emit("{} {} = atan({}, {});", dt_name, bin_name, lhs_name, rhs_name);
       }
       return;
-    } else if (bin->op_type == BinaryOpType::pow
-        && is_integral(bin->rhs->element_type())) {
-      // The GLSL `pow` is not so percise for `int`... e.g.: `pow(5, 3)` obtains 124
-      // So that we have to use some hack to make it percise.
-      // Discussion: https://github.com/taichi-dev/taichi/pull/943#issuecomment-626354902
+    } else if (bin->op_type == BinaryOpType::pow &&
+               is_integral(bin->rhs->element_type())) {
+      // The GLSL `pow` is not so percise for `int`... e.g.: `pow(5, 3)` obtains
+      // 124 So that we have to use some hack to make it percise. Discussion:
+      // https://github.com/taichi-dev/taichi/pull/943#issuecomment-626354902
       emit("{} {} = {}(fast_pow_{}({}, {}));", dt_name, bin_name, dt_name,
-          data_type_short_name(bin->lhs->element_type()),
-           lhs_name, rhs_name);
+           data_type_short_name(bin->lhs->element_type()), lhs_name, rhs_name);
       used.fast_pow = true;
       return;
     }
@@ -602,38 +601,32 @@ class KernelGen : public IRVisitor {
   }
 
   void visit(LoopIndexStmt *stmt) override {
-    TI_ASSERT(!stmt->is_struct_for);
     TI_ASSERT(stmt->index == 0);  // TODO: multiple indices
-    emit("int {} = _itv;", stmt->short_name());
+    if (stmt->loop->is<OffloadedStmt>()) {
+      TI_ASSERT(stmt->loop->as<OffloadedStmt>()->task_type ==
+                OffloadedStmt::TaskType::range_for);
+      emit("int {} = _itv;", stmt->short_name());
+    } else if (stmt->loop->is<RangeForStmt>()) {
+      emit("int {} = {};", stmt->short_name(), stmt->loop->short_name());
+    } else {
+      TI_NOT_IMPLEMENTED
+    }
   }
 
   void visit(RangeForStmt *for_stmt) override {
     TI_ASSERT(for_stmt->width() == 1);
-    auto *loop_var = for_stmt->loop_var;
-    if (loop_var->ret_type.data_type == DataType::i32) {
-      if (!for_stmt->reversed) {
-        emit("for (int {}_ = {}; {}_ < {}; {}_ = {}_ + {}) {{",
-             loop_var->short_name(), for_stmt->begin->short_name(),
-             loop_var->short_name(), for_stmt->end->short_name(),
-             loop_var->short_name(), loop_var->short_name(), 1);
-        // variable named `loop_var->short_name()` is already allocated by
-        // alloca
-        emit("  {} = {}_;", loop_var->short_name(), loop_var->short_name());
-      } else {
-        // reversed for loop
-        emit("for (int {}_ = {} - 1; {}_ >= {}; {}_ = {}_ - {}) {{",
-             loop_var->short_name(), for_stmt->end->short_name(),
-             loop_var->short_name(), for_stmt->begin->short_name(),
-             loop_var->short_name(), loop_var->short_name(), 1);
-        emit("  {} = {}_;", loop_var->short_name(), loop_var->short_name());
-      }
+    auto loop_var_name = for_stmt->short_name();
+    if (!for_stmt->reversed) {
+      emit("for (int {}_ = {}; {}_ < {}; {}_ = {}_ + {}) {{", loop_var_name,
+           for_stmt->begin->short_name(), loop_var_name,
+           for_stmt->end->short_name(), loop_var_name, loop_var_name, 1);
+      emit("  int {} = {}_;", loop_var_name, loop_var_name);
     } else {
-      TI_ASSERT(!for_stmt->reversed);
-      const auto type_name = opengl_data_type_name(loop_var->element_type());
-      emit("for ({} {} = {}; {} < {}; {} = {} + 1) {{", type_name,
-           loop_var->short_name(), for_stmt->begin->short_name(),
-           loop_var->short_name(), for_stmt->end->short_name(),
-           loop_var->short_name(), loop_var->short_name());
+      // reversed for loop
+      emit("for (int {}_ = {} - 1; {}_ >= {}; {}_ = {}_ - {}) {{",
+           loop_var_name, for_stmt->end->short_name(), loop_var_name,
+           for_stmt->begin->short_name(), loop_var_name, loop_var_name, 1);
+      emit("  int {} = {}_;", loop_var_name, loop_var_name);
     }
     for_stmt->body->accept(this);
     emit("}}");
@@ -730,12 +723,14 @@ void OpenglCodeGen::lower() {
   auto ir = kernel_->ir;
   auto &config = kernel_->program.config;
   config.demote_dense_struct_fors = true;
-  auto res = irpass::compile_to_offloads(ir, config,
-                              /*vectorize=*/false, kernel_->grad,
-                              /*ad_use_stack=*/false, config.print_ir,
-                              /*lower_global_access*/true);
+  auto res =
+      irpass::compile_to_offloads(ir, config,
+                                  /*vectorize=*/false, kernel_->grad,
+                                  /*ad_use_stack=*/false, config.print_ir,
+                                  /*lower_global_access*/ true);
   global_tmps_buffer_size_ = res.total_size;
-  TI_TRACE("[glsl] Global temporary buffer size {} B", global_tmps_buffer_size_);
+  TI_TRACE("[glsl] Global temporary buffer size {} B",
+           global_tmps_buffer_size_);
 #ifdef _GLSL_DEBUG
   irpass::print(ir);
 #endif

diff --git a/taichi/codegen/codegen_llvm.cpp b/taichi/codegen/codegen_llvm.cpp
@@ -790,13 +790,16 @@ void CodeGenLLVM::create_naive_range_for(RangeForStmt *for_stmt) {
   BasicBlock *after_loop = BasicBlock::Create(*llvm_context, "after_for", func);
   BasicBlock *loop_test =
       BasicBlock::Create(*llvm_context, "for_loop_test", func);
+
+  auto loop_var = create_entry_block_alloca(DataType::i32);
+  loop_vars_llvm[for_stmt].push_back(loop_var);
+
   if (!for_stmt->reversed) {
-    builder->CreateStore(llvm_val[for_stmt->begin],
-                         llvm_val[for_stmt->loop_var]);
+    builder->CreateStore(llvm_val[for_stmt->begin], loop_var);
   } else {
     builder->CreateStore(
         builder->CreateSub(llvm_val[for_stmt->end], tlctx->get_constant(1)),
-        llvm_val[for_stmt->loop_var]);
+        loop_var);
   }
   builder->CreateBr(loop_test);
 
@@ -805,15 +808,13 @@ void CodeGenLLVM::create_naive_range_for(RangeForStmt *for_stmt) {
     builder->SetInsertPoint(loop_test);
     llvm::Value *cond;
     if (!for_stmt->reversed) {
-      cond =
-          builder->CreateICmp(llvm::CmpInst::Predicate::ICMP_SLT,
-                              builder->CreateLoad(llvm_val[for_stmt->loop_var]),
-                              llvm_val[for_stmt->end]);
+      cond = builder->CreateICmp(llvm::CmpInst::Predicate::ICMP_SLT,
+                                 builder->CreateLoad(loop_var),
+                                 llvm_val[for_stmt->end]);
     } else {
-      cond =
-          builder->CreateICmp(llvm::CmpInst::Predicate::ICMP_SGE,
-                              builder->CreateLoad(llvm_val[for_stmt->loop_var]),
-                              llvm_val[for_stmt->begin]);
+      cond = builder->CreateICmp(llvm::CmpInst::Predicate::ICMP_SGE,
+                                 builder->CreateLoad(loop_var),
+                                 llvm_val[for_stmt->begin]);
     }
     builder->CreateCondBr(cond, body, after_loop);
   }
@@ -833,9 +834,9 @@ void CodeGenLLVM::create_naive_range_for(RangeForStmt *for_stmt) {
     builder->SetInsertPoint(loop_inc);
 
     if (!for_stmt->reversed) {
-      create_increment(llvm_val[for_stmt->loop_var], tlctx->get_constant(1));
+      create_increment(loop_var, tlctx->get_constant(1));
     } else {
-      create_increment(llvm_val[for_stmt->loop_var], tlctx->get_constant(-1));
+      create_increment(loop_var, tlctx->get_constant(-1));
     }
     builder->CreateBr(loop_test);
   }
@@ -1408,13 +1409,15 @@ void CodeGenLLVM::create_offload_struct_for(OffloadedStmt *stmt, bool spmd) {
 
 void CodeGenLLVM::visit(LoopIndexStmt *stmt) {
   TI_ASSERT(&module->getContext() == tlctx->get_this_thread_context());
-  if (stmt->is_struct_for) {
+  if (stmt->loop->is<OffloadedStmt>() &&
+      stmt->loop->as<OffloadedStmt>()->task_type ==
+          OffloadedStmt::TaskType::struct_for) {
     llvm_val[stmt] = builder->CreateLoad(builder->CreateGEP(
         current_coordinates, {tlctx->get_constant(0), tlctx->get_constant(0),
                               tlctx->get_constant(stmt->index)}));
   } else {
-    llvm_val[stmt] = builder->CreateLoad(
-        offloaded_loop_vars_llvm[current_offloaded_stmt][stmt->index]);
+    llvm_val[stmt] =
+        builder->CreateLoad(loop_vars_llvm[stmt->loop][stmt->index]);
   }
 }
 

diff --git a/taichi/codegen/codegen_llvm.h b/taichi/codegen/codegen_llvm.h
@@ -72,8 +72,7 @@ class CodeGenLLVM : public IRVisitor, public LLVMModuleBuilder {
   std::vector<OffloadedTask> offloaded_tasks;
   BasicBlock *func_body_bb;
 
-  std::unordered_map<OffloadedStmt *, std::vector<llvm::Value *>>
-      offloaded_loop_vars_llvm;
+  std::unordered_map<const Stmt *, std::vector<llvm::Value *>> loop_vars_llvm;
 
   using IRVisitor::visit;
   using LLVMModuleBuilder::call;

diff --git a/taichi/ir/ir.cpp b/taichi/ir/ir.cpp
@@ -655,10 +655,11 @@ void Block::replace_statements_in_range(int start,
 }
 
 void Block::replace_with(Stmt *old_statement,
-                         std::unique_ptr<Stmt> &&new_statement) {
+                         std::unique_ptr<Stmt> &&new_statement,
+                         bool replace_usages) {
   VecStatement vec;
   vec.push_back(std::move(new_statement));
-  replace_with(old_statement, std::move(vec));
+  replace_with(old_statement, std::move(vec), replace_usages);
 }
 
 Stmt *Block::lookup_var(const Identifier &ident) const {

diff --git a/taichi/ir/ir.h b/taichi/ir/ir.h
@@ -835,7 +835,9 @@ class Block : public IRNode {
   void insert(VecStatement &&stmt, int location = -1);
   void replace_statements_in_range(int start, int end, VecStatement &&stmts);
   void set_statements(VecStatement &&stmts);
-  void replace_with(Stmt *old_statement, std::unique_ptr<Stmt> &&new_statement);
+  void replace_with(Stmt *old_statement,
+                    std::unique_ptr<Stmt> &&new_statement,
+                    bool replace_usages = true);
   void insert_before(Stmt *old_statement, VecStatement &&new_statements);
   void replace_with(Stmt *old_statement,
                     VecStatement &&new_statements,