Merge branch 'PaddlePaddle:develop' into C18

cmake/cuda.cmake

@@ -294,7 +294,7 @@ select_nvcc_arch_flags(NVCC_FLAGS_EXTRA NVCC_ARCH_BIN)
 set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${NVCC_FLAGS_EXTRA}")
 message(STATUS "NVCC_FLAGS_EXTRA: ${NVCC_FLAGS_EXTRA}")
 
-# Set C++14 support
+# Set C++17 support
 set(CUDA_PROPAGATE_HOST_FLAGS OFF)
 # Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
 # So, don't set these flags here.

cmake/phi_header.cmake → cmake/export_paddle_header.cmake

@@ -15,33 +15,57 @@
 set(PADDLE_INFERENCE_INSTALL_DIR
     "${CMAKE_BINARY_DIR}/paddle_inference_install_dir")
 
-function(phi_header_path_compat TARGET_PATH)
-  message(STATUS "phi header path compat processing: ${TARGET_PATH}")
+function(header_path_compat TARGET_PATH)
+  message(STATUS "header path compat processing: ${TARGET_PATH}")
   file(GLOB HEADERS "${TARGET_PATH}/*" "*.h")
   foreach(header ${HEADERS})
     if(${header} MATCHES ".*.h$")
       file(READ ${header} HEADER_CONTENT)
       string(REPLACE "paddle/fluid/platform/" "paddle/phi/" HEADER_CONTENT
                      "${HEADER_CONTENT}")
+      string(REPLACE "paddle/pir/include/" "paddle/pir/" HEADER_CONTENT
+                     "${HEADER_CONTENT}")
+      string(REPLACE "paddle/fluid/pir/drr/include/" "paddle/pir/drr/"
+                     HEADER_CONTENT "${HEADER_CONTENT}")
+      string(REPLACE "paddle/fluid/pir/transforms/" "paddle/pir/transforms/"
+                     HEADER_CONTENT "${HEADER_CONTENT}")
       file(WRITE ${header} "${HEADER_CONTENT}")
-      message(STATUS "phi header path compat processing complete: ${header}")
+      message(STATUS "header path compat processing complete: ${header}")
     endif()
   endforeach()
 endfunction()
 
-phi_header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle)
-phi_header_path_compat(
-  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi)
-phi_header_path_compat(
+header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle)
+header_path_compat(${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi)
+header_path_compat(
   ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/api)
-phi_header_path_compat(
+header_path_compat(
   ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/api/ext)
-phi_header_path_compat(
+header_path_compat(
   ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/api/include)
-phi_header_path_compat(
+header_path_compat(
   ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/common)
-phi_header_path_compat(
+header_path_compat(
   ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/phi/core)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/core)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/core/parser)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/dialect/control_flow/ir
+)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/dialect/shape/ir)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/dialect/shape/utils)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/drr)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/pass)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/pattern_rewrite)
+header_path_compat(
+  ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/transforms)
 
 # NOTE(liuyuanle): In inference lib, no need include paddle/utils/pybind.h, so we delete this.
 file(READ ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/extension.h

cmake/inference_lib.cmake

@@ -354,12 +354,54 @@ copy(
   SRCS ${PADDLE_SOURCE_DIR}/paddle/extension.h
   DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/)
 
-# the include path of phi needs to be changed to adapt to inference api path
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/core/parser/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/core/parser/)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/core/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/core/)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/dialect/control_flow/ir/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/dialect/control_flow/ir/
+)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/dialect/shape/ir/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/dialect/shape/ir/
+)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/dialect/shape/utils/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/dialect/shape/utils/
+)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/pass/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/pass/)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/pir/include/pattern_rewrite/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/pattern_rewrite/
+)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/fluid/pir/drr/include/*.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/drr/)
+copy(
+  inference_lib_dist
+  SRCS ${PADDLE_SOURCE_DIR}/paddle/fluid/pir/transforms/transform_general_functions.h
+  DSTS ${PADDLE_INFERENCE_INSTALL_DIR}/paddle/include/paddle/pir/transforms/)
+
+# the include path of paddle needs to be changed to adapt to inference api path
 add_custom_command(
   TARGET inference_lib_dist
   POST_BUILD
-  COMMAND ${CMAKE_COMMAND} -P "${PADDLE_SOURCE_DIR}/cmake/phi_header.cmake"
-  COMMENT "Change phi header include path to adapt to inference api path")
+  COMMAND ${CMAKE_COMMAND} -P
+          "${PADDLE_SOURCE_DIR}/cmake/export_paddle_header.cmake"
+  COMMENT "Change paddle header include path to adapt to inference api path")
 
 # CAPI inference library for only inference
 set(PADDLE_INFERENCE_C_INSTALL_DIR

paddle/cinn/backends/codegen_c_test.cc

@@ -61,9 +61,9 @@ TEST(CodeGenC, module) {
   LOG(INFO) << "C.body: " << C->get_compute_op()->body.front();
 
   Target target;
-  target.arch = Target::Arch ::X86;
-  target.bits = Target::Bit ::k32;
-  target.os = Target::OS ::Linux;
+  target.arch = Target::Arch::X86;
+  target.bits = Target::Bit::k32;
+  target.os = Target::OS::Linux;
   Module::Builder builder("module1", target);
 
   ast_gen_ius::TensorGroup tensor_group({A, B, C});

paddle/cinn/hlir/dialect/operator/transforms/add_cinn_pass.cc

@@ -107,9 +107,9 @@ void ApplyCinnPreprocessPass(
 
   pass_manager->AddPass(
       cinn::dialect::ir::CreateMoveGenerateShapeOpsToProloguePass());
-  pass_manager->AddPass(cinn::dialect::ir::CreateDivideGroupOpToFusionOpPass());
   pass_manager->AddPass(cinn::dialect::ir::CreateDynamicReshapeOpPass());
   pass_manager->AddPass(cinn::dialect::ir::CreateReplaceDynamicExpandOpPass());
+  pass_manager->AddPass(cinn::dialect::ir::CreateDivideGroupOpToFusionOpPass());
   pass_manager->AddPass(pir::CreateDeadCodeEliminationPass());
 
   pass_manager->Run(program);

paddle/cinn/hlir/dialect/operator/transforms/dynamic_reshape_pass.cc

@@ -118,7 +118,7 @@ class DynamicReshapeOpPass : public pir::Pass {
     for (uint32_t i = 0; i < op->num_regions(); ++i) {
       for (auto& block : op->region(i)) {
         for (auto& op : block) {
-          if (op.isa<cinn::dialect::FusionOp>()) {
+          if (op.isa<cinn::dialect::GroupOp>()) {
             auto [_, num_rewrites] =
                 pir::ApplyPatternsGreedily(&op, patterns_, cfg);
             AddStatistics(num_rewrites);

paddle/cinn/hlir/dialect/operator/transforms/replace_dynamic_expand_pass.cc

@@ -52,7 +52,28 @@ class DynamicExpandOpPattern
       for (size_t i = 0; i < x_rank; ++i) {
         broadcast_axes[i] = i + index_gap;
       }
-      std::vector<int64_t> out_shape(out_rank, -1);
+
+      pir::ShapeConstraintIRAnalysis& shape_analysis =
+          pir::ShapeAnalysisManager::Instance().Get(op->GetParentProgram());
+
+      const auto& UpdateOutputShapeByDimExpr = [&]() -> std::vector<int64_t> {
+        std::vector<int64_t> out_shape(out_rank, -1);
+        if (shape_analysis.HasShapeOrDataForValue(op->result(0))) {
+          VLOG(3) << "found shape dialect";
+          auto shape_info =
+              shape_analysis.GetShapeOrDataForValue(op->result(0)).shape();
+
+          for (size_t i = 0; i < shape_info.size(); ++i) {
+            if (shape_info[i].isa<int64_t>()) {
+              out_shape[i] = shape_info[i].Get<int64_t>();
+            }
+          }
+        }
+        return out_shape;
+      };
+
+      auto out_shape = UpdateOutputShapeByDimExpr();
+
       return rewriter.Build<cinn::dialect::BroadcastOp>(
           op->operand_source(0), broadcast_axes, out_shape);
     }();
@@ -91,7 +112,7 @@ class ReplaceDynamicExpandOpPass : public pir::Pass {
     for (uint32_t i = 0; i < op->num_regions(); ++i) {
       for (auto& block : op->region(i)) {
         for (auto& op : block) {
-          if (op.isa<cinn::dialect::FusionOp>()) {
+          if (op.isa<cinn::dialect::GroupOp>()) {
             const auto& [_, num_rewrites] =
                 pir::ApplyPatternsGreedily(&op, patterns_, cfg);
             AddStatistics(num_rewrites);

paddle/cinn/hlir/framework/pir/op_mapper.h

@@ -13,9 +13,12 @@
 // limitations under the License.
 
 #pragma once
+
+#include <glog/logging.h>
 #include <string>
 #include <unordered_map>
 #include <vector>
+
 #include "paddle/cinn/utils/type_defs.h"
 #include "paddle/pir/include/core/operation.h"
 

paddle/cinn/ir/schedule/impl/base.cc

@@ -428,7 +428,7 @@ Expr DyScheduleImpl::SampleCategorical(
   std::string primitive = "SampleCategorical";
   std::ostringstream os;
   if (candidates.size() != probs.size()) {
-    os << "vector<int> params(candidates) and vector<int> prama(probs) must "
+    os << "vector<int> params(candidates) and vector<int> params(probs) must "
           "have same size in SampleCategorical!\n";
     throw IRScheduleErrorHandler(primitive, os.str(), module_expr_);
   }

paddle/cinn/ir/schedule/impl/compute_location.cc

@@ -42,11 +42,11 @@ void DyScheduleImpl::ComputeAt(const Expr& block,
   std::string primitive = "ComputeAt";
   std::ostringstream os;
   if (!block.As<ir::ScheduleBlockRealize>()) {
-    os << "Expr prama(block) should be a ScheduleBlockRealize!\n";
+    os << "Expr param(block) should be a ScheduleBlockRealize!\n";
     throw IRScheduleErrorHandler(primitive, os.str(), module_expr_);
   }
   if (!loop.As<ir::For>()) {
-    os << "Expr prama(loop) should be a For node!\n";
+    os << "Expr param(loop) should be a For node!\n";
     throw IRScheduleErrorHandler(primitive, os.str(), module_expr_);
   }
   Expr root = this->GetRootBlock(block);

paddle/cinn/ir/schedule/ir_schedule_error.cc

@@ -21,7 +21,7 @@ namespace ir {
 
 std::string IRScheduleErrorHandler::GeneralErrorMessage() const {
   std::ostringstream os;
-  os << "[IRScheduleError] An error occurred in the scheduel primitive < "
+  os << "[IRScheduleError] An error occurred in the schedule primitive < "
      << this->primitive_ << " >. " << std::endl;
   os << indent_str_ << "[Error info] " << this->err_msg_;
   return os.str();

paddle/cinn/ir/schedule/ir_schedule_util.cc

@@ -207,7 +207,7 @@ void ReplaceExpr(Expr* source,
                  const std::vector<Expr>& candidates) {
   CHECK_EQ(replaced.size(), candidates.size())
       << "In ReplaceExpr, the size of Vars to be replaced must be equal to the "
-         "size of cadidate Exprs! Please check.";
+         "size of candidate Exprs! Please check.";
   if (replaced.empty()) return;
   std::map<Var, Expr, CompVar> replacing_map;
   for (int i = 0; i < replaced.size(); ++i) {
@@ -764,7 +764,7 @@ Expr ConstructNewLoopChain(const std::vector<Expr>& chain,
   //   }                                             }
   // }                                             }
   //
-  // We go throuph origin loop and check other body stmts, adding it as another
+  // We go through origin loop and check other body stmts, adding it as another
   // chain, such as:
   //
   // for (i, 0, 32) {
@@ -1022,7 +1022,7 @@ void InsertBlock(Expr& for_loop, const Expr& insertion, int index) {  // NOLINT
     auto dst_it = dst_block->stmts.begin() + index;
     if (dst_it->As<IfThenElse>()) {
       auto* inserted_block = dst_it->As<IfThenElse>()->true_case.As<Block>();
-      CHECK(inserted_block) << "the IfThenElse node to be inserted shuold "
+      CHECK(inserted_block) << "the IfThenElse node to be inserted should "
                                "contain a true_case block";
       inserted_block->stmts.insert(inserted_block->stmts.begin(), insertion);
     } else {
@@ -1060,7 +1060,7 @@ std::vector<IterRange> CalculateRequiredRegions(
   }
 
   std::vector<IterRange> required_buffer_range;
-  // deduce accessed regions of the provided tensor in block by itering each
+  // deduce accessed regions of the provided tensor in block by iterating each
   // required block
   for (const Expr& pro_node : provided_nodes) {
     std::string provided_tensor_name =

paddle/cinn/ir/schedule/schedule_desc.cc

@@ -27,7 +27,7 @@
 namespace cinn {
 namespace ir {
 
-// ------ Following codes are about `Apply` functions registry of variaous types
+// ------ Following codes are about `Apply` functions registry of various types
 // of ScheduleDesc::Step
 class PackedStepContext;
 // uniformed function prototype of a scheduling operation in IRSchedule
@@ -118,7 +118,7 @@ class PackedStepContext {
       return absl::get<AttrType>(attrs_.at(idx));
     } catch (absl::bad_variant_access& ex) {
       LOG(FATAL) << "Attribute cast error, idx:" << idx
-                 << ", get tpye:" << typeid(AttrType).name()
+                 << ", get type:" << typeid(AttrType).name()
                  << ", real index:" << attrs_.at(idx).index();
       throw ex;
     }
@@ -197,7 +197,7 @@ struct FreeFuncConverter<Return (IRSchedule::*)(Args...) const, impl_fn> {
   }
 };
 
-// used for formatting scheduling functions with variaous function signatures to
+// used for formatting scheduling functions with various function signatures to
 // be uniformed form
 template <typename F, F f>
 struct ApplyFuncImpl;
@@ -689,8 +689,8 @@ proto::ScheduleDesc ScheduleDesc::ToProto() const {
       }
     }
 
-    // each output Expr is represented by a formatted name, to be refered by
-    // suceeding steps
+    // each output Expr is represented by a formatted name, to be referred by
+    // succeeding steps
     for (auto&& expr : step.outputs) {
       std::string local_name = "e" + std::to_string(expr2name.size());
       expr2name.emplace(expr, local_name);
@@ -722,7 +722,7 @@ std::vector<Expr> ScheduleDesc::ReplayWithProto(
   absl::flat_hash_map<std::string, Expr> name2expr;
   std::vector<Expr> last_outputs;
 
-  // resotre each scheduling step and apply to the new IRSchedule object
+  // restore each scheduling step and apply to the new IRSchedule object
   for (auto&& step_proto : desc_proto.steps()) {
     VLOG(4) << "Replay step:\n" << step_proto.DebugString();
     ScheduleDesc::Step step;

paddle/cinn/ir/test/tensor_test.cc

@@ -144,7 +144,7 @@ TEST(Tensor, ReshapeCopied) {
 
   stages->InsertLazily(B);
 
-  ir::Module::Builder builder("some_modue", cinn::common::DefaultHostTarget());
+  ir::Module::Builder builder("some_module", cinn::common::DefaultHostTarget());
   auto func = lang::Lower("fn", stages, {A, B}, {}, {}, &builder);
 
   backends::CodeGenC codegenc(cinn::common::DefaultHostTarget());

paddle/cinn/lang/lower_impl.h

@@ -150,8 +150,8 @@ class LowerImpl {
   std::vector<Tensor> CollectTemporaryTensors();
 
   /**
-   * \brief Check both the tensor_args and sclar_args not contain duplication
-   * (different arguemnt with the same name).
+   * \brief Check both the tensor_args and scalar_args not contain duplication
+   * (different argument with the same name).
    */
   void CheckArgsUnique();
 
@@ -304,7 +304,7 @@ struct MarkParallelMutator : public ir::IRMutator<Expr*> {
     auto it = parallels.find(tensor_n->name);
     if (it != parallels.end()) {
       for (int level : it->second) {
-        VLOG(1) << "Mark " << level << " Paralled";
+        VLOG(1) << "Mark " << level << " Parallelled";
         CHECK_LT(level, stack.size());
         stack[level]->set_parallel();
       }

paddle/cinn/optim/insert_debug_log_callee.cc

@@ -139,7 +139,7 @@ struct InsertDebugLogCalleeMutator : public ir::IRMutator<> {
     ir::IRMutator<>::Visit(&node->body, &node->body);
 
     auto deal_with_exprs =
-        [&](std::vector<Expr> *exprs) {  // deal with op->argument_preapre_exprs
+        [&](std::vector<Expr> *exprs) {  // deal with op->argument_prepare_exprs
           std::vector<Expr> new_stmts;
           for (auto &expr : *exprs) {
             auto msg =

paddle/cinn/optim/unroll_loops.cc

@@ -62,7 +62,7 @@ struct UnrollMutator : public ir::IRMutator<Expr*> {
   void Visit(const ir::For* op, Expr* expr) override {
     IRMutator<>::Visit(op, expr);
     if (op->extent.As<ir::IntImm>() == nullptr) {
-      VLOG(5) << "loop to be unrolled should have a contant extent";
+      VLOG(5) << "loop to be unrolled should have a constant extent";
       return;
     }
     int64_t extent = op->extent.as_int64();