[new-exec] enable the new standalone executor by default (#41179)

* enable new executor by default

* enable stream safe allocator

* test=document_fix;test=coverage

* do not use scope in op kernel

* fit empty program for new executor

* fix communication depend

* fix test_sync_batch_norm

* skip unsupported place

* refine datatransfer

* fit for dirtributed program

* fix dependencpy

* fix some ut

paddle/fluid/framework/new_executor/data_transfer.cc

@@ -319,6 +319,7 @@ void ApplyDataTransform(const OpKernelType& expected_kernel_key,
     }
   }
 
+  bool transfered = false;
   DataTranferHelper data_transfer_helper(place, var_scope);
   for (auto& var_name_item : *ins_map_temp) {
     bool should_skip_input =
@@ -334,6 +335,9 @@ void ApplyDataTransform(const OpKernelType& expected_kernel_key,
       if (var->IsType<LoDTensor>() || var->IsType<phi::SelectedRows>()) {
         tensor_in = GetLoDTensorOrSelectedRowsValueFromVar(*var);
       } else if (var->IsType<LoDTensorArray>()) {
+        if (var->Get<LoDTensorArray>().size() == 0) {
+          continue;
+        }
         tensor_in =
             static_cast<const Tensor*>(&(var->Get<LoDTensorArray>()[0]));
       } else {
@@ -389,6 +393,7 @@ void ApplyDataTransform(const OpKernelType& expected_kernel_key,
       }
 
       if (is_transferred) {
+        transfered = true;
         // update RuntimeContext.inputs and original op_func_node inputs
         op_func_node->input_index[var_name_item.first][i] =
             var_scope->VarId(new_var_name);
@@ -426,11 +431,13 @@ void ApplyDataTransform(const OpKernelType& expected_kernel_key,
     }
   }
 
+  if (transfered) {
     // NOTE(zhiqiu): UPDATE the corresponding OeratorBase to make it consistent
     // with instruction. (hot fix, it is not good design here)
     op_func_node->operator_base_ =
         std::shared_ptr<OperatorBase>(framework::OpRegistry::CreateOp(
             op_base->Type(), new_ins, new_outs, op_base->Attrs()));
+  }
   op_func_node->no_data_transform_index = std::move(no_data_transform_index);
 }
 

paddle/fluid/framework/new_executor/interpretercore.cc

@@ -300,8 +300,16 @@ void InterpreterCore::Convert(
     gc_event_.emplace_back(vec_instruction_[i].DeviceContext().GetPlace(),
                            platform::GenerateDeviceEventFlag());
   }
+  bool inplaced = false;
+  for (auto inst : vec_instruction_) {
+    if (inst.OpBase()->Type() == "share_buffer" ||
+        inst.OpBase()->Type() == "share_data") {
+      VLOG(4) << "Already inplaced, skip inplace now.";
+      inplaced = true;
+    }
+  }
 
-  if (FLAGS_new_executor_use_inplace) {
+  if (FLAGS_new_executor_use_inplace && !inplaced) {
     BuildInplace();
   }
 
@@ -565,12 +573,11 @@ void InterpreterCore::RunNextInstructions(
     const Instruction& instr, std::queue<size_t>* reserved_next_ops,
     std::vector<std::atomic<size_t>>* atomic_deps,
     std::vector<std::atomic<size_t>>* atomic_var_ref) {
-  VLOG(4) << "atomic 1:" << atomic_deps;
   auto& next_instr = instr.NextInstructions();
 
   auto IsReady = [atomic_deps](size_t next_id) {
-    VLOG(4) << "atomic:" << atomic_deps << " " << &(*atomic_deps)[next_id]
-            << " " << next_id;
+    VLOG(4) << "atomic:" << atomic_deps << " op_id: " << next_id
+            << ", remain deps: " << (*atomic_deps)[next_id];
     return (*atomic_deps)[next_id].fetch_sub(1, std::memory_order_relaxed) == 1;
   };
 

paddle/fluid/framework/new_executor/interpretercore_util.cc

@@ -428,19 +428,19 @@ void build_op_func_list(const platform::Place& place,
       op_func_node.dev_ctx_ = dev_ctx;
       VLOG(3) << op_with_kernel->Type()
               << " : expected_kernel_key : " << expected_kernel_key;
-      auto exec_ctx =
-          ExecutionContext(*op_with_kernel, scope, *dev_ctx, runtime_context);
 
       // see OperatorWithKernel::RunImpl in operator.cc for why
       if (!(op->HasAttr(kAllKernelsMustComputeRuntimeShape) &&
             op->Attr<bool>(kAllKernelsMustComputeRuntimeShape))) {
         InterpretercoreInferShapeContext infer_shape_ctx(*op, runtime_context);
         // TODO(Aurelius84): In case of control flow ops, they are NOT
-        // inheritted
-        // from OperatorWithKernel.
+        // inheritted from OperatorWithKernel.
         op_with_kernel->Info().infer_shape_(&infer_shape_ctx);
       }
 
+      auto exec_ctx =
+          ExecutionContext(*op_with_kernel, scope, *dev_ctx, runtime_context);
+
       auto run_phi_kernel = false;
       if (phi::KernelFactory::Instance().HasCompatiblePhiKernel(
               op_with_kernel->Type())) {
@@ -476,7 +476,6 @@ void build_op_func_list(const platform::Place& place,
         op_with_kernel->BuildPhiKernelContext(runtime_context, dev_ctx,
                                               &pt_kernel_context);
         op_func_node.pt_kernel_ = op_with_kernel->PhiKernel();
-
         (*op_func_node.pt_kernel_)(&pt_kernel_context);
       } else {
         auto kernels_iter = all_op_kernels.find(op->Type());
@@ -711,6 +710,7 @@ std::map<int, std::list<int>> build_op_downstream_map(
   const std::set<std::string> random_op_set = {
       "bernoulli",      "poisson", "multinomial", "gaussian_random",
       "uniform_random", "randint", "randperm",    "exponential"};
+
   int dependence_op_idx = -1;
   for (size_t op_idx = 0; op_idx < vec_instruction.size(); ++op_idx) {
     if (random_op_set.count(vec_instruction[op_idx].OpBase()->Type())) {
@@ -721,6 +721,147 @@ std::map<int, std::list<int>> build_op_downstream_map(
     }
   }
 
+  // add dependency for communication op
+  const std::string communication_op_prefix = "c_";
+  dependence_op_idx = -1;
+  for (size_t op_idx = 0; op_idx < vec_instruction.size(); ++op_idx) {
+    if (vec_instruction[op_idx].OpBase()->Type().find(
+            communication_op_prefix) != std::string::npos) {
+      if (dependence_op_idx != -1) {
+        op2dependences[op_idx].insert(dependence_op_idx);
+      }
+      dependence_op_idx = op_idx;
+    }
+  }
+
+  // TODO(zhiqiu): there still some cases not handled
+  // add dependency for c_sync_comm_stream
+
+  // in program, we can add only one c_sync_comm_stream to sync all
+  // communication ops.
+  // c_allreduce_sum(a)
+  // c_allreduce_sum(b)
+  // c_allreduce_sum(c)
+  // c_sync_comm_stream(a)
+  const std::string kSyncComm = "c_sync_comm_stream";
+  dependence_op_idx = -1;
+  for (size_t op_idx = 0; op_idx < vec_instruction.size(); ++op_idx) {
+    if (vec_instruction[op_idx].OpBase()->Type() == kSyncComm) {
+      dependence_op_idx = op_idx;
+    } else {
+      if (dependence_op_idx != -1) {
+        VLOG(4) << "Add depend from "
+                << vec_instruction[dependence_op_idx].OpBase()->Type() << " to "
+                << vec_instruction[op_idx].OpBase()->Type();
+        op2dependences[op_idx].insert(dependence_op_idx);
+      }
+    }
+  }
+
+  // add dependency for coalesce_tensor
+  const std::string kCoalesceTensor = "coalesce_tensor";
+  for (size_t op_idx = 0; op_idx < vec_instruction.size(); ++op_idx) {
+    if (vec_instruction[op_idx].OpBase()->Type() == kCoalesceTensor) {
+      VLOG(4) << "Add depend for " << kCoalesceTensor << " " << op_idx;
+      auto fused_out = vec_instruction[op_idx].Outputs().at("FusedOutput")[0];
+      auto outputs = vec_instruction[op_idx].Outputs().at("Output");
+
+      auto is_read = [](const Instruction& inst, int var_id) -> bool {
+        for (auto pair : inst.Inputs()) {
+          for (auto item : pair.second) {
+            if (item == var_id) {
+              return true;
+            }
+          }
+        }
+        return false;
+      };
+
+      auto is_write = [](const Instruction& inst, int var_id) -> bool {
+        for (auto pair : inst.Outputs()) {
+          for (auto item : pair.second) {
+            if (item == var_id) {
+              return true;
+            }
+          }
+        }
+        return false;
+      };
+
+      // find first op that reads fused_out
+      auto first_read_fused_out_op = -1;
+      for (auto j = op_idx + 1; j < vec_instruction.size(); ++j) {
+        if (is_read(vec_instruction[j], fused_out)) {
+          first_read_fused_out_op = j;
+          break;
+        }
+      }
+
+      if (UNLIKELY(first_read_fused_out_op == -1)) {
+        VLOG(4) << "No op read FusedOutput";
+        continue;
+      }
+
+      // find ops that write 'outputs' between (op_index,
+      // first_read_fused_out_op)
+      // add depend: them->first_read_fused_out_op
+      for (auto j = op_idx + 1;
+           j < static_cast<size_t>(first_read_fused_out_op); ++j) {
+        for (auto var_id : outputs) {
+          if (is_write(vec_instruction[j], var_id)) {
+            op2dependences[first_read_fused_out_op].insert(j);
+            VLOG(4) << j << " -> " << first_read_fused_out_op;
+            VLOG(4)
+                << "Add depend from " << vec_instruction[j].OpBase()->Type()
+                << " to "
+                << vec_instruction[first_read_fused_out_op].OpBase()->Type();
+          }
+        }
+      }
+
+      // find first op read 'outputs' between (first_read_fused_out_op, end)
+      // add depned:  first_read_fused_out_op -> first op that reads 'outputs'
+
+      // special case for consecutive communication ops, for example,
+      // FusedOutput = c_sync_calc_stream(FusedOutput)
+      // FusedOutput= c_allreduce_sum(FusedOutput)
+      // FusedOutput = c_sync_comm_stream(FusedOutput)
+      // we should take the last one to add depned instead of
+      // 'first_read_fused_out_op'
+      size_t target = first_read_fused_out_op;
+      for (size_t j = first_read_fused_out_op + 1; j < vec_instruction.size();
+           ++j) {
+        if (j == target + 1 &&
+            vec_instruction[target].OpBase()->Type().find(
+                communication_op_prefix) != std::string::npos &&
+            vec_instruction[j].OpBase()->Type().find(communication_op_prefix) !=
+                std::string::npos) {
+          VLOG(4) << "Found consecutive communication ops, "
+                  << vec_instruction[target].OpBase()->Type() << " -> "
+                  << vec_instruction[j].OpBase()->Type();
+          target = j;
+          continue;
+        }
+
+        for (auto var_id : outputs) {
+          if (is_read(vec_instruction[j], var_id)) {
+            op2dependences[j].insert(target);
+            VLOG(4) << target << " -> " << j;
+            VLOG(4) << "Add depend from "
+                    << vec_instruction[target].OpBase()->Type() << " to "
+                    << vec_instruction[j].OpBase()->Type();
+          }
+        }
+      }
+    }
+  }
+  for (auto pair : op2dependences) {
+    VLOG(10) << pair.first << " Depends on " << pair.second.size();
+    std::ostringstream oss;
+    std::copy(pair.second.begin(), pair.second.end(),
+              std::ostream_iterator<int>(oss, " "));
+    VLOG(10) << oss.str();
+  }
   return std::move(get_downstream_map(op2dependences));
 }
 

paddle/fluid/memory/allocation/allocator_facade.cc

@@ -85,7 +85,7 @@ PADDLE_DEFINE_EXPORTED_bool(use_virtual_memory_auto_growth, false,
 // NOTE(Ruibiao): This FLAGS is just to be compatibled with
 // the old single-stream CUDA allocator. It will be removed
 // after StreamSafeCudaAllocator has been fully tested.
-PADDLE_DEFINE_EXPORTED_bool(use_stream_safe_cuda_allocator, false,
+PADDLE_DEFINE_EXPORTED_bool(use_stream_safe_cuda_allocator, true,
                             "Enable StreamSafeCUDAAllocator");
 
 PADDLE_DEFINE_EXPORTED_bool(use_cuda_managed_memory, false,

python/paddle/fluid/executor.py

@@ -394,19 +394,10 @@ def _is_enable_standalone_executor():
     Whether to use experimental executor `StandaloneExecutor`.
     """
     flag = False
-    # NOTE(zhiqiu): enable STANDALONE_EXECUTOR on windows platform by default
-    # It should be enabled on all platform in the future.
-
-    import platform
-    sysstr = platform.system().lower()
-    if sysstr == 'windows':
-        env_val = os.environ.get('FLAGS_USE_STANDALONE_EXECUTOR', 1)
-    else:
-        env_val = os.environ.get('FLAGS_USE_STANDALONE_EXECUTOR', None)
 
+    env_val = os.environ.get('FLAGS_USE_STANDALONE_EXECUTOR', '1')
     if env_val in [1, '1', True, 'True', 'true']:
         flag = True
-        warnings.warn("STANDALONE_EXECUTOR is enabled.")
 
     return flag
 
@@ -1386,6 +1377,10 @@ class Executor(object):
             program = pruned_program
 
         def _can_use_interpreter_core(program, place):
+            if core.is_compiled_with_npu() or core.is_compiled_with_xpu(
+            ) or core.is_compiled_with_mlu() or core.is_compiled_with_ipu():
+                return False
+
             compiled = isinstance(program, compiler.CompiledProgram)
             # NOTE(zhiqiu): do not support compiled program now
             if compiled:
@@ -1396,6 +1391,8 @@ class Executor(object):
                 # else:
                 #     return False
             else:
+                if isinstance(program._graph, compiler.CompiledProgram):
+                    return False
                 assert isinstance(program, Program)
                 return True
 

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -951,7 +951,7 @@ endif()
 if (WITH_DISTRIBUTE AND NOT APPLE)
     if(WITH_GPU OR WITH_ROCM)
         set_tests_properties(test_c_comm_init_op PROPERTIES TIMEOUT 120)
-        set_tests_properties(test_dist_mnist_gradient_merge PROPERTIES TIMEOUT 160)
+        set_tests_properties(test_dist_mnist_gradient_merge PROPERTIES TIMEOUT 360)
     endif()
 endif()
 
@@ -1033,7 +1033,7 @@ set_tests_properties(test_parallel_executor_seresnext_with_reduce_gpu PROPERTIES
 set_tests_properties(test_dropout_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_argsort_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_gather_nd_op PROPERTIES TIMEOUT 120)
-set_tests_properties(test_nn_grad PROPERTIES TIMEOUT 120)
+set_tests_properties(test_nn_grad PROPERTIES TIMEOUT 120 ENVIRONMENT FLAGS_USE_STANDALONE_EXECUTOR=0)
 set_tests_properties(test_elementwise_sub_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_row_conv_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_parallel_executor_seresnext_with_fuse_all_reduce_gpu PROPERTIES TIMEOUT 120)
@@ -1072,7 +1072,7 @@ set_tests_properties(test_space_to_depth_op PROPERTIES TIMEOUT 200)
 set_tests_properties(test_dyn_rnn PROPERTIES TIMEOUT 120)
 set_tests_properties(test_sgd_op PROPERTIES TIMEOUT 250)
 set_tests_properties(test_parallel_executor_seresnext_base_gpu PROPERTIES TIMEOUT 120)
-set_tests_properties(test_norm_nn_grad PROPERTIES TIMEOUT 120)
+set_tests_properties(test_norm_nn_grad PROPERTIES TIMEOUT 120 ENVIRONMENT FLAGS_USE_STANDALONE_EXECUTOR=0)
 set_tests_properties(test_matrix_nms_op PROPERTIES TIMEOUT 120)
 set_tests_properties(test_generator_dataloader PROPERTIES TIMEOUT 120)
 set_tests_properties(test_partial_concat_op PROPERTIES TIMEOUT 120)

python/paddle/fluid/tests/unittests/collective_reducescatter.py

@@ -48,6 +48,7 @@ class TestCollectiveReduceScatter(TestCollectiveRunnerBase):
             tindata = layers.data(
                 name="tindata", shape=[10, 1000], dtype='float32')
             toutdata = fluid.layers.collective._c_reducescatter(tindata, nranks)
+            toutdata = fluid.layers.collective._c_sync_comm_stream(toutdata, 0)
             return toutdata
 
 

python/paddle/fluid/tests/unittests/distributed_passes/dist_pass_test_base.py

@@ -32,7 +32,7 @@ def prepare_python_path_and_return_module(path):
     assert filename.endswith(py_suffix), filename
 
     env_name = 'PYTHONPATH'
-    python_path = env_name
+    python_path = os.environ.get(env_name, '')
     if python_path:
         paths = [p for p in python_path.split(":") if p]
         if dirname not in paths:
@@ -41,6 +41,7 @@ def prepare_python_path_and_return_module(path):
     else:
         python_path = path
     os.environ[env_name] = python_path
+    print('GLOG_v=', os.environ.get('GLOG_v', None), flush=1)
     return filename[:-len(py_suffix)]
 
 

python/paddle/fluid/tests/unittests/ir/inference/CMakeLists.txt

@@ -91,7 +91,7 @@ set_tests_properties(test_trt_matmul_quant_dequant PROPERTIES TIMEOUT 100)
 set_tests_properties(test_trt_conv3d_op PROPERTIES TIMEOUT 60)
 set_tests_properties(test_trt_conv3d_transpose_op PROPERTIES TIMEOUT 60)
 set_tests_properties(test_trt_nearest_interp_v2_op PROPERTIES TIMEOUT 30)
-set_tests_properties(test_trt_multiclass_nms3_op PROPERTIES TIMEOUT 60)
+set_tests_properties(test_trt_multiclass_nms3_op PROPERTIES TIMEOUT 60 ENVIRONMENT FLAGS_USE_STANDALONE_EXECUTOR=0) 
 
 if (WITH_MKLDNN AND TENSORRT_FOUND AND WITH_GPU)
   set_tests_properties(test_emb_eltwise_layernorm_fuse_pass PROPERTIES TIMEOUT 120)

python/paddle/fluid/tests/unittests/test_nn_grad.py

@@ -27,6 +27,7 @@ paddle.enable_static()
 
 
 class TestSliceOpDoubleGradCheck(unittest.TestCase):
+    @prog_scope()
     def func(self, place):
         self.config()
 

python/paddle/fluid/tests/unittests/test_sync_batch_norm_op.py

@@ -30,6 +30,7 @@ from paddle.fluid import compiler
 from paddle.fluid import Program, program_guard
 
 from op_test import OpTest, _set_use_system_allocator
+from decorator_helper import prog_scope
 
 _set_use_system_allocator(True)
 
@@ -105,6 +106,7 @@ class TestSyncBatchNormOpTraining(unittest.TestCase):
                     sgd_opt.backward(out)
         return main, startup, [out, conv, bn]
 
+    @prog_scope()
     def _compare(self, place, layout, only_forward):
         """Compare results."""
         seed = 10