Merge pull request #1500 from zhanghonggeng/support_decompose_program

Support decompose pir program in paddle2onnx

paddle2onnx/command.py

@@ -92,6 +92,12 @@ def arg_parser():
         default=True,
         help="whether enable auto_update_opset, default is True",
     )
+    parser.add_argument(
+        "--enable_dist_prim_all",
+        type=ast.literal_eval,
+        default=False,
+        help="whether enable dist_prim_all, default is False",
+    )
     parser.add_argument(
         "--external_filename",
         type=str,
@@ -160,6 +166,7 @@ def main():
         save_file=args.save_file,
         opset_version=args.opset_version,
         auto_upgrade_opset=args.enable_auto_update_opset,
+        dist_prim_all=args.enable_dist_prim_all,
         verbose=True,
         enable_onnx_checker=args.enable_onnx_checker,
         enable_experimental_op=True,

paddle2onnx/convert.py

@@ -14,9 +14,96 @@
 
 import os
 import paddle
+import tempfile
 import paddle2onnx.paddle2onnx_cpp2py_export as c_p2o
 from paddle2onnx.utils import logging, paddle_jit_save_configs
 from contextlib import contextmanager
+from paddle.decomposition import decomp
+from paddle.base.executor import global_scope
+
+
+def load_model(model_filename):
+    """Loads the pir model from json file."""
+    assert os.path.exists(
+        model_filename
+    ), f"Model file {model_filename} does not exist."
+    if model_filename.endswith(".json"):
+        model_filename = model_filename[:-5]
+    return paddle.jit.load(model_filename)
+
+
+def compare_programs(original_program, new_program):
+    """Compares two pir programs' operations."""
+    original_ops = [op.name() for op in original_program.global_block().ops]
+    new_ops = [op.name() for op in new_program.global_block().ops]
+    return original_ops == new_ops
+
+
+def save_program(program, model_file):
+    """Saves the decomposed program to a file."""
+    place = paddle.CPUPlace()
+    exe = paddle.static.Executor(place)
+
+    tmp_dir = tempfile.mkdtemp()
+    filename = os.path.basename(model_file) + "_decompose"
+    filename_without_extension, _ = os.path.splitext(filename)
+    save_dir = os.path.join(tmp_dir, filename_without_extension)
+
+    # Find feed and fetch operations
+    feed, fetch = [], []
+    for op in program.global_block().ops:
+        if op.name() == "pd_op.feed":
+            feed.extend(op.results())
+        if op.name() == "pd_op.fetch" or op.name() == "builtin.shadow_output":
+            fetch.extend(op.operands_source())
+
+    with paddle.pir_utils.IrGuard():
+        paddle.static.save_inference_model(save_dir, feed, fetch, exe, program=program)
+
+    new_model_file = save_dir + ".json"
+    assert os.path.exists(
+        new_model_file
+    ), f"Pir Model file {new_model_file} does not exist."
+    logging.info(f"Decomposed Model file path: {new_model_file}")
+    return new_model_file
+
+
+def load_parameter(program):
+    params = []
+    opts = []
+    for var in program.list_vars():
+        if var.is_parameter or var.get_defining_op().name() == "builtin.parameter":
+            params.append(var)
+        elif var.persistable and var.get_defining_op().name() == "pd_op.data":
+            opts.append(var)
+    vars_list = params + opts
+    vars = [var for var in vars_list if var.persistable]
+
+    if vars is None:
+        return
+
+    place = paddle.CPUPlace()
+    exe = paddle.static.Executor(place)
+    paddle.base.libpaddle.pir.create_loaded_parameter(
+        vars, global_scope(), exe._default_executor
+    )
+
+
+def decompose_program(model_filename):
+    """Decomposes the given pir program."""
+    model = load_model(model_filename)
+    new_program = model.program().clone()
+    with decomp.prim_guard():
+        decomp.decompose_dist_program(new_program)
+
+    if compare_programs(model.program(), new_program):
+        return model_filename
+
+    # logging.info(f"Origin program: {model.program()}")
+    # logging.info(f"Decomposed program: {new_program}")
+
+    load_parameter(new_program)
+    return save_program(new_program, model_filename)
 
 
 def get_old_ir_guard():
@@ -39,6 +126,7 @@ def export(
     save_file=None,
     opset_version=7,
     auto_upgrade_opset=True,
+    dist_prim_all=False,
     verbose=True,
     enable_onnx_checker=True,
     enable_experimental_op=True,
@@ -49,6 +137,10 @@ def export(
     external_file="",
     export_fp16_model=False,
 ):
+    if paddle.get_flags("FLAGS_enable_pir_api")["FLAGS_enable_pir_api"]:
+        if dist_prim_all and auto_upgrade_opset:
+            model_filename = decompose_program(model_filename)
+
     deploy_backend = deploy_backend.lower()
     if custom_op_info is None:
         onnx_model_str = c_p2o.export(

paddle2onnx/mapper/activation/activation.cc

@@ -66,6 +66,7 @@ REGISTER_PIR_MAPPER(size, SizeMapper)
 REGISTER_MAPPER(softmax, SoftMaxMapper)
 REGISTER_PIR_MAPPER(softmax, SoftMaxMapper)
 REGISTER_MAPPER(softplus, ActivationMapper)
+REGISTER_PIR_MAPPER(rsqrt, RsqrtMapper)
 REGISTER_MAPPER(softshrink, SoftShrinkMapper)
 REGISTER_MAPPER(softsign, ActivationMapper)
 REGISTER_MAPPER(sqrt, ActivationMapper)

paddle2onnx/mapper/activation/activation.h

@@ -294,6 +294,11 @@ class RsqrtMapper : public Mapper {
               int64_t block_id,
               int64_t op_id)
       : Mapper(p, helper, block_id, op_id) {}
+  RsqrtMapper(const PaddlePirParser& p,
+              OnnxHelper* helper,
+              int64_t i,
+              bool c)
+      : Mapper(p, helper, i, c) {}
   void Opset7() override;
 };
 

paddle2onnx/mapper/tensor/elementwise.cc

@@ -34,7 +34,8 @@ REGISTER_PIR_MAPPER(elementwise_mod, ElementWiseModMapper)
 REGISTER_PIR_MAPPER(elementwise_floordiv, ElementWiseFloordivMapper)
 
 int32_t ElementwiseMapper::GetMinOpsetVersion(bool verbose) {
-  if (OpType() == "elementwise_min" || OpType() == "elementwise_max") {
+  if (convert_pir_op_name(OpType()) == "elementwise_min" ||
+      convert_pir_op_name(OpType()) == "elementwise_max") {
     Logger(verbose, 8) << RequireOpset(8) << std::endl;
     return 8;
   }

paddle2onnx/mapper/tensor/reduce_sum.cc

@@ -27,7 +27,7 @@ int32_t ReduceMapperSum::GetMinOpsetVersion(bool verbose) {
 void ReduceMapperSum::Opset13() {
   auto axis_name_ = "dim";
   GetAttr("keep_dim", &keep_dim_);
-if (!in_pir_mode) {
+  if (!in_pir_mode) {
     GetAttr("reduce_all", &reduce_all_);
     GetAttr("in_dtype", &in_dtype_);
     GetAttr("out_dtype", &out_dtype_);
@@ -47,6 +47,13 @@ if (!in_pir_mode) {
   }
 
   auto x_info = GetInput("X");
+  auto x_name = x_info[0].name;
+  auto x_tpye = x_info[0].dtype;
+  if (x_info[0].dtype == P2ODataType::BOOL) {
+    x_name = helper_->AutoCast(x_name, x_tpye, P2ODataType::INT32);
+    x_tpye = P2ODataType::INT32;
+  }
+
   std::string dims;
   if (IsAttrVar(axis_name_)) {
     auto info = GetAttrVar(axis_name_);
@@ -61,7 +68,7 @@ if (!in_pir_mode) {
   }
 
   // Add attribute
-  auto reduce_node = helper_->MakeNode("ReduceSum", {x_info[0].name, dims});
+  auto reduce_node = helper_->MakeNode("ReduceSum", {x_name, dims});
   AddAttribute(reduce_node, "keepdims", static_cast<int64_t>(keep_dim_));
   auto out_node_name = reduce_node->output(0);
 
@@ -73,7 +80,6 @@ if (!in_pir_mode) {
     out_node_name = helper_->Reshape(out_node_name, {-1});
   }
   auto out_info = GetOutput("Out");
-  helper_->AutoCast(out_node_name, out_info[0].name,
-                        x_info[0].dtype, out_info[0].dtype);
+  helper_->AutoCast(out_node_name, out_info[0].name, x_tpye, out_info[0].dtype);
 }
 }  // namespace paddle2onnx

tests/onnxbase.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import onnx
 from inspect import isfunction
 import logging
 from onnxruntime import InferenceSession
@@ -20,7 +21,7 @@
 import paddle
 import paddle.static as static
 import paddle2onnx.paddle2onnx_cpp2py_export as c_p2o
-from paddle2onnx.convert import dygraph2onnx
+from paddle2onnx.convert import dygraph2onnx, decompose_program
 import shutil
 from functools import wraps
 
@@ -231,6 +232,8 @@ def __init__(
         self.input_spec_shape = input_spec_shape
         self.input_dtype = []
         self.res_fict = {}
+        self.dist_prim_all = False
+        self.auto_upgrade_opset = False
 
         if isfunction(self.func):
             # self._func = self.BuildFunc(self.func, **self.kwargs_dict_dygraph["params_group1"])
@@ -351,10 +354,19 @@ def _mk_onnx_res(self, ver):
         """
         make onnx res
         """
+        model_path = os.path.join(
+            self.pwd, self.name, self.name + "_" + str(ver) + ".onnx"
+        )
+        model = onnx.load(model_path)
         sess = InferenceSession(
-            os.path.join(self.pwd, self.name, self.name + "_" + str(ver) + ".onnx"),
+            model_path,
             providers=["CPUExecutionProvider"],
         )
+
+        input_feed = {}
+        if len(model.graph.input) == 0:
+            return sess.run(output_names=None, input_feed=input_feed)
+
         ort_outs = sess.run(output_names=None, input_feed=self.input_feed)
         return ort_outs
 
@@ -473,7 +485,10 @@ def run(self):
             # clip extra
             model_file = None
             if paddle.get_flags("FLAGS_enable_pir_api")["FLAGS_enable_pir_api"]:
-                model_file = original_model_file
+                if self.dist_prim_all and self.auto_upgrade_opset:
+                    model_file = decompose_program(original_model_file)
+                else:
+                    model_file = original_model_file
             else:
                 model_file = os.path.join(self.name, "cliped_model.pdmodel")
                 self.clip_extra_program_only(original_model_file, model_file)