【PRIM】Support use operator's output metadata info in constructing static...

【PRIM】Support use operator's output metadata info  in constructing static backward composite (#50043)

* [prim] support custom target_gradients

* support infershape after append one gradop

* [prim] add simple net test

* fix test_loop segment fault bug

* [prim] fix infer shape segment fault bug when output of grad_op_desc is empty

paddle/fluid/prim/api/generated/prim_api/static_prim_api.cc

@@ -110,6 +110,7 @@ Tensor unsqueeze<DescTensor>(const Tensor& x, const IntArray& axis) {
   op->SetAttr("axes", new_shape);
   op->CheckAttrs();
   op->InferVarType(block);
+  op->InferShape(*block);
   return out;
 }
 
@@ -209,7 +210,7 @@ Tensor sum<DescTensor>(const Tensor& x,
       "Out", {std::static_pointer_cast<prim::DescTensor>(out.impl())->Name()});
   op->CheckAttrs();
   op->InferVarType(block);
-  // TODO(jiabin, cxxly): This may have runtime shape skip infershape for now.
+  op->InferShape(*block);
   return out;
 }
 
@@ -232,7 +233,7 @@ Tensor reshape<DescTensor>(const Tensor& x, const IntArray& shape) {
       "Out", {std::static_pointer_cast<prim::DescTensor>(out.impl())->Name()});
   op->CheckAttrs();
   op->InferVarType(block);
-  // TODO(jiabin, cxxly): This may have runtime shape skip infershape for now.
+  op->InferShape(*block);
   return out;
 }
 

paddle/fluid/prim/api/manual/backward/composite_backward_api.h

@@ -232,8 +232,8 @@ void multiply_grad(const Tensor& x,
                    Tensor* y_grad) {
   if (x_grad) {
     auto x_grad_unreduce = multiply<T>(out_grad, y);
-    if (x.dims() != y.dims()) {
-      auto axes = get_reduce_dims(x.dims(), y.dims());
+    if (x_grad_unreduce.dims() != x.dims()) {
+      auto axes = get_reduce_dims_from_out(x_grad_unreduce.dims(), x.dims());
       if (!axes.size()) {
         set_output<T>(x_grad_unreduce, x_grad);
       } else {
@@ -252,8 +252,8 @@ void multiply_grad(const Tensor& x,
   }
   if (y_grad) {
     auto y_grad_unreduce = multiply<T>(out_grad, x);
-    if (y.dims() != x.dims()) {
-      auto axes = get_reduce_dims(y.dims(), x.dims());
+    if (y_grad_unreduce.dims() != y.dims()) {
+      auto axes = get_reduce_dims_from_out(y_grad_unreduce.dims(), y.dims());
       if (!axes.size()) {
         set_output<T>(y_grad_unreduce, y_grad);
       } else {

paddle/fluid/prim/utils/static/composite_grad_desc_maker.h

@@ -318,6 +318,7 @@ class CompositeGradOpMakerBase {
       grad_var_name = framework::kEmptyVarName;
       if (drop_empty_grad) return nullptr;
     }
+
     if (original_block_->HasVar(grad_var_name)) {
       // Copy Var from original block to active block, or create a new one.
       CopyVarFromOrig(grad_var_name);
@@ -333,6 +334,12 @@ class CompositeGradOpMakerBase {
     auto grad_var_name = framework::GradVarName(var_name);
     (*this->grad_to_var_)[grad_var_name] = var_name;
     VLOG(8) << "Valid gradients: " << grad_var_name;
+
+    auto target_grad = StaticCompositeContext::Instance().GetTargetGradName();
+    if (target_grad.find(grad_var_name) != target_grad.end()) {
+      grad_var_name = target_grad.at(grad_var_name);
+    }
+
     if (original_block_->HasVar(grad_var_name)) {
       // Copy Var from original block to active block, or create a new one.
       CopyVarFromOrig(grad_var_name);
@@ -421,7 +428,11 @@ class CompositeGradOpMakerBase {
                      return g_name;
                    });
     std::vector<framework::VarDesc*> grad_out;
-    for (const auto& name : ret_val) {
+    for (auto name : ret_val) {
+      auto target_grad = StaticCompositeContext::Instance().GetTargetGradName();
+      if (target_grad.find(name) != target_grad.end()) {
+        name = target_grad.at(name);
+      }
       // TODO(jiabin): Will this cause fill zeros error?
       if (original_block_->HasVar(name)) {
         // Copy Var from original block to active block, or create a new one.

paddle/fluid/prim/utils/static/static_global_utils.h

@@ -69,12 +69,21 @@ class StaticCompositeContext {
     enable_bwd_prim_ = enable_prim;
   }
 
+  void SetTargetGradName(const std::map<std::string, std::string>& m) {
+    target_grad_name_ = m;
+  }
+
+  std::map<std::string, std::string> GetTargetGradName() {
+    return target_grad_name_;
+  }
+
  private:
   StaticCompositeContext()
       : current_block_desc_(nullptr), generator_(new UniqueNameGenerator()) {}
 
   framework::BlockDesc* current_block_desc_;
   std::unique_ptr<UniqueNameGenerator> generator_;
+  std::map<std::string, std::string> target_grad_name_;
   static thread_local bool enable_bwd_prim_;
   static thread_local bool enable_fwd_prim_;
   static StaticCompositeContext* static_composite_context_;

paddle/fluid/prim/utils/utils.cc

@@ -38,5 +38,11 @@ void PrimCommonUtils::SetFwdPrimEnabled(bool enable_prim) {
 void PrimCommonUtils::SetAllPrimEnabled(bool enable_prim) {
   return StaticCompositeContext::Instance().SetAllPrimEnabled(enable_prim);
 }
+
+void PrimCommonUtils::SetTargetGradName(
+    const std::map<std::string, std::string>& m) {
+  StaticCompositeContext::Instance().SetTargetGradName(m);
+}
+
 }  // namespace prim
 }  // namespace paddle

paddle/fluid/prim/utils/utils.h

@@ -14,6 +14,9 @@
 
 #pragma once
 
+#include <map>
+#include <string>
+
 namespace paddle {
 namespace prim {
 class PrimCommonUtils {
@@ -23,6 +26,7 @@ class PrimCommonUtils {
   static bool IsFwdPrimEnabled();
   static void SetFwdPrimEnabled(bool enabled);
   static void SetAllPrimEnabled(bool enabled);
+  static void SetTargetGradName(const std::map<std::string, std::string>& m);
 };
 }  // namespace prim
 }  // namespace paddle

paddle/fluid/pybind/pybind.cc

@@ -673,6 +673,8 @@ PYBIND11_MODULE(libpaddle, m) {
         &paddle::prim::PrimCommonUtils::IsFwdPrimEnabled);
   m.def("__set_all_prim_enabled",
         &paddle::prim::PrimCommonUtils::SetAllPrimEnabled);
+  m.def("_set_prim_target_grad_name",
+        &paddle::prim::PrimCommonUtils::SetTargetGradName);
   m.def("set_num_threads", &platform::SetNumThreads);
 
   m.def("disable_signal_handler", &DisableSignalHandler);

python/paddle/fluid/backward.py

@@ -1337,6 +1337,18 @@ def _append_backward_ops_(
         rename_var_map = {}
     assert isinstance(rename_var_map, dict)
 
+    if core._is_bwd_prim_enabled():
+        composite_block = program.clone().current_block()
+        # Infer shape for operators whose output haven't been created.
+        for op in composite_block.ops:
+            if not all(
+                tuple(
+                    composite_block._find_var_recursive(arg)
+                    for arg in op.output_arg_names
+                )
+            ):
+                infershape_for_composite(composite_block, op.desc)
+
     # add grad_op_desc by reversed ops
     for op in reversed(ops):
         grad_sub_block_list = []
@@ -1365,7 +1377,38 @@ def _append_backward_ops_(
 
             program._rollback()
             grad_sub_block_list.append(grad_sub_block.desc)
+        # In primitive mode, raw phi GradOp will be split into multiple small
+        # primitive operators, and the split rules are defined in c++ level,
+        # see detials: paddle/fluid/prim/api/manual/backward/composite_backward_api.h
+        # It means that the output's shape and dtype of previous operators which
+        # maybe used as the input of next operators must be known. Therefore,
+        # we infer shape and dtype in a sandbox block(named composite_block) for
+        # used in c++ level.
+        # For example:
+        #   forward:
+        #       z = multiply(x, y) //maybe broadcast in kernel
+        #   bcckward:
+        #       x_grad_unreduce = z_grad * y // maybe unreduce
+        #       reduced_axes = get_reduced_axes(x_grad.shape, x.shape) // need known shape
+        #       x_grad = reduce_sum(x_grad_unreduce)
+        grad_op_desc = []
+        op_grad_to_var = {}
+        if core._is_bwd_prim_enabled():
+
+            def find_op_index(block_desc, cur_op_desc):
+                for idx in range(block_desc.op_size()):
+                    if cur_op_desc == block_desc.op(idx):
+                        return idx
+                return -1
 
+            grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
+                composite_block.desc.op(find_op_index(block.desc, op.desc)),
+                no_grad_dict[composite_block.idx],
+                grad_sub_block_list,
+            )
+            for desc in grad_op_desc:
+                infershape_for_composite(composite_block, desc)
+        else:
             # Getting op's corresponding grad_op
             grad_op_desc, op_grad_to_var = core.get_grad_op_desc(
                 op.desc, no_grad_dict[block.idx], grad_sub_block_list
@@ -1655,7 +1698,43 @@ def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map):
         block.desc._remove_op(op_idx, op_idx + 1)
 
 
-def _rename_grad_(block, start_op_idx, grad_to_var, target_grad_map):
+def infershape_for_composite(block, grad_op_desc):
+    # pruning empty output
+    if len(grad_op_desc.output_arg_names()) == 0:
+        return
+
+    # append op to block
+    op_desc = block.desc.append_op()
+    op_desc.copy_from(grad_op_desc)
+    op_desc._set_attr(
+        core.op_proto_and_checker_maker.kOpRoleAttrName(),
+        core.op_proto_and_checker_maker.OpRole.Backward,
+    )
+
+    # create output var
+    new_vars = set()
+    # create new gradient variables
+    for grad_var_name in op_desc.output_arg_names():
+        if not (
+            block.desc.has_var_recursive(grad_var_name.encode())
+            or grad_var_name == core.empty_var_name()
+        ):
+            block.desc.var(grad_var_name.encode())
+            new_vars.add(grad_var_name)
+
+    # infer shape and infer dthype
+    op_desc.check_attrs()
+    op_desc.infer_var_type(block.desc)
+    op_desc.infer_shape(block.desc)
+
+    for arg in op_desc.output_arg_names():
+        if arg in new_vars:
+            _infer_var_data_type_shape_(arg, block)
+
+
+def _rename_grad_(
+    block, start_op_idx, grad_to_var, target_grad_map, skip_rename_var_list
+):
     var_map = copy.copy(target_grad_map)
     for op_idx in range(start_op_idx, block.desc.op_size()):
         op_desc = block.desc.op(op_idx)
@@ -1667,6 +1746,8 @@ def _rename_grad_(block, start_op_idx, grad_to_var, target_grad_map):
             if "@GRAD" not in name:
                 continue
             if block.desc.find_var(name.encode("ascii")):
+                if name in skip_rename_var_list:
+                    continue
                 new_name = unique_name.generate(name)
                 op_desc._rename_output(name, new_name)
                 var_map[name] = new_name
@@ -1993,7 +2074,7 @@ def append_backward(
     # Because append_backward may be called multiple times,
     # we need rename the internal gradient variables so that they have
     # different names.
-    _rename_grad_(target_grad_block, fwd_op_num, grad_to_var, {})
+    _rename_grad_(target_grad_block, fwd_op_num, grad_to_var, {}, [])
 
     _append_backward_vars_(
         target_grad_block, fwd_op_num, grad_to_var, grad_info_map
@@ -2297,33 +2378,24 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
 
     target_grad_map = {}
     rename_var_map = {}
+    skip_rename_var_list = []
     for i, grad in enumerate(target_gradients):
         target = targets[i]
         grad_name = _append_grad_suffix_(target.name)
         if grad is None:
-            target_shape = target.name + '_shape'
-            block.desc.append_op().copy_from(
-                _create_op_desc_(
-                    "shape",
-                    {'Input': [target.name]},
-                    {"Out": [target_shape]},
-                    {},
-                )
-            )
-            input_grad_names_set.add(target_shape)
             op_desc = _create_op_desc_(
-                "fill_constant",
-                {"ShapeTensor": [target_shape]},
+                "fill_any_like",
+                {"X": [target.name]},
                 {"Out": [grad_name]},
                 {
-                    "shape": target.shape,
                     "value": 1.0,
                     "dtype": target.dtype,
                 },
             )
-
             block.desc.append_op().copy_from(op_desc)
+            block.program._sync_with_cpp()
             input_grad_names_set.add(grad_name)
+            skip_rename_var_list.append(grad_name)
         else:
             if target.block.idx != block_idx or target.block.program != prog:
                 raise ValueError("all targets must be in the same block")
@@ -2336,6 +2408,9 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
             input_grad_names_set.add(grad.name)
             rename_var_map[grad_name] = grad.name
 
+    if core._is_bwd_prim_enabled():
+        core._set_prim_target_grad_name(target_grad_map)
+
     # For double backward, input_grad_names is used for filter
     # some non-used gradients op. rename_var_map is used to
     # associate target_grad var name with first grad_op input name.
@@ -2378,7 +2453,9 @@ def calc_gradient(targets, inputs, target_gradients=None, no_grad_set=None):
     # Because calc_gradient may be called multiple times,
     # we need rename the internal gradient variables so that they have
     # different names.
-    _rename_grad_(block, fwd_op_num, grad_to_var, target_grad_map)
+    _rename_grad_(
+        block, fwd_op_num, grad_to_var, target_grad_map, skip_rename_var_list
+    )
 
     _append_backward_vars_(block, fwd_op_num, grad_to_var, grad_info_map)
     prog._sync_with_cpp()

python/paddle/fluid/core.py

@@ -313,6 +313,7 @@ try:
     from .libpaddle import __set_fwd_prim_enabled
     from .libpaddle import _is_fwd_prim_enabled
     from .libpaddle import __set_all_prim_enabled
+    from .libpaddle import _set_prim_target_grad_name
 
     # custom devivce
     from .libpaddle import _get_current_custom_device_stream

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

@@ -9,3 +9,4 @@ foreach(TEST_OP ${TEST_OPS})
 endforeach()
 
 add_subdirectory(prim)
+add_subdirectory(model)

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

+file(
+  GLOB TEST_OPS
+  RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}"
+  "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+foreach(TEST_OP ${TEST_OPS})
+  py_test_modules(${TEST_OP} MODULES ${TEST_OP} ENVS ${GC_ENVS})
+endforeach()

python/paddle/fluid/tests/unittests/prim/model/test_comp_model_simple_net.py

+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+import parameterized as param
+
+import paddle
+from paddle.fluid import core, framework
+
+
+@param.parameterized_class(
+    ('name', 'primals', 'stop_gradients', 'cotangents', 'dtype'),
+    (
+        (
+            'test_normal_case',
+            (np.random.rand(2, 3, 4), np.random.rand(2, 3, 4)),
+            (False, False),
+            (np.random.rand(2, 3, 4),),
+            np.float32,
+        ),
+        (
+            'test_broadcast_diff_rank',
+            (np.random.rand(2, 3, 1, 4), np.random.rand(3, 3, 4)),
+            (False, False),
+            (np.random.rand(2, 3, 3, 4),),
+            np.float32,
+        ),
+        (
+            'test_broadcast_same_rank',
+            (np.random.rand(2, 3, 1, 4), np.random.rand(2, 1, 3, 4)),
+            (False, False),
+            (np.random.rand(2, 3, 3, 4),),
+            np.float32,
+        ),
+        (
+            'test_stop_gradient',
+            (np.random.rand(2, 3, 1, 4), np.random.rand(2, 1, 3, 4)),
+            (False, True),
+            (np.random.rand(2, 3, 3, 4),),
+            np.float32,
+        ),
+    ),
+)
+class TestMultiplyGradComp(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.primals = tuple(primal.astype(cls.dtype) for primal in cls.primals)
+        cls.cotangents = tuple(co.astype(cls.dtype) for co in cls.cotangents)
+
+    def setUp(self):
+        paddle.enable_static()
+
+    def tearDown(self):
+        paddle.disable_static()
+
+    def as_tuple(self, x):
+        return (x,) if isinstance(x, framework.Variable) else x
+
+    def net(self):
+        primals, cotangents = self.primals, self.cotangents
+        mp, sp = paddle.static.Program(), paddle.static.Program()
+        with paddle.static.program_guard(mp, sp):
+            primals = tuple(
+                paddle.static.data(f'primal{i}', primal.shape, primal.dtype)
+                for i, primal in enumerate(primals)
+            )
+            for primal, flag in zip(primals, self.stop_gradients):
+                primal.stop_gradient = flag
+            cotangents = tuple(
+                paddle.static.data(f'cotangent{i}', co.shape, co.dtype)
+                for i, co in enumerate(cotangents)
+            )
+            out = self.as_tuple(paddle.tanh(paddle.multiply(*primals)))
+            grads = paddle.static.gradients(out, primals)
+        exe = paddle.static.Executor()
+        exe.run(sp)
+        return exe.run(
+            program=mp,
+            feed={
+                **{
+                    f'primal{i}': primal
+                    for i, primal in enumerate(self.primals)
+                },
+                **{f'cotangent{i}': co for i, co in enumerate(self.cotangents)},
+            },
+            fetch_list=[g for g in grads if g is not None],
+        )
+
+    def test_comp(self):
+        core._set_prim_backward_enabled(True)
+        actual = self.net()
+
+        core._set_prim_backward_enabled(False)
+        desired = self.net()
+
+        self.assertEqual(len(actual), len(desired))
+        for i, j in zip(actual, desired):
+            np.testing.assert_allclose(
+                i,
+                j,
+                rtol=1e-6,
+                atol=0,
+            )
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/prim/prim/vjp/static/test_comp_multiply_grad.py

@@ -56,7 +56,7 @@ from paddle.fluid import core, framework
             'test_reduce_axe_empty',
             (np.random.rand(2, 3, 3, 4), np.random.rand(2, 1, 3, 4)),
             (False, False),
-            (np.random.rand(2, 1, 3, 1),),
+            (np.random.rand(2, 3, 3, 4),),
             np.float32,
         ),
     ),
@@ -91,7 +91,7 @@ class TestMultiplyGradComp(unittest.TestCase):
                 for i, co in enumerate(cotangents)
             )
             out = self.as_tuple(paddle.multiply(*primals))
-            grads = paddle.static.gradients(out, primals)
+            grads = paddle.static.gradients(out, primals, cotangents)
         exe = paddle.static.Executor()
         exe.run(sp)
         return exe.run(

python/paddle/fluid/tests/unittests/prim/test_comp_get_grad_op_desc_prim_enabled.py

@@ -75,7 +75,6 @@ class TestGetGradOpDescPrimEnabled(unittest.TestCase):
                 self.fwd, self.no_grad_var, self.grad_sub_block
             )[0]
         )
-        print(actual)
         self.assertEquals(actual, self.desired_ops)
         core._set_prim_backward_enabled(False)
 

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

@@ -105,7 +105,7 @@ class TestRunProgram(unittest.TestCase):
         )
         backward_program = _add_build_strategy_for(
             program,
-            main_program.desc.block(0).op_size() + 2,
+            main_program.desc.block(0).op_size() + 1,
             program.desc.block(0).op_size(),
         )
 

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

@@ -131,7 +131,7 @@ class RunProgramOpTest(unittest.TestCase):
         forward_program = _add_build_strategy_for(program, 0, forward_op_num)
         backward_program = _add_build_strategy_for(
             program,
-            forward_op_num + 2 * output_num,
+            forward_op_num + output_num,
             program.desc.block(0).op_size(),
         )
         return forward_program.desc, backward_program.desc

python/paddle/jit/dy2static/partial_program.py

@@ -576,9 +576,7 @@ class PartialProgramLayer:
                 core.check_and_set_prim_all_enabled()
             backward.gradients(targets=targets, inputs=[])
 
-        start_idx = len(main_program.block(0).ops) + 2 * len(
-            self._outputs.tolist()
-        )
+        start_idx = len(main_program.block(0).ops) + len(self._outputs.tolist())
 
         self.prepare_gradient_aggregation(start_idx, main_program, program)
 
@@ -753,7 +751,7 @@ class PartialProgramLayer:
     ):
         # NOTE(dev): We apply build_strategy for backward firstly to
         # avoid skipping more gc variables.
-        backward_start_op_index = forward_end_op_index + 2 * len(
+        backward_start_op_index = forward_end_op_index + len(
             self._outputs.var_ids
         )
         backward_end_op_index = whole_program.desc.block(0).op_size()

python/paddle/jit/dy2static/utils.py

@@ -1512,12 +1512,11 @@ def _out_grad_names(program_desc, fwd_end_op_index, out_size):
     """
     names = []
     for i in range(
-        fwd_end_op_index + 1,
-        min(fwd_end_op_index + 2 * out_size, program_desc.block(0).op_size()),
-        2,
+        fwd_end_op_index,
+        min(fwd_end_op_index + out_size, program_desc.block(0).op_size()),
     ):
         op = program_desc.block(0).op(i)
-        if op.type() == 'fill_constant':
+        if op.type() == 'fill_any_like':
             var_name = op.output('Out')[0]
             names.append(var_name)
     return names

python/paddle/jit/translated_layer.py

@@ -373,7 +373,7 @@ class _ProgramHolder:
     @switch_to_static_graph
     def _create_backward_train_program(self):
         whole_program = _build_program_by_desc(self._train_program_desc)
-        start_op_index = self._infer_program_desc.block(0).op_size() + 2 * len(
+        start_op_index = self._infer_program_desc.block(0).op_size() + len(
             self._output_descs
         )
         end_op_index = whole_program.desc.block(0).op_size()