【pir】modify Call vjp interface for controlflow grad (#58277)

xiaoguoguo626807 · 0x45f · web-flow · commit 376ace4fbaad · 2023-10-23T15:19:40.000+08:00
* tmp

* [PIR]Migrate maximum into pir

* Polish code

* add ir_grad of static_gradient

* add test

* modify bug

* modify

* add mean fill_constant test

* modify cpu int32 test

* get_shape_tensor

* delete

* add default place

* modify grad

* modify call_vjp

* Update python/paddle/autograd/ir_backward.py

* fix name

* new ci

---------

Co-authored-by: 0x45f &lt;wangzhen45@baidu.com&gt;
diff --git a/paddle/fluid/pir/dialect/op_generator/op_interface_gen.py b/paddle/fluid/pir/dialect/op_generator/op_interface_gen.py
@@ -111,7 +111,7 @@
     }"""
 
 OP_VJP_DEFINE_TEMPLATE = """
-std::vector<std::vector<pir::OpResult>> {op_class_name}::Vjp(pir::Operation* op, const std::vector<std::vector<pir::Value>>& out_grads, const std::vector<std::vector<bool>>& stop_gradients){{
+std::vector<std::vector<pir::OpResult>> {op_class_name}::Vjp(pir::Operation* op, const std::vector<std::vector<pir::Value>>& inputs_, const std::vector<std::vector<pir::OpResult>>& outputs, const std::vector<std::vector<pir::Value>>& out_grads, const std::vector<std::vector<bool>>& stop_gradients){{
     {op_class_name} op_obj = op->dyn_cast<{op_class_name}>(); (void)op_obj;
 
     VLOG(6) << "Prepare inputs of {op_grad_name}";
@@ -317,7 +317,7 @@ def gen_exclusive_interface_str(op_info, op_info_items):
                 "  static void InferMeta( phi::InferMetaContext *infer_meta );"
             )
     if op_info.op_phi_name[0] not in vjp_interface_black_list:
-        exclusive_interface_str += "\n  static std::vector<std::vector<pir::OpResult>> Vjp(pir::Operation* op, const std::vector<std::vector<pir::Value>>& out_grads, const std::vector<std::vector<bool>>& stop_gradients);"
+        exclusive_interface_str += "\n  static std::vector<std::vector<pir::OpResult>> Vjp(pir::Operation* op, const std::vector<std::vector<pir::Value>>& inputs_, const std::vector<std::vector<pir::OpResult>>& outputs, const std::vector<std::vector<pir::Value>>& out_grads, const std::vector<std::vector<bool>>& stop_gradients);"
     if op_info.op_phi_name[0] in decomp_interface_declare_gen_op_list:
         exclusive_interface_str += "\n  static std::vector<std::vector<pir::OpResult>> Decomp(pir::Operation* op);"
     return exclusive_interface_str
diff --git a/paddle/fluid/pir/dialect/operator/interface/interface.cc b/paddle/fluid/pir/dialect/operator/interface/interface.cc
@@ -20,6 +20,8 @@ namespace paddle {
 namespace dialect {
 std::vector<std::vector<pir::OpResult>> VjpInterface::Vjp(
     pir::Operation* op,
+    const std::vector<std::vector<pir::Value>>& inputs,
+    const std::vector<std::vector<pir::OpResult>>& outputs,
     const std::vector<std::vector<pir::OpResult>>& out_grads,
     const std::vector<std::vector<bool>>& stop_gradients) {
   std::vector<std::vector<pir::Value>> out_grads_value;
@@ -30,7 +32,7 @@ std::vector<std::vector<pir::OpResult>> VjpInterface::Vjp(
     }
     out_grads_value.emplace_back(std::move(grad_value));
   }
-  return impl_->vjp_(op, out_grads_value, stop_gradients);
+  return impl_->vjp_(op, inputs, outputs, out_grads_value, stop_gradients);
 }
 }  // namespace dialect
 }  // namespace paddle
diff --git a/paddle/fluid/pir/dialect/operator/interface/vjp.h b/paddle/fluid/pir/dialect/operator/interface/vjp.h
@@ -22,11 +22,15 @@ class VjpInterface : public pir::OpInterfaceBase<VjpInterface> {
   struct Concept {
     explicit Concept(std::vector<std::vector<pir::OpResult>> (*vjp)(
         pir::Operation* op,
+        const std::vector<std::vector<pir::Value>>& inputs,
+        const std::vector<std::vector<pir::OpResult>>& outputs,
         const std::vector<std::vector<pir::Value>>& out_grads,
         const std::vector<std::vector<bool>>& stop_gradients))
         : vjp_(vjp) {}
     std::vector<std::vector<pir::OpResult>> (*vjp_)(
         pir::Operation* op,
+        const std::vector<std::vector<pir::Value>>& inputs,
+        const std::vector<std::vector<pir::OpResult>>& outputs,
         const std::vector<std::vector<pir::Value>>& out_grads,
         const std::vector<std::vector<bool>>& stop_gradients);
   };
@@ -35,9 +39,11 @@ class VjpInterface : public pir::OpInterfaceBase<VjpInterface> {
   struct Model : public Concept {
     static std::vector<std::vector<pir::OpResult>> Vjp(
         pir::Operation* op,
+        const std::vector<std::vector<pir::Value>>& inputs,
+        const std::vector<std::vector<pir::OpResult>>& outputs,
         const std::vector<std::vector<pir::Value>>& out_grads,
         const std::vector<std::vector<bool>>& stop_gradients) {
-      return ConcreteOp::Vjp(op, out_grads, stop_gradients);
+      return ConcreteOp::Vjp(op, inputs, outputs, out_grads, stop_gradients);
     }
 
     Model() : Concept(Vjp) {}
@@ -49,13 +55,17 @@ class VjpInterface : public pir::OpInterfaceBase<VjpInterface> {
 
   std::vector<std::vector<pir::OpResult>> Vjp(
       pir::Operation* op,
+      const std::vector<std::vector<pir::Value>>& inputs,
+      const std::vector<std::vector<pir::OpResult>>& outputs,
       const std::vector<std::vector<pir::Value>>& out_grads,
       const std::vector<std::vector<bool>>& stop_gradients) {
-    return impl_->vjp_(op, out_grads, stop_gradients);
+    return impl_->vjp_(op, inputs, outputs, out_grads, stop_gradients);
   }
 
   std::vector<std::vector<pir::OpResult>> Vjp(
       pir::Operation* op,
+      const std::vector<std::vector<pir::Value>>& inputs,
+      const std::vector<std::vector<pir::OpResult>>& outputs,
       const std::vector<std::vector<pir::OpResult>>& out_grads,
       const std::vector<std::vector<bool>>& stop_gradients);
 
diff --git a/paddle/fluid/pir/dialect/operator/ir/manual_op.h b/paddle/fluid/pir/dialect/operator/ir/manual_op.h
@@ -51,6 +51,8 @@ class AddNOp : public pir::Op<AddNOp,
   static void InferMeta(phi::InferMetaContext *infer_meta);
   static std::vector<std::vector<pir::OpResult>> Vjp(
       pir::Operation *op,
+      const std::vector<std::vector<pir::Value>> &inputs_,
+      const std::vector<std::vector<pir::OpResult>> &outputs,
       const std::vector<std::vector<pir::Value>> &out_grads,
       const std::vector<std::vector<bool>> &stop_gradients);
 };
diff --git a/paddle/fluid/pir/dialect/operator/ir/manual_op_vjp.cc b/paddle/fluid/pir/dialect/operator/ir/manual_op_vjp.cc
@@ -29,6 +29,8 @@ using IntArray = paddle::experimental::IntArray;
 
 std::vector<std::vector<pir::OpResult>> AddNOp::Vjp(
     pir::Operation* op,
+    const std::vector<std::vector<pir::Value>>& inputs_,
+    const std::vector<std::vector<pir::OpResult>>& outputs,
     const std::vector<std::vector<pir::Value>>& out_grads,
     const std::vector<std::vector<bool>>& stop_gradients) {
   AddNOp op_obj = op->dyn_cast<AddNOp>();
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
@@ -695,6 +695,8 @@ void BindVjp(pybind11::module *m) {
   m->def(
       "call_vjp",
       [](pir::Operation &fwd_op,
+         const std::vector<std::vector<pir::Value>> &inputs,
+         const std::vector<std::vector<pir::OpResult>> &outputs,
          const std::vector<std::vector<pir::OpResult>> &out_grads,
          const std::vector<std::vector<bool>> &stop_gradients) {
         py::list res;
@@ -704,8 +706,8 @@ void BindVjp(pybind11::module *m) {
             vjp_interface,
             phi::errors::InvalidArgument(
                 "The vjp function is not registered in %s op ", fwd_op.name()));
-        std::vector<std::vector<pir::OpResult>> vjp_res =
-            vjp_interface.Vjp(&fwd_op, out_grads, stop_gradients);
+        std::vector<std::vector<pir::OpResult>> vjp_res = vjp_interface.Vjp(
+            &fwd_op, inputs, outputs, out_grads, stop_gradients);
         PADDLE_ENFORCE_EQ(
             stop_gradients.size(),
             vjp_res.size(),
diff --git a/python/paddle/autograd/ir_backward.py b/python/paddle/autograd/ir_backward.py
@@ -328,7 +328,7 @@ def append_backward_ops(
     if op has grad_op, prepare its grad_op's inputs by value_to_valuegrad,
         eg:
         value_to_valuegrad[v3] = [[v3_g]];
-        v2_g = call_vjp(op3, [v3_g], [v2_stopgradient])
+        v2_g = call_vjp(op3, [[v2]], [[v3]],[[v3_g]], [[v2_stopgradient]])
 
 
     special pattern 1:
@@ -339,7 +339,7 @@ def append_backward_ops(
 
         v1 is inside python api, we don't describe it in backward process(state)
         so v1_grad is inside vjp, we don't describe it in backward process(state)
-        [[v11_g, v12_g], v2_g] = call_vjp(combine_op, [v3_g], [[v11_stopgradient, v12_stopgradient], v2_stop_gradient)
+        [[v11_g, v12_g], v2_g] = call_vjp(combine_op, [[v11, v12]], [[v3]],[[v3_g]], [[v11_stopgradient, v12_stopgradient], v2_stop_gradient])
 
 
         op_vjp is:
@@ -358,8 +358,9 @@ def append_backward_ops(
     else continue to next op.
     '''
 
-    def make_output_grad(op):
+    def make_output_with_output_grad(op):
         zero_flag = [False] * op.num_results()
+        outputs = []
         output_grads = []
         for i, value in enumerate(op.results()):
             if (
@@ -396,12 +397,15 @@ def make_output_grad(op):
                     # pattern case:
                     # this fwd_op's output is vectorType, it will split to
                     # Type by builtin.split op, so need get from split op's ouput
-                    split_zero_flag, split_output_grad = make_output_grad(
-                        value.first_use().owner()
-                    )
+                    (
+                        split_zero_flag,
+                        split_outputs,
+                        split_output_grad,
+                    ) = make_output_with_output_grad(value.first_use().owner())
                     zero_flag[i] = all(split_zero_flag)
                     grad_values = [value[0] for value in split_output_grad]
                     state.value_to_valuegrad[value] = [grad_values]
+                    outputs.append([info[0] for info in split_outputs])
                 else:
                     # first case:
                     # this fwd_op's output didn't used by other fwd_op,
@@ -424,16 +428,19 @@ def make_output_grad(op):
 
                     state.value_to_valuegrad[value] = [[grad_value]]
 
+                    outputs.append([value])
             output_grads.append(state.value_to_valuegrad[value][0])
 
-        return zero_flag, output_grads
+        return zero_flag, outputs, output_grads
 
-    def make_input_stopgradient(op):
+    def make_input_with_input_stopgradient(op):
+        inputs = []
         input_grad_stopgradients = []
         if op.name() == "builtin.combine":
             grad_semantic_info = [True for _ in range(op.num_operands())]
         else:
             grad_semantic_info = op.get_input_grad_semantics()
+
         for input, grad_semantic in zip(
             op.operands_source(), grad_semantic_info
         ):
@@ -443,16 +450,22 @@ def make_input_stopgradient(op):
                 input.get_defining_op() is not None
                 and input.get_defining_op().name() == "builtin.combine"
             ):
-                stop_gradient = make_input_stopgradient(input.get_defining_op())
+                (
+                    combine_inputs,
+                    combine_stop_gradient,
+                ) = make_input_with_input_stopgradient(input.get_defining_op())
+                inputs.append([info[0] for info in combine_inputs])
                 input_grad_stopgradients.append(
-                    [info[0] for info in stop_gradient]
+                    [info[0] for info in combine_stop_gradient]
                 )
             else:
+                inputs.append([input])
                 if input.get_defining_op() is None or input in no_grad_set:
                     input_grad_stopgradients.append([True])
                 else:
                     input_grad_stopgradients.append([False])
-        return input_grad_stopgradients
+
+        return inputs, input_grad_stopgradients
 
     def update_input_grad_map(op, input_grads):
         i = 0
@@ -494,20 +507,23 @@ def update_input_grad_map(op, input_grads):
     for op in clear_effective_forward_ops:
         if paddle.framework.core.has_vjp(op):
             # prepare output_grad
-            zero_flag, output_grads = make_output_grad(op)
+            zero_flag, outputs, output_grads = make_output_with_output_grad(op)
 
             # all(zero_flag) support this op has no contribution for grad
             # should be delete (prune sub_graph)
             if len(output_grads) == 0 or all(zero_flag):
                 continue
 
             # prepare input_grad stop_gradient info.
-            input_grad_stopgradients = make_input_stopgradient(op)
+            (
+                inputs,
+                input_grad_stopgradients,
+            ) = make_input_with_input_stopgradient(op)
 
             # create grad_op
             before_ops_num = len(block.ops)
             input_grads = paddle.framework.core.call_vjp(
-                op, output_grads, input_grad_stopgradients
+                op, inputs, outputs, output_grads, input_grad_stopgradients
             )
             after_ops_num = len(block.ops)
 
diff --git a/test/cpp/prim/test_vjp.cc b/test/cpp/prim/test_vjp.cc
@@ -59,12 +59,15 @@ TEST(VJP, TanhBackwardTest) {
       std::vector<int64_t>{1}, 2.0, phi::DataType::FLOAT32, phi::CPUPlace());
 
   std::vector<std::vector<bool>> stop_gradients{{false}};
+  std::vector<std::vector<pir::Value>> inputs{{op1.out()}};
+  std::vector<std::vector<pir::OpResult>> outputs{{op2.out()}};
   std::vector<std::vector<pir::Value>> out_grads{{op3.out()}};
 
   pir::OpInfo op2_info = ctx->GetRegisteredOpInfo("pd_op.tanh");
   auto tanh_vjp_interface_impl =
       op2_info.GetInterfaceImpl<paddle::dialect::VjpInterface>();
-  tanh_vjp_interface_impl->vjp_(op2.operation(), out_grads, stop_gradients);
+  tanh_vjp_interface_impl->vjp_(
+      op2.operation(), inputs, outputs, out_grads, stop_gradients);
 
   auto kernel_program = paddle::dialect::PdOpLowerToKernelPass(&program);
 
@@ -114,12 +117,15 @@ TEST(VJP, Tanh_BackwardTest) {
       std::vector<int64_t>{1}, 2.0, phi::DataType::FLOAT32, phi::CPUPlace());
 
   std::vector<std::vector<bool>> stop_gradients{{false}};
+  std::vector<std::vector<pir::Value>> inputs{{op1.out()}};
+  std::vector<std::vector<pir::OpResult>> outputs{{op2.out()}};
   std::vector<std::vector<pir::Value>> out_grads{{op3.out()}};
 
   pir::OpInfo op2_info = ctx->GetRegisteredOpInfo("pd_op.tanh_");
   auto tanh_vjp_interface_impl =
       op2_info.GetInterfaceImpl<paddle::dialect::VjpInterface>();
-  tanh_vjp_interface_impl->vjp_(op2.operation(), out_grads, stop_gradients);
+  tanh_vjp_interface_impl->vjp_(
+      op2.operation(), inputs, outputs, out_grads, stop_gradients);
 
   auto kernel_program = paddle::dialect::PdOpLowerToKernelPass(&program);
 
@@ -169,12 +175,15 @@ TEST(VJP, MeanBackwardTest) {
       std::vector<int64_t>{}, 1.0, phi::DataType::FLOAT32, phi::CPUPlace());
 
   std::vector<std::vector<bool>> stop_gradients{{false}};
+  std::vector<std::vector<pir::Value>> inputs{{op1.out()}};
+  std::vector<std::vector<pir::OpResult>> outputs{{op2.out()}};
   std::vector<std::vector<pir::Value>> out_grads{{op3.out()}};
 
   pir::OpInfo op2_info = ctx->GetRegisteredOpInfo("pd_op.mean");
   auto mean_vjp_interface_impl =
       op2_info.GetInterfaceImpl<paddle::dialect::VjpInterface>();
-  mean_vjp_interface_impl->vjp_(op2.operation(), out_grads, stop_gradients);
+  mean_vjp_interface_impl->vjp_(
+      op2.operation(), inputs, outputs, out_grads, stop_gradients);
 
   auto kernel_program = paddle::dialect::PdOpLowerToKernelPass(&program);
 
@@ -227,11 +236,14 @@ TEST(VJP, ConcatBackwardTest) {
   paddle::dialect::FullOp op4 = builder->Build<paddle::dialect::FullOp>(
       std::vector<int64_t>{2, 2}, 1.0, phi::DataType::FLOAT32, phi::CPUPlace());
   std::vector<std::vector<bool>> stop_gradients{{false, false}};
+  std::vector<std::vector<pir::Value>> inputs{{op1.out(), op1.out()}};
+  std::vector<std::vector<pir::OpResult>> outputs{{op3.out()}};
   std::vector<std::vector<pir::Value>> out_grads{{op4.out()}};
   pir::OpInfo op2_info = ctx->GetRegisteredOpInfo("pd_op.concat");
   auto concat_vjp_interface_impl =
       op2_info.GetInterfaceImpl<paddle::dialect::VjpInterface>();
-  concat_vjp_interface_impl->vjp_(op3.operation(), out_grads, stop_gradients);
+  concat_vjp_interface_impl->vjp_(
+      op3.operation(), inputs, outputs, out_grads, stop_gradients);
   auto kernel_program = paddle::dialect::PdOpLowerToKernelPass(&program);
 
   auto place = platform::CPUPlace();
@@ -291,12 +303,15 @@ TEST(VJP, AddBackwardTest) {
       std::vector<int64_t>{1}, 1.0, phi::DataType::FLOAT32, phi::CPUPlace());
 
   std::vector<std::vector<bool>> stop_gradients{{false}, {false}};
+  std::vector<std::vector<pir::Value>> inputs{{op1.out(), op2.out()}};
+  std::vector<std::vector<pir::OpResult>> outputs{{op3.out()}};
   std::vector<std::vector<pir::Value>> out_grads{{op4.out()}};
 
   pir::OpInfo op3_info = ctx->GetRegisteredOpInfo("pd_op.add");
   auto add_vjp_interface_impl =
       op3_info.GetInterfaceImpl<paddle::dialect::VjpInterface>();
-  add_vjp_interface_impl->vjp_(op3.operation(), out_grads, stop_gradients);
+  add_vjp_interface_impl->vjp_(
+      op3.operation(), inputs, outputs, out_grads, stop_gradients);
 
   auto kernel_program = paddle::dialect::PdOpLowerToKernelPass(&program);
 
@@ -356,13 +371,15 @@ TEST(VJP, Add_BackwardTest) {
       std::vector<int64_t>{1}, 1.0, phi::DataType::FLOAT32, phi::CPUPlace());
 
   std::vector<std::vector<bool>> stop_gradients{{false}, {false}};
+  std::vector<std::vector<pir::Value>> inputs{{op1.out(), op2.out()}};
+  std::vector<std::vector<pir::OpResult>> outputs{{op3.out()}};
   std::vector<std::vector<pir::Value>> out_grads{{op4.out()}};
 
   pir::OpInfo op3_info = ctx->GetRegisteredOpInfo("pd_op.add_");
   auto add_inplace_vjp_interface_impl =
       op3_info.GetInterfaceImpl<paddle::dialect::VjpInterface>();
   add_inplace_vjp_interface_impl->vjp_(
-      op3.operation(), out_grads, stop_gradients);
+      op3.operation(), inputs, outputs, out_grads, stop_gradients);
 
   auto kernel_program = paddle::dialect::PdOpLowerToKernelPass(&program);
 
@@ -422,13 +439,16 @@ TEST(VJP, SplitBackwardTest) {
       std::vector<int64_t>{1, 2}, 1.0, phi::DataType::FLOAT32, phi::CPUPlace());
 
   std::vector<std::vector<bool>> stop_gradients{{false}};
+  std::vector<std::vector<pir::Value>> inputs{{op1.out()}};
+  std::vector<std::vector<pir::OpResult>> outputs{{op3.outputs()}};
   std::vector<std::vector<pir::Value>> out_grads{{op3.result(0), op4.out()}};
   pir::OpInfo op2_info = ctx->GetRegisteredOpInfo("pd_op.split");
 
   auto concat_vjp_interface_impl =
       op2_info.GetInterfaceImpl<paddle::dialect::VjpInterface>();
 
-  concat_vjp_interface_impl->vjp_(op2.operation(), out_grads, stop_gradients);
+  concat_vjp_interface_impl->vjp_(
+      op2.operation(), inputs, outputs, out_grads, stop_gradients);
   auto kernel_program = paddle::dialect::PdOpLowerToKernelPass(&program);
 
   auto place = platform::CPUPlace();
diff --git a/test/ir/new_ir/test_ir_vjp.py b/test/ir/new_ir/test_ir_vjp.py
diff --git a/test/prim/pir_prim/test_vjp_prim.py b/test/prim/pir_prim/test_vjp_prim.py

Original file line number	Diff line number	Diff line change
`@@ -20,6 +20,8 @@ namespace paddle {`
`20`	`20`	`namespace dialect {`
`21`	`21`	`std::vector<std::vector<pir::OpResult>> VjpInterface::Vjp(`
`22`	`22`	`pir::Operation* op,`
	`23`	`+ const std::vector<std::vector<pir::Value>>& inputs,`
	`24`	`+ const std::vector<std::vector<pir::OpResult>>& outputs,`
`23`	`25`	`const std::vector<std::vector<pir::OpResult>>& out_grads,`
`24`	`26`	`const std::vector<std::vector<bool>>& stop_gradients) {`
`25`	`27`	`std::vector<std::vector<pir::Value>> out_grads_value;`
`@@ -30,7 +32,7 @@ std::vector<std::vector<pir::OpResult>> VjpInterface::Vjp(`
`30`	`32`	`}`
`31`	`33`	`out_grads_value.emplace_back(std::move(grad_value));`
`32`	`34`	`}`
`33`		`- return impl_->vjp_(op, out_grads_value, stop_gradients);`
	`35`	`+ return impl_->vjp_(op, inputs, outputs, out_grads_value, stop_gradients);`
`34`	`36`	`}`
`35`	`37`	`} // namespace dialect`
`36`	`38`	`} // namespace paddle`