From d966301e620dd14dce82776d57402713e7d6f32b Mon Sep 17 00:00:00 2001
From: xiaoguoguo626807 <100397923+xiaoguoguo626807@users.noreply.github.com>
Date: Wed, 29 Mar 2023 14:28:54 +0800
Subject: [PATCH] tanh_double_grad_rules (#52192)

* tanh_double_grad_rules

* delete log got api_base

* modify composite yaml

* optimize rules
---
 .../generator/eager_gen.py                    |  5 ++++-
 .../composite_backward_api.h                  | 20 +++++++++++++++++++
 paddle/phi/api/yaml/backward.yaml             |  1 +
 paddle/phi/api/yaml/generator/api_base.py     |  1 +
 4 files changed, 26 insertions(+), 1 deletion(-)

diff --git a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
index 1325f7b1e5c..a8a4bcae481 100644
--- a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
+++ b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
@@ -63,7 +63,10 @@ black_ops_list = [
 # white ops list whose kernel can be deleted after performance analysis
 # original kernel and its derivative kernel can be deleted when composite_grad
 # kernel performs same to it.
-prim_white_list = ["matmul_double_grad"]
+prim_white_list = [
+    "matmul_double_grad",
+    "tanh_double_grad",
+]
 
 # dict of special api that forward api's output will affect bacward api's output
 # bacward api's output usually affected by backward api's input
diff --git a/paddle/fluid/prim/api/composite_backward/composite_backward_api.h b/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
index a90160f260a..35059ab2ddf 100644
--- a/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
+++ b/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
@@ -125,6 +125,26 @@ void tanh_grad(const Tensor& out, const Tensor& grad_out, Tensor* grad_x) {
   set_output<T>(grad_x_tmp, grad_x);
 }
 
+template <typename T>
+void tanh_double_grad(const Tensor& out,
+                      const Tensor& grad_out,
+                      const Tensor& grad_x_grad,
+                      Tensor* out_grad,
+                      Tensor* grad_out_grad) {
+  // tanh grad grad : ddout = (1 - out^2) * ddx, dout = - (dout_old * 2 * out *
+  // ddx)
+  auto out_m_grad_x_grad = out * grad_x_grad;
+  if (out_grad) {
+    auto out_grad_tmp = -2 * grad_out * out_m_grad_x_grad;
+    set_output<T>(out_grad_tmp, out_grad);
+  }
+
+  if (grad_out_grad) {
+    auto grad_out_grad_tmp = grad_x_grad - out * out_m_grad_x_grad;
+    set_output<T>(grad_out_grad_tmp, grad_out_grad);
+  }
+}
+
 template <typename T>
 void reshape_grad(const Tensor& x, const Tensor& grad_out, Tensor* grad_x) {
   if (grad_x) {
diff --git a/paddle/phi/api/yaml/backward.yaml b/paddle/phi/api/yaml/backward.yaml
index bc85748e23a..f4be9a37c62 100644
--- a/paddle/phi/api/yaml/backward.yaml
+++ b/paddle/phi/api/yaml/backward.yaml
@@ -1655,6 +1655,7 @@
     param : [out, out]
   kernel :
     func : tanh_double_grad
+  composite : tanh_double_grad(out, grad_out, grad_x_grad, out_grad, grad_out_grad)
   backward : tanh_triple_grad
   inplace : (grad_x_grad -> grad_out_grad)
 
diff --git a/paddle/phi/api/yaml/generator/api_base.py b/paddle/phi/api/yaml/generator/api_base.py
index 3f682ecfec1..93ead4586e8 100644
--- a/paddle/phi/api/yaml/generator/api_base.py
+++ b/paddle/phi/api/yaml/generator/api_base.py
@@ -1335,4 +1335,5 @@ PADDLE_API {self.get_return_type()} {self.api}({params_code}) {{
             else:
                 invoke_code = self.invoke
                 params_code = self.get_define_args()
+
             return self.gene_invoke_code(invoke_code, params_code)
-- 
GitLab