PaddlePaddle
diff --git a/‎paddle/phi/api/yaml/backward.yaml
Copy file name to clipboard
+11Lines changed: 11 additions & 0 deletions b/‎paddle/phi/api/yaml/backward.yaml
Copy file name to clipboard
+11Lines changed: 11 additions & 0 deletions
diff --git a/‎paddle/phi/api/yaml/generator/api_base.py
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions b/‎paddle/phi/api/yaml/generator/api_base.py
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions
diff --git a/‎paddle/phi/api/yaml/legacy_ops.yaml
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions b/‎paddle/phi/api/yaml/legacy_ops.yaml
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions
diff --git a/‎paddle/phi/api/yaml/ops.yaml
Copy file name to clipboard
+13Lines changed: 13 additions & 0 deletions b/‎paddle/phi/api/yaml/ops.yaml
Copy file name to clipboard
+13Lines changed: 13 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/cpu/gaussian_inplace_grad_kernel.cc
Copy file name to clipboard
+41Lines changed: 41 additions & 0 deletions b/‎paddle/phi/kernels/cpu/gaussian_inplace_grad_kernel.cc
Copy file name to clipboard
+41Lines changed: 41 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/cpu/gaussian_kernel.cc
Copy file name to clipboard
+31Lines changed: 31 additions & 0 deletions b/‎paddle/phi/kernels/cpu/gaussian_kernel.cc
Copy file name to clipboard
+31Lines changed: 31 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/gaussian_inplace_grad_kernel.h
Copy file name to clipboard
+29Lines changed: 29 additions & 0 deletions b/‎paddle/phi/kernels/gaussian_inplace_grad_kernel.h
Copy file name to clipboard
+29Lines changed: 29 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/gaussian_kernel.h
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions b/‎paddle/phi/kernels/gaussian_kernel.h
Copy file name to clipboard
+8Lines changed: 8 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/gpu/gaussian_inplace_grad_kernel.cu
Copy file name to clipboard
+44Lines changed: 44 additions & 0 deletions b/‎paddle/phi/kernels/gpu/gaussian_inplace_grad_kernel.cu
Copy file name to clipboard
+44Lines changed: 44 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/gpu/gaussian_kernel.cu
Copy file name to clipboard
+32Lines changed: 32 additions & 0 deletions b/‎paddle/phi/kernels/gpu/gaussian_kernel.cu
Copy file name to clipboard
+32Lines changed: 32 additions & 0 deletions
@@ -939,6 +939,17 @@
   composite : gather_nd_grad(x, index, out_grad, x_grad)
   no_need_buffer : x
 
+- backward_op : gaussian_inplace_grad
+  forward : gaussian_inplace(Tensor x, float mean=0, float std=1.0, int seed=0) -> Tensor(out)
+  args : (Tensor out_grad, float mean=0, float std=1.0, int seed=0)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [out_grad]
+  kernel :
+    func : gaussian_inplace_grad
+  inplace : (out_grad -> x_grad)
+
 - backward_op : gelu_grad
   forward : gelu(Tensor x,  bool approximate) -> Tensor(out)
   args : (Tensor x, Tensor out_grad,  bool approximate)
 
@@ -1223,6 +1223,7 @@ def gen_kernel_code(self, kernel_name, code_indent, inplace_flag=False):
             "unsqueeze",
             "reshape",
             "flatten",
+            "transpose",
         ]:
             i = 0
             for kernel_out in outputs_args:
 
@@ -1065,6 +1065,7 @@
     func : TransposeInferMeta
   kernel :
     func : transpose
+  inplace : (x -> out)
   backward : transpose_grad
 
 - op : tril
 
@@ -1052,6 +1052,19 @@
     func : gather_tree
     data_type : ids
 
+- op : gaussian_inplace
+  args: (Tensor x, float mean=0, float std=1.0, int seed=0)
+  output: Tensor(out)
+  infer_meta:
+    func: UnchangedInferMeta
+    param: [x]
+  kernel:
+    func: gaussian_inplace
+    data_type: x
+    backend : x
+  inplace: (x -> out)
+  backward: gaussian_inplace_grad
+
 - op : gelu
   args : (Tensor x,  bool approximate = false)
   output : Tensor(out)
 
@@ -0,0 +1,41 @@
+/* Copyright (c) 2023 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
+
+    https://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. */
+
+#include "paddle/phi/kernels/gaussian_inplace_grad_kernel.h"
+
+#include "paddle/phi/core/kernel_registry.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void GaussianInplaceGradKernel(const Context& ctx,
+                               const DenseTensor& out_grad UNUSED,
+                               float mean UNUSED,
+                               float std UNUSED,
+                               int seed UNUSED,
+                               DenseTensor* x_grad) {
+  if (x_grad) {
+    auto* data = ctx.template Alloc<T>(x_grad);
+    std::fill(data, data + x_grad->numel(), T(0));
+  }
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(gaussian_inplace_grad,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::GaussianInplaceGradKernel,
+                   float,
+                   double) {}
@@ -48,7 +48,38 @@ void GaussianKernel(const Context& dev_ctx,
   }
 }
 
+template <typename T, typename Context>
+void GaussianInplaceKernel(const Context& dev_ctx,
+                           const DenseTensor& x,
+                           float mean,
+                           float std,
+                           int seed,
+                           DenseTensor* out) {
+  T* data = dev_ctx.template Alloc<T>(out);
+  std::normal_distribution<T> dist(mean, std);
+
+  int64_t size = out->numel();
+  std::shared_ptr<std::mt19937_64> engine;
+  if (seed) {
+    engine = std::make_shared<std::mt19937_64>();
+    engine->seed(seed);
+  } else {
+    engine = dev_ctx.GetGenerator()->GetCPUEngine();
+  }
+
+  for (int64_t i = 0; i < size; ++i) {
+    data[i] = dist(*engine);
+  }
+}
+
 }  // namespace phi
 
 PD_REGISTER_KERNEL(
     gaussian, CPU, ALL_LAYOUT, phi::GaussianKernel, float, double) {}
+
+PD_REGISTER_KERNEL(gaussian_inplace,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::GaussianInplaceKernel,
+                   float,
+                   double) {}
@@ -0,0 +1,29 @@
+// Copyright (c) 2023 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
+//
+//     https://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.
+
+#pragma once
+
+#include "paddle/phi/core/dense_tensor.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void GaussianInplaceGradKernel(const Context& ctx,
+                               const DenseTensor& out_grad,
+                               float mean,
+                               float std,
+                               int seed,
+                               DenseTensor* x_grad);
+
+}  // namespace phi
@@ -29,4 +29,12 @@ void GaussianKernel(const Context& ctx,
                     DataType dtype,
                     DenseTensor* out);
 
+template <typename T, typename Context>
+void GaussianInplaceKernel(const Context& ctx,
+                           const DenseTensor& x,
+                           float mean,
+                           float std,
+                           int seed,
+                           DenseTensor* out);
+
 }  // namespace phi
@@ -0,0 +1,44 @@
+/* Copyright (c) 2021 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
+
+    https://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. */
+
+#include "paddle/phi/kernels/gaussian_inplace_grad_kernel.h"
+
+#include "paddle/phi/common/amp_type_traits.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/full_kernel.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void GaussianInplaceGradKernel(const Context& ctx,
+                               const DenseTensor& out_grad,
+                               float mean,
+                               float std,
+                               int seed,
+                               DenseTensor* x_grad) {
+  auto dims = vectorize(x_grad->dims());
+  float value = static_cast<float>(0.0f);
+  phi::FullKernel<T>(ctx, dims, value, phi::DataType::UNDEFINED, x_grad);
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(gaussian_inplace_grad,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::GaussianInplaceGradKernel,
+                   float,
+                   double,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16) {}
@@ -76,6 +76,29 @@ void GaussianKernel(const Context& dev_ctx,
   }
 }
 
+template <typename T, typename Context>
+void GaussianInpalceKernel(const Context& dev_ctx,
+                           const DenseTensor& x,
+                           float mean,
+                           float std,
+                           int seed,
+                           DenseTensor* out) {
+  dev_ctx.template Alloc<T>(out);
+  if (seed == 0) {
+    // use global Generator seed
+    using MT = typename phi::dtype::MPTypeTrait<T>::Type;
+    funcs::normal_distribution<MT> dist;
+    funcs::normal_transform<MT> trans(static_cast<MT>(mean),
+                                      static_cast<MT>(std));
+    funcs::distribution_and_transform<T>(dev_ctx, out, dist, trans);
+  } else {
+    // use OP seed
+    auto func =
+        GaussianGenerator<T>(static_cast<T>(mean), static_cast<T>(std), seed);
+    IndexKernel<T, GaussianGenerator<T>>(dev_ctx, out, func);
+  }
+}
+
 }  // namespace phi
 
 PD_REGISTER_KERNEL(gaussian,
@@ -86,3 +109,12 @@ PD_REGISTER_KERNEL(gaussian,
                    phi::dtype::bfloat16,
                    float,
                    double) {}
+
+PD_REGISTER_KERNEL(gaussian_inpalce,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::GaussianInpalceKernel,
+                   phi::dtype::float16,
+                   phi::dtype::bfloat16,
+                   float,
+                   double) {}