chore: import upstream snapshot with attribution

paddle/phi/kernels/cpu/lrn_grad_kernel.cc

@@ -0,0 +1,96 @@
+// Copyright (c) 2024 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.
+
+#include "paddle/phi/kernels/impl/lrn_kernel_impl.h"
+
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "paddle/phi/backends/onednn/onednn_helper.h"
+#include "paddle/phi/core/kernel_registry.h"
+#include "paddle/phi/kernels/funcs/blas/blas.h"
+#include "paddle/phi/kernels/funcs/math_function.h"
+
+namespace phi {
+
+template <typename T>
+struct LRNGradFunctor<CPUContext, T> {
+  void operator()(const CPUContext& dev_ctx,
+                  const DenseTensor& x,
+                  const DenseTensor& out,
+                  const DenseTensor& mid,
+                  DenseTensor* x_g,
+                  const DenseTensor& out_g,
+                  int64_t N,
+                  int64_t C,
+                  int64_t H,
+                  int64_t W,
+                  int n,
+                  T alpha,
+                  T beta,
+                  const DataLayout data_layout) {
+    T ratio = -2 * alpha * beta;
+    auto x_g_e = EigenVector<T>::Flatten(*x_g);
+    x_g_e = x_g_e.constant(0.0);
+
+    auto e_x = EigenTensor<T, 4>::From(x);
+    auto e_x_g = EigenTensor<T, 4>::From(*x_g);
+    auto e_out = EigenTensor<T, 4>::From(out);
+    auto e_out_g = EigenTensor<T, 4>::From(out_g);
+    auto e_mid = EigenTensor<T, 4>::From(mid);
+
+    const int start = -(n - 1) / 2;
+    const int end = start + n;
+    for (int64_t m = 0; m < N; m++) {
+      for (int64_t i = 0; i < C; i++) {
+        auto offsets = Eigen::array<int64_t, 4>({{m, i, 0, 0}});
+        auto extents = Eigen::array<int64_t, 4>({{1, 1, H, W}});
+        if (data_layout == DataLayout::NHWC) {
+          offsets = Eigen::array<int64_t, 4>({{m, 0, 0, i}});
+          extents = Eigen::array<int64_t, 4>({{1, H, W, 1}});
+        }
+
+        auto i_x = e_x.slice(offsets, extents);
+        auto i_x_g = e_x_g.slice(offsets, extents);
+        auto i_out_g = e_out_g.slice(offsets, extents);
+        auto i_mid = e_mid.slice(offsets, extents);
+
+        i_x_g = i_mid.pow(-beta) * i_out_g;
+        for (int c = start; c < end; c++) {
+          int64_t ch = i + c;
+          if (ch < 0 || ch >= C) {
+            continue;
+          }
+
+          if (data_layout != DataLayout::NHWC) {
+            offsets = Eigen::array<int64_t, 4>({{m, ch, 0, 0}});
+          } else {
+            offsets = Eigen::array<int64_t, 4>({{m, 0, 0, ch}});
+          }
+          auto c_out = e_out.slice(offsets, extents);
+          auto c_mid = e_mid.slice(offsets, extents);
+          auto c_out_g = e_out_g.slice(offsets, extents);
+
+          i_x_g += ratio * c_out_g * c_out * i_x / c_mid;
+        }
+      }
+    }
+  }
+};
+template struct LRNGradFunctor<CPUContext, float>;
+template struct LRNGradFunctor<CPUContext, double>;
+}  // namespace phi
+
+PD_REGISTER_KERNEL(lrn_grad, CPU, ALL_LAYOUT, phi::LRNGradKernel, float) {}