add check_memory_continue kernel (#45999)

paddle/phi/kernels/check_memory_continue_kernel.cc

+// 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.
+
+#include "paddle/phi/kernels/check_memory_continue_kernel.h"
+
+#include <sstream>
+#include <vector>
+#include "glog/logging.h"
+
+#include "paddle/phi/core/kernel_registry.h"
+
+#include "paddle/fluid/platform/device_memory_aligment.h"
+
+namespace phi {
+
+template <typename T, typename Context>
+void CheckMemoryContinueKernel(const Context &dev_ctx,
+                               const std::vector<const DenseTensor *> &input,
+                               DenseTensor *output,
+                               std::vector<DenseTensor *> xout) {
+  int64_t size_of_dtype = sizeof(T);
+  auto dtype = input.at(0)->dtype();
+  int64_t numel = 0;
+  // check address
+  for (size_t i = 1; i < input.size(); ++i) {
+    PADDLE_ENFORCE_EQ(
+        dtype,
+        input.at(i)->dtype(),
+        errors::InvalidArgument(
+            "The DataType of input tensors of fake_coalesce should be "
+            "consistent, current dtype is: %s, but the previous dtype is %s",
+            dtype,
+            input.at(i)->dtype()));
+    const void *cur_address = input.at(i - 1)->data();
+    int64_t len = input.at(i - 1)->numel();
+    auto offset =
+        paddle::platform::Alignment(len * size_of_dtype, dev_ctx.GetPlace());
+    void *infer_next_address = reinterpret_cast<void *>(
+        reinterpret_cast<uintptr_t>(cur_address) + offset);
+    const void *next_address = input.at(i)->data();
+    numel += offset;
+
+    VLOG(10) << ::paddle::string::Sprintf(
+        "Input[%d] address: 0X%02x, Input[%d] address: 0X%02x, Infer "
+        "input[%d] address: 0X%02x, offset: %d.",
+        i - 1,
+        cur_address,
+        i,
+        next_address,
+        i,
+        infer_next_address,
+        offset);
+    PADDLE_ENFORCE_EQ(
+        infer_next_address,
+        next_address,
+        errors::InvalidArgument(
+            "The infered address of the next tensor should be equal to the "
+            "real address of the next tensor. But got infered address is %p "
+            "and real address is %p.",
+            infer_next_address,
+            next_address));
+  }
+  numel += paddle::platform::Alignment(
+      (*input.rbegin())->numel() * size_of_dtype, dev_ctx.GetPlace());
+  // reset holder, do inplace
+  output->ShareBufferWith(*input.at(0));
+  output->Resize({numel / size_of_dtype});
+  VLOG(4) << "addr:" << output->data<T>();
+}
+
+}  // namespace phi
+
+PD_REGISTER_KERNEL(check_memory_continue,
+                   CPU,
+                   ALL_LAYOUT,
+                   phi::CheckMemoryContinueKernel,
+                   int,
+                   float,
+                   double) {}
+
+#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
+PD_REGISTER_KERNEL(check_memory_continue,
+                   GPU,
+                   ALL_LAYOUT,
+                   phi::CheckMemoryContinueKernel,
+                   phi::dtype::float16,
+                   int,
+                   float,
+                   double) {}
+#endif

paddle/phi/kernels/check_memory_continue_kernel.h

+// 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.
+
+#pragma once
+
+#include <vector>
+
+#include "paddle/phi/core/dense_tensor.h"
+
+namespace phi {
+
+// WHY add this op?
+// This op is used for convert fused_all_reduce_op_handle in Graph to Program.
+// i.e, fused_all_reduce_op_handle = check_memory_continue + c_allreduce_sum
+// There are two reasons that check_memory_continue is added:
+// 1. c_allreduce_sum takes single tensor as input, while
+// fused_all_reduce_op_handle takse tensor array as input, so we need a op to
+// convert tensor array into a single tensor
+// 2. fused_all_reduce_op_handle has a premise that all tensor's addresses are
+// continue, so we need a op to do the check.
+
+// see details in fused_all_reduce_op_handle.cc
+template <typename T, typename Context>
+void CheckMemoryContinueKernel(const Context &dev_ctx,
+                               const std::vector<const DenseTensor *> &input,
+                               DenseTensor *output,
+                               std::vector<DenseTensor *> xout);
+
+}  // namespace phi