coalesce_tensor_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/coalesce_tensor_kernel.h"

#include <sstream>
#include <vector>

#include "paddle/fluid/platform/device_memory_aligment.h"
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/math_function.h"

namespace phi {

template <typename Context>
struct FillConstantVisitor {
  FillConstantVisitor(const Context &dev_ctx,
                      DenseTensor *tensor,
                      const float value)
      : dev_ctx_(dev_ctx), tensor_(tensor), value_(value) {}

  template <typename T>
  void apply(typename std::enable_if<std::is_same<T, int8_t>::value ||
                                     std::is_same<T, int16_t>::value>::type * =
                 nullptr) const {
    PADDLE_THROW(
        errors::InvalidArgument("Not support data type for set_constant attr"));
  }

  template <typename T>
  void apply(typename std::enable_if<!(std::is_same<T, int8_t>::value ||
                                       std::is_same<T, int16_t>::value)>::type
                 * = nullptr) const {
    phi::funcs::SetConstant<Context, T> set_constant;
    set_constant(dev_ctx_, tensor_, static_cast<T>(value_));
  }

  const Context &dev_ctx_;
  DenseTensor *tensor_;
  float value_;
};

void GetMemSizeAndDtype(const std::vector<const DenseTensor *> &lod_tensors,
                        size_t *numel,
                        const size_t &size_of_dtype,
                        const phi::Place &place,
                        const bool use_align = true,
                        const int align_size = -1) {
  *numel = 0;
  std::stringstream ss;
  ss << "alloc_space_for_vars: ";
  for (size_t i = 0; i < lod_tensors.size(); ++i) {
    auto size = lod_tensors[i]->numel();
    PADDLE_ENFORCE_GT(size,
                      0,
                      errors::InvalidArgument(
                          "The number of `%d`-th tensor's elements is 0.", i));
    auto len = use_align ? paddle::platform::Alignment(
                               static_cast<size_t>(size) * size_of_dtype,
                               place,
                               align_size) /
                               size_of_dtype
                         : static_cast<size_t>(size);
    const void *ptr =
        lod_tensors[i]->IsInitialized() ? lod_tensors[i]->data() : nullptr;
    VLOG(4) << size << " " << len;
    ss << "input(" << i << "-th tensor) dim:(" << lod_tensors[i]->dims() << ") "
       << " addres:" << ptr << " len: " << len << ", ";
    *numel += len;
  }
  VLOG(10) << ss.str();
}

template <typename T, typename Context>
void CoalesceTensorKernel(const Context &dev_ctx,
                          const std::vector<const DenseTensor *> &input,
                          DataType dtype,
                          bool copy_data,
                          bool set_constant,
                          bool persist_output,
                          float constant,
                          bool use_align,
                          int align_size,
                          int size_of_dtype,
                          const std::vector<int64_t> &concated_shapes,
                          const std::vector<int64_t> &concated_ranks,
                          std::vector<DenseTensor *> output,
                          DenseTensor *fused_output) {
  PADDLE_ENFORCE_GT(
      input.size(),
      static_cast<size_t>(0),
      errors::InvalidArgument("The CoalesceTensor operator has no input."));
  PADDLE_ENFORCE_EQ(input.size(),
                    output.size(),
                    errors::InvalidArgument(
                        "The number of CoalesceTensor operator's input and "
                        "output is not match, "
                        "input number is %u, output number is %u.",
                        input.size(),
                        output.size()));

  // Input & Output check: only support LoDTensor
  bool has_not_init_in_vars = false;
  for (size_t i = 0; i < input.size(); ++i) {
    PADDLE_ENFORCE_NOT_NULL(
        input[i],
        errors::InvalidArgument("The %d-th input tensor cannot be nullptr.",
                                i));
    PADDLE_ENFORCE_NOT_NULL(
        output[i],
        errors::InvalidArgument("The %d-th output tensor cannot be nullptr.",
                                i));
    if (!input[i]->IsInitialized()) {
      has_not_init_in_vars = true;
    }
  }

  if (has_not_init_in_vars) {
    PADDLE_ENFORCE_EQ(
        concated_ranks.size(),
        output.size(),
        errors::InvalidArgument("The attribute(concated_ranks) length must be "
                                "equal to the output tensor number."));
    int64_t accumulated_ranks = 0;
    for (size_t i = 0; i < input.size(); ++i) {
      phi::DDim dims(concated_shapes.data() + accumulated_ranks,
                     concated_ranks[i]);
      if (!input[i]->IsInitialized()) {
        PADDLE_ENFORCE_EQ(
            input[i],
            output[i],
            errors::InvalidArgument(
                "The %d-th output tensor and %d-th input tensor when the "
                "%d-th input tensor is not initialized.",
                i,
                i,
                i));
        output[i]->Resize(dims);
      } else {
        PADDLE_ENFORCE_EQ(input[i]->dims(),
                          dims,
                          errors::InvalidArgument(
                              "The %d-th input tensor shape does not match the "
                              "attribute(concated_shapes) and "
                              "attribute(concated_ranks).",
                              i));
      }
      accumulated_ranks += concated_ranks[i];
      PADDLE_ENFORCE_LE(
          accumulated_ranks,
          concated_shapes.size(),
          errors::InvalidArgument("The attribute(concated_shapes) and "
                                  "attribute(concated_ranks) do not match."));
    }
    PADDLE_ENFORCE_EQ(
        accumulated_ranks,
        concated_shapes.size(),
        errors::InvalidArgument("The attribute(concated_shapes) and "
                                "attribute(concated_ranks) do not match."));
  }

  // Init the output as input
  for (size_t i = 0; i < input.size(); ++i) {
    output[i]->Resize(input[i]->dims());
  }

  // Get numel and dtype
  size_t numel = 0;

  if (size_of_dtype == -1) {
    size_of_dtype = paddle::experimental::SizeOf(dtype);
  }
  GetMemSizeAndDtype(
      input, &numel, size_of_dtype, dev_ctx.GetPlace(), use_align, align_size);

  // Alloc the continuous space
  void *fused_tensor_ptr = dev_ctx.Alloc(
      &fused_output->Resize(phi::make_ddim({static_cast<int64_t>(numel)})),
      dtype);
  VLOG(10) << "Fused tensor addr " << fused_tensor_ptr;

  // Init the continuous space
  size_t offset = 0;
  if (copy_data) {
    for (size_t i = 0; i < input.size(); ++i) {
      size_t len = static_cast<size_t>(input[i]->numel());
      auto sub_tensor = fused_output->Slice(static_cast<int64_t>(offset),
                                            static_cast<int64_t>(offset + len));
      phi::Copy(dev_ctx, *input[i], dev_ctx.GetPlace(), false, &sub_tensor);

      offset += use_align
                    ? paddle::platform::Alignment(
                          len * size_of_dtype, dev_ctx.GetPlace(), align_size) /
                          size_of_dtype
                    : len;
    }
  } else if (set_constant) {
    phi::VisitDataType(
        dtype, FillConstantVisitor<Context>(dev_ctx, fused_output, constant));
  } else if (persist_output) {
    for (size_t i = 0; i < output.size(); ++i) {
      size_t len = static_cast<size_t>(output[i]->numel());
      auto sub_tensor = fused_output->Slice(static_cast<int64_t>(offset),
                                            static_cast<int64_t>(offset + len));
      // some var may not persistable, or persistable var may not init
      if (output[i]->IsInitialized()) {
        phi::Copy(dev_ctx, *output[i], dev_ctx.GetPlace(), false, &sub_tensor);
      }
      offset += use_align
                    ? paddle::platform::Alignment(
                          len * size_of_dtype, dev_ctx.GetPlace(), align_size) /
                          size_of_dtype
                    : len;
    }
  }

  // Make the outputs point to the continuous space.
  offset = 0;
  std::stringstream ss;
  ss << "alloc_space_for_vars: ";

  for (size_t i = 0; i < output.size(); ++i) {
    size_t len = static_cast<size_t>(output[i]->numel());
    auto dim = output[i]->dims();
    VLOG(4) << len << " " << dim << " " << offset;
    output[i]
        ->ShareDataWith(fused_output->Slice(static_cast<int64_t>(offset),
                                            static_cast<int64_t>(offset + len)))
        .Resize(dim);
    len = use_align ? paddle::platform::Alignment(
                          len * size_of_dtype, dev_ctx.GetPlace(), align_size) /
                          size_of_dtype
                    : len;
    ss << "output(" << i << "-th tensor) dim:(" << dim << ")"
       << " address: " << output[i]->data() << " len: " << len << ", ";
    offset += len;
  }
  PADDLE_ENFORCE_EQ((int64_t)offset,
                    fused_output->numel(),
                    errors::InvalidArgument(
                        "The alloc_space_for_vars's offset: %s is unequal with "
                        "fused_output's numel: %s.",
                        offset,
                        fused_output->numel()));
  VLOG(10) << ss.str();
}

}  // namespace phi

PD_REGISTER_KERNEL(coalesce_tensor,
                   CPU,
                   ALL_LAYOUT,
                   phi::CoalesceTensorKernel,
                   int,
                   float,
                   double) {}

#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
PD_REGISTER_KERNEL(coalesce_tensor,
                   GPU,
                   ALL_LAYOUT,
                   phi::CoalesceTensorKernel,
                   phi::dtype::float16,
                   int,
                   float,
                   double) {
  kernel->InputAt(0).SetBackend(phi::Backend::ALL_BACKEND);
}
#endif