paddlepaddle--paddle/paddle/phi/kernels/xpu/top_k_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/top_k_kernel.h"

#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/full_kernel.h"
#include "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/phi/kernels/xpu/xpu_mem_util.h"
namespace phi {

template <typename T, typename Context>
void TopkKernel(const Context& dev_ctx,
                const DenseTensor& x,
                const Scalar& k_scalar,
                int axis,
                bool largest,
                bool sorted,
                DenseTensor* out,
                DenseTensor* indices) {
  using XPUType = typename XPUTypeTrait<T>::Type;
  using XPUTypeINT64 = typename XPUTypeTrait<int64_t>::Type;

  if (out && out->numel() == 0) {
    dev_ctx.template Alloc<T>(out);
    dev_ctx.template Alloc<int64_t>(indices);
    return;
  }

  const auto& in_dims = x.dims();
  if (in_dims.size() == 0) {
    Copy<Context>(dev_ctx, x, dev_ctx.GetPlace(), false, out);
    dev_ctx.template Alloc<int64_t>(indices);
    funcs::set_constant(dev_ctx, indices, static_cast<int64_t>(0));
    return;
  }

  // axis < 0, calculate the real axis
  if (axis < 0) {
    axis += in_dims.size();
  }

  int64_t k = k_scalar.to<int64_t>();
  // out shape [-1]
  if (k_scalar.FromTensor()) {
    auto out_dims_ = out->dims();
    // according to axis to set K value in the dim
    out_dims_[axis] = k;
    out->Resize(out_dims_);
    indices->Resize(out_dims_);
  }
  if (x.numel() == 0) {
    Full<T, Context>(dev_ctx, out->dims(), NAN, out);
    Full<int64_t, Context>(dev_ctx, indices->dims(), 0, indices);
    return;
  }
  PADDLE_ENFORCE_GE(
      x.numel(),
      k,
      errors::InvalidArgument(
          "x has only %d element, can not find %d top values.", x.numel(), k));

  const T* in_data = x.data<T>();
  int64_t* indices_data = dev_ctx.template Alloc<int64_t>(indices);
  T* output_data = dev_ctx.template Alloc<T>(out);

  const auto& out_dims = out->dims();

  // PADDLE_ENFORCE_EQ(
  //     sorted,
  //     true,
  //     errors::External(
  //         "XPU API does not support unsorted topk operation currently."
  //         " Operator will be supported in future update."));
  if (axis < 0) axis += in_dims.size();

  if (axis + 1 == in_dims.size()) {
    const int64_t row =
        common::product(slice_ddim(in_dims, 0, in_dims.size() - 1));
    const int64_t col = in_dims[in_dims.size() - 1];

    int r =
        xpu::sorted_topk<XPUType>(dev_ctx.x_context(),
                                  reinterpret_cast<const XPUType*>(in_data),
                                  reinterpret_cast<XPUType*>(output_data),
                                  reinterpret_cast<XPUTypeINT64*>(indices_data),
                                  row,
                                  col,
                                  k,
                                  largest);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "sorted_topk");

  } else {
    // do transpose if axis is not the last dim of input
    std::vector<int64_t> trans_axes;
    for (int i = 0; i < axis; i++) {
      trans_axes.emplace_back(i);
    }
    for (int i = axis + 1; i < in_dims.size(); i++) {
      trans_axes.emplace_back(i);
    }
    trans_axes.emplace_back(axis);
    // Get input and output dims for transpose
    DDim trans_dims(in_dims);
    DDim trans_out_dims(out->dims());
    for (size_t i = 0; i < trans_axes.size(); i++) {
      trans_dims[i] = in_dims[trans_axes[i]];
      trans_out_dims[i] = out_dims[trans_axes[i]];
    }

    std::vector<int64_t> x_shape_host(in_dims.size(), 0);
    for (int i = 0; i < in_dims.size(); ++i) {
      x_shape_host[i] = in_dims[i];
    }

    xpu::ctx_guard RAII_GUARD(dev_ctx.x_context());
    XPUType* trans_in_data = RAII_GUARD.alloc_l3_or_gm<XPUType>(x.numel());
    PADDLE_ENFORCE_XDNN_NOT_NULL(trans_in_data);

    // Transpose and save interval output to trans_in
    int r = xpu::transpose<XPUType>(dev_ctx.x_context(),
                                    reinterpret_cast<const XPUType*>(in_data),
                                    trans_in_data,
                                    x_shape_host,
                                    trans_axes);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "transpose");

    XPUType* trans_out_data = RAII_GUARD.alloc_l3_or_gm<XPUType>(out->numel());
    PADDLE_ENFORCE_XDNN_NOT_NULL(trans_out_data);

    int64_t* trans_idx_data = RAII_GUARD.alloc_l3_or_gm<int64_t>(out->numel());
    PADDLE_ENFORCE_XDNN_NOT_NULL(trans_idx_data);

    const int64_t row =
        common::product(slice_ddim(trans_dims, 0, trans_dims.size() - 1));
    const int64_t col = trans_dims[trans_dims.size() - 1];

    // Do top k on transposed input
    r = xpu::sorted_topk<XPUType>(
        dev_ctx.x_context(),
        reinterpret_cast<const XPUType*>(trans_in_data),
        reinterpret_cast<XPUType*>(trans_out_data),
        reinterpret_cast<XPUTypeINT64*>(trans_idx_data),
        row,
        col,
        k,
        largest);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "sorted_topk");

    PADDLE_ENFORCE_XDNN_SUCCESS(r, "cast");
    // Transpose back to original dims
    std::vector<int64_t> trans_back_axes;
    for (int i = 0; i < axis; i++) {
      trans_back_axes.emplace_back(i);
    }
    trans_back_axes.emplace_back(trans_out_dims.size() - 1);
    for (int i = axis; i < trans_out_dims.size() - 1; i++) {
      trans_back_axes.emplace_back(i);
    }

    std::vector<int64_t> trans_out_shape_host(trans_back_axes.size(), 0);
    for (size_t i = 0; i < trans_back_axes.size(); ++i) {
      trans_out_shape_host[i] = trans_out_dims[i];
    }
    r = xpu::transpose<XPUType>(
        dev_ctx.x_context(),
        reinterpret_cast<const XPUType*>(trans_out_data),
        reinterpret_cast<XPUType*>(output_data),
        trans_out_shape_host,
        trans_back_axes);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "transpose");
    r = xpu::transpose<int64_t>(dev_ctx.x_context(),
                                trans_idx_data,
                                indices_data,
                                trans_out_shape_host,
                                trans_back_axes);
    PADDLE_ENFORCE_XDNN_SUCCESS(r, "transpose");
  }
}

template <typename T, typename Context>
void TopkV1Kernel(const Context& dev_ctx,
                  const DenseTensor& x,
                  const Scalar& k_scalar,
                  DenseTensor* out,
                  DenseTensor* indices) {
  TopkKernel<T, Context>(dev_ctx, x, k_scalar, -1, true, true, out, indices);
}
}  // namespace phi

PD_REGISTER_KERNEL(
    topk, XPU, ALL_LAYOUT, phi::TopkKernel, float, phi::float16) {
  kernel->OutputAt(1).SetDataType(phi::DataType::INT64);
}

PD_REGISTER_KERNEL(
    topk_v1, XPU, ALL_LAYOUT, phi::TopkV1Kernel, float, phi::float16) {
  kernel->OutputAt(1).SetDataType(phi::DataType::INT64);
}