multinomial_kernel.cu

/* 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. */

#ifndef PADDLE_WITH_HIP
// To-do(qili93): fix this after issue resolved
// https://github.com/ROCmSoftwarePlatform/rocPRIM/issues/202

#include "paddle/phi/kernels/multinomial_kernel.h"

#include <thrust/execution_policy.h>
#include <thrust/random.h>
#include <thrust/scan.h>
#include <thrust/transform.h>

#ifdef __NVCC__
#include "cub/cub.cuh"
#endif
#ifdef __HIPCC__
#include <hipcub/hipcub.hpp>
namespace cub = hipcub;
#endif

#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/common/scalar.h"
#include "paddle/phi/core/ddim.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/arg_min_max_kernel.h"
#include "paddle/phi/kernels/empty_kernel.h"
#include "paddle/phi/kernels/funcs/distribution_helper.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"
#include "paddle/phi/kernels/funcs/for_range.h"
#include "paddle/phi/kernels/funcs/inclusive_scan.h"
#include "paddle/phi/kernels/funcs/multinomial_functor.h"
#include "paddle/phi/kernels/top_k_kernel.h"

// See Note [ Why still include the fluid headers? ]
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/platform/transform.h"

DECLARE_bool(use_curand);

namespace phi {

template <typename T>
__global__ void NormalizeProbability(T* norm_probs,
                                     const T* in_data,
                                     T* sum_rows,
                                     int64_t num_distributions,
                                     int64_t num_categories) {
  int id = threadIdx.x + blockIdx.x * blockDim.x +
           blockIdx.y * gridDim.x * blockDim.x;
  if (id < num_distributions * num_categories) {
    PADDLE_ENFORCE(
        in_data[id] >= 0.0,
        "The input of multinomial distribution should be >= 0, but got %f.",
        in_data[id]);
    int64_t row_id = id / num_categories;
    PADDLE_ENFORCE(sum_rows[row_id] > 0.0,
                   "The sum of one multinomial distribution probability should "
                   "be > 0, but got %f.",
                   sum_rows[row_id]);
    norm_probs[id] = in_data[id] / sum_rows[row_id];
  }
}

template <typename T>
__global__ void GetCumulativeProbs(T* norm_probs_data,
                                   int64_t num_distributions,
                                   int64_t num_categories,
                                   T* cumulative_probs_data) {
  int id = blockIdx.x;
  thrust::inclusive_scan(thrust::device,
                         norm_probs_data + id * num_categories,
                         norm_probs_data + (id + 1) * num_categories,
                         cumulative_probs_data + id * num_categories);
}

template <typename T>
struct RandomGeneratorCudaFunctor {
  unsigned int seed_;
  __host__ __device__ RandomGeneratorCudaFunctor(int seed) : seed_(seed) {}

  __host__ __device__ T operator()(const unsigned int n) const {
    thrust::minstd_rand rng;
    rng.seed(seed_);
    thrust::uniform_real_distribution<T> dist(0.0, 1.0);
    rng.discard(n);
    return dist(rng);
  }
};

template <typename T>
__device__ int binarySearchFunctor(T* cumulative_probs_data,
                                   T* norm_probs_data,
                                   int num_categories,
                                   T rng_number) {
  int left = 0;
  int right = num_categories;

  while (right - left > 0) {
    int mid = left + (right - left) / 2;

    T temp_prob = cumulative_probs_data[mid];
    if (temp_prob < rng_number) {
      left = mid + 1;
    } else {
      right = mid;
    }
  }

  if (left == num_categories) {
    left = num_categories - 1;
  }

  while (left >= 1 && norm_probs_data[left] == 0) left--;

  return left;
}

template <typename T>
__global__ void sampleMultinomialWithReplacement(
    T* rng_data,
    const int64_t num_samples,
    int64_t* out_data,
    const int64_t num_distributions,
    const int64_t num_categories,
    T* cumulative_probs_data,
    T* norm_probs_data,
    uint64_t seed,
    uint64_t offset,
    bool use_curand) {
  // use binary search to get the selected category sample id.
  // let cumulative_probs_data[id-1] < rng_data < cumulative_probs_data[id].
  size_t idx = gridDim.x * blockDim.x * blockIdx.y + blockDim.x * blockIdx.x +
               threadIdx.x;

  curandStatePhilox4_32_10_t state;
  curand_init(seed, idx, offset, &state);

  int sample = blockIdx.x * blockDim.x + threadIdx.x;
  for (int dist = blockIdx.y; dist < num_distributions; dist += gridDim.y) {
    if (sample < num_samples) {
      T rng_number = rng_data[sample + dist * num_samples];
      if (use_curand) {
        rng_number = static_cast<T>(curand_uniform4(&state).x);
      }
      // Find the bucket that a uniform random number lies in
      int selected_category =
          binarySearchFunctor<T>(cumulative_probs_data + dist * num_categories,
                                 norm_probs_data + dist * num_categories,
                                 num_categories,
                                 rng_number);

      out_data[sample + dist * num_samples] = selected_category;
    }
  }
}

template <typename T, typename Context>
void MultinomialKernel(const Context& dev_ctx,
                       const DenseTensor& x,
                       int num_samples,
                       bool replacement,
                       DenseTensor* out) {
  auto* in_data = x.data<T>();
  int64_t* out_data = dev_ctx.template Alloc<int64_t>(out);

  auto in_dims = x.dims();
  int64_t in_rank = in_dims.size();
  const int64_t num_categories = in_dims[in_rank - 1];
  const int64_t num_distributions = in_rank > 1 ? in_dims[in_rank - 2] : 1;

  // If replacement is False, it's not a replaceable sample. Every category
  // can
  // be used only once. So after every sample, probability of the distribution
  // will change. The implementation can't be parallelizable. Thus, call CPU
  // implementation ``funcs::MultinomialFunctor`` to sample the distribution.
  if (!replacement) {
    int64_t in_data_numel = x.numel();
    int64_t out_data_numel = out->numel();

    T* cpu_in_data = new T[in_data_numel];
    int64_t* cpu_out_data = new int64_t[out_data_numel];

#ifdef PADDLE_WITH_HIP
    hipMemcpy(
        cpu_in_data, in_data, in_data_numel * sizeof(T), hipMemcpyDeviceToHost);
#else
    cudaMemcpy(cpu_in_data,
               in_data,
               in_data_numel * sizeof(T),
               cudaMemcpyDeviceToHost);
#endif
    if (FLAGS_use_curand) {
      for (size_t i = 0; i < num_distributions; ++i) {
        int zero_num = 0;
        for (size_t j = 0; j < num_categories; ++j) {
          T weight = cpu_in_data[i * num_distributions + j];
          PADDLE_ENFORCE_GE(
              weight,
              0,
              errors::InvalidArgument(
                  "Each element of multinomial'input must >= 0, but got %f.",
                  weight));
          if (weight == static_cast<T>(0)) {
            zero_num++;
          }
        }
        int valid_samples = num_categories - zero_num;
        PADDLE_ENFORCE_LE(
            num_samples,
            valid_samples,
            errors::InvalidArgument("When replacement=False, 'num_samples' "
                                    "must less than or eaqual to the number of "
                                    "positive item of input"));
      }

      // Refer to [gumbel softmax algorithm]
      DenseTensor rand = EmptyLike<T, Context>(dev_ctx, x);
      T* rand_data = rand.data<T>();
      funcs::uniform_distribution<T> dist;
      funcs::exponential_transform<T> trans(1.0);
      funcs::distribution_and_transform<T>(dev_ctx, &rand, dist, trans);

      funcs::ForRange<Context> for_range(dev_ctx, x.numel());
      for_range([rand_data, in_data] __device__(size_t idx) {
        rand_data[idx] = in_data[idx] / rand_data[idx];
      });

      if (num_samples == 1) {
        ArgMaxKernel<T, Context>(
            dev_ctx, rand, -1, true, false, 3 /*proto::VarType::INT64*/, out);
      } else {
        std::vector<int64_t> out_dim_vec = vectorize<int64_t>(out->dims());
        DenseTensor value =
            Empty<T, Context>(dev_ctx, ScalarArray(out_dim_vec));
        TopkKernel<T, Context>(
            dev_ctx, rand, Scalar(num_samples), -1, true, true, &value, out);
      }
      return;
    }

    funcs::MultinomialFunctor<T>(dev_ctx,
                                 cpu_out_data,
                                 cpu_in_data,
                                 num_samples,
                                 replacement,
                                 num_categories,
                                 num_distributions);

#ifdef PADDLE_WITH_HIP
    hipMemcpy(out_data,
              cpu_out_data,
              out_data_numel * sizeof(int64_t),
              hipMemcpyHostToDevice);
#else
    cudaMemcpy(out_data,
               cpu_out_data,
               out_data_numel * sizeof(int64_t),
               cudaMemcpyHostToDevice);
#endif

    delete[] cpu_in_data;
    delete[] cpu_out_data;
    return;
  }

  // Sum of input may not be 1. To get probability in range [0, 1], calculate
  // sum of each row of input, and then use the sum to normalize the input.
  // sum_row_data: sum of each row
  DenseTensor sum_rows_tensor;
  sum_rows_tensor.Resize({num_distributions});
  auto* sum_rows_data = dev_ctx.template Alloc<T>(&sum_rows_tensor);

  auto& place = *dev_ctx.eigen_device();

  if (num_distributions == 1) {
    auto eigen_input = EigenVector<T>::Flatten(x);
    auto eigen_sum_rows = EigenVector<T>::Flatten(sum_rows_tensor);
    eigen_sum_rows.device(place) =
        eigen_input.sum(Eigen::DSizes<int, 1>(1))
            .eval()
            .reshape(Eigen::DSizes<int, 1>(sum_rows_tensor.dims()[0]));
  } else {
    auto eigen_input = EigenMatrix<T>::From(x);
    auto eigen_sum_rows = EigenVector<T>::Flatten(sum_rows_tensor);
    eigen_sum_rows.device(place) = eigen_input.sum(Eigen::DSizes<int, 1>(1));
  }

  // Normalize row of each distribution to get the probability in range [0,
  // 1].
  // norm_probs_data: probability of the distribution
  DenseTensor norm_probs_tensor;
  norm_probs_tensor.Resize({num_distributions, num_categories});
  auto* norm_probs_data = dev_ctx.template Alloc<T>(&norm_probs_tensor);

  // number of threads in a block is min(num_categories, 512)
  int block_size = num_categories < 512 ? num_categories : 512;
  dim3 block_norm(block_size);
  dim3 grid_norm((num_distributions * num_categories - 1) / block_norm.x + 1);
  NormalizeProbability<T><<<grid_norm, block_norm, 0, dev_ctx.stream()>>>(
      norm_probs_data,
      in_data,
      sum_rows_data,
      num_distributions,
      num_categories);

  // Get cumulative probability of each distribution. It's the same function
  // of ``cumsum`` op.
  DenseTensor cumulative_probs_tensor;
  cumulative_probs_tensor.Resize({num_distributions, num_categories});
  auto* cumulative_probs_data =
      dev_ctx.template Alloc<T>(&cumulative_probs_tensor);

  if (FLAGS_use_curand) {
    // 'phi::funcs::InclusiveScan' has higher accuracy than
    // 'thrust::inclusive_scan'
    funcs::InclusiveScan<T, std::plus<T>>(
        /*in*/ norm_probs_data,
        /*out*/ cumulative_probs_data,
        /*outer_dim*/ static_cast<size_t>(num_distributions),
        /*mid_dim*/ static_cast<size_t>(num_categories),
        /*inner_dim*/ static_cast<size_t>(1),
        /*init*/ static_cast<T>(0),
        std::plus<T>(),
        /*reverse=*/false,
        dev_ctx);
  } else {
    dim3 block_cumsum(1);
    dim3 grid_cumsum(num_distributions);
    GetCumulativeProbs<T><<<grid_cumsum, block_cumsum, 0, dev_ctx.stream()>>>(
        norm_probs_data,
        num_distributions,
        num_categories,
        cumulative_probs_data);
  }

  // Generate random number for each sample.
  std::random_device rd;
  auto seed = rd();

  DenseTensor rng_data_tensor;
  rng_data_tensor.Resize({num_distributions, num_samples});
  auto* rng_data = dev_ctx.template Alloc<T>(&rng_data_tensor);

  thrust::counting_iterator<int64_t> index_sequence_begin(0);
  paddle::platform::Transform<GPUContext> trans;
  trans(dev_ctx,
        index_sequence_begin,
        index_sequence_begin + num_distributions * num_samples,
        rng_data,
        RandomGeneratorCudaFunctor<T>(seed));

  // Sample the multinomial distributions.
  dim3 block(128);
  int64_t device_id = dev_ctx.GetPlace().GetDeviceId();
  const auto& prop = phi::backends::gpu::GetDeviceProperties(device_id);
  int grid_y = std::min<int64_t>(num_distributions, prop.maxGridSize[1]);
  dim3 grid((num_samples - 1) / block.x + 1, grid_y);

  auto gen_cuda = dev_ctx.GetGenerator();
  size_t curand4_loop_times =
      (num_distributions + 4 * grid_y - 1) / (4 * grid_y);
  // 'increment' shoulde be multiple of 4
  uint64_t increment = curand4_loop_times * 4;
  auto seed_offset = gen_cuda->IncrementOffset(increment);

  sampleMultinomialWithReplacement<T><<<grid, block, 0, dev_ctx.stream()>>>(
      rng_data,
      num_samples,
      out_data,
      num_distributions,
      num_categories,
      cumulative_probs_data,
      norm_probs_data,
      seed_offset.first,
      seed_offset.second,
      FLAGS_use_curand);
}

}  // namespace phi

PD_REGISTER_KERNEL(multinomial,  // cuda_only
                   GPU,
                   ALL_LAYOUT,
                   phi::MultinomialKernel,
                   float,
                   double) {}

#endif