uniform_random_kernel.cu 3.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// 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/uniform_random_kernel.h"

17
#include <thrust/random.h>
18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
#include "gflags/gflags.h"

#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/distribution_helper.h"
#include "paddle/phi/kernels/funcs/index_impl.cu.h"

namespace phi {

template <typename T>
struct UniformGenerator {
  T min_, max_;
  unsigned int seed_;
  T diag_val_;
  unsigned int diag_num_;
  unsigned int diag_step_;
  __host__ __device__ UniformGenerator(
      T min, T max, int seed, int diag_num, int diag_step, T diag_val)
      : min_(min),
        max_(max),
        seed_(seed),
        diag_num_(diag_num),
        diag_step_(diag_step),
        diag_val_(diag_val) {}

  __host__ __device__ T operator()(const unsigned int n) const {
    thrust::minstd_rand rng;
    rng.seed(seed_);
    thrust::uniform_real_distribution<T> dist(min_, max_);
    rng.discard(n);
    T out = dist(rng);
    unsigned int remainder = n % (diag_step_ + 1);
    if (remainder == 0 && diag_num_ > n / (diag_step_ + 1)) {
      out = diag_val_;
    }
    return out;
  }
};

template <typename T, typename Context>
void UniformRandomRawKernel(const Context& dev_ctx,
58
                            const IntArray& shape,
59 60 61 62 63 64 65 66 67
                            DataType dtype,
                            float min,
                            float max,
                            int seed,
                            int diag_num,
                            int diag_step,
                            float diag_val,
                            DenseTensor* out) {
  out->Resize(phi::make_ddim(shape.GetData()));
68
  dev_ctx.template Alloc<T>(out);
69
  if (seed == 0) {
70
    // Use global Generator seed
71 72 73 74
    using MT = typename kps::details::MPTypeTrait<T>::Type;
    funcs::uniform_distribution<MT> dist;
    funcs::uniform_real_transform<MT> trans(min, max);
    funcs::distribution_and_transform<T>(dev_ctx, out, dist, trans);
75
  } else {
76
    // Use OP seed
77 78 79 80 81 82 83 84
    auto func =
        UniformGenerator<T>(min, max, seed, diag_num, diag_step, diag_val);
    IndexKernel<T, UniformGenerator<T>>(dev_ctx, out, func);
  }
}

template <typename T, typename Context>
void UniformRandomKernel(const Context& dev_ctx,
85
                         const IntArray& shape,
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105
                         DataType dtype,
                         float min,
                         float max,
                         int seed,
                         DenseTensor* out) {
  UniformRandomRawKernel<T>(
      dev_ctx, shape, dtype, min, max, seed, 0, 0, 0.0f, out);
}

}  // namespace phi

PD_REGISTER_KERNEL(uniform_random_raw,
                   GPU,
                   ALL_LAYOUT,
                   phi::UniformRandomRawKernel,
                   float,
                   double) {}

PD_REGISTER_KERNEL(
    uniform_random, GPU, ALL_LAYOUT, phi::UniformRandomKernel, float, double) {}