auc_kernel.cu 10.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// 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.

15 16
#include "paddle/phi/kernels/auc_kernel.h"

17 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 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
#include "paddle/fluid/platform/device/gpu/gpu_primitives.h"
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/core/kernel_registry.h"

namespace phi {

using paddle::platform::PADDLE_CUDA_NUM_THREADS;

__global__ void ClearObsoleteDataKernel(int64_t *pos,
                                        int64_t *neg,
                                        const int bucket_length,
                                        const int slide_steps) {
  int cur_step_index =
      static_cast<int>(pos[(slide_steps + 1) * bucket_length]) % slide_steps;
  int cur_step_begin = cur_step_index * bucket_length;
  int sum_step_begin = slide_steps * bucket_length;
  CUDA_KERNEL_LOOP(i, bucket_length) {
    pos[sum_step_begin + i] -= pos[cur_step_begin + i];
    neg[sum_step_begin + i] -= neg[cur_step_begin + i];
    pos[cur_step_begin + i] = neg[cur_step_begin + i] = 0;
  }
}

__global__ void UpdateSumDataKernel(int64_t *pos,
                                    int64_t *neg,
                                    const int bucket_length,
                                    const int slide_steps) {
  int cur_step_index =
      static_cast<int>(pos[(slide_steps + 1) * bucket_length]) % slide_steps;
  int cur_step_begin = cur_step_index * bucket_length;
  int sum_step_begin = slide_steps * bucket_length;
  CUDA_KERNEL_LOOP(i, bucket_length) {
    pos[sum_step_begin + i] += pos[cur_step_begin + i];
    neg[sum_step_begin + i] += neg[cur_step_begin + i];
  }
}

template <typename T>
__global__ void AddDataKernel(const int64_t *label_data,
                              const T *pred_data,
                              const int inference_width,
                              const int num_thresholds,
                              int64_t *pos,
                              int64_t *neg,
                              const int numel,
                              const int slide_steps) {
  int cur_step_begin = 0;
  if (slide_steps > 0) {
    int cur_step_index =
        static_cast<int>(pos[(slide_steps + 1) * (1 + num_thresholds)]) %
        slide_steps;
    cur_step_begin = cur_step_index * (1 + num_thresholds);
  }
  CUDA_KERNEL_LOOP(i, numel) {
    auto predict_data = pred_data[i * inference_width + (inference_width - 1)];
    PADDLE_ENFORCE(predict_data <= 1, "The predict data must less or equal 1.");
    PADDLE_ENFORCE(predict_data >= 0,
                   "The predict data must gather or equal 0.");
    uint32_t binIdx = static_cast<uint32_t>(predict_data * num_thresholds);
    if (label_data[i]) {
      paddle::platform::CudaAtomicAdd(pos + cur_step_begin + binIdx, 1);
    } else {
      paddle::platform::CudaAtomicAdd(neg + cur_step_begin + binIdx, 1);
    }
  }
}

__global__ void CalcAucKernel(int64_t *stat_pos,
                              int64_t *stat_neg,
                              int num_thresholds,
                              double *auc,
                              bool need_add_batch_num) {
  *auc = 0.0f;
  double totPos = 0.0;
  double totNeg = 0.0;
  double totPosPrev = 0.0;
  double totNegPrev = 0.0;

  int idx = num_thresholds;

  while (idx >= 0) {
    totPosPrev = totPos;
    totNegPrev = totNeg;
    totPos += stat_pos[idx];
    totNeg += stat_neg[idx];
    *auc += (totNeg - totNegPrev) * (totPos + totPosPrev) / 2.0;
    --idx;
  }

  if (totPos > 0.0 && totNeg > 0.0) {
    *auc = *auc / totPos / totNeg;
  }
  if (need_add_batch_num) {
    stat_pos[num_thresholds + 1] += 1;
    stat_neg[num_thresholds + 1] += 1;
  }
}

inline static double trapezoidArea(double X1, double X2, double Y1, double Y2) {
  return (X1 > X2 ? (X1 - X2) : (X2 - X1)) * (Y1 + Y2) / 2.0;
}

template <typename T, typename Context>
void statAuc(const Context &dev_ctx,
             const DenseTensor &label,
             const DenseTensor &predict,
             const int num_thresholds,
             const int slide_steps,
             int64_t *origin_stat_pos,
126 127
             int64_t *origin_stat_neg,
             const bool is_fake_data) {
128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
  size_t batch_size = predict.dims()[0];
  size_t inference_width = predict.dims()[1];
  const T *inference_data = predict.data<T>();
  const auto *label_data = label.data<int64_t>();
  const int bucket_length = num_thresholds + 1;

  if (slide_steps == 0) {
    AddDataKernel<<<(batch_size + PADDLE_CUDA_NUM_THREADS - 1) /
                        PADDLE_CUDA_NUM_THREADS,
                    PADDLE_CUDA_NUM_THREADS,
                    0,
                    dev_ctx.stream()>>>(label_data,
                                        inference_data,
                                        inference_width,
                                        num_thresholds,
                                        origin_stat_pos,
                                        origin_stat_neg,
                                        batch_size,
                                        slide_steps);
    return;
  }
  // the last number of origin_stat_pos store the index should be used in
  // current step
  int cur_step_index =
      static_cast<int>(origin_stat_pos[(slide_steps + 1) * bucket_length]) %
      slide_steps;
  int cur_step_begin = cur_step_index * bucket_length;
  int sum_step_begin = slide_steps * bucket_length;

  ClearObsoleteDataKernel<<<(bucket_length + PADDLE_CUDA_NUM_THREADS - 1) /
                                PADDLE_CUDA_NUM_THREADS,
                            PADDLE_CUDA_NUM_THREADS,
                            0,
                            dev_ctx.stream()>>>(
      origin_stat_pos, origin_stat_neg, bucket_length, slide_steps);

  AddDataKernel<<<(batch_size + PADDLE_CUDA_NUM_THREADS - 1) /
                      PADDLE_CUDA_NUM_THREADS,
                  PADDLE_CUDA_NUM_THREADS,
                  0,
                  dev_ctx.stream()>>>(label_data,
                                      inference_data,
                                      inference_width,
                                      num_thresholds,
                                      origin_stat_pos,
                                      origin_stat_neg,
                                      batch_size,
                                      slide_steps);
176 177 178 179 180 181 182 183
  if (!is_fake_data) {
    UpdateSumDataKernel<<<(bucket_length + PADDLE_CUDA_NUM_THREADS - 1) /
                              PADDLE_CUDA_NUM_THREADS,
                          PADDLE_CUDA_NUM_THREADS,
                          0,
                          dev_ctx.stream()>>>(
        origin_stat_pos, origin_stat_neg, bucket_length, slide_steps);
  }
184 185 186 187 188 189 190 191
}

template <typename T, typename Context>
void AucKernel(const Context &dev_ctx,
               const DenseTensor &input,
               const DenseTensor &label,
               const DenseTensor &stat_pos,
               const DenseTensor &stat_neg,
192
               const paddle::optional<DenseTensor> &ins_tag_weight,
193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208
               const std::string &curve,
               int num_thresholds,
               int slide_steps,
               DenseTensor *auc,
               DenseTensor *stat_pos_out,
               DenseTensor *stat_neg_out) {
  // Only use output var for now, make sure it's persistable and
  // not cleaned up for each batch.
  auto *origin_stat_pos = dev_ctx.template Alloc<int64_t>(stat_pos_out);
  auto *origin_stat_neg = dev_ctx.template Alloc<int64_t>(stat_neg_out);
  auto *auc_value = dev_ctx.template Alloc<double>(auc);

  auto *stat_pos_in_tensor = &stat_pos;
  auto *stat_neg_in_tensor = &stat_neg;
  auto *pos_in_data = stat_pos.data<int64_t>();
  auto *neg_in_data = stat_neg.data<int64_t>();
209 210 211 212 213 214 215 216
  bool is_fake_data = false;
  if (ins_tag_weight.get_ptr() != nullptr) {
    const auto *ins_tag_weight_data = ins_tag_weight->data<float>();
    if (ins_tag_weight_data[0] == 0) {
      is_fake_data = true;
    }
  }

217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252
#ifdef PADDLE_WITH_CUDA
  if (stat_pos_in_tensor != stat_pos_out) {
    cudaMemcpy(
        origin_stat_pos,
        pos_in_data,
        ((1 + slide_steps) * (num_thresholds + 1) + (slide_steps > 0 ? 1 : 0)) *
            sizeof(int64_t),
        cudaMemcpyDeviceToDevice);
  }
  if (stat_neg_in_tensor != stat_neg_out) {
    cudaMemcpy(
        origin_stat_neg,
        neg_in_data,
        ((1 + slide_steps) * (num_thresholds + 1) + (slide_steps > 0 ? 1 : 0)) *
            sizeof(int64_t),
        cudaMemcpyDeviceToDevice);
  }
#else
  if (stat_pos_in_tensor != stat_pos_out) {
    hipMemcpy(
        origin_stat_pos,
        pos_in_data,
        ((1 + slide_steps) * (num_thresholds + 1) + (slide_steps > 0 ? 1 : 0)) *
            sizeof(int64_t),
        hipMemcpyDeviceToDevice);
  }
  if (stat_neg_in_tensor != stat_neg_out) {
    hipMemcpy(
        origin_stat_neg,
        neg_in_data,
        ((1 + slide_steps) * (num_thresholds + 1) + (slide_steps > 0 ? 1 : 0)) *
            sizeof(int64_t),
        hipMemcpyDeviceToDevice);
  }
#endif

253 254 255 256 257
  // when calculate global_auc && is fake data, just do nothing
  if (slide_steps == 0 && is_fake_data) {
    return;
  }

258 259 260 261 262 263
  statAuc<T, Context>(dev_ctx,
                      label,
                      input,
                      num_thresholds,
                      slide_steps,
                      origin_stat_pos,
264 265
                      origin_stat_neg,
                      is_fake_data);
266 267 268 269 270 271 272 273 274 275 276
  int sum_offset = slide_steps * (num_thresholds + 1);
  CalcAucKernel<<<1, 1, 0, dev_ctx.stream()>>>(origin_stat_pos + sum_offset,
                                               origin_stat_neg + sum_offset,
                                               num_thresholds,
                                               auc_value,
                                               slide_steps > 0);
}

}  // namespace phi

PD_REGISTER_KERNEL(auc, GPU, ALL_LAYOUT, phi::AucKernel, float) {}