math_function.cu 14.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 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 <algorithm>
#include <vector>
16

17 18 19 20 21
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/memory/malloc.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/platform/bfloat16.h"
#include "paddle/fluid/platform/float16.h"
22 23 24 25
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
#include "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/phi/kernels/funcs/math_function_impl.h"
26

27
namespace phi {
28 29
namespace funcs {

30 31
using float16 = phi::dtype::float16;
using bfloat16 = phi::dtype::bfloat16;
32

33 34 35 36 37 38 39 40 41 42 43
template struct SetConstant<phi::GPUContext, phi::dtype::float16>;
template struct SetConstant<phi::GPUContext, phi::dtype::bfloat16>;
template struct SetConstant<phi::GPUContext, float>;
template struct SetConstant<phi::GPUContext, double>;
template struct SetConstant<phi::GPUContext, uint8_t>;
template struct SetConstant<phi::GPUContext, int>;
template struct SetConstant<phi::GPUContext, int16_t>;
template struct SetConstant<phi::GPUContext, int64_t>;
template struct SetConstant<phi::GPUContext, bool>;
template struct SetConstant<phi::GPUContext, phi::dtype::complex<float>>;
template struct SetConstant<phi::GPUContext, phi::dtype::complex<double>>;
44 45

template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext,
46
                            phi::dtype::float16>;
47
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext,
48
                            phi::dtype::bfloat16>;
49 50 51 52 53 54 55 56
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext, float>;
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext, double>;
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext, uint8_t>;
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext, int>;
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext, int16_t>;
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext, int64_t>;
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext, bool>;
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext,
57
                            phi::dtype::complex<float>>;
58
template struct SetConstant<paddle::platform::CUDAPinnedDeviceContext,
59
                            phi::dtype::complex<double>>;
60

61 62 63 64 65 66 67 68 69 70 71 72
#define DEFINE_GPU_TRANS(RANK)                                \
  template struct Transpose<phi::GPUContext, bool, RANK>;     \
  template struct Transpose<phi::GPUContext, float, RANK>;    \
  template struct Transpose<phi::GPUContext, double, RANK>;   \
  template struct Transpose<phi::GPUContext, float16, RANK>;  \
  template struct Transpose<phi::GPUContext, bfloat16, RANK>; \
  template struct Transpose<phi::GPUContext, int8_t, RANK>;   \
  template struct Transpose<phi::GPUContext, int32_t, RANK>;  \
  template struct Transpose<phi::GPUContext, int64_t, RANK>;  \
  template struct Transpose<phi::GPUContext,                  \
                            phi::dtype::complex<float>,       \
                            RANK>;                            \
73
  template struct Transpose<phi::GPUContext, phi::dtype::complex<double>, RANK>;
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

DEFINE_GPU_TRANS(1);
DEFINE_GPU_TRANS(2);
DEFINE_GPU_TRANS(3);
DEFINE_GPU_TRANS(4);
DEFINE_GPU_TRANS(5);
DEFINE_GPU_TRANS(6);

#define REINTERPRET(T, DST_PTR, SRC_PTR) \
  T* DST_PTR = reinterpret_cast<T*>(SRC_PTR)

template <typename T>
__global__ void TransposeNormalKernel(const T* in_ptr,
                                      T* out_ptr,
                                      int64_t element,
                                      const int64_t* in_stride_ptr,
                                      const int64_t* out_stride_ptr,
                                      const int64_t* axis_ptr,
                                      int rank) {
  CUDA_KERNEL_LOOP(out_idx, element) {
    int64_t in_idx = 0;
    int64_t tmp_idx = out_idx;
    for (int i = 0; i < rank; ++i) {
      const int64_t coordinate = tmp_idx / out_stride_ptr[i];
      tmp_idx -= coordinate * out_stride_ptr[i];
      in_idx += coordinate * in_stride_ptr[axis_ptr[i]];
    }
    out_ptr[out_idx] = in_ptr[in_idx];
  }
}

105 106 107 108 109 110 111
template <typename DeviceContext, typename T>
void TransposeNormal<DeviceContext, T>::operator()(
    const DeviceContext& context,
    const paddle::framework::Tensor& in,
    paddle::framework::Tensor* out,
    const std::vector<int>& axis) {
  const int rank = axis.size();
112 113
  auto in_stride = phi::stride(in.dims());
  auto out_stride = phi::stride(out->dims());
114 115
  auto* in_ptr = in.data<T>();
  auto* out_ptr = out->data<T>();
116

117 118 119 120 121 122 123 124 125 126 127 128
  // copy in_stride, out_stride, axis to gpu device
  const paddle::platform::CUDAPlace& cuda_place = context.GetPlace();
  paddle::platform::CPUPlace cpu_place = paddle::platform::CPUPlace();
  size_t size = 3 * rank * sizeof(int64_t);
  auto cpu_buf_holder = paddle::memory::Alloc(cpu_place, size);
  auto cuda_buf_holder = paddle::memory::Alloc(cuda_place, size);
  REINTERPRET(int64_t, cpu_buf, cpu_buf_holder->ptr());
  REINTERPRET(int64_t, cuda_buf, cuda_buf_holder->ptr());
  for (int i = 0; i < rank; ++i) {
    cpu_buf[i] = in_stride[i];
    cpu_buf[rank + i] = out_stride[i];
    cpu_buf[2 * rank + i] = axis[i];
129
  }
130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146
  paddle::memory::Copy(
      cuda_place, cuda_buf, cpu_place, cpu_buf, size, context.stream());
  REINTERPRET(const int64_t, in_stride_ptr, cuda_buf);
  REINTERPRET(const int64_t, out_stride_ptr, cuda_buf + rank);
  REINTERPRET(const int64_t, axis_ptr, cuda_buf + 2 * rank);

  const int MAX_BLOCK_DIM = context.GetMaxThreadsPerBlock();
  const int MAX_GRID_DIM = context.GetMaxPhysicalThreadCount() / MAX_BLOCK_DIM;
  int64_t elements = in.numel();
  int block_size = (elements >= MAX_BLOCK_DIM)
                       ? MAX_BLOCK_DIM
                       : (1 << static_cast<int>(std::log2(elements)));
  int grid_size = elements / block_size;
  grid_size = (grid_size >= MAX_GRID_DIM) ? MAX_GRID_DIM : grid_size;
  TransposeNormalKernel<T><<<grid_size, block_size, 0, context.stream()>>>(
      in_ptr, out_ptr, elements, in_stride_ptr, out_stride_ptr, axis_ptr, rank);
}
147

H
hong 已提交
148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
template <typename T>
struct TransposeNormal<phi::GPUContext, T> {
  void operator()(const phi::GPUContext& context,
                  const DenseTensor& in,
                  DenseTensor* out,
                  const std::vector<int>& axis) {
    const int rank = axis.size();
    auto in_stride = stride(in.dims());
    auto out_stride = stride(out->dims());
    auto* in_ptr = in.data<T>();
    auto* out_ptr = out->data<T>();

    // copy in_stride, out_stride, axis to gpu device
    const phi::GPUPlace& cuda_place = context.GetPlace();
    phi::CPUPlace cpu_place = paddle::platform::CPUPlace();
    size_t size = 3 * rank * sizeof(int64_t);
    auto cpu_buf_holder = paddle::memory::Alloc(cpu_place, size);
    auto cuda_buf_holder = paddle::memory::Alloc(cuda_place, size);
    REINTERPRET(int64_t, cpu_buf, cpu_buf_holder->ptr());
    REINTERPRET(int64_t, cuda_buf, cuda_buf_holder->ptr());
    for (int i = 0; i < rank; ++i) {
      cpu_buf[i] = in_stride[i];
      cpu_buf[rank + i] = out_stride[i];
      cpu_buf[2 * rank + i] = axis[i];
    }
    paddle::memory::Copy(
        cuda_place, cuda_buf, cpu_place, cpu_buf, size, context.stream());
    REINTERPRET(const int64_t, in_stride_ptr, cuda_buf);
    REINTERPRET(const int64_t, out_stride_ptr, cuda_buf + rank);
    REINTERPRET(const int64_t, axis_ptr, cuda_buf + 2 * rank);

    const int MAX_BLOCK_DIM = context.GetMaxThreadsPerBlock();
    const int MAX_GRID_DIM =
        context.GetMaxPhysicalThreadCount() / MAX_BLOCK_DIM;
    int64_t elements = in.numel();
    int block_size = (elements >= MAX_BLOCK_DIM)
                         ? MAX_BLOCK_DIM
                         : (1 << static_cast<int>(std::log2(elements)));
    int grid_size = elements / block_size;
    grid_size = (grid_size >= MAX_GRID_DIM) ? MAX_GRID_DIM : grid_size;
188 189 190 191 192 193 194 195
    TransposeNormalKernel<T>
        <<<grid_size, block_size, 0, context.stream()>>>(in_ptr,
                                                         out_ptr,
                                                         elements,
                                                         in_stride_ptr,
                                                         out_stride_ptr,
                                                         axis_ptr,
                                                         rank);
H
hong 已提交
196 197 198
  }
};

199
// define transpose normal
200
#define DEFINE_GPU_TRANS_NORMAL(TYPE) \
201
  template struct TransposeNormal<phi::GPUContext, TYPE>
202 203 204 205 206 207 208 209 210 211 212

DEFINE_GPU_TRANS_NORMAL(float16);
DEFINE_GPU_TRANS_NORMAL(bfloat16);
DEFINE_GPU_TRANS_NORMAL(float);
DEFINE_GPU_TRANS_NORMAL(double);
DEFINE_GPU_TRANS_NORMAL(int);
DEFINE_GPU_TRANS_NORMAL(int64_t);
DEFINE_GPU_TRANS_NORMAL(bool);
DEFINE_GPU_TRANS_NORMAL(int16_t);
DEFINE_GPU_TRANS_NORMAL(uint8_t);
DEFINE_GPU_TRANS_NORMAL(int8_t);
213 214
DEFINE_GPU_TRANS_NORMAL(phi::dtype::complex<float>);
DEFINE_GPU_TRANS_NORMAL(phi::dtype::complex<double>);
215 216 217 218 219 220 221 222 223

struct TensorSetConstantGPU {
  TensorSetConstantGPU(const paddle::platform::DeviceContext& context,
                       paddle::framework::Tensor* tensor,
                       float value)
      : context_(context), tensor_(tensor), value_(value) {}

  template <typename T>
  void apply() const {
L
Leo Chen 已提交
224 225 226 227
    SetConstant<phi::GPUContext, T> functor;
    functor(reinterpret_cast<const phi::GPUContext&>(context_),
            tensor_,
            static_cast<T>(value_));
228 229 230 231 232 233 234 235 236 237 238 239
  }

  const paddle::platform::DeviceContext& context_;
  paddle::framework::Tensor* tensor_;
  float value_;
};

template <>
void set_constant_with_place<paddle::platform::CUDAPlace>(
    const paddle::platform::DeviceContext& context,
    paddle::framework::Tensor* tensor,
    float value) {
240 241
  phi::VisitDataType(tensor->dtype(),
                     TensorSetConstantGPU(context, tensor, value));
242 243 244 245 246 247 248 249 250 251 252 253 254 255
}

template <typename T>
__global__ void RowwiseAddKernel(
    const T* a, const T* b, T* c, int width, int num) {
  T tmp = 1.0 / width;
  CUDA_KERNEL_LOOP(i, num) {
    int h = i * tmp;
    int w = i - h * width;
    c[i] = a[i] + b[w];
  }
}

template <typename T>
L
Leo Chen 已提交
256 257
struct RowwiseAdd<phi::GPUContext, T> {
  void operator()(const phi::GPUContext& context,
258 259 260 261 262 263 264 265 266
                  const paddle::framework::Tensor& input,
                  const paddle::framework::Tensor& vector,
                  paddle::framework::Tensor* output) {
    auto in_dims = input.dims();
    auto out_dims = output->dims();
    auto size = input.numel() / in_dims[0];
    PADDLE_ENFORCE_EQ(
        vector.numel(),
        size,
267
        phi::errors::InvalidArgument(
268 269 270 271 272 273 274 275 276 277
            "The input vector size"
            " should be equal to the size of each row of input tensor."
            " Expected vector size=%d, but received %d",
            size,
            vector.numel()));
    const char* in_dims_cstr = in_dims.to_str().c_str();
    const char* out_dims_cstr = out_dims.to_str().c_str();
    PADDLE_ENFORCE_EQ(
        out_dims,
        in_dims,
278
        phi::errors::InvalidArgument(
279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
            "The output tensor shape should be same as the input tensor"
            " shape. Expected output tensor shape: %s,"
            " but received %s",
            in_dims_cstr,
            out_dims_cstr));
    int blocks = 512;
    int grids = (input.numel() + blocks - 1) / blocks;
    RowwiseAddKernel<T><<<grids, blocks, 0, context.stream()>>>(
        input.data<T>(),
        vector.data<T>(),
        output->data<T>(),
        static_cast<int>(in_dims[1]),
        static_cast<int>(input.numel()));
  }
};

L
Leo Chen 已提交
295 296 297 298 299 300 301
template struct RowwiseAdd<phi::GPUContext, float>;
template struct RowwiseAdd<phi::GPUContext, double>;
template struct ColwiseSum<phi::GPUContext, float>;
template struct ColwiseSum<phi::GPUContext, int>;
template struct ColwiseSum<phi::GPUContext, int64_t>;
// template struct ColwiseSum<phi::GPUContext, double>;
// The ColwiseSum<phi::GPUContext, double> failed in debug
302 303 304
// mode,
// and only failed for this case. So reimplemented it.
template <>
L
Leo Chen 已提交
305 306
void ColwiseSum<phi::GPUContext, double>::operator()(
    const phi::GPUContext& context,
307 308 309 310 311 312
    const paddle::framework::Tensor& input,
    paddle::framework::Tensor* vector) {
  auto in_dims = input.dims();
  auto size = input.numel() / in_dims[0];
  PADDLE_ENFORCE_EQ(vector->numel(),
                    size,
313
                    phi::errors::InvalidArgument(
314 315 316 317 318 319 320
                        "The size of input vector"
                        " should be equal to the size of input tensor column"
                        " dimension. Expected vector size=%d, but received %d",
                        size,
                        vector->numel()));
  paddle::framework::Tensor one;
  one.mutable_data<double>({in_dims[0]}, context.GetPlace());
L
Leo Chen 已提交
321
  SetConstant<phi::GPUContext, double> set;
322
  set(context, &one, static_cast<double>(1.0));
L
Leo Chen 已提交
323 324 325 326 327 328 329 330 331
  phi::funcs::GetBlas<phi::GPUContext, double>(context).GEMV(
      true,
      static_cast<int>(in_dims[0]),
      static_cast<int>(in_dims[1]),
      1.0,
      input.data<double>(),
      one.data<double>(),
      0.0,
      vector->data<double>());
332 333
}

L
Leo Chen 已提交
334 335
template struct RowwiseSum<phi::GPUContext, float>;
// template struct RowwiseSum<phi::GPUContext, double>;
336
// TODO(zcd): Following ColwiseSum format, need to confirm.
L
Leo Chen 已提交
337
// The RowwiseSum<phi::GPUContext, double> failed in debug
338 339 340
// mode,
// and only failed for this case. So reimplemented it.
template <>
L
Leo Chen 已提交
341 342
void RowwiseSum<phi::GPUContext, double>::operator()(
    const phi::GPUContext& context,
343 344 345 346 347 348
    const paddle::framework::Tensor& input,
    paddle::framework::Tensor* vector) {
  auto in_dims = input.dims();
  auto size = input.numel() / in_dims[0];
  PADDLE_ENFORCE_EQ(vector->numel(),
                    in_dims[0],
349
                    phi::errors::InvalidArgument(
350 351 352 353 354 355 356
                        "The size of input vector"
                        " should be equal to the size of input tensor row"
                        " dimension. Expected vector size=%d, but received %d",
                        in_dims[0],
                        vector->numel()));
  paddle::framework::Tensor one;
  one.mutable_data<double>({size}, context.GetPlace());
L
Leo Chen 已提交
357
  SetConstant<phi::GPUContext, double> set;
358
  set(context, &one, static_cast<double>(1.0));
L
Leo Chen 已提交
359 360 361 362 363 364 365 366 367
  phi::funcs::GetBlas<phi::GPUContext, double>(context).GEMV(
      true,
      static_cast<int>(in_dims[1]),
      static_cast<int>(in_dims[0]),
      1.0,
      one.data<double>(),
      input.data<double>(),
      0.0,
      vector->data<double>());
368 369
}

L
Leo Chen 已提交
370 371
template struct RowwiseMean<phi::GPUContext, float>;
template struct RowwiseMean<phi::GPUContext, double>;
372 373

}  // namespace funcs
374
}  // namespace phi