conv_op_npu.cc 16.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
// Copyright (c) 2021 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/fluid/operators/conv_op.h"
16
#include "paddle/fluid/platform/device/npu/npu_op_runner.h"
17 18 19 20 21

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
22
using NPUDeviceContext = platform::NPUDeviceContext;
23

24
template <typename T>
25 26
class DepthwiseConvNPUKernel : public framework::OpKernel<T> {
 public:
27 28 29 30 31
  void Compute(const framework::ExecutionContext& ctx) const override {
    const Tensor* input = ctx.Input<Tensor>("Input");
    const Tensor* filter = ctx.Input<Tensor>("Filter");
    Tensor* output = ctx.Output<Tensor>("Output");
    output->mutable_data<T>(ctx.GetPlace());
32

33 34 35 36 37 38
    const std::vector<int> stride = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> padding = ctx.Attr<std::vector<int>>("paddings");
    std::vector<int> dilation = ctx.Attr<std::vector<int>>("dilations");
    const std::string data_format = ctx.Attr<std::string>("data_format");
    const std::string padding_algorithm =
        ctx.Attr<std::string>("padding_algorithm");
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

    const bool channel_last = data_format == "NHWC";
    if (channel_last) {
      PADDLE_ENFORCE_EQ(
          output->dims()[output->dims().size() - 1],
          input->dims()[input->dims().size() - 1],
          platform::errors::InvalidArgument(
              "ShapeError: The output channels must be equal to the "
              "input channels. But receivced output channel number is %d "
              "and input channel number is %d",
              output->dims()[output->dims().size() - 1],
              input->dims()[input->dims().size() - 1]));
    } else {
      PADDLE_ENFORCE_EQ(
          output->dims()[1], input->dims()[1],
          platform::errors::InvalidArgument(
              "ShapeError: The output channels must be equal to the "
              "input channels. But receivced output channel number is %d "
              "and input channel number is %d",
              output->dims()[1], input->dims()[1]));
    }

    auto in_dims = input->dims();
    auto filter_dims = filter->dims();
    framework::DDim in_data_dims;
    framework::DDim filter_data_dims;

    if (channel_last) {
      in_data_dims = framework::slice_ddim(in_dims, 1, in_dims.size() - 1);
    } else {
      in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
    }
    filter_data_dims = framework::slice_ddim(filter_dims, 2, in_dims.size());

    std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
    UpdatePaddingAndDilation(&padding, &dilation, padding_algorithm,
                             in_data_dims, stride, ksize);

    std::vector<int> strides(4, 1);
    std::vector<int> dilations(4, 1);

    Tensor input_tensor, output_tensor;
    input_tensor.ShareDataWith(*input);
    output_tensor.ShareDataWith(*output);

    if (channel_last) {
      input_tensor.set_layout(DataLayout::kNHWC);
      output_tensor.set_layout(DataLayout::kNHWC);
      strides[1] = stride[0];
      strides[2] = stride[1];
      dilations[1] = dilation[0];
      dilations[2] = dilation[1];
    } else {
      strides[2] = stride[0];
      strides[3] = stride[1];
      dilations[2] = dilation[0];
      dilations[3] = dilation[1];
    }

98 99 100 101 102 103 104 105 106 107 108 109
    auto stream = ctx.template device_context<NPUDeviceContext>().stream();

    // Transform filter (n, 1, h, w) --> (1, n, h, w)
    Tensor transformed_filter(filter->type());
    transformed_filter.mutable_data<T>({filter->dims()[1], filter->dims()[0],
                                        filter->dims()[2], filter->dims()[3]},
                                       ctx.device_context().GetPlace());
    std::vector<int> perm = {1, 0, 2, 3};
    const auto& runner_trans = NpuOpRunner(
        "TransposeD", {*filter}, {transformed_filter}, {{"perm", perm}});
    runner_trans.Run(stream);

110 111 112 113 114 115 116 117 118 119
    const auto& runner =
        NpuOpRunner("DepthwiseConv2D", {input_tensor, transformed_filter},
                    {output_tensor}, {{"strides", strides},
                                      {"dilations", dilations},
                                      {"pads", padding},
                                      {"data_format", data_format}});
    runner.Run(stream);
  }
};

120 121 122
template <typename T>
class DepthwiseConvGradNPUKernel : public framework::OpKernel<T> {
 public:
123 124 125 126 127 128
  void Compute(const framework::ExecutionContext& ctx) const override {
    const Tensor* input = ctx.Input<Tensor>("Input");
    const Tensor* filter = ctx.Input<Tensor>("Filter");
    auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
    auto input_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
    auto filter_grad = ctx.Output<Tensor>(framework::GradVarName("Filter"));
129

130 131 132 133 134 135
    const std::vector<int> stride = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> padding = ctx.Attr<std::vector<int>>("paddings");
    std::vector<int> dilation = ctx.Attr<std::vector<int>>("dilations");
    const std::string data_format = ctx.Attr<std::string>("data_format");
    const std::string padding_algorithm =
        ctx.Attr<std::string>("padding_algorithm");
136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155

    const bool channel_last = data_format == "NHWC";

    // update padding and dilation
    auto in_dims = input->dims();
    auto filter_dims = filter->dims();
    framework::DDim in_data_dims;
    framework::DDim filter_data_dims;

    if (channel_last) {
      in_data_dims = framework::slice_ddim(in_dims, 1, in_dims.size() - 1);
    } else {
      in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
    }
    filter_data_dims = framework::slice_ddim(filter_dims, 2, in_dims.size());

    std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
    UpdatePaddingAndDilation(&padding, &dilation, padding_algorithm,
                             in_data_dims, stride, ksize);

156 157
    auto stream = ctx.template device_context<NPUDeviceContext>().stream();

158 159 160 161
    // Transform filter (n, 1, h, w) --> (1, n, h, w)
    Tensor transformed_filter(filter->type());
    transformed_filter.mutable_data<T>({filter->dims()[1], filter->dims()[0],
                                        filter->dims()[2], filter->dims()[3]},
162
                                       ctx.device_context().GetPlace());
163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189
    std::vector<int> perm = {1, 0, 2, 3};
    const auto& runner_trans = NpuOpRunner(
        "TransposeD", {*filter}, {transformed_filter}, {{"perm", perm}});
    runner_trans.Run(stream);

    // construct NPU attr
    std::vector<int> strides(4, 1);
    std::vector<int> dilations(4, 1);

    Tensor input_tensor, output_grad_tensor;
    input_tensor.ShareDataWith(*input);
    output_grad_tensor.ShareDataWith(*output_grad);
    if (channel_last) {
      input_tensor.set_layout(DataLayout::kNHWC);
      output_grad_tensor.set_layout(DataLayout::kNHWC);
      strides[1] = stride[0];
      strides[2] = stride[1];
      dilations[1] = dilation[0];
      dilations[2] = dilation[1];
    } else {
      strides[2] = stride[0];
      strides[3] = stride[1];
      dilations[2] = dilation[0];
      dilations[3] = dilation[1];
    }

    if (filter_grad) {
190
      filter_grad->mutable_data<T>(ctx.GetPlace());
191

192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210
      PADDLE_ENFORCE_EQ(
          (dilations[2] == 1 && dilations[3] == 1), true,
          platform::errors::InvalidArgument(
              "dilation_h and dilation_w in DepthwiseConv2DBackpropFilterD "
              "must be equal to 1, but got dilation_h %d, dilation_w %d",
              dilation[2], dilation[3]));

      NpuOpRunner runner;
      runner.SetType("DepthwiseConv2DBackpropFilterD")
          .AddInput(input_tensor)
          .AddInput(output_grad_tensor)
          .AddOutput(*filter_grad)
          .AddAttr("filter_size",
                   framework::vectorize(transformed_filter.dims()))
          .AddAttr("strides", strides)
          .AddAttr("dilations", dilations)
          .AddAttr("pads", padding)
          .AddAttr("data_format", data_format)
          .Run(stream);
211 212
    }
    if (input_grad) {
213
      input_grad->mutable_data<T>(ctx.GetPlace());
214 215 216 217 218
      Tensor input_grad_tensor;
      input_grad_tensor.ShareDataWith(*input_grad);
      if (channel_last) {
        input_grad_tensor.set_layout(DataLayout::kNHWC);
      }
219 220 221 222 223 224 225 226 227 228 229
      NpuOpRunner runner;
      runner.SetType("DepthwiseConv2DBackpropInputD")
          .AddInput(transformed_filter)
          .AddInput(output_grad_tensor)
          .AddOutput(input_grad_tensor)
          .AddAttr("input_size", framework::vectorize(input->dims()))
          .AddAttr("strides", strides)
          .AddAttr("dilations", dilations)
          .AddAttr("pads", padding)
          .AddAttr("data_format", data_format)
          .Run(stream);
230 231 232 233
    }
  }
};

234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288
template <typename T>
class NPUConvOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    const Tensor* input = ctx.Input<Tensor>("Input");
    auto* filter = ctx.Input<Tensor>("Filter");
    auto* output = ctx.Output<Tensor>("Output");
    output->mutable_data<T>(ctx.GetPlace());
    const std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
    int groups = ctx.Attr<int>("groups");
    const std::string padding_algorithm =
        ctx.Attr<std::string>("padding_algorithm");
    const std::string data_format = ctx.Attr<std::string>("data_format");

    const bool channel_last = data_format == "NHWC";

    // update padding and dilation
    auto in_dims = input->dims();
    auto filter_dims = filter->dims();
    framework::DDim in_data_dims;
    framework::DDim filter_data_dims;

    if (channel_last) {
      in_data_dims = framework::slice_ddim(in_dims, 1, in_dims.size() - 1);
    } else {
      in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
    }
    filter_data_dims = framework::slice_ddim(filter_dims, 2, in_dims.size());

    std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
    UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm,
                             in_data_dims, strides, ksize);

    std::vector<int> strides_vec(4, 1);
    std::vector<int> dilations_vec(4, 1);

    Tensor input_tensor, output_tensor;
    input_tensor.ShareDataWith(*input);
    output_tensor.ShareDataWith(*output);
    if (channel_last) {
      input_tensor.set_layout(DataLayout::kNHWC);
      output_tensor.set_layout(DataLayout::kNHWC);
      strides_vec[1] = strides[0];
      strides_vec[2] = strides[1];
      dilations_vec[1] = dilations[0];
      dilations_vec[2] = dilations[1];
    } else {
      strides_vec[2] = strides[0];
      strides_vec[3] = strides[1];
      dilations_vec[2] = dilations[0];
      dilations_vec[3] = dilations[1];
    }

289
    auto stream = ctx.template device_context<NPUDeviceContext>().stream();
290 291 292 293 294 295 296
    const auto& runner =
        NpuOpRunner("Conv2D", {input_tensor, *filter}, {output_tensor},
                    {{"strides", strides_vec},
                     {"pads", paddings},
                     {"dilations", dilations_vec},
                     {"groups", groups},
                     {"data_format", data_format}});
297
    runner.Run(stream);
298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357
  }
};

template <typename T>
class NPUConvGradOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto input = ctx.Input<Tensor>("Input");
    auto filter = ctx.Input<Tensor>("Filter");
    auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
    auto input_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
    auto filter_grad = ctx.Output<Tensor>(framework::GradVarName("Filter"));

    const std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
    int groups = ctx.Attr<int>("groups");
    const std::string padding_algorithm =
        ctx.Attr<std::string>("padding_algorithm");
    const std::string data_format = ctx.Attr<std::string>("data_format");

    const bool channel_last = data_format == "NHWC";

    // update padding and dilation
    auto in_dims = input->dims();
    auto filter_dims = filter->dims();
    framework::DDim in_data_dims;
    framework::DDim filter_data_dims;

    if (channel_last) {
      in_data_dims = framework::slice_ddim(in_dims, 1, in_dims.size() - 1);
    } else {
      in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
    }
    filter_data_dims = framework::slice_ddim(filter_dims, 2, in_dims.size());

    std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
    UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm,
                             in_data_dims, strides, ksize);

    std::vector<int> strides_vec(4, 1);
    std::vector<int> dilations_vec(4, 1);

    Tensor input_tensor, output_grad_tensor;
    input_tensor.ShareDataWith(*input);
    output_grad_tensor.ShareDataWith(*output_grad);
    if (channel_last) {
      input_tensor.set_layout(DataLayout::kNHWC);
      output_grad_tensor.set_layout(DataLayout::kNHWC);
      strides_vec[1] = strides[0];
      strides_vec[2] = strides[1];
      dilations_vec[1] = dilations[0];
      dilations_vec[2] = dilations[1];
    } else {
      strides_vec[2] = strides[0];
      strides_vec[3] = strides[1];
      dilations_vec[2] = dilations[0];
      dilations_vec[3] = dilations[1];
    }

358
    auto stream = ctx.template device_context<NPUDeviceContext>().stream();
359 360 361 362 363 364 365 366 367 368 369 370 371
    if (filter_grad) {
      filter_grad->mutable_data<T>(ctx.GetPlace());
      std::vector<int> filter_shape_vec =
          framework::vectorize<int>(filter->dims());

      const auto& runner = NpuOpRunner(
          "Conv2DBackpropFilterD", {input_tensor, output_grad_tensor},
          {*filter_grad}, {{"filter_size", filter_shape_vec},
                           {"strides", strides_vec},
                           {"pads", paddings},
                           {"dilations", dilations_vec},
                           {"groups", groups},
                           {"data_format", data_format}});
372
      runner.Run(stream);
373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391
    }
    if (input_grad) {
      input_grad->mutable_data<T>(ctx.GetPlace());
      std::vector<int> input_shape_vec =
          framework::vectorize<int>(input->dims());

      Tensor input_grad_tensor;
      input_grad_tensor.ShareDataWith(*input_grad);
      if (channel_last) {
        input_grad_tensor.set_layout(DataLayout::kNHWC);
      }
      const auto& runner =
          NpuOpRunner("Conv2DBackpropInputD", {*filter, output_grad_tensor},
                      {input_grad_tensor}, {{"input_size", input_shape_vec},
                                            {"strides", strides_vec},
                                            {"pads", paddings},
                                            {"dilations", dilations_vec},
                                            {"groups", groups},
                                            {"data_format", data_format}});
392
      runner.Run(stream);
393 394 395
    }
  }
};
396 397 398 399
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
400 401 402 403 404 405 406 407
namespace plat = paddle::platform;

REGISTER_OP_NPU_KERNEL(depthwise_conv2d, ops::DepthwiseConvNPUKernel<float>,
                       ops::DepthwiseConvNPUKernel<plat::float16>);

REGISTER_OP_NPU_KERNEL(depthwise_conv2d_grad,
                       ops::DepthwiseConvGradNPUKernel<float>,
                       ops::DepthwiseConvGradNPUKernel<plat::float16>);
408

409
REGISTER_OP_NPU_KERNEL(conv2d, ops::NPUConvOpKernel<float>,
410 411
                       ops::NPUConvOpKernel<plat::float16>);

412
REGISTER_OP_NPU_KERNEL(conv2d_grad, ops::NPUConvGradOpKernel<float>,
413
                       ops::NPUConvGradOpKernel<plat::float16>);