seqpool.cc 2.8 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2018 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/jit/gen/seqpool.h"
T
tensor-tang 已提交
16
#include <stddef.h>  // offsetof
T
tensor-tang 已提交
17 18 19 20 21 22 23 24
#include "paddle/fluid/operators/jit/registry.h"
#include "paddle/fluid/platform/cpu_info.h"

namespace paddle {
namespace operators {
namespace jit {
namespace gen {

T
tensor-tang 已提交
25 26 27
thread_local float ALIGN32_BEG float_h[1] ALIGN32_END = {
    1.f};  // TODO(TJ): try move to private

T
tensor-tang 已提交
28 29 30 31 32 33
void SeqPoolJitCode::genCode() {
  constexpr int block = YMM_FLOAT_BLOCK;
  constexpr int max_num_regs = 8;
  const int num_block = w_ / block;
  const int num_groups = num_block / max_num_regs;
  int rest_num_regs = num_block % max_num_regs;
T
tensor-tang 已提交
34 35 36 37 38 39
  mov(reg32_int_h, dword[param_attr]);
  if (type_ == SeqPoolType::kAvg || type_ == SeqPoolType::kSqrt) {
    mov(reg_tmp, reinterpret_cast<size_t>(float_h));
    fild(dword[param_attr]);
    fstp(dword[reg_tmp]);
    mov(reg32_fp_h, dword[reg_tmp]);
T
tensor-tang 已提交
40 41 42 43 44 45 46 47 48
  }
  const int group_len = max_num_regs * block * sizeof(float);
  for (int g = 0; g < num_groups; ++g) {
    pool_height<ymm_t>(g * group_len, block, max_num_regs);
  }
  if (rest_num_regs > 0) {
    pool_height<ymm_t>(num_groups * group_len, block, rest_num_regs);
  }

T
tensor-tang 已提交
49
  // part of rest_w * height
T
tensor-tang 已提交
50
  const int rest = w_ % block;
T
tensor-tang 已提交
51
  pool_height_of_rest_width(rest, (w_ - rest) * sizeof(float), max_num_regs);
T
tensor-tang 已提交
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
  ret();
}

class SeqPoolCreator : public JitCodeCreator<seq_pool_attr_t> {
 public:
  bool UseMe(const seq_pool_attr_t& attr) const override {
    return platform::MayIUse(platform::avx);
  }
  size_t CodeSize(const seq_pool_attr_t& attr) const override {
    // TODO(TJ): remove attr.h when enabled height
    bool yes =
        attr.type == SeqPoolType::kAvg || attr.type == SeqPoolType::kSqrt;
    return 96 /* basic */ +
           ((attr.w / YMM_FLOAT_BLOCK + 4 /* rest */) * 2 /* for sum */
            * (attr.h + (yes ? 3 : 1 /*for avg or sqrt*/))) *
               8;
  }
  std::unique_ptr<GenBase> CreateJitCode(
      const seq_pool_attr_t& attr) const override {
    PADDLE_ENFORCE_GT(attr.w, 0);
    PADDLE_ENFORCE_GT(attr.h, 0);
    return make_unique<SeqPoolJitCode>(attr, CodeSize(attr));
  }
};

}  // namespace gen
}  // namespace jit
}  // namespace operators
}  // namespace paddle

namespace gen = paddle::operators::jit::gen;

REGISTER_JITKERNEL_GEN(kSeqPool, gen::SeqPoolCreator);