jit_kernel_macro.h 4.7 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
/* 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. */

#pragma once
#include <string>
#include "paddle/fluid/platform/cpu_info.h"

namespace paddle {
namespace operators {
namespace math {
namespace jitkernel {

namespace jit = platform::jit;

T
tensor-tang 已提交
26
#define SEARCH_BLOCK(macro_, ker, dtype, isa)                 \
T
tensor-tang 已提交
27
  if (d < AVX_FLOAT_BLOCK) {                                  \
T
tensor-tang 已提交
28
    macro_(ker, dtype, isa, kLT8);                            \
T
tensor-tang 已提交
29
  } else if (d == AVX_FLOAT_BLOCK) {                          \
T
tensor-tang 已提交
30
    macro_(ker, dtype, isa, kEQ8);                            \
T
tensor-tang 已提交
31
  } else if (d > AVX_FLOAT_BLOCK && d < AVX512_FLOAT_BLOCK) { \
T
tensor-tang 已提交
32
    macro_(ker, dtype, isa, kGT8LT16);                        \
T
tensor-tang 已提交
33
  } else if (d == AVX512_FLOAT_BLOCK) {                       \
T
tensor-tang 已提交
34
    macro_(ker, dtype, isa, kEQ16);                           \
T
tensor-tang 已提交
35
  } else {                                                    \
T
tensor-tang 已提交
36
    macro_(ker, dtype, isa, kGT16);                           \
T
tensor-tang 已提交
37 38
  }

T
tensor-tang 已提交
39 40 41 42 43 44 45 46 47
#define SEARCH_ISA_BLOCK(macro_, ker, dtype)        \
  if (jit::MayIUse(jit::avx512f)) {                 \
    SEARCH_BLOCK(macro_, ker, dtype, jit::avx512f); \
  } else if (jit::MayIUse(jit::avx2)) {             \
    SEARCH_BLOCK(macro_, ker, dtype, jit::avx2);    \
  } else if (jit::MayIUse(jit::avx)) {              \
    SEARCH_BLOCK(macro_, ker, dtype, jit::avx);     \
  } else {                                          \
    SEARCH_BLOCK(macro_, ker, dtype, jit::isa_any); \
T
tensor-tang 已提交
48 49
  }

T
tensor-tang 已提交
50 51 52 53 54 55 56 57 58 59
#define JITKERNEL_DECLARE(ker_class, ker_dtype) \
  template <>                                   \
  std::shared_ptr<const ker_class<ker_dtype>>   \
  KernelPool::Get<ker_class<ker_dtype>, int>(int d)

#define JITKERNEL_KEY(ker_key, dtype_key) \
  #ker_key #dtype_key + std::to_string(d)

#define JITKERNEL_NEW_IMPL(ker, dtype, isa, k) \
  p = std::dynamic_pointer_cast<ker<dtype>>(   \
T
tensor-tang 已提交
60
      std::make_shared<ker##Impl<dtype, isa, k>>(d))
T
tensor-tang 已提交
61 62 63 64 65 66 67 68 69 70 71 72 73

#define JITKERNEL_WITH_DTYPE(ker_key, ker_class, ker_dtype, dtype_key, \
                             marco_declare, macro_key, macro_impl)     \
  marco_declare(ker_class, ker_dtype) {                                \
    std::string key = macro_key(ker_key, dtype_key);                   \
    if (kers_.find(key) == kers_.end()) {                              \
      std::shared_ptr<ker_class<ker_dtype>> p;                         \
      SEARCH_ISA_BLOCK(macro_impl, ker_class, ker_dtype);              \
      kers_.insert({key, std::dynamic_pointer_cast<Kernel>(p)});       \
      return p;                                                        \
    }                                                                  \
    return std::dynamic_pointer_cast<const ker_class<ker_dtype>>(      \
        kers_.at(key));                                                \
T
tensor-tang 已提交
74 75
  }

T
tensor-tang 已提交
76 77 78 79 80 81 82 83 84 85 86 87
#define REGISTER_JITKERNEL(ker_key, ker_class)                           \
  JITKERNEL_WITH_DTYPE(ker_key, ker_class, float, f, JITKERNEL_DECLARE,  \
                       JITKERNEL_KEY, JITKERNEL_NEW_IMPL);               \
  JITKERNEL_WITH_DTYPE(ker_key, ker_class, double, d, JITKERNEL_DECLARE, \
                       JITKERNEL_KEY, JITKERNEL_NEW_IMPL)

#define REGISTER_JITKERNEL_ARGS(ker_key, ker_class, marco_declare, macro_key,  \
                                macro_impl)                                    \
  JITKERNEL_WITH_DTYPE(ker_key, ker_class, float, f, marco_declare, macro_key, \
                       macro_impl);                                            \
  JITKERNEL_WITH_DTYPE(ker_key, ker_class, double, d, marco_declare,           \
                       macro_key, macro_impl)
T
tensor-tang 已提交
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111

#define FOR_EACH_ISA(macro_, block) \
  macro_(jit::avx512f, block);      \
  macro_(jit::avx2, block);         \
  macro_(jit::avx, block);          \
  macro_(jit::isa_any, block)

#define FOR_EACH_BLOCK(macro_, isa) \
  macro_(isa, kLT8);                \
  macro_(isa, kEQ8);                \
  macro_(isa, kGT8LT16);            \
  macro_(isa, kEQ16);               \
  macro_(isa, kGT16)

#define FOR_EACH_ISA_BLOCK(macro_)      \
  FOR_EACH_BLOCK(macro_, jit::avx512f); \
  FOR_EACH_BLOCK(macro_, jit::avx2);    \
  FOR_EACH_BLOCK(macro_, jit::avx);     \
  FOR_EACH_BLOCK(macro_, jit::isa_any)

}  // namespace jitkernel
}  // namespace math
}  // namespace operators
}  // namespace paddle