blas.h 3.7 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
/* 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>
18
#include "glog/logging.h"
T
tensor-tang 已提交
19
#include "paddle/fluid/operators/jit/gen/jitcode.h"
T
tensor-tang 已提交
20 21 22

namespace paddle {
namespace operators {
T
tensor-tang 已提交
23 24
namespace jit {
namespace gen {
T
tensor-tang 已提交
25 26 27 28

// function: vec = Operand(vec(or scalar), vec(or scalar)) (maybe with relu)
class VXXJitCode : public JitCode {
 public:
T
tensor-tang 已提交
29 30 31 32 33 34 35 36
  explicit VXXJitCode(int d, operand_type type, int scalar_index,
                      bool with_relu, size_t code_size = 256 * 1024,
                      void* code_ptr = nullptr)
      : JitCode(code_size, code_ptr),
        num_(d),
        type_(type),
        scalar_index_(scalar_index),
        with_relu_(with_relu) {
T
tensor-tang 已提交
37 38
    if (!(type_ == operand_type::MUL || type_ == operand_type::ADD ||
          type_ == operand_type::SUB)) {
39 40
      LOG(FATAL) << "Do not support this operand type: " << type_;
    }
T
tensor-tang 已提交
41 42 43 44
    this->genCode();
  }

  virtual const char* name() const {
T
tensor-tang 已提交
45 46 47 48 49 50
    std::string base = "VXXJitCode";
    if (scalar_index_ == 1) {
      base += "_Scalar";
    } else {
      base += "_Vec";
    }
51
    if (type_ == operand_type::MUL) {
T
tensor-tang 已提交
52
      base += "_Mul";
53
    } else if (type_ == operand_type::ADD) {
T
tensor-tang 已提交
54
      base += "_Add";
T
tensor-tang 已提交
55 56
    } else if (type_ == operand_type::SUB) {
      base += "_SUB";
T
tensor-tang 已提交
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
    }
    if (scalar_index_ == 2) {
      base += "_Scalar";
    } else {
      base += "_Vec";
    }
    base += (with_relu_ ? "_Relu" : "");
    return base.c_str();
  }
  void genCode() override;

 private:
  int num_;
  operand_type type_;
  int scalar_index_;
  bool with_relu_;
  reg64_t param1{abi_param1};
  reg64_t param2{abi_param2};
  reg64_t param3{abi_param3};

  xmm_t xmm_src1 = xmm_t(0);
  xmm_t xmm_src2 = xmm_t(1);
  xmm_t xmm_dst = xmm_t(2);
  xmm_t xmm_zero = xmm_t(3);

  ymm_t ymm_src1 = ymm_t(0);
  ymm_t ymm_src2 = ymm_t(1);
  ymm_t ymm_dst = ymm_t(2);
  ymm_t ymm_zero = ymm_t(3);
};

88 89 90 91 92 93 94 95
#define DECLARE_BLAS_JITCODE(name, op_type, scalar_idx, with_relu)             \
  class name##JitCode : public VXXJitCode {                                    \
   public:                                                                     \
    explicit name##JitCode(int d, size_t code_size, void* code_ptr = nullptr)  \
        : VXXJitCode(d, op_type, scalar_idx, with_relu, code_size, code_ptr) { \
    }                                                                          \
  };

96 97 98 99 100 101
DECLARE_BLAS_JITCODE(VMul, operand_type::MUL, 0, false);
DECLARE_BLAS_JITCODE(VAdd, operand_type::ADD, 0, false);
DECLARE_BLAS_JITCODE(VSub, operand_type::SUB, 0, false);
DECLARE_BLAS_JITCODE(VAddRelu, operand_type::ADD, 0, true);
DECLARE_BLAS_JITCODE(VScal, operand_type::MUL, 1, false);
DECLARE_BLAS_JITCODE(VAddBias, operand_type::ADD, 1, false);
102 103

#undef DECLARE_BLAS_JITCODE
T
tensor-tang 已提交
104

T
tensor-tang 已提交
105 106 107 108 109 110 111 112 113 114 115 116
// nChw16c = nChw16c .* NC
class NCHW16CMulNCJitCode : public JitCode {
 public:
  DECLARE_JIT_CODE(NCHW16CMulNCJitCode);
  explicit NCHW16CMulNCJitCode(int d /*unused*/, size_t code_size,
                               void* code_ptr = nullptr)
      : JitCode(code_size, code_ptr) {
    this->genCode();
  }
  void genCode() override;
};

T
tensor-tang 已提交
117 118
}  // namespace gen
}  // namespace jit
T
tensor-tang 已提交
119 120
}  // namespace operators
}  // namespace paddle