blas.cc 4.6 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
/* 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. */
T
tensor-tang 已提交
14 15 16

#include "paddle/fluid/operators/jit/gen/blas.h"
#include "paddle/fluid/operators/jit/registry.h"
T
tensor-tang 已提交
17
#include "paddle/fluid/platform/cpu_info.h"
T
tensor-tang 已提交
18 19 20

namespace paddle {
namespace operators {
T
tensor-tang 已提交
21 22
namespace jit {
namespace gen {
T
tensor-tang 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 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 98 99 100 101 102 103 104 105 106

void VXXJitCode::genCode() {
  // do not need push stack, and do not need save avx512reg if do not use avx512
  int offset = 0;
  if (with_relu_) {
    vxorps(ymm_zero, ymm_zero, ymm_zero);
  }
  if (scalar_index_ == 1) {
    vbroadcastss(ymm_src1, ptr[param1]);
  } else if (scalar_index_ == 2) {
    vbroadcastss(ymm_src2, ptr[param2]);
  }
  for (int i = 0; i < num_ / YMM_FLOAT_BLOCK; ++i) {
    if (scalar_index_ != 1) {
      vmovups(ymm_src1, ptr[param1 + offset]);
    }
    if (scalar_index_ != 2) {
      vmovups(ymm_src2, ptr[param2 + offset]);
    }
    if (type_ == operand_type::mul) {
      vmulps(ymm_dst, ymm_src1, ymm_src2);
    } else if (type_ == operand_type::add) {
      vaddps(ymm_dst, ymm_src1, ymm_src2);
    }
    if (with_relu_) {
      vmaxps(ymm_dst, ymm_zero, ymm_dst);
    }
    vmovups(ptr[param3 + offset], ymm_dst);
    offset += sizeof(float) * YMM_FLOAT_BLOCK;
  }
  int rest = num_ % YMM_FLOAT_BLOCK;
  while (rest > 0) {
    int block = XMM_FLOAT_BLOCK;
    if (rest >= 4) {
      block = 4;
      if (scalar_index_ != 1) {
        vmovups(xmm_src1, ptr[param1 + offset]);
      }
      if (scalar_index_ != 2) {
        vmovups(xmm_src2, ptr[param2 + offset]);
      }
    } else if (rest >= 2) {
      block = 2;
      if (scalar_index_ != 1) {
        vmovq(xmm_src1, ptr[param1 + offset]);
      }
      if (scalar_index_ != 2) {
        vmovq(xmm_src2, ptr[param2 + offset]);
      }
    } else {
      block = 1;
      if (scalar_index_ != 1) {
        vmovss(xmm_src1, ptr[param1 + offset]);
      }
      if (scalar_index_ != 2) {
        vmovss(xmm_src2, ptr[param2 + offset]);
      }
    }
    switch (type_) {
      case operand_type::mul:
        vmulps(xmm_dst, xmm_src1, xmm_src2);
        break;
      case operand_type::add:
        vaddps(xmm_dst, xmm_src1, xmm_src2);
        break;
      default:
        break;
    }
    if (with_relu_) {
      vmaxps(xmm_dst, xmm_zero, xmm_dst);
    }
    if (rest >= 4) {
      vmovups(ptr[param3 + offset], xmm_dst);
    } else if (rest >= 2) {
      vmovq(ptr[param3 + offset], xmm_dst);
    } else {
      vmovss(ptr[param3 + offset], xmm_dst);
    }
    offset += sizeof(float) * block;
    rest -= block;
  }
  ret();
}

107 108 109 110 111 112 113 114 115 116 117 118
#define DECLARE_BLAS_CREATOR(name)                                           \
  class name##Creator : public JitCodeCreator<int> {                         \
   public:                                                                   \
    bool UseMe(const int& attr) const override {                             \
      return platform::MayIUse(platform::avx);                               \
    }                                                                        \
    size_t CodeSize(const int& d) const override {                           \
      return 96 + d / YMM_FLOAT_BLOCK * 4 * 8;                               \
    }                                                                        \
    std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override { \
      return make_unique<name##JitCode>(attr, CodeSize(attr));               \
    }                                                                        \
T
tensor-tang 已提交
119
  }
120 121 122 123 124 125 126 127 128

DECLARE_BLAS_CREATOR(VMul);
DECLARE_BLAS_CREATOR(VAdd);
DECLARE_BLAS_CREATOR(VSub);
DECLARE_BLAS_CREATOR(VAddRelu);
DECLARE_BLAS_CREATOR(VScal);
DECLARE_BLAS_CREATOR(VAddBias);

#undef DECLARE_BLAS_CREATOR
T
tensor-tang 已提交
129

T
tensor-tang 已提交
130
}  // namespace gen
T
tensor-tang 已提交
131
}  // namespace jit
T
tensor-tang 已提交
132 133
}  // namespace operators
}  // namespace paddle
T
tensor-tang 已提交
134 135 136 137

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

REGISTER_JITKERNEL_GEN(vmul, gen::VMulCreator);
138 139 140 141 142 143
REGISTER_JITKERNEL_GEN(vadd, gen::VAddCreator);
// TODO(TJ): enable sub
// REGISTER_JITKERNEL_GEN(vsub, gen::VSubCreator);
REGISTER_JITKERNEL_GEN(vaddrelu, gen::VAddReluCreator);
REGISTER_JITKERNEL_GEN(vscal, gen::VScalCreator);
REGISTER_JITKERNEL_GEN(vaddbias, gen::VAddBiasCreator);