mix.cc 7.5 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 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
/* 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/more/mix/mix.h"
#include "paddle/fluid/operators/jit/kernels.h"
#include "paddle/fluid/operators/jit/registry.h"
#include "paddle/fluid/platform/cpu_info.h"

namespace paddle {
namespace operators {
namespace jit {
namespace more {
namespace mix {

void VSigmoid(const T* x, T* y, int n) {
  const float min = SIGMOID_THRESHOLD_MIN;
  const float max = SIGMOID_THRESHOLD_MAX;
  for (int i = 0; i < n; ++i) {
    y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
    y[i] = static_cast<T>(0) - y[i];
  }
  auto compute = Get<KernelType::vexp, XYNTuples<T>, platform::CPUPlace>(n);
  compute(y, y, n);
  for (int i = 0; i < n; ++i) {
    y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
  }
}

void VTanh(const T* x, T* y, int n) {
  const T a = 2, b = -1;
  auto compute_scal = Get<vscal, AXYNTuples<T>, platform::CPUPlace>(n);
  auto compute_addbias = Get<vaddbias, AXYNTuples<T>, platform::CPUPlace>(n);
  auto compute_sigmoid = Get<vsigmoid, XYNTuples<T>, platform::CPUPlace>(n);
  compute_scal(&a, x, y, n);
  compute_sigmoid(y, y, n);
  compute_scal(&a, y, y, n);
  compute_addbias(&b, y, y, n);
}

51 52 53 54 55 56 57 58 59 60 61 62
void (*getActFunc(KernelType type, int d))(const T*, T*, int) {  // NOLINT
  if (type == vsigmoid) {
    return Get<vsigmoid, XYNTuples<T>, platform::CPUPlace>(d);
  } else if (type == vrelu) {
    return Get<vrelu, XYNTuples<T>, platform::CPUPlace>(d);
  } else if (type == vtanh) {
    return Get<vtanh, XYNTuples<T>, platform::CPUPlace>(d);
  } else if (type == videntity) {
    return Get<videntity, XYNTuples<T>, platform::CPUPlace>(d);
  }
  PADDLE_THROW("Not support type: %s", type);
  return nullptr;
T
tensor-tang 已提交
63 64
}

65 66 67 68 69 70 71 72 73 74 75 76 77 78 79
void LSTMCtHt(lstm_t* step, const lstm_attr_t* attr) {
  T* gates = reinterpret_cast<T*>(step->gates);
  const T* ct_1 = reinterpret_cast<const T*>(step->ct_1);
  T* ct = reinterpret_cast<T*>(step->ct);
  T* ht = reinterpret_cast<T*>(step->ht);
  const T* wp = reinterpret_cast<const T*>(step->wp);
  T* checked = reinterpret_cast<T*>(step->checked);
  const int d = attr->d;
  const int d2 = d * 2;
  const int d3 = d * 3;
  auto vmul_d = Get<vmul, XYZNTuples<T>, platform::CPUPlace>(d);
  auto vadd_d = Get<vadd, XYZNTuples<T>, platform::CPUPlace>(d);
  auto vadd_d2 = Get<vadd, XYZNTuples<T>, platform::CPUPlace>(d2);
  auto act_gate_d = getActFunc(attr->act_gate, d);
  auto act_gate_d2 = getActFunc(attr->act_gate, d2);
T
tensor-tang 已提交
80
  auto act_gate_d3 = getActFunc(attr->act_gate, d3);
81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107
  auto act_cand_d = getActFunc(attr->act_cand, d);
  auto act_cell_d = getActFunc(attr->act_cell, d);

  if (attr->use_peephole) {
    vmul_d(wp, ct_1, checked, d);
    vmul_d(wp + d, ct_1, checked + d, d);
    vadd_d2(checked, gates + d, gates + d, d2);
    act_gate_d2(gates + d, gates + d, d2);
  } else {
    act_gate_d3(gates + d, gates + d, d3);
  }

  // C_t = C_t-1 * fgated + cand_gated * igated
  act_cand_d(gates, gates, d);
  vmul_d(gates, gates + d, gates + d, d);
  vmul_d(ct_1, gates + d2, gates + d2, d);
  vadd_d(gates + d, gates + d2, ct, d);

  if (attr->use_peephole) {
    // get ogated
    vmul_d(wp + d2, ct, gates + d, d);
    vadd_d(gates + d, gates + d3, gates + d3, d);
    act_gate_d(gates + d3, gates + d3, d);
  }
  // H_t = act_cell(C_t) * ogated
  act_cell_d(ct, gates + d2, d);
  vmul_d(gates + d2, gates + d3, ht, d);
T
tensor-tang 已提交
108 109
}

110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
void LSTMC1H1(lstm_t* step, const lstm_attr_t* attr) {
  T* gates = reinterpret_cast<T*>(step->gates);
  T* ct = reinterpret_cast<T*>(step->ct);
  T* ht = reinterpret_cast<T*>(step->ht);
  int d = attr->d;
  int d2 = d * 2;
  int d3 = d * 3;
  auto vmul_d = Get<vmul, XYZNTuples<T>, platform::CPUPlace>(d);
  auto vadd_d = Get<vadd, XYZNTuples<T>, platform::CPUPlace>(d);
  auto act_gate_d = getActFunc(attr->act_gate, d);
  auto act_cand_d = getActFunc(attr->act_cand, d);
  auto act_cell_d = getActFunc(attr->act_cell, d);
  /* C_t = igated * cgated*/
  act_gate_d(gates + d, gates + d, d);
  act_cand_d(gates, gates, d);
  vmul_d(gates, gates + d, ct, d);
  if (attr->use_peephole) {
    // get outgated, put W_oc * C_t on igated
    const T* wp = reinterpret_cast<const T*>(step->wp);
    vmul_d(wp + d2, ct, gates + d, d);
    vadd_d(gates + d, gates + d3, gates + d3, d);
T
tensor-tang 已提交
131
  }
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
  /* H_t = act_cell(C_t) * ogated */
  act_gate_d(gates + d3, gates + d3, d);
  act_cell_d(ct, gates + d2, d);
  vmul_d(gates + d2, gates + d3, ht, d);
}

// compute h1 without h0
void GRUH1(gru_t* step, const gru_attr_t* attr) {
  T* gates = reinterpret_cast<T*>(step->gates);
  T* ht = reinterpret_cast<T*>(step->ht);
  int d = attr->d;
  int d2 = d * 2;
  auto act_gate = getActFunc(attr->act_gate, d);
  auto act_cand = getActFunc(attr->act_cand, d);
  auto vmul_d = Get<vmul, XYZNTuples<T>, platform::CPUPlace>(d);
  act_gate(gates, gates, d);
  act_cand(gates + d2, gates + d2, d);
  vmul_d(gates, gates + d2, ht, d);
}

// compute the first part of GRU: ht = act_gate(r) * ht_1
void GRUHtPart1(gru_t* step, const gru_attr_t* attr) {
  // W: {W_update, W_reset; W_state}
  T* gates = reinterpret_cast<T*>(step->gates);
  T* ht = reinterpret_cast<T*>(step->ht);
  const T* ht_1 = reinterpret_cast<const T*>(step->ht_1);
  auto act_gate = getActFunc(attr->act_gate, attr->d);
  auto vmul_d = Get<vmul, XYZNTuples<T>, platform::CPUPlace>(attr->d);
  act_gate(gates + attr->d, gates + attr->d, attr->d);
  vmul_d(ht_1, gates + attr->d, ht, attr->d);
}

// compute the second part of GRU:
// ht = act_gate(u) * act_cand(s) + (1-act_gate(u)) * ht_1
void GRUHtPart2(gru_t* step, const gru_attr_t* attr) {
  T* gates = reinterpret_cast<T*>(step->gates);
  T* ht = reinterpret_cast<T*>(step->ht);
  const T* ht_1 = reinterpret_cast<const T*>(step->ht_1);
  int d = attr->d;
  auto act_gate = getActFunc(attr->act_gate, d);
  auto act_cand = getActFunc(attr->act_cand, d);
  T* y = gates + d * 2;
  act_gate(gates, gates, d);
  act_cand(y, y, d);
  // out = zt*ht~ + (1-zt)*ht_1
  for (int i = 0; i < d; ++i) {
    ht[i] = gates[i] * y[i] + (static_cast<T>(1) - gates[i]) * ht_1[i];
  }
}

// TODO(TJ): tuning me
T
tensor-tang 已提交
183
bool VSigmoidKernel::UseMe(const int& d) const { return true; }
184

T
tensor-tang 已提交
185
bool VTanhKernel::UseMe(const int& d) const { return true; }
186

T
tensor-tang 已提交
187
bool LSTMCtHtKernel::UseMe(const lstm_attr_t& attr) const { return true; }
188

T
tensor-tang 已提交
189
bool LSTMC1H1Kernel::UseMe(const lstm_attr_t& attr) const { return true; }
190

T
tensor-tang 已提交
191
bool GRUH1Kernel::UseMe(const gru_attr_t& attr) const { return true; }
T
tensor-tang 已提交
192

T
tensor-tang 已提交
193
bool GRUHtPart1Kernel::UseMe(const gru_attr_t& attr) const { return true; }
T
tensor-tang 已提交
194

T
tensor-tang 已提交
195
bool GRUHtPart2Kernel::UseMe(const gru_attr_t& attr) const { return true; }
T
tensor-tang 已提交
196 197 198 199 200 201 202 203 204

}  // namespace mix
}  // namespace more
}  // namespace jit
}  // namespace operators
}  // namespace paddle

namespace mix = paddle::operators::jit::more::mix;

205 206
#define REGISTER_MORE_KERNEL(key, func) \
  REGISTER_JITKERNEL_MORE(key, mix, mix::func##Kernel)
T
tensor-tang 已提交
207 208 209

REGISTER_MORE_KERNEL(vsigmoid, VSigmoid);
REGISTER_MORE_KERNEL(vtanh, VTanh);
210 211 212 213 214
REGISTER_MORE_KERNEL(lstmctht, LSTMCtHt);
REGISTER_MORE_KERNEL(lstmc1h1, LSTMC1H1);
REGISTER_MORE_KERNEL(gruh1, GRUH1);
REGISTER_MORE_KERNEL(gruhtpart1, GRUHtPart1);
REGISTER_MORE_KERNEL(gruhtpart2, GRUHtPart2);
T
tensor-tang 已提交
215 216

#undef REGISTER_MORE_KERNEL