cpp-taskflow/taskflow/algorithm/transform.hpp at master · EMinsight/cpp-taskflow · GitHub

History

190 lines (159 loc) · 5.8 KB

Raw

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

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

183

184

185

186

187

#pragma once

#include "../taskflow.hpp"

namespace tf {

// Function: make_transform_task

template <

typename B, typename E, typename O, typename C, typename P = DefaultPartitioner,

std::enable_if_t<is_partitioner_v<std::decay_t<P>>, void>* = nullptr

auto make_transform_task(B first1, E last1, O d_first, C c, P part = P()) {

using namespace std::string_literals;

using B_t = std::decay_t<unwrap_ref_decay_t<B>>;

using E_t = std::decay_t<unwrap_ref_decay_t<E>>;

using O_t = std::decay_t<unwrap_ref_decay_t<O>>;

return [=] (Runtime& rt) mutable {

// fetch the stateful values

B_t beg = first1;

E_t end = last1;

O_t d_beg = d_first;

size_t W = rt.executor().num_workers();

size_t N = std::distance(beg, end);

// only myself - no need to spawn another graph

if(W <= 1 || N <= part.chunk_size()) {

part([=]() mutable { std::transform(beg, end, d_beg, c); })();

return;

}

PreemptionGuard preemption_guard(rt);

if(N < W) {

W = N;

}

// static partitioner

if constexpr(part.type() == PartitionerType::STATIC) {

for(size_t w=0, curr_b=0; w<W && curr_b < N;) {

auto chunk_size = part.adjusted_chunk_size(N, W, w);

auto task = part([=] () mutable {

part.loop(N, W, curr_b, chunk_size, [=, prev_e=size_t{0}](size_t part_b, size_t part_e) mutable {

std::advance(beg, part_b - prev_e);

std::advance(d_beg, part_b - prev_e);

for(size_t x = part_b; x<part_e; x++) {

*d_beg++ = c(*beg++);

}

prev_e = part_e;

});

(++w == W || (curr_b += chunk_size) >= N) ? task() : rt.silent_async(task);

}

// dynamic partitioner

else {

auto next = std::make_shared<std::atomic<size_t>>(0);

for(size_t w=0; w<W;) {

auto task = part([=] () mutable {

part.loop(N, W, *next, [=, prev_e=size_t{0}](size_t part_b, size_t part_e) mutable {

std::advance(beg, part_b - prev_e);

std::advance(d_beg, part_b - prev_e);

for(size_t x = part_b; x<part_e; x++) {

*d_beg++ = c(*beg++);

}

prev_e = part_e;

});

(++w == W) ? task() : rt.silent_async(task);

}

};

}

// Function: make_transform_task

template <

typename B1, typename E1, typename B2, typename O, typename C, typename P = DefaultPartitioner,

std::enable_if_t<!is_partitioner_v<std::decay_t<C>>, void>* = nullptr

auto make_transform_task(B1 first1, E1 last1, B2 first2, O d_first, C c, P part = P()) {

using namespace std::string_literals;

using B1_t = std::decay_t<unwrap_ref_decay_t<B1>>;

using E1_t = std::decay_t<unwrap_ref_decay_t<E1>>;

using B2_t = std::decay_t<unwrap_ref_decay_t<B2>>;

using O_t = std::decay_t<unwrap_ref_decay_t<O>>;

return [=] (Runtime& rt) mutable {

// fetch the stateful values

B1_t beg1 = first1;

E1_t end1 = last1;

B2_t beg2 = first2;

O_t d_beg = d_first;

size_t W = rt.executor().num_workers();

size_t N = std::distance(beg1, end1);

// only myself - no need to spawn another graph

if(W <= 1 || N <= part.chunk_size()) {

part([=]() mutable { std::transform(beg1, end1, beg2, d_beg, c); })();

return;

}

PreemptionGuard preemption_guard(rt);

if(N < W) {

W = N;

}

// static partitioner

if constexpr(part.type() == PartitionerType::STATIC) {

for(size_t w=0, curr_b=0; w<W && curr_b < N;) {

auto chunk_size = part.adjusted_chunk_size(N, W, w);

auto task = part([=] () mutable {

part.loop(N, W, curr_b, chunk_size, [=, prev_e=size_t{0}](size_t part_b, size_t part_e) mutable {

std::advance(beg1, part_b - prev_e);

std::advance(beg2, part_b - prev_e);

std::advance(d_beg, part_b - prev_e);

for(size_t x = part_b; x<part_e; x++) {

*d_beg++ = c(*beg1++, *beg2++);

}

prev_e = part_e;

});

(++w == W || (curr_b += chunk_size) >= N) ? task() : rt.silent_async(task);

}

// dynamic partitioner

else {

auto next = std::make_shared<std::atomic<size_t>>(0);

for(size_t w=0; w<W;) {

auto task = part([=] () mutable {

part.loop(N, W, *next, [=, prev_e=size_t{0}](size_t part_b, size_t part_e) mutable {

std::advance(beg1, part_b - prev_e);

std::advance(beg2, part_b - prev_e);

std::advance(d_beg, part_b - prev_e);

for(size_t x = part_b; x<part_e; x++) {

*d_beg++ = c(*beg1++, *beg2++);

}

prev_e = part_e;

});

(++w == W) ? task() : rt.silent_async(task);

}

};

}

// ----------------------------------------------------------------------------

// transform

// ----------------------------------------------------------------------------

// Function: transform

template <typename B, typename E, typename O, typename C, typename P,

std::enable_if_t<is_partitioner_v<std::decay_t<P>>, void>*

Task FlowBuilder::transform(B first1, E last1, O d_first, C c, P part) {

return emplace(

make_transform_task(first1, last1, d_first, c, part)

);

}

// ----------------------------------------------------------------------------

// transform2

// ----------------------------------------------------------------------------

// Function: transform

template <

typename B1, typename E1, typename B2, typename O, typename C, typename P,

std::enable_if_t<!is_partitioner_v<std::decay_t<C>>, void>*

Task FlowBuilder::transform(

B1 first1, E1 last1, B2 first2, O d_first, C c, P part

) {

return emplace(make_transform_task(

first1, last1, first2, d_first, c, part

));

}

} // end of namespace tf -----------------------------------------------------

cpp-taskflow/taskflow/algorithm/transform.hpp at master · EMinsight/cpp-taskflow · GitHub

Expand file tree

Breadcrumbs

transform.hpp

Latest commit

History

Breadcrumbs

transform.hpp

File metadata and controls