9945915519

Committed 15 Jul 2024 08:07PM UTC coverage: 93.077% (-1.3%) from 94.362%

Build # 9945915519

Build Type

push

github

Committed by

fknorr

Commit Message

Rename existing backend / executor -> legacy_backend / legacy_executor

Names 'backend' and 'executor' will be re-used, but we want to keep the
old APIs around in the meantime to keep changesets small.

Run Details

3188 of 3687 branches covered (86.47%)

Branch coverage included in aggregate %.

17 of 23 new or added lines in 6 files covered. (73.91%)

95 existing lines in 8 files now uncovered.

7232 of 7508 relevant lines covered (96.32%)

169246.64 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

80.34

/include/task.h

#pragma once

#include <memory>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>

#include "device_queue.h"
#include "grid.h"
#include "hint.h"
#include "host_queue.h"
#include "intrusive_graph.h"
#include "launcher.h"
#include "lifetime_extending_state.h"
#include "range_mapper.h"
#include "types.h"

namespace celerity {

class handler;

namespace detail {

        class command_launcher_storage_base {
          public:
                command_launcher_storage_base() = default;
                command_launcher_storage_base(const command_launcher_storage_base&) = delete;
                command_launcher_storage_base(command_launcher_storage_base&&) = default;
                command_launcher_storage_base& operator=(const command_launcher_storage_base&) = delete;
                command_launcher_storage_base& operator=(command_launcher_storage_base&&) = default;
                virtual ~command_launcher_storage_base() = default;

                virtual sycl::event operator()(
                    device_queue& q, const subrange<3> execution_sr, const std::vector<void*>& reduction_ptrs, const bool is_reduction_initializer) const = 0;
                virtual std::future<host_queue::execution_info> operator()(host_queue& q, const subrange<3>& execution_sr) const = 0;
        };

        template <typename Functor>
        class command_launcher_storage : public command_launcher_storage_base {
          public:
                command_launcher_storage(Functor&& fun) : m_fun(std::move(fun)) {}

                sycl::event operator()(
                    device_queue& q, const subrange<3> execution_sr, const std::vector<void*>& reduction_ptrs, const bool is_reduction_initializer) const override {
                        return invoke<sycl::event>(q, execution_sr, reduction_ptrs, is_reduction_initializer);
                }

                std::future<host_queue::execution_info> operator()(host_queue& q, const subrange<3>& execution_sr) const override {
                        return invoke<std::future<host_queue::execution_info>>(q, execution_sr);
                }

          private:
                Functor m_fun;

                template <typename Ret, typename... Args>
                Ret invoke(Args&&... args) const {
                        if constexpr(std::is_invocable_v<Functor, Args...>) {
                                return m_fun(args...);
                        } else {
                                throw std::runtime_error("Cannot launch command function with provided arguments");
                        }
                }
        };

        class buffer_access_map {
          public:
                void add_access(buffer_id bid, std::unique_ptr<range_mapper_base>&& rm) { m_accesses.emplace_back(bid, std::move(rm)); }

                std::unordered_set<buffer_id> get_accessed_buffers() const;
                std::unordered_set<cl::sycl::access::mode> get_access_modes(buffer_id bid) const;
                size_t get_num_accesses() const { return m_accesses.size(); }
                std::pair<buffer_id, access_mode> get_nth_access(const size_t n) const {
                        const auto& [bid, rm] = m_accesses[n];
                        return {bid, rm->get_access_mode()};
                }

                /**
                 * @brief Computes the combined access-region for a given buffer, mode and subrange.
                 *
                 * @param bid
                 * @param mode
                 * @param sr The subrange to be passed to the range mappers (extended to a chunk using the global size of the task)
                 *
                 * @returns The region obtained by merging the results of all range-mappers for this buffer and mode
                 */
                region<3> get_mode_requirements(
                    const buffer_id bid, const access_mode mode, const int kernel_dims, const subrange<3>& sr, const range<3>& global_size) const;

                box<3> get_requirements_for_nth_access(const size_t n, const int kernel_dims, const subrange<3>& sr, const range<3>& global_size) const;

                std::vector<const range_mapper_base*> get_range_mappers(const buffer_id bid) const {
                        std::vector<const range_mapper_base*> rms;
                        for(const auto& [a_bid, a_rm] : m_accesses) {
                                if(a_bid == bid) { rms.push_back(a_rm.get()); }
                        }
                        return rms;
                }

                box_vector<3> get_required_contiguous_boxes(const buffer_id bid, const int kernel_dims, const subrange<3>& sr, const range<3>& global_size) const;

          private:
                std::vector<std::pair<buffer_id, std::unique_ptr<range_mapper_base>>> m_accesses;
        };

        using reduction_set = std::vector<reduction_info>;

        class side_effect_map : private std::unordered_map<host_object_id, experimental::side_effect_order> {
          private:
                using map_base = std::unordered_map<host_object_id, experimental::side_effect_order>;

          public:
                using typename map_base::const_iterator, map_base::value_type, map_base::key_type, map_base::mapped_type, map_base::const_reference,
                    map_base::const_pointer;
                using iterator = const_iterator;
                using reference = const_reference;
                using pointer = const_pointer;

                using map_base::size, map_base::count, map_base::empty, map_base::cbegin, map_base::cend, map_base::at;

                iterator begin() const { return cbegin(); }
                iterator end() const { return cend(); }
                iterator find(host_object_id key) const { return map_base::find(key); }

                void add_side_effect(host_object_id hoid, experimental::side_effect_order order);
        };

        class fence_promise {
          public:
                fence_promise() = default;
                fence_promise(const fence_promise&) = delete;
                fence_promise& operator=(const fence_promise&) = delete;
                virtual ~fence_promise() = default;

                virtual void fulfill() = 0;
                virtual allocation_id get_user_allocation_id() = 0;
        };

        struct task_geometry {
                int dimensions = 0;
                range<3> global_size{1, 1, 1};
                id<3> global_offset{};
                range<3> granularity{1, 1, 1};
        };

        class task : public intrusive_graph_node<task> {
          public:
                task_type get_type() const { return m_type; }

                task_id get_id() const { return m_tid; }

                collective_group_id get_collective_group_id() const { return m_cgid; }

                const buffer_access_map& get_buffer_access_map() const { return m_access_map; }

                const side_effect_map& get_side_effect_map() const { return m_side_effects; }

                const task_geometry& get_geometry() const { return m_geometry; }

                int get_dimensions() const { return m_geometry.dimensions; }

                range<3> get_global_size() const { return m_geometry.global_size; }

                id<3> get_global_offset() const { return m_geometry.global_offset; }

                range<3> get_granularity() const { return m_geometry.granularity; }

                void set_debug_name(const std::string& debug_name) { m_debug_name = debug_name; }
                const std::string& get_debug_name() const { return m_debug_name; }

                bool has_variable_split() const { return m_type == task_type::host_compute || m_type == task_type::device_compute; }

                execution_target get_execution_target() const {
                        switch(m_type) {
                        case task_type::epoch: return execution_target::none;
                        case task_type::device_compute: return execution_target::device;
                        case task_type::host_compute:
                        case task_type::collective:
                        case task_type::master_node: return execution_target::host;
                        case task_type::horizon:
                        case task_type::fence: return execution_target::none;
                        default: assert(!"Unhandled task type"); return execution_target::none;
                        }
                }

                const reduction_set& get_reductions() const { return m_reductions; }

                epoch_action get_epoch_action() const { return m_epoch_action; }

                fence_promise* get_fence_promise() const { return m_fence_promise.get(); }

                template <typename Launcher>
                Launcher get_launcher() const {
                        return {};
                } // placeholder

                template <typename... Args>
                auto launch(Args&&... args) const {
                        return (*m_launcher)(std::forward<Args>(args)...);
                }

                void extend_lifetime(std::shared_ptr<lifetime_extending_state> state) { m_attached_state.emplace_back(std::move(state)); }

                void add_hint(std::unique_ptr<hint_base>&& h) { m_hints.emplace_back(std::move(h)); }

                template <typename Hint>
                const Hint* get_hint() const {
                        static_assert(std::is_base_of_v<hint_base, Hint>, "Hint must extend hint_base");
                        for(auto& h : m_hints) {
                                if(auto* ptr = dynamic_cast<Hint*>(h.get()); ptr != nullptr) { return ptr; }
                        }
                        return nullptr;
                }

                static std::unique_ptr<task> make_epoch(task_id tid, detail::epoch_action action) {
                        return std::unique_ptr<task>(new task(tid, task_type::epoch, non_collective_group_id, task_geometry{}, nullptr, {}, {}, {}, action, nullptr));
                }

                static std::unique_ptr<task> make_host_compute(task_id tid, task_geometry geometry, std::unique_ptr<command_launcher_storage_base> launcher,
                    buffer_access_map access_map, side_effect_map side_effect_map, reduction_set reductions) {
                        return std::unique_ptr<task>(new task(tid, task_type::host_compute, non_collective_group_id, geometry, std::move(launcher), std::move(access_map),
                            std::move(side_effect_map), std::move(reductions), {}, nullptr));
                }

                static std::unique_ptr<task> make_device_compute(task_id tid, task_geometry geometry, std::unique_ptr<command_launcher_storage_base> launcher,
                    buffer_access_map access_map, reduction_set reductions) {
                        return std::unique_ptr<task>(new task(tid, task_type::device_compute, non_collective_group_id, geometry, std::move(launcher), std::move(access_map),
                            {}, std::move(reductions), {}, nullptr));
                }

                static std::unique_ptr<task> make_collective(task_id tid, collective_group_id cgid, size_t num_collective_nodes,
                    std::unique_ptr<command_launcher_storage_base> launcher, buffer_access_map access_map, side_effect_map side_effect_map) {
                        const task_geometry geometry{1, detail::range_cast<3>(range(num_collective_nodes)), {}, {1, 1, 1}};
                        return std::unique_ptr<task>(
                            new task(tid, task_type::collective, cgid, geometry, std::move(launcher), std::move(access_map), std::move(side_effect_map), {}, {}, nullptr));
                }

                static std::unique_ptr<task> make_master_node(
                    task_id tid, std::unique_ptr<command_launcher_storage_base> launcher, buffer_access_map access_map, side_effect_map side_effect_map) {
                        return std::unique_ptr<task>(new task(tid, task_type::master_node, non_collective_group_id, task_geometry{}, std::move(launcher),
                            std::move(access_map), std::move(side_effect_map), {}, {}, nullptr));
                }

                static std::unique_ptr<task> make_horizon(task_id tid) {
                        return std::unique_ptr<task>(new task(tid, task_type::horizon, non_collective_group_id, task_geometry{}, nullptr, {}, {}, {}, {}, nullptr));
                }

                static std::unique_ptr<task> make_fence(
                    task_id tid, buffer_access_map access_map, side_effect_map side_effect_map, std::unique_ptr<fence_promise> fence_promise) {
                        return std::unique_ptr<task>(new task(tid, task_type::fence, non_collective_group_id, task_geometry{}, nullptr, std::move(access_map),
                            std::move(side_effect_map), {}, {}, std::move(fence_promise)));
                }

          private:
                task_id m_tid;
                task_type m_type;
                collective_group_id m_cgid;
                task_geometry m_geometry;
                std::unique_ptr<command_launcher_storage_base> m_launcher;
                buffer_access_map m_access_map;
                detail::side_effect_map m_side_effects;
                reduction_set m_reductions;
                std::string m_debug_name;
                detail::epoch_action m_epoch_action;
                // TODO I believe that `struct task` should not store command_group_launchers, fence_promise or other state that is related to execution instead of
                // abstract DAG building. For user-initialized buffers we already notify the runtime -> executor of this state directly. Maybe also do that for these.
                std::unique_ptr<fence_promise> m_fence_promise;
                std::vector<std::shared_ptr<lifetime_extending_state>> m_attached_state;
                std::vector<std::unique_ptr<hint_base>> m_hints;

                task(task_id tid, task_type type, collective_group_id cgid, task_geometry geometry, std::unique_ptr<command_launcher_storage_base> launcher,
                    buffer_access_map access_map, detail::side_effect_map side_effects, reduction_set reductions, detail::epoch_action epoch_action,
                    std::unique_ptr<fence_promise> fence_promise)
                    : m_tid(tid), m_type(type), m_cgid(cgid), m_geometry(geometry), m_launcher(std::move(launcher)), m_access_map(std::move(access_map)),
                      m_side_effects(std::move(side_effects)), m_reductions(std::move(reductions)), m_epoch_action(epoch_action),
                      m_fence_promise(std::move(fence_promise)) {
                        assert(type == task_type::host_compute || type == task_type::device_compute || get_granularity().size() == 1);
                        // Only host tasks can have side effects
                        assert(this->m_side_effects.empty() || type == task_type::host_compute || type == task_type::collective || type == task_type::master_node
                               || type == task_type::fence);
                }
        };

        [[nodiscard]] std::string print_task_debug_label(const task& tsk, bool title_case = false);

        /// Determines which overlapping regions appear between write accesses when the iteration space of `tsk` is split into `chunks`.
        std::unordered_map<buffer_id, region<3>> detect_overlapping_writes(const task& tsk, const box_vector<3>& chunks);

} // namespace detail
} // namespace celerity

1	#pragma once
2
3	#include <memory>
4	#include <unordered_map>
5	#include <unordered_set>
6	#include <utility>
7	#include <vector>
8
9	#include "device_queue.h"
10	#include "grid.h"
11	#include "hint.h"
12	#include "host_queue.h"
13	#include "intrusive_graph.h"
14	#include "launcher.h"
15	#include "lifetime_extending_state.h"
16	#include "range_mapper.h"
17	#include "types.h"
18
19	namespace celerity {
20
21	class handler;
22
23	namespace detail {
24
25	class command_launcher_storage_base {
26	public:
27	command_launcher_storage_base() = default;	4,409✔
28	command_launcher_storage_base(const command_launcher_storage_base&) = delete;
29	command_launcher_storage_base(command_launcher_storage_base&&) = default;
30	command_launcher_storage_base& operator=(const command_launcher_storage_base&) = delete;
31	command_launcher_storage_base& operator=(command_launcher_storage_base&&) = default;
32	virtual ~command_launcher_storage_base() = default;	4,409✔
33
34	virtual sycl::event operator()(
35	device_queue& q, const subrange<3> execution_sr, const std::vector<void*>& reduction_ptrs, const bool is_reduction_initializer) const = 0;
36	virtual std::future<host_queue::execution_info> operator()(host_queue& q, const subrange<3>& execution_sr) const = 0;
37	};
38
39	template <typename Functor>
40	class command_launcher_storage : public command_launcher_storage_base {
41	public:
42	command_launcher_storage(Functor&& fun) : m_fun(std::move(fun)) {}	4,409✔
43
44	sycl::event operator()(	169✔
45	device_queue& q, const subrange<3> execution_sr, const std::vector<void*>& reduction_ptrs, const bool is_reduction_initializer) const override {
46	return invoke<sycl::event>(q, execution_sr, reduction_ptrs, is_reduction_initializer);	169✔
47	}
48
49	std::future<host_queue::execution_info> operator()(host_queue& q, const subrange<3>& execution_sr) const override {	3,157✔
50	return invoke<std::future<host_queue::execution_info>>(q, execution_sr);	3,157✔
51	}
52
53	private:
54	Functor m_fun;
55
56	template <typename Ret, typename... Args>
57	Ret invoke(Args&&... args) const {	3,326✔
58	if constexpr(std::is_invocable_v<Functor, Args...>) {
59	return m_fun(args...);	3,326✔
60	} else {
UNCOV 61	throw std::runtime_error("Cannot launch command function with provided arguments");	×
62	}
63	}
64	};
65
66	class buffer_access_map {
67	public:
68	void add_access(buffer_id bid, std::unique_ptr<range_mapper_base>&& rm) { m_accesses.emplace_back(bid, std::move(rm)); }	3,572✔
69
70	std::unordered_set<buffer_id> get_accessed_buffers() const;
71	std::unordered_set<cl::sycl::access::mode> get_access_modes(buffer_id bid) const;
72	size_t get_num_accesses() const { return m_accesses.size(); }	22,473✔
73	std::pair<buffer_id, access_mode> get_nth_access(const size_t n) const {	2,667✔
74	const auto& [bid, rm] = m_accesses[n];	2,667✔
75	return {bid, rm->get_access_mode()};	2,667✔
76	}
77
78	/**
79	* @brief Computes the combined access-region for a given buffer, mode and subrange.
80	*
81	* @param bid
82	* @param mode
83	* @param sr The subrange to be passed to the range mappers (extended to a chunk using the global size of the task)
84	*
85	* @returns The region obtained by merging the results of all range-mappers for this buffer and mode
86	*/
87	region<3> get_mode_requirements(
88	const buffer_id bid, const access_mode mode, const int kernel_dims, const subrange<3>& sr, const range<3>& global_size) const;
89
90	box<3> get_requirements_for_nth_access(const size_t n, const int kernel_dims, const subrange<3>& sr, const range<3>& global_size) const;
91
92	std::vector<const range_mapper_base*> get_range_mappers(const buffer_id bid) const {
93	std::vector<const range_mapper_base*> rms;
94	for(const auto& [a_bid, a_rm] : m_accesses) {
95	if(a_bid == bid) { rms.push_back(a_rm.get()); }
96	}
97	return rms;
98	}
99
100	box_vector<3> get_required_contiguous_boxes(const buffer_id bid, const int kernel_dims, const subrange<3>& sr, const range<3>& global_size) const;
101
102	private:
103	std::vector<std::pair<buffer_id, std::unique_ptr<range_mapper_base>>> m_accesses;
104	};
105
106	using reduction_set = std::vector<reduction_info>;
107
108	class side_effect_map : private std::unordered_map<host_object_id, experimental::side_effect_order> {
109	private:
110	using map_base = std::unordered_map<host_object_id, experimental::side_effect_order>;
111
112	public:
113	using typename map_base::const_iterator, map_base::value_type, map_base::key_type, map_base::mapped_type, map_base::const_reference,
114	map_base::const_pointer;
115	using iterator = const_iterator;
116	using reference = const_reference;
117	using pointer = const_pointer;
118
119	using map_base::size, map_base::count, map_base::empty, map_base::cbegin, map_base::cend, map_base::at;
120
121	iterator begin() const { return cbegin(); }	5,789✔
122	iterator end() const { return cend(); }	5,787✔
123	iterator find(host_object_id key) const { return map_base::find(key); }
124
125	void add_side_effect(host_object_id hoid, experimental::side_effect_order order);
126	};
127
128	class fence_promise {
129	public:
130	fence_promise() = default;	42✔
131	fence_promise(const fence_promise&) = delete;
132	fence_promise& operator=(const fence_promise&) = delete;
133	virtual ~fence_promise() = default;	42✔
134
135	virtual void fulfill() = 0;
136	virtual allocation_id get_user_allocation_id() = 0;
137	};
138
139	struct task_geometry {
140	int dimensions = 0;
141	range<3> global_size{1, 1, 1};
142	id<3> global_offset{};
143	range<3> granularity{1, 1, 1};
144	};
145
146	class task : public intrusive_graph_node<task> {
147	public:
148	task_type get_type() const { return m_type; }	36,208✔
149
150	task_id get_id() const { return m_tid; }	47,765✔
151
152	collective_group_id get_collective_group_id() const { return m_cgid; }	13,275✔
153
154	const buffer_access_map& get_buffer_access_map() const { return m_access_map; }	42,906✔
155
156	const side_effect_map& get_side_effect_map() const { return m_side_effects; }	14,986✔
157
158	const task_geometry& get_geometry() const { return m_geometry; }	3,941✔
159
160	int get_dimensions() const { return m_geometry.dimensions; }	58,862✔
161
162	range<3> get_global_size() const { return m_geometry.global_size; }	99,235✔
163
164	id<3> get_global_offset() const { return m_geometry.global_offset; }	28,935✔
165
166	range<3> get_granularity() const { return m_geometry.granularity; }	6,952✔
167
168	void set_debug_name(const std::string& debug_name) { m_debug_name = debug_name; }	4,409✔
169	const std::string& get_debug_name() const { return m_debug_name; }	4,923✔
170
171	bool has_variable_split() const { return m_type == task_type::host_compute \|\| m_type == task_type::device_compute; }	5,143✔
172
173	execution_target get_execution_target() const {	9,075✔
174	switch(m_type) {	9,075!
UNCOV 175	case task_type::epoch: return execution_target::none;	×
176	case task_type::device_compute: return execution_target::device;	2,087✔
177	case task_type::host_compute:	6,988✔
178	case task_type::collective:
179	case task_type::master_node: return execution_target::host;	6,988✔
UNCOV 180	case task_type::horizon:	×
UNCOV 181	case task_type::fence: return execution_target::none;	×
UNCOV 182	default: assert(!"Unhandled task type"); return execution_target::none;	×
183	}
184	}
185
186	const reduction_set& get_reductions() const { return m_reductions; }	51,506✔
187
188	epoch_action get_epoch_action() const { return m_epoch_action; }	417✔
189
190	fence_promise* get_fence_promise() const { return m_fence_promise.get(); }	69✔
191
192	template <typename Launcher>
193	Launcher get_launcher() const {	466✔
194	return {};	466✔
195	} // placeholder
196
197	template <typename... Args>
198	auto launch(Args&&... args) const {	3,704✔
199	return (*m_launcher)(std::forward<Args>(args)...);	3,704✔
200	}
201
202	void extend_lifetime(std::shared_ptr<lifetime_extending_state> state) { m_attached_state.emplace_back(std::move(state)); }	2,396✔
203
204	void add_hint(std::unique_ptr<hint_base>&& h) { m_hints.emplace_back(std::move(h)); }	46✔
205
206	template <typename Hint>
207	const Hint* get_hint() const {	3,397✔
208	static_assert(std::is_base_of_v<hint_base, Hint>, "Hint must extend hint_base");
209	for(auto& h : m_hints) {	3,484✔
210	if(auto* ptr = dynamic_cast<Hint*>(h.get()); ptr != nullptr) { return ptr; }	139!
211	}
212	return nullptr;	3,345✔
213	}
214
215	static std::unique_ptr<task> make_epoch(task_id tid, detail::epoch_action action) {	916✔
216	return std::unique_ptr<task>(new task(tid, task_type::epoch, non_collective_group_id, task_geometry{}, nullptr, {}, {}, {}, action, nullptr));	2,748!
217	}
218
219	static std::unique_ptr<task> make_host_compute(task_id tid, task_geometry geometry, std::unique_ptr<command_launcher_storage_base> launcher,	73✔
220	buffer_access_map access_map, side_effect_map side_effect_map, reduction_set reductions) {
221	return std::unique_ptr<task>(new task(tid, task_type::host_compute, non_collective_group_id, geometry, std::move(launcher), std::move(access_map),	146✔
222	std::move(side_effect_map), std::move(reductions), {}, nullptr));	219!
223	}
224
225	static std::unique_ptr<task> make_device_compute(task_id tid, task_geometry geometry, std::unique_ptr<command_launcher_storage_base> launcher,	957✔
226	buffer_access_map access_map, reduction_set reductions) {
227	return std::unique_ptr<task>(new task(tid, task_type::device_compute, non_collective_group_id, geometry, std::move(launcher), std::move(access_map),	1,914✔
228	{}, std::move(reductions), {}, nullptr));	2,871!
229	}
230
231	static std::unique_ptr<task> make_collective(task_id tid, collective_group_id cgid, size_t num_collective_nodes,	59✔
232	std::unique_ptr<command_launcher_storage_base> launcher, buffer_access_map access_map, side_effect_map side_effect_map) {
233	const task_geometry geometry{1, detail::range_cast<3>(range(num_collective_nodes)), {}, {1, 1, 1}};	59✔
234	return std::unique_ptr<task>(
235	new task(tid, task_type::collective, cgid, geometry, std::move(launcher), std::move(access_map), std::move(side_effect_map), {}, {}, nullptr));	118!
236	}
237
238	static std::unique_ptr<task> make_master_node(	3,320✔
239	task_id tid, std::unique_ptr<command_launcher_storage_base> launcher, buffer_access_map access_map, side_effect_map side_effect_map) {
240	return std::unique_ptr<task>(new task(tid, task_type::master_node, non_collective_group_id, task_geometry{}, std::move(launcher),	3,320✔
241	std::move(access_map), std::move(side_effect_map), {}, {}, nullptr));	13,280!
242	}
243
244	static std::unique_ptr<task> make_horizon(task_id tid) {	1,070✔
245	return std::unique_ptr<task>(new task(tid, task_type::horizon, non_collective_group_id, task_geometry{}, nullptr, {}, {}, {}, {}, nullptr));	3,210!
246	}
247
248	static std::unique_ptr<task> make_fence(	60✔
249	task_id tid, buffer_access_map access_map, side_effect_map side_effect_map, std::unique_ptr<fence_promise> fence_promise) {
250	return std::unique_ptr<task>(new task(tid, task_type::fence, non_collective_group_id, task_geometry{}, nullptr, std::move(access_map),	60✔
251	std::move(side_effect_map), {}, {}, std::move(fence_promise)));	240!
252	}
253
254	private:
255	task_id m_tid;
256	task_type m_type;
257	collective_group_id m_cgid;
258	task_geometry m_geometry;
259	std::unique_ptr<command_launcher_storage_base> m_launcher;
260	buffer_access_map m_access_map;
261	detail::side_effect_map m_side_effects;
262	reduction_set m_reductions;
263	std::string m_debug_name;
264	detail::epoch_action m_epoch_action;
265	// TODO I believe that `struct task` should not store command_group_launchers, fence_promise or other state that is related to execution instead of
266	// abstract DAG building. For user-initialized buffers we already notify the runtime -> executor of this state directly. Maybe also do that for these.
267	std::unique_ptr<fence_promise> m_fence_promise;
268	std::vector<std::shared_ptr<lifetime_extending_state>> m_attached_state;
269	std::vector<std::unique_ptr<hint_base>> m_hints;
270
271	task(task_id tid, task_type type, collective_group_id cgid, task_geometry geometry, std::unique_ptr<command_launcher_storage_base> launcher,	6,287✔
272	buffer_access_map access_map, detail::side_effect_map side_effects, reduction_set reductions, detail::epoch_action epoch_action,
273	std::unique_ptr<fence_promise> fence_promise)
274	: m_tid(tid), m_type(type), m_cgid(cgid), m_geometry(geometry), m_launcher(std::move(launcher)), m_access_map(std::move(access_map)),	6,287✔
275	m_side_effects(std::move(side_effects)), m_reductions(std::move(reductions)), m_epoch_action(epoch_action),	6,287✔
276	m_fence_promise(std::move(fence_promise)) {	12,574✔
277	assert(type == task_type::host_compute \|\| type == task_type::device_compute \|\| get_granularity().size() == 1);	6,287✔
278	// Only host tasks can have side effects
279	assert(this->m_side_effects.empty() \|\| type == task_type::host_compute \|\| type == task_type::collective \|\| type == task_type::master_node	6,287✔
280	\|\| type == task_type::fence);
281	}	6,287✔
282	};
283
284	[[nodiscard]] std::string print_task_debug_label(const task& tsk, bool title_case = false);
285
286	/// Determines which overlapping regions appear between write accesses when the iteration space of `tsk` is split into `chunks`.
287	std::unordered_map<buffer_id, region<3>> detect_overlapping_writes(const task& tsk, const box_vector<3>& chunks);
288
289	} // namespace detail
290	} // namespace celerity

celerity / celerity-runtime / 9945915519

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous