12047884020

Committed 27 Nov 2024 09:58AM UTC coverage: 94.956% (-0.02%) from 94.972%

Build # 12047884020

Build Type

push

github

Committed by

fknorr

Commit Message

Do not disable CGF disagnostics in test_utils::add_*_task

This eliminates dead code from an earlier incomplete refactoring.

The CGF teardown / reinit was only required by a single test, which was
coincidentally also broken and didn't test the feature advertised. This
commit splits up the test between runtime_ and runtime_deprecation_tests
and also moves runtime-independent sibling tests to task_graph_tests.

Run Details

3202 of 3633 branches covered (88.14%)

Branch coverage included in aggregate %.

7151 of 7270 relevant lines covered (98.36%)

1224689.38 hits per line

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

81.08

/include/task.h

#pragma once

#include "cgf.h"
#include "graph.h"
#include "grid.h"
#include "hint.h"
#include "intrusive_graph.h"
#include "ranges.h"
#include "reduction.h"
#include "sycl_wrappers.h"
#include "types.h"
#include "utils.h"

#include <cassert>
#include <cstddef>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <utility>
#include <vector>


namespace celerity {
namespace detail {

        class buffer_access_map {
          public:
                /// Default ctor for tasks w/o buffer accesses
                buffer_access_map() = default;

                buffer_access_map(std::vector<buffer_access>&& accesses, const task_geometry& geometry);

                const std::unordered_set<buffer_id>& get_accessed_buffers() const& { return m_accessed_buffers; }

                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, mode, _] = m_accesses[n];
                        return {bid, mode};
                }

                region<3> get_requirements_for_nth_access(const size_t n, const box<3>& execution_range) const;

                /// Returns the union of all consumer accesses made across the entire task (conceptually, the
                /// union of the set of regions obtained by calling get_consumed_region for each chunk).
                region<3> get_task_consumed_region(const buffer_id bid) const {
                        if(auto it = m_task_consumed_regions.find(bid); it != m_task_consumed_regions.end()) { return it->second; }
                        return {};
                }

                /// Returns the union of all producer accesses made across the entire task (conceptually, the
                /// union of the set of regions obtained by calling get_produced_region for each chunk).
                region<3> get_task_produced_region(const buffer_id bid) const {
                        if(auto it = m_task_produced_regions.find(bid); it != m_task_produced_regions.end()) { return it->second; }
                        return {};
                };

                /// Computes the union of all consumed regions (across multiple accesses) for a given execution range.
                region<3> compute_consumed_region(const buffer_id bid, const box<3>& execution_range) const;

                /// Computes the union of all produced regions (across multiple accesses) for a given execution range.
                region<3> compute_produced_region(const buffer_id bid, const box<3>& execution_range) const;

                /// Returns a set of bounding boxes, one for each accessed region, that must be allocated contiguously.
                box_vector<3> compute_required_contiguous_boxes(const buffer_id bid, const box<3>& execution_range) const;

          private:
                std::vector<buffer_access> m_accesses;
                std::unordered_set<buffer_id> m_accessed_buffers; ///< Cached set of buffer ids found in m_accesses
                task_geometry m_task_geometry;
                std::unordered_map<buffer_id, region<3>> m_task_consumed_regions;
                std::unordered_map<buffer_id, region<3>> m_task_produced_regions;
        };

        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;

                side_effect_map() = default;

                side_effect_map(const std::vector<host_object_effect>& side_effects) {
                        map_base::reserve(side_effects.size());
                        for(const auto& [hoid, order] : side_effects) {
                                map_base::emplace(hoid, order);
                        }
                }

                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(); }
        };

        // TODO refactor into an inheritance hierarchy
        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: utils::unreachable(); // LCOV_EXCL_LINE
                        }
                }

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

                epoch_action get_epoch_action() const { return m_epoch_action; }

                task_promise* get_task_promise() const { return m_promise.get(); }

                template <typename Launcher>
                Launcher get_launcher() const {
                        return std::get<Launcher>(m_launcher);
                }

                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, std::unique_ptr<task_promise> promise) {
                        return std::unique_ptr<task>(new task(tid, task_type::epoch, non_collective_group_id, task_geometry{}, {}, {}, {}, {}, action, std::move(promise)));
                }

                static std::unique_ptr<task> make_host_compute(task_id tid, task_geometry geometry, host_task_launcher 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, device_kernel_launcher 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, task_geometry geometry, collective_group_id cgid, size_t num_collective_nodes,
                    host_task_launcher launcher, buffer_access_map access_map, side_effect_map side_effect_map) {
                        // The geometry is required to construct the buffer_access_map, so we pass it in here even though it has to have a specific shape
                        assert(geometry.dimensions == 1 && geometry.global_size == detail::range_cast<3>(range(num_collective_nodes)) && geometry.global_offset == zeros);
                        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, host_task_launcher 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));
                }

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

          private:
                task_id m_tid;
                task_type m_type;
                collective_group_id m_cgid;
                task_geometry m_geometry;
                command_group_launcher 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;
                std::unique_ptr<task_promise> m_promise; // TODO keep user_allocation_id in struct task instead of inside task_promise
                std::vector<std::unique_ptr<hint_base>> m_hints;

                task(task_id tid, task_type type, collective_group_id cgid, task_geometry geometry, command_group_launcher launcher, buffer_access_map access_map,
                    detail::side_effect_map side_effects, reduction_set reductions, detail::epoch_action epoch_action, std::unique_ptr<task_promise> 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_promise(std::move(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);
                }
        };

        std::unique_ptr<detail::task> make_command_group_task(const detail::task_id tid, const size_t num_collective_nodes, raw_command_group&& cg);

        [[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);

        /// The task graph (TDAG) represents all cluster-wide operations, such as command group submissions and fences, and their interdependencies.
        class task_graph : public graph<task> {}; // inheritance instead of type alias so we can forward declare task_graph

} // namespace detail
} // namespace celerity

1	#pragma once
2
3	#include "cgf.h"
4	#include "graph.h"
5	#include "grid.h"
6	#include "hint.h"
7	#include "intrusive_graph.h"
8	#include "ranges.h"
9	#include "reduction.h"
10	#include "sycl_wrappers.h"
11	#include "types.h"
12	#include "utils.h"
13
14	#include <cassert>
15	#include <cstddef>
16	#include <memory>
17	#include <string>
18	#include <unordered_map>
19	#include <unordered_set>
20	#include <utility>
21	#include <vector>
22
23
24	namespace celerity {
25	namespace detail {
26
27	class buffer_access_map {
28	public:
29	/// Default ctor for tasks w/o buffer accesses
30	buffer_access_map() = default;	2,262✔
31
32	buffer_access_map(std::vector<buffer_access>&& accesses, const task_geometry& geometry);
33
34	const std::unordered_set<buffer_id>& get_accessed_buffers() const& { return m_accessed_buffers; }	29,167✔
35
36	size_t get_num_accesses() const { return m_accesses.size(); }	21,034✔
37
38	std::pair<buffer_id, access_mode> get_nth_access(const size_t n) const {	7,152✔
39	const auto& [bid, mode, _] = m_accesses[n];	7,152✔
40	return {bid, mode};	7,152✔
41	}
42
43	region<3> get_requirements_for_nth_access(const size_t n, const box<3>& execution_range) const;
44
45	/// Returns the union of all consumer accesses made across the entire task (conceptually, the
46	/// union of the set of regions obtained by calling get_consumed_region for each chunk).
47	region<3> get_task_consumed_region(const buffer_id bid) const {	8,004✔
48	if(auto it = m_task_consumed_regions.find(bid); it != m_task_consumed_regions.end()) { return it->second; }	8,004✔
49	return {};	753✔
50	}
51
52	/// Returns the union of all producer accesses made across the entire task (conceptually, the
53	/// union of the set of regions obtained by calling get_produced_region for each chunk).
54	region<3> get_task_produced_region(const buffer_id bid) const {	6,951✔
55	if(auto it = m_task_produced_regions.find(bid); it != m_task_produced_regions.end()) { return it->second; }	6,951✔
56	return {};	133✔
57	};
58
59	/// Computes the union of all consumed regions (across multiple accesses) for a given execution range.
60	region<3> compute_consumed_region(const buffer_id bid, const box<3>& execution_range) const;
61
62	/// Computes the union of all produced regions (across multiple accesses) for a given execution range.
63	region<3> compute_produced_region(const buffer_id bid, const box<3>& execution_range) const;
64
65	/// Returns a set of bounding boxes, one for each accessed region, that must be allocated contiguously.
66	box_vector<3> compute_required_contiguous_boxes(const buffer_id bid, const box<3>& execution_range) const;
67
68	private:
69	std::vector<buffer_access> m_accesses;
70	std::unordered_set<buffer_id> m_accessed_buffers; ///< Cached set of buffer ids found in m_accesses
71	task_geometry m_task_geometry;
72	std::unordered_map<buffer_id, region<3>> m_task_consumed_regions;
73	std::unordered_map<buffer_id, region<3>> m_task_produced_regions;
74	};
75
76	using reduction_set = std::vector<reduction_info>;
77
78	class side_effect_map : private std::unordered_map<host_object_id, experimental::side_effect_order> {
79	private:
80	using map_base = std::unordered_map<host_object_id, experimental::side_effect_order>;
81
82	public:
83	using typename map_base::const_iterator, map_base::value_type, map_base::key_type, map_base::mapped_type, map_base::const_reference,
84	map_base::const_pointer;
85	using iterator = const_iterator;
86	using reference = const_reference;
87	using pointer = const_pointer;
88
89	side_effect_map() = default;	3,924✔
90
91	side_effect_map(const std::vector<host_object_effect>& side_effects) {	1,840✔
92	map_base::reserve(side_effects.size());	1,840✔
93	for(const auto& [hoid, order] : side_effects) {	2,065✔
94	map_base::emplace(hoid, order);	225✔
95	}
96	}	1,840✔
97
98	using map_base::size, map_base::count, map_base::empty, map_base::cbegin, map_base::cend, map_base::at;
99
100	iterator begin() const { return cbegin(); }	8,380✔
101	iterator end() const { return cend(); }	8,361✔
102	};
103
104	// TODO refactor into an inheritance hierarchy
105	class task : public intrusive_graph_node<task> {
106	public:
107	task_type get_type() const { return m_type; }	39,662✔
108
109	task_id get_id() const { return m_tid; }	41,800✔
110
111	collective_group_id get_collective_group_id() const { return m_cgid; }	17,506✔
112
113	const buffer_access_map& get_buffer_access_map() const { return m_access_map; }	46,898✔
114
115	const side_effect_map& get_side_effect_map() const { return m_side_effects; }	16,058✔
116
117	const task_geometry& get_geometry() const { return m_geometry; }	2,213✔
118
119	int get_dimensions() const { return m_geometry.dimensions; }	1✔
120
121	range<3> get_global_size() const { return m_geometry.global_size; }	23,986✔
122
123	id<3> get_global_offset() const { return m_geometry.global_offset; }	11,706✔
124
125	range<3> get_granularity() const { return m_geometry.granularity; }	12,713✔
126
127	void set_debug_name(const std::string& debug_name) { m_debug_name = debug_name; }	1,284✔
128	const std::string& get_debug_name() const { return m_debug_name; }	9,071✔
129
130	bool has_variable_split() const { return m_type == task_type::host_compute \|\| m_type == task_type::device_compute; }	7,444✔
131
132	execution_target get_execution_target() const {	16,649✔
133	switch(m_type) {	16,649!
134	case task_type::epoch: return execution_target::none;	×
135	case task_type::device_compute: return execution_target::device;	8,959✔
136	case task_type::host_compute:	7,690✔
137	case task_type::collective:
138	case task_type::master_node: return execution_target::host;	7,690✔
139	case task_type::horizon:	×
140	case task_type::fence: return execution_target::none;	×
141	default: utils::unreachable(); // LCOV_EXCL_LINE
142	}
143	}
144
145	const reduction_set& get_reductions() const { return m_reductions; }	78,666✔
146
147	epoch_action get_epoch_action() const { return m_epoch_action; }	4,178✔
148
149	task_promise* get_task_promise() const { return m_promise.get(); }	1,334✔
150
151	template <typename Launcher>
152	Launcher get_launcher() const {	3,276✔
153	return std::get<Launcher>(m_launcher);	3,276✔
154	}
155
156	void add_hint(std::unique_ptr<hint_base>&& h) { m_hints.emplace_back(std::move(h)); }	142✔
157
158	template <typename Hint>
159	const Hint* get_hint() const {	12,145✔
160	static_assert(std::is_base_of_v<hint_base, Hint>, "Hint must extend hint_base");
161	for(auto& h : m_hints) {	12,532✔
162	if(auto* ptr = dynamic_cast<Hint*>(h.get()); ptr != nullptr) { return ptr; }	667!
163	}
164	return nullptr;	11,865✔
165	}
166
167	static std::unique_ptr<task> make_epoch(task_id tid, detail::epoch_action action, std::unique_ptr<task_promise> promise) {	1,389✔
168	return std::unique_ptr<task>(new task(tid, task_type::epoch, non_collective_group_id, task_geometry{}, {}, {}, {}, {}, action, std::move(promise)));	1,389!
169	}
170
171	static std::unique_ptr<task> make_host_compute(task_id tid, task_geometry geometry, host_task_launcher launcher, buffer_access_map access_map,	362✔
172	side_effect_map side_effect_map, reduction_set reductions) {
173	return std::unique_ptr<task>(new task(tid, task_type::host_compute, non_collective_group_id, geometry, std::move(launcher), std::move(access_map),	724✔
174	std::move(side_effect_map), std::move(reductions), {}, nullptr));	1,086!
175	}
176
177	static std::unique_ptr<task> make_device_compute(	1,615✔
178	task_id tid, task_geometry geometry, device_kernel_launcher launcher, buffer_access_map access_map, reduction_set reductions) {
179	return std::unique_ptr<task>(new task(tid, task_type::device_compute, non_collective_group_id, geometry, std::move(launcher), std::move(access_map),	3,230✔
180	{}, std::move(reductions), {}, nullptr));	4,845!
181	}
182
183	static std::unique_ptr<task> make_collective(task_id tid, task_geometry geometry, collective_group_id cgid, size_t num_collective_nodes,	68✔
184	host_task_launcher launcher, buffer_access_map access_map, side_effect_map side_effect_map) {
185	// The geometry is required to construct the buffer_access_map, so we pass it in here even though it has to have a specific shape
186	assert(geometry.dimensions == 1 && geometry.global_size == detail::range_cast<3>(range(num_collective_nodes)) && geometry.global_offset == zeros);	68✔
187	return std::unique_ptr<task>(
188	new task(tid, task_type::collective, cgid, geometry, std::move(launcher), std::move(access_map), std::move(side_effect_map), {}, {}, nullptr));	68!
189	}
190
191	static std::unique_ptr<task> make_master_node(task_id tid, host_task_launcher launcher, buffer_access_map access_map, side_effect_map side_effect_map) {	1,390✔
192	return std::unique_ptr<task>(new task(tid, task_type::master_node, non_collective_group_id, task_geometry{}, std::move(launcher),	1,390✔
193	std::move(access_map), std::move(side_effect_map), {}, {}, nullptr));	2,780!
194	}
195
196	static std::unique_ptr<task> make_horizon(task_id tid) {	850✔
197	return std::unique_ptr<task>(new task(tid, task_type::horizon, non_collective_group_id, task_geometry{}, {}, {}, {}, {}, {}, nullptr));	850!
198	}
199
200	static std::unique_ptr<task> make_fence(	90✔
201	task_id tid, buffer_access_map access_map, side_effect_map side_effect_map, std::unique_ptr<task_promise> promise) {
202	return std::unique_ptr<task>(new task(tid, task_type::fence, non_collective_group_id, task_geometry{}, {}, std::move(access_map),	90✔
203	std::move(side_effect_map), {}, {}, std::move(promise)));	180!
204	}
205
206	private:
207	task_id m_tid;
208	task_type m_type;
209	collective_group_id m_cgid;
210	task_geometry m_geometry;
211	command_group_launcher m_launcher;
212	buffer_access_map m_access_map;
213	detail::side_effect_map m_side_effects;
214	reduction_set m_reductions;
215	std::string m_debug_name;
216	detail::epoch_action m_epoch_action;
217	std::unique_ptr<task_promise> m_promise; // TODO keep user_allocation_id in struct task instead of inside task_promise
218	std::vector<std::unique_ptr<hint_base>> m_hints;
219
220	task(task_id tid, task_type type, collective_group_id cgid, task_geometry geometry, command_group_launcher launcher, buffer_access_map access_map,	5,764✔
221	detail::side_effect_map side_effects, reduction_set reductions, detail::epoch_action epoch_action, std::unique_ptr<task_promise> promise)
222	: m_tid(tid), m_type(type), m_cgid(cgid), m_geometry(geometry), m_launcher(std::move(launcher)), m_access_map(std::move(access_map)),	5,764✔
223	m_side_effects(std::move(side_effects)), m_reductions(std::move(reductions)), m_epoch_action(epoch_action), m_promise(std::move(promise)) {	11,528✔
224	assert(type == task_type::host_compute \|\| type == task_type::device_compute \|\| get_granularity().size() == 1);	5,764✔
225	// Only host tasks can have side effects
226	assert(this->m_side_effects.empty() \|\| type == task_type::host_compute \|\| type == task_type::collective \|\| type == task_type::master_node	5,764✔
227	\|\| type == task_type::fence);
228	}	5,764✔
229	};
230
231	std::unique_ptr<detail::task> make_command_group_task(const detail::task_id tid, const size_t num_collective_nodes, raw_command_group&& cg);
232
233	[[nodiscard]] std::string print_task_debug_label(const task& tsk, bool title_case = false);
234
235	/// Determines which overlapping regions appear between write accesses when the iteration space of `tsk` is split into `chunks`.
236	std::unordered_map<buffer_id, region<3>> detect_overlapping_writes(const task& tsk, const box_vector<3>& chunks);
237
238	/// The task graph (TDAG) represents all cluster-wide operations, such as command group submissions and fences, and their interdependencies.
239	class task_graph : public graph<task> {}; // inheritance instead of type alias so we can forward declare task_graph
240
241	} // namespace detail
242	} // namespace celerity

celerity / celerity-runtime / 12047884020

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