15142284085

Committed 20 May 2025 03:58PM UTC coverage: 60.528% (-0.02%) from 60.543%

Build # 15142284085

Build Type

push

github

Committed by

web-flow

Commit Message

Merge pull request #23 from daisytuner/loops

LoopTrees and Instrumentation Strategies

Run Details

80 of 182 new or added lines in 17 files covered. (43.96%)

11 existing lines in 6 files now uncovered.

7928 of 13098 relevant lines covered (60.53%)

104.16 hits per line

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

64.97

/src/codegen/code_generators/cuda_code_generator.cpp

#include "sdfg/codegen/code_generators/cuda_code_generator.h"

#include "sdfg/codegen/dispatchers/node_dispatcher_factory.h"

#include "sdfg/codegen/instrumentation/instrumentation.h"
#include "sdfg/codegen/instrumentation/outermost_loops_instrumentation.h"

namespace sdfg {
namespace codegen {

CUDACodeGenerator::CUDACodeGenerator(ConditionalSchedule& schedule)
    : CodeGenerator(schedule, InstrumentationStrategy::NONE){

      };

CUDACodeGenerator::CUDACodeGenerator(ConditionalSchedule& schedule, InstrumentationStrategy instrumentation_strategy)
    : CodeGenerator(schedule, instrumentation_strategy){

      };

bool CUDACodeGenerator::generate() {
    this->dispatch_includes();
    this->dispatch_structures();
    this->dispatch_globals();
    this->dispatch_schedule();
    return true;
};

std::string CUDACodeGenerator::function_definition() {
    // Define SDFG as a function
    auto& function = this->schedule_.schedule(0).sdfg();

    /********** Arglist **********/
    std::vector<std::string> args;
    for (auto& container : function.arguments()) {
        args.push_back(language_extension_.declaration(container, function.type(container)));
    }
    std::stringstream arglist;
    arglist << sdfg::helpers::join(args, ", ");

    return "extern \"C\" __global__ void " + function.name() + "(" + arglist.str() + ")";
};

bool CUDACodeGenerator::as_source(const std::filesystem::path& header_path,
                                  const std::filesystem::path& source_path,
                                  const std::filesystem::path& library_path) {
    std::ofstream ofs_header(header_path, std::ofstream::out);
    if (!ofs_header.is_open()) {
        return false;
    }

    std::ofstream ofs_source(source_path, std::ofstream::out);
    if (!ofs_source.is_open()) {
        return false;
    }

    std::ofstream ofs_library(library_path, std::ofstream::out);
    if (!ofs_library.is_open()) {
        return false;
    }

    ofs_header << "#pragma once" << std::endl;
    ofs_header << this->includes_stream_.str() << std::endl;
    ofs_header << this->classes_stream_.str() << std::endl;
    ofs_header.close();

    ofs_source << "#include \"" << header_path.filename().string() << "\"" << std::endl;
    ofs_source << this->globals_stream_.str() << std::endl;
    ofs_source << this->function_definition() << std::endl;
    ofs_source << "{" << std::endl;
    ofs_source << this->main_stream_.str() << std::endl;
    ofs_source << "}" << std::endl;
    ofs_source.close();

    auto library_content = this->library_stream_.str();
    if (library_content.empty()) {
        ofs_library.close();
        return true;
    }

    ofs_library << "#include \"" << header_path.filename().string() << "\"" << std::endl;
    ofs_library << library_content << std::endl;
    ofs_library.close();

    return true;
};

void CUDACodeGenerator::dispatch_includes() {
    this->includes_stream_ << "#define "
                           << "__DAISY_NVVM__" << std::endl;
    this->includes_stream_ << "#include "
                           << "\"daisyrtl.h\"" << std::endl;
    if (instrumentation_strategy_ != InstrumentationStrategy::NONE) this->includes_stream_ << "#include <daisy_rtl.h>" << std::endl;

    this->includes_stream_ << "#define __daisy_min(a,b) ((a)<(b)?(a):(b))" << std::endl;
    this->includes_stream_ << "#define __daisy_max(a,b) ((a)>(b)?(a):(b))" << std::endl;
    this->includes_stream_ << "#define __daisy_fma(a,b,c) a * b + c" << std::endl;
};

void CUDACodeGenerator::dispatch_structures() {
    auto& function = this->schedule_.schedule(0).sdfg();

    // Forward declarations
    for (auto& structure : function.structures()) {
        this->classes_stream_ << "struct " << structure << ";" << std::endl;
    }

    // Generate topology-sorted structure definitions
    typedef boost::adjacency_list<boost::vecS, boost::vecS, boost::bidirectionalS> structures_graph;
    typedef boost::graph_traits<structures_graph>::vertex_descriptor Vertex;
    std::vector<std::string> names;
    for (auto& structure : function.structures()) {
        names.push_back(structure);
    }
    structures_graph graph(names.size());

    for (auto& structure : names) {
        auto& definition = function.structure(structure);
        for (size_t i = 0; i < definition.num_members(); i++) {
            auto member_type = &definition.member_type(symbolic::integer(i));
            while (dynamic_cast<const types::Array*>(member_type)) {
                auto array_type = static_cast<const types::Array*>(member_type);
                member_type = &array_type->element_type();
            }

            if (auto member_structure = dynamic_cast<const sdfg::types::Structure*>(member_type)) {
                boost::add_edge(
                    std::find(names.begin(), names.end(), member_structure->name()) - names.begin(),
                    std::find(names.begin(), names.end(), structure) - names.begin(), graph);
            }
        }
    }

    std::list<Vertex> order;
    std::unordered_map<Vertex, boost::default_color_type> vertex_colors;
    boost::topological_sort(graph, std::back_inserter(order),
                            boost::color_map(boost::make_assoc_property_map(vertex_colors)));
    order.reverse();

    for (auto& structure_index : order) {
        std::string structure = names.at(structure_index);
        auto& definition = function.structure(structure);
        this->classes_stream_ << "struct ";
        if (definition.is_packed()) {
            this->classes_stream_ << "__attribute__((packed)) ";
        }
        this->classes_stream_ << structure << std::endl;
        this->classes_stream_ << "{\n";

        for (size_t i = 0; i < definition.num_members(); i++) {
            auto& member_type = definition.member_type(symbolic::integer(i));
            if (dynamic_cast<const sdfg::types::Structure*>(&member_type)) {
                this->classes_stream_ << "struct ";
            }
            this->classes_stream_ << language_extension_.declaration("member_" + std::to_string(i),
                                                                     member_type);
            this->classes_stream_ << ";" << std::endl;
        }

        this->classes_stream_ << "};" << std::endl;
    }
};

void CUDACodeGenerator::dispatch_globals() {
    auto& function = schedule_.schedule(0).sdfg();
    for (auto& container : function.externals()) {
        auto& type = function.type(container);
        if (type.address_space() != 3 && type.address_space() != 4) {
            this->globals_stream_ << "extern " << language_extension_.declaration(container, type)
                                  << ";" << std::endl;
        }
        if (type.address_space() == 4) {
            assert(type.initializer().empty());
            this->globals_stream_ << "__constant__ "
                                  << language_extension_.declaration(container, type, true) << ";"
                                  << std::endl;
        }
    }
};

void CUDACodeGenerator::dispatch_schedule() {
    auto& function = schedule_.schedule(0).sdfg();

    // Declare shared memory
    for (auto& container : function.externals()) {
        auto& type = function.type(container);
        if (type.address_space() == 3) {
            this->main_stream_ << language_extension_.declaration(container,
                                                                  function.type(container))
                               << ";" << std::endl;
        }
    }

    // Map external variables to internal variables
    for (auto& container : function.containers()) {
        if (!function.is_internal(container)) {
            continue;
        }

        std::string external_name =
            container.substr(0, container.length() - external_suffix.length());
        this->main_stream_ << language_extension_.declaration(container, function.type(container));
        this->main_stream_ << " = "
                           << "&" << external_name;
        this->main_stream_ << ";" << std::endl;
    }

    for (size_t i = 0; i < schedule_.size(); i++) {
        auto& schedule = schedule_.schedule(i);
        auto condition = schedule_.condition(i);
        
        // Add instrumentation
        auto instrumentation = create_instrumentation(instrumentation_strategy_, schedule);

        if (i > 0) {
            this->main_stream_ << "else ";
        }

        this->main_stream_ << "if (" << language_extension_.expression(condition)
                               << ") {\n";

        auto& function_i = schedule.builder().subject();
        auto dispatcher = create_dispatcher(language_extension_, schedule,
                                            function_i.root(), *instrumentation);
        dispatcher->dispatch(this->main_stream_, this->globals_stream_, this->library_stream_);

        this->main_stream_ << "}\n";
    }
};

}  // namespace codegen
}  // namespace sdfg

1	#include "sdfg/codegen/code_generators/cuda_code_generator.h"
2
3	#include "sdfg/codegen/dispatchers/node_dispatcher_factory.h"
4
5	#include "sdfg/codegen/instrumentation/instrumentation.h"
6	#include "sdfg/codegen/instrumentation/outermost_loops_instrumentation.h"
7
8	namespace sdfg {
9	namespace codegen {
10
11	CUDACodeGenerator::CUDACodeGenerator(ConditionalSchedule& schedule)	10✔
12	: CodeGenerator(schedule, InstrumentationStrategy::NONE){	5✔
13
14	};	5✔
15
NEW 16	CUDACodeGenerator::CUDACodeGenerator(ConditionalSchedule& schedule, InstrumentationStrategy instrumentation_strategy)	×
NEW 17	: CodeGenerator(schedule, instrumentation_strategy){	×
18
UNCOV 19	};	×
20
21	bool CUDACodeGenerator::generate() {	4✔
22	this->dispatch_includes();	4✔
23	this->dispatch_structures();	4✔
24	this->dispatch_globals();	4✔
25	this->dispatch_schedule();	4✔
26	return true;	4✔
27	};
28
29	std::string CUDACodeGenerator::function_definition() {	1✔
30	// Define SDFG as a function
31	auto& function = this->schedule_.schedule(0).sdfg();	1✔
32
33	/******** Arglist ********/
34	std::vector<std::string> args;	1✔
35	for (auto& container : function.arguments()) {	1✔
36	args.push_back(language_extension_.declaration(container, function.type(container)));	×
37	}
38	std::stringstream arglist;	1✔
39	arglist << sdfg::helpers::join(args, ", ");	1✔
40
41	return "extern \"C\" __global__ void " + function.name() + "(" + arglist.str() + ")";	1✔
42	};	1✔
43
44	bool CUDACodeGenerator::as_source(const std::filesystem::path& header_path,	×
45	const std::filesystem::path& source_path,
46	const std::filesystem::path& library_path) {
47	std::ofstream ofs_header(header_path, std::ofstream::out);	×
48	if (!ofs_header.is_open()) {	×
49	return false;	×
50	}
51
52	std::ofstream ofs_source(source_path, std::ofstream::out);	×
53	if (!ofs_source.is_open()) {	×
54	return false;	×
55	}
56
57	std::ofstream ofs_library(library_path, std::ofstream::out);	×
58	if (!ofs_library.is_open()) {	×
59	return false;	×
60	}
61
62	ofs_header << "#pragma once" << std::endl;	×
63	ofs_header << this->includes_stream_.str() << std::endl;	×
64	ofs_header << this->classes_stream_.str() << std::endl;	×
65	ofs_header.close();	×
66
67	ofs_source << "#include \"" << header_path.filename().string() << "\"" << std::endl;	×
68	ofs_source << this->globals_stream_.str() << std::endl;	×
69	ofs_source << this->function_definition() << std::endl;	×
70	ofs_source << "{" << std::endl;	×
71	ofs_source << this->main_stream_.str() << std::endl;	×
72	ofs_source << "}" << std::endl;	×
73	ofs_source.close();	×
74
75	auto library_content = this->library_stream_.str();	×
76	if (library_content.empty()) {	×
77	ofs_library.close();	×
78	return true;	×
79	}
80
81	ofs_library << "#include \"" << header_path.filename().string() << "\"" << std::endl;	×
82	ofs_library << library_content << std::endl;	×
83	ofs_library.close();	×
84
85	return true;	×
86	};	×
87
88	void CUDACodeGenerator::dispatch_includes() {	4✔
89	this->includes_stream_ << "#define "	4✔
90	<< "__DAISY_NVVM__" << std::endl;	4✔
91	this->includes_stream_ << "#include "	4✔
92	<< "\"daisyrtl.h\"" << std::endl;	4✔
93	if (instrumentation_strategy_ != InstrumentationStrategy::NONE) this->includes_stream_ << "#include <daisy_rtl.h>" << std::endl;	4✔
94
95	this->includes_stream_ << "#define __daisy_min(a,b) ((a)<(b)?(a):(b))" << std::endl;	4✔
96	this->includes_stream_ << "#define __daisy_max(a,b) ((a)>(b)?(a):(b))" << std::endl;	4✔
97	this->includes_stream_ << "#define __daisy_fma(a,b,c) a * b + c" << std::endl;	4✔
98	};	4✔
99
100	void CUDACodeGenerator::dispatch_structures() {	4✔
101	auto& function = this->schedule_.schedule(0).sdfg();	4✔
102
103	// Forward declarations
104	for (auto& structure : function.structures()) {	7✔
105	this->classes_stream_ << "struct " << structure << ";" << std::endl;	3✔
106	}
107
108	// Generate topology-sorted structure definitions
109	typedef boost::adjacency_list<boost::vecS, boost::vecS, boost::bidirectionalS> structures_graph;
110	typedef boost::graph_traits<structures_graph>::vertex_descriptor Vertex;
111	std::vector<std::string> names;	4✔
112	for (auto& structure : function.structures()) {	7✔
113	names.push_back(structure);	3✔
114	}
115	structures_graph graph(names.size());	4✔
116
117	for (auto& structure : names) {	7✔
118	auto& definition = function.structure(structure);	3✔
119	for (size_t i = 0; i < definition.num_members(); i++) {	6✔
120	auto member_type = &definition.member_type(symbolic::integer(i));	3✔
121	while (dynamic_cast<const types::Array*>(member_type)) {	3✔
122	auto array_type = static_cast<const types::Array*>(member_type);	×
123	member_type = &array_type->element_type();	×
124	}
125
126	if (auto member_structure = dynamic_cast<const sdfg::types::Structure*>(member_type)) {	3✔
127	boost::add_edge(	1✔
128	std::find(names.begin(), names.end(), member_structure->name()) - names.begin(),	1✔
129	std::find(names.begin(), names.end(), structure) - names.begin(), graph);	1✔
130	}	1✔
131	}	3✔
132	}
133
134	std::list<Vertex> order;	4✔
135	std::unordered_map<Vertex, boost::default_color_type> vertex_colors;	4✔
136	boost::topological_sort(graph, std::back_inserter(order),	8✔
137	boost::color_map(boost::make_assoc_property_map(vertex_colors)));	4✔
138	order.reverse();	4✔
139
140	for (auto& structure_index : order) {	7✔
141	std::string structure = names.at(structure_index);	3✔
142	auto& definition = function.structure(structure);	3✔
143	this->classes_stream_ << "struct ";	3✔
144	if (definition.is_packed()) {	3✔
145	this->classes_stream_ << "__attribute__((packed)) ";	×
146	}	×
147	this->classes_stream_ << structure << std::endl;	3✔
148	this->classes_stream_ << "{\n";	3✔
149
150	for (size_t i = 0; i < definition.num_members(); i++) {	6✔
151	auto& member_type = definition.member_type(symbolic::integer(i));	3✔
152	if (dynamic_cast<const sdfg::types::Structure*>(&member_type)) {	3✔
153	this->classes_stream_ << "struct ";	1✔
154	}	1✔
155	this->classes_stream_ << language_extension_.declaration("member_" + std::to_string(i),	3✔
156	member_type);	3✔
157	this->classes_stream_ << ";" << std::endl;	3✔
158	}	3✔
159
160	this->classes_stream_ << "};" << std::endl;	3✔
161	}	3✔
162	};	4✔
163
164	void CUDACodeGenerator::dispatch_globals() {	4✔
165	auto& function = schedule_.schedule(0).sdfg();	4✔
166	for (auto& container : function.externals()) {	5✔
167	auto& type = function.type(container);	1✔
168	if (type.address_space() != 3 && type.address_space() != 4) {	1✔
169	this->globals_stream_ << "extern " << language_extension_.declaration(container, type)	2✔
170	<< ";" << std::endl;	1✔
171	}	1✔
172	if (type.address_space() == 4) {	1✔
173	assert(type.initializer().empty());	×
174	this->globals_stream_ << "__constant__ "	×
175	<< language_extension_.declaration(container, type, true) << ";"	×
176	<< std::endl;	×
177	}	×
178	}
179	};	4✔
180
181	void CUDACodeGenerator::dispatch_schedule() {	4✔
182	auto& function = schedule_.schedule(0).sdfg();	4✔
183
184	// Declare shared memory
185	for (auto& container : function.externals()) {	5✔
186	auto& type = function.type(container);	1✔
187	if (type.address_space() == 3) {	1✔
188	this->main_stream_ << language_extension_.declaration(container,	×
189	function.type(container))	×
190	<< ";" << std::endl;	×
191	}	×
192	}
193
194	// Map external variables to internal variables
195	for (auto& container : function.containers()) {	5✔
196	if (!function.is_internal(container)) {	1✔
197	continue;	1✔
198	}
199
200	std::string external_name =
201	container.substr(0, container.length() - external_suffix.length());	×
202	this->main_stream_ << language_extension_.declaration(container, function.type(container));	×
203	this->main_stream_ << " = "	×
204	<< "&" << external_name;	×
205	this->main_stream_ << ";" << std::endl;	×
206	}	×
207
208	for (size_t i = 0; i < schedule_.size(); i++) {	8✔
209	auto& schedule = schedule_.schedule(i);	4✔
210	auto condition = schedule_.condition(i);	4✔
211
212	// Add instrumentation
213	auto instrumentation = create_instrumentation(instrumentation_strategy_, schedule);	4✔
214
215	if (i > 0) {	4✔
NEW 216	this->main_stream_ << "else ";	×
NEW 217	}	×
218
219	this->main_stream_ << "if (" << language_extension_.expression(condition)	8✔
220	<< ") {\n";	4✔
221
222	auto& function_i = schedule.builder().subject();	4✔
223	auto dispatcher = create_dispatcher(language_extension_, schedule,	4✔
224	function_i.root(), *instrumentation);	4✔
225	dispatcher->dispatch(this->main_stream_, this->globals_stream_, this->library_stream_);	4✔
226
227	this->main_stream_ << "}\n";	4✔
228	}	4✔
229	};	4✔
230
231	} // namespace codegen
232	} // namespace sdfg

daisytuner / sdfglib / 15142284085

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