18187030480

Committed 02 Oct 2025 08:00AM UTC coverage: 61.116% (+0.05%) from 61.068%

Build # 18187030480

Build Type

push

github

Committed by

web-flow

Commit Message

Merge pull request #258 from daisytuner/debug-print-macro

adds macro for debug prints

Run Details

0 of 16 new or added lines in 5 files covered. (0.0%)

10 existing lines in 2 files now uncovered.

9569 of 15657 relevant lines covered (61.12%)

111.31 hits per line

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

64.23

/src/codegen/code_generators/cuda_code_generator.cpp

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

#include "sdfg/codegen/dispatchers/node_dispatcher_registry.h"
#include "sdfg/codegen/instrumentation/instrumentation_plan.h"
#include "sdfg/helpers/helpers.h"

namespace sdfg {
namespace codegen {

CUDACodeGenerator::CUDACodeGenerator(
    StructuredSDFG& sdfg,
    InstrumentationPlan& instrumentation_plan,
    bool capture_args_results,
    const std::pair<std::filesystem::path, std::filesystem::path>* output_and_header_paths
)
    : CodeGenerator(sdfg, instrumentation_plan, capture_args_results, output_and_header_paths),
      language_extension_(sdfg.externals()) {
    if (sdfg.type() != FunctionType_NV_GLOBAL) {
        throw std::runtime_error("CUDACodeGenerator can only be used for GPU SDFGs");
    }
    if (capture_args_results) {
        DEBUG_PRINTLN("CUDACodeGenerator does not support capturing args/results!");
    }
};

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

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

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

bool CUDACodeGenerator::as_source(const std::filesystem::path& header_path, const std::filesystem::path& source_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;
    }

    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;

    append_function_source(ofs_source);

    ofs_source.close();

    return true;
};

void CUDACodeGenerator::append_function_source(std::ofstream& ofs_source) {
    ofs_source << this->function_definition() << std::endl;
    ofs_source << "{" << std::endl;
    ofs_source << this->main_stream_.str() << std::endl;
    ofs_source << "}" << std::endl;
}

void CUDACodeGenerator::dispatch_includes() {
    this->includes_stream_ << "#define "
                           << "__DAISY_NVVM__" << std::endl;
    this->includes_stream_ << "#include <daisy_rtl/daisy_rtl.h>" << std::endl;
};

void CUDACodeGenerator::dispatch_structures() {
    // Forward declarations
    for (auto& structure : sdfg_.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 : sdfg_.structures()) {
        names.push_back(structure);
    }
    structures_graph graph(names.size());

    for (auto& structure : names) {
        auto& definition = sdfg_.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 = sdfg_.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, false, true);
            this->classes_stream_ << ";" << std::endl;
        }

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

void CUDACodeGenerator::dispatch_globals() {
    for (auto& container : sdfg_.externals()) {
        auto& type = sdfg_.type(container);
        if (type.storage_type() == types::StorageType_NV_Global) {
            auto& base_type = dynamic_cast<const types::Pointer&>(type).pointee_type();
            if (sdfg_.linkage_type(container) == LinkageType_External) {
                this->globals_stream_ << "extern " << language_extension_.declaration(container, base_type) << ";"
                                      << std::endl;
            } else {
                this->globals_stream_ << "static " << language_extension_.declaration(container, base_type);
                if (!type.initializer().empty()) {
                    this->globals_stream_ << " = " << type.initializer();
                }
                this->globals_stream_ << ";" << std::endl;
            }
        }
        if (type.storage_type() == types::StorageType_NV_Constant) {
            assert(type.initializer().empty());
            auto& base_type = dynamic_cast<const types::Pointer&>(type).pointee_type();
            this->globals_stream_ << "__constant__ " << language_extension_.declaration(container, base_type, true)
                                  << ";" << std::endl;
        }
    }
};

void CUDACodeGenerator::dispatch_schedule() {
    // Declare shared memory
    for (auto& container : sdfg_.externals()) {
        auto& type = sdfg_.type(container);
        if (type.storage_type() == types::StorageType_NV_Shared) {
            this->main_stream_ << language_extension_.declaration(container, sdfg_.type(container)) << ";" << std::endl;
        }
    }

    // Declare transient containers
    for (auto& container : sdfg_.containers()) {
        if (!sdfg_.is_transient(container)) {
            continue;
        }

        std::string val = this->language_extension_.declaration(container, sdfg_.type(container), false, true);
        if (!val.empty()) {
            this->main_stream_ << val;
            this->main_stream_ << ";" << std::endl;
        }
    }

    auto dispatcher = create_dispatcher(language_extension_, sdfg_, sdfg_.root(), instrumentation_plan_);
    dispatcher->dispatch(this->main_stream_, this->globals_stream_, this->library_snippet_factory_);
};

} // namespace codegen
} // namespace sdfg

1	#include "sdfg/codegen/code_generators/cuda_code_generator.h"
2
3	#include "sdfg/codegen/dispatchers/node_dispatcher_registry.h"
4	#include "sdfg/codegen/instrumentation/instrumentation_plan.h"
5	#include "sdfg/helpers/helpers.h"
6
7	namespace sdfg {
8	namespace codegen {
9
10	CUDACodeGenerator::CUDACodeGenerator(	4✔
11	StructuredSDFG& sdfg,
12	InstrumentationPlan& instrumentation_plan,
13	bool capture_args_results,
14	const std::pair<std::filesystem::path, std::filesystem::path>* output_and_header_paths
15	)
16	: CodeGenerator(sdfg, instrumentation_plan, capture_args_results, output_and_header_paths),	4✔
17	language_extension_(sdfg.externals()) {	4✔
18	if (sdfg.type() != FunctionType_NV_GLOBAL) {	4✔
19	throw std::runtime_error("CUDACodeGenerator can only be used for GPU SDFGs");	×
20	}
21	if (capture_args_results) {	4✔
NEW 22	DEBUG_PRINTLN("CUDACodeGenerator does not support capturing args/results!");	×
23	}	×
24	};	4✔
25
26	bool CUDACodeGenerator::generate() {	3✔
27	this->dispatch_includes();	3✔
28	this->dispatch_structures();	3✔
29	this->dispatch_globals();	3✔
30	this->dispatch_schedule();	3✔
31	return true;	3✔
32	};
33
34	std::string CUDACodeGenerator::function_definition() {	1✔
35	/******** Arglist ********/
36	std::vector<std::string> args;	1✔
37	for (auto& container : sdfg_.arguments()) {	1✔
38	args.push_back(language_extension_.declaration(container, sdfg_.type(container)));	×
39	}
40	std::stringstream arglist;	1✔
41	arglist << sdfg::helpers::join(args, ", ");	1✔
42
43	return "extern \"C\" __global__ void " + sdfg_.name() + "(" + arglist.str() + ")";	1✔
44	};	1✔
45
46	bool CUDACodeGenerator::as_source(const std::filesystem::path& header_path, const std::filesystem::path& source_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	ofs_header << "#pragma once" << std::endl;	×
58	ofs_header << this->includes_stream_.str() << std::endl;	×
59	ofs_header << this->classes_stream_.str() << std::endl;	×
60	ofs_header.close();	×
61
62	ofs_source << "#include \"" << header_path.filename().string() << "\"" << std::endl;	×
63
64	ofs_source << this->globals_stream_.str() << std::endl;	×
65
66	append_function_source(ofs_source);	×
67
68	ofs_source.close();	×
69
70	return true;	×
71	};	×
72
73	void CUDACodeGenerator::append_function_source(std::ofstream& ofs_source) {	×
74	ofs_source << this->function_definition() << std::endl;	×
75	ofs_source << "{" << std::endl;	×
76	ofs_source << this->main_stream_.str() << std::endl;	×
77	ofs_source << "}" << std::endl;	×
78	}	×
79
80	void CUDACodeGenerator::dispatch_includes() {	3✔
81	this->includes_stream_ << "#define "	3✔
82	<< "__DAISY_NVVM__" << std::endl;	3✔
83	this->includes_stream_ << "#include <daisy_rtl/daisy_rtl.h>" << std::endl;	3✔
84	};	3✔
85
86	void CUDACodeGenerator::dispatch_structures() {	3✔
87	// Forward declarations
88	for (auto& structure : sdfg_.structures()) {	6✔
89	this->classes_stream_ << "struct " << structure << ";" << std::endl;	3✔
90	}
91
92	// Generate topology-sorted structure definitions
93	typedef boost::adjacency_list<boost::vecS, boost::vecS, boost::bidirectionalS> structures_graph;
94	typedef boost::graph_traits<structures_graph>::vertex_descriptor Vertex;
95	std::vector<std::string> names;	3✔
96	for (auto& structure : sdfg_.structures()) {	6✔
97	names.push_back(structure);	3✔
98	}
99	structures_graph graph(names.size());	3✔
100
101	for (auto& structure : names) {	6✔
102	auto& definition = sdfg_.structure(structure);	3✔
103	for (size_t i = 0; i < definition.num_members(); i++) {	6✔
104	auto member_type = &definition.member_type(symbolic::integer(i));	3✔
105	while (dynamic_cast<const types::Array*>(member_type)) {	3✔
106	auto array_type = static_cast<const types::Array*>(member_type);	×
107	member_type = &array_type->element_type();	×
108	}
109
110	if (auto member_structure = dynamic_cast<const sdfg::types::Structure*>(member_type)) {	3✔
111	boost::add_edge(	1✔
112	std::find(names.begin(), names.end(), member_structure->name()) - names.begin(),	1✔
113	std::find(names.begin(), names.end(), structure) - names.begin(),	1✔
114	graph
115	);
116	}	1✔
117	}	3✔
118	}
119
120	std::list<Vertex> order;	3✔
121	std::unordered_map<Vertex, boost::default_color_type> vertex_colors;	3✔
122	boost::topological_sort(	3✔
123	graph, std::back_inserter(order), boost::color_map(boost::make_assoc_property_map(vertex_colors))	3✔
124	);
125	order.reverse();	3✔
126
127	for (auto& structure_index : order) {	6✔
128	std::string structure = names.at(structure_index);	3✔
129	auto& definition = sdfg_.structure(structure);	3✔
130	this->classes_stream_ << "struct ";	3✔
131	if (definition.is_packed()) {	3✔
132	this->classes_stream_ << "__attribute__((packed)) ";	×
133	}	×
134	this->classes_stream_ << structure << std::endl;	3✔
135	this->classes_stream_ << "{\n";	3✔
136
137	for (size_t i = 0; i < definition.num_members(); i++) {	6✔
138	auto& member_type = definition.member_type(symbolic::integer(i));	3✔
139	if (dynamic_cast<const sdfg::types::Structure*>(&member_type)) {	3✔
140	this->classes_stream_ << "struct ";	1✔
141	}	1✔
142	this->classes_stream_	6✔
143	<< language_extension_.declaration("member_" + std::to_string(i), member_type, false, true);	3✔
144	this->classes_stream_ << ";" << std::endl;	3✔
145	}	3✔
146
147	this->classes_stream_ << "};" << std::endl;	3✔
148	}	3✔
149	};	3✔
150
151	void CUDACodeGenerator::dispatch_globals() {	3✔
152	for (auto& container : sdfg_.externals()) {	4✔
153	auto& type = sdfg_.type(container);	1✔
154	if (type.storage_type() == types::StorageType_NV_Global) {	1✔
155	auto& base_type = dynamic_cast<const types::Pointer&>(type).pointee_type();	1✔
156	if (sdfg_.linkage_type(container) == LinkageType_External) {	1✔
157	this->globals_stream_ << "extern " << language_extension_.declaration(container, base_type) << ";"	2✔
158	<< std::endl;	1✔
159	} else {	1✔
160	this->globals_stream_ << "static " << language_extension_.declaration(container, base_type);	×
161	if (!type.initializer().empty()) {	×
162	this->globals_stream_ << " = " << type.initializer();	×
163	}	×
164	this->globals_stream_ << ";" << std::endl;	×
165	}
166	}	1✔
167	if (type.storage_type() == types::StorageType_NV_Constant) {	1✔
168	assert(type.initializer().empty());	×
169	auto& base_type = dynamic_cast<const types::Pointer&>(type).pointee_type();	×
170	this->globals_stream_ << "__constant__ " << language_extension_.declaration(container, base_type, true)	×
171	<< ";" << std::endl;	×
172	}	×
173	}
174	};	3✔
175
176	void CUDACodeGenerator::dispatch_schedule() {	3✔
177	// Declare shared memory
178	for (auto& container : sdfg_.externals()) {	4✔
179	auto& type = sdfg_.type(container);	1✔
180	if (type.storage_type() == types::StorageType_NV_Shared) {	1✔
181	this->main_stream_ << language_extension_.declaration(container, sdfg_.type(container)) << ";" << std::endl;	×
182	}	×
183	}
184
185	// Declare transient containers
186	for (auto& container : sdfg_.containers()) {	4✔
187	if (!sdfg_.is_transient(container)) {	1✔
188	continue;	1✔
189	}
190
191	std::string val = this->language_extension_.declaration(container, sdfg_.type(container), false, true);	×
192	if (!val.empty()) {	×
193	this->main_stream_ << val;	×
194	this->main_stream_ << ";" << std::endl;	×
195	}	×
196	}	×
197
198	auto dispatcher = create_dispatcher(language_extension_, sdfg_, sdfg_.root(), instrumentation_plan_);	3✔
199	dispatcher->dispatch(this->main_stream_, this->globals_stream_, this->library_snippet_factory_);	3✔
200	};	3✔
201
202	} // namespace codegen
203	} // namespace sdfg

daisytuner / sdfglib / 18187030480

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