27480689803

Committed 13 Jun 2026 10:14PM UTC coverage: 61.318%. First build

Build # 27480689803

Build Type

Pull #760

github

Committed by

web-flow

Commit Message

Merge 01073dcd3 into 9754c0592

Pull Request Pull Request #760: Add hardware limit checks for gpu

Coverage Stats

32 of 52 new or added lines in 4 files covered. (61.54%)

36305 of 59208 relevant lines covered (61.32%)

1122.86 hits per line

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

44.37

/opt/src/transformations/offloading/cuda_transform.cpp

#include "sdfg/transformations/offloading/cuda_transform.h"

#include <unordered_set>

#include "sdfg/structured_control_flow/block.h"
#include "sdfg/symbolic/symbolic.h"
#include "sdfg/targets/cuda/cuda_data_offloading_node.h"
#include "sdfg/transformations/transformation.h"
#include "symengine/symengine_rcp.h"

namespace sdfg {
namespace cuda {

std::string CUDATransform::name() const { return "CUDATransform"; }

bool CUDATransform::can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {
    if (!OffloadTransform::can_be_applied(builder, analysis_manager)) {
        return false;
    }

    // Condition: Resulting CUDA grid X-dimension must not exceed hardware limits.
    // X grid dimension is limited to 2^31 - 1.
    auto num_iters = this->map_.num_iterations();
    if (!num_iters.is_null() && SymEngine::is_a<SymEngine::Integer>(*num_iters)) {
        int64_t iters = SymEngine::down_cast<const SymEngine::Integer&>(*num_iters).as_int();
        int64_t block = static_cast<int64_t>(block_size_);
        int64_t grid_size = (iters + block - 1) / block;

        constexpr int64_t max_grid_dim_x = 2147483647; // 2^31 - 1
        if (grid_size > max_grid_dim_x) {
            return false;
        }
    }

    return true;
}

void CUDATransform::add_device_buffer(
    builder::StructuredSDFGBuilder& builder,
    std::string host_arg_name,
    std::string device_arg_name,
    symbolic::Expression arg_size
) {
    // Allocate device pointer
    auto& sdfg = builder.subject();
    auto& type = sdfg.type(host_arg_name);
    auto new_type = type.clone();
    new_type->storage_type(global_device_storage_type(arg_size));
    builder.add_container(device_arg_name, *new_type);
}

void CUDATransform::allocate_device_arg(
    builder::StructuredSDFGBuilder& builder,
    Block& alloc_block,
    std::string host_arg_name,
    std::string device_arg_name,
    symbolic::Expression arg_size,
    symbolic::Expression page_size
) {
    auto& sdfg = builder.subject();
    if (!builder.subject().exists(device_arg_name)) {
        auto& type = sdfg.type(host_arg_name);
        auto new_type = type.clone();
        new_type->storage_type(global_device_storage_type(arg_size));
        new_type->storage_type().allocation(types::StorageType::AllocationType::Unmanaged);
        new_type->storage_type().deallocation(types::StorageType::AllocationType::Unmanaged);
        new_type->storage_type().allocation_size(SymEngine::null);

        std::unordered_set<std::string> container_set(sdfg.containers().begin(), sdfg.containers().end());
        if (container_set.find(device_arg_name) == container_set.end()) {
            builder.add_container(device_arg_name, *new_type);
        }
    }

    auto& out_type = builder.subject().type(device_arg_name);

    offloading::add_offloading_node<CUDADataOffloadingNode>(
        builder,
        alloc_block,
        host_arg_name,
        device_arg_name,
        offloading::DataTransferDirection::NONE,
        offloading::BufferLifecycle::ALLOC,
        out_type,
        out_type,
        this->map_.debug_info(),
        arg_size,
        symbolic::zero()
    );
}

void CUDATransform::deallocate_device_arg(
    builder::StructuredSDFGBuilder& builder,
    Block& dealloc_block,
    std::string device_arg_name,
    symbolic::Expression arg_size,
    symbolic::Expression page_size
) {
    auto& free_type = builder.subject().type(device_arg_name);
    offloading::add_offloading_node<CUDADataOffloadingNode>(
        builder,
        dealloc_block,
        device_arg_name,
        device_arg_name,
        offloading::DataTransferDirection::NONE,
        offloading::BufferLifecycle::FREE,
        free_type,
        free_type,
        this->map_.debug_info(),
        arg_size,
        symbolic::zero()
    );
}

void CUDATransform::copy_to_device(
    builder::StructuredSDFGBuilder& builder,
    const std::string host_arg_name,
    std::string device_arg_name,
    symbolic::Expression size,
    symbolic::Expression page_size,
    Block& copy_block
) {
    offloading::add_offloading_node<CUDADataOffloadingNode>(
        builder,
        copy_block,
        host_arg_name,
        device_arg_name,
        offloading::DataTransferDirection::H2D,
        offloading::BufferLifecycle::NO_CHANGE,
        builder.subject().type(host_arg_name),
        builder.subject().type(device_arg_name),
        this->map_.debug_info(),
        size,
        symbolic::integer(0)
    );
}

void CUDATransform::copy_to_device_with_allocation(
    builder::StructuredSDFGBuilder& builder,
    const std::string host_arg_name,
    std::string device_arg_name,
    symbolic::Expression size,
    symbolic::Expression page_size,
    Block& copy_block
) {
    offloading::add_offloading_node<CUDADataOffloadingNode>(
        builder,
        copy_block,
        host_arg_name,
        device_arg_name,
        offloading::DataTransferDirection::H2D,
        offloading::BufferLifecycle::ALLOC,
        builder.subject().type(host_arg_name),
        builder.subject().type(device_arg_name),
        this->map_.debug_info(),
        size,
        symbolic::integer(0)
    );
}

void CUDATransform::copy_from_device(
    builder::StructuredSDFGBuilder& builder,
    Block& copy_out_block,
    const std::string host_arg_name,
    std::string device_arg_name,
    symbolic::Expression size,
    symbolic::Expression page_size
) {
    offloading::add_offloading_node<CUDADataOffloadingNode>(
        builder,
        copy_out_block,
        host_arg_name,
        device_arg_name,
        offloading::DataTransferDirection::D2H,
        offloading::BufferLifecycle::NO_CHANGE,
        builder.subject().type(host_arg_name),
        builder.subject().type(device_arg_name),
        this->map_.debug_info(),
        size,
        symbolic::integer(0)
    );
}

void CUDATransform::copy_from_device_with_free(
    builder::StructuredSDFGBuilder& builder,
    Block& copy_out_block,
    const std::string host_arg_name,
    std::string device_arg_name,
    symbolic::Expression size,
    symbolic::Expression page_size
) {
    offloading::add_offloading_node<CUDADataOffloadingNode>(
        builder,
        copy_out_block,
        host_arg_name,
        device_arg_name,
        offloading::DataTransferDirection::D2H,
        offloading::BufferLifecycle::FREE,
        builder.subject().type(host_arg_name),
        builder.subject().type(device_arg_name),
        this->map_.debug_info(),
        size,
        symbolic::integer(0)
    );
}

void CUDATransform::to_json(nlohmann::json& j) const {
    std::string loop_type;
    if (dynamic_cast<structured_control_flow::Map*>(&map_)) {
        loop_type = "map";
    } else {
        throw std::runtime_error("Unsupported loop type for serialization of loop: " + map_.indvar()->get_name());
    }

    j["transformation_type"] = this->name();
    j["subgraph"] = {{"0", {{"element_id", this->map_.element_id()}, {"type", loop_type}}}};
    j["parameters"] = {{"block_size", block_size_}};
};

CUDATransform CUDATransform::from_json(builder::StructuredSDFGBuilder& builder, const nlohmann::json& desc) {
    auto loop_id = desc["subgraph"]["0"]["element_id"].get<size_t>();
    size_t block_size = desc["parameters"]["block_size"].get<size_t>();
    auto element = builder.find_element_by_id(loop_id);
    if (!element) {
        throw transformations::
            InvalidTransformationDescriptionException("Element with ID " + std::to_string(loop_id) + " not found.");
    }
    auto map = dynamic_cast<structured_control_flow::Map*>(element);

    return CUDATransform(*map, block_size);
};


} // namespace cuda
} // namespace sdfg

1	#include "sdfg/transformations/offloading/cuda_transform.h"
2
3	#include <unordered_set>
4
5	#include "sdfg/structured_control_flow/block.h"
6	#include "sdfg/symbolic/symbolic.h"
7	#include "sdfg/targets/cuda/cuda_data_offloading_node.h"
8	#include "sdfg/transformations/transformation.h"
9	#include "symengine/symengine_rcp.h"
10
11	namespace sdfg {
12	namespace cuda {
13
14	std::string CUDATransform::name() const { return "CUDATransform"; }	3✔
15
16	bool CUDATransform::can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	7✔
17	if (!OffloadTransform::can_be_applied(builder, analysis_manager)) {	7✔
NEW 18	return false;	×
NEW 19	}	×
20
21	// Condition: Resulting CUDA grid X-dimension must not exceed hardware limits.
22	// X grid dimension is limited to 2^31 - 1.
23	auto num_iters = this->map_.num_iterations();	7✔
24	if (!num_iters.is_null() && SymEngine::is_a<SymEngine::Integer>(*num_iters)) {	7✔
25	int64_t iters = SymEngine::down_cast<const SymEngine::Integer&>(*num_iters).as_int();	3✔
26	int64_t block = static_cast<int64_t>(block_size_);	3✔
27	int64_t grid_size = (iters + block - 1) / block;	3✔
28
29	constexpr int64_t max_grid_dim_x = 2147483647; // 2^31 - 1	3✔
30	if (grid_size > max_grid_dim_x) {	3✔
NEW 31	return false;	×
NEW 32	}	×
33	}	3✔
34
35	return true;	7✔
36	}	7✔
37
38	void CUDATransform::add_device_buffer(
39	builder::StructuredSDFGBuilder& builder,
40	std::string host_arg_name,
41	std::string device_arg_name,
42	symbolic::Expression arg_size
43	) {	3✔
44	// Allocate device pointer
45	auto& sdfg = builder.subject();	3✔
46	auto& type = sdfg.type(host_arg_name);	3✔
47	auto new_type = type.clone();	3✔
48	new_type->storage_type(global_device_storage_type(arg_size));	3✔
49	builder.add_container(device_arg_name, *new_type);	3✔
50	}	3✔
51
52	void CUDATransform::allocate_device_arg(
53	builder::StructuredSDFGBuilder& builder,
54	Block& alloc_block,
55	std::string host_arg_name,
56	std::string device_arg_name,
57	symbolic::Expression arg_size,
58	symbolic::Expression page_size
59	) {	×
60	auto& sdfg = builder.subject();	×
61	if (!builder.subject().exists(device_arg_name)) {	×
62	auto& type = sdfg.type(host_arg_name);	×
63	auto new_type = type.clone();	×
64	new_type->storage_type(global_device_storage_type(arg_size));	×
65	new_type->storage_type().allocation(types::StorageType::AllocationType::Unmanaged);	×
66	new_type->storage_type().deallocation(types::StorageType::AllocationType::Unmanaged);	×
67	new_type->storage_type().allocation_size(SymEngine::null);	×
68
69	std::unordered_set<std::string> container_set(sdfg.containers().begin(), sdfg.containers().end());	×
70	if (container_set.find(device_arg_name) == container_set.end()) {	×
71	builder.add_container(device_arg_name, *new_type);	×
72	}	×
73	}	×
74
75	auto& out_type = builder.subject().type(device_arg_name);	×
76
77	offloading::add_offloading_node<CUDADataOffloadingNode>(	×
78	builder,	×
79	alloc_block,	×
80	host_arg_name,	×
81	device_arg_name,	×
82	offloading::DataTransferDirection::NONE,	×
83	offloading::BufferLifecycle::ALLOC,	×
84	out_type,	×
85	out_type,	×
86	this->map_.debug_info(),	×
87	arg_size,	×
88	symbolic::zero()	×
89	);	×
90	}	×
91
92	void CUDATransform::deallocate_device_arg(
93	builder::StructuredSDFGBuilder& builder,
94	Block& dealloc_block,
95	std::string device_arg_name,
96	symbolic::Expression arg_size,
97	symbolic::Expression page_size
98	) {	×
99	auto& free_type = builder.subject().type(device_arg_name);	×
100	offloading::add_offloading_node<CUDADataOffloadingNode>(	×
101	builder,	×
102	dealloc_block,	×
103	device_arg_name,	×
104	device_arg_name,	×
105	offloading::DataTransferDirection::NONE,	×
106	offloading::BufferLifecycle::FREE,	×
107	free_type,	×
108	free_type,	×
109	this->map_.debug_info(),	×
110	arg_size,	×
111	symbolic::zero()	×
112	);	×
113	}	×
114
115	void CUDATransform::copy_to_device(
116	builder::StructuredSDFGBuilder& builder,
117	const std::string host_arg_name,
118	std::string device_arg_name,
119	symbolic::Expression size,
120	symbolic::Expression page_size,
121	Block& copy_block
122	) {	×
123	offloading::add_offloading_node<CUDADataOffloadingNode>(	×
124	builder,	×
125	copy_block,	×
126	host_arg_name,	×
127	device_arg_name,	×
128	offloading::DataTransferDirection::H2D,	×
129	offloading::BufferLifecycle::NO_CHANGE,	×
130	builder.subject().type(host_arg_name),	×
131	builder.subject().type(device_arg_name),	×
132	this->map_.debug_info(),	×
133	size,	×
134	symbolic::integer(0)	×
135	);	×
136	}	×
137
138	void CUDATransform::copy_to_device_with_allocation(
139	builder::StructuredSDFGBuilder& builder,
140	const std::string host_arg_name,
141	std::string device_arg_name,
142	symbolic::Expression size,
143	symbolic::Expression page_size,
144	Block& copy_block
145	) {	4✔
146	offloading::add_offloading_node<CUDADataOffloadingNode>(	4✔
147	builder,	4✔
148	copy_block,	4✔
149	host_arg_name,	4✔
150	device_arg_name,	4✔
151	offloading::DataTransferDirection::H2D,	4✔
152	offloading::BufferLifecycle::ALLOC,	4✔
153	builder.subject().type(host_arg_name),	4✔
154	builder.subject().type(device_arg_name),	4✔
155	this->map_.debug_info(),	4✔
156	size,	4✔
157	symbolic::integer(0)	4✔
158	);	4✔
159	}	4✔
160
161	void CUDATransform::copy_from_device(
162	builder::StructuredSDFGBuilder& builder,
163	Block& copy_out_block,
164	const std::string host_arg_name,
165	std::string device_arg_name,
166	symbolic::Expression size,
167	symbolic::Expression page_size
168	) {	×
169	offloading::add_offloading_node<CUDADataOffloadingNode>(	×
170	builder,	×
171	copy_out_block,	×
172	host_arg_name,	×
173	device_arg_name,	×
174	offloading::DataTransferDirection::D2H,	×
175	offloading::BufferLifecycle::NO_CHANGE,	×
176	builder.subject().type(host_arg_name),	×
177	builder.subject().type(device_arg_name),	×
178	this->map_.debug_info(),	×
179	size,	×
180	symbolic::integer(0)	×
181	);	×
182	}	×
183
184	void CUDATransform::copy_from_device_with_free(
185	builder::StructuredSDFGBuilder& builder,
186	Block& copy_out_block,
187	const std::string host_arg_name,
188	std::string device_arg_name,
189	symbolic::Expression size,
190	symbolic::Expression page_size
191	) {	4✔
192	offloading::add_offloading_node<CUDADataOffloadingNode>(	4✔
193	builder,	4✔
194	copy_out_block,	4✔
195	host_arg_name,	4✔
196	device_arg_name,	4✔
197	offloading::DataTransferDirection::D2H,	4✔
198	offloading::BufferLifecycle::FREE,	4✔
199	builder.subject().type(host_arg_name),	4✔
200	builder.subject().type(device_arg_name),	4✔
201	this->map_.debug_info(),	4✔
202	size,	4✔
203	symbolic::integer(0)	4✔
204	);	4✔
205	}	4✔
206
207	void CUDATransform::to_json(nlohmann::json& j) const {	1✔
208	std::string loop_type;	1✔
209	if (dynamic_cast<structured_control_flow::Map*>(&map_)) {	1✔
210	loop_type = "map";	1✔
211	} else {	1✔
212	throw std::runtime_error("Unsupported loop type for serialization of loop: " + map_.indvar()->get_name());	×
213	}	×
214
215	j["transformation_type"] = this->name();	1✔
216	j["subgraph"] = {{"0", {{"element_id", this->map_.element_id()}, {"type", loop_type}}}};	1✔
217	j["parameters"] = {{"block_size", block_size_}};	1✔
218	};	1✔
219
220	CUDATransform CUDATransform::from_json(builder::StructuredSDFGBuilder& builder, const nlohmann::json& desc) {	1✔
221	auto loop_id = desc["subgraph"]["0"]["element_id"].get<size_t>();	1✔
222	size_t block_size = desc["parameters"]["block_size"].get<size_t>();	1✔
223	auto element = builder.find_element_by_id(loop_id);	1✔
224	if (!element) {	1✔
225	throw transformations::	×
226	InvalidTransformationDescriptionException("Element with ID " + std::to_string(loop_id) + " not found.");	×
227	}	×
228	auto map = dynamic_cast<structured_control_flow::Map*>(element);	1✔
229
230	return CUDATransform(*map, block_size);	1✔
231	};	1✔
232
233
234	} // namespace cuda
235	} // namespace sdfg

daisytuner / docc / 27480689803

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