27704607727

Committed 17 Jun 2026 04:38PM UTC coverage: 61.63% (+0.1%) from 61.533%

Build # 27704607727

Build Type

Pull #764

github

Committed by

web-flow

Commit Message

Merge 52a992b2c into 5b779fb0a

Pull Request Pull Request #764: simplifies GPU transformations

Coverage Stats

255 of 329 new or added lines in 6 files covered. (77.51%)

5 existing lines in 2 files now uncovered.

36652 of 59471 relevant lines covered (61.63%)

1139.64 hits per line

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

69.49

/opt/src/transformations/offloading/offload_transform.cpp

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

#include <map>
#include <string>

#include "sdfg/analysis/type_analysis.h"
#include "sdfg/data_flow/access_node.h"
#include "sdfg/structured_control_flow/block.h"
#include "sdfg/structured_control_flow/if_else.h"
#include "sdfg/structured_control_flow/map.h"
#include "sdfg/symbolic/symbolic.h"

#include "sdfg/data_flow/library_node.h"
#include "sdfg/optimization_report/pass_report_consumer.h"
#include "sdfg/types/utils.h"
#include "sdfg/visitor/immutable_structured_sdfg_visitor.h"
#include "symengine/symengine_rcp.h"

namespace sdfg {
namespace transformations {

class SideEffectFinder : public visitor::ImmutableStructuredSDFGVisitor {
private:
    structured_control_flow::Map& map_;

public:
    SideEffectFinder(StructuredSDFG& sdfg, analysis::AnalysisManager& analysis_manager, structured_control_flow::Map& map)
        : visitor::ImmutableStructuredSDFGVisitor(sdfg, analysis_manager), map_(map) {}

    bool visit() override { return visit_internal(map_.root()); }

    bool accept(structured_control_flow::Block& node) override {
        for (const auto& lib_node : node.dataflow().library_nodes()) {
            if (lib_node->side_effect()) {
                return true;
            }
        }
        return false;
    }
};

OffloadTransform::OffloadTransform(structured_control_flow::Map& map, bool allow_dynamic_sizes)
    : map_(map), allow_dynamic_sizes_(allow_dynamic_sizes) {}


bool OffloadTransform::can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {
    auto& sdfg = builder.subject();

    auto& arguments_analysis = analysis_manager.get<analysis::ArgumentsAnalysis>();

    if (!arguments_analysis.inferred_types(analysis_manager, this->map_)) {
        if (report_) report_->transform_impossible(this, "unranged args");
        DEBUG_PRINTLN("Cannot apply transform: argument types not inferred");
        return false;
    }
    auto& arguments = arguments_analysis.arguments(analysis_manager, this->map_);

    // Criterion: arg Data Types must be continuous
    for (auto& [argument, meta] : arguments) {
        auto base_type = analysis::TypeAnalysis(sdfg, &map_, analysis_manager).get_outer_type(argument);
        if (base_type == nullptr) {
            if (report_) report_->transform_impossible(this, "cannot infer type");
            DEBUG_PRINTLN("Cannot apply transform: argument type cannot be inferred");
            return false;
        }
        if (!types::is_contiguous_type(*base_type, sdfg)) {
            if (report_) report_->transform_impossible(this, "type is not contiguous");
            DEBUG_PRINTLN("Cannot apply transform: argument type is not contiguous");
            return false;
        }
        if (meta.is_scalar && meta.is_output) {
            if (report_) report_->transform_impossible(this, "scalar output");
            DEBUG_PRINTLN("Cannot apply transform: map writes to scalar argument");
            return false;
        }
    }

    // Note: arbitrary `init` and `stride` are permitted on the kernel-boundary
    // Map. The CUDA/ROCm dispatchers emit
    //   `<map.indvar> = init + thread_flat_id * stride`,
    // and `num_iterations()` already accounts for both when computing the grid
    // geometry.
    if (map_.num_iterations().is_null()) {
        if (report_) report_->transform_impossible(this, "cannot determine num iterations");
        DEBUG_PRINTLN("Cannot apply transform: cannot determine number of iterations for map");
        return false;
    }

    // Criterion: Map cannot write to scalar arguments
    for (auto& [argument, meta] : arguments) {
        if (meta.is_scalar && meta.is_output) {
            if (report_) report_->transform_impossible(this, "scalar output");
            DEBUG_PRINTLN("Cannot apply transform: map writes to scalar argument");
            return false;
        }
    }

    if (!arguments_analysis.argument_size_known(analysis_manager, this->map_, allow_dynamic_sizes_)) {
        if (report_) report_->transform_impossible(this, "args not understood");
        DEBUG_PRINTLN("Cannot apply transform: argument sizes not known");
        return false;
    }

    // Criterion: Map cannot contain function calls with side effects (e.g. library nodes that write to memory)
    SideEffectFinder side_effect_finder(sdfg, analysis_manager, this->map_);
    if (side_effect_finder.visit()) {
        if (report_) report_->transform_impossible(this, "side effects");
        DEBUG_PRINTLN("Cannot apply transform: map contains library nodes with side effects");
        return false;
    }

    if (report_) report_->transform_possible(this);
    return true;
}

void OffloadTransform::apply(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {
    // Schedule
    builder.update_schedule_type(this->map_, transformed_schedule_type());

    auto& sdfg = builder.subject();

    // Identify arguments and locals
    auto& arguments_analysis = analysis_manager.get<analysis::ArgumentsAnalysis>();

    auto& arguments = arguments_analysis.arguments(analysis_manager, this->map_);
    auto& locals = arguments_analysis.locals(analysis_manager, this->map_);

    // Infer subsets for arguments
    auto& argument_sizes = arguments_analysis.argument_sizes(analysis_manager, this->map_, allow_dynamic_sizes_);

    auto parent_scope = static_cast<structured_control_flow::Sequence*>(this->map_.get_parent());

    std::string container_prefix = copy_prefix() + std::to_string(parent_scope->element_id()) + "_";

    // Allocate arguments and locals
    allocate_locals_on_device_stack(builder, analysis_manager, locals);
    handle_device_setup_and_teardown(builder, arguments, argument_sizes, container_prefix);

    // Copy-in arguments to device memory & allocation
    for (auto& [argument, meta] : arguments) {
        if (!meta.is_ptr) {
            continue;
        }
        auto argument_device = container_prefix + argument;
        auto& new_block = builder.add_block_before(*parent_scope, this->map_, {}, this->map_.debug_info());
        auto& size = argument_sizes.at(argument);
        copy_to_device_with_allocation(builder, argument, argument_device, size, SymEngine::null, new_block);
    }

    update_map_containers(arguments, container_prefix);

    // Copy-out arguments to host memory & free
    for (auto& [argument, meta] : arguments) {
        if (!meta.is_ptr) {
            continue;
        }
        auto argument_device = container_prefix + argument;
        auto& new_block = builder.add_block_after(*parent_scope, this->map_, {}, this->map_.debug_info());
        auto& size = argument_sizes.at(argument);
        if (!skip_unneeded_d2h_ || meta.is_output) {
            copy_from_device_with_free(builder, new_block, argument, argument_device, size, SymEngine::null);
        } else {
            deallocate_device_arg(builder, new_block, argument_device, size, SymEngine::null);
        }
    }

    if (report_) report_->transform_applied(this);
}

void OffloadTransform::handle_device_setup_and_teardown(
    builder::StructuredSDFGBuilder& builder,

    const std::map<std::string, analysis::RegionArgument>& arguments,
    const std::unordered_map<std::string, symbolic::Expression>& argument_sizes,
    std::string prefix
) {
    // Add managed buffers for pointer arguments
    for (auto& [argument, meta] : arguments) {
        if (!meta.is_ptr || builder.subject().exists(prefix + argument)) {
            continue;
        }
        auto argument_device = prefix + argument;

        auto arg_size = argument_sizes.at(argument);

        add_device_buffer(builder, argument, argument_device, arg_size);
    }
}

void OffloadTransform::
    update_map_containers(const std::map<std::string, analysis::RegionArgument>& arguments, std::string prefix) {
    for (auto& [argument, meta] : arguments) {
        if (meta.is_ptr) {
            auto argument_device = prefix + argument;
            this->map_.replace(symbolic::symbol(argument), symbolic::symbol(argument_device));
        }
    }
}

} // namespace transformations
} // namespace sdfg

1	#include "sdfg/transformations/offloading/offload_transform.h"
2
3	#include <map>
4	#include <string>
5
6	#include "sdfg/analysis/type_analysis.h"
7	#include "sdfg/data_flow/access_node.h"
8	#include "sdfg/structured_control_flow/block.h"
9	#include "sdfg/structured_control_flow/if_else.h"
10	#include "sdfg/structured_control_flow/map.h"
11	#include "sdfg/symbolic/symbolic.h"
12
13	#include "sdfg/data_flow/library_node.h"
14	#include "sdfg/optimization_report/pass_report_consumer.h"
15	#include "sdfg/types/utils.h"
16	#include "sdfg/visitor/immutable_structured_sdfg_visitor.h"
17	#include "symengine/symengine_rcp.h"
18
19	namespace sdfg {
20	namespace transformations {
21
22	class SideEffectFinder : public visitor::ImmutableStructuredSDFGVisitor {
23	private:
24	structured_control_flow::Map& map_;
25
26	public:
27	SideEffectFinder(StructuredSDFG& sdfg, analysis::AnalysisManager& analysis_manager, structured_control_flow::Map& map)
28	: visitor::ImmutableStructuredSDFGVisitor(sdfg, analysis_manager), map_(map) {}	20✔
29
30	bool visit() override { return visit_internal(map_.root()); }	20✔
31
32	bool accept(structured_control_flow::Block& node) override {	21✔
33	for (const auto& lib_node : node.dataflow().library_nodes()) {	21✔
34	if (lib_node->side_effect()) {	×
35	return true;	×
36	}	×
37	}	×
38	return false;	21✔
39	}	21✔
40	};
41
42	OffloadTransform::OffloadTransform(structured_control_flow::Map& map, bool allow_dynamic_sizes)
43	: map_(map), allow_dynamic_sizes_(allow_dynamic_sizes) {}	38✔
44
45
46	bool OffloadTransform::can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	20✔
47	auto& sdfg = builder.subject();	20✔
48
49	auto& arguments_analysis = analysis_manager.get<analysis::ArgumentsAnalysis>();	20✔
50
51	if (!arguments_analysis.inferred_types(analysis_manager, this->map_)) {	20✔
52	if (report_) report_->transform_impossible(this, "unranged args");	×
53	DEBUG_PRINTLN("Cannot apply transform: argument types not inferred");	×
54	return false;	×
55	}	×
56	auto& arguments = arguments_analysis.arguments(analysis_manager, this->map_);	20✔
57
58	// Criterion: arg Data Types must be continuous
59	for (auto& [argument, meta] : arguments) {	41✔
60	auto base_type = analysis::TypeAnalysis(sdfg, &map_, analysis_manager).get_outer_type(argument);	41✔
61	if (base_type == nullptr) {	41✔
62	if (report_) report_->transform_impossible(this, "cannot infer type");	×
63	DEBUG_PRINTLN("Cannot apply transform: argument type cannot be inferred");	×
64	return false;	×
65	}	×
66	if (!types::is_contiguous_type(*base_type, sdfg)) {	41✔
67	if (report_) report_->transform_impossible(this, "type is not contiguous");	×
68	DEBUG_PRINTLN("Cannot apply transform: argument type is not contiguous");	×
69	return false;	×
70	}	×
71	if (meta.is_scalar && meta.is_output) {	41✔
72	if (report_) report_->transform_impossible(this, "scalar output");	×
73	DEBUG_PRINTLN("Cannot apply transform: map writes to scalar argument");	×
74	return false;	×
75	}	×
76	}	41✔
77
78	// Note: arbitrary `init` and `stride` are permitted on the kernel-boundary
79	// Map. The CUDA/ROCm dispatchers emit
80	// `<map.indvar> = init + thread_flat_id * stride`,
81	// and `num_iterations()` already accounts for both when computing the grid
82	// geometry.
83	if (map_.num_iterations().is_null()) {	20✔
NEW 84	if (report_) report_->transform_impossible(this, "cannot determine num iterations");	×
NEW 85	DEBUG_PRINTLN("Cannot apply transform: cannot determine number of iterations for map");	×
86	return false;	×
87	}	×
88
89	// Criterion: Map cannot write to scalar arguments
90	for (auto& [argument, meta] : arguments) {	41✔
91	if (meta.is_scalar && meta.is_output) {	41✔
92	if (report_) report_->transform_impossible(this, "scalar output");	×
93	DEBUG_PRINTLN("Cannot apply transform: map writes to scalar argument");	×
94	return false;	×
95	}	×
96	}	41✔
97
98	if (!arguments_analysis.argument_size_known(analysis_manager, this->map_, allow_dynamic_sizes_)) {	20✔
99	if (report_) report_->transform_impossible(this, "args not understood");	×
100	DEBUG_PRINTLN("Cannot apply transform: argument sizes not known");	×
101	return false;	×
102	}	×
103
104	// Criterion: Map cannot contain function calls with side effects (e.g. library nodes that write to memory)
105	SideEffectFinder side_effect_finder(sdfg, analysis_manager, this->map_);	20✔
106	if (side_effect_finder.visit()) {	20✔
107	if (report_) report_->transform_impossible(this, "side effects");	×
108	DEBUG_PRINTLN("Cannot apply transform: map contains library nodes with side effects");	×
109	return false;	×
110	}	×
111
112	if (report_) report_->transform_possible(this);	20✔
113	return true;	20✔
114	}	20✔
115
116	void OffloadTransform::apply(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	18✔
117	// Schedule
118	builder.update_schedule_type(this->map_, transformed_schedule_type());	18✔
119
120	auto& sdfg = builder.subject();	18✔
121
122	// Identify arguments and locals
123	auto& arguments_analysis = analysis_manager.get<analysis::ArgumentsAnalysis>();	18✔
124
125	auto& arguments = arguments_analysis.arguments(analysis_manager, this->map_);	18✔
126	auto& locals = arguments_analysis.locals(analysis_manager, this->map_);	18✔
127
128	// Infer subsets for arguments
129	auto& argument_sizes = arguments_analysis.argument_sizes(analysis_manager, this->map_, allow_dynamic_sizes_);	18✔
130
131	auto parent_scope = static_cast<structured_control_flow::Sequence*>(this->map_.get_parent());	18✔
132
133	std::string container_prefix = copy_prefix() + std::to_string(parent_scope->element_id()) + "_";	18✔
134
135	// Allocate arguments and locals
136	allocate_locals_on_device_stack(builder, analysis_manager, locals);	18✔
137	handle_device_setup_and_teardown(builder, arguments, argument_sizes, container_prefix);	18✔
138
139	// Copy-in arguments to device memory & allocation
140	for (auto& [argument, meta] : arguments) {	37✔
141	if (!meta.is_ptr) {	37✔
142	continue;	18✔
143	}	18✔
144	auto argument_device = container_prefix + argument;	19✔
145	auto& new_block = builder.add_block_before(*parent_scope, this->map_, {}, this->map_.debug_info());	19✔
146	auto& size = argument_sizes.at(argument);	19✔
147	copy_to_device_with_allocation(builder, argument, argument_device, size, SymEngine::null, new_block);	19✔
148	}	19✔
149
150	update_map_containers(arguments, container_prefix);	18✔
151
152	// Copy-out arguments to host memory & free
153	for (auto& [argument, meta] : arguments) {	37✔
154	if (!meta.is_ptr) {	37✔
155	continue;	18✔
156	}	18✔
157	auto argument_device = container_prefix + argument;	19✔
158	auto& new_block = builder.add_block_after(*parent_scope, this->map_, {}, this->map_.debug_info());	19✔
159	auto& size = argument_sizes.at(argument);	19✔
160	if (!skip_unneeded_d2h_ \|\| meta.is_output) {	19✔
161	copy_from_device_with_free(builder, new_block, argument, argument_device, size, SymEngine::null);	17✔
162	} else {	17✔
163	deallocate_device_arg(builder, new_block, argument_device, size, SymEngine::null);	2✔
164	}	2✔
165	}	19✔
166
167	if (report_) report_->transform_applied(this);	18✔
168	}	18✔
169
170	void OffloadTransform::handle_device_setup_and_teardown(
171	builder::StructuredSDFGBuilder& builder,
172
173	const std::map<std::string, analysis::RegionArgument>& arguments,
174	const std::unordered_map<std::string, symbolic::Expression>& argument_sizes,
175	std::string prefix
176	) {	18✔
177	// Add managed buffers for pointer arguments
178	for (auto& [argument, meta] : arguments) {	37✔
179	if (!meta.is_ptr \|\| builder.subject().exists(prefix + argument)) {	37✔
180	continue;	27✔
181	}	27✔
182	auto argument_device = prefix + argument;	10✔
183
184	auto arg_size = argument_sizes.at(argument);	10✔
185
186	add_device_buffer(builder, argument, argument_device, arg_size);	10✔
187	}	10✔
188	}	18✔
189
190	void OffloadTransform::
191	update_map_containers(const std::map<std::string, analysis::RegionArgument>& arguments, std::string prefix) {	18✔
192	for (auto& [argument, meta] : arguments) {	37✔
193	if (meta.is_ptr) {	37✔
194	auto argument_device = prefix + argument;	19✔
195	this->map_.replace(symbolic::symbol(argument), symbolic::symbol(argument_device));	19✔
196	}	19✔
197	}	37✔
198	}	18✔
199
200	} // namespace transformations
201	} // namespace sdfg

daisytuner / docc / 27704607727

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