29438562417

Committed 15 Jul 2026 05:56PM UTC coverage: 63.155% (-0.02%) from 63.174%

Build # 29438562417

Build Type

Pull #853

github

Committed by

web-flow

Commit Message

Merge c9df0ad93 into 74b3dd197

Pull Request Pull Request #853: Make LoopSchedulers available from DoccTarget

Coverage Stats

22 of 43 new or added lines in 4 files covered. (51.16%)

6 existing lines in 1 file now uncovered.

40628 of 64331 relevant lines covered (63.15%)

980.85 hits per line

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

70.0

/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 {


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());

    // Key the device-buffer names by THIS map's element id so every offloaded loop nest gets its own SSA device
    // buffer. Keying by the parent scope instead would make two maps under the same sequence that offload the same
    // host container resolve to one device name, allocated and freed once per map -- a single device name carrying
    // multiple alloc/free lifetimes. DataTransferMinimization's reuse reconciliation bails on exactly that
    // (a device buffer with more than one free cannot be proven double-free safe), which blocks the D2H->H2D
    // device aliasing and leaves a redundant host round-trip. One buffer per loop nest keeps each name single-alloc
    // single-free, so the reuse fires and DeadDataElimination can drop the now-userless staging container.
    std::string container_prefix = copy_prefix() + std::to_string(this->map_.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));
        }
    }
}

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

bool ::sdfg::transformations::SideEffectFinder::visit() {
    return visitor::ImmutableStructuredSDFGVisitor::visit_internal(map_.StructuredLoop::root());
}

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

} // 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
23	OffloadTransform::OffloadTransform(structured_control_flow::Map& map, bool allow_dynamic_sizes)
24	: map_(map), allow_dynamic_sizes_(allow_dynamic_sizes) {}	38✔
25
26
27	bool OffloadTransform::can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	20✔
28	auto& sdfg = builder.subject();	20✔
29
30	auto& arguments_analysis = analysis_manager.get<analysis::ArgumentsAnalysis>();	20✔
31
32	if (!arguments_analysis.inferred_types(analysis_manager, this->map_)) {	20✔
33	if (report_) report_->transform_impossible(this, "unranged args");	×
34	DEBUG_PRINTLN("Cannot apply transform: argument types not inferred");	×
35	return false;	×
36	}	×
37	auto& arguments = arguments_analysis.arguments(analysis_manager, this->map_);	20✔
38
39	// Criterion: arg Data Types must be continuous
40	for (auto& [argument, meta] : arguments) {	41✔
41	auto base_type = analysis::TypeAnalysis(sdfg, &map_, analysis_manager).get_outer_type(argument);	41✔
42	if (base_type == nullptr) {	41✔
43	if (report_) report_->transform_impossible(this, "cannot infer type");	×
44	DEBUG_PRINTLN("Cannot apply transform: argument type cannot be inferred");	×
45	return false;	×
46	}	×
47	if (!types::is_contiguous_type(*base_type, sdfg)) {	41✔
48	if (report_) report_->transform_impossible(this, "type is not contiguous");	×
49	DEBUG_PRINTLN("Cannot apply transform: argument type is not contiguous");	×
50	return false;	×
51	}	×
52	if (meta.is_scalar && meta.is_output) {	41✔
53	if (report_) report_->transform_impossible(this, "scalar output");	×
54	DEBUG_PRINTLN("Cannot apply transform: map writes to scalar argument");	×
55	return false;	×
56	}	×
57	}	41✔
58
59	// Note: arbitrary `init` and `stride` are permitted on the kernel-boundary
60	// Map. The CUDA/ROCm dispatchers emit
61	// `<map.indvar> = init + thread_flat_id * stride`,
62	// and `num_iterations()` already accounts for both when computing the grid
63	// geometry.
64	if (map_.num_iterations().is_null()) {	20✔
65	if (report_) report_->transform_impossible(this, "cannot determine num iterations");	×
66	DEBUG_PRINTLN("Cannot apply transform: cannot determine number of iterations for map");	×
67	return false;	×
68	}	×
69
70	// Criterion: Map cannot write to scalar arguments
71	for (auto& [argument, meta] : arguments) {	41✔
72	if (meta.is_scalar && meta.is_output) {	41✔
73	if (report_) report_->transform_impossible(this, "scalar output");	×
74	DEBUG_PRINTLN("Cannot apply transform: map writes to scalar argument");	×
75	return false;	×
76	}	×
77	}	41✔
78
79	if (!arguments_analysis.argument_size_known(analysis_manager, this->map_, allow_dynamic_sizes_)) {	20✔
80	if (report_) report_->transform_impossible(this, "args not understood");	×
81	DEBUG_PRINTLN("Cannot apply transform: argument sizes not known");	×
82	return false;	×
83	}	×
84
85	// Criterion: Map cannot contain function calls with side effects (e.g. library nodes that write to memory)
86	SideEffectFinder side_effect_finder(sdfg, analysis_manager, this->map_);	20✔
87	if (side_effect_finder.visit()) {	20✔
88	if (report_) report_->transform_impossible(this, "side effects");	×
89	DEBUG_PRINTLN("Cannot apply transform: map contains library nodes with side effects");	×
90	return false;	×
91	}	×
92
93	if (report_) report_->transform_possible(this);	20✔
94	return true;	20✔
95	}	20✔
96
97	void OffloadTransform::apply(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	18✔
98	// Schedule
99	builder.update_schedule_type(this->map_, transformed_schedule_type());	18✔
100
101	auto& sdfg = builder.subject();	18✔
102
103	// Identify arguments and locals
104	auto& arguments_analysis = analysis_manager.get<analysis::ArgumentsAnalysis>();	18✔
105
106	auto& arguments = arguments_analysis.arguments(analysis_manager, this->map_);	18✔
107	auto& locals = arguments_analysis.locals(analysis_manager, this->map_);	18✔
108
109	// Infer subsets for arguments
110	auto& argument_sizes = arguments_analysis.argument_sizes(analysis_manager, this->map_, allow_dynamic_sizes_);	18✔
111
112	auto parent_scope = static_cast<structured_control_flow::Sequence*>(this->map_.get_parent());	18✔
113
114	// Key the device-buffer names by THIS map's element id so every offloaded loop nest gets its own SSA device
115	// buffer. Keying by the parent scope instead would make two maps under the same sequence that offload the same
116	// host container resolve to one device name, allocated and freed once per map -- a single device name carrying
117	// multiple alloc/free lifetimes. DataTransferMinimization's reuse reconciliation bails on exactly that
118	// (a device buffer with more than one free cannot be proven double-free safe), which blocks the D2H->H2D
119	// device aliasing and leaves a redundant host round-trip. One buffer per loop nest keeps each name single-alloc
120	// single-free, so the reuse fires and DeadDataElimination can drop the now-userless staging container.
121	std::string container_prefix = copy_prefix() + std::to_string(this->map_.element_id()) + "_";	18✔
122
123	// Allocate arguments and locals
124	allocate_locals_on_device_stack(builder, analysis_manager, locals);	18✔
125	handle_device_setup_and_teardown(builder, arguments, argument_sizes, container_prefix);	18✔
126
127	// Copy-in arguments to device memory & allocation
128	for (auto& [argument, meta] : arguments) {	37✔
129	if (!meta.is_ptr) {	37✔
130	continue;	18✔
131	}	18✔
132	auto argument_device = container_prefix + argument;	19✔
133	auto& new_block = builder.add_block_before(*parent_scope, this->map_, {}, this->map_.debug_info());	19✔
134	auto& size = argument_sizes.at(argument);	19✔
135	copy_to_device_with_allocation(builder, argument, argument_device, size, SymEngine::null, new_block);	19✔
136	}	19✔
137
138	update_map_containers(arguments, container_prefix);	18✔
139
140	// Copy-out arguments to host memory & free
141	for (auto& [argument, meta] : arguments) {	37✔
142	if (!meta.is_ptr) {	37✔
143	continue;	18✔
144	}	18✔
145	auto argument_device = container_prefix + argument;	19✔
146	auto& new_block = builder.add_block_after(*parent_scope, this->map_, {}, this->map_.debug_info());	19✔
147	auto& size = argument_sizes.at(argument);	19✔
148	if (!skip_unneeded_d2h_ \|\| meta.is_output) {	19✔
149	copy_from_device_with_free(builder, new_block, argument, argument_device, size, SymEngine::null);	17✔
150	} else {	17✔
151	deallocate_device_arg(builder, new_block, argument_device, size, SymEngine::null);	2✔
152	}	2✔
153	}	19✔
154
155	if (report_) report_->transform_applied(this);	18✔
156	}	18✔
157
158	void OffloadTransform::handle_device_setup_and_teardown(
159	builder::StructuredSDFGBuilder& builder,
160
161	const std::map<std::string, analysis::RegionArgument>& arguments,
162	const std::unordered_map<std::string, symbolic::Expression>& argument_sizes,
163	std::string prefix
164	) {	18✔
165	// Add managed buffers for pointer arguments
166	for (auto& [argument, meta] : arguments) {	37✔
167	if (!meta.is_ptr \|\| builder.subject().exists(prefix + argument)) {	37✔
168	continue;	18✔
169	}	18✔
170	auto argument_device = prefix + argument;	19✔
171
172	auto arg_size = argument_sizes.at(argument);	19✔
173
174	add_device_buffer(builder, argument, argument_device, arg_size);	19✔
175	}	19✔
176	}	18✔
177
178	void OffloadTransform::
179	update_map_containers(const std::map<std::string, analysis::RegionArgument>& arguments, std::string prefix) {	18✔
180	for (auto& [argument, meta] : arguments) {	37✔
181	if (meta.is_ptr) {	37✔
182	auto argument_device = prefix + argument;	19✔
183	this->map_.replace(symbolic::symbol(argument), symbolic::symbol(argument_device));	19✔
184	}	19✔
185	}	37✔
186	}	18✔
187
188	bool ::sdfg::transformations::SideEffectFinder::accept(structured_control_flow::Block& node) {	21✔
189	for (const auto& lib_node : node.dataflow().library_nodes()) {	21✔
NEW 190	if (lib_node->side_effect()) {	×
NEW 191	return true;	×
NEW 192	}	×
NEW 193	}	×
194	return false;	21✔
195	}	21✔
196
197	bool ::sdfg::transformations::SideEffectFinder::visit() {	20✔
198	return visitor::ImmutableStructuredSDFGVisitor::visit_internal(map_.StructuredLoop::root());	20✔
199	}	20✔
200
201	::sdfg::transformations::SideEffectFinder::SideEffectFinder(
202	StructuredSDFG& sdfg, analysis::AnalysisManager& analysis_manager, structured_control_flow::Map& map
203	)
204	: visitor::ImmutableStructuredSDFGVisitor(sdfg, analysis_manager), map_(map) {}	20✔
205
206	} // namespace transformations
207	} // namespace sdfg

daisytuner / docc / 29438562417

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