31009380971

Committed 05 Aug 2026 01:15PM UTC coverage: 65.103% (+0.1%) from 65.005%

Build # 31009380971

Build Type

Pull #814

github

Committed by

web-flow

Commit Message

Merge 268a080ae into 7d5b198bd

Pull Request Pull Request #814: Adds GPU reduce dispatchers

Coverage Stats

409 of 663 new or added lines in 18 files covered. (61.69%)

145 existing lines in 6 files now uncovered.

46693 of 71722 relevant lines covered (65.1%)

713.1 hits per line

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

67.44

/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::StructuredLoop& loop, bool allow_dynamic_sizes)
    : loop_(loop), allow_dynamic_sizes_(allow_dynamic_sizes) {}


bool OffloadTransform::can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {
    if (dynamic_cast<structured_control_flow::Map*>(&loop_) == nullptr &&
        dynamic_cast<structured_control_flow::Reduce*>(&loop_) == nullptr) {
        return false;
    }

    if (dynamic_cast<structured_control_flow::Reduce*>(&loop_)) {
        if (report_) report_->transform_impossible(this, "reduce");
        DEBUG_PRINTLN("Cannot apply transform: Reduce nodes are not offloaded yet");
        return false;
    }

    auto& sdfg = builder.subject();

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

    if (!arguments_analysis.inferred_types(analysis_manager, this->loop_)) {
        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->loop_);

    // Criterion: arg Data Types must be continuous
    for (auto& [argument, meta] : arguments) {
        auto base_type = analysis::TypeAnalysis(sdfg, &loop_, 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 (loop_.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->loop_, 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->loop_);
    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->loop_, 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->loop_);
    auto& locals = arguments_analysis.locals(analysis_manager, this->loop_);

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

    auto parent_scope = static_cast<structured_control_flow::Sequence*>(this->loop_.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->loop_.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->loop_, this->loop_.debug_info());
        auto& size = argument_sizes.at(argument);
        copy_to_device_with_allocation(builder, argument, argument_device, size, SymEngine::null, new_block);
    }

    update_loop_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->loop_, this->loop_.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_loop_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->loop_.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(loop_.StructuredLoop::root());
}

::sdfg::transformations::SideEffectFinder::SideEffectFinder(
    StructuredSDFG& sdfg, analysis::AnalysisManager& analysis_manager, structured_control_flow::StructuredLoop& loop
)
    : visitor::ImmutableStructuredSDFGVisitor(sdfg, analysis_manager), loop_(loop) {}

} // 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::StructuredLoop& loop, bool allow_dynamic_sizes)
24	: loop_(loop), allow_dynamic_sizes_(allow_dynamic_sizes) {}	38✔
25
26
27	bool OffloadTransform::can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	20✔
28	if (dynamic_cast<structured_control_flow::Map*>(&loop_) == nullptr &&	20✔
29	dynamic_cast<structured_control_flow::Reduce*>(&loop_) == nullptr) {	20✔
NEW 30	return false;	×
NEW 31	}	×
32
33	if (dynamic_cast<structured_control_flow::Reduce*>(&loop_)) {	20✔
NEW 34	if (report_) report_->transform_impossible(this, "reduce");	×
NEW 35	DEBUG_PRINTLN("Cannot apply transform: Reduce nodes are not offloaded yet");	×
NEW 36	return false;	×
NEW 37	}	×
38
39	auto& sdfg = builder.subject();	20✔
40
41	auto& arguments_analysis = analysis_manager.get<analysis::ArgumentsAnalysis>();	20✔
42
43	if (!arguments_analysis.inferred_types(analysis_manager, this->loop_)) {	20✔
44	if (report_) report_->transform_impossible(this, "unranged args");	×
45	DEBUG_PRINTLN("Cannot apply transform: argument types not inferred");	×
46	return false;	×
47	}	×
48	auto& arguments = arguments_analysis.arguments(analysis_manager, this->loop_);	20✔
49
50	// Criterion: arg Data Types must be continuous
51	for (auto& [argument, meta] : arguments) {	41✔
52	auto base_type = analysis::TypeAnalysis(sdfg, &loop_, analysis_manager).get_outer_type(argument);	41✔
53	if (base_type == nullptr) {	41✔
54	if (report_) report_->transform_impossible(this, "cannot infer type");	×
55	DEBUG_PRINTLN("Cannot apply transform: argument type cannot be inferred");	×
56	return false;	×
57	}	×
58	if (!types::is_contiguous_type(*base_type, sdfg)) {	41✔
59	if (report_) report_->transform_impossible(this, "type is not contiguous");	×
60	DEBUG_PRINTLN("Cannot apply transform: argument type is not contiguous");	×
61	return false;	×
62	}	×
63	if (meta.is_scalar && meta.is_output) {	41✔
64	if (report_) report_->transform_impossible(this, "scalar output");	×
65	DEBUG_PRINTLN("Cannot apply transform: map writes to scalar argument");	×
66	return false;	×
67	}	×
68	}	41✔
69
70	// Note: arbitrary `init` and `stride` are permitted on the kernel-boundary
71	// Map. The CUDA/ROCm dispatchers emit
72	// `<map.indvar> = init + thread_flat_id * stride`,
73	// and `num_iterations()` already accounts for both when computing the grid
74	// geometry.
75	if (loop_.num_iterations().is_null()) {	20✔
76	if (report_) report_->transform_impossible(this, "cannot determine num iterations");	×
77	DEBUG_PRINTLN("Cannot apply transform: cannot determine number of iterations for map");	×
78	return false;	×
79	}	×
80
81	// Criterion: Map cannot write to scalar arguments
82	for (auto& [argument, meta] : arguments) {	41✔
83	if (meta.is_scalar && meta.is_output) {	41✔
84	if (report_) report_->transform_impossible(this, "scalar output");	×
85	DEBUG_PRINTLN("Cannot apply transform: map writes to scalar argument");	×
86	return false;	×
87	}	×
88	}	41✔
89
90	if (!arguments_analysis.argument_size_known(analysis_manager, this->loop_, allow_dynamic_sizes_)) {	20✔
91	if (report_) report_->transform_impossible(this, "args not understood");	×
92	DEBUG_PRINTLN("Cannot apply transform: argument sizes not known");	×
93	return false;	×
94	}	×
95
96	// Criterion: Map cannot contain function calls with side effects (e.g. library nodes that write to memory)
97	SideEffectFinder side_effect_finder(sdfg, analysis_manager, this->loop_);	20✔
98	if (side_effect_finder.visit()) {	20✔
99	if (report_) report_->transform_impossible(this, "side effects");	×
100	DEBUG_PRINTLN("Cannot apply transform: map contains library nodes with side effects");	×
101	return false;	×
102	}	×
103
104	if (report_) report_->transform_possible(this);	20✔
105	return true;	20✔
106	}	20✔
107
108	void OffloadTransform::apply(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	18✔
109	// Schedule
110	builder.update_schedule_type(this->loop_, transformed_schedule_type());	18✔
111
112	auto& sdfg = builder.subject();	18✔
113
114	// Identify arguments and locals
115	auto& arguments_analysis = analysis_manager.get<analysis::ArgumentsAnalysis>();	18✔
116
117	auto& arguments = arguments_analysis.arguments(analysis_manager, this->loop_);	18✔
118	auto& locals = arguments_analysis.locals(analysis_manager, this->loop_);	18✔
119
120	// Infer subsets for arguments
121	auto& argument_sizes = arguments_analysis.argument_sizes(analysis_manager, this->loop_, allow_dynamic_sizes_);	18✔
122
123	auto parent_scope = static_cast<structured_control_flow::Sequence*>(this->loop_.get_parent());	18✔
124
125	// Key the device-buffer names by THIS map's element id so every offloaded loop nest gets its own SSA device
126	// buffer. Keying by the parent scope instead would make two maps under the same sequence that offload the same
127	// host container resolve to one device name, allocated and freed once per map -- a single device name carrying
128	// multiple alloc/free lifetimes. DataTransferMinimization's reuse reconciliation bails on exactly that
129	// (a device buffer with more than one free cannot be proven double-free safe), which blocks the D2H->H2D
130	// device aliasing and leaves a redundant host round-trip. One buffer per loop nest keeps each name single-alloc
131	// single-free, so the reuse fires and DeadDataElimination can drop the now-userless staging container.
132	std::string container_prefix = copy_prefix() + std::to_string(this->loop_.element_id()) + "_";	18✔
133
134	// Allocate arguments and locals
135	allocate_locals_on_device_stack(builder, analysis_manager, locals);	18✔
136	handle_device_setup_and_teardown(builder, arguments, argument_sizes, container_prefix);	18✔
137
138	// Copy-in arguments to device memory & allocation
139	for (auto& [argument, meta] : arguments) {	37✔
140	if (!meta.is_ptr) {	37✔
141	continue;	18✔
142	}	18✔
143	auto argument_device = container_prefix + argument;	19✔
144	auto& new_block = builder.add_block_before(*parent_scope, this->loop_, this->loop_.debug_info());	19✔
145	auto& size = argument_sizes.at(argument);	19✔
146	copy_to_device_with_allocation(builder, argument, argument_device, size, SymEngine::null, new_block);	19✔
147	}	19✔
148
149	update_loop_containers(arguments, container_prefix);	18✔
150
151	// Copy-out arguments to host memory & free
152	for (auto& [argument, meta] : arguments) {	37✔
153	if (!meta.is_ptr) {	37✔
154	continue;	18✔
155	}	18✔
156	auto argument_device = container_prefix + argument;	19✔
157	auto& new_block = builder.add_block_after(*parent_scope, this->loop_, this->loop_.debug_info());	19✔
158	auto& size = argument_sizes.at(argument);	19✔
159	if (!skip_unneeded_d2h_ \|\| meta.is_output) {	19✔
160	copy_from_device_with_free(builder, new_block, argument, argument_device, size, SymEngine::null);	17✔
161	} else {	17✔
162	deallocate_device_arg(builder, new_block, argument_device, size, SymEngine::null);	2✔
163	}	2✔
164	}	19✔
165
166	if (report_) report_->transform_applied(this);	18✔
167	}	18✔
168
169	void OffloadTransform::handle_device_setup_and_teardown(
170	builder::StructuredSDFGBuilder& builder,
171
172	const std::map<std::string, analysis::RegionArgument>& arguments,
173	const std::unordered_map<std::string, symbolic::Expression>& argument_sizes,
174	std::string prefix
175	) {	18✔
176	// Add managed buffers for pointer arguments
177	for (auto& [argument, meta] : arguments) {	37✔
178	if (!meta.is_ptr \|\| builder.subject().exists(prefix + argument)) {	37✔
179	continue;	18✔
180	}	18✔
181	auto argument_device = prefix + argument;	19✔
182
183	auto arg_size = argument_sizes.at(argument);	19✔
184
185	add_device_buffer(builder, argument, argument_device, arg_size);	19✔
186	}	19✔
187	}	18✔
188
189	void OffloadTransform::
190	update_loop_containers(const std::map<std::string, analysis::RegionArgument>& arguments, std::string prefix) {	18✔
191	for (auto& [argument, meta] : arguments) {	37✔
192	if (meta.is_ptr) {	37✔
193	auto argument_device = prefix + argument;	19✔
194	this->loop_.replace(symbolic::symbol(argument), symbolic::symbol(argument_device));	19✔
195	}	19✔
196	}	37✔
197	}	18✔
198
199	bool ::sdfg::transformations::SideEffectFinder::accept(structured_control_flow::Block& node) {	21✔
200	for (const auto& lib_node : node.dataflow().library_nodes()) {	21✔
201	if (lib_node->side_effect()) {	×
202	return true;	×
203	}	×
204	}	×
205	return false;	21✔
206	}	21✔
207
208	bool ::sdfg::transformations::SideEffectFinder::visit() {	20✔
209	return visitor::ImmutableStructuredSDFGVisitor::visit_internal(loop_.StructuredLoop::root());	20✔
210	}	20✔
211
212	::sdfg::transformations::SideEffectFinder::SideEffectFinder(
213	StructuredSDFG& sdfg, analysis::AnalysisManager& analysis_manager, structured_control_flow::StructuredLoop& loop
214	)
215	: visitor::ImmutableStructuredSDFGVisitor(sdfg, analysis_manager), loop_(loop) {}	20✔
216
217	} // namespace transformations
218	} // namespace sdfg

daisytuner / docc / 31009380971

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