28302975093

Committed 27 Jun 2026 09:55PM UTC coverage: 61.929% (+0.2%) from 61.754%

Build # 28302975093

Build Type

Pull #814

github

Committed by

web-flow

Commit Message

Merge d9618e50c into 8322f5994

Pull Request Pull Request #814: Adds GPU reduce dispatchers

Coverage Stats

567 of 854 new or added lines in 16 files covered. (66.39%)

17 existing lines in 1 file now uncovered.

39450 of 63702 relevant lines covered (61.93%)

968.19 hits per line

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

75.61

/opt/src/passes/offloading/gpu_nested_parallelization_pass.cpp

#include "sdfg/passes/offloading/gpu_nested_parallelization_pass.h"

#include "sdfg/analysis/loop_analysis.h"
#include "sdfg/structured_control_flow/reduce.h"
#include "sdfg/structured_control_flow/structured_loop.h"
#include "sdfg/transformations/offloading/cuda_parallelize_nested_map.h"
#include "sdfg/transformations/offloading/rocm_parallelize_nested_map.h"

namespace sdfg {
namespace passes {

GPUNestedParallelizationPass::GPUNestedParallelizationPass(
    const std::vector<structured_control_flow::Map*>& maps, GPUTarget target, size_t block_size
)
    : maps_(maps), target_(target), block_size_(block_size) {}

bool GPUNestedParallelizationPass::
    run_pass(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {
    if (maps_.empty()) {
        return false;
    }

    auto& loop_analysis = analysis_manager.get<analysis::LoopAnalysis>();

    // Phase 1: Collect all applicable nested loops (maps or reduces)
    std::vector<structured_control_flow::StructuredLoop*> candidates;

    for (auto* map : maps_) {
        auto descendants = loop_analysis.descendants(map);
        for (auto* descendant : descendants) {
            auto* nested_loop = dynamic_cast<structured_control_flow::Map*>(descendant);
            if (nested_loop == nullptr) {
                continue;
            }

            bool applicable = false;
            if (target_ == GPUTarget::CUDA) {
                transformations::CUDAParallelizeNestedMap transform(*nested_loop, block_size_);
                applicable = transform.can_be_applied(builder, analysis_manager);
            } else {
                transformations::ROCMParallelizeNestedMap transform(*nested_loop, block_size_);
                applicable = transform.can_be_applied(builder, analysis_manager);
            }
            if (applicable) {
                candidates.push_back(nested_loop);
            }
        }
    }

    if (candidates.empty()) {
        return false;
    }

    // Phase 2: Apply all parallelizations
    for (auto* nested_loop : candidates) {
        if (target_ == GPUTarget::CUDA) {
            transformations::CUDAParallelizeNestedMap transform(*nested_loop, block_size_);
            transform.apply(builder, analysis_manager);
        } else {
            transformations::ROCMParallelizeNestedMap transform(*nested_loop, block_size_);
            transform.apply(builder, analysis_manager);
        }
    }

    return true;
}

} // namespace passes
} // namespace sdfg

1	#include "sdfg/passes/offloading/gpu_nested_parallelization_pass.h"
2
3	#include "sdfg/analysis/loop_analysis.h"
4	#include "sdfg/structured_control_flow/reduce.h"
5	#include "sdfg/structured_control_flow/structured_loop.h"
6	#include "sdfg/transformations/offloading/cuda_parallelize_nested_map.h"
7	#include "sdfg/transformations/offloading/rocm_parallelize_nested_map.h"
8
9	namespace sdfg {
10	namespace passes {
11
12	GPUNestedParallelizationPass::GPUNestedParallelizationPass(
13	const std::vector<structured_control_flow::Map*>& maps, GPUTarget target, size_t block_size
14	)
15	: maps_(maps), target_(target), block_size_(block_size) {}	6✔
16
17	bool GPUNestedParallelizationPass::
18	run_pass(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	6✔
19	if (maps_.empty()) {	6✔
20	return false;	×
21	}	×
22
23	auto& loop_analysis = analysis_manager.get<analysis::LoopAnalysis>();	6✔
24
25	// Phase 1: Collect all applicable nested loops (maps or reduces)
26	std::vector<structured_control_flow::StructuredLoop*> candidates;	6✔
27
28	for (auto* map : maps_) {	16✔
29	auto descendants = loop_analysis.descendants(map);	16✔
30	for (auto* descendant : descendants) {	16✔
31	auto* nested_loop = dynamic_cast<structured_control_flow::Map*>(descendant);	2✔
32	if (nested_loop == nullptr) {	2✔
NEW 33	continue;	×
NEW 34	}	×
35
36	bool applicable = false;	2✔
37	if (target_ == GPUTarget::CUDA) {	2✔
38	transformations::CUDAParallelizeNestedMap transform(*nested_loop, block_size_);	2✔
39	applicable = transform.can_be_applied(builder, analysis_manager);	2✔
40	} else {	2✔
NEW 41	transformations::ROCMParallelizeNestedMap transform(*nested_loop, block_size_);	×
NEW 42	applicable = transform.can_be_applied(builder, analysis_manager);	×
NEW 43	}	×
44	if (applicable) {	2✔
45	candidates.push_back(nested_loop);	2✔
46	}	2✔
47	}	2✔
48	}	16✔
49
50	if (candidates.empty()) {	6✔
51	return false;	4✔
52	}	4✔
53
54	// Phase 2: Apply all parallelizations
55	for (auto* nested_loop : candidates) {	2✔
56	if (target_ == GPUTarget::CUDA) {	2✔
57	transformations::CUDAParallelizeNestedMap transform(*nested_loop, block_size_);	2✔
58	transform.apply(builder, analysis_manager);	2✔
59	} else {	2✔
NEW 60	transformations::ROCMParallelizeNestedMap transform(*nested_loop, block_size_);	×
61	transform.apply(builder, analysis_manager);	×
62	}	×
63	}	2✔
64
65	return true;	2✔
66	}	6✔
67
68	} // namespace passes
69	} // namespace sdfg

daisytuner / docc / 28302975093

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