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

79.49

/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::StructuredLoop*>& loops, GPUTarget target, size_t block_size
)
    : loops_(loops), target_(target), block_size_(block_size) {}

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

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

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

    for (auto* loop : loops_) {
        auto descendants = loop_analysis.descendants(loop);
        for (auto* descendant : descendants) {
            if (auto* nested_loop = dyn_cast<structured_control_flow::StructuredLoop*>(descendant)) {
                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::StructuredLoop*>& loops, GPUTarget target, size_t block_size
14	)
15	: loops_(loops), target_(target), block_size_(block_size) {}	6✔
16
17	bool GPUNestedParallelizationPass::
18	run_pass(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	6✔
19	if (loops_.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 (loops or reduces)
26	std::vector<structured_control_flow::StructuredLoop*> candidates;	6✔
27
28	for (auto* loop : loops_) {	16✔
29	auto descendants = loop_analysis.descendants(loop);	16✔
30	for (auto* descendant : descendants) {	16✔
31	if (auto* nested_loop = dyn_cast<structured_control_flow::StructuredLoop*>(descendant)) {	2✔
32	bool applicable = false;	2✔
33	if (target_ == GPUTarget::CUDA) {	2✔
34	transformations::CUDAParallelizeNestedMap transform(*nested_loop, block_size_);	2✔
35	applicable = transform.can_be_applied(builder, analysis_manager);	2✔
36	} else {	2✔
NEW 37	transformations::ROCMParallelizeNestedMap transform(*nested_loop, block_size_);	×
38	applicable = transform.can_be_applied(builder, analysis_manager);	×
39	}	×
40	if (applicable) {	2✔
41	candidates.push_back(nested_loop);	2✔
42	}	2✔
43	}	2✔
44	}	2✔
45	}	16✔
46
47	if (candidates.empty()) {	6✔
48	return false;	4✔
49	}	4✔
50
51	// Phase 2: Apply all parallelizations
52	for (auto* nested_loop : candidates) {	2✔
53	if (target_ == GPUTarget::CUDA) {	2✔
54	transformations::CUDAParallelizeNestedMap transform(*nested_loop, block_size_);	2✔
55	transform.apply(builder, analysis_manager);	2✔
56	} else {	2✔
NEW 57	transformations::ROCMParallelizeNestedMap transform(*nested_loop, block_size_);	×
58	transform.apply(builder, analysis_manager);	×
59	}	×
60	}	2✔
61
62	return true;	2✔
63	}	6✔
64
65	} // namespace passes
66	} // 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