24102763944

Committed 07 Apr 2026 08:26PM UTC coverage: 64.814% (+0.02%) from 64.797%

Build # 24102763944

Build Type

push

github

Committed by

web-flow

Commit Message

Separate can_be_applied and apply for GPUTilling during Loop Scheduling (#632)

* restructure stride minimization pass

* remove iterative bottleneck and add/fix tests

* loop distribute rework

* Loop distributionfix and unit tests

* Refactor loop scheduler to split the can be applied from the apply to save quadratic analysis recalculation cost.

* enforce can be applied for polly scheduler

* add branching

* guard rpc scheduler

* remove redundant analysis

* remove multi invalidation

* add resnet subset tests

* remove redundant invalidates

* light weight invalidation of scope analysis

* revert stride minimization and perfect loop distribution to main

* update tenstorrent scheduler

* Drop submodels from PR

* extend candidate selection to skip impossible candidates fast

* Stream line kls criterions to reuse them for the scheduler.
Saperating can be applied and apply calls

* remove polly header

* Use new RTX label

Coverage Stats

53 of 67 new or added lines in 2 files covered. (79.1%)

23 existing lines in 4 files now uncovered.

28868 of 44540 relevant lines covered (64.81%)

605.56 hits per line

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

48.94

/opt/src/passes/offloading/gpu_tiling_pass.cpp

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

#include "sdfg/analysis/loop_analysis.h"
#include "sdfg/analysis/users.h"
#include "sdfg/structured_control_flow/structured_loop.h"
#include "sdfg/targets/cuda/cuda.h"
#include "sdfg/targets/rocm/rocm.h"
#include "sdfg/transformations/loop_tiling.h"
#include "sdfg/transformations/offloading/gpu_tiling.h"
#include "sdfg/transformations/offloading/kernel_local_storage.h"

namespace sdfg {
namespace passes {

GPUTilingPass::GPUTilingPass(const std::vector<structured_control_flow::Map*>& maps, size_t tile_size)
    : maps_(maps), tile_size_(tile_size) {}

bool GPUTilingPass::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 tiling targets
    std::vector<structured_control_flow::StructuredLoop*> candidates;
    auto& users = analysis_manager.get<analysis::Users>();

    // Find targets for normal loop tiling
    for (auto* map : maps_) {
        for (auto* descendant : loop_analysis.descendants(map)) {
            auto* struc_loop = dynamic_cast<structured_control_flow::StructuredLoop*>(descendant);
            if (!struc_loop) continue;


            // Pre-filter: check if any container used in the loop is a KLS candidate
            analysis::UsersView users_view(users, struc_loop->root());
            bool has_kls_candidate = false;
            for (const auto& container : users_view.all_containers_in_order()) {
                if (transformations::KernelLocalStorage::is_candidate(*struc_loop, container, builder, analysis_manager)) {
                    has_kls_candidate = true;
                    break;
                }
            }
            if (!has_kls_candidate) continue;

            transformations::LoopTiling tiling(*struc_loop, tile_size_);
            if (tiling.can_be_applied(builder, analysis_manager)) {
                candidates.push_back(struc_loop);
            }
        }
    }

    std::vector<structured_control_flow::StructuredLoop*> tilable_loops;

    for (auto* map : candidates) {
        for (auto* descendant : loop_analysis.descendants(map)) {
            if (auto* target_loop = dynamic_cast<structured_control_flow::StructuredLoop*>(descendant)) {
                transformations::GPUTiling tiling(*target_loop, tile_size_);
                if (tiling.can_be_applied(builder, analysis_manager)) {
                    tilable_loops.push_back(target_loop);
                }
            }
        }
    }

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

    // Phase 2: Apply all tilings
    for (auto* target_loop : tilable_loops) {
        transformations::GPUTiling tiling(*target_loop, tile_size_);
        tiling.apply(builder, analysis_manager);
    }

    return true;
}

} // namespace passes
} // namespace sdfg

1	#include "sdfg/passes/offloading/gpu_tiling_pass.h"
2
3	#include "sdfg/analysis/loop_analysis.h"
4	#include "sdfg/analysis/users.h"
5	#include "sdfg/structured_control_flow/structured_loop.h"
6	#include "sdfg/targets/cuda/cuda.h"
7	#include "sdfg/targets/rocm/rocm.h"
8	#include "sdfg/transformations/loop_tiling.h"
9	#include "sdfg/transformations/offloading/gpu_tiling.h"
10	#include "sdfg/transformations/offloading/kernel_local_storage.h"
11
12	namespace sdfg {
13	namespace passes {
14
15	GPUTilingPass::GPUTilingPass(const std::vector<structured_control_flow::Map*>& maps, size_t tile_size)
16	: maps_(maps), tile_size_(tile_size) {}	3✔
17
18	bool GPUTilingPass::run_pass(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	3✔
19	if (maps_.empty()) {	3✔
20	return false;	×
21	}	×
22
23	auto& loop_analysis = analysis_manager.get<analysis::LoopAnalysis>();	3✔
24
25	// Phase 1: Collect all applicable tiling targets
26	std::vector<structured_control_flow::StructuredLoop*> candidates;	3✔
27	auto& users = analysis_manager.get<analysis::Users>();	3✔
28
29	// Find targets for normal loop tiling
30	for (auto* map : maps_) {	4✔
31	for (auto* descendant : loop_analysis.descendants(map)) {	4✔
32	auto* struc_loop = dynamic_cast<structured_control_flow::StructuredLoop*>(descendant);	2✔
33	if (!struc_loop) continue;	2✔
34
35
36	// Pre-filter: check if any container used in the loop is a KLS candidate
37	analysis::UsersView users_view(users, struc_loop->root());	2✔
38	bool has_kls_candidate = false;	2✔
39	for (const auto& container : users_view.all_containers_in_order()) {	6✔
40	if (transformations::KernelLocalStorage::is_candidate(*struc_loop, container, builder, analysis_manager)) {	6✔
NEW 41	has_kls_candidate = true;	×
NEW 42	break;	×
NEW 43	}	×
44	}	6✔
45	if (!has_kls_candidate) continue;	2✔
46
NEW 47	transformations::LoopTiling tiling(*struc_loop, tile_size_);	×
NEW 48	if (tiling.can_be_applied(builder, analysis_manager)) {	×
NEW 49	candidates.push_back(struc_loop);	×
NEW 50	}	×
NEW 51	}	×
52	}	4✔
53
54	std::vector<structured_control_flow::StructuredLoop*> tilable_loops;	3✔
55
56	for (auto* map : candidates) {	3✔
UNCOV 57	for (auto* descendant : loop_analysis.descendants(map)) {	×
UNCOV 58	if (auto* target_loop = dynamic_cast<structured_control_flow::StructuredLoop*>(descendant)) {	×
UNCOV 59	transformations::GPUTiling tiling(*target_loop, tile_size_);	×
UNCOV 60	if (tiling.can_be_applied(builder, analysis_manager)) {	×
NEW 61	tilable_loops.push_back(target_loop);	×
62	}	×
UNCOV 63	}	×
UNCOV 64	}	×
UNCOV 65	}	×
66
67	if (tilable_loops.empty()) {	3✔
68	return false;	3✔
69	}	3✔
70
71	// Phase 2: Apply all tilings
NEW 72	for (auto* target_loop : tilable_loops) {	×
73	transformations::GPUTiling tiling(*target_loop, tile_size_);	×
74	tiling.apply(builder, analysis_manager);	×
75	}	×
76
77	return true;	×
78	}	3✔
79
80	} // namespace passes
81	} // namespace sdfg

daisytuner / docc / 24102763944

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