27471076166

Committed 13 Jun 2026 03:33PM UTC coverage: 61.254% (-0.02%) from 61.274%

Build # 27471076166

Build Type

Pull #760

github

Committed by

web-flow

Commit Message

Merge d6c6a15bd into db7d71ecc

Pull Request Pull Request #760: Add hardware limit checks for gpu

Coverage Stats

20 of 52 new or added lines in 4 files covered. (38.46%)

13 existing lines in 2 files now uncovered.

36267 of 59208 relevant lines covered (61.25%)

1122.95 hits per line

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

0.0

/opt/src/transformations/offloading/rocm_parallelize_nested_map.cpp

#include "sdfg/transformations/offloading/rocm_parallelize_nested_map.h"

#include <sdfg/analysis/loop_analysis.h>
#include "sdfg/exceptions.h"
#include "sdfg/symbolic/symbolic.h"
#include "sdfg/targets/rocm/rocm.h"

namespace sdfg {
namespace transformations {

ROCMParallelizeNestedMap::ROCMParallelizeNestedMap(structured_control_flow::Map& loop, size_t block_size)
    : loop_(loop), block_size_(block_size) {}

std::string ROCMParallelizeNestedMap::name() const { return "ROCMParallelizeNestedMap"; }

bool ROCMParallelizeNestedMap::
    can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {
    auto& loop_analysis = analysis_manager.get<analysis::LoopAnalysis>();

    // Condition: Check if map is not yet parallelized with ROCM
    if (loop_.schedule_type().value() != ScheduleType_Sequential::value()) {
        return false;
    }

    // Condition: Check if parent loop exists
    auto parent = loop_analysis.parent_loop(&loop_);
    if (parent == nullptr) {
        return false;
    }

    // Condition: Check if parent loop is a ROCM map, and not Z dimension (final dimension)
    if (auto map = dynamic_cast<structured_control_flow::Map*>(parent)) {
        if (map->schedule_type().value() != rocm::ScheduleType_ROCM::value()) {
            return false;
        }
        if (rocm::ScheduleType_ROCM::dimension(map->schedule_type()) == rocm::ROCMDimension::Z) {
            return false;
        }
        auto parent_indvar = map->indvar();
        auto ancestor = parent;
        while (ancestor) {
            if (auto map_ancestor = dynamic_cast<structured_control_flow::Map*>(ancestor)) {
                parent_indvar = map_ancestor->indvar();
                for (auto& arg : symbolic::atoms(loop_.condition())) {
                    if (symbolic::eq(arg, parent_indvar)) {
                        return false;
                    }
                }
            }
            ancestor = loop_analysis.parent_loop(ancestor);
        }
    } else {
        return false;
    }

    // Condition: Check if current loop starts from 0
    if (!symbolic::eq(loop_.init(), symbolic::zero())) {
        return false;
    }

    // Condition: Loop has a stride of 1
    auto stride = loop_.stride();
    if (!symbolic::eq(stride, symbolic::one())) {
        return false;
    }

    // Condition: Resulting ROCm grid dimension must not exceed hardware limits.
    // Y and Z grid dimensions are limited to 65535.
    auto num_iters = loop_.num_iterations();
    if (!num_iters.is_null() && SymEngine::is_a<SymEngine::Integer>(*num_iters)) {
        int64_t iters = SymEngine::down_cast<const SymEngine::Integer&>(*num_iters).as_int();
        int64_t block = static_cast<int64_t>(block_size_);
        int64_t grid_size = (iters + block - 1) / block;

        constexpr int64_t max_grid_dim_yz = 65535;
        if (grid_size > max_grid_dim_yz) {
            return false;
        }
    }

    return true;
}

void ROCMParallelizeNestedMap::apply(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {
    auto& loop_analysis = analysis_manager.get<analysis::LoopAnalysis>();
    auto parent = loop_analysis.parent_loop(&loop_);

    auto parent_dim =
        rocm::ScheduleType_ROCM::dimension(static_cast<structured_control_flow::Map*>(parent)->schedule_type());

    rocm::ROCMDimension child_dim;
    if (parent_dim == rocm::ROCMDimension::X) {
        child_dim = rocm::ROCMDimension::Y;
    } else if (parent_dim == rocm::ROCMDimension::Y) {
        child_dim = rocm::ROCMDimension::Z;
    } else {
        throw InvalidSDFGException("Parent loop is Z dimension, cannot parallelize nested map.");
    }

    auto new_schedule = rocm::ScheduleType_ROCM::create();
    rocm::ScheduleType_ROCM::dimension(new_schedule, child_dim);
    rocm::ScheduleType_ROCM::block_size(new_schedule, symbolic::integer(block_size_));

    builder.update_schedule_type(loop_, new_schedule);
}

void ROCMParallelizeNestedMap::to_json(nlohmann::json& j) const {
    if (dynamic_cast<structured_control_flow::For*>(&loop_)) {
        throw std::runtime_error("ROCMParallelizeNestedMap transformation does not support for-loops.");
    }
    j["transformation_type"] = this->name();

    // Describe the subgraph in a form compatible with EmbeddingRecorder/EmbeddingReplayer.
    // Keep the existing "loop" and "block_size" fields for backward compatibility.
    j["subgraph"] = {{"0", {{"element_id", loop_.element_id()}, {"type", "map"}}}};

    j["parameters"] = {{"block_size", block_size_}};

    j["loop"] = loop_.element_id();
    j["block_size"] = block_size_;
}

ROCMParallelizeNestedMap ROCMParallelizeNestedMap::
    from_json(builder::StructuredSDFGBuilder& builder, const nlohmann::json& j) {
    // Prefer the embedding-compatible representation (subgraph/parameters),
    // but fall back to legacy fields (loop/block_size) if needed.
    size_t loop_id;
    if (j.contains("subgraph")) {
        const auto& subgraph = j.at("subgraph");
        const auto& node_desc = subgraph.at("0");
        loop_id = node_desc.at("element_id").get<size_t>();
    } else {
        loop_id = j.at("loop").get<size_t>();
    }

    size_t block_size;
    if (j.contains("parameters") && j.at("parameters").contains("block_size")) {
        block_size = j.at("parameters").at("block_size").get<size_t>();
    } else {
        block_size = j.at("block_size").get<size_t>();
    }
    auto element = builder.find_element_by_id(loop_id);
    if (!element) {
        throw InvalidTransformationDescriptionException("Element with ID " + std::to_string(loop_id) + " not found.");
    }
    auto loop = dynamic_cast<structured_control_flow::Map*>(element);
    if (!loop) {
        throw InvalidTransformationDescriptionException("Element with ID " + std::to_string(loop_id) + " is not a loop.");
    }
    return ROCMParallelizeNestedMap(*loop, block_size);
}

} // namespace transformations
} // namespace sdfg

1	#include "sdfg/transformations/offloading/rocm_parallelize_nested_map.h"
2
3	#include <sdfg/analysis/loop_analysis.h>
4	#include "sdfg/exceptions.h"
5	#include "sdfg/symbolic/symbolic.h"
6	#include "sdfg/targets/rocm/rocm.h"
7
8	namespace sdfg {
9	namespace transformations {
10
11	ROCMParallelizeNestedMap::ROCMParallelizeNestedMap(structured_control_flow::Map& loop, size_t block_size)
12	: loop_(loop), block_size_(block_size) {}	×
13
14	std::string ROCMParallelizeNestedMap::name() const { return "ROCMParallelizeNestedMap"; }	×
15
16	bool ROCMParallelizeNestedMap::
17	can_be_applied(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	×
18	auto& loop_analysis = analysis_manager.get<analysis::LoopAnalysis>();	×
19
20	// Condition: Check if map is not yet parallelized with ROCM
21	if (loop_.schedule_type().value() != ScheduleType_Sequential::value()) {	×
22	return false;	×
23	}	×
24
25	// Condition: Check if parent loop exists
26	auto parent = loop_analysis.parent_loop(&loop_);	×
27	if (parent == nullptr) {	×
28	return false;	×
29	}	×
30
31	// Condition: Check if parent loop is a ROCM map, and not Z dimension (final dimension)
32	if (auto map = dynamic_cast<structured_control_flow::Map*>(parent)) {	×
33	if (map->schedule_type().value() != rocm::ScheduleType_ROCM::value()) {	×
34	return false;	×
35	}	×
36	if (rocm::ScheduleType_ROCM::dimension(map->schedule_type()) == rocm::ROCMDimension::Z) {	×
37	return false;	×
38	}	×
39	auto parent_indvar = map->indvar();	×
40	auto ancestor = parent;	×
41	while (ancestor) {	×
42	if (auto map_ancestor = dynamic_cast<structured_control_flow::Map*>(ancestor)) {	×
43	parent_indvar = map_ancestor->indvar();	×
44	for (auto& arg : symbolic::atoms(loop_.condition())) {	×
45	if (symbolic::eq(arg, parent_indvar)) {	×
46	return false;	×
47	}	×
48	}	×
49	}	×
50	ancestor = loop_analysis.parent_loop(ancestor);	×
51	}	×
52	} else {	×
53	return false;	×
54	}	×
55
56	// Condition: Check if current loop starts from 0
57	if (!symbolic::eq(loop_.init(), symbolic::zero())) {	×
58	return false;	×
59	}	×
60
61	// Condition: Loop has a stride of 1
62	auto stride = loop_.stride();	×
63	if (!symbolic::eq(stride, symbolic::one())) {	×
64	return false;	×
65	}	×
66
67	// Condition: Resulting ROCm grid dimension must not exceed hardware limits.
68	// Y and Z grid dimensions are limited to 65535.
NEW 69	auto num_iters = loop_.num_iterations();	×
NEW 70	if (!num_iters.is_null() && SymEngine::is_a<SymEngine::Integer>(*num_iters)) {	×
NEW 71	int64_t iters = SymEngine::down_cast<const SymEngine::Integer&>(*num_iters).as_int();	×
NEW 72	int64_t block = static_cast<int64_t>(block_size_);	×
NEW 73	int64_t grid_size = (iters + block - 1) / block;	×
74
NEW 75	constexpr int64_t max_grid_dim_yz = 65535;	×
NEW 76	if (grid_size > max_grid_dim_yz) {	×
NEW 77	return false;	×
NEW 78	}	×
NEW 79	}	×
80
81	return true;	×
82	}	×
83
84	void ROCMParallelizeNestedMap::apply(builder::StructuredSDFGBuilder& builder, analysis::AnalysisManager& analysis_manager) {	×
85	auto& loop_analysis = analysis_manager.get<analysis::LoopAnalysis>();	×
86	auto parent = loop_analysis.parent_loop(&loop_);	×
87
88	auto parent_dim =	×
89	rocm::ScheduleType_ROCM::dimension(static_cast<structured_control_flow::Map*>(parent)->schedule_type());	×
90
91	rocm::ROCMDimension child_dim;	×
92	if (parent_dim == rocm::ROCMDimension::X) {	×
93	child_dim = rocm::ROCMDimension::Y;	×
94	} else if (parent_dim == rocm::ROCMDimension::Y) {	×
95	child_dim = rocm::ROCMDimension::Z;	×
96	} else {	×
97	throw InvalidSDFGException("Parent loop is Z dimension, cannot parallelize nested map.");	×
98	}	×
99
100	auto new_schedule = rocm::ScheduleType_ROCM::create();	×
101	rocm::ScheduleType_ROCM::dimension(new_schedule, child_dim);	×
102	rocm::ScheduleType_ROCM::block_size(new_schedule, symbolic::integer(block_size_));	×
103
104	builder.update_schedule_type(loop_, new_schedule);	×
105	}	×
106
107	void ROCMParallelizeNestedMap::to_json(nlohmann::json& j) const {	×
108	if (dynamic_cast<structured_control_flow::For*>(&loop_)) {	×
109	throw std::runtime_error("ROCMParallelizeNestedMap transformation does not support for-loops.");	×
110	}	×
111	j["transformation_type"] = this->name();	×
112
113	// Describe the subgraph in a form compatible with EmbeddingRecorder/EmbeddingReplayer.
114	// Keep the existing "loop" and "block_size" fields for backward compatibility.
115	j["subgraph"] = {{"0", {{"element_id", loop_.element_id()}, {"type", "map"}}}};	×
116
117	j["parameters"] = {{"block_size", block_size_}};	×
118
119	j["loop"] = loop_.element_id();	×
120	j["block_size"] = block_size_;	×
121	}	×
122
123	ROCMParallelizeNestedMap ROCMParallelizeNestedMap::
124	from_json(builder::StructuredSDFGBuilder& builder, const nlohmann::json& j) {	×
125	// Prefer the embedding-compatible representation (subgraph/parameters),
126	// but fall back to legacy fields (loop/block_size) if needed.
127	size_t loop_id;	×
128	if (j.contains("subgraph")) {	×
129	const auto& subgraph = j.at("subgraph");	×
130	const auto& node_desc = subgraph.at("0");	×
131	loop_id = node_desc.at("element_id").get<size_t>();	×
132	} else {	×
133	loop_id = j.at("loop").get<size_t>();	×
134	}	×
135
136	size_t block_size;	×
137	if (j.contains("parameters") && j.at("parameters").contains("block_size")) {	×
138	block_size = j.at("parameters").at("block_size").get<size_t>();	×
139	} else {	×
140	block_size = j.at("block_size").get<size_t>();	×
141	}	×
142	auto element = builder.find_element_by_id(loop_id);	×
143	if (!element) {	×
144	throw InvalidTransformationDescriptionException("Element with ID " + std::to_string(loop_id) + " not found.");	×
145	}	×
146	auto loop = dynamic_cast<structured_control_flow::Map*>(element);	×
147	if (!loop) {	×
148	throw InvalidTransformationDescriptionException("Element with ID " + std::to_string(loop_id) + " is not a loop.");	×
149	}	×
150	return ROCMParallelizeNestedMap(*loop, block_size);	×
151	}	×
152
153	} // namespace transformations
154	} // namespace sdfg

daisytuner / docc / 27471076166

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