17300165647

Committed 28 Aug 2025 03:14PM UTC coverage: 60.049% (+0.3%) from 59.781%

Build # 17300165647

Build Type

Pull #210

github

Committed by

web-flow

Commit Message

Merge f6109d03a into 18d34db1e

Pull Request Pull Request #210: New debug info

Run Details

377 of 593 new or added lines in 37 files covered. (63.58%)

15 existing lines in 8 files now uncovered.

9588 of 15967 relevant lines covered (60.05%)

114.92 hits per line

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

0.0

/src/data_flow/library_nodes/math/ml/softmax.cpp

#include "sdfg/data_flow/library_nodes/math/ml/softmax.h"

#include "sdfg/analysis/analysis.h"
#include "sdfg/analysis/scope_analysis.h"
#include "sdfg/builder/structured_sdfg_builder.h"

namespace sdfg {
namespace math {
namespace ml {

SoftmaxNode::SoftmaxNode(
    size_t element_id,
    const DebugInfoRegion &debug_info,
    const graph::Vertex vertex,
    data_flow::DataFlowGraph &parent,
    int axis
)
    : MathNode(
          element_id,
          debug_info,
          vertex,
          parent,
          LibraryNodeType_Softmax,
          {"output"},
          {"input"},
          data_flow::ImplementationType_NONE
      ),
      axis_(axis) {}

void SoftmaxNode::validate(const Function &) const { /* TODO */ }

bool SoftmaxNode::expand(builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager) {
    auto &dataflow = this->get_parent();
    auto &block = static_cast<structured_control_flow::Block &>(*dataflow.get_parent());

    auto &scope_analysis = analysis_manager.get<analysis::ScopeAnalysis>();
    auto &parent = static_cast<structured_control_flow::Sequence &>(*scope_analysis.parent_scope(&block));

    // Locate edges
    const data_flow::Memlet *iedge_input = nullptr;
    const data_flow::Memlet *oedge_output = nullptr;
    for (const auto &edge : dataflow.in_edges(*this)) {
        if (edge.dst_conn() == "input") {
            iedge_input = &edge;
        }
    }
    for (const auto &edge : dataflow.out_edges(*this)) {
        if (edge.src_conn() == "output") {
            oedge_output = &edge;
        }
    }
    if (!iedge_input || !oedge_output) return false;

    std::string input_name = static_cast<const data_flow::AccessNode &>(iedge_input->src()).data();
    std::string output_name = static_cast<const data_flow::AccessNode &>(oedge_output->dst()).data();

    // Create new sequence before
    auto &new_sequence = builder.add_sequence_before(parent, block, block.debug_info()).first;
    structured_control_flow::Sequence *last_scope = &new_sequence;

    // Create maps over output subset dims (parallel dims)
    data_flow::Subset domain_begin = oedge_output->begin_subset();
    data_flow::Subset domain_end = oedge_output->end_subset();

    std::vector<symbolic::Expression> loop_syms;
    structured_control_flow::Map *last_map = nullptr;
    for (size_t d = 0; d < domain_begin.size(); ++d) {
        std::string indvar_str = builder.find_new_name("_i");
        builder.add_container(indvar_str, types::Scalar(types::PrimitiveType::UInt64));
        auto indvar = symbolic::symbol(indvar_str);
        auto init = domain_begin[d];
        auto update = symbolic::add(indvar, symbolic::one());
        auto cond = symbolic::Lt(indvar, symbolic::add(domain_end[d], symbolic::one()));
        last_map = &builder.add_map(
            *last_scope,
            indvar,
            cond,
            init,
            update,
            structured_control_flow::ScheduleType_Sequential,
            {},
            block.debug_info()
        );
        last_scope = &last_map->root();
        loop_syms.push_back(indvar);
    }

    // Initialize temp variable to zero
    std::string temp_name = builder.find_new_name("_tmp");
    std::string temp_name2 = builder.find_new_name("_tmp");
    types::Scalar temp_type(types::PrimitiveType::Float);
    builder.add_container(temp_name, temp_type);
    builder.add_container(temp_name2, temp_type);

    auto &init_block = builder.add_block(*last_scope);
    auto &init_tasklet = builder.add_tasklet(init_block, data_flow::TaskletCode::assign, "_out", {"0.0f"});
    auto &tmp_access_init = builder.add_access(init_block, temp_name);
    builder.add_computational_memlet(init_block, init_tasklet, "_out", tmp_access_init, {}, temp_type);

    // add reduction for loop
    symbolic::Expression red_begin;
    symbolic::Expression red_end;
    if (axis_ >= 0) {
        red_begin = iedge_input->begin_subset()[axis_];
        red_end = iedge_input->end_subset()[axis_];
    } else {
        red_begin = iedge_input->begin_subset().back();
        red_end = iedge_input->end_subset().back();
    }
    std::string red_name = builder.find_new_name("_i");
    builder.add_container(red_name, types::Scalar(types::PrimitiveType::UInt64));
    auto red_indvar = symbolic::symbol(red_name);
    auto red_init = red_begin;
    auto red_update = symbolic::add(red_indvar, symbolic::one());
    auto red_cond = symbolic::Lt(red_indvar, symbolic::add(red_end, symbolic::one()));
    auto red_map = &builder.add_for(*last_scope, red_indvar, red_cond, red_init, red_update, {}, block.debug_info());

    // Create innermost block
    auto &code_block = builder.add_block(red_map->root());

    // Create access nodes for input and output
    auto &input_access = builder.add_access(code_block, input_name);
    auto &tmp2_access = builder.add_access(code_block, temp_name2);
    auto &tmp_access_out = builder.add_access(code_block, temp_name);
    auto &tmp_access_in = builder.add_access(code_block, temp_name2);

    // Create index expressions for input and output
    std::vector<symbolic::Expression> input_subset = loop_syms;

    // Replace the reduction axis index with the reduction variable for input
    if (axis_ >= 0 && axis_ < static_cast<int>(input_subset.size())) {
        input_subset.insert(input_subset.begin() + axis_, red_indvar);
    } else if (axis_ < 0) {
        input_subset.push_back(red_indvar);
    }

    // Compute exponential
    auto &exp_tasklet = builder.add_tasklet(code_block, data_flow::TaskletCode::expf, "_out", {"_in"});
    builder
        .add_computational_memlet(code_block, input_access, exp_tasklet, "_in", input_subset, iedge_input->base_type());
    builder.add_computational_memlet(code_block, exp_tasklet, "_out", tmp2_access, {}, temp_type);

    // Add to temp (reduction)
    auto &add_tasklet = builder.add_tasklet(code_block, data_flow::TaskletCode::add, "_out", {"_in1", "_in2"});
    builder.add_computational_memlet(code_block, tmp_access_in, add_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(code_block, tmp2_access, add_tasklet, "_in2", {}, temp_type);
    builder.add_computational_memlet(code_block, add_tasklet, "_out", tmp_access_out, {}, temp_type);

    // Create writeback - assign the accumulated sum to output
    auto &writeback_block = builder.add_block(*last_scope);
    auto &tmp_access_wb = builder.add_access(writeback_block, temp_name);
    auto &output_access_wb = builder.add_access(writeback_block, output_name);
    auto &writeback_tasklet = builder.add_tasklet(writeback_block, data_flow::TaskletCode::assign, "_out", {"_in"});
    builder.add_computational_memlet(writeback_block, tmp_access_wb, writeback_tasklet, "_in", {}, temp_type);
    builder.add_computational_memlet(
        writeback_block, writeback_tasklet, "_out", output_access_wb, loop_syms, oedge_output->base_type()
    );

    // Cleanup old block
    builder.remove_memlet(block, *iedge_input);
    builder.remove_memlet(block, *oedge_output);
    builder.remove_node(block, *this);
    builder.remove_child(parent, block);

    return true;
}

std::unique_ptr<data_flow::DataFlowNode> SoftmaxNode::
    clone(size_t element_id, const graph::Vertex vertex, data_flow::DataFlowGraph &parent) const {
    return std::unique_ptr<data_flow::DataFlowNode>(new SoftmaxNode(element_id, this->debug_info(), vertex, parent, axis_)
    );
}

nlohmann::json SoftmaxNodeSerializer::serialize(const data_flow::LibraryNode &library_node) {
    const SoftmaxNode &node = static_cast<const SoftmaxNode &>(library_node);
    nlohmann::json j;

    j["code"] = node.code().value();
    j["axis"] = node.axis();

    return j;
}

data_flow::LibraryNode &SoftmaxNodeSerializer::deserialize(
    const nlohmann::json &j, builder::StructuredSDFGBuilder &builder, structured_control_flow::Block &parent
) {
    auto code = j["code"].get<std::string>();
    if (code != LibraryNodeType_Softmax.value()) {
        throw std::runtime_error("Invalid library node code");
    }

    sdfg::serializer::JSONSerializer serializer;
    DebugInfoRegion debug_info = serializer.json_to_debug_info_region(j["debug_info_region"], builder.debug_info());

    auto axis = j["axis"].get<int>();

    return builder.add_library_node<SoftmaxNode>(parent, debug_info, axis);
}

} // namespace ml
} // namespace math
} // namespace sdfg

1	#include "sdfg/data_flow/library_nodes/math/ml/softmax.h"
2
3	#include "sdfg/analysis/analysis.h"
4	#include "sdfg/analysis/scope_analysis.h"
5	#include "sdfg/builder/structured_sdfg_builder.h"
6
7	namespace sdfg {
8	namespace math {
9	namespace ml {
10
11	SoftmaxNode::SoftmaxNode(	×
12	size_t element_id,
13	const DebugInfoRegion &debug_info,
14	const graph::Vertex vertex,
15	data_flow::DataFlowGraph &parent,
16	int axis
17	)
18	: MathNode(	×
NEW 19	element_id,	×
NEW 20	debug_info,	×
NEW 21	vertex,	×
NEW 22	parent,	×
23	LibraryNodeType_Softmax,
NEW 24	{"output"},	×
NEW 25	{"input"},	×
26	data_flow::ImplementationType_NONE
27	),
28	axis_(axis) {}	×
29
30	void SoftmaxNode::validate(const Function &) const { /* TODO */ }	×
31
32	bool SoftmaxNode::expand(builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager) {	×
33	auto &dataflow = this->get_parent();	×
34	auto &block = static_cast<structured_control_flow::Block &>(*dataflow.get_parent());	×
35
36	auto &scope_analysis = analysis_manager.get<analysis::ScopeAnalysis>();	×
37	auto &parent = static_cast<structured_control_flow::Sequence &>(*scope_analysis.parent_scope(&block));	×
38
39	// Locate edges
40	const data_flow::Memlet *iedge_input = nullptr;	×
41	const data_flow::Memlet *oedge_output = nullptr;	×
42	for (const auto &edge : dataflow.in_edges(*this)) {	×
43	if (edge.dst_conn() == "input") {	×
44	iedge_input = &edge;	×
45	}	×
46	}
47	for (const auto &edge : dataflow.out_edges(*this)) {	×
48	if (edge.src_conn() == "output") {	×
49	oedge_output = &edge;	×
50	}	×
51	}
52	if (!iedge_input \|\| !oedge_output) return false;	×
53
54	std::string input_name = static_cast<const data_flow::AccessNode &>(iedge_input->src()).data();	×
55	std::string output_name = static_cast<const data_flow::AccessNode &>(oedge_output->dst()).data();	×
56
57	// Create new sequence before
58	auto &new_sequence = builder.add_sequence_before(parent, block, block.debug_info()).first;	×
59	structured_control_flow::Sequence *last_scope = &new_sequence;	×
60
61	// Create maps over output subset dims (parallel dims)
62	data_flow::Subset domain_begin = oedge_output->begin_subset();	×
63	data_flow::Subset domain_end = oedge_output->end_subset();	×
64
65	std::vector<symbolic::Expression> loop_syms;	×
66	structured_control_flow::Map *last_map = nullptr;	×
67	for (size_t d = 0; d < domain_begin.size(); ++d) {	×
68	std::string indvar_str = builder.find_new_name("_i");	×
69	builder.add_container(indvar_str, types::Scalar(types::PrimitiveType::UInt64));	×
70	auto indvar = symbolic::symbol(indvar_str);	×
71	auto init = domain_begin[d];	×
72	auto update = symbolic::add(indvar, symbolic::one());	×
73	auto cond = symbolic::Lt(indvar, symbolic::add(domain_end[d], symbolic::one()));	×
74	last_map = &builder.add_map(	×
75	*last_scope,	×
76	indvar,
77	cond,
78	init,
79	update,
80	structured_control_flow::ScheduleType_Sequential,
81	{},	×
82	block.debug_info()	×
83	);
84	last_scope = &last_map->root();	×
85	loop_syms.push_back(indvar);	×
86	}	×
87
88	// Initialize temp variable to zero
89	std::string temp_name = builder.find_new_name("_tmp");	×
90	std::string temp_name2 = builder.find_new_name("_tmp");	×
91	types::Scalar temp_type(types::PrimitiveType::Float);	×
92	builder.add_container(temp_name, temp_type);	×
93	builder.add_container(temp_name2, temp_type);	×
94
95	auto &init_block = builder.add_block(*last_scope);	×
96	auto &init_tasklet = builder.add_tasklet(init_block, data_flow::TaskletCode::assign, "_out", {"0.0f"});	×
97	auto &tmp_access_init = builder.add_access(init_block, temp_name);	×
98	builder.add_computational_memlet(init_block, init_tasklet, "_out", tmp_access_init, {}, temp_type);	×
99
100	// add reduction for loop
101	symbolic::Expression red_begin;	×
102	symbolic::Expression red_end;	×
103	if (axis_ >= 0) {	×
104	red_begin = iedge_input->begin_subset()[axis_];	×
105	red_end = iedge_input->end_subset()[axis_];	×
106	} else {	×
107	red_begin = iedge_input->begin_subset().back();	×
108	red_end = iedge_input->end_subset().back();	×
109	}
110	std::string red_name = builder.find_new_name("_i");	×
111	builder.add_container(red_name, types::Scalar(types::PrimitiveType::UInt64));	×
112	auto red_indvar = symbolic::symbol(red_name);	×
113	auto red_init = red_begin;	×
114	auto red_update = symbolic::add(red_indvar, symbolic::one());	×
115	auto red_cond = symbolic::Lt(red_indvar, symbolic::add(red_end, symbolic::one()));	×
NEW 116	auto red_map = &builder.add_for(*last_scope, red_indvar, red_cond, red_init, red_update, {}, block.debug_info());	×
117
118	// Create innermost block
119	auto &code_block = builder.add_block(red_map->root());	×
120
121	// Create access nodes for input and output
122	auto &input_access = builder.add_access(code_block, input_name);	×
123	auto &tmp2_access = builder.add_access(code_block, temp_name2);	×
124	auto &tmp_access_out = builder.add_access(code_block, temp_name);	×
125	auto &tmp_access_in = builder.add_access(code_block, temp_name2);	×
126
127	// Create index expressions for input and output
128	std::vector<symbolic::Expression> input_subset = loop_syms;	×
129
130	// Replace the reduction axis index with the reduction variable for input
131	if (axis_ >= 0 && axis_ < static_cast<int>(input_subset.size())) {	×
132	input_subset.insert(input_subset.begin() + axis_, red_indvar);	×
133	} else if (axis_ < 0) {	×
134	input_subset.push_back(red_indvar);	×
135	}	×
136
137	// Compute exponential
138	auto &exp_tasklet = builder.add_tasklet(code_block, data_flow::TaskletCode::expf, "_out", {"_in"});	×
NEW 139	builder	×
NEW 140	.add_computational_memlet(code_block, input_access, exp_tasklet, "_in", input_subset, iedge_input->base_type());	×
UNCOV 141	builder.add_computational_memlet(code_block, exp_tasklet, "_out", tmp2_access, {}, temp_type);	×
142
143	// Add to temp (reduction)
144	auto &add_tasklet = builder.add_tasklet(code_block, data_flow::TaskletCode::add, "_out", {"_in1", "_in2"});	×
145	builder.add_computational_memlet(code_block, tmp_access_in, add_tasklet, "_in1", {}, temp_type);	×
146	builder.add_computational_memlet(code_block, tmp2_access, add_tasklet, "_in2", {}, temp_type);	×
147	builder.add_computational_memlet(code_block, add_tasklet, "_out", tmp_access_out, {}, temp_type);	×
148
149	// Create writeback - assign the accumulated sum to output
150	auto &writeback_block = builder.add_block(*last_scope);	×
151	auto &tmp_access_wb = builder.add_access(writeback_block, temp_name);	×
152	auto &output_access_wb = builder.add_access(writeback_block, output_name);	×
153	auto &writeback_tasklet = builder.add_tasklet(writeback_block, data_flow::TaskletCode::assign, "_out", {"_in"});	×
154	builder.add_computational_memlet(writeback_block, tmp_access_wb, writeback_tasklet, "_in", {}, temp_type);	×
NEW 155	builder.add_computational_memlet(	×
NEW 156	writeback_block, writeback_tasklet, "_out", output_access_wb, loop_syms, oedge_output->base_type()	×
157	);
158
159	// Cleanup old block
160	builder.remove_memlet(block, *iedge_input);	×
161	builder.remove_memlet(block, *oedge_output);	×
162	builder.remove_node(block, *this);	×
163	builder.remove_child(parent, block);	×
164
165	return true;	×
166	}	×
167
168	std::unique_ptr<data_flow::DataFlowNode> SoftmaxNode::
169	clone(size_t element_id, const graph::Vertex vertex, data_flow::DataFlowGraph &parent) const {	×
NEW 170	return std::unique_ptr<data_flow::DataFlowNode>(new SoftmaxNode(element_id, this->debug_info(), vertex, parent, axis_)	×
171	);
UNCOV 172	}	×
173
174	nlohmann::json SoftmaxNodeSerializer::serialize(const data_flow::LibraryNode &library_node) {	×
175	const SoftmaxNode &node = static_cast<const SoftmaxNode &>(library_node);	×
176	nlohmann::json j;	×
177
178	j["code"] = node.code().value();	×
179	j["axis"] = node.axis();	×
180
181	return j;	×
182	}	×
183
184	data_flow::LibraryNode &SoftmaxNodeSerializer::deserialize(	×
185	const nlohmann::json &j, builder::StructuredSDFGBuilder &builder, structured_control_flow::Block &parent
186	) {
187	auto code = j["code"].get<std::string>();	×
188	if (code != LibraryNodeType_Softmax.value()) {	×
189	throw std::runtime_error("Invalid library node code");	×
190	}
191
192	sdfg::serializer::JSONSerializer serializer;	×
NEW 193	DebugInfoRegion debug_info = serializer.json_to_debug_info_region(j["debug_info_region"], builder.debug_info());	×
194
195	auto axis = j["axis"].get<int>();	×
196
197	return builder.add_library_node<SoftmaxNode>(parent, debug_info, axis);	×
198	}	×
199
200	} // namespace ml
201	} // namespace math
202	} // namespace sdfg

daisytuner / sdfglib / 17300165647

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