17602788814

Committed 05 Sep 2025 11:40AM UTC coverage: 59.145% (+0.09%) from 59.057%

Build # 17602788814

Build Type

push

github

Committed by

web-flow

Commit Message

Schedule type extension (#221)

* Initial Draft

* Simplify schedule type class for serialization

* string ref

* fix and = operator

Run Details

60 of 72 new or added lines in 19 files covered. (83.33%)

4 existing lines in 2 files now uncovered.

9274 of 15680 relevant lines covered (59.15%)

115.92 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/layer_normalization.cpp

#include "sdfg/data_flow/library_nodes/math/ml/layer_normalization.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 {

LayerNormalizationNode::LayerNormalizationNode(
    size_t element_id,
    const DebugInfo &debug_info,
    const graph::Vertex vertex,
    data_flow::DataFlowGraph &parent,
    int axis,
    const std::string &epsilon
)
    : MathNode(
          element_id,
          debug_info,
          vertex,
          parent,
          LibraryNodeType_LayerNormalization,
          {"Y"},
          {"X", "Scale", "B"},
          data_flow::ImplementationType_NONE
      ),
      axis_(axis), epsilon_(epsilon) {}

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

bool LayerNormalizationNode::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));
    int index = parent.index(block);
    auto &transition = parent.at(index).second;

    // Locate edges
    const data_flow::Memlet *iedge_input = nullptr;
    const data_flow::Memlet *iedge_scale = nullptr;
    const data_flow::Memlet *iedge_bias = nullptr;
    const data_flow::Memlet *oedge_output = nullptr;
    for (const auto &edge : dataflow.in_edges(*this)) {
        if (edge.dst_conn() == "X") {
            iedge_input = &edge;
        } else if (edge.dst_conn() == "Scale") {
            iedge_scale = &edge;
        } else if (edge.dst_conn() == "B") {
            iedge_bias = &edge;
        }
    }
    for (const auto &edge : dataflow.out_edges(*this)) {
        if (edge.src_conn() == "Y") {
            oedge_output = &edge;
        }
    }
    if (!iedge_input || !iedge_scale || !oedge_output) return false;

    std::string input_name = static_cast<const data_flow::AccessNode &>(iedge_input->src()).data();
    std::string scale_name = static_cast<const data_flow::AccessNode &>(iedge_scale->src()).data();
    std::string bias_name;
    if (iedge_bias) {
        bias_name = static_cast<const data_flow::AccessNode &>(iedge_bias->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, transition.assignments(), block.debug_info());
    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::create(),
            {},
            block.debug_info()
        );
        last_scope = &last_map->root();
        loop_syms.push_back(indvar);
    }

    // Initialize temp variables for mean and variance
    std::string mean_temp_name = builder.find_new_name("_mean_tmp");
    std::string var_temp_name = builder.find_new_name("_var_tmp");
    std::string count_temp_name = builder.find_new_name("_count_tmp");
    types::Scalar temp_type(types::PrimitiveType::Float);
    builder.add_container(mean_temp_name, temp_type);
    builder.add_container(var_temp_name, temp_type);
    builder.add_container(count_temp_name, temp_type);

    auto &init_block = builder.add_block(*last_scope);
    auto &init_mean_tasklet = builder.add_tasklet(init_block, data_flow::TaskletCode::assign, "_out", {"0.0f"});
    auto &init_var_tasklet = builder.add_tasklet(init_block, data_flow::TaskletCode::assign, "_out", {"0.0f"});
    auto &init_count_tasklet = builder.add_tasklet(init_block, data_flow::TaskletCode::assign, "_out", {"0.0f"});
    auto &mean_access_init = builder.add_access(init_block, mean_temp_name);
    auto &var_access_init = builder.add_access(init_block, var_temp_name);
    auto &count_access_init = builder.add_access(init_block, count_temp_name);
    builder.add_computational_memlet(init_block, init_mean_tasklet, "_out", mean_access_init, {}, temp_type);
    builder.add_computational_memlet(init_block, init_var_tasklet, "_out", var_access_init, {}, temp_type);
    builder.add_computational_memlet(init_block, init_count_tasklet, "_out", count_access_init, {}, temp_type);

    // Add reduction for loop to compute mean and variance
    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 for mean and variance computation
    auto &compute_block = builder.add_block(red_map->root());

    // Create access nodes
    auto &input_access = builder.add_access(compute_block, input_name);
    auto &mean_access_in = builder.add_access(compute_block, mean_temp_name);
    auto &mean_access_out = builder.add_access(compute_block, mean_temp_name);
    auto &var_access_in = builder.add_access(compute_block, var_temp_name);
    auto &var_access_out = builder.add_access(compute_block, var_temp_name);
    auto &count_access_in = builder.add_access(compute_block, count_temp_name);
    auto &count_access_out = builder.add_access(compute_block, count_temp_name);

    // Create index expressions for input
    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);
    }

    // Add to mean (reduction)
    auto &add_mean_tasklet = builder.add_tasklet(compute_block, data_flow::TaskletCode::add, "_out", {"_in1", "_in2"});
    builder.add_computational_memlet(
        compute_block, input_access, add_mean_tasklet, "_in1", input_subset, iedge_input->base_type()
    );
    builder.add_computational_memlet(compute_block, mean_access_in, add_mean_tasklet, "_in2", {}, temp_type);
    builder.add_computational_memlet(compute_block, add_mean_tasklet, "_out", mean_access_out, {}, temp_type);

    // Add to variance (reduction of squared values)
    auto &square_tasklet = builder.add_tasklet(compute_block, data_flow::TaskletCode::mul, "_out", {"_in1", "_in2"});
    auto &square_temp_access = builder.add_access(compute_block, builder.find_new_name("_square_tmp"));
    builder.add_computational_memlet(
        compute_block, input_access, square_tasklet, "_in1", input_subset, iedge_input->base_type()
    );
    builder.add_computational_memlet(
        compute_block, input_access, square_tasklet, "_in2", input_subset, iedge_input->base_type()
    );
    builder.add_computational_memlet(compute_block, square_tasklet, "_out", square_temp_access, {}, temp_type);

    auto &add_var_tasklet = builder.add_tasklet(compute_block, data_flow::TaskletCode::add, "_out", {"_in1", "_in2"});
    builder.add_computational_memlet(compute_block, square_temp_access, add_var_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(compute_block, var_access_in, add_var_tasklet, "_in2", {}, temp_type);
    builder.add_computational_memlet(compute_block, add_var_tasklet, "_out", var_access_out, {}, temp_type);

    // Increment count
    auto &add_count_tasklet = builder.add_tasklet(compute_block, data_flow::TaskletCode::add, "_out", {"_in1", "1.0f"});
    builder.add_computational_memlet(compute_block, count_access_in, add_count_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(compute_block, add_count_tasklet, "_out", count_access_out, {}, temp_type);

    // Create normalization block
    auto &norm_block = builder.add_block(*last_scope);

    // Create access nodes for normalization
    auto &mean_access_norm = builder.add_access(norm_block, mean_temp_name);
    auto &var_access_norm = builder.add_access(norm_block, var_temp_name);
    auto &count_access_norm = builder.add_access(norm_block, count_temp_name);
    auto &input_access_norm = builder.add_access(norm_block, input_name);
    auto &scale_access_norm = builder.add_access(norm_block, scale_name);
    auto &output_access_norm = builder.add_access(norm_block, output_name);

    // Compute mean by dividing sum by count
    auto &div_mean_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::div, "_out", {"_in1", "_in2"});
    auto &mean_result_access = builder.add_access(norm_block, builder.find_new_name("_mean_result"));
    builder.add_computational_memlet(norm_block, mean_access_norm, div_mean_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(norm_block, count_access_norm, div_mean_tasklet, "_in2", {}, temp_type);
    builder.add_computational_memlet(norm_block, div_mean_tasklet, "_out", mean_result_access, {}, temp_type);

    // Compute variance by dividing sum of squares by count and subtracting mean squared
    auto &div_var_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::div, "_out", {"_in1", "_in2"});
    auto &var_div_access = builder.add_access(norm_block, builder.find_new_name("_var_div"));
    builder.add_computational_memlet(norm_block, var_access_norm, div_var_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(norm_block, count_access_norm, div_var_tasklet, "_in2", {}, temp_type);
    builder.add_computational_memlet(norm_block, div_var_tasklet, "_out", var_div_access, {}, temp_type);

    auto &square_mean_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::mul, "_out", {"_in", "_in"});
    auto &mean_squared_access = builder.add_access(norm_block, builder.find_new_name("_mean_squared"));
    builder.add_computational_memlet(norm_block, mean_result_access, square_mean_tasklet, "_in", {}, temp_type);
    builder.add_computational_memlet(norm_block, square_mean_tasklet, "_out", mean_squared_access, {}, temp_type);

    auto &sub_var_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::sub, "_out", {"_in1", "_in2"});
    auto &var_result_access = builder.add_access(norm_block, builder.find_new_name("_var_result"));
    builder.add_computational_memlet(norm_block, var_div_access, sub_var_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(norm_block, mean_squared_access, sub_var_tasklet, "_in2", {}, temp_type);
    builder.add_computational_memlet(norm_block, sub_var_tasklet, "_out", var_result_access, {}, temp_type);

    // Add epsilon to variance and compute standard deviation
    auto &add_epsilon_tasklet =
        builder.add_tasklet(norm_block, data_flow::TaskletCode::add, "_out", {"_in1", epsilon_});
    auto &var_eps_access = builder.add_access(norm_block, builder.find_new_name("_var_eps"));
    builder.add_computational_memlet(norm_block, var_result_access, add_epsilon_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(norm_block, add_epsilon_tasklet, "_out", var_eps_access, {}, temp_type);

    auto &sqrt_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::sqrt, "_out", {"_in"});
    auto &std_dev_access = builder.add_access(norm_block, builder.find_new_name("_std_dev"));
    builder.add_computational_memlet(norm_block, var_eps_access, sqrt_tasklet, "_in", {}, temp_type);
    builder.add_computational_memlet(norm_block, sqrt_tasklet, "_out", std_dev_access, {}, temp_type);

    // Normalize: (x - mean) / std_dev
    auto &sub_norm_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::sub, "_out", {"_in1", "_in2"});
    auto &centered_access = builder.add_access(norm_block, builder.find_new_name("_centered"));
    builder.add_computational_memlet(
        norm_block, input_access_norm, sub_norm_tasklet, "_in1", loop_syms, iedge_input->base_type()
    );
    builder.add_computational_memlet(norm_block, mean_result_access, sub_norm_tasklet, "_in2", {}, temp_type);
    builder.add_computational_memlet(norm_block, sub_norm_tasklet, "_out", centered_access, {}, temp_type);

    auto &div_norm_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::div, "_out", {"_in1", "_in2"});
    auto &normalized_access = builder.add_access(norm_block, builder.find_new_name("_normalized"));
    builder.add_computational_memlet(norm_block, centered_access, div_norm_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(norm_block, std_dev_access, div_norm_tasklet, "_in2", {}, temp_type);
    builder.add_computational_memlet(norm_block, div_norm_tasklet, "_out", normalized_access, {}, temp_type);

    // Apply scale and bias: scale * normalized + bias (if bias is provided)
    auto &mul_scale_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::mul, "_out", {"_in1", "_in2"});
    auto &scaled_access = builder.add_access(norm_block, builder.find_new_name("_scaled"));
    builder.add_computational_memlet(norm_block, normalized_access, mul_scale_tasklet, "_in1", {}, temp_type);
    builder.add_computational_memlet(
        norm_block, scale_access_norm, mul_scale_tasklet, "_in2", loop_syms, iedge_scale->base_type()
    );
    builder.add_computational_memlet(norm_block, mul_scale_tasklet, "_out", scaled_access, {}, temp_type);

    if (iedge_bias) {
        // Add bias if provided
        auto &bias_access_norm = builder.add_access(norm_block, bias_name);
        auto &add_bias_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::add, "_out", {"_in1", "_in2"});
        builder.add_computational_memlet(norm_block, scaled_access, add_bias_tasklet, "_in1", {}, temp_type);
        builder.add_computational_memlet(
            norm_block, bias_access_norm, add_bias_tasklet, "_in2", loop_syms, iedge_bias->base_type()
        );
        builder.add_computational_memlet(
            norm_block, add_bias_tasklet, "_out", output_access_norm, loop_syms, oedge_output->base_type()
        );
    } else {
        // No bias, just assign scaled result to output
        auto &assign_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::assign, "_out", {"_in"});
        builder.add_computational_memlet(norm_block, scaled_access, assign_tasklet, "_in", {}, temp_type);
        builder.add_computational_memlet(
            norm_block, assign_tasklet, "_out", output_access_norm, loop_syms, oedge_output->base_type()
        );
    }

    // Cleanup old block
    builder.remove_memlet(block, *iedge_input);
    builder.remove_memlet(block, *iedge_scale);
    if (iedge_bias) {
        builder.remove_memlet(block, *iedge_bias);
    }
    builder.remove_memlet(block, *oedge_output);
    builder.remove_node(block, *this);
    builder.remove_child(parent, index + 1);

    return true;
}

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

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

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

    return j;
}

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

    sdfg::serializer::JSONSerializer serializer;
    DebugInfo debug_info = serializer.json_to_debug_info(j["debug_info"]);

    auto axis = j["axis"].get<int>();
    auto epsilon = j["epsilon"].get<std::string>();

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

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

1	#include "sdfg/data_flow/library_nodes/math/ml/layer_normalization.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	LayerNormalizationNode::LayerNormalizationNode(	×
12	size_t element_id,
13	const DebugInfo &debug_info,
14	const graph::Vertex vertex,
15	data_flow::DataFlowGraph &parent,
16	int axis,
17	const std::string &epsilon
18	)
19	: MathNode(	×
20	element_id,	×
21	debug_info,	×
22	vertex,	×
23	parent,	×
24	LibraryNodeType_LayerNormalization,
25	{"Y"},	×
26	{"X", "Scale", "B"},	×
27	data_flow::ImplementationType_NONE
28	),
29	axis_(axis), epsilon_(epsilon) {}	×
30
31	void LayerNormalizationNode::validate(const Function &) const { /* TODO */ }	×
32
33	bool LayerNormalizationNode::expand(builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager) {	×
34	auto &dataflow = this->get_parent();	×
35	auto &block = static_cast<structured_control_flow::Block &>(*dataflow.get_parent());	×
36
37	auto &scope_analysis = analysis_manager.get<analysis::ScopeAnalysis>();	×
38	auto &parent = static_cast<structured_control_flow::Sequence &>(*scope_analysis.parent_scope(&block));	×
39	int index = parent.index(block);	×
40	auto &transition = parent.at(index).second;	×
41
42	// Locate edges
43	const data_flow::Memlet *iedge_input = nullptr;	×
44	const data_flow::Memlet *iedge_scale = nullptr;	×
45	const data_flow::Memlet *iedge_bias = nullptr;	×
46	const data_flow::Memlet *oedge_output = nullptr;	×
47	for (const auto &edge : dataflow.in_edges(*this)) {	×
48	if (edge.dst_conn() == "X") {	×
49	iedge_input = &edge;	×
50	} else if (edge.dst_conn() == "Scale") {	×
51	iedge_scale = &edge;	×
52	} else if (edge.dst_conn() == "B") {	×
53	iedge_bias = &edge;	×
54	}	×
55	}
56	for (const auto &edge : dataflow.out_edges(*this)) {	×
57	if (edge.src_conn() == "Y") {	×
58	oedge_output = &edge;	×
59	}	×
60	}
61	if (!iedge_input \|\| !iedge_scale \|\| !oedge_output) return false;	×
62
63	std::string input_name = static_cast<const data_flow::AccessNode &>(iedge_input->src()).data();	×
64	std::string scale_name = static_cast<const data_flow::AccessNode &>(iedge_scale->src()).data();	×
65	std::string bias_name;	×
66	if (iedge_bias) {	×
67	bias_name = static_cast<const data_flow::AccessNode &>(iedge_bias->src()).data();	×
68	}	×
69	std::string output_name = static_cast<const data_flow::AccessNode &>(oedge_output->dst()).data();	×
70
71	// Create new sequence before
72	auto &new_sequence = builder.add_sequence_before(parent, block, transition.assignments(), block.debug_info());	×
73	structured_control_flow::Sequence *last_scope = &new_sequence;	×
74
75	// Create maps over output subset dims (parallel dims)
76	data_flow::Subset domain_begin = oedge_output->begin_subset();	×
77	data_flow::Subset domain_end = oedge_output->end_subset();	×
78
79	std::vector<symbolic::Expression> loop_syms;	×
80	structured_control_flow::Map *last_map = nullptr;	×
81	for (size_t d = 0; d < domain_begin.size(); ++d) {	×
82	std::string indvar_str = builder.find_new_name("_i");	×
83	builder.add_container(indvar_str, types::Scalar(types::PrimitiveType::UInt64));	×
84	auto indvar = symbolic::symbol(indvar_str);	×
85	auto init = domain_begin[d];	×
86	auto update = symbolic::add(indvar, symbolic::one());	×
87	auto cond = symbolic::Lt(indvar, symbolic::add(domain_end[d], symbolic::one()));	×
88	last_map = &builder.add_map(	×
89	*last_scope,	×
90	indvar,
91	cond,
92	init,
93	update,
NEW 94	structured_control_flow::ScheduleType_Sequential::create(),	×
95	{},	×
96	block.debug_info()	×
97	);
98	last_scope = &last_map->root();	×
99	loop_syms.push_back(indvar);	×
100	}	×
101
102	// Initialize temp variables for mean and variance
103	std::string mean_temp_name = builder.find_new_name("_mean_tmp");	×
104	std::string var_temp_name = builder.find_new_name("_var_tmp");	×
105	std::string count_temp_name = builder.find_new_name("_count_tmp");	×
106	types::Scalar temp_type(types::PrimitiveType::Float);	×
107	builder.add_container(mean_temp_name, temp_type);	×
108	builder.add_container(var_temp_name, temp_type);	×
109	builder.add_container(count_temp_name, temp_type);	×
110
111	auto &init_block = builder.add_block(*last_scope);	×
112	auto &init_mean_tasklet = builder.add_tasklet(init_block, data_flow::TaskletCode::assign, "_out", {"0.0f"});	×
113	auto &init_var_tasklet = builder.add_tasklet(init_block, data_flow::TaskletCode::assign, "_out", {"0.0f"});	×
114	auto &init_count_tasklet = builder.add_tasklet(init_block, data_flow::TaskletCode::assign, "_out", {"0.0f"});	×
115	auto &mean_access_init = builder.add_access(init_block, mean_temp_name);	×
116	auto &var_access_init = builder.add_access(init_block, var_temp_name);	×
117	auto &count_access_init = builder.add_access(init_block, count_temp_name);	×
118	builder.add_computational_memlet(init_block, init_mean_tasklet, "_out", mean_access_init, {}, temp_type);	×
119	builder.add_computational_memlet(init_block, init_var_tasklet, "_out", var_access_init, {}, temp_type);	×
120	builder.add_computational_memlet(init_block, init_count_tasklet, "_out", count_access_init, {}, temp_type);	×
121
122	// Add reduction for loop to compute mean and variance
123	symbolic::Expression red_begin;	×
124	symbolic::Expression red_end;	×
125	if (axis_ >= 0) {	×
126	red_begin = iedge_input->begin_subset()[axis_];	×
127	red_end = iedge_input->end_subset()[axis_];	×
128	} else {	×
129	red_begin = iedge_input->begin_subset().back();	×
130	red_end = iedge_input->end_subset().back();	×
131	}
132	std::string red_name = builder.find_new_name("_i");	×
133	builder.add_container(red_name, types::Scalar(types::PrimitiveType::UInt64));	×
134	auto red_indvar = symbolic::symbol(red_name);	×
135	auto red_init = red_begin;	×
136	auto red_update = symbolic::add(red_indvar, symbolic::one());	×
137	auto red_cond = symbolic::Lt(red_indvar, symbolic::add(red_end, symbolic::one()));	×
138	auto red_map = &builder.add_for(*last_scope, red_indvar, red_cond, red_init, red_update, {}, block.debug_info());	×
139
140	// Create innermost block for mean and variance computation
141	auto &compute_block = builder.add_block(red_map->root());	×
142
143	// Create access nodes
144	auto &input_access = builder.add_access(compute_block, input_name);	×
145	auto &mean_access_in = builder.add_access(compute_block, mean_temp_name);	×
146	auto &mean_access_out = builder.add_access(compute_block, mean_temp_name);	×
147	auto &var_access_in = builder.add_access(compute_block, var_temp_name);	×
148	auto &var_access_out = builder.add_access(compute_block, var_temp_name);	×
149	auto &count_access_in = builder.add_access(compute_block, count_temp_name);	×
150	auto &count_access_out = builder.add_access(compute_block, count_temp_name);	×
151
152	// Create index expressions for input
153	std::vector<symbolic::Expression> input_subset = loop_syms;	×
154
155	// Replace the reduction axis index with the reduction variable for input
156	if (axis_ >= 0 && axis_ < static_cast<int>(input_subset.size())) {	×
157	input_subset.insert(input_subset.begin() + axis_, red_indvar);	×
158	} else if (axis_ < 0) {	×
159	input_subset.push_back(red_indvar);	×
160	}	×
161
162	// Add to mean (reduction)
163	auto &add_mean_tasklet = builder.add_tasklet(compute_block, data_flow::TaskletCode::add, "_out", {"_in1", "_in2"});	×
164	builder.add_computational_memlet(	×
165	compute_block, input_access, add_mean_tasklet, "_in1", input_subset, iedge_input->base_type()	×
166	);
167	builder.add_computational_memlet(compute_block, mean_access_in, add_mean_tasklet, "_in2", {}, temp_type);	×
168	builder.add_computational_memlet(compute_block, add_mean_tasklet, "_out", mean_access_out, {}, temp_type);	×
169
170	// Add to variance (reduction of squared values)
171	auto &square_tasklet = builder.add_tasklet(compute_block, data_flow::TaskletCode::mul, "_out", {"_in1", "_in2"});	×
172	auto &square_temp_access = builder.add_access(compute_block, builder.find_new_name("_square_tmp"));	×
173	builder.add_computational_memlet(	×
174	compute_block, input_access, square_tasklet, "_in1", input_subset, iedge_input->base_type()	×
175	);
176	builder.add_computational_memlet(	×
177	compute_block, input_access, square_tasklet, "_in2", input_subset, iedge_input->base_type()	×
178	);
179	builder.add_computational_memlet(compute_block, square_tasklet, "_out", square_temp_access, {}, temp_type);	×
180
181	auto &add_var_tasklet = builder.add_tasklet(compute_block, data_flow::TaskletCode::add, "_out", {"_in1", "_in2"});	×
182	builder.add_computational_memlet(compute_block, square_temp_access, add_var_tasklet, "_in1", {}, temp_type);	×
183	builder.add_computational_memlet(compute_block, var_access_in, add_var_tasklet, "_in2", {}, temp_type);	×
184	builder.add_computational_memlet(compute_block, add_var_tasklet, "_out", var_access_out, {}, temp_type);	×
185
186	// Increment count
187	auto &add_count_tasklet = builder.add_tasklet(compute_block, data_flow::TaskletCode::add, "_out", {"_in1", "1.0f"});	×
188	builder.add_computational_memlet(compute_block, count_access_in, add_count_tasklet, "_in1", {}, temp_type);	×
189	builder.add_computational_memlet(compute_block, add_count_tasklet, "_out", count_access_out, {}, temp_type);	×
190
191	// Create normalization block
192	auto &norm_block = builder.add_block(*last_scope);	×
193
194	// Create access nodes for normalization
195	auto &mean_access_norm = builder.add_access(norm_block, mean_temp_name);	×
196	auto &var_access_norm = builder.add_access(norm_block, var_temp_name);	×
197	auto &count_access_norm = builder.add_access(norm_block, count_temp_name);	×
198	auto &input_access_norm = builder.add_access(norm_block, input_name);	×
199	auto &scale_access_norm = builder.add_access(norm_block, scale_name);	×
200	auto &output_access_norm = builder.add_access(norm_block, output_name);	×
201
202	// Compute mean by dividing sum by count
203	auto &div_mean_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::div, "_out", {"_in1", "_in2"});	×
204	auto &mean_result_access = builder.add_access(norm_block, builder.find_new_name("_mean_result"));	×
205	builder.add_computational_memlet(norm_block, mean_access_norm, div_mean_tasklet, "_in1", {}, temp_type);	×
206	builder.add_computational_memlet(norm_block, count_access_norm, div_mean_tasklet, "_in2", {}, temp_type);	×
207	builder.add_computational_memlet(norm_block, div_mean_tasklet, "_out", mean_result_access, {}, temp_type);	×
208
209	// Compute variance by dividing sum of squares by count and subtracting mean squared
210	auto &div_var_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::div, "_out", {"_in1", "_in2"});	×
211	auto &var_div_access = builder.add_access(norm_block, builder.find_new_name("_var_div"));	×
212	builder.add_computational_memlet(norm_block, var_access_norm, div_var_tasklet, "_in1", {}, temp_type);	×
213	builder.add_computational_memlet(norm_block, count_access_norm, div_var_tasklet, "_in2", {}, temp_type);	×
214	builder.add_computational_memlet(norm_block, div_var_tasklet, "_out", var_div_access, {}, temp_type);	×
215
216	auto &square_mean_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::mul, "_out", {"_in", "_in"});	×
217	auto &mean_squared_access = builder.add_access(norm_block, builder.find_new_name("_mean_squared"));	×
218	builder.add_computational_memlet(norm_block, mean_result_access, square_mean_tasklet, "_in", {}, temp_type);	×
219	builder.add_computational_memlet(norm_block, square_mean_tasklet, "_out", mean_squared_access, {}, temp_type);	×
220
221	auto &sub_var_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::sub, "_out", {"_in1", "_in2"});	×
222	auto &var_result_access = builder.add_access(norm_block, builder.find_new_name("_var_result"));	×
223	builder.add_computational_memlet(norm_block, var_div_access, sub_var_tasklet, "_in1", {}, temp_type);	×
224	builder.add_computational_memlet(norm_block, mean_squared_access, sub_var_tasklet, "_in2", {}, temp_type);	×
225	builder.add_computational_memlet(norm_block, sub_var_tasklet, "_out", var_result_access, {}, temp_type);	×
226
227	// Add epsilon to variance and compute standard deviation
228	auto &add_epsilon_tasklet =	×
229	builder.add_tasklet(norm_block, data_flow::TaskletCode::add, "_out", {"_in1", epsilon_});	×
230	auto &var_eps_access = builder.add_access(norm_block, builder.find_new_name("_var_eps"));	×
231	builder.add_computational_memlet(norm_block, var_result_access, add_epsilon_tasklet, "_in1", {}, temp_type);	×
232	builder.add_computational_memlet(norm_block, add_epsilon_tasklet, "_out", var_eps_access, {}, temp_type);	×
233
234	auto &sqrt_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::sqrt, "_out", {"_in"});	×
235	auto &std_dev_access = builder.add_access(norm_block, builder.find_new_name("_std_dev"));	×
236	builder.add_computational_memlet(norm_block, var_eps_access, sqrt_tasklet, "_in", {}, temp_type);	×
237	builder.add_computational_memlet(norm_block, sqrt_tasklet, "_out", std_dev_access, {}, temp_type);	×
238
239	// Normalize: (x - mean) / std_dev
240	auto &sub_norm_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::sub, "_out", {"_in1", "_in2"});	×
241	auto &centered_access = builder.add_access(norm_block, builder.find_new_name("_centered"));	×
242	builder.add_computational_memlet(	×
243	norm_block, input_access_norm, sub_norm_tasklet, "_in1", loop_syms, iedge_input->base_type()	×
244	);
245	builder.add_computational_memlet(norm_block, mean_result_access, sub_norm_tasklet, "_in2", {}, temp_type);	×
246	builder.add_computational_memlet(norm_block, sub_norm_tasklet, "_out", centered_access, {}, temp_type);	×
247
248	auto &div_norm_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::div, "_out", {"_in1", "_in2"});	×
249	auto &normalized_access = builder.add_access(norm_block, builder.find_new_name("_normalized"));	×
250	builder.add_computational_memlet(norm_block, centered_access, div_norm_tasklet, "_in1", {}, temp_type);	×
251	builder.add_computational_memlet(norm_block, std_dev_access, div_norm_tasklet, "_in2", {}, temp_type);	×
252	builder.add_computational_memlet(norm_block, div_norm_tasklet, "_out", normalized_access, {}, temp_type);	×
253
254	// Apply scale and bias: scale * normalized + bias (if bias is provided)
255	auto &mul_scale_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::mul, "_out", {"_in1", "_in2"});	×
256	auto &scaled_access = builder.add_access(norm_block, builder.find_new_name("_scaled"));	×
257	builder.add_computational_memlet(norm_block, normalized_access, mul_scale_tasklet, "_in1", {}, temp_type);	×
258	builder.add_computational_memlet(	×
259	norm_block, scale_access_norm, mul_scale_tasklet, "_in2", loop_syms, iedge_scale->base_type()	×
260	);
261	builder.add_computational_memlet(norm_block, mul_scale_tasklet, "_out", scaled_access, {}, temp_type);	×
262
263	if (iedge_bias) {	×
264	// Add bias if provided
265	auto &bias_access_norm = builder.add_access(norm_block, bias_name);	×
266	auto &add_bias_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::add, "_out", {"_in1", "_in2"});	×
267	builder.add_computational_memlet(norm_block, scaled_access, add_bias_tasklet, "_in1", {}, temp_type);	×
268	builder.add_computational_memlet(	×
269	norm_block, bias_access_norm, add_bias_tasklet, "_in2", loop_syms, iedge_bias->base_type()	×
270	);
271	builder.add_computational_memlet(	×
272	norm_block, add_bias_tasklet, "_out", output_access_norm, loop_syms, oedge_output->base_type()	×
273	);
274	} else {	×
275	// No bias, just assign scaled result to output
276	auto &assign_tasklet = builder.add_tasklet(norm_block, data_flow::TaskletCode::assign, "_out", {"_in"});	×
277	builder.add_computational_memlet(norm_block, scaled_access, assign_tasklet, "_in", {}, temp_type);	×
278	builder.add_computational_memlet(	×
279	norm_block, assign_tasklet, "_out", output_access_norm, loop_syms, oedge_output->base_type()	×
280	);
281	}
282
283	// Cleanup old block
284	builder.remove_memlet(block, *iedge_input);	×
285	builder.remove_memlet(block, *iedge_scale);	×
286	if (iedge_bias) {	×
287	builder.remove_memlet(block, *iedge_bias);	×
288	}	×
289	builder.remove_memlet(block, *oedge_output);	×
290	builder.remove_node(block, *this);	×
291	builder.remove_child(parent, index + 1);	×
292
293	return true;	×
294	}	×
295
296	std::unique_ptr<data_flow::DataFlowNode> LayerNormalizationNode::
297	clone(size_t element_id, const graph::Vertex vertex, data_flow::DataFlowGraph &parent) const {	×
298	return std::unique_ptr<data_flow::DataFlowNode>(	×
299	new LayerNormalizationNode(element_id, this->debug_info(), vertex, parent, axis_, epsilon_)	×
300	);
301	}	×
302
303	nlohmann::json LayerNormalizationNodeSerializer::serialize(const data_flow::LibraryNode &library_node) {	×
304	const LayerNormalizationNode &node = static_cast<const LayerNormalizationNode &>(library_node);	×
305	nlohmann::json j;	×
306
307	j["code"] = node.code().value();	×
308	j["axis"] = node.axis();	×
309	j["epsilon"] = node.epsilon();	×
310
311	return j;	×
312	}	×
313
314	data_flow::LibraryNode &LayerNormalizationNodeSerializer::deserialize(	×
315	const nlohmann::json &j, builder::StructuredSDFGBuilder &builder, structured_control_flow::Block &parent
316	) {
317	auto code = j["code"].get<std::string>();	×
318	if (code != LibraryNodeType_LayerNormalization.value()) {	×
319	throw std::runtime_error("Invalid library node code");	×
320	}
321
322	sdfg::serializer::JSONSerializer serializer;	×
323	DebugInfo debug_info = serializer.json_to_debug_info(j["debug_info"]);	×
324
325	auto axis = j["axis"].get<int>();	×
326	auto epsilon = j["epsilon"].get<std::string>();	×
327
328	return builder.add_library_node<LayerNormalizationNode>(parent, debug_info, axis, epsilon);	×
329	}	×
330
331	} // namespace ml
332	} // namespace math
333	} // namespace sdfg

daisytuner / sdfglib / 17602788814

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