CastTess.h File

CastTess Node — fused CVU kernel that casts FP32→BF16 then tessellates. More...

Included Headers

#include "builder/Node.h" #include "builder/NodeContractConfigurable.h" #include "builder/NodeContractProvider.h" #include <nlohmann/json.hpp> #include <memory> #include <optional> #include <string> #include <vector>

Namespaces Index

namespace	simaai
namespace	neat
namespace	nodes

Classes Index

struct	CastTessOptions
	Construction options for a CastTess Node. More...
class	CastTess
	Fused CVU kernel Node: Cast (FP32→BF16) followed by Tess. More...

Description

CastTess Node — fused CVU kernel that casts FP32→BF16 then tessellates.

Combines a BF16 cast and a tessellate layout shuffle into a single CVU pass. Inserted by the route planner before the MLA when the model expects BF16 input and MLA-side tessellation is not part of the compiled MLA kernel — fusing the two stages saves a DDR round-trip versus running Cast and Tess back to back.

See Also

"The dtype contract" page in /concepts/dtype_contract

File Listing

The file content with the documentation metadata removed is:

1

13#pragma once

14

15#include "builder/Node.h"

16#include "builder/NodeContractConfigurable.h"

17#include "builder/NodeContractProvider.h"

18#ifdef SIMA_NEAT_INTERNAL

19#include "model/internal/ModelRouteRetarget.h"

20#endif

21

22#include <nlohmann/json.hpp>

23

24#include <memory>

25#include <optional>

26#include <string>

27#include <vector>

28

29namespace simaai::neat {

30class Model;

31struct CompiledProcessCvuContract;

32

38struct CastTessOptions {

40 CastTessOptions() = default;

42 explicit CastTessOptions(const simaai::neat::Model& model);

43

44 std::string config_path;

45 std::optional<nlohmann::json>

46 config_json;

47 std::string element_name;

48 std::shared_ptr<const CompiledProcessCvuContract>

49 compiled_contract;

50 int num_buffers = 0;

51 int num_buffers_model = 0;

52 bool num_buffers_locked = false;

53#ifdef SIMA_NEAT_INTERNAL

54 std::shared_ptr<const simaai::neat::internal::ModelLineageBinding> model_lineage;

55#endif

56};

57

69class CastTess final : public Node, public NodeContractProvider, public NodeContractConfigurable {

70public:

72 explicit CastTess(CastTessOptions opt = {});

73 struct ConfigHolder;

74

76 std::string kind() const override {

77 return "CastTess";

78 }

80 NodeCapsBehavior caps_behavior() const override {

81 return NodeCapsBehavior::Static;

82 }

84 NodeContractDefinition contract_definition() const override;

86 bool compile_node_contract(const ContractCompileInput& input, CompiledNodeContract* out,

87 std::string* err) const override;

89 void apply_compiled_contract(const CompiledNodeContract& contract, std::string* err) override;

91 std::string backend_fragment(int node_index) const override;

93 std::vector<std::string> element_names(int node_index) const override;

94

96 const nlohmann::json* config_json() const;

97

99 const CastTessOptions& options() const {

100 return opt_;

101 }

103 const std::string& config_path() const {

104 return config_path_;

105 }

106

107#ifdef SIMA_NEAT_INTERNAL

108 const std::optional<CompiledProcessCvuContract>& compiled_contract_internal() const;

109#endif

110

111private:

112 CastTessOptions opt_;

113 std::shared_ptr<ConfigHolder> config_holder_;

114 std::string config_path_;

115};

116

117} // namespace simaai::neat

118

119namespace simaai::neat::nodes {

121std::shared_ptr<simaai::neat::Node> CastTess(CastTessOptions opt = {});

122} // namespace simaai::neat::nodes

---

Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.