DetessCast.h File

DetessCast Node — fused CVU kernel that detessellates and casts BF16→FP32. More...

Included Headers

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

Namespaces Index

namespace	simaai
namespace	neat
namespace	nodes

Classes Index

struct	DetessCastOptions
	Construction options for a DetessCast Node. More...
class	DetessCast
	Fused CVU kernel Node: Detess followed by Cast (BF16→FP32). More...

Description

DetessCast Node — fused CVU kernel that detessellates and casts BF16→FP32.

Combines an MLA-layout untessellate with a BF16→FP32 cast in a single CVU pass. Inserted after MLA stages on the BF16 output path when downstream wants natural HWC/CHW float tensors — fusing the two stages saves a DDR round-trip versus running Detess then Cast.

See Also

"The dtype contract" page in /concepts/dtype_contract

File Listing

The file content with the documentation metadata removed is:

1

12#pragma once

13

14#include "builder/Node.h"

15#include "builder/NodeContractConfigurable.h"

16#include "builder/NodeContractProvider.h"

17#include "builder/OutputSpec.h"

18

19#include <nlohmann/json.hpp>

20

21#include <memory>

22#include <optional>

23#include <string>

24#include <vector>

25

26namespace simaai::neat {

27class Model;

28struct CompiledProcessCvuContract;

29

35struct DetessCastOptions {

37 DetessCastOptions() = default;

39 explicit DetessCastOptions(const simaai::neat::Model& model);

40

41 std::string config_path;

42 std::optional<nlohmann::json>

43 config_json;

44 std::string upstream_name;

45 std::string element_name;

46 std::shared_ptr<const CompiledProcessCvuContract>

47 compiled_contract;

48 int num_buffers = 0;

49 int num_buffers_model = 0;

50 bool num_buffers_locked = false;

51};

52

64class DetessCast final : public Node,

65 public NodeContractProvider,

66 public NodeContractConfigurable,

67 public OutputSpecProvider {

68public:

70 explicit DetessCast(DetessCastOptions opt = {});

71 struct ConfigHolder;

72

74 std::string kind() const override {

75 return "DetessCast";

76 }

78 NodeCapsBehavior caps_behavior() const override {

79 return NodeCapsBehavior::Static;

80 }

82 NodeContractDefinition contract_definition() const override;

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

85 std::string* err) const override;

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

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

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

93 OutputSpec output_spec(const OutputSpec& input) const override;

94

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

97

99 const DetessCastOptions& 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 DetessCastOptions 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> DetessCast(DetessCastOptions opt = {});

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

---

Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.