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 |
| 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
Generated via doxygen2docusaurus 2.0.0 by Doxygen 1.9.8.