8b790434e5
Load and Store Linalg operations are converter to their LLVM IR counterparts
preceded by a sequence of operations that recover the effective address of the
accessed element. The address is computed given the subscripts and the view
descriptor as
base_pointer + base_offset + SUM_i subscript_i * stride_i.
Manual testing shows that the resulting LLVM IR for the matrix multiplication
example can be compiled and executed, producing correct results.
--
PiperOrigin-RevId: 241889003
|
||
|---|---|---|
| .. | ||
| Analysis.cpp | ||
| ConvertToLLVMDialect.cpp | ||
| DialectRegistration.cpp | ||
| LoadStoreOps.cpp | ||
| TensorOps.cpp | ||
| Transforms.cpp | ||