Sarek_codegen.Sarek_ir_cuda_coopmat_preambleThe CUDA/HIP cooperative-matrix helpers, emitted before extern "C" when the kernel has a cooperative-matrix statement (Sarek_ir_cuda_coopmat).
The AMD helpers have C++ linkage and one pointer type each, so a buffer of the wrong element type is a compile error rather than a reinterpretation.
Each operand helper names an access pattern, not an operand: load_a_* reads 16 contiguous elements per lane starting at p + i + l16 * s, which is an A fragment row-major and a B fragment column-major; load_b_* reads 16 elements s apart starting at p + i + l16, a B fragment row-major and an A fragment column-major. The contiguous form is the one the compiler turns into vector loads.
The stride is long long, so every offset product (l16 * s, e * s, 2 * j * s) is computed in 64 bits: an int product overflows once the stride passes 2^31 / 15 elements, although every element the fragment touches is addressable.