Helper to create a tensor filled with the given scalar. Scalar would be converted the to the element type of the given tensor type.

Print the params of a fusion with tensor shape at runtime. For debugging.

Helper to create a rank 0 tensor filled with the given `scalar`. `scalar` would be converted to the element type of the given `inputType`.

Deduce cuda major and minor version and cuda directory.

Check whether a depends on b.

return a tuple with major and minor version

Do the first level (tile level) codegen for stitch fusion pattern. An example is: Input IR ``` lmhlo.fusion() { lmhlo.exp(%in, %exp_out) : (memref<?x?

we can emit them in one kernel based on the fact that all the fused col reduction should be in the same shape.

* The generated kernel for row-reduce is as following. The code is augly and * will be removed. * * template <typename T> * __global__ void row_re

For concat op having many operands, we use following schedule: parallel.for %idx in range(num_operands_of_concat) { if (%idx == 0) { copy operand #0