Determine remaining `COp`s for Numba conversion

[brandonwillard]

Before we can replace the C backend, we need to make sure all the existing C implementations for Ops have Numba equivalents. Let's look through the codebase and list the remaining Ops here.

[kc611]

Not sure which of these are redundant with respect to Numba but here's a list of all COp and COp derivatives that don't have a corresponding registration in Numba backend. Doesn't include things from aesara/gpuarray since that is an entirely different matter.

aesara/assert_op.py

• Assert(COp)

aesara/sandbox/multinomial.py

• ~MultinomialFromUniform(COp)~

aesara/sparse/basic.py

• StructuredDotGradCSC(COp)
• StructuredDotGradCSR(COp)

aesara/sparse/opt.py

• AddSD_ccode(_NoPythonCOp)
• StructuredDotCSC(COp)
• StructuredDotCSR(COp)
• UsmmCscDense(_NoPythonCOp)
• CSMGradC(_NoPythonCOp)
• MulSDCSC(_NoPythonCOp)
• MulSDCSR(_NoPythonCOp)
• MulSVCSR(_NoPythonCOp)
• StructuredAddSVCSR(_NoPythonCOp)
• SamplingDotCSR(_NoPythonCOp)

aesara/tensor/basic.py

• Split(COp)
• Flatten(COp)

aesara/tensor/blas_c.py

• CGer(BaseBLAS(COp), Ger)
• CGemv(BaseBLAS(COp), Gemv)

aesara/tensor/blas.py

• Gemm(GemmRelated(COp))
• Dot22(GemmRelated(COp))
• Dot22Scalar(GemmRelated(COp))

aesara/tensor/extra_ops.py

• ~CpuContiguous(COp)~

aesara/tensor/math.py

• Argmax(COp)

~aesara/typed_list/basic.py~

• ~GetItem(COp)~
• ~Append(COp)~
• ~Extend(COp)~
• ~Insert(COp)~
• ~Reverse(COp)~
• ~Length(COp)~

There are also these NoPythonOps:

aesara/ifelse.py

• IfElse(_NoPythonOp) - In progress

Edit: Removed the Ops in nnet since we're discontinuing support for this module

[ferrine]

is this list updated anyhow?

[brandonwillard]

is this list updated anyhow?

I've crossed out the Ops for which we don't really need conversions. The aesara.sparse sub-package requires an entirely different approach, and it has its own issue, so those probably shouldn't be considered here.