LiteralFloat with value Zero not supported in compilation of integral

[jorgensd]

MWE:

import basix.ufl
import ufl

c_el = basix.ufl.element("Lagrange", "interval", 1, shape=(1,))
mesh = ufl.Mesh(c_el)
C = ufl.as_tensor([[0, 0], [0, 1]])

integrand = ufl.transpose(C)[0, 0]

dx = ufl.Measure("dx", domain=mesh)
J = integrand * dx
forms = [J]

Traceback

Traceback (most recent call last):
  File "/usr/lib/python3.10/runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/usr/lib/python3.10/runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "/root/shared/ffcx/ffcx/__main__.py", line 14, in <module>
    sys.exit(main())
  File "/root/shared/ffcx/ffcx/main.py", line 75, in main
    code_h, code_c = compiler.compile_ufl_objects(
  File "/root/shared/ffcx/ffcx/compiler.py", line 118, in compile_ufl_objects
    code = generate_code(ir, options)
  File "/root/shared/ffcx/ffcx/codegeneration/codegeneration.py", line 48, in generate_code
    code_integrals = [integral_generator(integral_ir, options) for integral_ir in ir.integrals]
  File "/root/shared/ffcx/ffcx/codegeneration/codegeneration.py", line 48, in <listcomp>
    code_integrals = [integral_generator(integral_ir, options) for integral_ir in ir.integrals]
  File "/root/shared/ffcx/ffcx/codegeneration/C/integrals.py", line 42, in generator
    parts = ig.generate()
  File "/root/shared/ffcx/ffcx/codegeneration/integral_generator.py", line 168, in generate
    all_quadparts += self.generate_quadrature_loop(rule)
  File "/root/shared/ffcx/ffcx/codegeneration/integral_generator.py", line 273, in generate_quadrature_loop
    tensor_comp, intermediates_fw = self.generate_dofblock_partition(quadrature_rule)
  File "/root/shared/ffcx/ffcx/codegeneration/integral_generator.py", line 387, in generate_dofblock_partition
    block_quadparts, intermediate = self.generate_block_parts(
  File "/root/shared/ffcx/ffcx/codegeneration/integral_generator.py", line 506, in generate_block_parts
    var = fw if isinstance(fw, L.Symbol) else fw.array
AttributeError: 'LiteralFloat' object has no attribute 'array'

Issue first reported at
https://fenicsproject.discourse.group/t/asemble-scalar-error-literalfloat-no-array-attribute/14807?u=dokken

[jorgensd]

The issue might be:

ffcx/ffcx/ir/integral.py

Line 225 in 17045ea

block_is_transposed = False # FIXME: Handle transposes for these block types

[jorgensd]

Or it seems like some optimization of zeros is not applied:

import basix.ufl
import ufl

c_el = basix.ufl.element("Lagrange", "triangle", 1, shape=(2,))
mesh = ufl.Mesh(c_el)
C = ufl.as_tensor([[0, 0], [0, 1]])

int = ufl.transpose(C)[0, 0]
# int = C[0, 0]
dx = ufl.Measure("dx", domain=mesh)
J = int * dx

forms = [J]

This should be reduced to 0s, but it doesn't seem like it is.

[chrisrichardson]

Is this a problem only because the value is zero. The kernel works with non-zero values in C[0,0]?

[jorgensd]

Yes, it is only an issue when C[0,0] is 0.

[michalhabera]

An optimization won't be applied in this case, that is expected (UFL does not zero parts of expression tree at the granularity of each component). For expression you should see the 0.0 multiplication in the generated code?

I think what happens is https://github.com/FEniCS/ffcx/blob/main/ffcx/codegeneration/lnodes.py#L98 triggers and fw becomes LiteralFloat(0.0) which is neither a temporary symbol nor it has .array. A quick fix might be to check for literals and access .value in that case. Or in float_product return 0.0 for zero literal?

[chrisrichardson]

I don't think this will work, as the value 0.0 is still neither a Symbol, nor have an array property, and it will fail in the same place. We could cut out those lines in https://github.com/FEniCS/ffcx/blob/main/ffcx/codegeneration/lnodes.py#L98 - then it will just return Product() which will actually multiply by zero.

[michalhabera]

I don't think this will work, as the value 0.0 is still neither a Symbol, nor have an array property, and it will fail in the same place. We could cut out those lines in https://github.com/FEniCS/ffcx/blob/main/ffcx/codegeneration/lnodes.py#L98 - then it will just return Product() which will actually multiply by zero.

Thats sounds good. It will avoid extra work to check for possibly unused quadrature tables.