[MaxFreedomPollard]

ExprToISPCExprMapper.map_subscript calls self.kernel.get_grid_size_upper_bounds_as_exprs() with no arguments (loopy/target/ispc.py:172). callables_table has been a required parameter of that method since the kernel callables work in 2021 (loopy/kernel/__init__.py:1108), so ISPC code generation fails for any kernel that subscripts a private temporary:

TypeError: LoopKernel.get_grid_size_upper_bounds_as_exprs() missing 1 required positional argument: 'callables_table'

A kernel does not need an explicitly private array to hit this. An ilp iname privatizes the temporaries in its loop into arrays, so this fails the same way on 067ffc8:

knl = lp.make_kernel(
    "{[i]: 0<=i<n}",
    """
    <> tmp = 2*a[i]
    out[i] = tmp + 1
    """,
    [lp.GlobalArg("a,out", np.float32, shape="n"), lp.ValueArg("n", np.int32)],
    target=lp.ISPCTarget(), lang_version=(2018, 2))
knl = lp.split_iname(knl, "i", 32, outer_tag="g.0")
knl = lp.split_iname(knl, "i_inner", 4, outer_tag="l.0", inner_tag="ilp")
lp.generate_code_v2(knl)

The other three calls to this method in loopy/target/ispc.py already pass the table: map_variable since fbdda4d (#764, the scalar case from #763), the private temporary declarator since 34c7344, and pre_codegen_entrypoint_check. This call was missed. The change passes self.codegen_state.callables_table, as map_variable does twenty lines above. basedpyright flags the call as reportCallIssue ("Argument types: ()"), but that entry sits in .basedpyright/baseline.json.

With the fix, the accesses match the declaration that get_temporary_var_declarator emits, with one copy of the temporary per program instance (lsize + shape). For the kernel in the new test (tmp[j] private, i split by 8 into l.0 and g.0):

float tmp[8 * 4];
...
tmp[programIndex + 8 * (0 + j)] = 2.0 * a[4 * (varying int32) programIndex + 32 * ((uniform int32) taskIndex0) + j];
out[4 * (varying int32) programIndex + 32 * ((uniform int32) taskIndex0) + j] = tmp[programIndex + 8 * (0 + 3 + -1 * j)];

The ilp example above now generates float tmp[8 * 4];, with the unrolled accesses tmp[programIndex + 8 * (0 + 0)] through tmp[programIndex + 8 * (0 + 3)].

test_ispc_private_array in test/test_target.py is the subscripted counterpart of test_ispc_private_var. It checks the declaration and that both the store and the load are offset by programIndex. Like the other ISPC tests it only generates code, so it needs no ISPC compiler and no OpenCL device. I did not compile the generated ISPC, because no ISPC compiler was available here.

Verified on macOS 26.7.1 (arm64), Python 3.12.15, numpy 2.5.4, islpy 2026.2.2, namedisl 2026.1.1, pymbolic 2025.1, pyopencl 2026.1.4. This machine has no usable OpenCL device, so a local pytest plugin made pyopencl's fixture report zero devices, and the tests that need a context are skipped:

$ LOOPY_NO_CACHE=1 python -m pytest test/test_target.py -k ispc -q
4 passed, 49 deselected                    # this branch
1 failed, 3 passed, 49 deselected          # loopy/target/ispc.py reverted to 067ffc8: test_ispc_private_array
                                           # fails with the TypeError above at loopy/target/ispc.py:172

$ LOOPY_NO_CACHE=1 python -m pytest test/test_target.py test/test_loopy.py -q
1 failed, 98 passed, 134 skipped           # 067ffc8
1 failed, 99 passed, 134 skipped           # this branch

The per-test outcomes are the same on both trees except for the new test. The one failure, test_dgemm_with_rectangular_tile_prefetch, calls cl.create_some_context() itself and gets clGetDeviceIDs failed: INVALID_VALUE on both trees.

$ uv run --only-dev ruff check                       # ruff 0.17.2
All checks passed!

$ typos loopy/target/ispc.py test/test_target.py     # typos-cli 1.51.1
(no findings)

$ basedpyright loopy test                            # basedpyright 1.40.2
163 errors on 067ffc8 and on this branch. The error lists are identical apart from line numbers in loopy/target/ispc.py.

🤖 Generated with Claude Code

[MaxFreedomPollard]

The basedpyright failure is in loopy/symbolic.py, which this PR doesn't touch; it fails the same way on main (067ffc8).

[inducer]

Thanks!