Improving context stack

[sklam]

based on #6762

Improves the context stack implementation in #6762 and refactor the target options and flags.

[sklam]

This will need CUDA smoketest when done

[stuartarchibald]

stuartarchibald@4666c77 hacks through the flags for operator.add.

[sklam]

There're known cuda error due to changes to error msg. To be fixed later.

[sklam]

@stuartarchibald,

stuartarchibald/numba@4666c77 hacks through the flags for operator.add.

looks promising. I added a print of the function name and the flags in the lowerer and it showing the right things:

----- jit_wrapper__built_in_function_add_ Flags(enable_looplift=True, enable_pyobject=False, enable_pyobject_looplift=True, debuginfo=False, boundscheck=False, no_cpython_wrapper=True, nrt=True, fastmath=FastMathOptions({'fast'}))
----- foo Flags(enable_looplift=True, enable_pyobject=False, enable_pyobject_looplift=True, debuginfo=False, boundscheck=False, nrt=True, fastmath=FastMathOptions({'fast'}))

for

from numba import njit

@njit(fastmath=True)
def foo(x, y):
  a = x + y
  return a


x, y = 1, 2
r = foo(x, y)
print(r)

[stuartarchibald]

Thanks for the patch @sklam, looks good, just a few minor things to resolve.

[stuartarchibald]

Are both of these conditions necessary?

[sklam]

tls_flags.nrt by itself can be reading the default. Here, it is checking that "nrt" is not default and its value is True.

[stuartarchibald]

Docstr? Particularly explaining what this is going to do.

[stuartarchibald]

I've also checked that this does transmit the flags correctly, applying this patch: stuartarchibald@2f63896 will get the flags to the right place and also disables the fastmath visitor. Then running something like:

from numba import njit, types
import numpy as np
import operator

def foo(n):
    return operator.add(n, 3.14)

# no fast
jfoo = njit(foo)
jfoo.compile((types.float64,),)
print(jfoo.inspect_llvm(jfoo.signatures[0]))

with and without fastmath=True in the jit decorator demonstrates the presence (or not) of a fadd fast in the LLVM IR.

[stuartarchibald]

Buildfarm ID: numba_smoketest_cpu_yaml_13.

[stuartarchibald]

Both CPU and CUDA tests are running through the farm.

[stuartarchibald]

CUDA tests are failing on all platforms with:

ERROR: test_target_cuda_unrecognized_arg (numba.cuda.tests.cudapy.test_vectorize_decor.TestVectorizeUnrecognizedArg)
<traceback>
KeyError: "cuda vectorize target does not support option: 'nonexistent'"

[sklam]

started new cuda smoketest: numba_smoketest_cuda_yaml_35

[stuartarchibald]

Thanks for the fixes, change set from ec69a18 onwards is checked, few minor things to resolve else looks good.

[stuartarchibald]

Thanks for the fixes!

[stuartarchibald]

started new cuda smoketest: numba_smoketest_cuda_yaml_35

This passed on the 5/6 of the CUDA test matrix that ran (the rest suffered an intermittent unrelated problem) and given the nature of the patch/previous failures, that any part of the matrix passed is sufficient confirmation of the fixes. Thanks again for all your work on this, it should unlock a load more optimisation opportunities :) Marking ready to merge...!

[sklam]

Azure CI build 24b16a6 twice. The first time is all green https://dev.azure.com/numba/numba/_build/results?buildId=8387&view=results. The second time it got stuck in internet traffic jam.