[REQ] wp.printf supports stderr

[erikfrey]

Description

I would like to pass an argument to wp.printf so that I can print to stderr instead of stdout.

Context

mjwarp-testspeed is a command line tool that is expected to produce parseable results to stdout. MJWarp currently prints runtime warnings to stdout that mangle the parseable results. I would like to print those warnings to stderr instead.

[christophercrouzet]

Hi @erikfrey,

Unless I'm missing something, CUDA only supports writing to stdout, so we'd need to do this redirection in Python around wp.launch()/wp.capture_launch(), but it's unclear how to restore that redirection since these calls are async.

Additionally:

• This seems quite fragile to implement such a redirection approach as part of a library.
• wp.printf() should probably only be used for debugging purposes, not for release builds.

@adenzler-nvidia mentioned that these prints are likely related to a workaround he introduced, and might try to address this differently instead. He will follow up in this thread soon.

If you still would like to implement such a redirection, you could use something along the lines of what we do in the unit tests:

warp/warp/tests/unittest_utils.py

Line 141 in 81677b2

class StreamCapture:

[erikfrey]

Thanks @christophercrouzet - definitely open to alternative suggestions. The main user requirement is to warn users about conditions encountered when running a graph that cannot be anticipated ahead of time.

Whether that's wp.printf or some other kind of warn mechanism, happy to discuss.

[christophercrouzet]

One general approach could be to output error values from the kernel, and processes these on the host, like so:

import sys
from enum import IntEnum

import warp as wp

wp.init()


class Status(IntEnum):
    ERROR_FOO = 0
    ERROR_BAR = 1


@wp.kernel
def kernel(i: int, out: wp.array(dtype=int)):
    tid = wp.tid()
    if tid % 2 == 0:
        if i % 2 == 0:
            wp.atomic_add(out, Status.ERROR_FOO, 1)
        else:
            wp.atomic_add(out, Status.ERROR_BAR, 1)


class Example:
    def __init__(self):
        self.out = wp.zeros(2, dtype=int)

    def step(self):
        for i in range(8):
            wp.launch(kernel=kernel, dim=4, inputs=(i,), outputs=(self.out,))


def run():
    example = Example()
    with wp.ScopedCapture() as capture:
        example.step()

    graph = capture.graph
    wp.capture_launch(graph)
    wp.synchronize()

    statuses = example.out.numpy()
    if statuses[Status.ERROR_FOO]:
        print("foo happened", file=sys.stderr)
    if statuses[Status.ERROR_BAR]:
        print("bar happened", file=sys.stderr)


if __name__ == "__main__":
    with wp.ScopedDevice("cuda"):
        run()

But let's wait for @adenzler-nvidia's input as he might have a suggestion better suited to MJWarp's use case.

[adenzler-nvidia]

Proposal here: google-deepmind/mujoco_warp#1089

Avoids the problem entirely by not doing any printfs anymore and relying on host-side errors. This also enables downstream implementers to implement their own error handling based on the data we support. I like that.