[SYCL][USM] Make host device return nullptr for bad mallocs (huge, 0, etc)

[jbrodman]

Host device wasn't properly handling alloc::unknown.
Also, aligned allocator wasn't failing when it really should, so check to throw an error.

Signed-off-by: James Brodman [email protected]

[bader]

@jbrodman, do you know why new test fails on Windows/GPU?

[jbrodman]

@jbrodman, do you know why new test fails on Windows/GPU?

I do not, but building on windows takes forever. Looking into it.

[jbrodman]

@jbrodman, do you know why new test fails on Windows/GPU?

I do not, but building on windows takes forever. Looking into it.

@bader Ok - I was able to get everything built on windows again. I updated my DCH Windows drivers to version 26.20.100.7584 from 12/18/2019 (aka latest public version) and the tests pass, so I dunno what the deal is.

[bader]

Interesting... buildbot machines uses 26.20.100.7372 driver and @jbrodman validated that 26.20.100.7584 driver has no issues. I don't understand why driver version might be the problem as new test should not call any driver API, unless I missing something. The tested functions should return at argument validation before calling driver API. Right?

[jbrodman]

Interesting... buildbot machines uses 26.20.100.7372 driver and @jbrodman validated that 26.20.100.7584 driver has no issues. I don't understand why driver version might be the problem as new test should not call any driver API, unless I missing something. The tested functions should return at argument validation before calling driver API. Right?

The changes only checked for overflow conditions really. The older driver might still have not been dying when it should've? Modified the test so we can get a better idea of where it's failing.

[jbrodman]

Yeah I'm pretty sure this is dying in NEO (which is barfing on sz = -1). When are we updating the Win GPU drivers?

Can we blacklist this test on windows for now?

[bader]

Yeah I'm pretty sure this is dying in NEO (which is barfing on sz = -1).

Why? Don't this change prevents SYCL runtime from using lower level API?

    if (Size > NumBytes)
      throw std::bad_alloc();

Maybe we missing similar check in other functions...

[jbrodman]

Yeah I'm pretty sure this is dying in NEO (which is barfing on sz = -1).

Why? Don't this change prevents SYCL runtime from using lower level API?

    if (Size > NumBytes)
      throw std::bad_alloc();

Maybe we missing similar check in other functions...

The test passes on the HOST device. It fails on the GPU device b/c of old windows drivers. Native device support doesn't do the funky alignment massaging your aligned_allocator does, so that fix is not relevant for those devices. (or rather, they need to do the right thing inside their own APIs)

[romanovvlad]

Yeah I'm pretty sure this is dying in NEO (which is barfing on sz = -1).

Why? Don't this change prevents SYCL runtime from using lower level API?

    if (Size > NumBytes)
      throw std::bad_alloc();

Maybe we missing similar check in other functions...

The test passes on the HOST device. It fails on the GPU device b/c of old windows drivers. Native device support doesn't do the funky alignment massaging your aligned_allocator does, so that fix is not relevant for those devices. (or rather, they need to do the right thing inside their own APIs)

Suggest disabling the test to unblock commit.
You can disable entire LIST test on windows by adding
// UNSUPPORTED: windows
or disable run on GPU on both OSes by breaking RUN word
// RUNx: %GPU_RUN_PLACEHOLDER %t.out

[jbrodman]

Yeah I'm pretty sure this is dying in NEO (which is barfing on sz = -1).

Why? Don't this change prevents SYCL runtime from using lower level API?

    if (Size > NumBytes)
      throw std::bad_alloc();

Maybe we missing similar check in other functions...

The test passes on the HOST device. It fails on the GPU device b/c of old windows drivers. Native device support doesn't do the funky alignment massaging your aligned_allocator does, so that fix is not relevant for those devices. (or rather, they need to do the right thing inside their own APIs)

Suggest disabling the test to unblock commit.
You can disable entire LIST test on windows by adding
// UNSUPPORTED: windows
or disable run on GPU on both OSes by breaking RUN word
// RUNx: %GPU_RUN_PLACEHOLDER %t.out

Thanks. Will just disable on windows for now.