Fix wrong device_id computation in identifier_from_devid

[janciesko]

Fixes: #4455

[dalg24]

How do we know this is less wrong than the code that was there before?

[janciesko]

I'll add a code comment that explains the bit range assignments.

[janciesko]

#include <stdio.h>

struct ExecutionSpaceIdentifier {
  uint32_t type;
  uint32_t device_id;
  uint32_t instance_id;
};
// Type your code here, or load an example.
ExecutionSpaceIdentifier one(const uint32_t in) {
      return {(in >> 24),  // first 8 bits
          ((in & 0x00FFFFFF) >> 17),  // next 7 bits
           (in & 0x0001FFFF)}; // last 17 bits
}

// Type your code here, or load an example.
ExecutionSpaceIdentifier two(const uint32_t in) {
return {(in >> 24),
          (~((uint32_t(-1)) << 7)) & (in >> 17),
          (~((uint32_t(-1)) << 17)) & in};
}

int main(int c, char ** args)
{
ExecutionSpaceIdentifier a = one(0x12342334);
ExecutionSpaceIdentifier b = two(0x12342334);
printf("%i, %i, %i, %i, %i, %i\n", a.type, a.device_id, a.instance_id, b.type, b.device_id, b.instance_id);
return 1;
}

@dalg24 This shows equivalence.

[janciesko]

retest this please

[janciesko]

Added unit test coverage

[cz4rs]

Looks good to me (only minor nitpicking in the comments).

[khuck]

@janciesko I am working with the current development branch of Kokkos and the ROCm 5.0.2 compiler... and for my HIP kernels, I am now getting the wrong device type and id. Instead of getting HIP and device 0, I am getting Serial and device 3. I have updated my code to match what is in the PR test example, but I'm still getting the wrong info. Did something else change?

[khuck]

btw, I know that Kokkos git hash e5a8acc gives what I am expecting...

[janciesko]

@Rombur - do you maybe know what might have changed in the HIP backend in respect to obtaining device Id and device type?

[Rombur]

No I don't know anything that could have change that

[khuck]

@janciesko @Rombur With the current master, I am running HIP kernels on device 0. The "device ID" passed to our profiler has the value 393217. That is being translated (by the above code) as a Serial kernel (0), on device 3, instance 1. That is incorrect...is the device getting constructed differently by Kokkos, or being interpreted incorrectly?

393217 is:

0000 0000 0000 0110 0000 0000 0000 0001
| type  | |device||      instance     |

The value should be:
50331649:

0000 0011 0000 0000 0000 0000 0000 0001
| type  | |device||      instance     |

...right? This has changed since the last Kokkos release (3.5).