Level zero codegen

[diptorupd]

Fixes few left over things from level zero changes.

[diptorupd]

@oleksandr-pavlyk I made the change you suggested by dynamically opening libze_loader.so. To test if it works, I tried importing dpctl without having libze_loader.so as a target_link_library for libDPCTLSyclInterface.so and it works (earlier it used to throw an import error).

[oleksandr-pavlyk]

Used

# sasha_test.py
with open('dpctl/tests/input_files/multi_kernel.spv', 'br') as fh:
    sprv_pr = fh.read()
import dpctl, dpctl.memory as dpmem, dpctl.program as dppro
import numpy as np
import ctypes

q = dpctl.get_current_queue()

pr = dppro.create_program_from_spirv(q, sprv_pr)

k = pr.get_sycl_kernel('axpy')

s = np.dtype('i4').itemsize
a = dpmem.MemoryUSMShared(1024 * s)
b = dpmem.MemoryUSMShared(1024 * s)
c = dpmem.MemoryUSMShared(1024 * s)

ha = np.arange(0, 1024, dtype='i4')
hb = np.arange(1024-1, -1, -1, dtype='i4')
hc = np.zeros(1024, dtype='i4')

a.copy_from_host(ha.view(dtype='|u1'))
b.copy_from_host(hb.view(dtype='|u1'))
c.copy_from_host(hc.view(dtype='|u1'))

ev = q.submit(k, [a, b, c, ctypes.c_int(3)], [1024])
ev.wait()

res = c.copy_to_host().view(dtype='i4')
print(np.array_equal(res, ha + 3*hb))

q.get_sycl_device().dump_device_info()

Verified that libze_loader has been loaded and initialized:

(idp) [15:20:02 nuc dpctl]$ LD_DEBUG=libs python sasha_test.py |& grep ze_lo | grep init
   4085652:     calling init: /lib/x86_64-linux-gnu/libze_loader.so.1

The test passed, outputting

$ python sasha_test.py
True
    Name            Intel(R) Graphics [0x9bca]
    Driver version  1.0.18820
    Vendor          Intel(R) Corporation
    Profile         FULL_PROFILE
    Device type     gpu

[oleksandr-pavlyk]

I have not tried to build and run dpctl on a machine without level-zero installed, say any older machine with older than Gen9 GPU.