cumsum() performance on Arc Pro B70 is lower than on Arc Pro B60

[caijimin]

With below code, the output on Intel Arc B60 is
memalloc_time=18.989502819720656, calculation_time=25.28791649499908
On B70
memalloc_time=18.39343382231891, calculation_time=40.41595908906311
The performance on B70 should be higher than B60, could you help to tell how to pinpoint the problem?
Do we have some tools to analyze?
Thanks.

import dpnp as np
import time

start_time = time.perf_counter()
a = np.array([0.1, 0.01, 0.001, 0.0001] * 4096*4096, dtype=np.float32)
end_time = time.perf_counter()
memalloc_time = end_time - start_time

i = 0
start_time = time.perf_counter()
while i < 10000:
    b = a.cumsum()
    i += 1
end_time = time.perf_counter()
calculation_time = end_time - start_time

print(
    f"memalloc_time={memalloc_time}, calculation_time={calculation_time}"
)

Tried some other functions like cumprod()/linalg.pinv()/linalg.solve() all have similar symptom.

[antonwolfy]

@caijimin, could you please provide output of python -m dpctl -f in both testing environments to have more info on exact device used?

Also, could you please provide the list of installed dependent packages through conda list or pip show command? Or exact command, how you installed dpnp package?
It'd be very helpful to understand the used versions.

Also, let me update the code you used a bit to have more fare scenario:

import dpnp as np
import time

start_time = time.perf_counter()
a = np.array([0.1, 0.01, 0.001, 0.0001] * 4096*4096, dtype=np.float32)
end_time = time.perf_counter()
memalloc_time = end_time - start_time

# decouple a cold run to build the kernel
b = a.cumsum()
np.synchronize_array_data(a)

i = 0
start_time = time.perf_counter()
while i < 10000:
    b = a.cumsum()
    i += 1
# need to wait until all SYCL kernels are completed
np.synchronize_array_data(a)
end_time = time.perf_counter()
calculation_time = end_time - start_time

print(
    f"memalloc_time={memalloc_time}, calculation_time={calculation_time}"
)