[ACL] Performance behavior of using aclrtMallocHost() vs malloc()

I developed a single-device micro-benchmark code, implementing a matrix-matrix multiplication operation, based on ACL and targeting Atlas 800-3000. The workflow is the following:

1. Allocate data on the host

2. Allocate data on the device

3. Initialize data on the host

4. Transfer data from the host to the device

5. Launch aclblasGemmEx()

6. Synchronize to wait for the completion of the task in 5)

7. Copy back data from the device to the host

I measured the elapsed time for the step 5) and I experience a very low performance when I use normal libc malloc() for allocating host memory in step 1) instead of aclrtMallocHost(). This is surprising for me since the time measurement does not take into account the data transfers as I do consider only the step 5). I ran the application through the profiler and found that when I use malloc() the runtime is doing unexpected calls to HostMalloc during kernel launch (see the attachment for a screenshot of the profiling from MindStudio).

Using aclrtMallocHost leads to 35% of performance improvement compared to when malloc() is used.

Do you have any feedback on this behavior? Especially on what aclrtMallocHost() is doing differently

compared to malloc() ? Why do we have HostMalloc calls during kernel launch (note that these calls happen only when malloc() is used)?