[USM] Will sycl::aligned_alloc_host allocate pinned host momery?

[GHGmc2]

• code example

#include "CL/sycl.hpp"

int main() {
  sycl::queue q(sycl::gpu_selector{});
  void *h = sycl::aligned_alloc_host(64, 256, q);
  sycl::free(h, q);
  return 0;
}

I notice that there is an extension for sycl buffer "Use Pinned Host Memory Property", but how about USM host allocator, will sycl::aligned_alloc_host allocate pinned host momery?

[GHGmc2]

A clue from Level Zero spec on "zeMemAllocHost":

• "Host allocations are frequently used as staging areas to transfer data to or from devices."

[zjin-lcf]

Thank you for providing the example. Running the example with Nvprof shows pinned host memory. I think this example also shows that the CUDA backend adds a lot of other runtime API calls. Do you think they are runtime overhead ?

No kernels were profiled.
Type Time(%) Time Calls Avg Min Max Name
API calls: 70.95% 179.94ms 1 179.94ms 179.94ms 179.94ms cuCtxCreate
28.76% 72.943ms 1 72.943ms 72.943ms 72.943ms cuCtxDestroy
0.16% 405.47us 1 405.47us 405.47us 405.47us cuMemAllocHost
0.10% 247.62us 1 247.62us 247.62us 247.62us cuMemFreeHost
0.01% 15.124us 1 15.124us 15.124us 15.124us cuStreamCreate
0.01% 14.804us 1 14.804us 14.804us 14.804us cuStreamDestroy
0.00% 11.401us 6 1.9000us 1.0600us 4.6500us cuDeviceGetPCIBusId
0.00% 10.350us 42 246ns 201ns 504ns cuDeviceGetAttribute
0.00% 9.6150us 1 9.6150us 9.6150us 9.6150us cuEventRecord
0.00% 8.1220us 1 8.1220us 8.1220us 8.1220us cuStreamSynchronize
0.00% 7.9980us 1 7.9980us 7.9980us 7.9980us cuCtxSynchronize
0.00% 3.7810us 1 3.7810us 3.7810us 3.7810us cuPointerGetAttributes
0.00% 3.2670us 12 272ns 210ns 481ns cuDeviceGet
0.00% 2.7630us 7 394ns 243ns 644ns cuCtxGetCurrent
0.00% 2.2660us 1 2.2660us 2.2660us 2.2660us cuEventCreate
0.00% 1.9570us 1 1.9570us 1.9570us 1.9570us cuEventDestroy
0.00% 1.2620us 3 420ns 303ns 611ns cuDeviceGetCount
0.00% 888ns 1 888ns 888ns 888ns cuCtxPopCurrent
~

[GHGmc2]

@zjin-lcf Just to confirm if SYCL supports allocating pinned host memory, since sometimes it brings much benefit when doing memcpy like CUDA docs say: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#page-locked-host-memory

[zjin-lcf]

Ok. You may write a SYCL program to confirm my answer using the example in https://developer.nvidia.com/blog/how-optimize-data-transfers-cuda-cc/. That program will be great.