Provide span abstractions for host and device memory

[jrhemstad]

I would like Thrust to provide the equivalent of a std::span, i.e., a non-owning view of a contiguous sequence of objects.

Like host_vector and device_vector, I believe there should be separate host_span and device_span (and universal?) classes to indicate if the data is safe to touch from host or device.

A key requirement is that device_span must work to pass by value and is usable directly into CUDA kernels, e.g.,

__global__ void kernel(device_span<int> a, device_span<const int> b){
   if(tid < b.size()){
      a[tid] = b[tid] * 42;
   }
}

[hcedwar]

Another strategy is to follow the cuda::atomic<T,scope> vs. cuda::std::atomic<T> pattern for extension; for example, cuda::span<T,Extent,Property> where Property denotes the memory space.

[griwes]

@hcedwar yeah, I think we'll end up doing something like that for libcu++; however, the Thrust types are more involved and should follow the pattern of allowing access to device-only memory from the host by copying it when used with non-managed memory. These would serve a slightly different usage pattern than the libcu++ ones.

[brycelelbach]

I think Carter is correct that we will have tagged span types, just as vector_base and pointer are tagged.

…

On Mon, Nov 16, 2020 at 8:23 PM Michał Dominiak ***@***.***> wrote: @hcedwar <https://github.com/hcedwar> yeah, I think we'll end up doing something like that for libcu++; however, the Thrust types are more involved and should follow the pattern of allowing access to device-only memory from the host by copying it when used with non-managed memory. These would serve a slightly different usage pattern than the libcu++ ones. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#752>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AADBG4WTIEEJCZJZPIZT2CDSQH3E5ANCNFSM4TNB46EQ> .

-- Bryce Adelstein Lelbach aka wash (he/him/his) US Programming Language Standards (PL22) Chair ISO C++ Library Evolution Chair CppCon and C++Now Program Chair C++ Core Compute Libraries (Thrust, CUB, libcu++) Lead @ NVIDIA --

[alliepiper]

@hschwane submitted NVIDIA/thrust#1407, which implements a vector_reference that provides a non-owning view, similar to span. We'll wait for the proper span implementation to settle instead of merging a temporary workaround into main, but folks may be interested in that patch in the meantime.

[fortminors]

Hello! Are there any news on implementing this?

[jrhemstad]

Hello! Are there any news on implementing this?

libcu++ now provides <cuda/std/span>: https://godbolt.org/z/113Mhfv9h

However, there is no device or host specific span type yet. That is something we're still talking about though.