exec_space¶
Where a procedure runs.
Declaración¶
Sintaxis¶
exec_space = Host | Device | HostDevice | Kernel
Values¶
Value |
Meaning |
|---|---|
|
ordinary CPU code. This is what every procedure is until something says otherwise. |
|
runs on the GPU, and is called from other device code. |
|
reachable from the host and from the device alike, and emitted for both. |
|
the entry point of a GPU kernel: the host launches it with a GpuKernelLaunch. |
Valores devueltos¶
None. An enumeration value is not evaluated.
Descripción¶
The execution space is a member of FunctionType, so it belongs to the signature of the procedure rather than to its symbol.
Device and Kernel answer different questions. Both run on the GPU, but a
Device procedure is an ordinary callee, while a Kernel is what a launch
names: it takes the __global__ qualifier of CUDA or the kernel qualifier
of Metal, its arguments are bound as buffers, and it returns nothing.
Only the kernels are marked by the frontend. The device_partition pass takes
the closure of the call graph from them and gives every procedure it reaches
the space it belongs in, so a procedure called from both a kernel and the host
becomes HostDevice.