On Fri, Oct 07, 2022 at 04:26:58PM +0200, Tobias Burnus wrote: > libgomp/nvptx: Prepare for reverse-offload callback handling > > This patch adds a stub 'gomp_target_rev' in the host's target.c, which will > later handle the reverse offload. > For nvptx, it adds support for forwarding the offload gomp_target_ext call > to the host by setting values in a struct on the device and querying it on > the host - invoking gomp_target_rev on the result. > > For host-device consistency guarantee reasons, reverse offload is currently > limited -march=sm_70 (for libgomp). > > gcc/ChangeLog: > > * config/nvptx/mkoffload.cc (process): Warn if the linked-in libgomp.a > has not been compiled with sm_70 or higher and disable code gen then. > > include/ChangeLog: > > * cuda/cuda.h (enum CUdevice_attribute): Add > CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING. > (CU_MEMHOSTALLOC_DEVICEMAP): Define. > (cuMemHostAlloc): Add prototype. > > libgomp/ChangeLog: > > * config/nvptx/icv-device.c (GOMP_DEVICE_NUM_VAR): Remove > 'static' for this variable. > * config/nvptx/libgomp-nvptx.h: New file. > * config/nvptx/target.c: Include it. > (GOMP_ADDITIONAL_ICVS): Declare extern var. > (GOMP_REV_OFFLOAD_VAR): Declare var. > (GOMP_target_ext): Handle reverse offload. > * libgomp-plugin.h (GOMP_PLUGIN_target_rev): New prototype. > * libgomp-plugin.c (GOMP_PLUGIN_target_rev): New, call ... > * target.c (gomp_target_rev): ... this new stub function. > * libgomp.h (gomp_target_rev): Declare. > * libgomp.map (GOMP_PLUGIN_1.4): New; add GOMP_PLUGIN_target_rev. > * plugin/cuda-lib.def (cuMemHostAlloc): Add. > * plugin/plugin-nvptx.c: Include libgomp-nvptx.h. > (struct ptx_device): Add rev_data member. > (nvptx_open_device): #if 0 unused check; add > unified address assert check. > (GOMP_OFFLOAD_get_num_devices): Claim unified address > support. > (GOMP_OFFLOAD_load_image): Free rev_fn_table if no > offload functions exist. Make offload var available > on host and device. > (rev_off_dev_to_host_cpy, rev_off_host_to_dev_cpy): New. > (GOMP_OFFLOAD_run): Handle reverse offload.
So, does this mean one has to have gcc configured --with-arch=sm_70 or later to make reverse offloading work (and then on the other side no support for older PTX arches at all)? If yes, I was kind of hoping we could arrange for it to be more user-friendly, build libgomp.a normally (sm_35 or what is the default), build the single TU in libgomp that needs the sm_70 stuff with -march=sm_70 and arrange for mkoffload to link in the sm_70 stuff only if the user wants reverse offload (or has requires reverse_offload?). In that case ignore sm_60 and older devices, if reverse offload isn't wanted, don't link in the part that needs sm_70 and make stuff working on sm_35 and later. Or perhaps have 2 versions of target.o, one sm_35 and one sm_70 and let mkoffload choose among them. > + /* The code for nvptx for GOMP_target_ext in > libgomp/config/nvptx/target.c > + for < sm_70 exists but is disabled here as it is unclear whether there > + is the required consistency between host and device. > + See https://gcc.gnu.org/pipermail/gcc-patches/2022-October/602715.html > + for details. */ > + warning_at (input_location, 0, > + "Disabling offload-code generation for this device type: " > + "%<omp requires reverse_offload%> can only be fulfilled " > + "for %<sm_70%> or higher"); > + inform (UNKNOWN_LOCATION, > + "Reverse offload requires that GCC is configured with " > + "%<--with-arch=sm_70%> or higher and not overridden by a lower " > + "value for %<-foffload-options=nvptx-none=-march=%>"); Diagnostics (sure, Fortran FE is an exception) shouldn't start with capital letters). > @@ -519,10 +523,20 @@ nvptx_open_device (int n) > CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR, dev); > ptx_dev->max_threads_per_multiprocessor = pi; > > +#if 0 > + int async_engines; > r = CUDA_CALL_NOCHECK (cuDeviceGetAttribute, &async_engines, > CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT, dev); > if (r != CUDA_SUCCESS) > async_engines = 1; > +#endif Please avoid #if 0 code. > + > + /* Required below for reverse offload as implemented, but with compute > + capability >= 2.0 and 64bit device processes, this should be > universally be > + the case; hence, an assert. */ > + r = CUDA_CALL_NOCHECK (cuDeviceGetAttribute, &pi, > + CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING, dev); > + assert (r == CUDA_SUCCESS && pi); > > for (int i = 0; i != GOMP_DIM_MAX; i++) > ptx_dev->default_dims[i] = 0; > @@ -1179,8 +1193,10 @@ GOMP_OFFLOAD_get_num_devices (unsigned int > omp_requires_mask) > { > int num_devices = nvptx_get_num_devices (); > /* Return -1 if no omp_requires_mask cannot be fulfilled but > - devices were present. */ > - if (num_devices > 0 && omp_requires_mask != 0) > + devices were present. Unified-shared address: see comment in 2 spaces after . rather than 1. > --- a/libgomp/target.c > +++ b/libgomp/target.c > @@ -2925,6 +2925,25 @@ GOMP_target_ext (int device, void (*fn) (void *), > size_t mapnum, > htab_free (refcount_set); > } > > +/* Handle reverse offload. This is called by the device plugins for a > + reverse offload; it is not called if the outer target runs on the host. > */ Likewise. Jakub
