|
| 1 | +# Overview |
| 2 | + |
| 3 | +The indirect clause enables **indirect device invocation** for a procedure: |
| 4 | +> 19 An indirect call to the device version of a procedure on a device other than the host<br> |
| 5 | +> 20 device, through a function pointer (C/C++), a pointer to a member function (C++) or<br> |
| 6 | +> 21 a procedure pointer (Fortran) that refers to the host version of the procedure. |
| 7 | +
|
| 8 | +# Compiler support |
| 9 | +### Offload entry metadata (C++ FE) |
| 10 | + |
| 11 | +For each function declared as **declare target indirect** C++ FE generates the following offload metadata: |
| 12 | + |
| 13 | +```c++ |
| 14 | +// Entry 0 -> Kind of this type of metadata (2) |
| 15 | +// Entry 1 -> Mangled name of the function. |
| 16 | +// Entry 2 -> Order the entry was created. |
| 17 | +``` |
| 18 | + |
| 19 | +The offloading metadata uses new `OffloadEntriesInfoManagerTy::OffloadingEntryInfoKinds::OffloadingEntryInfoDeviceIndirectFunc` metadata kind. |
| 20 | + |
| 21 | +### Offload entries table |
| 22 | + |
| 23 | +The offload entries table that is created for the host and for each of the device images currently have entries for **declare target** global variables, **omp target** outlined functions and constructor/destructor thunks for **declare target** global variables. |
| 24 | + |
| 25 | + |
| 26 | +Compiler will also produce an entry for each procedure listed in **indirect** clause of **declare target** construct: |
| 27 | +```C++ |
| 28 | +struct __tgt_offload_entry { |
| 29 | + void *addr; // Pointer to the function |
| 30 | + char *name; // Name of the function |
| 31 | + size_t size; // 0 for function |
| 32 | + int32_t flags; // OpenMPOffloadingDeclareTargetFlags::OMP_DECLARE_TARGET_FPTR |
| 33 | + int32_t reserved; // Reserved |
| 34 | +}; |
| 35 | +``` |
| 36 | +
|
| 37 | +### Run-time dispatch in device code |
| 38 | +
|
| 39 | +When an indirect function call is generated by a FE in **device code** it translates the original function pointer (which may be an address of a host function) into the device function pointer using a translation API, and uses the resulting function pointer for the call. |
| 40 | +
|
| 41 | +Original call code: |
| 42 | +
|
| 43 | +``` |
| 44 | + %0 = load void ()*, void ()** %fptr.addr |
| 45 | + call void %0() |
| 46 | +``` |
| 47 | +
|
| 48 | +Becomes this: |
| 49 | +
|
| 50 | +``` |
| 51 | + %0 = load void ()*, void ()** %fptr.addr |
| 52 | + %1 = bitcast void ()* %0 to i8* |
| 53 | + %call = call i8* @__kmpc_target_translate_fptr(i8* %1) |
| 54 | + %fptr_device = bitcast i8* %call to void ()* |
| 55 | + call void %fptr_device() |
| 56 | +``` |
| 57 | +
|
| 58 | +Device RTLs must provide the translation API: |
| 59 | +
|
| 60 | +```c++ |
| 61 | +// Translate \p FnPtr identifying a host function into a function pointer |
| 62 | +// identifying its device counterpart. |
| 63 | +// If \p FnPtr matches an address of any host function |
| 64 | +// declared as 'declare target indirect', then the API |
| 65 | +// must return an address of the same function compiled |
| 66 | +// for the device. If \p FnPtr does not match an address |
| 67 | +// of any host function, then the API returns \p FnPtr |
| 68 | +// unchanged. |
| 69 | +EXTERN void *__kmpc_target_translate_fptr(void *FnPtr); |
| 70 | +``` |
| 71 | + |
| 72 | +# Runtime handling of function pointers |
| 73 | + |
| 74 | +`OpenMPOffloadingDeclareTargetFlags::OMP_DECLARE_TARGET_FPTR` is a new flag to distinguish offload entries for function pointers from other function entries. Unlike other function entries (with `size` equal to 0) `omptarget::InitLibrary()` will establish mapping for function pointer entries in `Device.HostDataToTargetMap`. |
| 75 | + |
| 76 | +For each `OMP_DECLARE_TARGET_FPTR` entry in the offload entries table `libomptarget` creates an entry of the following type: |
| 77 | + |
| 78 | +```c++ |
| 79 | +struct __omp_offloading_fptr_map_ty { |
| 80 | + int64_t host_ptr; // key |
| 81 | + int64_t tgt_ptr; // value |
| 82 | +}; |
| 83 | +``` |
| 84 | +
|
| 85 | +Where `host_ptr` is `__tgt_offload_entry::addr` in a **host** offload entry, and `tgt_ptr` is `__tgt_offload_entry::addr` in the corresponding **device** offload entry (which may be found using the populated `Device.HostDataToTargetMap`). |
| 86 | +
|
| 87 | +When all `__omp_offloading_function_ptr_map_ty` entries are collected in a single host array, `libomptarget` sorts the table by `host_ptr` values and passes it to the device plugin for registration, if plugin supports optional `__tgt_rtl_set_function_ptr_map` API. |
| 88 | +
|
| 89 | +Plugins may provide the following API, if they want to support **declare target indirect** functionality: |
| 90 | +
|
| 91 | +```c++ |
| 92 | +// Register in a target implementation defined way a table |
| 93 | +// of __omp_offloading_function_ptr_map_ty entries providing |
| 94 | +// mapping between host and device addresses of 'declare target indirect' |
| 95 | +// functions. \p table_size is the number of elements in \p table_host_ptr |
| 96 | +// array. |
| 97 | +EXTERN void __tgt_rtl_set_function_ptr_map( |
| 98 | + int32_t device_id, uint64_t table_size, __omp_offloading_fptr_map_ty *table_host_ptr); |
| 99 | +``` |
| 100 | + |
| 101 | +# Sample implementation |
| 102 | + |
| 103 | +This section describes one of potential implementations. |
| 104 | + |
| 105 | +A FE may define the following global symbols for each translation module containing **declare target indirect**, when compiling this module for a device: |
| 106 | + |
| 107 | +```c++ |
| 108 | +// Mapping between host and device functions declared as |
| 109 | +// 'declare target indirect'. |
| 110 | +__attribute__((weak)) struct __omp_offloading_fptr_map_ty { |
| 111 | + int64_t host_ptr; // key |
| 112 | + int64_t tgt_ptr; // value |
| 113 | +} *__omp_offloading_fptr_map_p = 0; |
| 114 | + |
| 115 | +// Number of elements in __omp_offloading_fptr_map_p table. |
| 116 | +__attribute__((weak)) uint64_t __omp_offloading_fptr_map_size = 0; |
| 117 | +``` |
| 118 | +
|
| 119 | +`__tgt_rtl_set_function_ptr_map(int32_t device_id, uint64_t table_size, __omp_offloading_fptr_map_ty *table_host_ptr)` allocates device memory of size `sizeof(__omp_offloading_fptr_map_ty) * table_size`, and transfers the contents of `table_host_ptr` array into this device memory. An address of the allocated device memory area is then assigned to `__omp_offloading_fptr_map_p` global variables on the device. For example, in **CUDA**, a device address of `__omp_offloading_fptr_map_p` may be taken by calling `cuModuleGetGlobal`, and then a pointer-sized data transfer will initialize `__omp_offloading_fptr_map_p` to point to the device copy of `table_host_ptr` array. `__omp_offloading_fptr_map_size` is assigned to `table_size` the same way. |
| 120 | +
|
| 121 | +An alternative implementation of `__tgt_rtl_set_function_ptr_map` may invoke a device kernel that will do the assignments. |
| 122 | +
|
| 123 | +`__kmpc_target_translate_fptr(void *FnPtr)` API uses binary search to match `FnPtr` against `host_ptr` inside the device table pointed to by `__omp_offloading_fptr_map_p`. If the matching key is found, it returns the corresponding `tgt_ptr`, otherwise, it returns `FnPtr`. |
| 124 | +
|
| 125 | +# TODO: Optimization for non-unified_shared_memory |
| 126 | +
|
| 127 | +If a program does not use **required unified_shared_memory**, and all function pointers are mapped (not a requirement by OpenMP spec), then an implementation may avoid the runtime dispatch code for indirect function calls (i.e. `__kmpc_target_translate_fptr` is not needed) and also `__tgt_rtl_set_function_ptr_map` is not needed. `libomptarget` will just map the function pointers as regular data pointers via `Device.HostDataToTargetMap`. |
0 commit comments