src/arch/cuda/hybridAPI/hapi_functions.h

   1 /*
   2  * This file is separate from hapi.h because the Hybrid API is included as part
   3  * of AMPI's extensions to the MPI standard, and certain global variable
   4  * privatization methods require the AMPI API to be exposed as function pointers
   5  * through a shim and loader mechanism that needs to list the entire set of
   6  * provided functions at multiple points in its implementation.
   7  *
   8  * See src/libs/ck-libs/ampi/ampi_functions.h for mandatory procedures.
   9  *
  10  * For ease of reading: AMPI_CUSTOM_FUNC(ReturnType, FunctionName, Parameters...)
  11  */
  12
  13 /******************** DEPRECATED ********************/
  14 // Create a hapiWorkRequest object for the user. The runtime manages the associated
  15 // memory, so the user only needs to set it up properly.
  16 AMPI_CUSTOM_FUNC(hapiWorkRequest*, hapiCreateWorkRequest, void)
  17
  18 /******************** DEPRECATED ********************/
  19 // Add a work request into the "queue". Currently all specified data transfers
  20 // and kernel execution are directly put into a CUDA stream.
  21 AMPI_CUSTOM_FUNC(void, hapiEnqueue, hapiWorkRequest* wr)
  22
  23 // The runtime queries the compute capability of the device, and creates as
  24 // many streams as the maximum number of concurrent kernels.
  25 AMPI_CUSTOM_FUNC(int, hapiCreateStreams, void)
  26
  27 // Get a CUDA stream that was created by the runtime. Current scheme is to
  28 // hand out streams in a round-robin fashion.
  29 AMPI_CUSTOM_FUNC(cudaStream_t, hapiGetStream, void)
  30
  31 // Add a Charm++ callback function to be invoked after the previous operation
  32 // in the stream completes. This call should be placed after data transfers or
  33 // a kernel invocation.
  34 AMPI_CUSTOM_FUNC(void, hapiAddCallback, cudaStream_t, void*, void*)
  35
  36 // Thin wrappers for memory related CUDA API calls.
  37 AMPI_CUSTOM_FUNC(cudaError_t, hapiMalloc, void**, size_t)
  38 AMPI_CUSTOM_FUNC(cudaError_t, hapiFree, void*)
  39 AMPI_CUSTOM_FUNC(cudaError_t, hapiMallocHost, void**, size_t)
  40 AMPI_CUSTOM_FUNC(cudaError_t, hapiFreeHost, void*)
  41 AMPI_CUSTOM_FUNC(cudaError_t, hapiMallocHostPool, void**, size_t)
  42 AMPI_CUSTOM_FUNC(cudaError_t, hapiFreeHostPool, void*)
  43 AMPI_CUSTOM_FUNC(cudaError_t, hapiMemcpyAsync, void*, const void*, size_t, enum cudaMemcpyKind, cudaStream_t)
  44
  45 // Explicit memory allocations using pinned memory pool.
  46 AMPI_CUSTOM_FUNC(void*, hapiPoolMalloc, size_t)
  47 AMPI_CUSTOM_FUNC(void, hapiPoolFree, void*)
  48
  49 // Provides support for detecting errors with CUDA API calls.
  50 AMPI_CUSTOM_FUNC(void, hapiErrorDie, cudaError_t, const char*, const char*, int)
  51
  52 #ifdef HAPI_INSTRUMENT_WRS
  53 AMPI_CUSTOM_FUNC(void, hapiInitInstrument, int n_chares, char n_types)
  54 AMPI_CUSTOM_FUNC(hapiRequestTimeInfo*, hapiQueryInstrument, int chare, char type, char phase)
  55 #endif