include/corecel/DeviceRuntimeApi.hh

0001 //------------------------------- -*- C++ -*- -------------------------------//
0002 // Copyright Celeritas contributors: see top-level COPYRIGHT file for details
0003 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
0004 //---------------------------------------------------------------------------//
0005 //! \file corecel/DeviceRuntimeApi.hh
0006 //! \brief Include CUDA or HIP runtime APIs for compiling with host/cc
0007 //! compiler.
0008 //---------------------------------------------------------------------------//
0009 #pragma once
0010
0011 #include "corecel/Config.hh"
0012
0013 #if CELERITAS_USE_CUDA
0014 #    include <cuda_runtime_api.h>
0015 #elif CELERITAS_USE_HIP
0016 #    ifndef __HIPCC__
0017 /* Assume we're on an AMD system but not being invoked by the rocm compiler */
0018 #        define __HIP_PLATFORM_AMD__ 1
0019 #        define __HIP_PLATFORM_HCC__ 1
0020 #    endif
0021 #    include <hip/hip_runtime.h>
0022 #endif
0023
0024 /*!
0025  * \def CELER_DEVICE_PLATFORM
0026  *
0027  * API prefix token for the device offloading type.
0028  */
0029 /*!
0030  * \def CELER_DEVICE_API_SYMBOL
0031  *
0032  * Add a prefix "hip" or "cuda" to a code token.
0033  */
0034 /*!
0035  * \def CELER_EU_PER_CU
0036  *
0037  * Execution units per compute unit.  AMD multiprocessors each have 4 SIMD
0038  * units per compute unit, but there is no device attribute or
0039  * compiler definition that provides this information.
0040  * For CUDA, each streaming multiprocessor (MP) is a single "execution unit"
0041  * and a "compute unit".
0042  */
0043 #if CELERITAS_USE_CUDA
0044 #    define CELER_DEVICE_PLATFORM cuda
0045 #    define CELER_DEVICE_PLATFORM_UPPER_STR "CUDA"
0046 #    define CELER_DEVICE_API_SYMBOL(TOK) cuda##TOK
0047 #    define CELER_EU_PER_CU 1
0048 #elif CELERITAS_USE_HIP
0049 #    define CELER_DEVICE_PLATFORM hip
0050 #    define CELER_DEVICE_PLATFORM_UPPER_STR "HIP"
0051 #    define CELER_DEVICE_API_SYMBOL(TOK) hip##TOK
0052 #    if defined(__HIP_PLATFORM_AMD__) || defined(__HIP_PLATFORM_HCC__)
0053 #        define CELER_EU_PER_CU 4
0054 #    elif defined(__HIP_PLATFORM_NVIDIA__) || defined(__HIP_PLATFORM_NVCC__)
0055 #        define CELER_EU_PER_CU 1
0056 #    else
0057 #        define CELER_EU_PER_CU 0
0058 #    endif
0059 #else
0060 #    define CELER_DEVICE_PLATFORM none
0061 #    define CELER_DEVICE_PLATFORM_UPPER_STR ""
0062 #    define CELER_DEVICE_API_SYMBOL(TOK) void
0063 #    define CELER_EU_PER_CU 0
0064 #endif
0065
0066 /*!
0067  * This macro informs downstream Celeritas code (namely, Stream) that it's safe
0068  * to use types from the device APIs.
0069  */
0070 #define CELER_DEVICE_RUNTIME_INCLUDED
0071
0072 /*!
0073  * Declare a dummy variable for disabled \c CELER_DEVICE_API_CALL calls.
0074  *
0075  * With this declaration, the build will fail if this include is missing.
0076  * (Unfortunately, since the use of this symbol is embedded in a macro, IWYU
0077  * won't include this file automatically.)
0078  */
0079 extern int const CorecelDeviceRuntimeApiHh;