Fixes and improvements from LLVM/XLC work. (#716)

Code fixes for XLC and Clang execution without build system changes.
This mainly adds missing OpenMP pragmas and makes cnrn_target_
wrappers visible to NMODL.

Author: olupton

CMake/MakefileBuildOptions.cmake

@@ -75,6 +75,7 @@ string(TOUPPER "${CMAKE_BUILD_TYPE}" _BUILD_TYPE)
 set(CORENRN_CXX_FLAGS
     "${CMAKE_CXX_FLAGS} ${CMAKE_CXX_FLAGS_${_BUILD_TYPE}} ${CXX14_STD_FLAGS} ${NVHPC_ACC_COMP_FLAGS} ${NVHPC_CXX_INLINE_FLAGS}"
 )
+set(CORENRN_LD_FLAGS "${NVHPC_ACC_LINK_FLAGS}")
 
 # =============================================================================
 # nmodl/mod2c related options : TODO

coreneuron/gpu/nrn_acc_manager.cpp

@@ -24,16 +24,14 @@
 #include "coreneuron/mpi/nrnmpidec.h"
 #include "coreneuron/utils/utils.hpp"
 
+#ifdef CRAYPAT
+#include <pat_api.h>
+#endif
+
 #ifdef _OPENACC
 #include <openacc.h>
 #endif
-#ifdef CORENEURON_PREFER_OPENMP_OFFLOAD
-#include <omp.h>
-#endif
 
-#ifdef CRAYPAT
-#include <pat_api.h>
-#endif
 namespace coreneuron {
 extern InterleaveInfo* interleave_info;
 void copy_ivoc_vect_to_device(const IvocVect& iv, IvocVect& div);
@@ -43,51 +41,6 @@ void nrn_ion_global_map_delete_from_device();
 void nrn_VecPlay_copyto_device(NrnThread* nt, void** d_vecplay);
 void nrn_VecPlay_delete_from_device(NrnThread* nt);
 
-template <typename T>
-T* cnrn_target_deviceptr(const T* h_ptr) {
-#if defined(CORENEURON_ENABLE_GPU) && !defined(CORENEURON_PREFER_OPENMP_OFFLOAD) && defined(_OPENACC)
-    return static_cast<T*>(acc_deviceptr(const_cast<T*>(h_ptr)));
-#elif defined(CORENEURON_ENABLE_GPU) && defined(CORENEURON_PREFER_OPENMP_OFFLOAD) && defined(_OPENMP)
-    return static_cast<T*>(omp_get_mapped_ptr(const_cast<T*>(h_ptr), omp_get_default_device()));
-#else
-    throw std::runtime_error("cnrn_target_deviceptr() not implemented without OpenACC/OpenMP and gpu build");
-#endif
-}
-
-template <typename T>
-T* cnrn_target_copyin(const T* h_ptr, std::size_t len = 1) {
-#if defined(CORENEURON_ENABLE_GPU) && !defined(CORENEURON_PREFER_OPENMP_OFFLOAD) && defined(_OPENACC)
-    return static_cast<T*>(acc_copyin(const_cast<T*>(h_ptr), len * sizeof(T)));
-#elif defined(CORENEURON_ENABLE_GPU) && defined(CORENEURON_PREFER_OPENMP_OFFLOAD) && defined(_OPENMP)
-    #pragma omp target enter data map(to:h_ptr[:len])
-    return cnrn_target_deviceptr(const_cast<T*>(h_ptr));
-#else
-    throw std::runtime_error("cnrn_target_copyin() not implemented without OpenACC/OpenMP and gpu build");
-#endif
-}
-
-template <typename T>
-void cnrn_target_delete(T* h_ptr, std::size_t len = 1) {
-#if defined(CORENEURON_ENABLE_GPU) && !defined(CORENEURON_PREFER_OPENMP_OFFLOAD) && defined(_OPENACC)
-    acc_delete(h_ptr, len * sizeof(T));
-#elif defined(CORENEURON_ENABLE_GPU) && defined(CORENEURON_PREFER_OPENMP_OFFLOAD) && defined(_OPENMP)
-    #pragma omp target exit data map(delete: h_ptr[:len])
-#else
-    throw std::runtime_error("cnrn_target_delete() not implemented without OpenACC/OpenMP and gpu build");
-#endif
-}
-
-template <typename T>
-void cnrn_target_memcpy_to_device(T* d_ptr, const T* h_ptr, std::size_t len = 1) {
-#if defined(CORENEURON_ENABLE_GPU) && !defined(CORENEURON_PREFER_OPENMP_OFFLOAD) && defined(_OPENACC)
-    acc_memcpy_to_device(d_ptr, const_cast<T*>(h_ptr), len * sizeof(T));
-#elif defined(CORENEURON_ENABLE_GPU) && defined(CORENEURON_PREFER_OPENMP_OFFLOAD) && defined(_OPENMP)
-    omp_target_memcpy(d_ptr, const_cast<T*>(h_ptr), len* sizeof(T), 0, 0, omp_get_default_device(), omp_get_initial_device());
-#else
-    throw std::runtime_error("cnrn_target_memcpy_to_device() not implemented without OpenACC/OpenMP and gpu build");
-#endif
-}
-
 /* note: threads here are corresponding to global nrn_threads array */
 void setup_nrnthreads_on_device(NrnThread* threads, int nthreads) {
 #ifdef _OPENACC

coreneuron/kinderiv.py

@@ -59,6 +59,9 @@ def write_out_kinderiv(fout):
     fout.write("\n/* declarations */\n")
     fout.write("\nnamespace coreneuron {\n")
 
+    if deriv or kin or euler:
+        fout.write('nrn_pragma_omp(declare target)\n')
+
     for item in deriv:
         fout.write('#pragma acc routine seq\n')
         fout.write('extern int %s%s(_threadargsproto_);\n' % (item[0], item[1]))
@@ -73,6 +76,9 @@ def write_out_kinderiv(fout):
         fout.write('#pragma acc routine seq\n')
         fout.write('extern int %s%s(_threadargsproto_);\n' % (item[0], item[1]))
 
+    if deriv or kin or euler:
+        fout.write('nrn_pragma_omp(end declare target)\n')
+
     fout.write("\n/* callback indices */\n")
     derivoffset = 1
     kinoffset = 1

coreneuron/mechanism/eion.cpp

@@ -177,6 +177,7 @@ double nrn_nernst(double ci, double co, double z, double celsius) {
     }
 }
 
+nrn_pragma_omp(declare target)
 void nrn_wrote_conc(int type,
                     double* p1,
                     int p2,
@@ -193,6 +194,7 @@ void nrn_wrote_conc(int type,
         pe[0] = nrn_nernst(pe[1 * _STRIDE], pe[2 * _STRIDE], gimap[type][2], celsius);
     }
 }
+nrn_pragma_omp(end declare target)
 
 static double efun(double x) {
     if (fabs(x) < 1e-4) {

coreneuron/mechanism/mech/dimplic.cpp

@@ -24,6 +24,7 @@
 #include "coreneuron/mechanism/mech/mod2c_core_thread.hpp"
 #include "_kinderiv.h"
 namespace coreneuron {
+nrn_pragma_omp(declare target)
 int derivimplicit_thread(int n, int* slist, int* dlist, DIFUN fun, _threadargsproto_) {
     difun(fun);
     return 0;
@@ -48,5 +49,6 @@ int nrn_kinetic_steer(int fun, SparseObj* so, double* rhs, _threadargsproto_) {
     switch (fun) { _NRN_KINETIC_CASES }
     return 0;
 }
+nrn_pragma_omp(end declare target)
 
 }  // namespace coreneuron

coreneuron/mechanism/register_mech.cpp

@@ -19,7 +19,9 @@
 
 namespace coreneuron {
 int secondorder = 0;
+nrn_pragma_omp(declare target)
 double t, dt, celsius, pi;
+nrn_pragma_omp(end declare target)
 int rev_dt;
 
 using Pfrv = void (*)();

coreneuron/network/cvodestb.cpp

@@ -86,12 +86,14 @@ void fixed_play_continuous(NrnThread* nt) {
 
 // NOTE : this implementation is duplicated in "coreneuron/mechanism/nrnoc_ml.ispc"
 // for the ISPC backend. If changes are required, make sure to change ISPC as well.
+nrn_pragma_omp(declare target)
 int at_time(NrnThread* nt, double te) {
     double x = te - 1e-11;
     if (x <= nt->_t && x > (nt->_t - nt->_dt)) {
         return 1;
     }
     return 0;
 }
+nrn_pragma_omp(end declare target)
 
 }  // namespace coreneuron

coreneuron/network/netcvode.cpp

@@ -537,7 +537,7 @@ void NetCvode::check_thresh(NrnThread* nt) {  // for default method
     nrn_pragma_acc(parallel loop present(
         nt [0:1], presyns_helper [0:nt->n_presyn], presyns [0:nt->n_presyn], actual_v [0:nt->end])
                        copy(net_send_buf_count) if (nt->compute_gpu) async(nt->stream_id))
-    nrn_pragma_omp(target teams distribute parallel for simd map(tofrom: net_send_buf_count) if(nt->compute_gpu))
+    nrn_pragma_omp(target teams distribute parallel for map(tofrom: net_send_buf_count) if(nt->compute_gpu))
     for (int i = 0; i < nt->ncell; ++i) {
         PreSyn* ps = presyns + i;
         PreSynHelper* psh = presyns_helper + i;

coreneuron/sim/scopmath/crout_thread.cpp

@@ -50,6 +50,7 @@ namespace coreneuron {
 #define ix(arg) ((arg) *_STRIDE)
 
 /* having a differnt permutation per instance may not be a good idea */
+nrn_pragma_omp(declare target)
 int nrn_crout_thread(NewtonSpace* ns, int n, double** a, int* perm, _threadargsproto_) {
     int save_i = 0;
 
@@ -224,4 +225,5 @@ void nrn_scopmath_solve_thread(int n,
         }
     }
 }
+nrn_pragma_omp(end declare target)
 }  // namespace coreneuron

coreneuron/sim/scopmath/newton_thread.cpp

@@ -59,6 +59,7 @@ namespace coreneuron {
 #define ix(arg) ((arg) *_STRIDE)
 #define s_(arg) _p[s[arg] * _STRIDE]
 
+nrn_pragma_omp(declare target)
 int nrn_newton_thread(NewtonSpace* ns,
                       int n,
                       int* s,
@@ -136,6 +137,7 @@ int nrn_newton_thread(NewtonSpace* ns,
 
     return (error);
 }
+nrn_pragma_omp(end declare target)
 
 /*------------------------------------------------------------*/
 /*                                                            */