diff options
Diffstat (limited to 'libgomp/testsuite')
149 files changed, 9840 insertions, 44 deletions
diff --git a/libgomp/testsuite/lib/libgomp.exp b/libgomp/testsuite/lib/libgomp.exp index bc38e3c..fd475ac 100644 --- a/libgomp/testsuite/lib/libgomp.exp +++ b/libgomp/testsuite/lib/libgomp.exp @@ -30,6 +30,7 @@ load_gcc_lib scandump.exp load_gcc_lib scanlang.exp load_gcc_lib scanrtl.exp load_gcc_lib scansarif.exp +load_gcc_lib scanhtml.exp load_gcc_lib scantree.exp load_gcc_lib scanltrans.exp load_gcc_lib scanoffload.exp @@ -553,7 +554,23 @@ int main() { } } "-lcuda" ] } -# Return 1 if cublas_v2.h and -lcublas are available. +# Return 1 if -lcuda is available (header not required). + +proc check_effective_target_openacc_libcuda { } { + return [check_no_compiler_messages openacc_libcuda executable { +typedef enum { CUDA_SUCCESS } CUresult; +typedef int CUdevice; +CUresult cuDeviceGet (CUdevice *, int); +int main() { + CUdevice dev; + CUresult r = cuDeviceGet (&dev, 0); + if (r != CUDA_SUCCESS) + return 1; + return 0; +} } "-lcuda" ] +} + +# Return 1 if cublas_v2.h, cuda.h, -lcublas and -lcuda are available. proc check_effective_target_openacc_cublas { } { return [check_no_compiler_messages openacc_cublas executable { @@ -573,7 +590,25 @@ int main() { } } "-lcuda -lcublas" ] } -# Return 1 if cuda_runtime_api.h and -lcudart are available. +# Return 1 if -lcublas is available header not required). + +proc check_effective_target_openacc_libcublas { } { + return [check_no_compiler_messages openacc_libcublas executable { +typedef enum { CUBLAS_STATUS_SUCCESS } cublasStatus_t; +typedef struct cublasContext* cublasHandle_t; +#define cublasCreate cublasCreate_v2 +cublasStatus_t cublasCreate_v2 (cublasHandle_t *); +int main() { + cublasStatus_t s; + cublasHandle_t h; + s = cublasCreate (&h); + if (s != CUBLAS_STATUS_SUCCESS) + return 1; + return 0; +} } "-lcublas" ] +} + +# Return 1 if cuda_runtime_api.h, cuda.h, -lcuda and -lcudart are available. proc check_effective_target_openacc_cudart { } { return [check_no_compiler_messages openacc_cudart executable { @@ -592,3 +627,98 @@ int main() { return 0; } } "-lcuda -lcudart" ] } + +# Return 1 if -lcudart is available (no header required). + +proc check_effective_target_openacc_libcudart { } { + return [check_no_compiler_messages openacc_libcudart executable { +typedef int cudaError_t; +cudaError_t cudaGetDevice(int *); +enum { cudaSuccess }; +int main() { + cudaError_t e; + int devn; + e = cudaGetDevice (&devn); + if (e != cudaSuccess) + return 1; + return 0; +} } "-lcudart" ] +} + +# Return 1 if hip.h is available (no link check; AMD platform). + +proc check_effective_target_gomp_hip_header_amd { } { + return [check_no_compiler_messages gomp_hip_header_amd assembly { +#define __HIP_PLATFORM_AMD__ +#include <hip/hip_runtime_api.h> +int main() { + hipDevice_t dev; + hipError_t r = hipDeviceGet (&dev, 0); + if (r != hipSuccess) + return 1; + return 0; +} }] +} + +# Return 1 if hip.h is available (no link check; Nvidia/CUDA platform). + +proc check_effective_target_gomp_hip_header_nvidia { } { + return [check_no_compiler_messages gomp_hip_header_nvidia assembly { +#define __HIP_PLATFORM_NVIDIA__ +#include <hip/hip_runtime_api.h> +int main() { + hipDevice_t dev; + hipError_t r = hipDeviceGet (&dev, 0); + if (r != hipSuccess) + return 1; + return 0; +} } "-Wno-deprecated-declarations"] +} + +# Return 1 if the Fortran hipfort module is available (no link check) + +proc check_effective_target_gomp_hipfort_module { } { + return [check_no_compiler_messages gomp_hipfort_module assembly { +! Fortran +use hipfort +implicit none +integer(kind(hipSuccess)) :: r +integer(c_int) :: dev +r = hipDeviceGet (dev, 0) +if (r /= hipSuccess) error stop +end +}] +} + +# Return 1 if AMD HIP's -lamdhip64 is available (no header required). + +proc check_effective_target_gomp_libamdhip64 { } { + return [check_no_compiler_messages gomp_libamdhip64 executable { +typedef int hipError_t; +typedef int hipDevice_t; +enum { hipSuccess = 0 }; +hipError_t hipDeviceGet(hipDevice_t*, int); +int main() { + hipDevice_t dev; + hipError_t r = hipDeviceGet (&dev, 0); + if (r != hipSuccess) + return 1; + return 0; +} } "-lamdhip64" ] +} + +# Return 1 if AMD HIP's -lamdhip64 is available (no header required). + +proc check_effective_target_gomp_libhipblas { } { + return [check_no_compiler_messages gomp_libhipblas executable { +typedef enum { HIPBLAS_STATUS_SUCCESS = 0 } hipblasStatus_t; +typedef void* hipblasHandle_t; +hipblasStatus_t hipblasCreate (hipblasHandle_t*); +int main() { + hipblasHandle_t handle; + hipblasStatus_t stat = hipblasCreate (&handle); + if (stat != HIPBLAS_STATUS_SUCCESS) + return 1; + return 0; +} } "-lhipblas" ] +} diff --git a/libgomp/testsuite/libgomp.c++/allocator-1.C b/libgomp/testsuite/libgomp.c++/allocator-1.C index f820722..49425386 100644 --- a/libgomp/testsuite/libgomp.c++/allocator-1.C +++ b/libgomp/testsuite/libgomp.c++/allocator-1.C @@ -78,7 +78,10 @@ void test_inequality () CHECK_INEQUALITY (omp::allocator::cgroup_mem, void); CHECK_INEQUALITY (omp::allocator::pteam_mem, void); CHECK_INEQUALITY (omp::allocator::thread_mem, void); +#ifdef __gnu_linux__ + /* Pinning not implemented on other targets. */ CHECK_INEQUALITY (ompx::allocator::gnu_pinned_mem, void); +#endif /* And again with the same type passed to the allocator. */ CHECK_INEQUALITY (omp::allocator::null_allocator, T); CHECK_INEQUALITY (omp::allocator::default_mem, T); @@ -89,7 +92,9 @@ void test_inequality () CHECK_INEQUALITY (omp::allocator::cgroup_mem, T); CHECK_INEQUALITY (omp::allocator::pteam_mem, T); CHECK_INEQUALITY (omp::allocator::thread_mem, T); +#ifdef __gnu_linux__ CHECK_INEQUALITY (ompx::allocator::gnu_pinned_mem, T); +#endif } #undef CHECK_INEQUALITY @@ -121,7 +126,9 @@ int main () test<int, omp::allocator::cgroup_mem>(42); test<int, omp::allocator::pteam_mem>(42); test<int, omp::allocator::thread_mem>(42); +#ifdef __gnu_linux__ test<int, ompx::allocator::gnu_pinned_mem>(42); +#endif test<long long, omp::allocator::null_allocator>(42); test<long long, omp::allocator::default_mem>(42); @@ -132,7 +139,9 @@ int main () test<long long, omp::allocator::cgroup_mem>(42); test<long long, omp::allocator::pteam_mem>(42); test<long long, omp::allocator::thread_mem>(42); +#ifdef __gnu_linux__ test<long long, ompx::allocator::gnu_pinned_mem>(42); +#endif test<S, omp::allocator::null_allocator>( S{42, true, 128.f}); test<S, omp::allocator::default_mem>( S{42, true, 128.f}); @@ -143,7 +152,9 @@ int main () test<S, omp::allocator::cgroup_mem>( S{42, true, 128.f}); test<S, omp::allocator::pteam_mem>( S{42, true, 128.f}); test<S, omp::allocator::thread_mem>( S{42, true, 128.f}); +#ifdef __gnu_linux__ test<S, ompx::allocator::gnu_pinned_mem>(S{42, true, 128.f}); +#endif test_inequality<int, omp::allocator::null_allocator>(); test_inequality<int, omp::allocator::default_mem>(); @@ -154,5 +165,7 @@ int main () test_inequality<int, omp::allocator::cgroup_mem>(); test_inequality<int, omp::allocator::pteam_mem>(); test_inequality<int, omp::allocator::thread_mem>(); +#ifdef __gnu_linux__ test_inequality<int, ompx::allocator::gnu_pinned_mem>(); +#endif } diff --git a/libgomp/testsuite/libgomp.c++/allocator-2.C b/libgomp/testsuite/libgomp.c++/allocator-2.C index d25b755..ca94fc7 100644 --- a/libgomp/testsuite/libgomp.c++/allocator-2.C +++ b/libgomp/testsuite/libgomp.c++/allocator-2.C @@ -86,7 +86,10 @@ int main () f<int, omp::allocator::cgroup_mem >(0, 1, 2, 3); f<int, omp::allocator::pteam_mem >(0, 1, 2, 3); f<int, omp::allocator::thread_mem >(0, 1, 2, 3); +#ifdef __gnu_linux__ + /* Pinning not implemented on other targets. */ f<int, ompx::allocator::gnu_pinned_mem>(0, 1, 2, 3); +#endif f<long long, omp::allocator::null_allocator >(0, 1, 2, 3); f<long long, omp::allocator::default_mem >(0, 1, 2, 3); @@ -97,7 +100,9 @@ int main () f<long long, omp::allocator::cgroup_mem >(0, 1, 2, 3); f<long long, omp::allocator::pteam_mem >(0, 1, 2, 3); f<long long, omp::allocator::thread_mem >(0, 1, 2, 3); +#ifdef __gnu_linux__ f<long long, ompx::allocator::gnu_pinned_mem>(0, 1, 2, 3); +#endif S0 s0_0{ 42, true, 111128.f}; S0 s0_1{ 142, false, 11128.f}; @@ -112,7 +117,9 @@ int main () f<S0, omp::allocator::cgroup_mem >(s0_0, s0_1, s0_2, s0_3); f<S0, omp::allocator::pteam_mem >(s0_0, s0_1, s0_2, s0_3); f<S0, omp::allocator::thread_mem >(s0_0, s0_1, s0_2, s0_3); +#ifdef __gnu_linux__ f<S0, ompx::allocator::gnu_pinned_mem>(s0_0, s0_1, s0_2, s0_3); +#endif S1 s1_0{ 42, true, 111128.f}; S1 s1_1{ 142, false, 11128.f}; @@ -128,5 +135,7 @@ int main () f<S1, omp::allocator::cgroup_mem >(s1_0, s1_1, s1_2, s1_3); f<S1, omp::allocator::pteam_mem >(s1_0, s1_1, s1_2, s1_3); f<S1, omp::allocator::thread_mem >(s1_0, s1_1, s1_2, s1_3); +#ifdef __gnu_linux__ f<S1, ompx::allocator::gnu_pinned_mem>(s1_0, s1_1, s1_2, s1_3); +#endif } diff --git a/libgomp/testsuite/libgomp.c++/declare-mapper-1.C b/libgomp/testsuite/libgomp.c++/declare-mapper-1.C new file mode 100644 index 0000000..aba4f42 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare-mapper-1.C @@ -0,0 +1,87 @@ +// { dg-do run } + +#include <cstdlib> +#include <cassert> + +#define N 64 + +struct points +{ + double *x; + double *y; + double *z; + size_t len; +}; + +#pragma omp declare mapper(points p) map(to:p.x, p.y, p.z) \ + map(p.x[0:p.len]) \ + map(p.y[0:p.len]) \ + map(p.z[0:p.len]) + +struct shape +{ + points tmp; + points *pts; + int metadata[128]; +}; + +#pragma omp declare mapper(shape s) map(tofrom:s.pts, *s.pts) map(alloc:s.tmp) + +void +alloc_points (points *pts, size_t sz) +{ + pts->x = new double[sz]; + pts->y = new double[sz]; + pts->z = new double[sz]; + pts->len = sz; + for (int i = 0; i < sz; i++) + pts->x[i] = pts->y[i] = pts->z[i] = 0; +} + +int main (int argc, char *argv[]) +{ + shape myshape; + points mypts; + + myshape.pts = &mypts; + + alloc_points (&myshape.tmp, N); + myshape.pts = new points; + alloc_points (myshape.pts, N); + + #pragma omp target map(myshape) + { + for (int i = 0; i < N; i++) + { + myshape.pts->x[i]++; + myshape.pts->y[i]++; + myshape.pts->z[i]++; + } + } + + for (int i = 0; i < N; i++) + { + assert (myshape.pts->x[i] == 1); + assert (myshape.pts->y[i] == 1); + assert (myshape.pts->z[i] == 1); + } + + #pragma omp target + { + for (int i = 0; i < N; i++) + { + myshape.pts->x[i]++; + myshape.pts->y[i]++; + myshape.pts->z[i]++; + } + } + + for (int i = 0; i < N; i++) + { + assert (myshape.pts->x[i] == 2); + assert (myshape.pts->y[i] == 2); + assert (myshape.pts->z[i] == 2); + } + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/declare-mapper-2.C b/libgomp/testsuite/libgomp.c++/declare-mapper-2.C new file mode 100644 index 0000000..d848fdb --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare-mapper-2.C @@ -0,0 +1,55 @@ +// { dg-do run } + +#include <cassert> + +#define N 256 + +struct doublebuf +{ + int buf_a[N][N]; + int buf_b[N][N]; +}; + +#pragma omp declare mapper(lo:doublebuf b) map(b.buf_a[0:N/2][0:N]) \ + map(b.buf_b[0:N/2][0:N]) + +#pragma omp declare mapper(hi:doublebuf b) map(b.buf_a[N/2:N/2][0:N]) \ + map(b.buf_b[N/2:N/2][0:N]) + +int main (int argc, char *argv[]) +{ + doublebuf db; + + for (int i = 0; i < N; i++) + for (int j = 0; j < N; j++) + db.buf_a[i][j] = db.buf_b[i][j] = 0; + + #pragma omp target map(mapper(lo), tofrom:db) + { + for (int i = 0; i < N / 2; i++) + for (int j = 0; j < N; j++) + { + db.buf_a[i][j]++; + db.buf_b[i][j]++; + } + } + + #pragma omp target map(mapper(hi), tofrom:db) + { + for (int i = N / 2; i < N; i++) + for (int j = 0; j < N; j++) + { + db.buf_a[i][j]++; + db.buf_b[i][j]++; + } + } + + for (int i = 0; i < N; i++) + for (int j = 0; j < N; j++) + { + assert (db.buf_a[i][j] == 1); + assert (db.buf_b[i][j] == 1); + } + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/declare-mapper-3.C b/libgomp/testsuite/libgomp.c++/declare-mapper-3.C new file mode 100644 index 0000000..ea9b7de --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare-mapper-3.C @@ -0,0 +1,63 @@ +// { dg-do run } + +#include <cstdlib> +#include <cassert> + +struct S { + int *myarr; +}; + +#pragma omp declare mapper (S s) map(to:s.myarr) map (tofrom: s.myarr[0:20]) + +namespace A { +#pragma omp declare mapper (S s) map(to:s.myarr) map (tofrom: s.myarr[0:100]) +} + +namespace B { +#pragma omp declare mapper (S s) map(to:s.myarr) map (tofrom: s.myarr[100:100]) +} + +namespace A +{ + void incr_a (S my_s) + { +#pragma omp target + { + for (int i = 0; i < 100; i++) + my_s.myarr[i]++; + } + } +} + +namespace B +{ + void incr_b (S my_s) + { +#pragma omp target + { + for (int i = 100; i < 200; i++) + my_s.myarr[i]++; + } + } +} + +int main (int argc, char *argv[]) +{ + S my_s; + + my_s.myarr = (int *) calloc (200, sizeof (int)); + +#pragma omp target + { + for (int i = 0; i < 20; i++) + my_s.myarr[i]++; + } + + A::incr_a (my_s); + B::incr_b (my_s); + + for (int i = 0; i < 200; i++) + assert (my_s.myarr[i] == (i < 20) ? 2 : 1); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/declare-mapper-4.C b/libgomp/testsuite/libgomp.c++/declare-mapper-4.C new file mode 100644 index 0000000..f194e63 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare-mapper-4.C @@ -0,0 +1,63 @@ +// { dg-do run } + +#include <cstdlib> +#include <cassert> + +struct S { + int *myarr; +}; + +#pragma omp declare mapper (S s) map(to:s.myarr) map (tofrom: s.myarr[0:20]) + +namespace A { +#pragma omp declare mapper (S s) map(to:s.myarr) map (tofrom: s.myarr[0:100]) +} + +namespace B { +#pragma omp declare mapper (S s) map(to:s.myarr) map (tofrom: s.myarr[100:100]) +} + +namespace A +{ + void incr_a (S &my_s) + { +#pragma omp target + { + for (int i = 0; i < 100; i++) + my_s.myarr[i]++; + } + } +} + +namespace B +{ + void incr_b (S &my_s) + { +#pragma omp target + { + for (int i = 100; i < 200; i++) + my_s.myarr[i]++; + } + } +} + +int main (int argc, char *argv[]) +{ + S my_s; + + my_s.myarr = (int *) calloc (200, sizeof (int)); + +#pragma omp target + { + for (int i = 0; i < 20; i++) + my_s.myarr[i]++; + } + + A::incr_a (my_s); + B::incr_b (my_s); + + for (int i = 0; i < 200; i++) + assert (my_s.myarr[i] == (i < 20) ? 2 : 1); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/declare-mapper-5.C b/libgomp/testsuite/libgomp.c++/declare-mapper-5.C new file mode 100644 index 0000000..0030de8 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare-mapper-5.C @@ -0,0 +1,52 @@ +// { dg-do run } + +#include <cassert> + +struct S +{ + int *myarr; + int len; +}; + +class C +{ + S smemb; +#pragma omp declare mapper (custom:S s) map(to:s.myarr) \ + map(tofrom:s.myarr[0:s.len]) + +public: + C(int l) + { + smemb.myarr = new int[l]; + smemb.len = l; + for (int i = 0; i < l; i++) + smemb.myarr[i] = 0; + } + void bump(); + void check(); +}; + +void +C::bump () +{ +#pragma omp target map(mapper(custom), tofrom: smemb) + { + for (int i = 0; i < smemb.len; i++) + smemb.myarr[i]++; + } +} + +void +C::check () +{ + for (int i = 0; i < smemb.len; i++) + assert (smemb.myarr[i] == 1); +} + +int main (int argc, char *argv[]) +{ + C test (100); + test.bump (); + test.check (); + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/declare-mapper-6.C b/libgomp/testsuite/libgomp.c++/declare-mapper-6.C new file mode 100644 index 0000000..14ed10d --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare-mapper-6.C @@ -0,0 +1,37 @@ +// { dg-do run } + +#include <cassert> + +template <typename T> +void adjust (T param) +{ +#pragma omp declare mapper (T x) map(to:x.len, x.base) \ + map(tofrom:x.base[0:x.len]) + +#pragma omp target + for (int i = 0; i < param.len; i++) + param.base[i]++; +} + +struct S { + int len; + int *base; +}; + +int main (int argc, char *argv[]) +{ + S a; + + a.len = 100; + a.base = new int[a.len]; + + for (int i = 0; i < a.len; i++) + a.base[i] = 0; + + adjust (a); + + for (int i = 0; i < a.len; i++) + assert (a.base[i] == 1); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/declare-mapper-7.C b/libgomp/testsuite/libgomp.c++/declare-mapper-7.C new file mode 100644 index 0000000..ba4792a --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare-mapper-7.C @@ -0,0 +1,59 @@ +// { dg-do run } + +#include <cassert> + +struct S +{ + int *myarr; +}; + +struct T +{ + S *s; +}; + +#pragma omp declare mapper (s100: S x) map(to: x.myarr) \ + map(tofrom: x.myarr[0:100]) +// Define this because ... +#pragma omp declare mapper (default: S x) map(to: x.myarr) \ + map(tofrom: x.myarr[0:100]) + + +void +bump (T t) +{ + /* Here we have an implicit/default mapper invoking a named mapper. We + need to make sure that can be located properly at gimplification + time. */ + +// ... the following is invalid in OpenMP - albeit supported by GCC +// (after disabling: error: in ‘declare mapper’ directives, parameter to ‘mapper’ modifier must be ‘default’ ) + +// #pragma omp declare mapper (T t) map(to:t.s) map(mapper(s100), tofrom: t.s[0]) + +// ... thus, we now use ... +#pragma omp declare mapper (T t) map(to:t.s) map(mapper(default), tofrom: t.s[0]) + +#pragma omp target + for (int i = 0; i < 100; i++) + t.s->myarr[i]++; +} + +int main (int argc, char *argv[]) +{ + S my_s; + T my_t; + + my_s.myarr = new int[100]; + my_t.s = &my_s; + + for (int i = 0; i < 100; i++) + my_s.myarr[i] = 0; + + bump (my_t); + + for (int i = 0; i < 100; i++) + assert (my_s.myarr[i] == 1); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/declare-mapper-8.C b/libgomp/testsuite/libgomp.c++/declare-mapper-8.C new file mode 100644 index 0000000..3818e52 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare-mapper-8.C @@ -0,0 +1,61 @@ +// { dg-do run } + +#include <cassert> + +struct S +{ + int *myarr; + int len; +}; + +template<typename T> +class C +{ + T memb; +#pragma omp declare mapper (T t) map(to:t.len, t.myarr) \ + map(tofrom:t.myarr[0:t.len]) + +public: + C(int sz); + ~C(); + void bump(); + void check(); +}; + +template<typename T> +C<T>::C(int sz) +{ + memb.myarr = new int[sz]; + for (int i = 0; i < sz; i++) + memb.myarr[i] = 0; + memb.len = sz; +} + +template<typename T> +C<T>::~C() +{ + delete[] memb.myarr; +} + +template<typename T> +void C<T>::bump() +{ +#pragma omp target map(memb) + for (int i = 0; i < memb.len; i++) + memb.myarr[i]++; +} + +template<typename T> +void C<T>::check() +{ + for (int i = 0; i < memb.len; i++) + assert (memb.myarr[i] == 1); +} + +int main(int argc, char *argv[]) +{ + C<S> c_int(100); + c_int.bump(); + c_int.check(); + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/declare_target-2.C b/libgomp/testsuite/libgomp.c++/declare_target-2.C new file mode 100644 index 0000000..ab94a55 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/declare_target-2.C @@ -0,0 +1,25 @@ +// { dg-do link } + +// Actually not needed: -fipa-cp is default with -O2: +// { dg-additional-options "-O2 -fipa-cp" } + +// The code failed because 'std::endl' becoḿes implicitly 'declare target' +// but not the 'widen' function it calls. While the linker had no issues +// (endl is never called, either because it is inlined or optimized away), +// the IPA-CP (enabled by -O2 and higher) failed as the definition for +// 'widen' did not exist on the offload side. + +#include <iostream> + +void func (int m) +{ + if (m < 0) + std::cout << "should not happen" << std::endl; +} + + +int main() +{ + #pragma omp target + func (1); +} diff --git a/libgomp/testsuite/libgomp.c++/pr106445-1-O0.C b/libgomp/testsuite/libgomp.c++/pr106445-1-O0.C new file mode 100644 index 0000000..bcd499c --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/pr106445-1-O0.C @@ -0,0 +1,3 @@ +// { dg-additional-options -O0 } + +#include "pr106445-1.C" diff --git a/libgomp/testsuite/libgomp.c++/pr106445-1.C b/libgomp/testsuite/libgomp.c++/pr106445-1.C new file mode 100644 index 0000000..329ce62 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/pr106445-1.C @@ -0,0 +1,18 @@ +#include <vector> + +int main() +{ +#pragma omp target + { + { + std::vector<int> v; + if (!v.empty()) + __builtin_abort(); + } + { + std::vector<int> v(100); + if (v.capacity() < 100) + __builtin_abort(); + } + } +} diff --git a/libgomp/testsuite/libgomp.c++/pr119692-1-4.C b/libgomp/testsuite/libgomp.c++/pr119692-1-4.C index 6995f26..af9fe1c 100644 --- a/libgomp/testsuite/libgomp.c++/pr119692-1-4.C +++ b/libgomp/testsuite/libgomp.c++/pr119692-1-4.C @@ -3,6 +3,9 @@ /* { dg-additional-options -DDEFAULT=defaultmap(firstprivate) } Wrong code for offloading execution. { dg-xfail-run-if PR119692 { offload_device } } */ +/* There are configurations where we 'WARNING: program timed out.' while in + 'dynamic_cast', see <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119692#c6>. + { dg-timeout 10 } ... to make sure that happens quickly. */ /* { dg-additional-options -fdump-tree-gimple } */ #include "pr119692-1-1.C" diff --git a/libgomp/testsuite/libgomp.c++/pr119692-1-5.C b/libgomp/testsuite/libgomp.c++/pr119692-1-5.C index 02121b6..e5c6e07 100644 --- a/libgomp/testsuite/libgomp.c++/pr119692-1-5.C +++ b/libgomp/testsuite/libgomp.c++/pr119692-1-5.C @@ -3,6 +3,9 @@ /* { dg-additional-options -DDEFAULT=defaultmap(to) } Wrong code for offloading execution. { dg-xfail-run-if PR119692 { offload_device } } */ +/* There are configurations where we 'WARNING: program timed out.' while in + 'dynamic_cast', see <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119692#c6>. + { dg-timeout 10 } ... to make sure that happens quickly. */ /* { dg-additional-options -fdump-tree-gimple } */ #include "pr119692-1-1.C" diff --git a/libgomp/testsuite/libgomp.c++/pr96390.C b/libgomp/testsuite/libgomp.c++/pr96390.C index 1f3c3e0..be19601 100644 --- a/libgomp/testsuite/libgomp.c++/pr96390.C +++ b/libgomp/testsuite/libgomp.c++/pr96390.C @@ -1,6 +1,4 @@ /* { dg-additional-options "-O0 -fdump-tree-omplower" } */ -/* { dg-additional-options "-foffload=-Wa,--verify" { target offload_target_nvptx } } */ -/* { dg-xfail-if "PR 97106/PR 97102 - .alias not (yet) supported for nvptx" { offload_target_nvptx } } */ #include <cstdlib> #include <type_traits> diff --git a/libgomp/testsuite/libgomp.c++/target-cdtor-1.C b/libgomp/testsuite/libgomp.c++/target-cdtor-1.C new file mode 100644 index 0000000..ecb029e --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-cdtor-1.C @@ -0,0 +1,104 @@ +/* Offloaded C++ objects construction and destruction. */ + +/* { dg-additional-options -fdump-tree-optimized-raw-asmname } + { dg-additional-options -foffload-options=-fdump-tree-optimized-raw-asmname } */ + +#include <omp.h> +#include <vector> + +#pragma omp declare target + +struct S +{ + int x; + + S() + : x(-1) + { + __builtin_printf("%s, %d, %d\n", __FUNCTION__, x, omp_is_initial_device()); + } + S(int x) + : x(x) + { + __builtin_printf("%s, %d, %d\n", __FUNCTION__, x, omp_is_initial_device()); + } + ~S() + { + __builtin_printf("%s, %d, %d\n", __FUNCTION__, x, omp_is_initial_device()); + } +}; + +#pragma omp end declare target + +S sH1(7); + +#pragma omp declare target + +S sHD1(5); + +std::vector<S> svHD1(2); + +#pragma omp end declare target + +S sH2(3); + +int main() +{ + int c = 0; + + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + +#pragma omp target map(c) + { + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + } + +#pragma omp target map(c) + { + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + } + + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + + return 0; +} + +/* Verify '__cxa_atexit' calls. + + For the host, there are four expected calls: + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, } 4 optimized { target cxa_atexit } } } + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZN1SD1Ev, \&sH1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZN1SD1Ev, \&sHD1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZNSt6vectorI1SSaIS0_EED1Ev, \&svHD1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZN1SD1Ev, \&sH2, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + + For the device, there are two expected calls: + { dg-final { scan-offload-tree-dump-times {gimple_call <__cxa_atexit, } 2 optimized { target cxa_atexit } } } + { dg-final { scan-offload-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZN1SD1Ev, \&sHD1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + { dg-final { scan-offload-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZNSt6vectorI1SSaIS0_EED1Ev, \&svHD1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } +*/ + +/* C++ objects are constructed in order of appearance (..., and destructed in reverse order). + { dg-output {S, 7, 1[\r\n]+} } + { dg-output {S, 5, 1[\r\n]+} } + { dg-output {S, -1, 1[\r\n]+} } + { dg-output {S, -1, 1[\r\n]+} } + { dg-output {S, 3, 1[\r\n]+} } + { dg-output {main:1, 1[\r\n]+} } + { dg-output {S, 5, 0[\r\n]+} { target offload_device } } + { dg-output {S, -1, 0[\r\n]+} { target offload_device } } + { dg-output {S, -1, 0[\r\n]+} { target offload_device } } + { dg-output {main:2, 1[\r\n]+} { target { ! offload_device } } } + { dg-output {main:2, 0[\r\n]+} { target offload_device } } + { dg-output {main:3, 1[\r\n]+} { target { ! offload_device } } } + { dg-output {main:3, 0[\r\n]+} { target offload_device } } + { dg-output {main:4, 1[\r\n]+} } + { dg-output {~S, -1, 0[\r\n]+} { target offload_device } } + { dg-output {~S, -1, 0[\r\n]+} { target offload_device } } + { dg-output {~S, 5, 0[\r\n]+} { target offload_device } } + { dg-output {~S, 3, 1[\r\n]+} } + { dg-output {~S, -1, 1[\r\n]+} } + { dg-output {~S, -1, 1[\r\n]+} } + { dg-output {~S, 5, 1[\r\n]+} } + { dg-output {~S, 7, 1[\r\n]+} } +*/ diff --git a/libgomp/testsuite/libgomp.c++/target-cdtor-2.C b/libgomp/testsuite/libgomp.c++/target-cdtor-2.C new file mode 100644 index 0000000..75e48ca --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-cdtor-2.C @@ -0,0 +1,140 @@ +/* Offloaded 'constructor' and 'destructor' functions, and C++ objects construction and destruction. */ + +/* { dg-require-effective-target init_priority } */ + +/* { dg-additional-options -fdump-tree-optimized-raw-asmname } + { dg-additional-options -foffload-options=-fdump-tree-optimized-raw-asmname } */ + +#include <omp.h> +#include <vector> + +#pragma omp declare target + +struct S +{ + int x; + + S() + : x(-1) + { + __builtin_printf("%s, %d, %d\n", __FUNCTION__, x, omp_is_initial_device()); + } + S(int x) + : x(x) + { + __builtin_printf("%s, %d, %d\n", __FUNCTION__, x, omp_is_initial_device()); + } + ~S() + { + __builtin_printf("%s, %d, %d\n", __FUNCTION__, x, omp_is_initial_device()); + } +}; + +#pragma omp end declare target + +S sH1 __attribute__((init_priority(1500))) (7); + +#pragma omp declare target + +S sHD1 __attribute__((init_priority(2000))) (5); + +std::vector<S> svHD1 __attribute__((init_priority(1000))) (2); + +static void +__attribute__((constructor(20000))) +initDH1() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +static void +__attribute__((destructor(20000))) +finiDH1() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +#pragma omp end declare target + +S sH2 __attribute__((init_priority(500))) (3); + +static void +__attribute__((constructor(10000))) +initH1() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +static void +__attribute__((destructor(10000))) +finiH1() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +int main() +{ + int c = 0; + + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + +#pragma omp target map(c) + { + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + } + +#pragma omp target map(c) + { + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + } + + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + + return 0; +} + +/* Verify '__cxa_atexit' calls. + + For the host, there are four expected calls: + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, } 4 optimized { target cxa_atexit } } } + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZN1SD1Ev, \&sH1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZN1SD1Ev, \&sHD1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZNSt6vectorI1SSaIS0_EED1Ev, \&svHD1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + { dg-final { scan-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZN1SD1Ev, \&sH2, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + + For the device, there are two expected calls: + { dg-final { scan-offload-tree-dump-times {gimple_call <__cxa_atexit, } 2 optimized { target cxa_atexit } } } + { dg-final { scan-offload-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZN1SD1Ev, \&sHD1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } + { dg-final { scan-offload-tree-dump-times {gimple_call <__cxa_atexit, NULL, _ZNSt6vectorI1SSaIS0_EED1Ev, \&svHD1, \&__dso_handle>} 1 optimized { target cxa_atexit } } } +*/ + +/* Defined order in which 'constructor' functions, and 'destructor' functions are run, and C++ objects are constructed (..., and destructed in reverse order). + { dg-output {S, 3, 1[\r\n]+} } + { dg-output {S, -1, 1[\r\n]+} } + { dg-output {S, -1, 1[\r\n]+} } + { dg-output {S, 7, 1[\r\n]+} } + { dg-output {S, 5, 1[\r\n]+} } + { dg-output {initH1, 1[\r\n]+} } + { dg-output {initDH1, 1[\r\n]+} } + { dg-output {main:1, 1[\r\n]+} } + { dg-output {S, -1, 0[\r\n]+} { target offload_device } } + { dg-output {S, -1, 0[\r\n]+} { target offload_device } } + { dg-output {S, 5, 0[\r\n]+} { target offload_device } } + { dg-output {initDH1, 0[\r\n]+} { target offload_device } } + { dg-output {main:2, 1[\r\n]+} { target { ! offload_device } } } + { dg-output {main:2, 0[\r\n]+} { target offload_device } } + { dg-output {main:3, 1[\r\n]+} { target { ! offload_device } } } + { dg-output {main:3, 0[\r\n]+} { target offload_device } } + { dg-output {main:4, 1[\r\n]+} } + { dg-output {~S, 5, 0[\r\n]+} { target offload_device } } + { dg-output {~S, -1, 0[\r\n]+} { target offload_device } } + { dg-output {~S, -1, 0[\r\n]+} { target offload_device } } + { dg-output {finiDH1, 0[\r\n]+} { target offload_device } } + { dg-output {~S, 5, 1[\r\n]+} } + { dg-output {~S, 7, 1[\r\n]+} } + { dg-output {~S, -1, 1[\r\n]+} } + { dg-output {~S, -1, 1[\r\n]+} } + { dg-output {~S, 3, 1[\r\n]+} } + { dg-output {finiDH1, 1[\r\n]+} } + { dg-output {finiH1, 1[\r\n]+} } +*/ diff --git a/libgomp/testsuite/libgomp.c++/target-exceptions-bad_cast-1.C b/libgomp/testsuite/libgomp.c++/target-exceptions-bad_cast-1.C index 3848295..a862652 100644 --- a/libgomp/testsuite/libgomp.c++/target-exceptions-bad_cast-1.C +++ b/libgomp/testsuite/libgomp.c++/target-exceptions-bad_cast-1.C @@ -23,3 +23,6 @@ PR119692. { dg-shouldfail {'std::bad_cast' exception} } */ +/* There are configurations where we 'WARNING: program timed out.' while in + 'dynamic_cast', see <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119692#c6>. + { dg-timeout 10 } ... to make sure that happens quickly. */ diff --git a/libgomp/testsuite/libgomp.c++/target-exceptions-bad_cast-2.C b/libgomp/testsuite/libgomp.c++/target-exceptions-bad_cast-2.C index 8861740..ff15c9f 100644 --- a/libgomp/testsuite/libgomp.c++/target-exceptions-bad_cast-2.C +++ b/libgomp/testsuite/libgomp.c++/target-exceptions-bad_cast-2.C @@ -22,3 +22,6 @@ For GCN, nvptx offload execution, there is no 'catch'ing; any exception is fatal. { dg-shouldfail {'MyException' exception} { offload_device } } */ +/* There are configurations where we 'WARNING: program timed out.' while in + 'dynamic_cast', see <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119692#c6>. + { dg-timeout 10 } ... to make sure that happens quickly. */ diff --git a/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-GCN.C b/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-GCN.C index 3cdedf4..d4dccf1 100644 --- a/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-GCN.C +++ b/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-GCN.C @@ -14,8 +14,10 @@ /* In this specific C++ arrangement, distilled from PR118794, GCC synthesizes '__builtin_eh_pointer', '__builtin_unwind_resume' calls as dead code in 'f': - { dg-final { scan-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized } } - { dg-final { scan-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized } } + { dg-final { scan-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized { target { ! { arm_eabi || tic6x-*-* } } } } } + { dg-final { scan-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized { target { ! { arm_eabi || tic6x-*-* } } } } } + ..., just 'targetm.arm_eabi_unwinder' is different: + { dg-final { scan-tree-dump-times {gimple_call <__builtin_cxa_end_cleanup, } 1 optimized { target { arm_eabi || tic6x-*-* } } } } { dg-final { only_for_offload_target amdgcn-amdhsa scan-offload-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized } } { dg-final { only_for_offload_target amdgcn-amdhsa scan-offload-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized } } Given '-O0' and '-foffload-options=-mno-fake-exceptions', offload compilation fails: diff --git a/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-nvptx.C b/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-nvptx.C index ef996cf..724e34b 100644 --- a/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-nvptx.C +++ b/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1-offload-sorry-nvptx.C @@ -14,8 +14,10 @@ /* In this specific C++ arrangement, distilled from PR118794, GCC synthesizes '__builtin_eh_pointer', '__builtin_unwind_resume' calls as dead code in 'f': - { dg-final { scan-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized } } - { dg-final { scan-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized } } + { dg-final { scan-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized { target { ! { arm_eabi || tic6x-*-* } } } } } + { dg-final { scan-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized { target { ! { arm_eabi || tic6x-*-* } } } } } + ..., just 'targetm.arm_eabi_unwinder' is different: + { dg-final { scan-tree-dump-times {gimple_call <__builtin_cxa_end_cleanup, } 1 optimized { target { arm_eabi || tic6x-*-* } } } } { dg-final { only_for_offload_target nvptx-none scan-offload-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized } } { dg-final { only_for_offload_target nvptx-none scan-offload-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized } } Given '-O0' and '-foffload-options=-mno-fake-exceptions', offload compilation fails: diff --git a/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1.C b/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1.C index a73e7f8..24eb7a5 100644 --- a/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1.C +++ b/libgomp/testsuite/libgomp.c++/target-exceptions-pr118794-1.C @@ -9,10 +9,6 @@ /* See also '../../../gcc/testsuite/g++.target/gcn/exceptions-pr118794-1.C', '../../../gcc/testsuite/g++.target/nvptx/exceptions-pr118794-1.C'. */ -/* Help nvptx offloading overcome a code generation issue; - PR106445, PR118518. */ -#define ALWAYS_INLINE __attribute__((always_inline)) - #pragma omp begin declare target bool ok = false; @@ -20,12 +16,10 @@ bool ok = false; template <typename T> struct C { - ALWAYS_INLINE C() { ok = true; } - ALWAYS_INLINE C(int) {}; ~C() {}; @@ -57,7 +51,9 @@ int main() /* In this specific C++ arrangement, distilled from PR118794, GCC synthesizes '__builtin_eh_pointer', '__builtin_unwind_resume' calls as dead code in 'f': - { dg-final { scan-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized } } - { dg-final { scan-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized } } + { dg-final { scan-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized { target { ! { arm_eabi || tic6x-*-* } } } } } + { dg-final { scan-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized { target { ! { arm_eabi || tic6x-*-* } } } } } + ..., just 'targetm.arm_eabi_unwinder' is different: + { dg-final { scan-tree-dump-times {gimple_call <__builtin_cxa_end_cleanup, } 1 optimized { target { arm_eabi || tic6x-*-* } } } } { dg-final { scan-offload-tree-dump-times {gimple_call <__builtin_eh_pointer, } 1 optimized } } { dg-final { scan-offload-tree-dump-times {gimple_call <__builtin_unwind_resume, } 1 optimized } } */ diff --git a/libgomp/testsuite/libgomp.c++/target-exceptions-throw-1.C b/libgomp/testsuite/libgomp.c++/target-exceptions-throw-1.C index 2467061..a4e7a10 100644 --- a/libgomp/testsuite/libgomp.c++/target-exceptions-throw-1.C +++ b/libgomp/testsuite/libgomp.c++/target-exceptions-throw-1.C @@ -4,9 +4,6 @@ { dg-additional-options -fexceptions } */ /* { dg-additional-options -fdump-tree-optimized-raw } { dg-additional-options -foffload-options=-fdump-tree-optimized-raw } */ -/* { dg-bogus {Size expression must be absolute\.} PR119737 { target offload_target_amdgcn xfail *-*-* } 0 } - { dg-ice PR119737 { offload_target_amdgcn } } - { dg-excess-errors {'mkoffload' failures etc.} { xfail offload_target_amdgcn } } */ #include "../libgomp.oacc-c++/exceptions-throw-1.C" diff --git a/libgomp/testsuite/libgomp.c++/target-exceptions-throw-2.C b/libgomp/testsuite/libgomp.c++/target-exceptions-throw-2.C index e85e6c3..97f4845 100644 --- a/libgomp/testsuite/libgomp.c++/target-exceptions-throw-2.C +++ b/libgomp/testsuite/libgomp.c++/target-exceptions-throw-2.C @@ -4,9 +4,6 @@ { dg-additional-options -fexceptions } */ /* { dg-additional-options -fdump-tree-optimized-raw } { dg-additional-options -foffload-options=-fdump-tree-optimized-raw } */ -/* { dg-bogus {Size expression must be absolute\.} PR119737 { target offload_target_amdgcn xfail *-*-* } 0 } - { dg-ice PR119737 { offload_target_amdgcn } } - { dg-excess-errors {'mkoffload' failures etc.} { xfail offload_target_amdgcn } } */ #include "../libgomp.oacc-c++/exceptions-throw-2.C" diff --git a/libgomp/testsuite/libgomp.c++/target-flex-10.C b/libgomp/testsuite/libgomp.c++/target-flex-10.C new file mode 100644 index 0000000..8fa9af7 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-10.C @@ -0,0 +1,215 @@ +/* Basic container usage. */ + +#include <vector> +#include <deque> +#include <list> +#include <set> +#include <map> +#if __cplusplus >= 201103L +#include <array> +#include <forward_list> +#include <unordered_set> +#include <unordered_map> +#endif + +bool vector_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::vector<int> vector; + ok = vector.empty(); + } + return ok; +} + +bool deque_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::deque<int> deque; + ok = deque.empty(); + } + return ok; +} + +bool list_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::list<int> list; + ok = list.empty(); + } + return ok; +} + +bool map_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::map<int, int> map; + ok = map.empty(); + } + return ok; +} + +bool set_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::set<int> set; + ok = set.empty(); + } + return ok; +} + +bool multimap_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::multimap<int, int> multimap; + ok = multimap.empty(); + } + return ok; +} + +bool multiset_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::multiset<int, int> multiset; + ok = multiset.empty(); + } + return ok; +} + +#if __cplusplus >= 201103L + +bool array_test() +{ + static constexpr std::size_t array_size = 42; + bool ok; + #pragma omp target map(from: ok) + { + std::array<int, array_size> array{}; + ok = array[0] == 0 + && array[array_size - 1] == 0; + } + return ok; +} + +bool forward_list_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::forward_list<int> forward_list; + ok = forward_list.empty(); + } + return ok; +} + +bool unordered_map_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::unordered_map<int, int> unordered_map; + ok = unordered_map.empty(); + } + return ok; +} + +bool unordered_set_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + std::unordered_set<int> unordered_set; + ok = unordered_set.empty(); + } + return ok; +} + +bool unordered_multimap_test() +{ + + bool ok; + #pragma omp target map(from: ok) + { + std::unordered_multimap<int, int> unordered_multimap; + ok = unordered_multimap.empty(); + } + return ok; +} + +bool unordered_multiset_test() +{ + + bool ok; + #pragma omp target map(from: ok) + { + std::unordered_multiset<int> unordered_multiset; + ok = unordered_multiset.empty(); + } + return ok; +} + +#else +bool array_test() { return true; } +bool forward_list_test() { return true; } +bool unordered_map_test() { return true; } +bool unordered_set_test() { return true; } +bool unordered_multimap_test() { return true; } +bool unordered_multiset_test() { return true; } +#endif + +int main() +{ + const bool vec_res = vector_test(); + __builtin_printf("vector : %s\n", vec_res ? "PASS" : "FAIL"); + const bool deque_res = deque_test(); + __builtin_printf("deque : %s\n", deque_res ? "PASS" : "FAIL"); + const bool list_res = list_test(); + __builtin_printf("list : %s\n", list_res ? "PASS" : "FAIL"); + const bool map_res = map_test(); + __builtin_printf("map : %s\n", map_res ? "PASS" : "FAIL"); + const bool set_res = set_test(); + __builtin_printf("set : %s\n", set_res ? "PASS" : "FAIL"); + const bool multimap_res = multimap_test(); + __builtin_printf("multimap : %s\n", multimap_res ? "PASS" : "FAIL"); + const bool multiset_res = multiset_test(); + __builtin_printf("multiset : %s\n", multiset_res ? "PASS" : "FAIL"); + const bool array_res = array_test(); + __builtin_printf("array : %s\n", array_res ? "PASS" : "FAIL"); + const bool forward_list_res = forward_list_test(); + __builtin_printf("forward_list : %s\n", forward_list_res ? "PASS" : "FAIL"); + const bool unordered_map_res = unordered_map_test(); + __builtin_printf("unordered_map : %s\n", unordered_map_res ? "PASS" : "FAIL"); + const bool unordered_set_res = unordered_set_test(); + __builtin_printf("unordered_set : %s\n", unordered_set_res ? "PASS" : "FAIL"); + const bool unordered_multimap_res = unordered_multimap_test(); + __builtin_printf("unordered_multimap: %s\n", unordered_multimap_res ? "PASS" : "FAIL"); + const bool unordered_multiset_res = unordered_multiset_test(); + __builtin_printf("unordered_multiset: %s\n", unordered_multiset_res ? "PASS" : "FAIL"); + const bool ok = vec_res + && deque_res + && list_res + && map_res + && set_res + && multimap_res + && multiset_res + && array_res + && forward_list_res + && unordered_map_res + && unordered_set_res + && unordered_multimap_res + && unordered_multiset_res; + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-100.C b/libgomp/testsuite/libgomp.c++/target-flex-100.C new file mode 100644 index 0000000..7ab047f --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-100.C @@ -0,0 +1,210 @@ +/* Container adaptors in target region. + Does not test comparison operators other than equality to allow these tests + to be generalized to arbitrary input data. */ + +#include <algorithm> +#include <cstdio> +#include <deque> +#include <queue> +#include <stack> +#include <vector> + +#include "target-flex-common.h" + +template<typename T, std::size_t Size> +bool test_stack(T (&arr)[Size]) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arr[:Size]) + { + bool inner_ok = true; + const std::size_t half_size = Size / 2; + const T first_element = arr[0]; + const T middle_element = arr[half_size - 1]; + const T last_element = arr[Size - 1]; + typedef std::stack<T, std::vector<T> > stack_type; + stack_type stack; + VERIFY (stack.empty()); + VERIFY (stack.size() == 0); + { + /* Do half with push. */ + std::size_t idx = 0; + for (; idx < half_size; ++idx) + { + stack.push(arr[idx]); + VERIFY (stack.top() == arr[idx]); + } + VERIFY (stack.size() == half_size); + VERIFY (static_cast<const stack_type&>(stack).size() == half_size); + for (; idx < Size; ++idx) + { + #if __cplusplus >= 201103L + /* Do the rest with emplace if C++11 or higher. */ + stack.emplace(arr[idx]); + #else + /* Otherwise just use push again. */ + stack.push(arr[idx]); + #endif + VERIFY (stack.top() == arr[idx]); + } + VERIFY (stack.size() == Size); + VERIFY (static_cast<const stack_type&>(stack).size() == Size); + + const stack_type stack_orig = stack_type(std::vector<T>(arr, arr + Size)); + VERIFY (stack == stack_orig); + /* References are contained in their own scope so we don't accidently + add tests referencing them after they have been invalidated. */ + { + const T& const_top = static_cast<const stack_type&>(stack).top(); + VERIFY (const_top == last_element); + T& mutable_top = stack.top(); + mutable_top = first_element; + VERIFY (const_top == first_element); + } + /* Will only compare inequal if the first and last elements are different. */ + VERIFY (first_element != last_element || stack != stack_orig); + for (std::size_t count = Size - half_size; count != 0; --count) + stack.pop(); + VERIFY (stack.top() == middle_element); + const stack_type stack_half_orig = stack_type(std::vector<T>(arr, arr + half_size)); + VERIFY (stack == stack_half_orig); + } + end: + ok = inner_ok; + } + return ok; +} + +template<typename T, std::size_t Size> +bool test_queue(T (&arr)[Size]) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arr[:Size]) + { + bool inner_ok = true; + const std::size_t half_size = Size / 2; + const T first_element = arr[0]; + const T last_element = arr[Size - 1]; + typedef std::queue<T, std::deque<T> > queue_type; + queue_type queue; + VERIFY (queue.empty()); + VERIFY (queue.size() == 0); + { + /* Do half with push. */ + std::size_t idx = 0; + for (; idx < half_size; ++idx) + { + queue.push(arr[idx]); + VERIFY (queue.back() == arr[idx]); + VERIFY (queue.front() == first_element); + } + VERIFY (queue.size() == half_size); + VERIFY (static_cast<const queue_type&>(queue).size() == half_size); + for (; idx < Size; ++idx) + { + #if __cplusplus >= 201103L + /* Do the rest with emplace if C++11 or higher. */ + queue.emplace(arr[idx]); + #else + /* Otherwise just use push again. */ + queue.push(arr[idx]); + #endif + VERIFY (queue.back() == arr[idx]); + } + VERIFY (queue.size() == Size); + VERIFY (static_cast<const queue_type&>(queue).size() == Size); + + const queue_type queue_orig = queue_type(std::deque<T>(arr, arr + Size)); + VERIFY (queue == queue_orig); + + /* References are contained in their own scope so we don't accidently + add tests referencing them after they have been invalidated. */ + { + const T& const_front = static_cast<const queue_type&>(queue).front(); + VERIFY (const_front == first_element); + T& mutable_front = queue.front(); + + const T& const_back = static_cast<const queue_type&>(queue).back(); + VERIFY (const_back == last_element); + T& mutable_back = queue.back(); + { + using std::swap; + swap(mutable_front, mutable_back); + } + VERIFY (const_front == last_element); + VERIFY (const_back == first_element); + /* Will only compare inequal if the first and last elements are different. */ + VERIFY (first_element != last_element || queue != queue_orig); + /* Return the last element to normal for the next comparison. */ + mutable_back = last_element; + } + + const T middle_element = arr[half_size]; + for (std::size_t count = Size - half_size; count != 0; --count) + queue.pop(); + VERIFY (queue.front() == middle_element); + const queue_type queue_upper_half = queue_type(std::deque<T>(arr + half_size, arr + Size)); + VERIFY (queue == queue_upper_half); + } + end: + ok = inner_ok; + } + return ok; +} + +template<typename T, std::size_t Size> +bool test_priority_queue(T (&arr)[Size], const T min_value, const T max_value) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arr[:Size]) + { + bool inner_ok = true; + typedef std::priority_queue<T, std::vector<T> > priority_queue_type; + { + priority_queue_type pqueue; + VERIFY (pqueue.empty()); + VERIFY (pqueue.size() == 0); + } + { + priority_queue_type pqueue(arr, arr + Size); + VERIFY (!pqueue.empty()); + VERIFY (pqueue.size() == Size); + VERIFY (static_cast<const priority_queue_type&>(pqueue).size() == Size); + + const T old_max = pqueue.top(); + + #if __cplusplus >= 201103L + pqueue.emplace(max_value); + #else + pqueue.push(max_value); + #endif + VERIFY (pqueue.top() == max_value); + pqueue.pop(); + VERIFY (pqueue.top() == old_max); + pqueue.push(min_value); + VERIFY (pqueue.top() == old_max); + pqueue.push(max_value); + VERIFY (pqueue.top() == max_value); + pqueue.pop(); + VERIFY (pqueue.top() == old_max); + VERIFY (pqueue.size() == Size + 1); + + for (std::size_t count = Size; count != 0; --count) + pqueue.pop(); + VERIFY (pqueue.size() == 1); + VERIFY (pqueue.top() == min_value); + } + end: + ok = inner_ok; + } + return ok; +} + +int main() +{ + int arr[10] = {0,1,2,3,4,5,6,7,8,9}; + + return test_stack(arr) + && test_queue(arr) + && test_priority_queue(arr, 0, 1000) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-101.C b/libgomp/testsuite/libgomp.c++/target-flex-101.C new file mode 100644 index 0000000..be9037e --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-101.C @@ -0,0 +1,136 @@ +/* { dg-additional-options -std=c++23 } */ + +/* C++23 container adaptors in target region. + Severely needs additional tests. */ + +#include <cstdio> +#include <utility> +#include <version> + +#if __cpp_lib_flat_map >= 202207L +#define ENABLE_FLAT_MAP 1 +#endif +#if __cpp_lib_flat_set >= 202207L +#define ENABLE_FLAT_SET 1 +#endif + +#ifdef ENABLE_FLAT_MAP +#include <flat_map> +#endif +#ifdef ENABLE_FLAT_SET +#include <flat_set> +#endif + +#include "target-flex-common.h" + +#ifdef ENABLE_FLAT_MAP +template<typename K, typename V, typename std::size_t Size> +bool test_flat_map(std::pair<K, V> (&arr)[Size]) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arr[:Size]) + { + bool inner_ok = true; + { + using flat_map_type = std::flat_map<K, V>; + flat_map_type map = {arr, arr + Size}; + + VERIFY (!map.empty()); + for (const auto& element : arr) + VERIFY (map.contains(element.first)); + } + end: + ok = inner_ok; + } + return ok; +} + +template<typename K, typename V, typename std::size_t Size> +bool test_flat_multimap(std::pair<K, V> (&arr)[Size]) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arr[:Size]) + { + bool inner_ok = true; + { + using flat_map_type = std::flat_map<K, V>; + flat_map_type map = {arr, arr + Size}; + + VERIFY (!map.empty()); + for (const auto& element : arr) + VERIFY (map.contains(element.first)); + } + end: + ok = inner_ok; + } + return ok; +} +#else +template<typename K, typename V, typename std::size_t Size> +bool test_flat_map(std::pair<K, V> (&arr)[Size]) { return true; } + +template<typename K, typename V, typename std::size_t Size> +bool test_flat_multimap(std::pair<K, V> (&arr)[Size]) { return true; } +#endif + +#ifdef ENABLE_FLAT_SET +template<typename T, typename std::size_t Size> +bool test_flat_set(T (&arr)[Size]) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arr[:Size]) + { + bool inner_ok = true; + { + using flat_set_type = std::flat_set<T>; + flat_set_type set = {arr, arr + Size}; + + VERIFY (!set.empty()); + for (const auto& element : arr) + VERIFY (set.contains(element)); + } + end: + ok = inner_ok; + } + return ok; +} + +template<typename T, typename std::size_t Size> +bool test_flat_multiset(T (&arr)[Size]) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arr[:Size]) + { + bool inner_ok = true; + { + using flat_multiset_type = std::flat_multiset<T>; + flat_multiset_type multiset = {arr, arr + Size}; + + VERIFY (!multiset.empty()); + for (const auto& element : arr) + VERIFY (multiset.contains(element)); + } + end: + ok = inner_ok; + } + return ok; +} +#else +template<typename T, typename std::size_t Size> +bool test_flat_set(T (&arr)[Size]) { return true; } + +template<typename T, typename std::size_t Size> +bool test_flat_multiset(T (&arr)[Size]) { return true; } +#endif + +int main() +{ + int arr[10] = {0,1,2,3,4,5,6,7,8,9}; + std::pair<int, int> pairs[10] = {{ 1, 2}, { 2, 4}, { 3, 6}, { 4, 8}, { 5, 10}, + { 6, 12}, { 7, 14}, { 8, 16}, { 9, 18}, {10, 20}}; + + return test_flat_set(arr) + && test_flat_multiset(arr) + && test_flat_map(pairs) + && test_flat_multimap(pairs) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-11.C b/libgomp/testsuite/libgomp.c++/target-flex-11.C new file mode 100644 index 0000000..6d55129 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-11.C @@ -0,0 +1,444 @@ +/* Check constructors/destructors are called in containers. */ + +#include <vector> +#include <deque> +#include <list> +#include <set> +#include <map> +#include <utility> +#if __cplusplus >= 201103L +#include <array> +#include <forward_list> +#include <unordered_set> +#include <unordered_map> +#endif + +#include "target-flex-common.h" + +struct indirect_counter +{ + typedef int counter_value_type; + counter_value_type *_count_ptr; + + indirect_counter(counter_value_type *count_ptr) BL_NOEXCEPT : _count_ptr(count_ptr) { + ++(*_count_ptr); + } + indirect_counter(const indirect_counter& other) BL_NOEXCEPT : _count_ptr(other._count_ptr) { + ++(*_count_ptr); + } + /* Don't declare a move constructor, we want to copy no matter what. */ + ~indirect_counter() { + --(*_count_ptr); + } +}; + +bool operator==(indirect_counter const& lhs, indirect_counter const& rhs) BL_NOEXCEPT + { return lhs._count_ptr == rhs._count_ptr; } +bool operator<(indirect_counter const& lhs, indirect_counter const& rhs) BL_NOEXCEPT + { return lhs._count_ptr < rhs._count_ptr; } + +#if __cplusplus >= 201103L +template<> +struct std::hash<indirect_counter> +{ + std::size_t operator()(const indirect_counter& ic) const noexcept + { return std::hash<indirect_counter::counter_value_type *>{}(ic._count_ptr); } +}; +#endif + +/* Not a container, just a sanity check really. */ +bool automatic_lifetime_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + indirect_counter c = indirect_counter(&counter); + indirect_counter(static_cast<int*>(&counter)); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool vector_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + std::vector<indirect_counter> vec(42, indirect_counter(&counter)); + VERIFY (counter == 42); + vec.resize(32, indirect_counter(&counter)); + VERIFY (counter == 32); + vec.push_back(indirect_counter(&counter)); + VERIFY (counter == 33); + vec.pop_back(); + VERIFY (counter == 32); + vec.pop_back(); + VERIFY (counter == 31); + vec.resize(100, indirect_counter(&counter)); + VERIFY (counter == 100); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool deque_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + std::deque<indirect_counter> vec(42, indirect_counter(&counter)); + VERIFY (counter == 42); + vec.resize(32, indirect_counter(&counter)); + VERIFY (counter == 32); + vec.push_back(indirect_counter(&counter)); + VERIFY (counter == 33); + vec.pop_back(); + VERIFY (counter == 32); + vec.pop_back(); + VERIFY (counter == 31); + vec.resize(100, indirect_counter(&counter)); + VERIFY (counter == 100); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool list_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + std::list<indirect_counter> list(42, indirect_counter(&counter)); + VERIFY (counter == 42); + list.resize(32, indirect_counter(&counter)); + VERIFY (counter == 32); + list.push_back(indirect_counter(&counter)); + VERIFY (counter == 33); + list.pop_back(); + VERIFY (counter == 32); + list.pop_back(); + VERIFY (counter == 31); + list.resize(100, indirect_counter(&counter)); + VERIFY (counter == 100); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool map_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + std::map<int, indirect_counter> map; + map.insert(std::make_pair(1, indirect_counter(&counter))); + VERIFY (counter == 1); + map.insert(std::make_pair(1, indirect_counter(&counter))); + VERIFY (counter == 1); + map.insert(std::make_pair(2, indirect_counter(&counter))); + VERIFY (counter == 2); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool set_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter0 = 0; + int counter1 = 0; + { + std::set<indirect_counter> set; + set.insert(indirect_counter(&counter0)); + VERIFY (counter0 == 1); + set.insert(indirect_counter(&counter0)); + VERIFY (counter0 == 1); + set.insert(indirect_counter(&counter1)); + VERIFY (counter0 == 1 && counter1 == 1); + } + VERIFY (counter0 == 0 && counter1 == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool multimap_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + std::multimap<int, indirect_counter> multimap; + multimap.insert(std::make_pair(1, indirect_counter(&counter))); + VERIFY (counter == 1); + multimap.insert(std::make_pair(1, indirect_counter(&counter))); + VERIFY (counter == 2); + multimap.insert(std::make_pair(2, indirect_counter(&counter))); + VERIFY (counter == 3); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool multiset_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter0 = 0; + int counter1 = 0; + { + std::multiset<indirect_counter> multiset; + multiset.insert(indirect_counter(&counter0)); + VERIFY (counter0 == 1); + multiset.insert(indirect_counter(&counter0)); + VERIFY (counter0 == 2); + multiset.insert(indirect_counter(&counter1)); + VERIFY (counter0 == 2 && counter1 == 1); + } + VERIFY (counter0 == 0 && counter1 == 0); + end: + ok = inner_ok; + } + return ok; +} + +#if __cplusplus >= 201103L + +bool array_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + indirect_counter ic(&counter); + std::array<indirect_counter, 10> array{ic, ic, ic, ic, ic, + ic, ic, ic, ic, ic}; + VERIFY (counter == 11); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool forward_list_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + std::forward_list<indirect_counter> forward_list(42, indirect_counter(&counter)); + VERIFY (counter == 42); + forward_list.resize(32, indirect_counter(&counter)); + VERIFY (counter == 32); + forward_list.push_front(indirect_counter(&counter)); + VERIFY (counter == 33); + forward_list.pop_front(); + VERIFY (counter == 32); + forward_list.pop_front(); + VERIFY (counter == 31); + forward_list.resize(100, indirect_counter(&counter)); + VERIFY (counter == 100); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool unordered_map_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + std::unordered_map<int, indirect_counter> unordered_map; + unordered_map.insert({1, indirect_counter(&counter)}); + VERIFY (counter == 1); + unordered_map.insert({1, indirect_counter(&counter)}); + VERIFY (counter == 1); + unordered_map.insert({2, indirect_counter(&counter)}); + VERIFY (counter == 2); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool unordered_set_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter0 = 0; + int counter1 = 0; + { + std::unordered_set<indirect_counter> unordered_set; + unordered_set.insert(indirect_counter(&counter0)); + VERIFY (counter0 == 1); + unordered_set.insert(indirect_counter(&counter0)); + VERIFY (counter0 == 1); + unordered_set.insert(indirect_counter(&counter1)); + VERIFY (counter0 == 1 && counter1 == 1); + } + VERIFY (counter0 == 0 && counter1 == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool unordered_multimap_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter = 0; + { + std::unordered_multimap<int, indirect_counter> unordered_multimap; + unordered_multimap.insert({1, indirect_counter(&counter)}); + VERIFY (counter == 1); + unordered_multimap.insert({1, indirect_counter(&counter)}); + VERIFY (counter == 2); + unordered_multimap.insert({2, indirect_counter(&counter)}); + VERIFY (counter == 3); + } + VERIFY (counter == 0); + end: + ok = inner_ok; + } + return ok; +} + +bool unordered_multiset_test() +{ + bool ok; + #pragma omp target map(from: ok) + { + bool inner_ok = true; + int counter0 = 0; + int counter1 = 0; + { + std::unordered_multiset<indirect_counter> unordered_multiset; + unordered_multiset.insert(indirect_counter(&counter0)); + VERIFY (counter0 == 1); + unordered_multiset.insert(indirect_counter(&counter0)); + VERIFY (counter0 == 2); + unordered_multiset.insert(indirect_counter(&counter1)); + VERIFY (counter0 == 2 && counter1 == 1); + } + VERIFY (counter0 == 0 && counter1 == 0); + end: + ok = inner_ok; + } + return ok; +} + +#else +bool array_test() { return true; } +bool forward_list_test() { return true; } +bool unordered_map_test() { return true; } +bool unordered_set_test() { return true; } +bool unordered_multimap_test() { return true; } +bool unordered_multiset_test() { return true; } +#endif + +int main() +{ + const bool auto_res = automatic_lifetime_test(); + const bool vec_res = vector_test(); + const bool deque_res = deque_test(); + const bool list_res = list_test(); + const bool map_res = map_test(); + const bool set_res = set_test(); + const bool multimap_res = multimap_test(); + const bool multiset_res = multiset_test(); + const bool array_res = array_test(); + const bool forward_list_res = forward_list_test(); + const bool unordered_map_res = unordered_map_test(); + const bool unordered_set_res = unordered_set_test(); + const bool unordered_multimap_res = unordered_multimap_test(); + const bool unordered_multiset_res = unordered_multiset_test(); + std::printf("sanity check : %s\n", auto_res ? "PASS" : "FAIL"); + std::printf("vector : %s\n", vec_res ? "PASS" : "FAIL"); + std::printf("deque : %s\n", deque_res ? "PASS" : "FAIL"); + std::printf("list : %s\n", list_res ? "PASS" : "FAIL"); + std::printf("map : %s\n", map_res ? "PASS" : "FAIL"); + std::printf("set : %s\n", set_res ? "PASS" : "FAIL"); + std::printf("multimap : %s\n", multimap_res ? "PASS" : "FAIL"); + std::printf("multiset : %s\n", multiset_res ? "PASS" : "FAIL"); + std::printf("array : %s\n", array_res ? "PASS" : "FAIL"); + std::printf("forward_list : %s\n", forward_list_res ? "PASS" : "FAIL"); + std::printf("unordered_map : %s\n", unordered_map_res ? "PASS" : "FAIL"); + std::printf("unordered_set : %s\n", unordered_set_res ? "PASS" : "FAIL"); + std::printf("unordered_multimap: %s\n", unordered_multimap_res ? "PASS" : "FAIL"); + std::printf("unordered_multiset: %s\n", unordered_multiset_res ? "PASS" : "FAIL"); + const bool ok = auto_res + && vec_res + && deque_res + && list_res + && map_res + && set_res + && multimap_res + && multiset_res + && array_res + && forward_list_res + && unordered_map_res + && unordered_set_res + && unordered_multimap_res + && unordered_multiset_res; + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-12.C b/libgomp/testsuite/libgomp.c++/target-flex-12.C new file mode 100644 index 0000000..024fb73 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-12.C @@ -0,0 +1,736 @@ +/* Populated with mapped data, validate, mutate, validate again. + The cases using sets do not mutate. + Note: Some of the code in here really sucks due to being made to be + compatible with c++98. */ + +#include <vector> +#include <deque> +#include <list> +#include <set> +#include <map> +#if __cplusplus >= 201103L +#include <array> +#include <forward_list> +#include <unordered_set> +#include <unordered_map> +#endif + +#include <limits> +#include <iterator> + +#include "target-flex-common.h" + +template<bool B, class T = void> +struct enable_if {}; + +template<class T> +struct enable_if<true, T> { typedef T type; }; + +struct identity_func +{ +#if __cplusplus < 201103L + template<typename T> + T& operator()(T& arg) const BL_NOEXCEPT { return arg; } + template<typename T> + T const& operator()(T const& arg) const BL_NOEXCEPT { return arg; } +#else + template<typename T> + constexpr T&& operator()(T&& arg) const BL_NOEXCEPT { return std::forward<T>(arg); } +#endif +}; + +/* Applies projection to the second iterator. */ +template<typename It0, typename It1, typename Proj> +bool validate_sequential_elements(const It0 begin0, const It0 end0, + const It1 begin1, const It1 end1, + Proj proj) BL_NOEXCEPT +{ + It0 it0 = begin0; + It1 it1 = begin1; + for (; it0 != end0; ++it0, ++it1) + { + /* Sizes mismatch, don't bother aborting though just fail the test. */ + if (it1 == end1) + return false; + if (*it0 != proj(*it1)) + return false; + } + /* Sizes mismatch, do as above. */ + if (it1 != end1) + return false; + return true; +} + +template<typename It0, typename It1> +bool validate_sequential_elements(const It0 begin0, const It0 end0, + const It1 begin1, const It1 end1) BL_NOEXCEPT +{ + return validate_sequential_elements(begin0, end0, begin1, end1, identity_func()); +} + +/* Inefficient, but simple. */ +template<typename It, typename OutIt> +void simple_copy(const It begin, const It end, OutIt out) BL_NOEXCEPT +{ + for (It it = begin; it != end; ++it, ++out) + *out = *it; +} + +template<typename It, typename MutateFn> +void simple_mutate(const It begin, const It end, MutateFn mut_fn) BL_NOEXCEPT +{ + for (It it = begin; it != end; ++it) + *it = mut_fn(*it); +} + +template<typename MutationFunc, typename T, std::size_t Size> +bool vector_test(const T (&arr)[Size]) +{ + bool ok; + T out_arr[Size]; + T out_mut_arr[Size]; + #pragma omp target map(from: ok, out_arr[:Size], out_mut_arr[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::vector<T> vector(arr, arr + Size); + VERIFY (validate_sequential_elements(vector.begin(), vector.end(), + arr, arr + Size)); + simple_copy(vector.begin(), vector.end(), out_arr); + simple_mutate(vector.begin(), vector.end(), MutationFunc()); + VERIFY (validate_sequential_elements(vector.begin(), vector.end(), + arr, arr + Size, MutationFunc())); + simple_copy(vector.begin(), vector.end(), out_mut_arr); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_sequential_elements(out_arr, out_arr + Size, + arr, arr + Size)); + VERIFY_NON_TARGET (validate_sequential_elements(out_mut_arr, out_mut_arr + Size, + arr, arr + Size, MutationFunc())); + return true; +} + +template<typename MutationFunc, typename T, std::size_t Size> +bool deque_test(const T (&arr)[Size]) +{ + bool ok; + T out_arr[Size]; + T out_mut_arr[Size]; + #pragma omp target map(from: ok, out_arr[:Size], out_mut_arr[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::deque<T> deque(arr, arr + Size); + VERIFY (validate_sequential_elements(deque.begin(), deque.end(), + arr, arr + Size)); + simple_copy(deque.begin(), deque.end(), out_arr); + simple_mutate(deque.begin(), deque.end(), MutationFunc()); + VERIFY (validate_sequential_elements(deque.begin(), deque.end(), + arr, arr + Size, MutationFunc())); + simple_copy(deque.begin(), deque.end(), out_mut_arr); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_sequential_elements(out_arr, out_arr + Size, + arr, arr + Size)); + VERIFY_NON_TARGET (validate_sequential_elements(out_mut_arr, out_mut_arr + Size, + arr, arr + Size, MutationFunc())); + return true; +} + +template<typename MutationFunc, typename T, std::size_t Size> +bool list_test(const T (&arr)[Size]) +{ + bool ok; + T out_arr[Size]; + T out_mut_arr[Size]; + #pragma omp target map(from: ok, out_arr[:Size], out_mut_arr[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::list<T> list(arr, arr + Size); + VERIFY (validate_sequential_elements(list.begin(), list.end(), + arr, arr + Size)); + simple_copy(list.begin(), list.end(), out_arr); + simple_mutate(list.begin(), list.end(), MutationFunc()); + VERIFY (validate_sequential_elements(list.begin(), list.end(), + arr, arr + Size, MutationFunc())); + simple_copy(list.begin(), list.end(), out_mut_arr); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_sequential_elements(out_arr, out_arr + Size, + arr, arr + Size)); + VERIFY_NON_TARGET (validate_sequential_elements(out_mut_arr, out_mut_arr + Size, + arr, arr + Size, MutationFunc())); + return true; +} + +template<typename T> +const T& get_key(const T& arg) BL_NOEXCEPT + { return arg; } +template<typename K, typename V> +const K& get_key(const std::pair<K, V>& pair) BL_NOEXCEPT + { return pair.first; } +template<typename T> +const T& get_value(const T& arg) BL_NOEXCEPT + { return arg; } +template<typename K, typename V> +const K& get_value(const std::pair<K, V>& pair) BL_NOEXCEPT + { return pair.second; } + +template<typename T> +struct key_type { typedef T type; }; +template<typename K, typename V> +struct key_type<std::pair<K, V> > { typedef K type; }; + +template<typename Proj, typename Container, typename It> +bool validate_associative(const Container& container, + const It compare_begin, + const It compare_end, + Proj proj) BL_NOEXCEPT +{ + const typename Container::const_iterator elem_end = container.end(); + for (It compare_it = compare_begin; compare_it != compare_end; ++compare_it) + { + const typename Container::const_iterator elem_it = container.find(get_key(*compare_it)); + VERIFY_NON_TARGET (elem_it != elem_end); + VERIFY_NON_TARGET (proj(get_value(*compare_it)) == get_value(*elem_it)); + } + return true; +} + +template<typename Container, typename It> +bool validate_associative(const Container& container, + const It compare_begin, + const It compare_end) BL_NOEXCEPT +{ + return validate_associative(container, compare_begin, compare_end, identity_func()); +} + +template<typename It, typename MutateFn> +void simple_mutate_map(const It begin, const It end, MutateFn mut_fn) BL_NOEXCEPT +{ + for (It it = begin; it != end; ++it) + it->second = mut_fn(it->second); +} + +template<typename It, typename OutIter> +void simple_copy_unique(const It begin, const It end, OutIter out) BL_NOEXCEPT +{ + /* In case anyone reads this, I want it to be known that I hate c++98. */ + typedef typename key_type<typename std::iterator_traits<It>::value_type>::type key_t; + std::set<key_t> already_seen; + for (It it = begin; it != end; ++it, ++out) + { + key_t key = get_key(*it); + if (already_seen.find(key) != already_seen.end()) + continue; + already_seen.insert(key); + *out = *it; + } +} + +template<typename MutationFunc, typename K, typename V, std::size_t Size> +bool map_test(const std::pair<K, V> (&arr)[Size]) +{ + std::map<K, V> reference_map(arr, arr + Size); + bool ok; + /* Both sizes should be the same. */ + std::pair<K, V> out_pairs[Size]; + std::size_t out_size; + std::pair<K, V> out_pairs_mut[Size]; + std::size_t out_size_mut; + #pragma omp target map(from: ok, out_pairs[:Size], out_size, \ + out_pairs_mut[:Size], out_size_mut) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::vector<std::pair<K, V> > unique_elems; + simple_copy_unique(arr, arr + Size, + std::back_insert_iterator<std::vector<std::pair<K, V> > >(unique_elems)); + + std::map<K, V> map(arr, arr + Size); + VERIFY (validate_associative(map, unique_elems.begin(), unique_elems.end())); + simple_copy(map.begin(), map.end(), out_pairs); + out_size = map.size(); + simple_mutate_map(map.begin(), map.end(), MutationFunc()); + VERIFY (validate_associative(map, unique_elems.begin(), unique_elems.end(), + MutationFunc())); + simple_copy(map.begin(), map.end(), out_pairs_mut); + out_size_mut = map.size(); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (out_size == out_size_mut); + VERIFY_NON_TARGET (validate_associative(reference_map, + out_pairs, out_pairs + out_size)); + simple_mutate_map(reference_map.begin(), reference_map.end(), MutationFunc()); + VERIFY_NON_TARGET (validate_associative(reference_map, + out_pairs_mut, out_pairs_mut + out_size_mut)); + return true; +} + +template<typename T, std::size_t Size> +bool set_test(const T (&arr)[Size]) +{ + std::set<T> reference_set(arr, arr + Size); + bool ok; + /* Both sizes should be the same. */ + T out_arr[Size]; + std::size_t out_size; + #pragma omp target map(from: ok, out_arr[:Size], out_size) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::vector<T> unique_elems; + simple_copy_unique(arr, arr + Size, + std::back_insert_iterator<std::vector<T> >(unique_elems)); + + std::set<T> set(arr, arr + Size); + VERIFY (validate_associative(set, unique_elems.begin(), unique_elems.end())); + simple_copy(set.begin(), set.end(), out_arr); + out_size = set.size(); + /* Sets can't be mutated, we could create another set with mutated + but it gets a little annoying and probably isn't an interesting test. */ + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_associative(reference_set, + out_arr, out_arr + out_size)); + return true; +} + +template<typename Proj, typename Container, typename It> +bool validate_multi_associative(const Container& container, + const It compare_begin, + const It compare_end, + Proj proj) BL_NOEXCEPT +{ + /* Once again, for the poor soul reviewing these, I hate c++98. */ + typedef typename key_type<typename std::iterator_traits<It>::value_type>::type key_t; + typedef std::map<key_t, std::size_t> counter_map; + counter_map key_count_map; + for (It it = compare_begin; it != compare_end; ++it) + { + const key_t& key = get_key(*it); + typename counter_map::iterator counter_it + = key_count_map.find(key); + if (counter_it != key_count_map.end()) + ++counter_it->second; + else + key_count_map.insert(std::pair<const key_t, std::size_t>(key, std::size_t(1))); + } + const typename Container::const_iterator elem_end = container.end(); + for (It compare_it = compare_begin; compare_it != compare_end; ++compare_it) + { + const key_t& key = get_key(*compare_it); + typename counter_map::iterator count_it = key_count_map.find(key); + std::size_t key_count = count_it != key_count_map.end() ? count_it->second + : std::size_t(0); + VERIFY_NON_TARGET (key_count > std::size_t(0) && "this will never happen"); + /* This gets tested multiple times but that should be fine. */ + VERIFY_NON_TARGET (key_count == container.count(key)); + typename Container::const_iterator elem_it = container.find(key); + /* This will never happen if the previous case passed. */ + VERIFY_NON_TARGET (elem_it != elem_end); + bool found_element = false; + for (; elem_it != elem_end; ++elem_it) + if (proj(get_value(*compare_it)) == get_value(*elem_it)) + { + found_element = true; + break; + } + VERIFY_NON_TARGET (found_element); + } + return true; +} + +template<typename Container, typename It> +bool validate_multi_associative(const Container& container, + const It compare_begin, + const It compare_end) BL_NOEXCEPT +{ + return validate_multi_associative(container, compare_begin, compare_end, identity_func()); +} + +template<typename MutationFunc, typename K, typename V, std::size_t Size> +bool multimap_test(const std::pair<K, V> (&arr)[Size]) +{ + std::multimap<K, V> reference_multimap(arr, arr + Size); + bool ok; + std::pair<K, V> out_pairs[Size]; + std::pair<K, V> out_pairs_mut[Size]; + #pragma omp target map(from: ok, out_pairs[:Size], out_pairs_mut[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::multimap<K, V> multimap(arr, arr + Size); + VERIFY (validate_multi_associative(multimap, arr, arr + Size)); + simple_copy(multimap.begin(), multimap.end(), out_pairs); + simple_mutate_map(multimap.begin(), multimap.end(), MutationFunc()); + VERIFY (validate_multi_associative(multimap, arr, arr + Size, MutationFunc())); + simple_copy(multimap.begin(), multimap.end(), out_pairs_mut); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_multi_associative(reference_multimap, + out_pairs, out_pairs + Size)); + simple_mutate_map(reference_multimap.begin(), reference_multimap.end(), MutationFunc()); + VERIFY_NON_TARGET (validate_multi_associative(reference_multimap, + out_pairs_mut, out_pairs_mut + Size)); + return true; +} + +template<typename T, std::size_t Size> +bool multiset_test(const T (&arr)[Size]) +{ + std::multiset<T> reference_multiset(arr, arr + Size); + bool ok; + T out_arr[Size]; + #pragma omp target map(from: ok, out_arr[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::multiset<T> set(arr, arr + Size); + VERIFY (validate_multi_associative(set, arr, arr + Size)); + simple_copy(set.begin(), set.end(), out_arr); + /* Sets can't be mutated, we could create another set with mutated + but it gets a little annoying and probably isn't an interesting test. */ + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_multi_associative(reference_multiset, + out_arr, out_arr + Size)); + return true; +} + +#if __cplusplus >= 201103L + +template<typename MutationFunc, typename T, std::size_t Size> +bool array_test(const T (&arr)[Size]) +{ + bool ok; + T out_arr[Size]; + T out_mut_arr[Size]; + #pragma omp target map(from: ok, out_arr[:Size], out_mut_arr[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::array<T, Size> std_array{}; + /* Special case for std::array since it can't be initialized + with iterators. */ + { + T zero_val = T{}; + for (auto it = std_array.begin(); it != std_array.end(); ++it) + VERIFY (*it == zero_val); + } + simple_copy(arr, arr + Size, std_array.begin()); + VERIFY (validate_sequential_elements(std_array.begin(), std_array.end(), + arr, arr + Size)); + simple_copy(std_array.begin(), std_array.end(), out_arr); + simple_mutate(std_array.begin(), std_array.end(), MutationFunc()); + VERIFY (validate_sequential_elements(std_array.begin(), std_array.end(), + arr, arr + Size, MutationFunc())); + simple_copy(std_array.begin(), std_array.end(), out_mut_arr); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_sequential_elements(out_arr, out_arr + Size, + arr, arr + Size)); + VERIFY_NON_TARGET (validate_sequential_elements(out_mut_arr, out_mut_arr + Size, + arr, arr + Size, MutationFunc())); + return true; +} + +template<typename MutationFunc, typename T, std::size_t Size> +bool forward_list_test(const T (&arr)[Size]) +{ + bool ok; + T out_arr[Size]; + T out_mut_arr[Size]; + #pragma omp target map(from: ok, out_arr[:Size], out_mut_arr[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::forward_list<T> fwd_list(arr, arr + Size); + VERIFY (validate_sequential_elements(fwd_list.begin(), fwd_list.end(), + arr, arr + Size)); + simple_copy(fwd_list.begin(), fwd_list.end(), out_arr); + simple_mutate(fwd_list.begin(), fwd_list.end(), MutationFunc()); + VERIFY (validate_sequential_elements(fwd_list.begin(), fwd_list.end(), + arr, arr + Size, MutationFunc())); + simple_copy(fwd_list.begin(), fwd_list.end(), out_mut_arr); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_sequential_elements(out_arr, out_arr + Size, + arr, arr + Size)); + VERIFY_NON_TARGET (validate_sequential_elements(out_mut_arr, out_mut_arr + Size, + arr, arr + Size, MutationFunc())); + return true; +} + +template<typename MutationFunc, typename K, typename V, std::size_t Size> +bool unordered_map_test(const std::pair<K, V> (&arr)[Size]) +{ + std::unordered_map<K, V> reference_map(arr, arr + Size); + bool ok; + /* Both sizes should be the same. */ + std::pair<K, V> out_pairs[Size]; + std::size_t out_size; + std::pair<K, V> out_pairs_mut[Size]; + std::size_t out_size_mut; + #pragma omp target map(from: ok, out_pairs[:Size], out_size, \ + out_pairs_mut[:Size], out_size_mut) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::vector<std::pair<K, V> > unique_elems; + simple_copy_unique(arr, arr + Size, + std::back_insert_iterator<std::vector<std::pair<K, V> > >(unique_elems)); + + std::unordered_map<K, V> map(arr, arr + Size); + VERIFY (validate_associative(map, unique_elems.begin(), unique_elems.end())); + simple_copy(map.begin(), map.end(), out_pairs); + out_size = map.size(); + simple_mutate_map(map.begin(), map.end(), MutationFunc()); + VERIFY (validate_associative(map, unique_elems.begin(), unique_elems.end(), + MutationFunc())); + simple_copy(map.begin(), map.end(), out_pairs_mut); + out_size_mut = map.size(); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (out_size == out_size_mut); + VERIFY_NON_TARGET (validate_associative(reference_map, + out_pairs, out_pairs + out_size)); + simple_mutate_map(reference_map.begin(), reference_map.end(), MutationFunc()); + VERIFY_NON_TARGET (validate_associative(reference_map, + out_pairs_mut, out_pairs_mut + out_size_mut)); + return true; +} + +template<typename T, std::size_t Size> +bool unordered_set_test(const T (&arr)[Size]) +{ + std::unordered_set<T> reference_set(arr, arr + Size); + bool ok; + /* Both sizes should be the same. */ + T out_arr[Size]; + std::size_t out_size; + #pragma omp target map(from: ok, out_arr[:Size], out_size) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::vector<T> unique_elems; + simple_copy_unique(arr, arr + Size, + std::back_insert_iterator<std::vector<T> >(unique_elems)); + + std::unordered_set<T> set(arr, arr + Size); + VERIFY (validate_associative(set, unique_elems.begin(), unique_elems.end())); + simple_copy(set.begin(), set.end(), out_arr); + out_size = set.size(); + /* Sets can't be mutated, we could create another set with mutated + but it gets a little annoying and probably isn't an interesting test. */ + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_associative(reference_set, + out_arr, out_arr + out_size)); + return true; +} + +template<typename MutationFunc, typename K, typename V, std::size_t Size> +bool unordered_multimap_test(const std::pair<K, V> (&arr)[Size]) +{ + std::unordered_multimap<K, V> reference_multimap(arr, arr + Size); + bool ok; + std::pair<K, V> out_pairs[Size]; + std::pair<K, V> out_pairs_mut[Size]; + #pragma omp target map(from: ok, out_pairs[:Size], out_pairs_mut[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::unordered_multimap<K, V> multimap(arr, arr + Size); + VERIFY (validate_multi_associative(multimap, arr, arr + Size)); + simple_copy(multimap.begin(), multimap.end(), out_pairs); + simple_mutate_map(multimap.begin(), multimap.end(), MutationFunc()); + VERIFY (validate_multi_associative(multimap, arr, arr + Size, MutationFunc())); + simple_copy(multimap.begin(), multimap.end(), out_pairs_mut); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_multi_associative(reference_multimap, + out_pairs, out_pairs + Size)); + simple_mutate_map(reference_multimap.begin(), reference_multimap.end(), MutationFunc()); + VERIFY_NON_TARGET (validate_multi_associative(reference_multimap, + out_pairs_mut, out_pairs_mut + Size)); + return true; +} + +template<typename T, std::size_t Size> +bool unordered_multiset_test(const T (&arr)[Size]) +{ + std::unordered_multiset<T> reference_multiset(arr, arr + Size); + bool ok; + T out_arr[Size]; + #pragma omp target map(from: ok, out_arr[:Size]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::unordered_multiset<T> set(arr, arr + Size); + VERIFY (validate_multi_associative(set, arr, arr + Size)); + simple_copy(set.begin(), set.end(), out_arr); + /* Sets can't be mutated, we could create another set with mutated + but it gets a little annoying and probably isn't an interesting test. */ + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (validate_multi_associative(reference_multiset, + out_arr, out_arr + Size)); + return true; +} + +#else +template<typename, typename T, std::size_t Size> bool array_test(const T (&arr)[Size]) { return true; } +template<typename, typename T, std::size_t Size> bool forward_list_test(const T (&arr)[Size]) { return true; } +template<typename, typename T, std::size_t Size> bool unordered_map_test(const T (&arr)[Size]) { return true; } +template<typename T, std::size_t Size> bool unordered_set_test(const T (&arr)[Size]) { return true; } +template<typename, typename T, std::size_t Size> bool unordered_multimap_test(const T (&arr)[Size]) { return true; } +template<typename T, std::size_t Size> bool unordered_multiset_test(const T (&arr)[Size]) { return true; } +#endif + +/* This clamps to the maximum value to guard against overflowing, + assuming std::numeric_limits is specialized for T. */ +struct multiply_by_2 +{ + template<typename T> + typename enable_if<std::numeric_limits<T>::is_specialized, T>::type + operator()(T arg) const BL_NOEXCEPT { + if (arg < static_cast<T>(0)) + { + if (std::numeric_limits<T>::min() / static_cast<T>(2) >= arg) + return std::numeric_limits<T>::min(); + } + else + { + if (std::numeric_limits<T>::max() / static_cast<T>(2) <= arg) + return std::numeric_limits<T>::max(); + } + return arg * 2; + } + template<typename T> + typename enable_if<!std::numeric_limits<T>::is_specialized, T>::type + operator()(T arg) const BL_NOEXCEPT { + return arg * 2; + } +}; + +int main() +{ + int data[8] = {0, 1, 2, 3, 4, 5, 6, 7}; + std::pair<int, int> pairs[10] = {std::pair<int, int>( 1, 2), + std::pair<int, int>( 2, 4), + std::pair<int, int>( 3, 6), + std::pair<int, int>( 4, 8), + std::pair<int, int>( 5, 10), + std::pair<int, int>( 6, 12), + std::pair<int, int>( 7, 14), + std::pair<int, int>( 8, 16), + std::pair<int, int>( 9, 18), + std::pair<int, int>(10, 20)}; + const bool vec_res = vector_test<multiply_by_2>(data); + const bool deque_res = deque_test<multiply_by_2>(data); + const bool list_res = list_test<multiply_by_2>(data); + const bool map_res = map_test<multiply_by_2>(pairs); + const bool set_res = set_test(data); + const bool multimap_res = multimap_test<multiply_by_2>(pairs); + const bool multiset_res = multiset_test(data); + const bool array_res = array_test<multiply_by_2>(data); + const bool forward_list_res = forward_list_test<multiply_by_2>(data); + const bool unordered_map_res = unordered_map_test<multiply_by_2>(pairs); + const bool unordered_set_res = unordered_set_test(data); + const bool unordered_multimap_res = unordered_multimap_test<multiply_by_2>(pairs); + const bool unordered_multiset_res = unordered_multiset_test(data); + std::printf("vector : %s\n", vec_res ? "PASS" : "FAIL"); + std::printf("deque : %s\n", deque_res ? "PASS" : "FAIL"); + std::printf("list : %s\n", list_res ? "PASS" : "FAIL"); + std::printf("map : %s\n", map_res ? "PASS" : "FAIL"); + std::printf("set : %s\n", set_res ? "PASS" : "FAIL"); + std::printf("multimap : %s\n", multimap_res ? "PASS" : "FAIL"); + std::printf("multiset : %s\n", multiset_res ? "PASS" : "FAIL"); + std::printf("array : %s\n", array_res ? "PASS" : "FAIL"); + std::printf("forward_list : %s\n", forward_list_res ? "PASS" : "FAIL"); + std::printf("unordered_map : %s\n", unordered_map_res ? "PASS" : "FAIL"); + std::printf("unordered_set : %s\n", unordered_set_res ? "PASS" : "FAIL"); + std::printf("unordered_multimap: %s\n", unordered_multimap_res ? "PASS" : "FAIL"); + std::printf("unordered_multiset: %s\n", unordered_multiset_res ? "PASS" : "FAIL"); + const bool ok = vec_res + && deque_res + && list_res + && map_res + && set_res + && multimap_res + && multiset_res + && array_res + && forward_list_res + && unordered_map_res + && unordered_set_res + && unordered_multimap_res + && unordered_multiset_res; + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-2000.C b/libgomp/testsuite/libgomp.c++/target-flex-2000.C new file mode 100644 index 0000000..688c014 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-2000.C @@ -0,0 +1,32 @@ +/* Tiny tuple test. */ + +#include <tuple> + +#include "target-flex-common.h" + +bool test(int arg) +{ + bool ok; + int out; + std::tuple tup = {'a', arg, 3.14f}; + #pragma omp target map(from: ok, out) map(to: tup) + { + bool inner_ok = true; + { + VERIFY (std::get<0>(tup) == 'a'); + out = std::get<1>(tup); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (out == arg); + return true; +} + +int main() +{ + volatile int arg = 42u; + return test(arg) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-2001.C b/libgomp/testsuite/libgomp.c++/target-flex-2001.C new file mode 100644 index 0000000..f1a6c12 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-2001.C @@ -0,0 +1,61 @@ +/* { dg-additional-options "-std=c++20" } */ + +/* Functional */ + +#include <functional> +#include <utility> + +#include "target-flex-common.h" + +template<typename T,typename Fn> +auto invoke_unary(T&& a, Fn&& fn) noexcept +{ + return std::invoke(std::forward<Fn>(fn), + std::forward<T>(a)); +} + +template<typename T, typename U, typename Fn> +auto invoke_binary(T&& a, U&& b, Fn&& fn) noexcept +{ + return std::invoke(std::forward<Fn>(fn), + std::forward<T>(a), + std::forward<U>(b)); +} + +bool test(unsigned arg) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arg) + { + bool inner_ok = true; + { + VERIFY (std::plus{}(arg, 2) == arg + 2); + auto bound_plus_arg = std::bind_front(std::plus{}, arg); + VERIFY (bound_plus_arg(10) == arg + 10); + VERIFY (bound_plus_arg(20) == arg + 20); + + VERIFY (std::not_fn(std::not_equal_to{})(arg, arg)); + VERIFY (invoke_binary(arg, arg, std::not_fn(std::not_equal_to{}))); + auto bound_equals_arg = std::bind_front(std::not_fn(std::not_equal_to{}), arg); + VERIFY (bound_equals_arg(arg)); + VERIFY (std::not_fn(bound_equals_arg)(arg + 1)); + VERIFY (invoke_unary(arg, bound_equals_arg)); + + VERIFY (std::not_fn(std::ranges::not_equal_to{})(arg, arg)); + VERIFY (invoke_binary(arg, arg, std::not_fn(std::ranges::not_equal_to{}))); + auto bound_ranges_equals_arg = std::bind_front(std::not_fn(std::ranges::not_equal_to{}), arg); + VERIFY (bound_ranges_equals_arg(arg)); + VERIFY (std::not_fn(bound_ranges_equals_arg)(arg + 1)); + VERIFY (invoke_unary(arg, bound_ranges_equals_arg)); + } + end: + ok = inner_ok; + } + return ok; +} + +int main() +{ + volatile unsigned arg = 42u; + return test(arg) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-2002.C b/libgomp/testsuite/libgomp.c++/target-flex-2002.C new file mode 100644 index 0000000..f738806 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-2002.C @@ -0,0 +1,97 @@ +/* { dg-additional-options "-std=c++23" } */ + +/* expected/optional */ + +#include <optional> +#include <expected> + +#include "target-flex-common.h" + +std::optional<unsigned> make_optional(bool b, unsigned arg = 0u) noexcept +{ + if (!b) + return std::nullopt; + return {arg}; +} + +bool test_optional(unsigned arg) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arg) + { + bool inner_ok = true; + { + auto null_opt = make_optional(false); + VERIFY (!null_opt); + VERIFY (!null_opt.has_value()); + VERIFY (null_opt.value_or(arg * 2u) == arg * 2u); + VERIFY (null_opt.or_else([&](){ return std::optional<unsigned>{arg}; }) + .transform([](int a){ return a * 2u; }) + .value_or(0) == arg * 2u); + + auto opt = make_optional(true, arg); + VERIFY (opt); + VERIFY (opt.has_value()); + VERIFY (opt.value() == arg); + VERIFY (*opt == arg); + VERIFY (opt.value_or(arg + 42) == arg); + VERIFY (opt.or_else([&](){ return std::optional<unsigned>{arg + 42}; }) + .transform([](int a){ return a * 2u; }) + .value_or(0) == arg * 2u); + } + end: + ok = inner_ok; + } + return ok; +} + +struct my_error +{ + int _e; +}; + +std::expected<unsigned, my_error> make_expected(bool b, unsigned arg = 0u) noexcept +{ + if (!b) + return std::unexpected{my_error{-1}}; + return {arg}; +} + +bool test_expected(unsigned arg) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arg) + { + bool inner_ok = true; + { + auto unexpected = make_expected(false); + VERIFY (!unexpected); + VERIFY (!unexpected.has_value()); + VERIFY (unexpected.error()._e == -1); + VERIFY (unexpected.value_or(arg * 2u) == arg * 2u); + VERIFY (unexpected.or_else([&](my_error e){ return std::expected<unsigned, my_error>{arg}; }) + .transform([](int a){ return a * 2u; }) + .value_or(0) == arg * 2u); + + auto expected = make_expected(true, arg); + VERIFY (expected); + VERIFY (expected.has_value()); + VERIFY (expected.value() == arg); + VERIFY (*expected == arg); + VERIFY (expected.value_or(arg + 42) == arg); + VERIFY (expected.or_else([&](my_error e){ return std::expected<unsigned, my_error>{std::unexpected{e}}; }) + .transform([](int a){ return a * 2u; }) + .value_or(0) == arg * 2u); + } + end: + ok = inner_ok; + } + return ok; +} + +int main() +{ + volatile unsigned arg = 42; + return test_optional(arg) + && test_expected(arg) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-2003.C b/libgomp/testsuite/libgomp.c++/target-flex-2003.C new file mode 100644 index 0000000..8e8ca8e --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-2003.C @@ -0,0 +1,176 @@ +/* { dg-additional-options "-std=c++20" } */ + +/* bit_cast and memcpy */ + +#include <bit> +#include <cstring> + +#include "target-flex-common.h" + +struct S0 +{ + int _v0; + char _v1; + long long _v2; +}; + +struct S1 +{ + int _v0; + char _v1; + long long _v2; +}; + +bool test_bit_cast(int arg) +{ + bool ok; + S1 s1_out; + #pragma omp target map(from: ok, s1_out) map(to: arg) + { + bool inner_ok = true; + { + long long v = static_cast<long long>(arg + 42ll); + S0 s = {arg, 'a', v}; + VERIFY (std::bit_cast<S1>(s)._v0 == arg); + VERIFY (std::bit_cast<S1>(s)._v1 == 'a'); + VERIFY (std::bit_cast<S1>(s)._v2 == v); + s1_out = std::bit_cast<S1>(s); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + long long v = static_cast<long long>(arg + 42ll); + VERIFY_NON_TARGET (std::bit_cast<S0>(s1_out)._v0 == arg); + VERIFY_NON_TARGET (std::bit_cast<S0>(s1_out)._v1 == 'a'); + VERIFY_NON_TARGET (std::bit_cast<S0>(s1_out)._v2 == v); + return true; +} + + +struct OutStruct +{ + std::size_t _id; + void *_next; +}; + +struct Extendable1 +{ + std::size_t _id; + void *_next; + int _v; +}; + +struct Extendable2 +{ + std::size_t _id; + void *_next; + char _str[256]; +}; + +struct Extendable3 +{ + std::size_t _id; + void *_next; + const int *_nums; + std::size_t _size; +}; + +struct ExtendableUnknown +{ + std::size_t _id; + void *_next; +}; + +template<typename To, std::size_t Id> +To *get_extendable(void *p) +{ + while (p != nullptr) + { + OutStruct out; + std::memcpy(&out, p, sizeof(OutStruct)); + if (out._id == Id) + return static_cast<To *>(p); + p = out._next; + } + return nullptr; +} + +bool test_memcpy(int arg, const int *nums, std::size_t nums_size) +{ + bool ok; + Extendable2 e2_out; + #pragma omp target map(from: ok, e2_out) map(to: arg, nums[:nums_size], nums_size) + { + bool inner_ok = true; + { + Extendable3 e3 = {3u, nullptr, nums, nums_size}; + ExtendableUnknown u1 = {100u, &e3}; + Extendable2 e2 = {2u, &u1, {'H', 'e', 'l', 'l', 'o', '!', '\000'}}; + ExtendableUnknown u2 = {101u, &e2}; + ExtendableUnknown u3 = {102u, &u2}; + ExtendableUnknown u4 = {142u, &u3}; + Extendable1 e1 = {1u, &u4, arg}; + + void *p = &e1; + while (p != nullptr) + { + /* You can always cast a pointer to a struct to a pointer to + the type of it's first member. */ + switch (*static_cast<std::size_t *>(p)) + { + case 1: + { + Extendable1 *e1_p = static_cast<Extendable1 *>(p); + p = e1_p->_next; + VERIFY (e1_p->_v == arg); + break; + } + case 2: + { + Extendable2 *e2_p = static_cast<Extendable2 *>(p); + p = e2_p->_next; + VERIFY (std::strcmp(e2_p->_str, "Hello!") == 0); + break; + } + case 3: + { + Extendable3 *e3_p = static_cast<Extendable3 *>(p); + p = e3_p->_next; + VERIFY (nums == e3_p->_nums); + VERIFY (nums_size == e3_p->_size); + break; + } + default: + { + /* Casting to a pointer to OutStruct invokes undefined + behavior though, memcpy is required to extract the _next + member. */ + OutStruct out; + std::memcpy(&out, p, sizeof(OutStruct)); + p = out._next; + } + } + } + Extendable2 *e2_p = get_extendable<Extendable2, 2u>(&e1); + VERIFY (e2_p != nullptr); + e2_out = *e2_p; + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (e2_out._id == 2u); + VERIFY_NON_TARGET (std::strcmp(e2_out._str, "Hello!") == 0); + return true; +} + +int main() +{ + volatile int arg = 42; + int arr[8] = {0, 1, 2, 3, 4, 5, 6, 7}; + return test_bit_cast(arg) + && test_memcpy(arg, arr, 8) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-30.C b/libgomp/testsuite/libgomp.c++/target-flex-30.C new file mode 100644 index 0000000..c66075b --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-30.C @@ -0,0 +1,51 @@ +/* std::initializer_list in target region. */ + +#include <initializer_list> +#include <array> + +#include "target-flex-common.h" + +bool test_initializer_list(int arg) +{ + static constexpr std::size_t out_arr_size = 7; + int out_arr[out_arr_size]; + bool ok; + #pragma omp target map(from: ok, out_arr[:out_arr_size]) map(to: arg) + { + bool inner_ok = true; + { + auto il = {0, 1, 2, 3, 4, 5, arg}; + + int sum = 0; + for (auto const& e : il) + sum += e; + VERIFY (sum == 0 + 1 + 2 + 3 + 4 + 5 + arg); + + auto* out_it = out_arr; + const auto* const out_end = out_arr + out_arr_size; + for (auto const& e : il) + { + VERIFY (out_it != out_end); + *out_it = e; + ++out_it; + } + } + end: + ok = inner_ok; + } + if (!ok) + return false; + + std::array<int, out_arr_size> reference_array = {0, 1, 2, 3, 4, 5, arg}; + const auto *out_arr_it = out_arr; + for (auto const& e : reference_array) + VERIFY_NON_TARGET (e == *(out_arr_it++)); + + return true; +} + +int main() +{ + volatile int arg = 42; + return test_initializer_list(arg) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-300.C b/libgomp/testsuite/libgomp.c++/target-flex-300.C new file mode 100644 index 0000000..ef9e5a9 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-300.C @@ -0,0 +1,49 @@ +/* { dg-additional-options -std=c++23 } */ + +/* numerics */ + +#include <algorithm> +#include <numeric> +#include <ranges> +#include <span> +#include <vector> + +//TODO PR120454 "C++ constexpr vs. OpenMP implicit mapping" +#pragma omp declare target(std::ranges::all_of, std::ranges::iota) + +#include "target-flex-common.h" + +namespace stdr = std::ranges; + +bool test(std::size_t arg) +{ + bool ok; + int midpoint_out; + std::vector<int> vec(arg); + int *data = vec.data(); + std::size_t size = vec.size(); + #pragma omp target defaultmap(none) map(from: ok, midpoint_out) map(tofrom: data[:size]) map(to: arg, size) + { + std::span span = {data, size}; + bool inner_ok = true; + { + VERIFY (stdr::all_of(span, [](int v){ return v == int{}; })); + stdr::iota(span, 0); + midpoint_out = *std::midpoint(span.data(), span.data() + span.size()); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (stdr::equal(vec, std::views::iota(0, static_cast<int>(vec.size())))); + VERIFY_NON_TARGET (*std::midpoint(vec.data(), vec.data() + vec.size()) + == midpoint_out); + return true; +} + +int main() +{ + volatile std::size_t arg = 42; + return test(arg) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-31.C b/libgomp/testsuite/libgomp.c++/target-flex-31.C new file mode 100644 index 0000000..adaf18f --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-31.C @@ -0,0 +1,80 @@ +/* std::initializer_list in target region. */ + +#include <initializer_list> + +#include "target-flex-common.h" + +struct S0 +{ + int _v; + S0(std::initializer_list<int> il) + : _v(0) + { + for (auto const& e : il) + _v += e; + } +}; + +struct S1 +{ + int _v; + template<typename T> + S1(std::initializer_list<T> il) + : _v(0) + { + for (auto const& e : il) + _v += e; + } +}; + +template<typename T> +struct S2 +{ + T _v; + S2(std::initializer_list<T> il) + : _v(0) + { + for (auto const& e : il) + _v += e; + } +}; + +#if __cplusplus >= 201703L +template<typename T> +S2(std::initializer_list<T>) -> S2<T>; +#endif + +bool test_initializer_list(int arg) +{ + bool ok; + #pragma omp target map(from: ok) map(to: arg) + { + bool inner_ok = true; + { + static constexpr int partial_sum = 0 + 1 + 2 + 3 + 4 + 5; + + S0 s0{0, 1, 2, 3, 4, 5, arg}; + VERIFY (s0._v == partial_sum + arg); + + S1 s1{0, 1, 2, 3, 4, 5, arg}; + VERIFY (s1._v == partial_sum + arg); + + S2<int> s2{0, 1, 2, 3, 4, 5, arg}; + VERIFY (s2._v == partial_sum + arg); + + #if __cplusplus >= 201703L + S2 s2_ctad{0, 1, 2, 3, 4, 5, arg}; + VERIFY (s2_ctad._v == partial_sum + arg); + #endif + } + end: + ok = inner_ok; + } + return ok; +} + +int main() +{ + volatile int arg = 42; + return test_initializer_list(arg) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-32.C b/libgomp/testsuite/libgomp.c++/target-flex-32.C new file mode 100644 index 0000000..7f74401a --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-32.C @@ -0,0 +1,50 @@ +/* std::initializer_list constructor of std::vector (explicit template arg) */ + +#include <vector> +#include <array> + +#include "target-flex-common.h" + +bool test_initializer_list(int arg) +{ + static constexpr std::size_t out_arr_size = 7; + int out_arr[out_arr_size]; + bool ok; + #pragma omp target map(from: ok, out_arr[:out_arr_size]) map(to: arg) + { + bool inner_ok = true; + { + std::vector<int> vec{0, 1, 2, 3, 4, 5, arg}; + int sum = 0; + for (auto const& e : vec) + sum += e; + VERIFY (sum == 0 + 1 + 2 + 3 + 4 + 5 + arg); + + auto* out_it = out_arr; + const auto* const out_end = out_arr + out_arr_size; + for (auto const& e : vec) + { + VERIFY (out_it != out_end); + *out_it = e; + ++out_it; + } + } + end: + ok = inner_ok; + } + if (!ok) + return false; + + std::array<int, out_arr_size> reference_array = {0, 1, 2, 3, 4, 5, arg}; + const auto *out_arr_it = out_arr; + for (auto const& e : reference_array) + VERIFY_NON_TARGET (e == *(out_arr_it++)); + + return true; +} + +int main() +{ + volatile int arg = 42; + return test_initializer_list(arg) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-33.C b/libgomp/testsuite/libgomp.c++/target-flex-33.C new file mode 100644 index 0000000..bb8a39b --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-33.C @@ -0,0 +1,52 @@ +/* { dg-additional-options "-std=c++17" } */ + +/* deduced std::initializer_list constructor of std::vector (CTAD) */ + +#include <vector> +#include <array> + +#include "target-flex-common.h" + +bool test_initializer_list(int arg) +{ + static constexpr std::size_t out_arr_size = 7; + int out_arr[out_arr_size]; + bool ok; + #pragma omp target map(from: ok, out_arr[:out_arr_size]) map(to: arg) + { + bool inner_ok = true; + { + std::vector vec{0, 1, 2, 3, 4, 5, arg}; + int sum = 0; + for (auto const& e : vec) + sum += e; + VERIFY (sum == 0 + 1 + 2 + 3 + 4 + 5 + arg); + + auto* out_it = out_arr; + const auto* const out_end = out_arr + out_arr_size; + for (auto const& e : vec) + { + VERIFY (out_it != out_end); + *out_it = e; + ++out_it; + } + } + end: + ok = inner_ok; + } + if (!ok) + return false; + + std::array<int, out_arr_size> reference_array = {0, 1, 2, 3, 4, 5, arg}; + const auto *out_arr_it = out_arr; + for (auto const& e : reference_array) + VERIFY_NON_TARGET (e == *(out_arr_it++)); + + return true; +} + +int main() +{ + volatile int arg = 42; + return test_initializer_list(arg) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-41.C b/libgomp/testsuite/libgomp.c++/target-flex-41.C new file mode 100644 index 0000000..4d36341 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-41.C @@ -0,0 +1,94 @@ +/* { dg-additional-options "-std=c++20" } */ + +/* <iterator> c++20 */ + +/* std::common_iterator uses std::variant. */ + +#include <vector> +#include <iterator> +#include <span> + +//TODO PR120454 "C++ constexpr vs. OpenMP implicit mapping" +#pragma omp declare target(std::ranges::distance, std::ranges::next) + +#include "target-flex-common.h" + +namespace stdr = std::ranges; + +template<typename It0, typename It1> +bool simple_equal(const It0 begin0, const It0 end0, + const It1 begin1, const It1 end1) BL_NOEXCEPT +{ + It0 it0 = begin0; + It1 it1 = begin1; + for (; it0 != end0; ++it0, ++it1) + if (it1 == end1 || *it0 != *it1) + return false; + return true; +} + +template<typename It, typename OutIt> +void simple_copy(const It begin, const It end, OutIt out) BL_NOEXCEPT +{ + for (It it = begin; it != end; ++it, ++out) + *out = *it; +} + +template<typename T, std::size_t Size> +bool test(const T (&arr)[Size]) +{ + bool ok; + T out_rev_arr[Size]; + T out_fwd_arr[Size]; + T out_first_half_arr[Size / 2]; + #pragma omp target defaultmap(none) \ + map(from: ok, out_rev_arr[:Size], out_fwd_arr[:Size], \ + out_first_half_arr[:Size / 2]) \ + map(to: arr[:Size]) + { + bool inner_ok = true; + { + std::span<const T> span = {arr, Size}; + std::vector<T> rev_vec(std::reverse_iterator{span.end()}, + std::reverse_iterator{span.begin()}); + VERIFY (std::distance(span.begin(), span.end()) + == std::distance(rev_vec.begin(), rev_vec.end())); + VERIFY (stdr::distance(span.begin(), span.end()) + == stdr::distance(rev_vec.begin(), rev_vec.end())); + VERIFY (stdr::distance(span) == stdr::distance(rev_vec)); + VERIFY (simple_equal(span.begin(), span.end(), + std::reverse_iterator{rev_vec.end()}, + std::reverse_iterator{rev_vec.begin()})); + simple_copy(rev_vec.begin(), rev_vec.end(), out_rev_arr); + simple_copy(std::reverse_iterator{rev_vec.end()}, + std::reverse_iterator{rev_vec.begin()}, + out_fwd_arr); + using counted_iter = std::counted_iterator<decltype(span.begin())>; + using common_iter = std::common_iterator<counted_iter, + std::default_sentinel_t>; + std::vector<T> front_half; + simple_copy(common_iter{counted_iter{span.begin(), Size / 2}}, + common_iter{std::default_sentinel}, + std::back_insert_iterator{front_half}); + VERIFY (simple_equal(span.begin(), stdr::next(span.begin(), Size / 2), + front_half.begin(), front_half.end())); + simple_copy(front_half.begin(), front_half.end(), out_first_half_arr); + } + end: + ok = inner_ok; + } + VERIFY_NON_TARGET (simple_equal(std::reverse_iterator{arr + Size}, + std::reverse_iterator{arr}, + out_rev_arr, out_rev_arr + Size)); + VERIFY_NON_TARGET (simple_equal(arr, arr + Size, + out_fwd_arr, out_fwd_arr + Size)); + VERIFY_NON_TARGET (simple_equal(arr, arr + Size / 2, + out_first_half_arr, out_first_half_arr + Size / 2)); + return ok; +} + +int main() +{ + int arr[] = {0, 1, 2, 3, 4, 5, 6, 7}; + return test(arr) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-60.C b/libgomp/testsuite/libgomp.c++/target-flex-60.C new file mode 100644 index 0000000..014b9f5 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-60.C @@ -0,0 +1,46 @@ +/* algorithms pre c++20 */ + +#include <algorithm> +#include <vector> + +#include "target-flex-common.h" + +template<typename T, std::size_t Size> +bool test(const T (&arr)[Size]) +{ + bool ok; + T out_2x_arr[Size]; + T out_shifted_arr[Size]; + #pragma omp target map(from: ok, out_2x_arr[:Size], out_shifted_arr[:Size]) \ + map(to: arr[:Size]) + { + std::vector<T> vec(Size); + std::vector<T> mutated(Size); + bool inner_ok = true; + { + std::copy(arr, arr + Size, vec.begin()); + VERIFY (std::equal(arr, arr + Size, vec.begin())); + std::transform(vec.begin(), vec.end(), mutated.begin(), + [](const T& v){ return v * 2; }); + std::copy(mutated.begin(), mutated.end(), out_2x_arr); + std::rotate(vec.begin(), std::next(vec.begin(), Size / 2), vec.end()); + std::copy(vec.begin(), vec.end(), out_shifted_arr); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (std::equal(arr, arr + Size, out_2x_arr, + [](const T& a, const T& b){ return a * 2 == b; })); + std::vector<T> shifted(arr, arr + Size); + std::rotate(shifted.begin(), std::next(shifted.begin(), Size / 2), shifted.end()); + VERIFY_NON_TARGET (std::equal(out_shifted_arr, out_shifted_arr + Size, shifted.begin())); + return true; +} + +int main() +{ + int arr[] = {0, 1, 2, 3, 4, 5, 6, 7}; + return test(arr) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-61.C b/libgomp/testsuite/libgomp.c++/target-flex-61.C new file mode 100644 index 0000000..9070c2d --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-61.C @@ -0,0 +1,54 @@ +/* { dg-additional-options "-std=c++20" } */ + +/* ranged algorithms c++20 */ + +#include <algorithm> +#include <ranges> +#include <vector> + +//TODO PR120454 "C++ constexpr vs. OpenMP implicit mapping" +#pragma omp declare target(std::ranges::copy, std::ranges::equal, std::ranges::rotate, std::ranges::transform) + +#include "target-flex-common.h" + +namespace stdr = std::ranges; + +template<typename T, std::size_t Size> +bool test(const T (&arr)[Size]) +{ + bool ok; + T out_2x_arr[Size]; + T out_shifted_arr[Size]; + #pragma omp target defaultmap(none) \ + map(from: ok, out_2x_arr[:Size], out_shifted_arr[:Size]) \ + map(to: arr[:Size]) + { + std::vector<T> vec(Size); + std::vector<T> mutated(Size); + bool inner_ok = true; + { + stdr::copy(arr, vec.begin()); + VERIFY (stdr::equal(arr, vec)); + stdr::transform(vec, mutated.begin(), + [](const T& v){ return v * 2; }); + stdr::copy(mutated, out_2x_arr); + stdr::rotate(vec, std::next(vec.begin(), Size / 2)); + stdr::copy(vec, out_shifted_arr); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (stdr::equal(arr, out_2x_arr, stdr::equal_to{}, [](const T& v){ return v * 2; })); + std::vector<T> shifted(arr, arr + Size); + stdr::rotate(shifted, std::next(shifted.begin(), Size / 2)); + VERIFY_NON_TARGET (stdr::equal(out_shifted_arr, shifted)); + return true; +} + +int main() +{ + int arr[] = {0, 1, 2, 3, 4, 5, 6, 7}; + return test(arr) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-62.C b/libgomp/testsuite/libgomp.c++/target-flex-62.C new file mode 100644 index 0000000..ef6b942 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-62.C @@ -0,0 +1,50 @@ +/* { dg-additional-options -std=c++23 } */ + +/* std::views stuff. Also tests std::tuple with std::views::zip. */ + +#include <algorithm> +#include <ranges> +#include <span> + +//TODO PR120454 "C++ constexpr vs. OpenMP implicit mapping" +#pragma omp declare target(std::ranges::all_of, std::ranges::equal, std::ranges::fold_left, std::views::reverse, std::views::zip) + +#include "target-flex-common.h" + +namespace stdr = std::ranges; +namespace stdv = std::views; + +bool f() +{ + const int arr_fwd[8] = {0, 1, 2, 3, 4, 5, 6, 7}; + const int arr_rev[8] = {7, 6, 5, 4, 3, 2, 1, 0}; + + bool ok; + #pragma omp target defaultmap(none) map(from: ok) map(to: arr_fwd[:8], arr_rev[:8]) + { + std::span<const int> fwd = {arr_fwd, 8}; + std::span<const int> rev = {arr_rev, 8}; + bool inner_ok = true; + { + VERIFY(stdr::equal(fwd, rev | stdv::reverse)); + VERIFY(stdr::equal(fwd | stdv::drop(4) | stdv::reverse, + rev | stdv::take(4))); + for (auto [first, second] : stdv::zip(fwd, rev)) + VERIFY(first + second == 7); + auto plus = [](int a, int b){ return a + b; }; + auto is_even = [](int v){ return v % 2 == 0; }; + VERIFY(stdr::fold_left(fwd | stdv::filter(is_even), 0, plus) + == 12); + VERIFY(stdr::all_of(fwd | stdv::transform([](int v){ return v * 2; }), + is_even)); + } + end: + ok = inner_ok; + } + return ok; +} + +int main() +{ + return f() ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-70.C b/libgomp/testsuite/libgomp.c++/target-flex-70.C new file mode 100644 index 0000000..9e9383d --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-70.C @@ -0,0 +1,26 @@ +/* CTAD in target regions. */ + +template<typename T> +struct S +{ + T _v; +}; + +template<typename T> +S(T) -> S<T>; + +bool f() +{ + bool ok; + #pragma omp target map(from: ok) + { + S s{42}; + ok = s._v == 42; + } + return ok; +} + +int main() +{ + return f() ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-80.C b/libgomp/testsuite/libgomp.c++/target-flex-80.C new file mode 100644 index 0000000..f41a1bb --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-80.C @@ -0,0 +1,49 @@ +// { dg-additional-options "-std=c++20" } + +/* std::span */ + +#include <span> + +#include "target-flex-common.h" + +template<typename It0, typename It1> +bool simple_equal(It0 it0, const It0 end0, + It1 it1, const It1 end1) noexcept +{ + for (; it0 != end0; ++it0, ++it1) + if (it1 == end1 || *it0 != *it1) + return false; + return true; +} + +template<typename T, std::size_t Size> +bool test(const T (&arr)[Size]) +{ + bool ok; + T out_arr[Size]; + #pragma omp target map(from: ok) map(to: arr[:Size]) + { + std::span span = {arr, Size}; + bool inner_ok = true; + { + VERIFY (!span.empty()); + VERIFY (span.size() == Size); + auto out_it = out_arr; + for (auto elem : span) + *(out_it++) = elem; + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (simple_equal(arr, arr + Size, + out_arr, out_arr + Size)); + return true; +} + +int main() +{ + int arr[8] = {0, 1, 2, 3, 4, 5, 6, 7}; + return test(arr) ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-81.C b/libgomp/testsuite/libgomp.c++/target-flex-81.C new file mode 100644 index 0000000..a86fefb --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-81.C @@ -0,0 +1,75 @@ +/* { dg-additional-options "-std=c++20" } */ + +#include <ranges> +#include <span> +#include <type_traits> +#include <vector> + +#include "target-flex-common.h" + +namespace stdr = std::ranges; + +template<typename It0, typename It1> +bool simple_equal(It0 it0, const It0 end0, + It1 it1, const It1 end1) noexcept +{ + for (; it0 != end0; ++it0, ++it1) + if (it1 == end1 || *it0 != *it1) + return false; + return true; +} + +template<typename Rn0, typename Rn1> +bool simple_equal(Rn0&& rn0, Rn1&& rn1) noexcept +{ + return simple_equal(stdr::begin(rn0), stdr::end(rn0), + stdr::begin(rn1), stdr::end(rn1)); +} + +template<typename Rn> +bool test(Rn&& range) +{ + using value_type = stdr::range_value_t<std::remove_cvref_t<Rn>>; + std::vector<value_type> vec = {stdr::begin(range), stdr::end(range)}; + value_type *data = vec.data(); + std::size_t size = vec.size(); + bool ok; + #pragma omp target map(from: ok) map(tofrom: data[:size]) map(to: size) + { + std::vector<value_type> orig = {data, data + size}; + std::span<value_type> span = {data, size}; + bool inner_ok = true; + { + auto mul_by_2 = [](const value_type& v){ return v * 2; }; + VERIFY (simple_equal(orig, span)); + for (auto& elem : span) + elem = mul_by_2(elem); + VERIFY (simple_equal(orig | std::views::transform(mul_by_2), span)); + } + end: + ok = inner_ok; + } + if (!ok) + return false; + auto mul_by_2 = [](const value_type& v){ return v * 2; }; + VERIFY_NON_TARGET (simple_equal(range | std::views::transform(mul_by_2), vec)); + return true; +} + +struct my_int +{ + int _v; + bool operator==(my_int const&) const = default; + my_int operator*(int rhs) const noexcept { + return {_v * rhs}; + } +}; + +int main() +{ + std::vector<int> ints = {1, 2, 3, 4, 5}; + const bool ints_res = test(ints); + std::vector<my_int> my_ints = {my_int{1}, my_int{2}, my_int{3}, my_int{4}, my_int{5}}; + const bool my_ints_res = test(my_ints); + return ints_res && my_ints_res ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-90.C b/libgomp/testsuite/libgomp.c++/target-flex-90.C new file mode 100644 index 0000000..b3f1197 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-90.C @@ -0,0 +1,107 @@ +/* structured bindings */ + +#include <array> +#include <tuple> + +#include "target-flex-common.h" + +template<typename Array, typename Tuple, typename Struct> +bool test(Array array, Tuple tuple, Struct s) +{ + bool ok; + auto array_2nd_in = std::get<2>(array); + auto tuple_2nd_in = std::get<2>(tuple); + auto s_2nd_in = s._2; + decltype(array_2nd_in) array_2nd_out_0; + decltype(tuple_2nd_in) tuple_2nd_out_0; + decltype(s_2nd_in) s_2nd_out_0; + decltype(array_2nd_in) array_2nd_out_1; + decltype(tuple_2nd_in) tuple_2nd_out_1; + decltype(s_2nd_in) s_2nd_out_1; + decltype(array_2nd_in) array_2nd_out_2; + decltype(tuple_2nd_in) tuple_2nd_out_2; + decltype(s_2nd_in) s_2nd_out_2; + #pragma omp target map(from: ok, \ + array_2nd_out_0, tuple_2nd_out_0, s_2nd_out_0, \ + array_2nd_out_1, tuple_2nd_out_1, s_2nd_out_1, \ + array_2nd_out_2, tuple_2nd_out_2, s_2nd_out_2) \ + map(to: array_2nd_in, tuple_2nd_in, s_2nd_in, array, tuple, s) + { + bool inner_ok = true; + { + { + auto [array_0th, array_1st, array_2nd] = array; + VERIFY (array_2nd_in == array_2nd); + VERIFY (std::get<2>(array) == array_2nd); + array_2nd_out_0 = array_2nd; + auto [tuple_0th, tuple_1st, tuple_2nd] = tuple; + VERIFY (tuple_2nd_in == tuple_2nd); + VERIFY (std::get<2>(tuple) == tuple_2nd); + tuple_2nd_out_0 = tuple_2nd; + auto [s_0th, s_1st, s_2nd] = s; + VERIFY (s_2nd_in == s_2nd); + VERIFY (s._2 == s_2nd); + s_2nd_out_0 = s_2nd; + } + { + auto& [array_0th, array_1st, array_2nd] = array; + VERIFY (array_2nd_in == array_2nd); + VERIFY (std::get<2>(array) == array_2nd); + array_2nd_out_1 = array_2nd; + auto& [tuple_0th, tuple_1st, tuple_2nd] = tuple; + VERIFY (tuple_2nd_in == tuple_2nd); + VERIFY (std::get<2>(tuple) == tuple_2nd); + tuple_2nd_out_1 = tuple_2nd; + auto& [s_0th, s_1st, s_2nd] = s; + VERIFY (s_2nd_in == s_2nd); + VERIFY (s._2 == s_2nd); + s_2nd_out_1 = s_2nd; + } + { + const auto& [array_0th, array_1st, array_2nd] = array; + VERIFY (array_2nd_in == array_2nd); + VERIFY (std::get<2>(array) == array_2nd); + array_2nd_out_2 = array_2nd; + const auto& [tuple_0th, tuple_1st, tuple_2nd] = tuple; + VERIFY (tuple_2nd_in == tuple_2nd); + VERIFY (std::get<2>(tuple) == tuple_2nd); + tuple_2nd_out_2 = tuple_2nd; + const auto& [s_0th, s_1st, s_2nd] = s; + VERIFY (s_2nd_in == s_2nd); + VERIFY (s._2 == s_2nd); + s_2nd_out_2 = s_2nd; + } + } + end: + ok = inner_ok; + } + if (!ok) + return false; + VERIFY_NON_TARGET (array_2nd_out_0 == array_2nd_in); + VERIFY_NON_TARGET (tuple_2nd_out_0 == tuple_2nd_in); + VERIFY_NON_TARGET (s_2nd_out_0 == s_2nd_in); + VERIFY_NON_TARGET (array_2nd_out_1 == array_2nd_in); + VERIFY_NON_TARGET (tuple_2nd_out_1 == tuple_2nd_in); + VERIFY_NON_TARGET (s_2nd_out_1 == s_2nd_in); + VERIFY_NON_TARGET (array_2nd_out_2 == array_2nd_in); + VERIFY_NON_TARGET (tuple_2nd_out_2 == tuple_2nd_in); + VERIFY_NON_TARGET (s_2nd_out_2 == s_2nd_in); + + return true; +} + +struct S +{ + char _0; + float _1; + int _2; +}; + +int main() +{ + const bool test_res + = test(std::array{0, 1, 2}, + std::tuple{'a', 3.14f, 42}, + S{'a', 3.14f, 42}); + return test_res ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-flex-common.h b/libgomp/testsuite/libgomp.c++/target-flex-common.h new file mode 100644 index 0000000..14523c4 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-flex-common.h @@ -0,0 +1,40 @@ +#include <cstdio> + +#if __cplusplus >= 201103L + #define BL_NOEXCEPT noexcept +#else + #define BL_NOEXCEPT throw() +#endif + +#if defined __has_builtin +# if __has_builtin (__builtin_LINE) +# define VERIFY_LINE __builtin_LINE () +# endif +#endif +#if !defined VERIFY_LINE +# define VERIFY_LINE __LINE__ +#endif + +/* I'm not a huge fan of macros but in the interest of keeping the code that + isn't being tested as simple as possible, we use them. */ + +#define VERIFY(EXPR) \ + do { \ + if (!(EXPR)) \ + { \ + std::printf("VERIFY ln: %d `" #EXPR "` evaluated to false\n", \ + VERIFY_LINE); \ + inner_ok = false; \ + goto end; \ + } \ + } while (false) + +#define VERIFY_NON_TARGET(EXPR) \ + do { \ + if (!(EXPR)) \ + { \ + std::printf("VERIFY ln: %d `" #EXPR "` evaluated to false\n", \ + VERIFY_LINE); \ + return false; \ + } \ + } while (false) diff --git a/libgomp/testsuite/libgomp.c++/target-std__array-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__array-concurrent-usm.C new file mode 100644 index 0000000..9923783 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__array-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__array-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__array-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__array-concurrent.C new file mode 100644 index 0000000..c42105a --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__array-concurrent.C @@ -0,0 +1,62 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <array> +#include <algorithm> + +#define N 50000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand (); +} + +#pragma omp declare target +bool validate (const std::array<int,N> &arr, int data[]) +{ + for (int i = 0; i < N; ++i) + if (arr[i] != data[i] * data[i]) + return false; + return true; +} +#pragma omp end declare target + +int main (void) +{ + int data[N]; + bool ok; + std::array<int,N> arr; + + srand (time (NULL)); + init (data); + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: arr) +#endif + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&arr) std::array<int,N> (); +#endif + std::copy (data, data + N, arr.begin ()); + } + + #pragma omp target teams distribute parallel for + for (int i = 0; i < N; ++i) + arr[i] *= arr[i]; + + #pragma omp target map (from: ok) + { + ok = validate (arr, data); +#ifndef MEM_SHARED + arr.~array (); +#endif + } + } + + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__bitset-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__bitset-concurrent-usm.C new file mode 100644 index 0000000..9023ef8 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__bitset-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__bitset-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__bitset-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__bitset-concurrent.C new file mode 100644 index 0000000..4fcce93 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__bitset-concurrent.C @@ -0,0 +1,69 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <bitset> +#include <set> +#include <algorithm> + +#define N 4000 +#define MAX 16384 + +void init (int data[]) +{ + std::set<int> _set; + for (int i = 0; i < N; ++i) + { + // Avoid duplicates in data array. + do + data[i] = rand () % MAX; + while (_set.find (data[i]) != _set.end ()); + _set.insert (data[i]); + } +} + +bool validate (int sum, int data[]) +{ + int total = 0; + for (int i = 0; i < N; ++i) + total += data[i]; + return sum == total; +} + +int main (void) +{ + int data[N]; + std::bitset<MAX> _set; + int sum = 0; + + srand (time (NULL)); + init (data); + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: _set) +#endif + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&_set) std::bitset<MAX> (); +#endif + for (int i = 0; i < N; ++i) + _set[data[i]] = true; + } + + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < MAX; ++i) + if (_set[i]) + sum += i; + +#ifndef MEM_SHARED + #pragma omp target + _set.~bitset (); +#endif + } + + bool ok = validate (sum, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__cmath.C b/libgomp/testsuite/libgomp.c++/target-std__cmath.C new file mode 100644 index 0000000..aaf7152 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__cmath.C @@ -0,0 +1,340 @@ +// { dg-do run } +// { dg-additional-options "-std=c++20" } + +#include <cmath> +#include <numbers> + +#define FP_EQUAL(x,y) (std::abs ((x) - (y)) < 1E-6) + +#pragma omp declare target +template<typename T> bool test_basic () +{ + T x = -3.456789; + T y = 1.234567; + T z = 5.678901; + + if (std::abs (x) != -x) + return false; + if (!FP_EQUAL (std::trunc (x / y) * y + std::fmod (x, y), x)) + return false; + if (!FP_EQUAL (x - std::round (x / y) * y, std::remainder (x, y))) + return false; + if (!FP_EQUAL (std::fma (x, y, z), x * y + z)) + return false; + if (std::fmax (x, y) != (x > y ? x : y)) + return false; + if (std::fmin (x, y) != (x < y ? x : y)) + return false; + if (std::fdim (x, y) != std::max(x - y, (T) 0.0)) + return false; + if (std::fdim (y, x) != std::max(y - x, (T) 0.0)) + return false; + return true; +} + +template<typename T> bool test_exp () +{ + T x = -4.567890; + T y = 2.345678; + + if (!FP_EQUAL (std::exp (x), std::pow (std::numbers::e_v<T>, x))) + return false; + if (!FP_EQUAL (std::exp2 (y), std::pow ((T) 2.0, y))) + return false; + if (!FP_EQUAL (std::expm1 (y), std::exp (y) - (T) 1.0)) + return false; + if (!FP_EQUAL (std::log (std::exp (x)), x)) + return false; + if (!FP_EQUAL (std::log10 (std::pow ((T) 10.0, y)), y)) + return false; + if (!FP_EQUAL (std::log2 (std::exp2 (y)), y)) + return false; + if (!FP_EQUAL (std::log1p (std::expm1 (y)), y)) + return false; + return true; +} + +template<typename T> bool test_power () +{ + T x = 7.234251; + T y = 0.340128; + + if (!FP_EQUAL (std::log (std::pow (x, y)) / std::log (x), y)) + return false; + if (!FP_EQUAL (std::sqrt (x) * std::sqrt (x), x)) + return false; + if (!FP_EQUAL (std::cbrt (x) * std::cbrt (x) * std::cbrt (x), x)) + return false; + if (!FP_EQUAL (std::hypot (x, y), std::sqrt (x * x + y * y))) + return false; + return true; +} + +template<typename T> bool test_trig () +{ + T theta = std::numbers::pi / 4; + T phi = std::numbers::pi / 6; + + if (!FP_EQUAL (std::sin (theta), std::sqrt ((T) 2) / 2)) + return false; + if (!FP_EQUAL (std::sin (phi), 0.5)) + return false; + if (!FP_EQUAL (std::cos (theta), std::sqrt ((T) 2) / 2)) + return false; + if (!FP_EQUAL (std::cos (phi), std::sqrt ((T) 3) / 2)) + return false; + if (!FP_EQUAL (std::tan (theta), 1.0)) + return false; + if (!FP_EQUAL (std::tan (phi), std::sqrt ((T) 3) / 3)) + return false; + + T x = 0.33245623; + + if (!FP_EQUAL (std::asin (std::sin (x)), x)) + return false; + if (!FP_EQUAL (std::acos (std::cos (x)), x)) + return false; + if (!FP_EQUAL (std::atan (std::tan (x)), x)) + return false; + if (!FP_EQUAL (std::atan2 (std::sin (x), std::cos (x)), x)) + return false; + return true; +} + +template<typename T> bool test_hyperbolic () +{ + T x = 0.7423532; + + if (!FP_EQUAL (std::sinh (x), (std::exp (x) - std::exp (-x)) / (T) 2.0)) + return false; + if (!FP_EQUAL (std::cosh (x), (std::exp (x) + std::exp (-x)) / (T) 2.0)) + return false; + if (!FP_EQUAL (std::tanh (x), std::sinh (x) / std::cosh (x))) + return false; + if (!FP_EQUAL (std::asinh (std::sinh (x)), x)) + return false; + if (!FP_EQUAL (std::acosh (std::cosh (x)), x)) + return false; + if (!FP_EQUAL (std::atanh (std::tanh (x)), x)) + return false; + return true; +} + +template<typename T> bool test_erf () +{ + if (!FP_EQUAL (std::erf ((T) 0), 0)) + return false; + if (!FP_EQUAL (std::erf ((T) INFINITY), 1)) + return false; + if (!FP_EQUAL (std::erf ((T) -INFINITY), -1)) + return false; + + if (!FP_EQUAL (std::erfc (0), 1)) + return false; + if (!FP_EQUAL (std::erfc ((T) INFINITY), 0)) + return false; + if (!FP_EQUAL (std::erfc ((T) -INFINITY), 2)) + return false; + + return true; +} + +template<typename T> bool test_gamma () +{ + if (!FP_EQUAL (std::tgamma ((T) 5), 4*3*2*1)) + return false; + if (!FP_EQUAL (std::tgamma ((T) 0.5), std::sqrt (std::numbers::pi_v<T>))) + return false; + if (!FP_EQUAL (std::tgamma ((T) -0.5), (T) -2 * std::sqrt (std::numbers::pi_v<T>))) + return false; + if (!FP_EQUAL (std::tgamma ((T) 2.5), (T) 0.75 * std::sqrt (std::numbers::pi_v<T>))) + return false; + if (!FP_EQUAL (std::tgamma ((T) -2.5), (T) -8.0/15 * std::sqrt (std::numbers::pi_v<T>))) + return false; + + if (!FP_EQUAL (std::lgamma ((T) 5), std::log ((T) 4*3*2*1))) + return false; + if (!FP_EQUAL (std::lgamma ((T) 0.5), std::log (std::sqrt (std::numbers::pi_v<T>)))) + return false; + if (!FP_EQUAL (std::lgamma ((T) 2.5), + std::log ((T) 0.75 * std::sqrt (std::numbers::pi_v<T>)))) + return false; + + return true; +} + +template<typename T> bool test_rounding () +{ + T x = -2.5678; + T y = 3.6789; + + if (std::ceil (x) != -2) + return false; + if (std::floor (x) != -3) + return false; + if (std::trunc (x) != -2) + return false; + if (std::round (x) != -3) + return false; + + if (std::ceil (y) != 4) + return false; + if (std::floor (y) != 3) + return false; + if (std::trunc (y) != 3) + return false; + if (std::round (y) != 4) + return false; + + /* Not testing std::rint and std::nearbyint due to dependence on + floating-point environment. */ + + return true; +} + +template<typename T> bool test_fpmanip () +{ + T x = -2.3456789; + T y = 3.6789012; + int exp; + + T mantissa = std::frexp (x, &exp); + if (std::ldexp (mantissa, exp) != x) + return false; + if (std::logb (x) + 1 != exp) + return false; + if (std::ilogb (x) + 1 != exp) + return false; + if (std::scalbn (x, -exp) != mantissa) + return false; + + T next = std::nextafter (x, y); + if (!(next > x && next < y)) + return false; + +#if 0 + /* TODO Due to 'std::nexttoward' using 'long double to', this triggers a + '80-bit-precision floating-point numbers unsupported (mode ‘XF’)' error + with x86_64 host and nvptx, GCN offload compilers, or + '128-bit-precision floating-point numbers unsupported (mode ‘TF’)' error + with powerpc64le host and nvptx offload compiler, for example; + PR71064 'nvptx offloading: "long double" data type'. + It ought to work on systems where the host's 'long double' is the same as + 'double' ('DF'): aarch64, for example? */ + next = std::nexttoward (x, y); + if (!(next > x && next < y)) + return false; +#endif + + if (std::copysign (x, y) != std::abs (x)) + return false; + if (std::copysign (y, x) != -y) + return false; + + return true; +} + +template<typename T> bool test_classify () +{ + T x = -2.3456789; + T y = 3.6789012; + + if (std::fpclassify (x) != FP_NORMAL || std::fpclassify (y) != FP_NORMAL) + return false; + if (std::fpclassify ((T) INFINITY) != FP_INFINITE + || std::fpclassify ((T) -INFINITY) != FP_INFINITE) + return false; + if (std::fpclassify ((T) 0.0) != FP_ZERO) + return false; + if (std::fpclassify ((T) NAN) != FP_NAN) + return false; + if (!std::isfinite (x) || !std::isfinite (y)) + return false; + if (std::isfinite ((T) INFINITY) || std::isfinite ((T) -INFINITY)) + return false; + if (std::isinf (x) || std::isinf (y)) + return false; + if (!std::isinf ((T) INFINITY) || !std::isinf ((T) -INFINITY)) + return false; + if (std::isnan (x) || std::isnan (y)) + return false; + if (!std::isnan ((T) 0.0 / (T) 0.0)) + return false; + if (std::isnan (x) || std::isnan (y)) + return false; + if (!std::isnormal (x) || !std::isnormal (y)) + return false; + if (std::isnormal ((T) 0.0) || std::isnormal ((T) INFINITY) || std::isnormal ((T) NAN)) + return false; + if (!std::signbit (x) || std::signbit (y)) + return false; + + return true; +} + +template<typename T> bool test_compare () +{ + T x = 5.6789012; + T y = 8.9012345; + + if (std::isgreater (x, y)) + return false; + if (std::isgreater (x, x)) + return false; + if (std::isgreaterequal (x, y)) + return false; + if (!std::isgreaterequal (x, x)) + return false; + if (!std::isless (x, y)) + return false; + if (std::isless (x, x)) + return false; + if (!std::islessequal (x, y)) + return false; + if (!std::islessequal (x, x)) + return false; + if (!std::islessgreater (x, y)) + return false; + if (std::islessgreater (x, x)) + return false; + if (std::isunordered (x, y)) + return false; + if (!std::isunordered (x, NAN)) + return false; + return true; +} +#pragma omp end declare target + +#define RUN_TEST(func) \ +{ \ + pass++; \ + bool ok = test_##func<float> (); \ + if (!ok) { result = pass; break; } \ + pass++; \ + ok = test_##func<double> (); \ + if (!ok) { result = pass; break; } \ +} + +int main (void) +{ + int result = 0; + + #pragma omp target map (tofrom: result) + do { + int pass = 0; + + RUN_TEST (basic); + RUN_TEST (exp); + RUN_TEST (power); + RUN_TEST (trig); + RUN_TEST (hyperbolic); + RUN_TEST (erf); + RUN_TEST (gamma); + RUN_TEST (rounding); + RUN_TEST (fpmanip); + RUN_TEST (classify); + RUN_TEST (compare); + } while (false); + + return result; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__complex.C b/libgomp/testsuite/libgomp.c++/target-std__complex.C new file mode 100644 index 0000000..e392d17 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__complex.C @@ -0,0 +1,175 @@ +// { dg-do run } +// { dg-additional-options "-std=c++20" } + +#include <cmath> +#include <complex> +#include <numbers> + +using namespace std::complex_literals; + +#define FP_EQUAL(x,y) (std::abs ((x) - (y)) < 1E-6) +#define COMPLEX_EQUAL(x,y) (FP_EQUAL ((x).real (), (y).real ()) \ + && FP_EQUAL ((x).imag (), (y).imag ())) + +#pragma omp declare target +template<typename T> bool test_complex () +{ + std::complex<T> z (-1.334, 5.763); + + if (!FP_EQUAL (z.real (), (T) -1.334)) + return false; + if (!FP_EQUAL (z.imag (), (T) 5.763)) + return false; + if (!FP_EQUAL (std::abs (z), + std::sqrt (z.real () * z.real () + z.imag () * z.imag ()))) + return false; + if (!FP_EQUAL (std::arg (z), std::atan2 (z.imag (), z.real ()))) + return false; + if (!FP_EQUAL (std::norm (z), z.real () * z.real () + z.imag () * z.imag ())) + return false; + + auto conj = std::conj (z); + if (!FP_EQUAL (conj.real (), z.real ()) + || !FP_EQUAL (conj.imag (), -z.imag ())) + return false; + + if (std::proj (z) != z) + return false; + + auto infz1 = std::proj (std::complex<float> (INFINITY, -1)); + if (infz1.real () != INFINITY || infz1.imag () != (T) -0.0) + return false; + auto infz2 = std::proj (std::complex<float> (0, -INFINITY)); + if (infz2.real () != INFINITY || infz2.imag () != (T) -0.0) + return false; + + auto polarz = std::polar ((T) 1.5, std::numbers::pi_v<T> / 4); + if (!FP_EQUAL (polarz.real (), (T) 1.5 * std::cos (std::numbers::pi_v<T> / 4)) + || !FP_EQUAL (polarz.imag (), + (T) 1.5* std::sin (std::numbers::pi_v<T> / 4))) + return false; + + return true; +} + +template<typename T> bool test_complex_exp_log () +{ + std::complex<T> z (-1.724, -3.763); + + // Euler's identity + auto eulerz = std::exp (std::complex<T> (0, std::numbers::pi)); + eulerz += 1.0; + if (!COMPLEX_EQUAL (eulerz, std::complex<T> ())) + return false; + + auto my_exp_z + = std::complex<T> (std::exp (z.real ()) * std::cos (z.imag ()), + std::exp (z.real ()) * std::sin (z.imag ())); + if (!COMPLEX_EQUAL (std::exp (z), my_exp_z)) + return false; + + if (!COMPLEX_EQUAL (std::log10 (z), + std::log (z) / std::log (std::complex<T> (10)))) + return false; + + return true; +} + +template<typename T> bool test_complex_trig () +{ + std::complex<T> z (std::numbers::pi / 8, std::numbers::pi / 10); + const std::complex<T> i (0, 1); + + auto my_sin_z + = std::complex<T> (std::sin (z.real ()) * std::cosh (z.imag ()), + std::cos (z.real ()) * std::sinh (z.imag ())); + if (!COMPLEX_EQUAL (std::sin (z), my_sin_z)) + return false; + + auto my_cos_z + = std::complex<T> (std::cos (z.real ()) * std::cosh (z.imag ()), + -std::sin (z.real ()) * std::sinh (z.imag ())); + if (!COMPLEX_EQUAL (std::cos (z), my_cos_z)) + return false; + + auto my_tan_z + = std::complex<T> (std::sin (2*z.real ()), std::sinh (2*z.imag ())) + / (std::cos (2*z.real ()) + std::cosh (2*z.imag ())); + if (!COMPLEX_EQUAL (std::tan (z), my_tan_z)) + return false; + + auto my_sinh_z + = std::complex<T> (std::sinh (z.real ()) * std::cos (z.imag ()), + std::cosh (z.real ()) * std::sin (z.imag ())); + if (!COMPLEX_EQUAL (std::sinh (z), my_sinh_z)) + return false; + + auto my_cosh_z + = std::complex<T> (std::cosh (z.real ()) * std::cos (z.imag ()), + std::sinh (z.real ()) * std::sin (z.imag ())); + if (!COMPLEX_EQUAL (std::cosh (z), my_cosh_z)) + return false; + + auto my_tanh_z + = std::complex<T> (std::sinh (2*z.real ()), + std::sin (2*z.imag ())) + / (std::cosh (2*z.real ()) + std::cos (2*z.imag ())); + if (!COMPLEX_EQUAL (std::tanh (z), my_tanh_z)) + return false; + + auto my_asin_z = -i * std::log (i * z + std::sqrt ((T) 1.0 - z*z)); + if (!COMPLEX_EQUAL (std::asin (z), my_asin_z)) + return false; + + auto my_acos_z + = std::complex<T> (std::numbers::pi / 2) + + i * std::log (i * z + std::sqrt ((T) 1.0 - z*z)); + if (!COMPLEX_EQUAL (std::acos (z), my_acos_z)) + return false; + + auto my_atan_z = std::complex<T> (0, -0.5) * (std::log ((i - z) / (i + z))); + if (!COMPLEX_EQUAL (std::atan (z), my_atan_z)) + return false; + + auto my_asinh_z = std::log (z + std::sqrt (z*z + (T) 1.0)); + if (!COMPLEX_EQUAL (std::asinh (z), my_asinh_z)) + return false; + + auto my_acosh_z = std::log (z + std::sqrt (z*z - (T) 1.0)); + if (!COMPLEX_EQUAL (std::acosh (z), my_acosh_z)) + return false; + + auto my_atanh_z + = std::complex<T> (0.5) * (std::log ((T) 1.0 + z) - std::log ((T) 1.0 - z)); + if (!COMPLEX_EQUAL (std::atanh (z), my_atanh_z)) + return false; + + return true; +} +#pragma omp end declare target + +#define RUN_TEST(func) \ +{ \ + pass++; \ + bool ok = test_##func<float> (); \ + if (!ok) { result = pass; break; } \ + pass++; \ + ok = test_##func<double> (); \ + if (!ok) { result = pass; break; } \ +} + +int main (void) +{ + int result = 0; + + #pragma omp target map (tofrom: result) + do { + int pass = 0; + + RUN_TEST (complex); + RUN_TEST (complex_exp_log); + RUN_TEST (complex_trig); + } while (false); + + return result; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__deque-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__deque-concurrent-usm.C new file mode 100644 index 0000000..863a1de --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__deque-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__deque-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__deque-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__deque-concurrent.C new file mode 100644 index 0000000..9c2d6fa --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__deque-concurrent.C @@ -0,0 +1,64 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <deque> +#include <algorithm> + +#define N 50000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand (); +} + +#pragma omp declare target +bool validate (const std::deque<int> &_deque, int data[]) +{ + for (int i = 0; i < N; ++i) + if (_deque[i] != data[i] * data[i]) + return false; + return true; +} +#pragma omp end declare target + +int main (void) +{ + int data[N]; + bool ok; + + srand (time (NULL)); + init (data); + +#ifdef MEM_SHARED + std::deque<int> _deque (std::begin (data), std::end (data)); +#else + std::deque<int> _deque; +#endif + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: _deque) +#endif + { +#ifndef MEM_SHARED + #pragma omp target + new (&_deque) std::deque<int> (std::begin (data), std::end (data)); +#endif + + #pragma omp target teams distribute parallel for + for (int i = 0; i < N; ++i) + _deque[i] *= _deque[i]; + + #pragma omp target map (from: ok) + { + ok = validate (_deque, data); +#ifndef MEM_SHARED + _deque.~deque (); +#endif + } + } + + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__flat_map-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__flat_map-concurrent.C new file mode 100644 index 0000000..9e59907 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__flat_map-concurrent.C @@ -0,0 +1,71 @@ +// { dg-do run } +// { dg-additional-options "-std=c++23" } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +/* { dg-ice {TODO PR120450} { offload_target_amdgcn && { ! offload_device_shared_as } } } + { dg-excess-errors {'mkoffload' failure etc.} { xfail { offload_target_amdgcn && { ! offload_device_shared_as } } } } + (For effective-target 'offload_device_shared_as', we've got '-DMEM_SHARED', and therefore don't invoke the constructor with placement new.) */ + +#include <stdlib.h> +#include <time.h> +#include <set> +#include <flat_map> + +#define N 3000 + +void init (int data[], bool unique) +{ + std::set<int> _set; + for (int i = 0; i < N; ++i) + { + // Avoid duplicates in data array if unique is true. + do + data[i] = rand (); + while (unique && _set.count (data[i]) > 0); + _set.insert (data[i]); + } +} + +bool validate (long long sum, int keys[], int data[]) +{ + long long total = 0; + for (int i = 0; i < N; ++i) + total += (long long) keys[i] * data[i]; + return sum == total; +} + +int main (void) +{ + int keys[N], data[N]; + std::flat_map<int,int> _map; + + srand (time (NULL)); + init (keys, true); + init (data, false); + + #pragma omp target enter data map (to: keys[:N], data[:N]) map (alloc: _map) + + #pragma omp target + { +#ifndef MEM_SHARED + new (&_map) std::flat_map<int,int> (); +#endif + for (int i = 0; i < N; ++i) + _map[keys[i]] = data[i]; + } + + long long sum = 0; + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < N; ++i) + sum += (long long) keys[i] * _map[keys[i]]; + +#ifndef MEM_SHARED + #pragma omp target + _map.~flat_map (); +#endif + + #pragma omp target exit data map (release: _map) + + bool ok = validate (sum, keys, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__flat_multimap-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__flat_multimap-concurrent.C new file mode 100644 index 0000000..1dc60c8 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__flat_multimap-concurrent.C @@ -0,0 +1,70 @@ +// { dg-do run } +// { dg-additional-options "-std=c++23" } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +/* { dg-ice {TODO PR120450} { offload_target_amdgcn && { ! offload_device_shared_as } } } + { dg-excess-errors {'mkoffload' failure etc.} { xfail { offload_target_amdgcn && { ! offload_device_shared_as } } } } + (For effective-target 'offload_device_shared_as', we've got '-DMEM_SHARED', and therefore don't invoke the constructor with placement new.) */ + +#include <stdlib.h> +#include <time.h> +#include <flat_map> + +// Make sure that KEY_MAX is less than N to ensure some duplicate keys. +#define N 3000 +#define KEY_MAX 1000 + +void init (int data[], int max) +{ + for (int i = 0; i < N; ++i) + data[i] = i % max; +} + +bool validate (long long sum, int keys[], int data[]) +{ + long long total = 0; + for (int i = 0; i < N; ++i) + total += (long long) keys[i] * data[i]; + return sum == total; +} + +int main (void) +{ + int keys[N], data[N]; + std::flat_multimap<int,int> _map; + + srand (time (NULL)); + init (keys, KEY_MAX); + init (data, RAND_MAX); + + #pragma omp target enter data map (to: keys[:N], data[:N]) map (alloc: _map) + + #pragma omp target + { +#ifndef MEM_SHARED + new (&_map) std::flat_multimap<int,int> (); +#endif + for (int i = 0; i < N; ++i) + _map.insert({keys[i], data[i]}); + } + + long long sum = 0; + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < KEY_MAX; ++i) + { + auto range = _map.equal_range (i); + for (auto it = range.first; it != range.second; ++it) { + sum += (long long) it->first * it->second; + } + } + +#ifndef MEM_SHARED + #pragma omp target + _map.~flat_multimap (); +#endif + + #pragma omp target exit data map (release: _map) + + bool ok = validate (sum, keys, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__flat_multiset-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__flat_multiset-concurrent.C new file mode 100644 index 0000000..59b59bf --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__flat_multiset-concurrent.C @@ -0,0 +1,60 @@ +// { dg-do run } +// { dg-additional-options "-std=c++23" } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <flat_set> +#include <algorithm> + +// MAX should be less than N to ensure that some duplicates occur. +#define N 4000 +#define MAX 1000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand () % MAX; +} + +bool validate (int sum, int data[]) +{ + int total = 0; + for (int i = 0; i < N; ++i) + total += data[i]; + return sum == total; +} + +int main (void) +{ + int data[N]; + std::flat_multiset<int> set; + int sum = 0; + + srand (time (NULL)); + init (data); + + #pragma omp target data map (to: data[:N]) map (alloc: set) + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&set) std::flat_multiset<int> (); +#endif + for (int i = 0; i < N; ++i) + set.insert (data[i]); + } + + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < MAX; ++i) + sum += i * set.count (i); + +#ifndef MEM_SHARED + #pragma omp target + set.~flat_multiset (); +#endif + } + + bool ok = validate (sum, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__flat_set-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__flat_set-concurrent.C new file mode 100644 index 0000000..b255cd5 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__flat_set-concurrent.C @@ -0,0 +1,67 @@ +// { dg-do run } +// { dg-additional-options "-std=c++23" } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <flat_set> +#include <algorithm> + +#define N 4000 +#define MAX 16384 + +void init (int data[]) +{ + std::flat_set<int> _set; + for (int i = 0; i < N; ++i) + { + // Avoid duplicates in data array. + do + data[i] = rand () % MAX; + while (_set.count (data[i]) != 0); + _set.insert (data[i]); + } +} + +bool validate (int sum, int data[]) +{ + int total = 0; + for (int i = 0; i < N; ++i) + total += data[i]; + return sum == total; +} + +int main (void) +{ + int data[N]; + std::flat_set<int> _set; + int sum = 0; + + srand (time (NULL)); + init (data); + + #pragma omp target data map (to: data[:N]) map (alloc: _set) + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&_set) std::flat_set<int> (); +#endif + for (int i = 0; i < N; ++i) + _set.insert (data[i]); + } + + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < MAX; ++i) + if (_set.count (i) > 0) + sum += i; + +#ifndef MEM_SHARED + #pragma omp target + _set.~flat_set (); +#endif + } + + bool ok = validate (sum, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__forward_list-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__forward_list-concurrent-usm.C new file mode 100644 index 0000000..60d5cee --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__forward_list-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__forward_list-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__forward_list-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__forward_list-concurrent.C new file mode 100644 index 0000000..6b0ee65 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__forward_list-concurrent.C @@ -0,0 +1,83 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <omp.h> +#include <forward_list> +#include <algorithm> + +#define N 3000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand (); +} + +#pragma omp declare target +bool validate (const std::forward_list<int> &list, int data[]) +{ + int i = 0; + for (auto &v : list) + { + if (v != data[i] * data[i]) + return false; + ++i; + } + return true; +} +#pragma omp end declare target + +int main (void) +{ + int data[N]; + bool ok; + + srand (time (NULL)); + init (data); + +#ifdef MEM_SHARED + std::forward_list<int> list (std::begin (data), std::end (data)); +#else + std::forward_list<int> list; +#endif + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: list) +#endif + { +#ifndef MEM_SHARED + #pragma omp target + new (&list) std::forward_list<int> (std::begin (data), std::end (data)); +#endif + + #pragma omp target teams + do + { + int len = N / omp_get_num_teams () + (N % omp_get_num_teams () > 0); + int start = len * omp_get_team_num (); + if (start >= N) + break; + if (start + len >= N) + len = N - start; + auto it = list.begin (); + std::advance (it, start); + for (int i = 0; i < len; ++i) + { + *it *= *it; + ++it; + } + } while (false); + + #pragma omp target map (from: ok) + { + ok = validate (list, data); +#ifndef MEM_SHARED + list.~forward_list (); +#endif + } + } + + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__list-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__list-concurrent-usm.C new file mode 100644 index 0000000..5057bf9 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__list-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__list-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__list-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__list-concurrent.C new file mode 100644 index 0000000..1f44a17 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__list-concurrent.C @@ -0,0 +1,83 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <omp.h> +#include <list> +#include <algorithm> + +#define N 3000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand (); +} + +#pragma omp declare target +bool validate (const std::list<int> &_list, int data[]) +{ + int i = 0; + for (auto &v : _list) + { + if (v != data[i] * data[i]) + return false; + ++i; + } + return true; +} +#pragma omp end declare target + +int main (void) +{ + int data[N]; + bool ok; + + srand (time (NULL)); + init (data); + +#ifdef MEM_SHARED + std::list<int> _list (std::begin (data), std::end (data)); +#else + std::list<int> _list; +#endif + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: _list) +#endif + { +#ifndef MEM_SHARED + #pragma omp target + new (&_list) std::list<int> (std::begin (data), std::end (data)); +#endif + + #pragma omp target teams + do + { + int len = N / omp_get_num_teams () + (N % omp_get_num_teams () > 0); + int start = len * omp_get_team_num (); + if (start >= N) + break; + if (start + len >= N) + len = N - start; + auto it = _list.begin (); + std::advance (it, start); + for (int i = 0; i < len; ++i) + { + *it *= *it; + ++it; + } + } while (false); + + #pragma omp target map (from: ok) + { + ok = validate (_list, data); +#ifndef MEM_SHARED + _list.~list (); +#endif + } + } + + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__map-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__map-concurrent-usm.C new file mode 100644 index 0000000..fe37426 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__map-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__map-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__map-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__map-concurrent.C new file mode 100644 index 0000000..36556ef --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__map-concurrent.C @@ -0,0 +1,70 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <set> +#include <map> + +#define N 3000 + +void init (int data[], bool unique) +{ + std::set<int> _set; + for (int i = 0; i < N; ++i) + { + // Avoid duplicates in data array if unique is true. + do + data[i] = rand (); + while (unique && _set.find (data[i]) != _set.end ()); + _set.insert (data[i]); + } +} + +bool validate (long long sum, int keys[], int data[]) +{ + long long total = 0; + for (int i = 0; i < N; ++i) + total += (long long) keys[i] * data[i]; + return sum == total; +} + +int main (void) +{ + int keys[N], data[N]; + std::map<int,int> _map; + + srand (time (NULL)); + init (keys, true); + init (data, false); + +#ifndef MEM_SHARED + #pragma omp target enter data map (to: keys[:N], data[:N]) map (alloc: _map) +#endif + + #pragma omp target + { +#ifndef MEM_SHARED + new (&_map) std::map<int,int> (); +#endif + for (int i = 0; i < N; ++i) + _map[keys[i]] = data[i]; + } + + long long sum = 0; + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < N; ++i) + sum += (long long) keys[i] * _map[keys[i]]; + +#ifndef MEM_SHARED + #pragma omp target + _map.~map (); +#endif + +#ifndef MEM_SHARED + #pragma omp target exit data map (release: _map) +#endif + + bool ok = validate (sum, keys, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__multimap-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__multimap-concurrent-usm.C new file mode 100644 index 0000000..79f9245 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__multimap-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__multimap-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__multimap-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__multimap-concurrent.C new file mode 100644 index 0000000..6a4a4e8 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__multimap-concurrent.C @@ -0,0 +1,68 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <map> + +// Make sure that KEY_MAX is less than N to ensure some duplicate keys. +#define N 3000 +#define KEY_MAX 1000 + +void init (int data[], int max) +{ + for (int i = 0; i < N; ++i) + data[i] = rand () % max; +} + +bool validate (long long sum, int keys[], int data[]) +{ + long long total = 0; + for (int i = 0; i < N; ++i) + total += (long long) keys[i] * data[i]; + return sum == total; +} + +int main (void) +{ + int keys[N], data[N]; + std::multimap<int,int> _map; + + srand (time (NULL)); + init (keys, KEY_MAX); + init (data, RAND_MAX); + +#ifndef MEM_SHARED + #pragma omp target enter data map (to: keys[:N], data[:N]) map (alloc: _map) +#endif + + #pragma omp target + { +#ifndef MEM_SHARED + new (&_map) std::multimap<int,int> (); +#endif + for (int i = 0; i < N; ++i) + _map.insert({keys[i], data[i]}); + } + + long long sum = 0; + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < KEY_MAX; ++i) + { + auto range = _map.equal_range (i); + for (auto it = range.first; it != range.second; ++it) + sum += (long long) it->first * it->second; + } + +#ifndef MEM_SHARED + #pragma omp target + _map.~multimap (); +#endif + +#ifndef MEM_SHARED + #pragma omp target exit data map (release: _map) +#endif + + bool ok = validate (sum, keys, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__multiset-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__multiset-concurrent-usm.C new file mode 100644 index 0000000..2d80756 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__multiset-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__multiset-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__multiset-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__multiset-concurrent.C new file mode 100644 index 0000000..b12402e --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__multiset-concurrent.C @@ -0,0 +1,62 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <stdio.h> +#include <time.h> +#include <set> +#include <algorithm> + +// MAX should be less than N to ensure that some duplicates occur. +#define N 4000 +#define MAX 1000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand () % MAX; +} + +bool validate (int sum, int data[]) +{ + int total = 0; + for (int i = 0; i < N; ++i) + total += data[i]; + return sum == total; +} + +int main (void) +{ + int data[N]; + std::multiset<int> set; + int sum = 0; + + srand (time (NULL)); + init (data); + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: set) +#endif + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&set) std::multiset<int> (); +#endif + for (int i = 0; i < N; ++i) + set.insert (data[i]); + } + + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < MAX; ++i) + sum += i * set.count (i); + +#ifndef MEM_SHARED + #pragma omp target + set.~multiset (); +#endif + } + + bool ok = validate (sum, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__numbers.C b/libgomp/testsuite/libgomp.c++/target-std__numbers.C new file mode 100644 index 0000000..a6b3665 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__numbers.C @@ -0,0 +1,93 @@ +// { dg-do run } +// { dg-additional-options "-std=c++20" } + +#include <cmath> +#include <numbers> + +#define FP_EQUAL(x,y) (std::abs ((x) - (y)) < 1E-6) + +#pragma omp declare target +template<typename T> bool test_pi () +{ + if (!FP_EQUAL (std::sin (std::numbers::pi_v<T>), (T) 0.0)) + return false; + if (!FP_EQUAL (std::cos (std::numbers::pi_v<T>), (T) -1.0)) + return false; + if (!FP_EQUAL (std::numbers::pi_v<T> * std::numbers::inv_pi_v<T>, (T) 1.0)) + return false; + if (!FP_EQUAL (std::numbers::pi_v<T> * std::numbers::inv_sqrtpi_v<T> + * std::numbers::inv_sqrtpi_v<T>, (T) 1.0)) + return false; + return true; +} + +template<typename T> bool test_sqrt () +{ + if (!FP_EQUAL (std::numbers::sqrt2_v<T> * std::numbers::sqrt2_v<T>, (T) 2.0)) + return false; + if (!FP_EQUAL (std::numbers::sqrt3_v<T> * std::numbers::sqrt3_v<T>, (T) 3.0)) + return false; + return true; +} + +template<typename T> bool test_phi () +{ + T myphi = ((T) 1.0 + std::sqrt ((T) 5.0)) / (T) 2.0; + if (!FP_EQUAL (myphi, std::numbers::phi_v<T>)) + return false; + return true; +} + +template<typename T> bool test_log () +{ + if (!FP_EQUAL (std::log ((T) 2.0), std::numbers::ln2_v<T>)) + return false; + if (!FP_EQUAL (std::log ((T) 10.0), std::numbers::ln10_v<T>)) + return false; + if (!FP_EQUAL (std::log2 ((T) std::numbers::e), std::numbers::log2e_v<T>)) + return false; + if (!FP_EQUAL (std::log10 ((T) std::numbers::e), std::numbers::log10e_v<T>)) + return false; + return true; +} + +template<typename T> bool test_egamma () +{ + T myegamma = 0.0; + #pragma omp parallel for reduction(+:myegamma) + for (int k = 2; k < 100000; ++k) + myegamma += (std::riemann_zeta (k) - 1) / k; + myegamma = (T) 1 - myegamma; + if (!FP_EQUAL (myegamma, std::numbers::egamma_v<T>)) + return false; + return true; +} +#pragma omp end declare target + +#define RUN_TEST(func) \ +{ \ + pass++; \ + bool ok = test_##func<float> (); \ + if (!ok) { result = pass; break; } \ + pass++; \ + ok = test_##func<double> (); \ + if (!ok) { result = pass; break; } \ +} + +int main (void) +{ + int result = 0; + + #pragma omp target map (tofrom: result) + do { + int pass = 0; + + RUN_TEST (pi); + RUN_TEST (sqrt); + RUN_TEST (phi); + RUN_TEST (log); + RUN_TEST (egamma); + } while (false); + + return result; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__set-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__set-concurrent-usm.C new file mode 100644 index 0000000..54f62e3 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__set-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__set-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__set-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__set-concurrent.C new file mode 100644 index 0000000..cd23128 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__set-concurrent.C @@ -0,0 +1,68 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <set> +#include <algorithm> + +#define N 4000 +#define MAX 16384 + +void init (int data[]) +{ + std::set<int> _set; + for (int i = 0; i < N; ++i) + { + // Avoid duplicates in data array. + do + data[i] = rand () % MAX; + while (_set.find (data[i]) != _set.end ()); + _set.insert (data[i]); + } +} + +bool validate (int sum, int data[]) +{ + int total = 0; + for (int i = 0; i < N; ++i) + total += data[i]; + return sum == total; +} + +int main (void) +{ + int data[N]; + std::set<int> _set; + int sum = 0; + + srand (time (NULL)); + init (data); + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: _set) +#endif + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&_set) std::set<int> (); +#endif + for (int i = 0; i < N; ++i) + _set.insert (data[i]); + } + + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < MAX; ++i) + if (_set.find (i) != _set.end ()) + sum += i; + +#ifndef MEM_SHARED + #pragma omp target + _set.~set (); +#endif + } + + bool ok = validate (sum, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__span-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__span-concurrent-usm.C new file mode 100644 index 0000000..7ef16bf --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__span-concurrent-usm.C @@ -0,0 +1,7 @@ +// { dg-additional-options "-std=c++20" } + +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__span-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__span-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__span-concurrent.C new file mode 100644 index 0000000..046b3c1 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__span-concurrent.C @@ -0,0 +1,66 @@ +// { dg-do run } +// { dg-additional-options "-std=c++20" } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <span> + +#define N 64 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand (); +} + +#pragma omp declare target +bool validate (const std::span<int, N> &span, int data[]) +{ + for (int i = 0; i < N; ++i) + if (span[i] != data[i] * data[i]) + return false; + return true; +} +#pragma omp end declare target + +int main (void) +{ + int data[N]; + bool ok; + int elements[N]; + std::span<int, N> span(elements); + + srand (time (NULL)); + init (data); + +#ifndef MEM_SHARED + #pragma omp target enter data map (to: data[:N]) map (alloc: elements, span) +#endif + + #pragma omp target + { +#ifndef MEM_SHARED + new (&span) std::span<int, N> (elements); +#endif + std::copy (data, data + N, span.begin ()); + } + + #pragma omp target teams distribute parallel for + for (int i = 0; i < N; ++i) + span[i] *= span[i]; + + #pragma omp target map (from: ok) + { + ok = validate (span, data); +#ifndef MEM_SHARED + span.~span (); +#endif + } + +#ifndef MEM_SHARED + #pragma omp target exit data map (release: elements, span) +#endif + + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__unordered_map-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__unordered_map-concurrent.C new file mode 100644 index 0000000..00d7943 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__unordered_map-concurrent.C @@ -0,0 +1,66 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <set> +#include <unordered_map> + +#define N 3000 + +void init (int data[], bool unique) +{ + std::set<int> _set; + for (int i = 0; i < N; ++i) + { + // Avoid duplicates in data array if unique is true. + do + data[i] = rand (); + while (unique && _set.count (data[i]) > 0); + _set.insert (data[i]); + } +} + +bool validate (long long sum, int keys[], int data[]) +{ + long long total = 0; + for (int i = 0; i < N; ++i) + total += (long long) keys[i] * data[i]; + return sum == total; +} + +int main (void) +{ + int keys[N], data[N]; + std::unordered_map<int,int> _map; + + srand (time (NULL)); + init (keys, true); + init (data, false); + + #pragma omp target enter data map (to: keys[:N], data[:N]) map (alloc: _map) + + #pragma omp target + { +#ifndef MEM_SHARED + new (&_map) std::unordered_map<int,int> (); +#endif + for (int i = 0; i < N; ++i) + _map[keys[i]] = data[i]; + } + + long long sum = 0; + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < N; ++i) + sum += (long long) keys[i] * _map[keys[i]]; + +#ifndef MEM_SHARED + #pragma omp target + _map.~unordered_map (); +#endif + + #pragma omp target exit data map (release: _map) + + bool ok = validate (sum, keys, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__unordered_multimap-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__unordered_multimap-concurrent.C new file mode 100644 index 0000000..2567634 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__unordered_multimap-concurrent.C @@ -0,0 +1,65 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <unordered_map> + +// Make sure that KEY_MAX is less than N to ensure some duplicate keys. +#define N 3000 +#define KEY_MAX 1000 + +void init (int data[], int max) +{ + for (int i = 0; i < N; ++i) + data[i] = i % max; +} + +bool validate (long long sum, int keys[], int data[]) +{ + long long total = 0; + for (int i = 0; i < N; ++i) + total += (long long) keys[i] * data[i]; + return sum == total; +} + +int main (void) +{ + int keys[N], data[N]; + std::unordered_multimap<int,int> _map; + + srand (time (NULL)); + init (keys, KEY_MAX); + init (data, RAND_MAX); + + #pragma omp target enter data map (to: keys[:N], data[:N]) map (alloc: _map) + + #pragma omp target + { +#ifndef MEM_SHARED + new (&_map) std::unordered_multimap<int,int> (); +#endif + for (int i = 0; i < N; ++i) + _map.insert({keys[i], data[i]}); + } + + long long sum = 0; + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < KEY_MAX; ++i) + { + auto range = _map.equal_range (i); + for (auto it = range.first; it != range.second; ++it) { + sum += (long long) it->first * it->second; + } + } + +#ifndef MEM_SHARED + #pragma omp target + _map.~unordered_multimap (); +#endif + + #pragma omp target exit data map (release: _map) + + bool ok = validate (sum, keys, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__unordered_multiset-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__unordered_multiset-concurrent.C new file mode 100644 index 0000000..da6c875 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__unordered_multiset-concurrent.C @@ -0,0 +1,59 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <unordered_set> +#include <algorithm> + +// MAX should be less than N to ensure that some duplicates occur. +#define N 4000 +#define MAX 1000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand () % MAX; +} + +bool validate (int sum, int data[]) +{ + int total = 0; + for (int i = 0; i < N; ++i) + total += data[i]; + return sum == total; +} + +int main (void) +{ + int data[N]; + std::unordered_multiset<int> set; + int sum = 0; + + srand (time (NULL)); + init (data); + + #pragma omp target data map (to: data[:N]) map (alloc: set) + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&set) std::unordered_multiset<int> (); +#endif + for (int i = 0; i < N; ++i) + set.insert (data[i]); + } + + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < MAX; ++i) + sum += i * set.count (i); + +#ifndef MEM_SHARED + #pragma omp target + set.~unordered_multiset (); +#endif + } + + bool ok = validate (sum, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__unordered_set-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__unordered_set-concurrent.C new file mode 100644 index 0000000..b7bd935 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__unordered_set-concurrent.C @@ -0,0 +1,66 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <unordered_set> +#include <algorithm> + +#define N 4000 +#define MAX 16384 + +void init (int data[]) +{ + std::unordered_set<int> _set; + for (int i = 0; i < N; ++i) + { + // Avoid duplicates in data array. + do + data[i] = rand () % MAX; + while (_set.count (data[i]) != 0); + _set.insert (data[i]); + } +} + +bool validate (int sum, int data[]) +{ + int total = 0; + for (int i = 0; i < N; ++i) + total += data[i]; + return sum == total; +} + +int main (void) +{ + int data[N]; + std::unordered_set<int> _set; + int sum = 0; + + srand (time (NULL)); + init (data); + + #pragma omp target data map (to: data[:N]) map (alloc: _set) + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&_set) std::unordered_set<int> (); +#endif + for (int i = 0; i < N; ++i) + _set.insert (data[i]); + } + + #pragma omp target teams distribute parallel for reduction (+:sum) + for (int i = 0; i < MAX; ++i) + if (_set.count (i) > 0) + sum += i; + +#ifndef MEM_SHARED + #pragma omp target + _set.~unordered_set (); +#endif + } + + bool ok = validate (sum, data); + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__valarray-1.C b/libgomp/testsuite/libgomp.c++/target-std__valarray-1.C new file mode 100644 index 0000000..865cde2 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__valarray-1.C @@ -0,0 +1,179 @@ +// { dg-additional-options -std=c++20 } +// { dg-output-file target-std__valarray-1.output } + +#include <valarray> +#include <ostream> +#include <sstream> + + +/*TODO Work around PR118484 "ICE during IPA pass: cp, segfault in determine_versionability ipa-cp.cc:467". + +We can't: + + #pragma omp declare target(std::basic_streambuf<char, std::char_traits<char>>::basic_streambuf) + +... because: + + error: overloaded function name ‘std::basic_streambuf<char>::__ct ’ in clause ‘enter’ + +Therefore, use dummy classes in '#pragma omp declare target': +*/ + +#pragma omp declare target + +// For 'std::basic_streambuf<char, std::char_traits<char> >::basic_streambuf': + +class dummy_basic_streambuf__char + : public std::basic_streambuf<char> +{ +public: + dummy_basic_streambuf__char() {} +}; + +// For 'std::basic_ios<char, std::char_traits<char> >::basic_ios()': + +class dummy_basic_ios__char + : public std::basic_ios<char> +{ +public: + dummy_basic_ios__char() {} +}; + +#pragma omp end declare target + + +int main() +{ + // Due to PR120021 "Offloading vs. C++ 'std::initializer_list'", we can't construct these on the device. + std::initializer_list<int> v1_i = {10, 20, 30, 40, 50}; + const int *v1_i_data = std::data(v1_i); + size_t v1_i_size = v1_i.size(); + std::initializer_list<int> v2_i = {5, 4, 3, 2, 1}; + const int *v2_i_data = std::data(v2_i); + size_t v2_i_size = v2_i.size(); + std::initializer_list<int> shiftData_i = {1, 2, 3, 4, 5}; + const int *shiftData_i_data = std::data(shiftData_i); + size_t shiftData_i_size = shiftData_i.size(); +#pragma omp target \ + defaultmap(none) \ + map(to: v1_i_data[:v1_i_size], v1_i_size, \ + v2_i_data[:v2_i_size], v2_i_size, \ + shiftData_i_data[:shiftData_i_size], shiftData_i_size) + { + /* Manually set up a buffer we can stream into, similar to 'cout << [...]', and print it at the end of region. */ + std::stringbuf out_b; + std::ostream out(&out_b); + + std::valarray<int> v1(v1_i_data, v1_i_size); + out << "\nv1:"; + for (auto val : v1) + out << " " << val; + + std::valarray<int> v2(v2_i_data, v2_i_size); + out << "\nv2:"; + for (auto val : v2) + out << " " << val; + + std::valarray<int> sum = v1 + v2; + out << "\nv1 + v2:"; + for (auto val : sum) + out << " " << val; + + std::valarray<int> diff = v1 - v2; + out << "\nv1 - v2:"; + for (auto val : diff) + out << " " << val; + + std::valarray<int> product = v1 * v2; + out << "\nv1 * v2:"; + for (auto val : product) + out << " " << val; + + std::valarray<int> quotient = v1 / v2; + out << "\nv1 / v2:"; + for (auto val : quotient) + out << " " << val; + + std::valarray<int> squares = pow(v1, 2); + out << "\npow(v1, 2):"; + for (auto val : squares) + out << " " << val; + + std::valarray<int> sinhs = sinh(v2); + out << "\nsinh(v2):"; + for (auto val : sinhs) + out << " " << val; + + std::valarray<int> logs = log(v1 * v2); + out << "\nlog(v1 * v2):"; + for (auto val : logs) + out << " " << val; + + std::valarray<int> data(12); + for (size_t i = 0; i < data.size(); ++i) + data[i] = i; + out << "\nOriginal array:"; + for (auto val : data) + out << " " << val; + + std::slice slice1(2, 5, 1); + std::valarray<int> sliced1 = data[slice1]; + out << "\nSlice(2, 5, 1):"; + for (auto val : sliced1) + out << " " << val; + + std::slice slice2(1, 4, 3); + std::valarray<int> sliced2 = data[slice2]; + out << "\nSlice(1, 4, 3):"; + for (auto val : sliced2) + out << " " << val; + + data[slice1] = 99; + out << "\nArray after slice modification:"; + for (auto val : data) + out << " " << val; + + std::valarray<bool> mask = (v1 > 20); + out << "\nElements of v1 > 20:"; + for (size_t i = 0; i < v1.size(); ++i) + { + if (mask[i]) + out << " " << v1[i]; + } + + std::valarray<int> masked = v1[mask]; + out << "\nMasked array:"; + for (auto val : masked) + out << " " << val; + + std::valarray<int> shiftData(shiftData_i_data, shiftData_i_size); + out << "\nOriginal shiftData:"; + for (auto val : shiftData) + out << " " << val; + + std::valarray<int> shifted = shiftData.shift(2); + out << "\nshift(2):"; + for (auto val : shifted) + out << " " << val; + + std::valarray<int> cshifted = shiftData.cshift(-1); + out << "\ncshift(-1):"; + for (auto val : cshifted) + out << " " << val; + + out << "\nSum(v1): " << v1.sum(); + out << "\nMin(v1): " << v1.min(); + out << "\nMax(v1): " << v1.max(); + + out << "\n"; + + /* Terminate with a NUL. Otherwise, we'd have to use: + __builtin_printf("%.*s", (int) out_b_sv.size(), out_b_sv.data()); + ... which nvptx 'printf', as implemented via PTX 'vprintf', doesn't support (TODO). */ + out << '\0'; + std::string_view out_b_sv = out_b.view(); + __builtin_printf("%s", out_b_sv.data()); + } + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__valarray-1.output b/libgomp/testsuite/libgomp.c++/target-std__valarray-1.output new file mode 100644 index 0000000..c441e06 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__valarray-1.output @@ -0,0 +1,22 @@ + +v1: 10 20 30 40 50 +v2: 5 4 3 2 1 +v1 + v2: 15 24 33 42 51 +v1 - v2: 5 16 27 38 49 +v1 * v2: 50 80 90 80 50 +v1 / v2: 2 5 10 20 50 +pow(v1, 2): 100 400 900 1600 2500 +sinh(v2): 74 27 10 3 1 +log(v1 * v2): 3 4 4 4 3 +Original array: 0 1 2 3 4 5 6 7 8 9 10 11 +Slice(2, 5, 1): 2 3 4 5 6 +Slice(1, 4, 3): 1 4 7 10 +Array after slice modification: 0 1 99 99 99 99 99 7 8 9 10 11 +Elements of v1 > 20: 30 40 50 +Masked array: 30 40 50 +Original shiftData: 1 2 3 4 5 +shift(2): 3 4 5 0 0 +cshift(-1): 5 1 2 3 4 +Sum(v1): 150 +Min(v1): 10 +Max(v1): 50 diff --git a/libgomp/testsuite/libgomp.c++/target-std__valarray-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__valarray-concurrent-usm.C new file mode 100644 index 0000000..41ec80e --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__valarray-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__valarray-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__valarray-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__valarray-concurrent.C new file mode 100644 index 0000000..8933072b --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__valarray-concurrent.C @@ -0,0 +1,66 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <valarray> + +#define N 50000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand (); +} + +#pragma omp declare target +bool validate (const std::valarray<int> &arr, int data[]) +{ + for (int i = 0; i < N; ++i) + if (arr[i] != data[i] * data[i] + i) + return false; + return true; +} +#pragma omp end declare target + +int main (void) +{ + int data[N]; + bool ok; + + srand (time (NULL)); + init (data); + +#ifdef MEM_SHARED + std::valarray<int> arr (data, N); +#else + std::valarray<int> arr; +#endif + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: arr) +#endif + { + #pragma omp target + { +#ifndef MEM_SHARED + new (&arr) std::valarray<int> (data, N); +#endif + arr *= arr; + } + + #pragma omp target teams distribute parallel for + for (int i = 0; i < N; ++i) + arr[i] += i; + + #pragma omp target map (from: ok) + { + ok = validate (arr, data); +#ifndef MEM_SHARED + arr.~valarray (); +#endif + } + } + + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c++/target-std__vector-concurrent-usm.C b/libgomp/testsuite/libgomp.c++/target-std__vector-concurrent-usm.C new file mode 100644 index 0000000..967bff3 --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__vector-concurrent-usm.C @@ -0,0 +1,5 @@ +#pragma omp requires unified_shared_memory self_maps + +#define MEM_SHARED + +#include "target-std__vector-concurrent.C" diff --git a/libgomp/testsuite/libgomp.c++/target-std__vector-concurrent.C b/libgomp/testsuite/libgomp.c++/target-std__vector-concurrent.C new file mode 100644 index 0000000..a94b4cf --- /dev/null +++ b/libgomp/testsuite/libgomp.c++/target-std__vector-concurrent.C @@ -0,0 +1,63 @@ +// { dg-do run } +// { dg-additional-options -DMEM_SHARED { target offload_device_shared_as } } + +#include <stdlib.h> +#include <time.h> +#include <vector> + +#define N 50000 + +void init (int data[]) +{ + for (int i = 0; i < N; ++i) + data[i] = rand (); +} + +#pragma omp declare target +bool validate (const std::vector<int> &vec, int data[]) +{ + for (int i = 0; i < N; ++i) + if (vec[i] != data[i] * data[i]) + return false; + return true; +} +#pragma omp end declare target + +int main (void) +{ + int data[N]; + bool ok; + + srand (time (NULL)); + init (data); + +#ifdef MEM_SHARED + std::vector<int> vec (data, data + N); +#else + std::vector<int> vec; +#endif + +#ifndef MEM_SHARED + #pragma omp target data map (to: data[:N]) map (alloc: vec) +#endif + { +#ifndef MEM_SHARED + #pragma omp target + new (&vec) std::vector<int> (data, data + N); +#endif + + #pragma omp target teams distribute parallel for + for (int i = 0; i < N; ++i) + vec[i] *= vec[i]; + + #pragma omp target map (from: ok) + { + ok = validate (vec, data); +#ifndef MEM_SHARED + vec.~vector (); +#endif + } + } + + return ok ? 0 : 1; +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-10.c b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-10.c new file mode 100644 index 0000000..00eb48b --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-10.c @@ -0,0 +1,64 @@ +/* { dg-do run } */ + +#include <string.h> +#include <stdlib.h> +#include <assert.h> + +#define N 64 + +typedef struct { + int *arr; + int size; +} B; + +#pragma omp declare mapper (mapB : B myb) map(to: myb.size, myb.arr) \ + map(tofrom: myb.arr[0:myb.size]) +// While GCC handles more, only default is ... +#pragma omp declare mapper (default : B myb) map(to: myb.size, myb.arr) \ + map(tofrom: myb.arr[0:myb.size]) + +struct A { + int *arr1; + B *arr2; + int arr3[N]; +}; + +int +main (int argc, char *argv[]) +{ + struct A var; + + memset (&var, 0, sizeof var); + var.arr1 = (int *) calloc (N, sizeof (int)); + var.arr2 = (B *) malloc (sizeof (B)); + var.arr2->arr = (int *) calloc (N, sizeof (float)); + var.arr2->size = N; + + { + // ... permitted here: + #pragma omp declare mapper (struct A x) map(to: x.arr1, x.arr2) \ + map(tofrom: x.arr1[0:N]) \ + map(mapper(default), tofrom: x.arr2[0:1]) + #pragma omp target + { + for (int i = 0; i < N; i++) + { + var.arr1[i]++; + var.arr2->arr[i]++; + } + } + } + + for (int i = 0; i < N; i++) + { + assert (var.arr1[i] == 1); + assert (var.arr2->arr[i] == 1); + assert (var.arr3[i] == 0); + } + + free (var.arr1); + free (var.arr2->arr); + free (var.arr2); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-11.c b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-11.c new file mode 100644 index 0000000..942d6a5 --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-11.c @@ -0,0 +1,59 @@ +/* { dg-do run } */ + +#include <string.h> +#include <stdlib.h> +#include <assert.h> + +#define N 64 + +typedef struct B_tag { + int *arr; + int size; +} B; + +#pragma omp declare mapper (B myb) map(to: myb.size, myb.arr) \ + map(tofrom: myb.arr[0:myb.size]) + +struct A { + int *arr1; + B *arr2; + int arr3[N]; +}; + +int +main (int argc, char *argv[]) +{ + struct A var; + + memset (&var, 0, sizeof var); + var.arr1 = (int *) calloc (N, sizeof (int)); + var.arr2 = (B *) malloc (sizeof (B)); + var.arr2->arr = (int *) calloc (N, sizeof (int)); + var.arr2->size = N; + + { + #pragma omp declare mapper (struct A x) map(to: x.arr1, x.arr2) \ + map(tofrom: x.arr1[0:N]) map(tofrom: x.arr2[0:1]) + #pragma omp target + { + for (int i = 0; i < N; i++) + { + var.arr1[i]++; + var.arr2->arr[i]++; + } + } + } + + for (int i = 0; i < N; i++) + { + assert (var.arr1[i] == 1); + assert (var.arr2->arr[i] == 1); + assert (var.arr3[i] == 0); + } + + free (var.arr1); + free (var.arr2->arr); + free (var.arr2); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-12.c b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-12.c new file mode 100644 index 0000000..cfc6a91 --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-12.c @@ -0,0 +1,94 @@ +/* { dg-do run } */ + +#include <string.h> +#include <stdlib.h> +#include <assert.h> + +#define N 64 + +typedef struct { + int *arr; + int size; +} B; + +#pragma omp declare mapper (samename : B myb) map(to: myb.size, myb.arr) \ + map(tofrom: myb.arr[0:myb.size]) +// While GCC handles more, only default is ... +#pragma omp declare mapper (default : B myb) map(to: myb.size, myb.arr) \ + map(tofrom: myb.arr[0:myb.size]) +typedef struct { + int *arr; + int size; +} C; + + +struct A { + int *arr1; + B *arr2; + C *arr3; +}; + +int +main (int argc, char *argv[]) +{ + struct A var; + + memset (&var, 0, sizeof var); + var.arr1 = (int *) calloc (N, sizeof (int)); + var.arr2 = (B *) malloc (sizeof (B)); + var.arr2->arr = (int *) calloc (N, sizeof (int)); + var.arr2->size = N; + var.arr3 = (C *) malloc (sizeof (C)); + var.arr3->arr = (int *) calloc (N, sizeof (int)); + var.arr3->size = N; + + { + // ... permitted here. + #pragma omp declare mapper (struct A x) map(to: x.arr1, x.arr2) \ + map(tofrom: x.arr1[0:N]) \ + map(mapper(default), tofrom: x.arr2[0:1]) + #pragma omp target + { + for (int i = 0; i < N; i++) + { + var.arr1[i]++; + var.arr2->arr[i]++; + } + } + } + + { + #pragma omp declare mapper (samename : C myc) map(to: myc.size, myc.arr) \ + map(tofrom: myc.arr[0:myc.size]) + // While GCC handles more, only default is ... + #pragma omp declare mapper (default : C myc) map(to: myc.size, myc.arr) \ + map(tofrom: myc.arr[0:myc.size]) + // ... permitted here. + #pragma omp declare mapper (struct A x) map(to: x.arr1, x.arr3) \ + map(tofrom: x.arr1[0:N]) \ + map(mapper( default ) , tofrom: *x.arr3) + #pragma omp target + { + for (int i = 0; i < N; i++) + { + var.arr1[i]++; + var.arr3->arr[i]++; + } + } + } + + for (int i = 0; i < N; i++) + { + assert (var.arr1[i] == 2); + assert (var.arr2->arr[i] == 1); + assert (var.arr3->arr[i] == 1); + } + + free (var.arr1); + free (var.arr2->arr); + free (var.arr2); + free (var.arr3->arr); + free (var.arr3); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-13.c b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-13.c new file mode 100644 index 0000000..c4784eb --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-13.c @@ -0,0 +1,55 @@ +/* { dg-do run } */ + +#include <assert.h> + +struct T { + int a; + int b; + int c; +}; + +void foo (void) +{ + struct T x; + x.a = x.b = x.c = 0; + +#pragma omp target + { + x.a++; + x.c++; + } + + assert (x.a == 1); + assert (x.b == 0); + assert (x.c == 1); +} + +// An identity mapper. This should do the same thing as the default! +#pragma omp declare mapper (struct T v) map(v) + +void bar (void) +{ + struct T x; + x.a = x.b = x.c = 0; + +#pragma omp target + { + x.b++; + } + +#pragma omp target map(x) + { + x.a++; + } + + assert (x.a == 1); + assert (x.b == 1); + assert (x.c == 0); +} + +int main (int argc, char *argv[]) +{ + foo (); + bar (); + return 0; +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-14.c b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-14.c new file mode 100644 index 0000000..3e6027e --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-14.c @@ -0,0 +1,57 @@ +/* { dg-do run } */ + +#include <stdlib.h> +#include <assert.h> + +struct Z { + int *arr; +}; + +void baz (struct Z *zarr, int len) +{ +#pragma omp declare mapper (struct Z myvar) map(to: myvar.arr) \ + map(tofrom: myvar.arr[0:len]) + zarr[0].arr = (int *) calloc (len, sizeof (int)); + zarr[5].arr = (int *) calloc (len, sizeof (int)); + +#pragma omp target map(zarr, *zarr) + { + for (int i = 0; i < len; i++) + zarr[0].arr[i]++; + } + +#pragma omp target map(zarr, zarr[5]) + { + for (int i = 0; i < len; i++) + zarr[5].arr[i]++; + } + +#pragma omp target map(zarr[5]) + { + for (int i = 0; i < len; i++) + zarr[5].arr[i]++; + } + +#pragma omp target map(zarr, zarr[5:1]) + { + for (int i = 0; i < len; i++) + zarr[5].arr[i]++; + } + + for (int i = 0; i < len; i++) + assert (zarr[0].arr[i] == 1); + + for (int i = 0; i < len; i++) + assert (zarr[5].arr[i] == 3); + + free (zarr[5].arr); + free (zarr[0].arr); +} + +int +main (int argc, char *argv[]) +{ + struct Z myzarr[10]; + baz (myzarr, 256); + return 0; +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-9.c b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-9.c new file mode 100644 index 0000000..324d535 --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/declare-mapper-9.c @@ -0,0 +1,62 @@ +/* { dg-do run } */ + +#include <string.h> +#include <stdlib.h> +#include <assert.h> + +#define N 64 + +struct A { + int *arr1; + float *arr2; + int arr3[N]; +}; + +int +main (int argc, char *argv[]) +{ + struct A var; + + memset (&var, 0, sizeof var); + var.arr1 = (int *) calloc (N, sizeof (int)); + var.arr2 = (float *) calloc (N, sizeof (float)); + + { + #pragma omp declare mapper (struct A x) map(to: x.arr1) \ + map(tofrom: x.arr1[0:N]) + #pragma omp target + { + for (int i = 0; i < N; i++) + var.arr1[i]++; + } + } + + { + #pragma omp declare mapper (struct A x) map(to: x.arr2) \ + map(tofrom: x.arr2[0:N]) + #pragma omp target + { + for (int i = 0; i < N; i++) + var.arr2[i]++; + } + } + + { + #pragma omp declare mapper (struct A x) map(tofrom: x.arr3[0:N]) + #pragma omp target + { + for (int i = 0; i < N; i++) + var.arr3[i]++; + } + } + + for (int i = 0; i < N; i++) + { + assert (var.arr1[i] == 1); + assert (var.arr2[i] == 1); + assert (var.arr3[i] == 1); + } + + free (var.arr1); + free (var.arr2); +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/interop-2.c b/libgomp/testsuite/libgomp.c-c++-common/interop-2.c new file mode 100644 index 0000000..a7526dc --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/interop-2.c @@ -0,0 +1,129 @@ +/* { dg-do run } */ +/* { dg-additional-options "-lm" } */ + +/* Note: At the time this program was written, Nvptx was not asynchronous + enough to trigger the issue (with a 'nowait' added); however, one + AMD GPUs, it triggered. */ + +/* Test whether nowait / dependency is handled correctly. + Motivated by OpenMP_VV's 5.1/interop/test_interop_target.c + + The code actually only creates a streaming object without actually using it, + except for dependency tracking. + + Note that there is a difference between having a steaming (targetsync) object + and not (= omp_interop_none); at least if one assumes that omp_interop_none + does not include 'targetsync' as (effective) interop type - in that case, + 'nowait' has no effect and the 'depend' is active as included task, otherwise + the code continues with the depend being active only for the about to be + destroyed or used thread. + + The OpenMP spec states (here 6.0): + "If the interop-type set includes 'targetsync', an empty mergeable task is + generated. If the 'nowait' clause is not present on the construct then + the task is also an included task. If the interop-type set does not + include 'targetsync', the 'nowait' clause has no effect. Any depend + clauses that are present on the construct apply to the generated task. */ + +#include <omp.h> + +void +test_async (const int dev) +{ + constexpr int N = 2048; + constexpr int ulp = 4; + constexpr double M_PI = 2.0 * __builtin_acos (0.0); + omp_interop_t obj1, obj2; + double A[N] = { }; + int B[N] = { }; + + /* Create interop object. */ + #pragma omp interop device(dev) init(targetsync : obj1, obj2) + + if (dev == omp_initial_device || dev == omp_get_num_devices ()) + { + if (obj1 != omp_interop_none || obj2 != omp_interop_none) + __builtin_abort (); + } + else + { + if (obj1 == omp_interop_none || obj2 == omp_interop_none) + __builtin_abort (); + } + + /* DOUBLE */ + + /* Now in the background update it, slowly enough that the + code afterwards is reached while still running asynchronously. + As OpenMP_VV's Issue #863 shows, the overhead is high enough to + fail even when only doing an atomic integer increment. */ + + #pragma omp target device(dev) map(A) depend(out: A[:N]) nowait + for (int i = 0; i < N; i++) + #pragma omp atomic update + A[i] += __builtin_sin (2*i*M_PI/N); + + /* DESTROY take care of the dependeny such that ... */ + + if (obj1 == omp_interop_none) + { + // Same as below as 'nowait' is ignored. + #pragma omp interop destroy(obj1) depend(in: A[:N]) nowait + } + else + { + #pragma omp interop destroy(obj1) depend(in: A[:N]) + } + + /* ... this code is only executed once the dependency as been fulfilled. */ + + /* Check the value - part I: quick, avoid A[0] == sin(0) = 0. */ + for (int i = 1; i < N; i++) + if (A[i] == 0.0) + __builtin_abort (); + + /* Check the value - part II: throughly */ + for (int i = 0; i < N; i++) + { + double x = A[i]; + double y = __builtin_sin (2*i*M_PI/N); + if (__builtin_fabs (x - y) > ulp * __builtin_fabs (x+y) * __DBL_EPSILON__) + __builtin_abort (); + } + + /* Integer */ + + #pragma omp target device(dev) map(B) depend(out: B[:N]) nowait + for (int i = 0; i < N; i++) + #pragma omp atomic update + B[i] += 42; + + /* Same - but using USE. */ + if (obj2 == omp_interop_none) + { + // Same as below as 'nowait' is ignored. + #pragma omp interop use(obj2) depend(in: B[:N]) nowait + } + else + { + #pragma omp interop use(obj2) depend(in: B[:N]) + } + + for (int i = 0; i < N; i++) + if (B[i] != 42) + __builtin_abort (); + + #pragma omp interop destroy(obj2) +} + +int +main () +{ + int ndev = omp_get_num_devices (); + + for (int dev = 0; dev <= ndev; dev++) + test_async (dev); + test_async (omp_initial_device); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/metadirective-1.c b/libgomp/testsuite/libgomp.c-c++-common/metadirective-1.c index a57d6fd..fbe4ac3 100644 --- a/libgomp/testsuite/libgomp.c-c++-common/metadirective-1.c +++ b/libgomp/testsuite/libgomp.c-c++-common/metadirective-1.c @@ -1,4 +1,5 @@ -/* { dg-do run } */ +/* { dg-do run { target { ! offload_target_nvptx } } } */ +/* { dg-do compile { target offload_target_nvptx } } */ #define N 100 @@ -7,12 +8,17 @@ f (int x[], int y[], int z[]) { int i; + // The following fails as on the host the target side cannot be + // resolved - and the 'teams' or not status affects how 'target' + // is called. + // Note also the dg-do compile above for offload_target_nvptx #pragma omp target map(to: x[0:N], y[0:N]) map(from: z[0:N]) #pragma omp metadirective \ when (device={arch("nvptx")}: teams loop) \ default (parallel loop) for (i = 0; i < N; i++) z[i] = x[i] * y[i]; + /* { dg-bogus "'target' construct with nested 'teams' construct contains directives outside of the 'teams' construct" "PR118694" { xfail offload_target_nvptx } .-6 } */ } int diff --git a/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset-2.c b/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset-2.c new file mode 100644 index 0000000..b36d2f5 --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset-2.c @@ -0,0 +1,62 @@ +// PR libgomp/120444 +// Async version + +#include <omp.h> + +int main() +{ + #pragma omp parallel for + for (int dev = omp_initial_device; dev <= omp_get_num_devices (); dev++) + { + char *ptr = (char *) omp_target_alloc (sizeof(int) * 1024, dev); + + omp_depend_t dep; + #pragma omp depobj(dep) depend(inout: ptr) + + /* Play also around with the alignment - as hsa_amd_memory_fill operates + on multiples of 4 bytes (uint32_t). */ + + for (int start = 0; start < 32; start++) + for (int tail = 0; tail < 32; tail++) + { + unsigned char val = '0' + start + tail; +#if __cplusplus + void *ptr2 = omp_target_memset_async (ptr + start, val, + 1024 - start - tail, dev, 0); +#else + void *ptr2 = omp_target_memset_async (ptr + start, val, + 1024 - start - tail, dev, 0, nullptr); +#endif + if (ptr + start != ptr2) + __builtin_abort (); + + #pragma omp taskwait + + #pragma omp target device(dev) is_device_ptr(ptr) depend(depobj: dep) nowait + for (int i = start; i < 1024 - start - tail; i++) + { + if (ptr[i] != val) + __builtin_abort (); + ptr[i] += 2; + } + + omp_target_memset_async (ptr + start, val + 3, + 1024 - start - tail, dev, 1, &dep); + + #pragma omp target device(dev) is_device_ptr(ptr) depend(depobj: dep) nowait + for (int i = start; i < 1024 - start - tail; i++) + { + if (ptr[i] != val + 3) + __builtin_abort (); + ptr[i] += 1; + } + + omp_target_memset_async (ptr + start, val - 3, + 1024 - start - tail, dev, 1, &dep); + + #pragma omp taskwait depend (depobj: dep) + } + #pragma omp depobj(dep) destroy + omp_target_free (ptr, dev); + } +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset-3.c b/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset-3.c new file mode 100644 index 0000000..c0e4fa9 --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset-3.c @@ -0,0 +1,80 @@ +#include <stddef.h> +#include <stdint.h> +#include <omp.h> + +#define MIN(x,y) ((x) < (y) ? x : y) + +enum { N = 524288 + 8 }; + +static void +init_val (int8_t *ptr, int val, size_t count) +{ + #pragma omp target is_device_ptr(ptr) firstprivate(val, count) + __builtin_memset (ptr, val, count); +} + +static void +check_val (int8_t *ptr, int val, size_t count) +{ + if (count == 0) + return; + #pragma omp target is_device_ptr(ptr) firstprivate(val, count) + for (size_t i = 0; i < count; i++) + if (ptr[i] != val) __builtin_abort (); +} + +static void +test_it (int8_t *ptr, int lshift, size_t count) +{ + if (N < count + lshift) __builtin_abort (); + if (lshift >= 4) __builtin_abort (); + ptr += lshift; + + init_val (ptr, 'z', MIN (count + 32, N - lshift)); + + omp_target_memset (ptr, '1', count, omp_get_default_device()); + + check_val (ptr, '1', count); + check_val (ptr + count, 'z', MIN (32, N - lshift - count)); +} + + +int main() +{ + size_t size; + int8_t *ptr = (int8_t *) omp_target_alloc (N + 3, omp_get_default_device()); + ptr += (4 - (uintptr_t) ptr % 4) % 4; + if ((uintptr_t) ptr % 4 != 0) __builtin_abort (); + + test_it (ptr, 0, 1); + test_it (ptr, 3, 1); + test_it (ptr, 0, 4); + test_it (ptr, 3, 4); + test_it (ptr, 0, 5); + test_it (ptr, 3, 5); + test_it (ptr, 0, 6); + test_it (ptr, 3, 6); + + for (int i = 1; i <= 9; i++) + { + switch (i) + { + case 1: size = 16; break; // = 2^4 bytes + case 2: size = 32; break; // = 2^5 bytes + case 3: size = 64; break; // = 2^7 bytes + case 4: size = 128; break; // = 2^7 bytes + case 5: size = 256; break; // = 2^8 bytes + case 6: size = 512; break; // = 2^9 bytes + case 7: size = 65536; break; // = 2^16 bytes + case 8: size = 262144; break; // = 2^18 bytes + case 9: size = 524288; break; // = 2^20 bytes + default: __builtin_abort (); + } + test_it (ptr, 0, size); + test_it (ptr, 3, size); + test_it (ptr, 0, size + 1); + test_it (ptr, 3, size + 1); + test_it (ptr, 3, size + 2); + } + omp_target_free (ptr, omp_get_default_device()); +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset.c b/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset.c new file mode 100644 index 0000000..01909f8 --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/omp_target_memset.c @@ -0,0 +1,62 @@ +// PR libgomp/120444 + +#include <omp.h> + +int main() +{ + for (int dev = omp_initial_device; dev < omp_get_num_devices (); dev++) + { + char *ptr = (char *) omp_target_alloc (sizeof(int) * 1024, dev); + + /* Play also around with the alignment - as hsa_amd_memory_fill operates + on multiples of 4 bytes (uint32_t). */ + + for (int start = 0; start < 32; start++) + for (int tail = 0; tail < 32; tail++) + { + unsigned char val = '0' + start + tail; + void *ptr2 = omp_target_memset (ptr + start, val, + 1024 - start - tail, dev); + if (ptr + start != ptr2) + __builtin_abort (); + + #pragma omp target device(dev) is_device_ptr(ptr) + for (int i = start; i < 1024 - start - tail; i++) + if (ptr[i] != val) + __builtin_abort (); + + } + + /* Check 'small' values for correctness. */ + + for (int start = 0; start < 32; start++) + for (int size = 0; size <= 64 + 32; size++) + { + omp_target_memset (ptr, 'a' - 2, 1024, dev); + + unsigned char val = '0' + start + size % 32; + void *ptr2 = omp_target_memset (ptr + start, val, size, dev); + + if (ptr + start != ptr2) + __builtin_abort (); + + if (size == 0) + continue; + + #pragma omp target device(dev) is_device_ptr(ptr) + { + for (int i = 0; i < start; i++) + if (ptr[i] != 'a' - 2) + __builtin_abort (); + for (int i = start; i < start + size; i++) + if (ptr[i] != val) + __builtin_abort (); + for (int i = start + size + 1; i < 1024; i++) + if (ptr[i] != 'a' - 2) + __builtin_abort (); + } + } + + omp_target_free (ptr, dev); + } +} diff --git a/libgomp/testsuite/libgomp.c-c++-common/pr96390.c b/libgomp/testsuite/libgomp.c-c++-common/pr96390.c index b89f934..ca7865d 100644 --- a/libgomp/testsuite/libgomp.c-c++-common/pr96390.c +++ b/libgomp/testsuite/libgomp.c-c++-common/pr96390.c @@ -1,7 +1,7 @@ /* { dg-additional-options "-O0 -fdump-tree-omplower" } */ /* { dg-additional-options "-foffload=-Wa,--verify" { target offload_target_nvptx } } */ /* { dg-require-alias "" } */ -/* { dg-xfail-if "PR 97102/PR 97106 - .alias not (yet) supported for nvptx" { offload_target_nvptx } } */ +/* { dg-xfail-if PR105018 { offload_target_nvptx } } */ #ifdef __cplusplus extern "C" { diff --git a/libgomp/testsuite/libgomp.c-c++-common/target-abi-struct-1-O0.c b/libgomp/testsuite/libgomp.c-c++-common/target-abi-struct-1-O0.c new file mode 100644 index 0000000..9bf949a --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/target-abi-struct-1-O0.c @@ -0,0 +1,3 @@ +/* { dg-additional-options -O0 } */ + +#include "target-abi-struct-1.c" diff --git a/libgomp/testsuite/libgomp.c-c++-common/target-abi-struct-1.c b/libgomp/testsuite/libgomp.c-c++-common/target-abi-struct-1.c new file mode 100644 index 0000000..d9268af --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/target-abi-struct-1.c @@ -0,0 +1 @@ +#include "../libgomp.oacc-c-c++-common/abi-struct-1.c" diff --git a/libgomp/testsuite/libgomp.c-c++-common/target-cdtor-1.c b/libgomp/testsuite/libgomp.c-c++-common/target-cdtor-1.c new file mode 100644 index 0000000..e6099cf --- /dev/null +++ b/libgomp/testsuite/libgomp.c-c++-common/target-cdtor-1.c @@ -0,0 +1,89 @@ +/* Offloaded 'constructor' and 'destructor' functions. */ + +#include <omp.h> + +#pragma omp declare target + +static void +__attribute__((constructor)) +initHD1() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +static void +__attribute__((constructor)) +initHD2() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +static void +__attribute__((destructor)) +finiHD1() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +static void +__attribute__((destructor)) +finiHD2() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +#pragma omp end declare target + +static void +__attribute__((constructor)) +initH1() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +static void +__attribute__((destructor)) +finiH2() +{ + __builtin_printf("%s, %d\n", __FUNCTION__, omp_is_initial_device()); +} + +int main() +{ + int c = 0; + + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + +#pragma omp target map(c) + { + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + } + +#pragma omp target map(c) + { + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + } + + __builtin_printf("%s:%d, %d\n", __FUNCTION__, ++c, omp_is_initial_device()); + + return 0; +} + +/* The order is undefined, in which same-priority 'constructor' functions, and 'destructor' functions are run. + { dg-output {init[^,]+, 1[\r\n]+} } + { dg-output {init[^,]+, 1[\r\n]+} } + { dg-output {init[^,]+, 1[\r\n]+} } + { dg-output {main:1, 1[\r\n]+} } + { dg-output {initHD[^,]+, 0[\r\n]+} { target offload_device } } + { dg-output {initHD[^,]+, 0[\r\n]+} { target offload_device } } + { dg-output {main:2, 1[\r\n]+} { target { ! offload_device } } } + { dg-output {main:2, 0[\r\n]+} { target offload_device } } + { dg-output {main:3, 1[\r\n]+} { target { ! offload_device } } } + { dg-output {main:3, 0[\r\n]+} { target offload_device } } + { dg-output {main:4, 1[\r\n]+} } + { dg-output {finiHD[^,]+, 0[\r\n]+} { target offload_device } } + { dg-output {finiHD[^,]+, 0[\r\n]+} { target offload_device } } + { dg-output {fini[^,]+, 1[\r\n]+} } + { dg-output {fini[^,]+, 1[\r\n]+} } + { dg-output {fini[^,]+, 1[\r\n]+} } +*/ diff --git a/libgomp/testsuite/libgomp.c-target/aarch64/udr-sve.c b/libgomp/testsuite/libgomp.c-target/aarch64/udr-sve.c index 03d93cc..02e02dc 100644 --- a/libgomp/testsuite/libgomp.c-target/aarch64/udr-sve.c +++ b/libgomp/testsuite/libgomp.c-target/aarch64/udr-sve.c @@ -9,8 +9,8 @@ void __attribute__ ((noipa)) parallel_reduction () { - int a[8] = {1 ,1, 1, 1, 1, 1, 1, 1}; - int b[8] = {0 ,0, 0, 0, 0, 0, 0, 0}; + int a[8] = {1, 1, 1, 1, 1, 1, 1, 1}; + int b[8] = {0, 0, 0, 0, 0, 0, 0, 0}; svint32_t va = svld1_s32 (svptrue_b32 (), b); int i = 0; int64_t res; @@ -30,8 +30,8 @@ parallel_reduction () void __attribute__ ((noipa)) for_reduction () { - int a[8] = {1 ,1, 1, 1, 1, 1, 1, 1}; - int b[8] = {0 ,0, 0, 0, 0, 0, 0, 0}; + int a[8] = {1, 1, 1, 1, 1, 1, 1, 1}; + int b[8] = {0, 0, 0, 0, 0, 0, 0, 0}; svint32_t va = svld1_s32 (svptrue_b32 (), b); int j; int64_t res; @@ -58,13 +58,13 @@ simd_reduction () for (j = 0; j < 8; j++) a[j] = 1; - #pragma omp simd reduction (+:va, i) + #pragma omp simd reduction (+:va) for (j = 0; j < 16; j++) - va = svld1_s32 (svptrue_b32 (), a); + va += svld1_s32 (svptrue_b32 (), a); res = svaddv_s32 (svptrue_b32 (), va); - if (res != 8) + if (res != 128) __builtin_abort (); } @@ -72,22 +72,57 @@ void __attribute__ ((noipa)) inscan_reduction_incl () { svint32_t va = svindex_s32 (0, 0); + int a[8] = {1, 1, 1, 1, 1, 1, 1, 1}; + int b[64] = { 0 }; int j; int64_t res = 0; - #pragma omp parallel - #pragma omp for reduction (inscan,+:va) firstprivate (res) lastprivate (res) + #pragma omp parallel for reduction (inscan, +:va) for (j = 0; j < 8; j++) { - va = svindex_s32 (1, 0); + va += svld1_s32 (svptrue_b32 (), a); #pragma omp scan inclusive (va) - res += svaddv_s32 (svptrue_b32 (), va); + svst1_s32 (svptrue_b32 (), b + j * 8, va); + } + + res = svaddv_s32 (svptrue_b32 (), va); + + if (res != 64) + __builtin_abort (); + + for (j = 0; j < 64; j+=8) + if (b[j] != (j / 8 + 1)) + __builtin_abort (); +} + +void __attribute__ ((noipa)) +inscan_reduction_excl () +{ + svint32_t va = svindex_s32 (0, 0); + int a[8] = {1, 1, 1, 1, 1, 1, 1, 1}; + int b[64] = { 0 }; + int j; + int64_t res = 0; + + #pragma omp parallel for reduction (inscan, +:va) + for (j = 0; j < 8; j++) + { + svst1_s32 (svptrue_b32 (), b + j * 8, va); + #pragma omp scan exclusive (va) + va += svld1_s32 (svptrue_b32 (), a); } + res = svaddv_s32 (svptrue_b32 (), va); + if (res != 64) __builtin_abort (); + + for (j = 0; j < 64; j+=8) + if (b[j] != j / 8) + __builtin_abort (); } + int main () { @@ -95,4 +130,5 @@ main () for_reduction (); simd_reduction (); inscan_reduction_incl (); + inscan_reduction_excl (); } diff --git a/libgomp/testsuite/libgomp.c/declare-variant-3-sm61.c b/libgomp/testsuite/libgomp.c/declare-variant-3-sm61.c new file mode 100644 index 0000000..e6941d3 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/declare-variant-3-sm61.c @@ -0,0 +1,8 @@ +/* { dg-do link { target { offload_target_nvptx } } } */ +/* { dg-additional-options -foffload=nvptx-none } */ +/* { dg-additional-options "-foffload=-misa=sm_61 -foffload=-mptx=_" } */ +/* { dg-additional-options "-foffload=-fdump-tree-optimized" } */ + +#include "declare-variant-3.h" + +/* { dg-final { only_for_offload_target nvptx-none scan-offload-tree-dump "= f61 \\(\\);" "optimized" } } */ diff --git a/libgomp/testsuite/libgomp.c/declare-variant-3.h b/libgomp/testsuite/libgomp.c/declare-variant-3.h index c9c8f4a..f5695a2 100644 --- a/libgomp/testsuite/libgomp.c/declare-variant-3.h +++ b/libgomp/testsuite/libgomp.c/declare-variant-3.h @@ -37,6 +37,13 @@ f53 (void) __attribute__ ((noipa)) int +f61 (void) +{ + return 61; +} + +__attribute__ ((noipa)) +int f70 (void) { return 70; @@ -68,6 +75,7 @@ f89 (void) #pragma omp declare variant (f37) match (device={isa("sm_37")}) #pragma omp declare variant (f52) match (device={isa("sm_52")}) #pragma omp declare variant (f53) match (device={isa("sm_53")}) +#pragma omp declare variant (f61) match (device={isa("sm_61")}) #pragma omp declare variant (f70) match (device={isa("sm_70")}) #pragma omp declare variant (f75) match (device={isa("sm_75")}) #pragma omp declare variant (f80) match (device={isa("sm_80")}) diff --git a/libgomp/testsuite/libgomp.c/declare-variant-4-gfx942.c b/libgomp/testsuite/libgomp.c/declare-variant-4-gfx942.c new file mode 100644 index 0000000..d1df550 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/declare-variant-4-gfx942.c @@ -0,0 +1,8 @@ +/* { dg-do link { target { offload_target_amdgcn } } } */ +/* { dg-additional-options -foffload=amdgcn-amdhsa } */ +/* { dg-additional-options -foffload=-march=gfx942 } */ +/* { dg-additional-options "-foffload=-fdump-tree-optimized" } */ + +#include "declare-variant-4.h" + +/* { dg-final { only_for_offload_target amdgcn-amdhsa scan-offload-tree-dump "= gfx942 \\(\\);" "optimized" } } */ diff --git a/libgomp/testsuite/libgomp.c/declare-variant-4.h b/libgomp/testsuite/libgomp.c/declare-variant-4.h index 53788d2..2257f4c 100644 --- a/libgomp/testsuite/libgomp.c/declare-variant-4.h +++ b/libgomp/testsuite/libgomp.c/declare-variant-4.h @@ -37,6 +37,13 @@ gfx90c (void) __attribute__ ((noipa)) int +gfx942 (void) +{ + return 0x942; +} + +__attribute__ ((noipa)) +int gfx1030 (void) { return 0x1030; @@ -68,6 +75,7 @@ gfx1103 (void) #pragma omp declare variant(gfx908) match(device = {isa("gfx908")}) #pragma omp declare variant(gfx90a) match(device = {isa("gfx90a")}) #pragma omp declare variant(gfx90c) match(device = {isa("gfx90c")}) +#pragma omp declare variant(gfx942) match(device = {isa("gfx942")}) #pragma omp declare variant(gfx1030) match(device = {isa("gfx1030")}) #pragma omp declare variant(gfx1036) match(device = {isa("gfx1036")}) #pragma omp declare variant(gfx1100) match(device = {isa("gfx1100")}) diff --git a/libgomp/testsuite/libgomp.c/interop-cublas-full.c b/libgomp/testsuite/libgomp.c/interop-cublas-full.c new file mode 100644 index 0000000..2df5277 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-cublas-full.c @@ -0,0 +1,176 @@ +/* { dg-require-effective-target openacc_cublas } */ +/* { dg-additional-options "-lcublas" } */ + +/* NOTE: This file is also included by libgomp.c-c++-common/interop-cudablas-libonly.c + to test the fallback version. */ + +/* Check whether cuBlas' daxpy works with an interop object. + daxpy(N, DA, DX, INCX, DY, INCY) + calculates (for DX = DY = 1): + DY(1:N) = DY(1:N) + DA * DX(1:N) + and otherwise N array elements, taking every INCX-th or INCY-th one, repectively. + +Based on the interop example in OpenMP's example document */ + +/* Minimal check whether CUDA works - by checking whether the API routines + seem to work. This includes a fallback if the header is not + available. */ + +#include <assert.h> +#include <omp.h> +#include "../libgomp.c-c++-common/on_device_arch.h" + + +#if __has_include(<cuda.h>) && __has_include(<cudaTypedefs.h>) && __has_include(<cuda_runtime.h>) && __has_include(<cublas_v2.h>) && !defined(USE_CUDA_FALLBACK_HEADER) + #include <cuda.h> + #include <cudaTypedefs.h> + #include <cuda_runtime.h> + #include <cublas_v2.h> + +#else + /* Add a poor man's fallback declaration. */ + #if USE_CUDA_FALLBACK_HEADER + // Don't warn. + #elif !__has_include(<cuda.h>) + #warning "Using GCC's cuda.h as fallback for cuda.h" + #elif !__has_include(<cudaTypedefs.h>) + #warning "Using GCC's cuda.h as fallback for cudaTypedefs.h" + #elif !__has_include(<cuda_runtime.h>) + #warning "Using GCC's cuda.h as fallback for cuda_runtime.h" + #else + #warning "Using GCC's cuda.h as fallback for cublas_v2.h" + #endif + #include "../../../include/cuda/cuda.h" + + typedef enum { + CUBLAS_STATUS_SUCCESS = 0, + } cublasStatus_t; + + typedef CUstream cudaStream_t; + typedef struct cublasContext* cublasHandle_t; + + #define cublasCreate cublasCreate_v2 + cublasStatus_t cublasCreate_v2 (cublasHandle_t *); + + #define cublasSetStream cublasSetStream_v2 + cublasStatus_t cublasSetStream_v2 (cublasHandle_t, cudaStream_t); + + #define cublasDaxpy cublasDaxpy_v2 + cublasStatus_t cublasDaxpy_v2(cublasHandle_t, int, const double*, const double*, int, double*, int); +#endif + +static int used_variant = 0; + +void +run_cuBlasdaxpy (int n, double da, const double *dx, int incx, double *dy, int incy, omp_interop_t obj) +{ + used_variant = 1; + + omp_interop_rc_t res; + cublasStatus_t stat; + + omp_intptr_t fr = omp_get_interop_int(obj, omp_ipr_fr_id, &res); + assert (res == omp_irc_success && fr == omp_ifr_cuda); + + cudaStream_t stream = (cudaStream_t) omp_get_interop_ptr (obj, omp_ipr_targetsync, &res); + assert (res == omp_irc_success); + + cublasHandle_t handle; + stat = cublasCreate (&handle); + assert (stat == CUBLAS_STATUS_SUCCESS); + + stat = cublasSetStream (handle, stream); + assert (stat == CUBLAS_STATUS_SUCCESS); + + /* 'da' can be in host or device space, 'dx' and 'dy' must be in device space. */ + stat = cublasDaxpy (handle, n, &da, dx, 1, dy, 1) ; + assert (stat == CUBLAS_STATUS_SUCCESS); +} + + +#pragma omp declare variant(run_cuBlasdaxpy) \ + match(construct={dispatch}, target_device={kind(nohost), arch("nvptx")}) \ + adjust_args(need_device_ptr : dx, dy) \ + append_args(interop(targetsync, prefer_type("cuda"))) + +void +run_daxpy (int n, double da, const double *dx, int incx, double *dy, int incy) +{ + used_variant = 2; + + if (incx == 1 && incy == 1) + #pragma omp simd + for (int i = 0; i < n; i++) + dy[i] += da * dx[i]; + else + { + int ix = 0; + int iy = 0; + for (int i = 0; i < n; i++) + { + dy[iy] += da * dx[ix]; + ix += incx; + iy += incy; + } + } +} + + +void +run_test (int dev) +{ + constexpr int N = 1024; + + // A = {1,2,...,N} + // B = {-1, -2, ..., N} + // B' = daxpy (N, 3, A, incx=1, B, incy=1) + // = B + 3*A + // -> B' = {0, 2, 4, 6, ... } + + double A[N], B[N]; + double factor = 3.0; + for (int i = 0; i < N; i++) + { + A[i] = i; + B[i] = -i; + } + + if (dev != omp_initial_device && dev != omp_get_num_devices ()) + { + #pragma omp target enter data device(dev) map(A, B) + } + + used_variant = 99; + #pragma omp dispatch device(dev) + run_daxpy (N, factor, A, 1, B, 1); + + if (dev != omp_initial_device && dev != omp_get_num_devices ()) + { + #pragma omp target exit data device(dev) map(release: A) map(from: B) + + int tmp = omp_get_default_device (); + omp_set_default_device (dev); + if (on_device_arch_nvptx ()) + assert (used_variant == 1); + else + assert (used_variant == 2); + omp_set_default_device (tmp); + } + else + assert (used_variant == 2); + + for (int i = 0; i < N; i++) + assert (B[i] == 2*i); +} + +int +main () +{ + int ndev = omp_get_num_devices (); + + for (int dev = 0; dev <= ndev; dev++) + run_test (dev); + run_test (omp_initial_device); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c/interop-cublas-libonly.c b/libgomp/testsuite/libgomp.c/interop-cublas-libonly.c new file mode 100644 index 0000000..89c0652 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-cublas-libonly.c @@ -0,0 +1,7 @@ +/* { dg-require-effective-target openacc_libcublas } */ +/* { dg-additional-options "-lcublas" } */ + +/* Same as interop-cudablas-full.c, but also works if the header is not available. */ + +#define USE_CUDA_FALLBACK_HEADER 1 +#include "interop-cublas-full.c" diff --git a/libgomp/testsuite/libgomp.c/interop-cuda-full.c b/libgomp/testsuite/libgomp.c/interop-cuda-full.c new file mode 100644 index 0000000..c48a934 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-cuda-full.c @@ -0,0 +1,162 @@ +/* { dg-do run { target { offload_device_nvptx } } } */ +/* { dg-do link { target { ! offload_device_nvptx } } } */ + +/* { dg-require-effective-target openacc_cuda } */ +/* { dg-require-effective-target openacc_cudart } */ +/* { dg-additional-options "-lcuda -lcudart" } */ + +/* NOTE: This file is also included by libgomp.c-c++-common/interop-cuda-libonly.c + to test the fallback version, which defines USE_CUDA_FALLBACK_HEADER. */ + +/* Minimal check whether CUDA works - by checking whether the API routines + seem to work. This includes a fallback if the header is not + available. */ + +#include <assert.h> +#include <omp.h> + +#if __has_include(<cuda.h>) && __has_include(<cudaTypedefs.h>) && __has_include(<cuda_runtime.h>) && !defined(USE_CUDA_FALLBACK_HEADER) + #include <cuda.h> + #include <cudaTypedefs.h> + #include <cuda_runtime.h> + +#else + /* Add a poor man's fallback declaration. */ + #if USE_CUDA_FALLBACK_HEADER + // Don't warn. + #elif !__has_include(<cuda.h>) + #warning "Using GCC's cuda.h as fallback for cuda.h" + #elif !__has_include(<cudaTypedefs.h>) + #warning "Using GCC's cuda.h as fallback for cudaTypedefs.h" + #else + #warning "Using GCC's cuda.h as fallback for cuda_runtime.h" + #endif + #include "../../../include/cuda/cuda.h" + + typedef int cudaError_t; + typedef CUstream cudaStream_t; + enum { + cudaSuccess = 0 + }; + + enum cudaDeviceAttr { + cudaDevAttrClockRate = 13, + cudaDevAttrMaxGridDimX = 5 + }; + + cudaError_t cudaDeviceGetAttribute (int *, enum cudaDeviceAttr, int); + cudaError_t cudaStreamQuery(cudaStream_t); + CUresult cuCtxGetApiVersion(CUcontext, unsigned int *); + CUresult cuStreamGetCtx (CUstream, CUcontext *); +#endif + +int +main () +{ + int ivar; + unsigned uvar; + omp_interop_rc_t res; + omp_interop_t obj_cuda = omp_interop_none; + omp_interop_t obj_cuda_driver = omp_interop_none; + cudaError_t cuda_err; + CUresult cu_err; + + #pragma omp interop init(target, targetsync, prefer_type("cuda") : obj_cuda) \ + init(target, targetsync, prefer_type("cuda_driver") : obj_cuda_driver) \ + + omp_interop_fr_t fr = (omp_interop_fr_t) omp_get_interop_int (obj_cuda, omp_ipr_fr_id, &res); + assert (res == omp_irc_success); + assert (fr == omp_ifr_cuda); + + fr = (omp_interop_fr_t) omp_get_interop_int (obj_cuda_driver, omp_ipr_fr_id, &res); + assert (res == omp_irc_success); + assert (fr == omp_ifr_cuda_driver); + + ivar = (int) omp_get_interop_int (obj_cuda, omp_ipr_vendor, &res); + assert (res == omp_irc_success); + assert (ivar == 11); + + ivar = (int) omp_get_interop_int (obj_cuda_driver, omp_ipr_vendor, &res); + assert (res == omp_irc_success); + assert (ivar == 11); + + + /* Check whether the omp_ipr_device -> cudaDevice_t yields a valid device. */ + + CUdevice cu_dev = (int) omp_get_interop_int (obj_cuda_driver, omp_ipr_device, &res); + assert (res == omp_irc_success); + + /* Assume a clock size is available and > 1 GHz; value is in kHz. */ + cu_err = cuDeviceGetAttribute (&ivar, cudaDevAttrClockRate, cu_dev); + assert (cu_err == CUDA_SUCCESS); + assert (ivar > 1000000 /* kHz */); + + /* Assume that the MaxGridDimX is available and > 1024. */ + cu_err = cuDeviceGetAttribute (&ivar, cudaDevAttrMaxGridDimX, cu_dev); + assert (cu_err == CUDA_SUCCESS); + assert (ivar > 1024); + + int cuda_dev = (int) omp_get_interop_int (obj_cuda, omp_ipr_device, &res); + assert (res == omp_irc_success); + assert (cuda_dev == (CUdevice) cu_dev); // Assume they are the same ... + + /* Assume a clock size is available and > 1 GHz; value is in kHz. */ + cuda_err = cudaDeviceGetAttribute (&ivar, cudaDevAttrClockRate, cuda_dev); + assert (cuda_err == cudaSuccess); + assert (ivar > 1000000 /* kHz */); + + /* Assume that the MaxGridDimX is available and > 1024. */ + cuda_err = cudaDeviceGetAttribute (&ivar, cudaDevAttrMaxGridDimX, cuda_dev); + assert (cuda_err == cudaSuccess); + assert (ivar > 1024); + + + + + /* Check whether the omp_ipr_device_context -> CUcontext yields a context. */ + + CUcontext cu_ctx = (CUcontext) omp_get_interop_ptr (obj_cuda_driver, omp_ipr_device_context, &res); + assert (res == omp_irc_success); + + /* Assume API Version > 0 for Nvidia, cudaErrorNotSupported for AMD. */ + uvar = 99; + cu_err = cuCtxGetApiVersion (cu_ctx, &uvar); + assert (cu_err == CUDA_SUCCESS); + assert (uvar > 0); + + + /* Check whether the omp_ipr_targetsync -> cudaStream_t yields a stream. */ + + cudaStream_t cuda_sm = (cudaStream_t) omp_get_interop_ptr (obj_cuda, omp_ipr_targetsync, &res); + assert (res == omp_irc_success); + + CUstream cu_sm = (cudaStream_t) omp_get_interop_ptr (obj_cuda_driver, omp_ipr_targetsync, &res); + assert (res == omp_irc_success); + + assert ((void*) cu_sm != (void*) cuda_sm); // Type compatible but should have created two streams + + int dev_stream = 99; +#if CUDA_VERSION >= 12080 + cuda_err = cudaStreamGetDevice (cuda_sm, &dev_stream); + assert (cuda_err == cudaSuccess); +#else + cu_err = cuStreamGetCtx (cu_sm, &cu_ctx) != CUDA_SUCCESS; + if (cu_err == CUDA_SUCCESS) + cuda_err = cuCtxPushCurrent (cu_ctx) != CUDA_SUCCESS; + if (cu_err == CUDA_SUCCESS) + cuda_err = cuCtxGetDevice (&dev_stream) != CUDA_SUCCESS; + if (cu_err == CUDA_SUCCESS) + cu_err = cuCtxPopCurrent (&cu_ctx) != CUDA_SUCCESS; + assert (cu_err == CUDA_SUCCESS); +#endif + assert (dev_stream == cuda_dev); + + /* All jobs should have been completed (as there were none none) */ + cuda_err = cudaStreamQuery (cuda_sm); + assert (cuda_err == cudaSuccess); + + cu_err = cuStreamQuery (cu_sm); + assert (cu_err == CUDA_SUCCESS); + + #pragma omp interop destroy(obj_cuda, obj_cuda_driver) +} diff --git a/libgomp/testsuite/libgomp.c/interop-cuda-libonly.c b/libgomp/testsuite/libgomp.c/interop-cuda-libonly.c new file mode 100644 index 0000000..bc257a2 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-cuda-libonly.c @@ -0,0 +1,11 @@ +/* { dg-do run { target { offload_device_nvptx } } } */ +/* { dg-do link { target { ! offload_device_nvptx } } } */ + +/* { dg-require-effective-target openacc_libcudart } */ +/* { dg-require-effective-target openacc_libcuda } */ +/* { dg-additional-options "-lcuda -lcudart" } */ + +/* Same as interop-cuda-full.c, but also works if the header is not available. */ + +#define USE_CUDA_FALLBACK_HEADER 1 +#include "interop-cuda-full.c" diff --git a/libgomp/testsuite/libgomp.c/interop-hip-amd-full.c b/libgomp/testsuite/libgomp.c/interop-hip-amd-full.c new file mode 100644 index 0000000..bd44f44 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hip-amd-full.c @@ -0,0 +1,10 @@ +/* { dg-do run { target { offload_device_gcn } } } */ +/* { dg-do link { target { ! offload_device_gcn } } } */ + +/* { dg-require-effective-target gomp_hip_header_amd } */ +/* { dg-require-effective-target gomp_libamdhip64 } */ +/* { dg-additional-options "-lamdhip64" } */ + +#define __HIP_PLATFORM_AMD__ 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hip-amd-no-hip-header.c b/libgomp/testsuite/libgomp.c/interop-hip-amd-no-hip-header.c new file mode 100644 index 0000000..91ad987 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hip-amd-no-hip-header.c @@ -0,0 +1,11 @@ +/* { dg-do run { target { offload_device_gcn } } } */ +/* { dg-do link { target { ! offload_device_gcn } } } */ + +/* { dg-require-effective-target gomp_libamdhip64 } */ +/* { dg-additional-options "-lamdhip64" } */ + +#define __HIP_PLATFORM_AMD__ 1 + +#define USE_HIP_FALLBACK_HEADER 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hip-nvidia-full.c b/libgomp/testsuite/libgomp.c/interop-hip-nvidia-full.c new file mode 100644 index 0000000..d5dc236 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hip-nvidia-full.c @@ -0,0 +1,11 @@ +/* { dg-do run { target { offload_device_nvptx } } } */ +/* { dg-do link { target { ! offload_device_nvptx } } } */ + +/* { dg-require-effective-target openacc_cudart } */ +/* { dg-require-effective-target openacc_cuda } */ +/* { dg-require-effective-target gomp_hip_header_nvidia } */ +/* { dg-additional-options "-lcuda -lcudart -Wno-deprecated-declarations" } */ + +#define __HIP_PLATFORM_NVIDIA__ 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hip-nvidia-no-headers.c b/libgomp/testsuite/libgomp.c/interop-hip-nvidia-no-headers.c new file mode 100644 index 0000000..7cff2cb --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hip-nvidia-no-headers.c @@ -0,0 +1,13 @@ +/* { dg-do run { target { offload_device_nvptx } } } */ +/* { dg-do link { target { ! offload_device_nvptx } } } */ + +/* { dg-require-effective-target openacc_libcudart } */ +/* { dg-require-effective-target openacc_libcuda } */ +/* { dg-additional-options "-lcuda -lcudart" } */ + +#define __HIP_PLATFORM_NVIDIA__ 1 + +#define USE_HIP_FALLBACK_HEADER 1 +#define USE_CUDA_FALLBACK_HEADER 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hip-nvidia-no-hip-header.c b/libgomp/testsuite/libgomp.c/interop-hip-nvidia-no-hip-header.c new file mode 100644 index 0000000..7b7dc74 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hip-nvidia-no-hip-header.c @@ -0,0 +1,12 @@ +/* { dg-do run { target { offload_device_nvptx } } } */ +/* { dg-do link { target { ! offload_device_nvptx } } } */ + +/* { dg-require-effective-target openacc_cudart } */ +/* { dg-require-effective-target openacc_cuda } */ +/* { dg-additional-options "-lcuda -lcudart" } */ + +#define __HIP_PLATFORM_NVIDIA__ 1 + +#define USE_HIP_FALLBACK_HEADER 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hip.h b/libgomp/testsuite/libgomp.c/interop-hip.h new file mode 100644 index 0000000..20a1ccb --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hip.h @@ -0,0 +1,234 @@ +/* Minimal check whether HIP works - by checking whether the API routines + seem to work. This includes various fallbacks if the header is not + available. */ + +#include <assert.h> +#include <omp.h> + +#if !defined(__HIP_PLATFORM_AMD__) && !defined(__HIP_PLATFORM_NVIDIA__) + #error "Either __HIP_PLATFORM_AMD__ or __HIP_PLATFORM_NVIDIA__ must be defined" +#endif + +#if defined(__HIP_PLATFORM_AMD__) && defined(__HIP_PLATFORM_NVIDIA__) + #error "Either __HIP_PLATFORM_AMD__ or __HIP_PLATFORM_NVIDIA__ must be defined" +#endif + +#if __has_include(<hip/hip_runtime_api.h>) && !defined(USE_HIP_FALLBACK_HEADER) + #include <hip/hip_runtime_api.h> + +#elif defined(__HIP_PLATFORM_AMD__) + /* Add a poor man's fallback declaration. */ + #if !defined(USE_HIP_FALLBACK_HEADER) + #warning "Using fallback declaration for <hip/hip_runtime_api.h> for __HIP_PLATFORM_AMD__" + #endif + + typedef struct ihipStream_t* hipStream_t; + typedef struct ihipCtx_t* hipCtx_t; + typedef int hipError_t; + typedef int hipDevice_t; + enum { + hipSuccess = 0, + hipErrorNotSupported = 801 + }; + + typedef enum hipDeviceAttribute_t { + hipDeviceAttributeClockRate = 5, + hipDeviceAttributeMaxGridDimX = 29 + } hipDeviceAttribute_t; + + hipError_t hipDeviceGetAttribute (int *, hipDeviceAttribute_t, hipDevice_t); + hipError_t hipCtxGetApiVersion (hipCtx_t, int *); + hipError_t hipStreamGetDevice (hipStream_t, hipDevice_t *); + hipError_t hipStreamQuery (hipStream_t); + +#elif defined(__HIP_PLATFORM_NVIDIA__) + /* Add a poor man's fallback declaration. */ + #if !defined(USE_HIP_FALLBACK_HEADER) + #warning "Using fallback declaration for <hip/hip_runtime_api.h> for __HIP_PLATFORM_NVIDIA__" + #endif + + #if __has_include(<cuda.h>) && __has_include(<cudaTypedefs.h>) && __has_include(<cuda_runtime.h>) && !defined(USE_CUDA_FALLBACK_HEADER) + #include <cuda.h> + #include <cudaTypedefs.h> + #include <cuda_runtime.h> + #else + #if defined(USE_CUDA_FALLBACK_HEADER) + // no warning + #elif !__has_include(<cuda.h>) + #warning "Using GCC's cuda.h as fallback for cuda.h" + #elif !__has_include(<cudaTypedefs.h>) + #warning "Using GCC's cuda.h as fallback for cudaTypedefs.h" + #else + #warning "Using GCC's cuda.h as fallback for cuda_runtime.h" + #endif + + #include "../../../include/cuda/cuda.h" + + typedef int cudaError_t; + enum { + cudaSuccess = 0 + }; + + enum cudaDeviceAttr { + cudaDevAttrClockRate = 13, + cudaDevAttrMaxGridDimX = 5 + }; + + cudaError_t cudaDeviceGetAttribute (int *, enum cudaDeviceAttr, int); + CUresult cuCtxGetApiVersion(CUcontext, unsigned int *); + CUresult cuStreamGetCtx (CUstream, CUcontext *); + #endif + + typedef CUstream hipStream_t; + typedef CUcontext hipCtx_t; + typedef CUdevice hipDevice_t; + + typedef int hipError_t; + typedef int hipDevice_t; + enum { + hipSuccess = 0, + hipErrorNotSupported = 801 + }; + + + typedef enum hipDeviceAttribute_t { + hipDeviceAttributeClockRate = 5, + hipDeviceAttributeMaxGridDimX = 29 + } hipDeviceAttribute_t; + + inline static hipError_t + hipDeviceGetAttribute (int *ival, hipDeviceAttribute_t attr, hipDevice_t dev) + { + enum cudaDeviceAttr cuattr; + switch (attr) + { + case hipDeviceAttributeClockRate: + cuattr = cudaDevAttrClockRate; + break; + case hipDeviceAttributeMaxGridDimX: + cuattr = cudaDevAttrMaxGridDimX; + break; + default: + assert (0); + } + return cudaDeviceGetAttribute (ival, cuattr, dev) != cudaSuccess; + } + + inline static hipError_t + hipCtxGetApiVersion (hipCtx_t ctx, int *ver) + { + unsigned uver; + hipError_t err; + err = cuCtxGetApiVersion (ctx, &uver) != CUDA_SUCCESS; + *ver = (int) uver; + return err; + } + + inline static hipError_t + hipStreamGetDevice (hipStream_t stream, hipDevice_t *dev) + { +#if CUDA_VERSION >= 12080 + return cudaStreamGetDevice (stream, dev); +#else + hipError_t err; + CUcontext ctx; + err = cuStreamGetCtx (stream, &ctx) != CUDA_SUCCESS; + if (err == hipSuccess) + err = cuCtxPushCurrent (ctx) != CUDA_SUCCESS; + if (err == hipSuccess) + err = cuCtxGetDevice (dev) != CUDA_SUCCESS; + if (err == hipSuccess) + err = cuCtxPopCurrent (&ctx) != CUDA_SUCCESS; + return err; +#endif + } + + inline static hipError_t + hipStreamQuery (hipStream_t stream) + { + return cuStreamQuery (stream) != CUDA_SUCCESS; + } + +#else + #error "should be unreachable" +#endif + +int +main () +{ + int ivar; + omp_interop_rc_t res; + omp_interop_t obj = omp_interop_none; + hipError_t hip_err; + + #pragma omp interop init(target, targetsync, prefer_type("hip") : obj) + + omp_interop_fr_t fr = (omp_interop_fr_t) omp_get_interop_int (obj, omp_ipr_fr_id, &res); + assert (res == omp_irc_success); + assert (fr == omp_ifr_hip); + + ivar = (int) omp_get_interop_int (obj, omp_ipr_vendor, &res); + assert (res == omp_irc_success); + int vendor_is_amd = ivar == 1; + #if defined(__HIP_PLATFORM_AMD__) + assert (ivar == 1); + #elif defined(__HIP_PLATFORM_NVIDIA__) + assert (ivar == 11); + #else + assert (0); + #endif + + + /* Check whether the omp_ipr_device -> hipDevice_t yields a valid device. */ + + hipDevice_t hip_dev = (int) omp_get_interop_int (obj, omp_ipr_device, &res); + assert (res == omp_irc_success); + + /* Assume a clock size is available and > 1 GHz; value is in kHz. */ + hip_err = hipDeviceGetAttribute (&ivar, hipDeviceAttributeClockRate, hip_dev); + assert (hip_err == hipSuccess); + assert (ivar > 1000000 /* kHz */); + + /* Assume that the MaxGridDimX is available and > 1024. */ + hip_err = hipDeviceGetAttribute (&ivar, hipDeviceAttributeMaxGridDimX, hip_dev); + assert (hip_err == hipSuccess); + assert (ivar > 1024); + + + /* Check whether the omp_ipr_device_context -> hipCtx_t yields a context. */ + + hipCtx_t hip_ctx = (hipCtx_t) omp_get_interop_ptr (obj, omp_ipr_device_context, &res); + assert (res == omp_irc_success); + + /* Assume API Version > 0 for Nvidia, hipErrorNotSupported for AMD. */ + ivar = -99; + #pragma GCC diagnostic push + #pragma GCC diagnostic ignored "-Wdeprecated-declarations" + hip_err = hipCtxGetApiVersion (hip_ctx, &ivar); + #pragma GCC diagnostic pop + + if (vendor_is_amd) + assert (hip_err == hipErrorNotSupported && ivar == -99); + else + { + assert (hip_err == hipSuccess); + assert (ivar > 0); + } + + + /* Check whether the omp_ipr_targetsync -> hipStream_t yields a stream. */ + + hipStream_t hip_sm = (hipStream_t) omp_get_interop_ptr (obj, omp_ipr_targetsync, &res); + assert (res == omp_irc_success); + + hipDevice_t dev_stream = 99; + hip_err = hipStreamGetDevice (hip_sm, &dev_stream); + assert (hip_err == hipSuccess); + assert (dev_stream == hip_dev); + + /* All jobs should have been completed (as there were none none) */ + hip_err = hipStreamQuery (hip_sm); + assert (hip_err == hipSuccess); + + #pragma omp interop destroy(obj) +} diff --git a/libgomp/testsuite/libgomp.c/interop-hipblas-amd-full.c b/libgomp/testsuite/libgomp.c/interop-hipblas-amd-full.c new file mode 100644 index 0000000..53c05bd --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hipblas-amd-full.c @@ -0,0 +1,7 @@ +/* { dg-require-effective-target gomp_hip_header_amd } */ +/* { dg-require-effective-target gomp_libhipblas } */ +/* { dg-additional-options "-lhipblas" } */ + +#define __HIP_PLATFORM_AMD__ 1 + +#include "interop-hipblas.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hipblas-amd-no-hip-header.c b/libgomp/testsuite/libgomp.c/interop-hipblas-amd-no-hip-header.c new file mode 100644 index 0000000..0ea3133 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hipblas-amd-no-hip-header.c @@ -0,0 +1,8 @@ +/* { dg-require-effective-target gomp_libhipblas } */ +/* { dg-additional-options "-lhipblas" } */ + +#define __HIP_PLATFORM_AMD__ 1 + +#define USE_HIP_FALLBACK_HEADER 1 + +#include "interop-hipblas.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-full.c b/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-full.c new file mode 100644 index 0000000..ed428c6 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-full.c @@ -0,0 +1,7 @@ +/* { dg-require-effective-target openacc_cublas } */ +/* { dg-require-effective-target gomp_hip_header_nvidia } */ +/* { dg-additional-options "-lcublas -Wno-deprecated-declarations" } */ + +#define __HIP_PLATFORM_NVIDIA__ 1 + +#include "interop-hipblas.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-no-headers.c b/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-no-headers.c new file mode 100644 index 0000000..1a31b30 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-no-headers.c @@ -0,0 +1,9 @@ +/* { dg-require-effective-target openacc_libcublas } */ +/* { dg-additional-options "-lcublas" } */ + +#define __HIP_PLATFORM_NVIDIA__ 1 + +#define USE_HIP_FALLBACK_HEADER 1 +#define USE_CUDA_FALLBACK_HEADER 1 + +#include "interop-hipblas.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-no-hip-header.c b/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-no-hip-header.c new file mode 100644 index 0000000..f85c13b --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hipblas-nvidia-no-hip-header.c @@ -0,0 +1,8 @@ +/* { dg-require-effective-target openacc_cublas } */ +/* { dg-additional-options "-lcublas" } */ + +#define __HIP_PLATFORM_NVIDIA__ 1 + +#define USE_HIP_FALLBACK_HEADER 1 + +#include "interop-hipblas.h" diff --git a/libgomp/testsuite/libgomp.c/interop-hipblas.h b/libgomp/testsuite/libgomp.c/interop-hipblas.h new file mode 100644 index 0000000..d7cb174 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hipblas.h @@ -0,0 +1,240 @@ +/* Check whether hipBlas' daxpy works with an interop object. + daxpy(N, DA, DX, INCX, DY, INCY) + calculates (for DX = DY = 1): + DY(1:N) = DY(1:N) + DA * DX(1:N) + and otherwise N array elements, taking every INCX-th or INCY-th one, repectively. + +Based on the interop example in OpenMP's example document */ + +/* Minimal check whether HIP works - by checking whether the API routines + seem to work. This includes a fallback if the header is not + available. */ + +#if !defined(__HIP_PLATFORM_AMD__) && !defined(__HIP_PLATFORM_NVIDIA__) + #error "Either __HIP_PLATFORM_AMD__ or __HIP_PLATFORM_NVIDIA__ must be defined" +#endif + +#if defined(__HIP_PLATFORM_AMD__) && defined(__HIP_PLATFORM_NVIDIA__) + #error "Either __HIP_PLATFORM_AMD__ or __HIP_PLATFORM_NVIDIA__ must be defined" +#endif + + +#include <assert.h> +#include <omp.h> +#include "../libgomp.c-c++-common/on_device_arch.h" + + +#if __has_include(<hipblas/hipblas.h>) && (__has_include(<library_types.h>) || !defined(__HIP_PLATFORM_NVIDIA__)) && !defined(USE_HIP_FALLBACK_HEADER) + #ifdef __HIP_PLATFORM_NVIDIA__ + /* There seems to be an issue with hip/library_types.h including + CUDA's "library_types.h". Include CUDA's one explicitly here. + Could possibly worked around by using -isystem vs. -I. */ + #include <library_types.h> + + /* For some reasons, the following symbols do not seem to get + mapped from HIP to CUDA, causing link errors. */ + #define hipblasSetStream cublasSetStream_v2 + #define hipblasDaxpy cublasDaxpy_v2 + #define hipblasCreate cublasCreate_v2 + #endif + #include <hipblas/hipblas.h> + +#elif defined(__HIP_PLATFORM_AMD__) + /* Add a poor man's fallback declaration. */ + #if !defined(USE_HIP_FALLBACK_HEADER) + #warning "Using fallback declaration for <hipblas/hipblas.h> for __HIP_PLATFORM_AMD__" + #endif + + typedef enum + { + HIPBLAS_STATUS_SUCCESS = 0 + + } hipblasStatus_t; + + typedef struct ihipStream_t* hipStream_t; + typedef void* hipblasHandle_t; + + hipblasStatus_t hipblasCreate (hipblasHandle_t*); + hipblasStatus_t hipblasSetStream (hipblasHandle_t, hipStream_t); + hipblasStatus_t hipblasDaxpy (hipblasHandle_t, int, const double*, const double*, int, double*, int); + +#else + /* Add a poor man's fallback declaration. */ + #if !defined(USE_HIP_FALLBACK_HEADER) + #warning "Using fallback declaration for <hipblas/hipblas.h> for __HIP_PLATFORM_NVIDA__" + #endif + + #if __has_include(<cuda.h>) && __has_include(<cudaTypedefs.h>) && __has_include(<cuda_runtime.h>) && __has_include(<cublas_v2.h>) && !defined(USE_CUDA_FALLBACK_HEADER) + #include <cuda.h> + #include <cudaTypedefs.h> + #include <cuda_runtime.h> + #include <cublas_v2.h> + + #else + /* Add a poor man's fallback declaration. */ + #if defined(USE_CUDA_FALLBACK_HEADER) + // no warning + #elif !__has_include(<cuda.h>) + #warning "Using GCC's cuda.h as fallback for cuda.h" + #elif !__has_include(<cudaTypedefs.h>) + #warning "Using GCC's cuda.h as fallback for cudaTypedefs.h" + #elif !__has_include(<cuda_runtime.h>) + #warning "Using GCC's cuda.h as fallback for cuda_runtime.h" + #else + #warning "Using GCC's cuda.h as fallback for cublas_v2.h" + #endif + #include "../../../include/cuda/cuda.h" + + typedef enum { + CUBLAS_STATUS_SUCCESS = 0, + } cublasStatus_t; + + typedef CUstream cudaStream_t; + typedef struct cublasContext* cublasHandle_t; + + #define cublasCreate cublasCreate_v2 + cublasStatus_t cublasCreate_v2 (cublasHandle_t *); + + #define cublasSetStream cublasSetStream_v2 + cublasStatus_t cublasSetStream_v2 (cublasHandle_t, cudaStream_t); + + #define cublasDaxpy cublasDaxpy_v2 + cublasStatus_t cublasDaxpy_v2(cublasHandle_t, int, const double*, const double*, int, double*, int); + #endif + + #define HIPBLAS_STATUS_SUCCESS CUBLAS_STATUS_SUCCESS + #define hipblasStatus_t cublasStatus_t + #define hipStream_t cudaStream_t + #define hipblasHandle_t cublasHandle_t + #define hipblasCreate cublasCreate + #define hipblasSetStream cublasSetStream + #define hipblasDaxpy cublasDaxpy +#endif + +static int used_variant = 0; + +void +run_hipBlasdaxpy (int n, double da, const double *dx, int incx, double *dy, int incy, omp_interop_t obj) +{ + used_variant = 1; + + omp_interop_rc_t res; + hipblasStatus_t stat; + + omp_intptr_t fr = omp_get_interop_int(obj, omp_ipr_fr_id, &res); + assert (res == omp_irc_success && fr == omp_ifr_hip); + + hipStream_t stream = (hipStream_t) omp_get_interop_ptr (obj, omp_ipr_targetsync, &res); + assert (res == omp_irc_success); + + hipblasHandle_t handle; + stat = hipblasCreate (&handle); + assert (stat == HIPBLAS_STATUS_SUCCESS); + + stat = hipblasSetStream (handle, stream); + assert (stat == HIPBLAS_STATUS_SUCCESS); + + /* 'da' can be in host or device space, 'dx' and 'dy' must be in device space. */ + stat = hipblasDaxpy (handle, n, &da, dx, 1, dy, 1) ; + assert (stat == HIPBLAS_STATUS_SUCCESS); +} + +#if defined(__HIP_PLATFORM_AMD__) +#pragma omp declare variant(run_hipBlasdaxpy) \ + match(construct={dispatch}, target_device={kind(nohost), arch("amdgcn")}) \ + adjust_args(need_device_ptr : dx, dy) \ + append_args(interop(targetsync, prefer_type("hip"))) +#elif defined(__HIP_PLATFORM_NVIDIA__) +#pragma omp declare variant(run_hipBlasdaxpy) \ + match(construct={dispatch}, target_device={kind(nohost), arch("nvptx")}) \ + adjust_args(need_device_ptr : dx, dy) \ + append_args(interop(targetsync, prefer_type("hip"))) +#else + #error "wrong platform" +#endif + +void +run_daxpy (int n, double da, const double *dx, int incx, double *dy, int incy) +{ + used_variant = 2; + + if (incx == 1 && incy == 1) + #pragma omp simd + for (int i = 0; i < n; i++) + dy[i] += da * dx[i]; + else + { + int ix = 0; + int iy = 0; + for (int i = 0; i < n; i++) + { + dy[iy] += da * dx[ix]; + ix += incx; + iy += incy; + } + } +} + + +void +run_test (int dev) +{ + constexpr int N = 1024; + + // A = {1,2,...,N} + // B = {-1, -2, ..., N} + // B' = daxpy (N, 3, A, incx=1, B, incy=1) + // = B + 3*A + // -> B' = {0, 2, 4, 6, ... } + + double A[N], B[N]; + double factor = 3.0; + for (int i = 0; i < N; i++) + { + A[i] = i; + B[i] = -i; + } + + if (dev != omp_initial_device && dev != omp_get_num_devices ()) + { + #pragma omp target enter data device(dev) map(A, B) + } + + used_variant = 99; + #pragma omp dispatch device(dev) + run_daxpy (N, factor, A, 1, B, 1); + + if (dev != omp_initial_device && dev != omp_get_num_devices ()) + { + #pragma omp target exit data device(dev) map(release: A) map(from: B) + + int tmp = omp_get_default_device (); + omp_set_default_device (dev); +#if defined(__HIP_PLATFORM_AMD__) + if (on_device_arch_gcn ()) +#else + if (on_device_arch_nvptx ()) +#endif + assert (used_variant == 1); + else + assert (used_variant == 2); + omp_set_default_device (tmp); + } + else + assert (used_variant == 2); + + for (int i = 0; i < N; i++) + assert (B[i] == 2*i); +} + +int +main () +{ + int ndev = omp_get_num_devices (); + + for (int dev = 0; dev <= ndev; dev++) + run_test (dev); + run_test (omp_initial_device); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c/interop-hsa.c b/libgomp/testsuite/libgomp.c/interop-hsa.c new file mode 100644 index 0000000..21ac91c --- /dev/null +++ b/libgomp/testsuite/libgomp.c/interop-hsa.c @@ -0,0 +1,205 @@ +/* { dg-additional-options "-ldl" } */ +/* { dg-require-effective-target offload_device_gcn } + The 'asm' insert is valid for GCN only: + { dg-additional-options -foffload=amdgcn-amdhsa } */ + +#include <stdio.h> +#include <stdlib.h> +#include <string.h> +#include <omp.h> +#include <assert.h> +#include <dlfcn.h> +#include "../../../include/hsa.h" +#include "../../config/gcn/libgomp-gcn.h" + +#define STACKSIZE (100 * 1024) +#define HEAPSIZE (10 * 1024 * 1024) +#define ARENASIZE HEAPSIZE + +/* This code fragment must be optimized or else the host-fallback kernel has + * invalid ASM inserts. The rest of the file can be compiled safely at -O0. */ +#pragma omp declare target +uintptr_t __attribute__((optimize("O1"))) +get_kernel_ptr () +{ + uintptr_t val; + if (!omp_is_initial_device ()) + /* "main._omp_fn.0" is the name GCC gives the first OpenMP target + * region in the "main" function. + * The ".kd" suffix is added by the LLVM assembler when it creates the + * kernel meta-data, and this is what we need to launch a kernel. */ + asm ("s_getpc_b64 %0\n\t" + "s_add_u32 %L0, %L0, main._omp_fn.0.kd@rel32@lo+4\n\t" + "s_addc_u32 %H0, %H0, main._omp_fn.0.kd@rel32@hi+4" + : "=Sg"(val)); + return val; +} +#pragma omp end declare target + +int +main(int argc, char** argv) +{ + + /* Load the HSA runtime DLL. */ + void *hsalib = dlopen ("libhsa-runtime64.so.1", RTLD_LAZY); + assert (hsalib); + + hsa_status_t (*hsa_signal_create) (hsa_signal_value_t initial_value, + uint32_t num_consumers, + const hsa_agent_t *consumers, + hsa_signal_t *signal) + = dlsym (hsalib, "hsa_signal_create"); + assert (hsa_signal_create); + + uint64_t (*hsa_queue_load_write_index_relaxed) (const hsa_queue_t *queue) + = dlsym (hsalib, "hsa_queue_load_write_index_relaxed"); + assert (hsa_queue_load_write_index_relaxed); + + void (*hsa_signal_store_relaxed) (hsa_signal_t signal, + hsa_signal_value_t value) + = dlsym (hsalib, "hsa_signal_store_relaxed"); + assert (hsa_signal_store_relaxed); + + hsa_signal_value_t (*hsa_signal_wait_relaxed) (hsa_signal_t signal, + hsa_signal_condition_t condition, + hsa_signal_value_t compare_value, + uint64_t timeout_hint, + hsa_wait_state_t wait_state_hint) + = dlsym (hsalib, "hsa_signal_wait_relaxed"); + assert (hsa_signal_wait_relaxed); + + void (*hsa_queue_store_write_index_relaxed) (const hsa_queue_t *queue, + uint64_t value) + = dlsym (hsalib, "hsa_queue_store_write_index_relaxed"); + assert (hsa_queue_store_write_index_relaxed); + + hsa_status_t (*hsa_signal_destroy) (hsa_signal_t signal) + = dlsym (hsalib, "hsa_signal_destroy"); + assert (hsa_signal_destroy); + + /* Set up the device data environment. */ + int test_data_value = 0; +#pragma omp target enter data map(test_data_value) + + /* Get the interop details. */ + int device_num = omp_get_default_device(); + hsa_agent_t *gpu_agent; + hsa_queue_t *hsa_queue = NULL; + + omp_interop_t interop = omp_interop_none; +#pragma omp interop init(target, targetsync, prefer_type("hsa"): interop) device(device_num) + assert (interop != omp_interop_none); + + omp_interop_rc_t retcode; + omp_interop_fr_t fr = omp_get_interop_int (interop, omp_ipr_fr_id, &retcode); + assert (retcode == omp_irc_success); + assert (fr == omp_ifr_hsa); + + gpu_agent = omp_get_interop_ptr(interop, omp_ipr_device, &retcode); + assert (retcode == omp_irc_success); + + hsa_queue = omp_get_interop_ptr(interop, omp_ipr_targetsync, &retcode); + assert (retcode == omp_irc_success); + assert (hsa_queue); + + /* Call an offload kernel via OpenMP/libgomp. + * + * This kernel serves two purposes: + * 1) Lookup the device-side load-address of itself (thus avoiding the + * need to access the libgomp internals). + * 2) Count how many times it is called. + * We then call it once using OpenMP, and once manually, and check + * the counter reads "2". */ + uint64_t kernel_object = 0; +#pragma omp target map(from:kernel_object) map(present,alloc:test_data_value) + { + kernel_object = get_kernel_ptr (); + ++test_data_value; + } + + assert (kernel_object != 0); + + /* Configure the same kernel to run again, using HSA manually this time. */ + hsa_status_t status; + hsa_signal_t signal; + status = hsa_signal_create(1, 0, NULL, &signal); + assert (status == HSA_STATUS_SUCCESS); + + /* The kernel is built by GCC for OpenMP, so we need to pass the same + * data pointers that libgomp would pass in. */ + struct { + uintptr_t test_data_value; + uintptr_t kernel_object; + } tgtaddrs; + +#pragma omp target data use_device_addr(test_data_value) + { + tgtaddrs.test_data_value = (uintptr_t)&test_data_value; + tgtaddrs.kernel_object = (uintptr_t)omp_target_alloc (8, device_num); + } + + /* We also need to duplicate the launch ABI used by plugin-gcn.c. */ + struct kernargs_abi args; /* From libgomp-gcn.h. */ + args.dummy1 = (int64_t)&tgtaddrs; + args.out_ptr = (int64_t)malloc (sizeof (struct output)); /* Host side. */ + args.heap_ptr = (int64_t)omp_target_alloc (HEAPSIZE, device_num); + args.arena_ptr = (int64_t)omp_target_alloc (ARENASIZE, device_num); + args.stack_ptr = (int64_t)omp_target_alloc (STACKSIZE, device_num); + args.arena_size_per_team = ARENASIZE; + args.stack_size_per_thread = STACKSIZE; + + /* Build the HSA dispatch packet, and insert it into the queue. */ + uint64_t packet_id = hsa_queue_load_write_index_relaxed (hsa_queue); + const uint32_t queueMask = hsa_queue->size - 1; + hsa_kernel_dispatch_packet_t *dispatch_packet = + &(((hsa_kernel_dispatch_packet_t *) + (hsa_queue->base_address))[packet_id & queueMask]); + + dispatch_packet->setup = 3 << HSA_KERNEL_DISPATCH_PACKET_SETUP_DIMENSIONS; + dispatch_packet->workgroup_size_x = 1; + dispatch_packet->workgroup_size_y = 64; + dispatch_packet->workgroup_size_z = 1; + dispatch_packet->grid_size_x = 1; + dispatch_packet->grid_size_y = 64; + dispatch_packet->grid_size_z = 1; + dispatch_packet->completion_signal = signal; + dispatch_packet->kernel_object = kernel_object; + dispatch_packet->kernarg_address = &args; + dispatch_packet->private_segment_size = 0; + dispatch_packet->group_segment_size = 1536; + + uint16_t header = 0; + header |= HSA_PACKET_TYPE_KERNEL_DISPATCH << HSA_PACKET_HEADER_TYPE; + header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_ACQUIRE_FENCE_SCOPE; + header |= HSA_FENCE_SCOPE_SYSTEM << HSA_PACKET_HEADER_RELEASE_FENCE_SCOPE; + + /* Finish writing the packet header with an atomic release. */ + __atomic_store_n((uint16_t*)dispatch_packet, header, __ATOMIC_RELEASE); + + hsa_queue_store_write_index_relaxed (hsa_queue, packet_id + 1); + + ;/* Run the kernel and wait for it to complete. */ + hsa_signal_store_relaxed(hsa_queue->doorbell_signal, packet_id); + while (hsa_signal_wait_relaxed(signal, HSA_SIGNAL_CONDITION_LT, 1, + UINT64_MAX, HSA_WAIT_STATE_ACTIVE) != 0) + ; + + /* Clean up HSA. */ + hsa_signal_destroy(signal); + free ((void*)args.out_ptr); + omp_target_free ((void*)args.heap_ptr, device_num); + omp_target_free ((void*)args.arena_ptr, device_num); + omp_target_free ((void*)args.stack_ptr, device_num); + omp_target_free ((void*)tgtaddrs.kernel_object, device_num); + + /* Clean up OpenMP. */ + #pragma omp interop destroy(interop) + + /* Bring the data back from the device. */ +#pragma omp target exit data map(test_data_value) + + /* Ensure the kernel was called twice. Once by OpenMP, once by HSA. */ + assert (test_data_value == 2); + + return 0; +} diff --git a/libgomp/testsuite/libgomp.c/target-map-zero-sized-2.c b/libgomp/testsuite/libgomp.c/target-map-zero-sized-2.c new file mode 100644 index 0000000..3220828 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/target-map-zero-sized-2.c @@ -0,0 +1,74 @@ +int +main () +{ + int i, n; + int data[] = {1,2}; + struct S { int **ptrset; }; + +// ----------------------------------- + +/* The produced mapping for sptr1->ptrset[i][:n] + + GOMP_MAP_STRUCT (size = 1) + GOMP_MAP_ZERO_LEN_ARRAY_SECTION + GOMP_MAP_ZERO_LEN_ARRAY_SECTION + GOMP_MAP_ATTACH + GOMP_MAP_ATTACH -> attaching to 2nd GOMP_MAP_ZERO_LEN_ARRAY_SECTION + +which get split into 3 separate map_vars call; in particular, +the latter is separate and points to an unmpapped variable. + +Thus, it failed with: + libgomp: pointer target not mapped for attach */ + + struct S s1, *sptr1; + s1.ptrset = (int **) __builtin_malloc (sizeof(void*) * 3); + s1.ptrset[0] = data; + s1.ptrset[1] = data; + s1.ptrset[2] = data; + sptr1 = &s1; + + i = 1; + n = 0; + #pragma omp target enter data map(sptr1[:1], sptr1->ptrset[:3]) + #pragma omp target enter data map(sptr1->ptrset[i][:n]) + + #pragma omp target exit data map(sptr1->ptrset[i][:n]) + #pragma omp target exit data map(sptr1[:1], sptr1->ptrset[:3]) + + __builtin_free (s1.ptrset); + +// ----------------------------------- + +/* The produced mapping for sptr2->ptrset[i][:n] is similar: + + GOMP_MAP_STRUCT (size = 1) + GOMP_MAP_ZERO_LEN_ARRAY_SECTION + GOMP_MAP_TO ! this one has now a finite size + GOMP_MAP_ATTACH + GOMP_MAP_ATTACH -> attach to the GOMP_MAP_TO + +As the latter GOMP_MAP_ATTACH has now a pointer target, +the attachment worked. */ + + struct S s2, *sptr2; + s2.ptrset = (int **) __builtin_malloc (sizeof(void*) * 3); + s2.ptrset[0] = data; + s2.ptrset[1] = data; + s2.ptrset[2] = data; + sptr2 = &s2; + + i = 1; + n = 2; + #pragma omp target enter data map(sptr2[:1], sptr2->ptrset[:3]) + #pragma omp target enter data map(sptr2->ptrset[i][:n]) + + #pragma omp target + if (sptr2->ptrset[1][0] != 1 || sptr2->ptrset[1][1] != 2) + __builtin_abort (); + + #pragma omp target exit data map(sptr2->ptrset[i][:n]) + #pragma omp target exit data map(sptr2[:1], sptr2->ptrset[:3]) + + __builtin_free (s2.ptrset); +} diff --git a/libgomp/testsuite/libgomp.c/target-map-zero-sized-3.c b/libgomp/testsuite/libgomp.c/target-map-zero-sized-3.c new file mode 100644 index 0000000..580c6ad --- /dev/null +++ b/libgomp/testsuite/libgomp.c/target-map-zero-sized-3.c @@ -0,0 +1,50 @@ +int +main () +{ + int i, n; + int data[] = {1,2}; + struct S { + int **ptrset; + int **ptrset2; + }; + + /* This is the same as target-map-zero-sized-3.c, but by mixing + mapped and non-mapped items, the mapping before the ATTACH + might (or here: is) not actually associated with the the + pointer used for attaching. Thus, if one does a simple + + if (openmp_p + && (pragma_kind & GOMP_MAP_VARS_ENTER_DATA) + && mapnum == 1) + check in target.c's gomp_map_vars_internal will fail + as mapnum > 1 but still the map associated with this + ATTACH is in a different set. */ + + struct S s1, *sptr1; + s1.ptrset = (int **) __builtin_malloc (sizeof(void*) * 3); + s1.ptrset2 = (int **) __builtin_malloc (sizeof(void*) * 3); + s1.ptrset[0] = data; + s1.ptrset[1] = data; + s1.ptrset[2] = data; + s1.ptrset2[0] = data; + s1.ptrset2[1] = data; + s1.ptrset2[2] = data; + sptr1 = &s1; + + i = 1; + n = 0; + #pragma omp target enter data map(data) + #pragma omp target enter data map(sptr1[:1], sptr1->ptrset[:3], sptr1->ptrset2[:3]) + #pragma omp target enter data map(sptr1->ptrset[i][:n], sptr1->ptrset2[i][:n]) + + #pragma omp target map(sptr1->ptrset[i][:n], sptr1->ptrset2[i][:n]) + if (sptr1->ptrset2[1][0] != 1 || sptr1->ptrset2[1][1] != 2) + __builtin_abort (); + + #pragma omp target exit data map(sptr1->ptrset[i][:n], sptr1->ptrset2[i][:n]) + #pragma omp target exit data map(sptr1[:1], sptr1->ptrset[:3], sptr1->ptrset2[:3]) + #pragma omp target exit data map(data) + + __builtin_free (s1.ptrset); + __builtin_free (s1.ptrset2); +} diff --git a/libgomp/testsuite/libgomp.c/target-map-zero-sized.c b/libgomp/testsuite/libgomp.c/target-map-zero-sized.c new file mode 100644 index 0000000..7c4ab80 --- /dev/null +++ b/libgomp/testsuite/libgomp.c/target-map-zero-sized.c @@ -0,0 +1,107 @@ +/* { dg-do run } */ +/* { dg-additional-options "-O0" } */ + +/* Issue showed up in the real world when large data was distributed + over multiple MPI progresses - such that for one process n == 0 + happend at run time. + + Before map(var[:0]) and map(var[:n]) with n > 0 was handled, + this patch now also handles map(var[:n]) with n == 0. + + Failed before with "libgomp: pointer target not mapped for attach". */ + +/* Here, the base address is shifted - which should have no effect, + but must work as well. */ +void +with_offset () +{ + struct S { + int *ptr1, *ptr2; + }; + struct S s1, s2; + int *a, *b, *c, *d; + s1.ptr1 = (int *) 0L; + s1.ptr2 = (int *) 0xdeedbeef; + s2.ptr1 = (int *) 0L; + s2.ptr2 = (int *) 0xdeedbeef; + a = (int *) 0L; + b = (int *) 0xdeedbeef; + c = (int *) 0L; + d = (int *) 0xdeedbeef; + + int n1, n2, n3, n4; + n1 = n2 = n3 = n4 = 0; + + #pragma omp target enter data map(s1.ptr1[4:n1], s1.ptr2[6:n2], a[3:n3], b[2:n4]) + + #pragma omp target map(s2.ptr1[4:n1], s2.ptr2[2:n2], c[6:n3], d[9:n4]) + { + if (s2.ptr1 != (void *) 0L || s2.ptr2 != (void *) 0xdeedbeef + || c != (void *) 0L || d != (void *) 0xdeedbeef) + __builtin_abort (); + } + + #pragma omp target map(s1.ptr1[4:n1], s1.ptr2[6:n2], a[3:n3], b[2:n4]) + { + if (s1.ptr1 != (void *) 0L || s1.ptr2 != (void *) 0xdeedbeef + || a != (void *) 0L || b != (void *) 0xdeedbeef) + __builtin_abort (); + } + + #pragma omp target + { + if (s1.ptr1 != (void *) 0L || s1.ptr2 != (void *) 0xdeedbeef + || a != (void *) 0L || b != (void *) 0xdeedbeef) + __builtin_abort (); + } + + #pragma omp target exit data map(s1.ptr1[4:n1], s1.ptr2[6:n2], a[3:n3], b[2:n4]) +} + +int +main () +{ + struct S { + int *ptr1, *ptr2; + }; + struct S s1, s2; + int *a, *b, *c, *d; + s1.ptr1 = (int *) 0L; + s1.ptr2 = (int *) 0xdeedbeef; + s2.ptr1 = (int *) 0L; + s2.ptr2 = (int *) 0xdeedbeef; + a = (int *) 0L; + b = (int *) 0xdeedbeef; + c = (int *) 0L; + d = (int *) 0xdeedbeef; + + int n1, n2, n3, n4; + n1 = n2 = n3 = n4 = 0; + + #pragma omp target enter data map(s1.ptr1[:n1], s1.ptr2[:n2], a[:n3], b[:n4]) + + #pragma omp target map(s2.ptr1[:n1], s2.ptr2[:n2], c[:n3], d[:n4]) + { + if (s2.ptr1 != (void *) 0L || s2.ptr2 != (void *) 0xdeedbeef + || c != (void *) 0L || d != (void *) 0xdeedbeef) + __builtin_abort (); + } + + #pragma omp target map(s1.ptr1[:n1], s1.ptr2[:n2], a[:n3], b[:n4]) + { + if (s1.ptr1 != (void *) 0L || s1.ptr2 != (void *) 0xdeedbeef + || a != (void *) 0L || b != (void *) 0xdeedbeef) + __builtin_abort (); + } + + #pragma omp target + { + if (s1.ptr1 != (void *) 0L || s1.ptr2 != (void *) 0xdeedbeef + || a != (void *) 0L || b != (void *) 0xdeedbeef) + __builtin_abort (); + } + + #pragma omp target exit data map(s1.ptr1[:n1], s1.ptr2[:n2], a[:n3], b[:n4]) + + with_offset (); +} diff --git a/libgomp/testsuite/libgomp.fortran/alloc-comp-4.f90 b/libgomp/testsuite/libgomp.fortran/alloc-comp-4.f90 new file mode 100644 index 0000000..d5e982b --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/alloc-comp-4.f90 @@ -0,0 +1,75 @@ +! +! Check that mapping with map(var%tiles(1)) works. +! +! This uses deep mapping to handle the allocatable +! derived-type components +! +! The tricky part is that GCC generates intermittently +! an SSA_NAME that needs to be resolved. +! +module m +type t + integer, allocatable :: den1(:,:), den2(:,:) +end type t + +type t2 + type(t), allocatable :: tiles(:) +end type t2 +end + +use m +use iso_c_binding +implicit none (type, external) +type(t2), target :: var +logical :: is_self_map +type(C_ptr) :: pden1, pden2, ptiles, ptiles1 + +allocate(var%tiles(1)) +var%tiles(1)%den1 = reshape([1,2,3,4],[2,2]) +var%tiles(1)%den2 = reshape([11,22,33,44],[2,2]) + +ptiles = c_loc(var%tiles) +ptiles1 = c_loc(var%tiles(1)) +pden1 = c_loc(var%tiles(1)%den1) +pden2 = c_loc(var%tiles(1)%den2) + + +is_self_map = .false. +!$omp target map(to: is_self_map) + is_self_map = .true. +!$omp end target + +!$omp target enter data map(var%tiles(1)) + +!$omp target firstprivate(ptiles, ptiles1, pden1, pden2) + if (any (var%tiles(1)%den1 /= reshape([1,2,3,4],[2,2]))) stop 1 + if (any (var%tiles(1)%den2 /= reshape([11,22,33,44],[2,2]))) stop 2 + var%tiles(1)%den1 = var%tiles(1)%den1 + 5 + var%tiles(1)%den2 = var%tiles(1)%den2 + 7 + + if (is_self_map) then + if (.not. c_associated (ptiles, c_loc(var%tiles))) stop 3 + if (.not. c_associated (ptiles1, c_loc(var%tiles(1)))) stop 4 + if (.not. c_associated (pden1, c_loc(var%tiles(1)%den1))) stop 5 + if (.not. c_associated (pden2, c_loc(var%tiles(1)%den2))) stop 6 + else + if (c_associated (ptiles, c_loc(var%tiles))) stop 3 + if (c_associated (ptiles1, c_loc(var%tiles(1)))) stop 4 + if (c_associated (pden1, c_loc(var%tiles(1)%den1))) stop 5 + if (c_associated (pden2, c_loc(var%tiles(1)%den2))) stop 6 + endif +!$omp end target + +if (is_self_map) then + if (any (var%tiles(1)%den1 /= 5 + reshape([1,2,3,4],[2,2]))) stop 7 + if (any (var%tiles(1)%den2 /= 7 + reshape([11,22,33,44],[2,2]))) stop 8 +else + if (any (var%tiles(1)%den1 /= reshape([1,2,3,4],[2,2]))) stop 7 + if (any (var%tiles(1)%den2 /= reshape([11,22,33,44],[2,2]))) stop 8 +endif + +!$omp target exit data map(var%tiles(1)) + +if (any (var%tiles(1)%den1 /= 5 + reshape([1,2,3,4],[2,2]))) stop 7 +if (any (var%tiles(1)%den2 /= 7 + reshape([11,22,33,44],[2,2]))) stop 8 +end diff --git a/libgomp/testsuite/libgomp.fortran/allocate-8a.f90 b/libgomp/testsuite/libgomp.fortran/allocate-8a.f90 new file mode 100644 index 0000000..5f6c8c1 --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/allocate-8a.f90 @@ -0,0 +1,45 @@ +! { dg-additional-options "-fopenmp-allocators" } +! { dg-additional-options "-fdump-tree-omplower" } +program main + use iso_c_binding + use omp_lib + implicit none (type, external) + integer(omp_allocator_handle_kind):: alloc_h + integer :: i, N + integer(c_intptr_t) :: intptr + integer, allocatable :: A(:) + type(omp_alloctrait):: traits(1) = [omp_alloctrait(omp_atk_alignment, 128)] + + N = 10 + alloc_h = omp_init_allocator(omp_default_mem_space, 1, traits) + + !$omp allocate(A) allocator(alloc_h) + allocate(A(N)) + a(:) = [(i, i=1,N)] + if (mod (transfer (loc(a), intptr),128) /= 0) & + stop 1 + if (any (a /= [(i, i=1,N)])) & + stop 2 + deallocate(A) + !$omp allocate(A) allocator(alloc_h) align(512) + allocate(A(N)) + block + integer, allocatable :: B(:) + !$omp allocators allocate(allocator(alloc_h), align(256) : B) + allocate(B(N)) + B(:) = [(2*i, i=1,N)] + A(:) = B + if (mod (transfer (loc(B), intptr), 256) /= 0) & + stop 1 + ! end of scope deallocation + end block + if (mod (transfer (loc(a), intptr),512) /= 0) & + stop 1 + if (any (a /= [(2*i, i=1,N)])) & + stop 2 + deallocate(A) ! Must deallocate here - before deallocator is destroyed + call omp_destroy_allocator(alloc_h) + ! No auto dealloc of A because it is SAVE +end +! { dg-final { scan-tree-dump-times "__builtin_GOMP_alloc \\(" 3 "omplower" } } +! { dg-final { scan-tree-dump-times "__builtin_GOMP_free \\(" 3 "omplower" } } diff --git a/libgomp/testsuite/libgomp.fortran/interop-hip-amd-full.F90 b/libgomp/testsuite/libgomp.fortran/interop-hip-amd-full.F90 new file mode 100644 index 0000000..eb2f437 --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/interop-hip-amd-full.F90 @@ -0,0 +1,10 @@ +! { dg-do run { target { offload_device_gcn } } } +! { dg-do link { target { ! offload_device_gcn } } } + +! { dg-require-effective-target gomp_hipfort_module } +! { dg-require-effective-target gomp_libamdhip64 } +! { dg-additional-options "-lamdhip64" } + +#define HAVE_HIPFORT 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.fortran/interop-hip-amd-no-module.F90 b/libgomp/testsuite/libgomp.fortran/interop-hip-amd-no-module.F90 new file mode 100644 index 0000000..0ebbe80 --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/interop-hip-amd-no-module.F90 @@ -0,0 +1,9 @@ +! { dg-do run { target { offload_device_gcn } } } +! { dg-do link { target { ! offload_device_gcn } } } + +! { dg-require-effective-target gomp_libamdhip64 } +! { dg-additional-options "-lamdhip64" } + +#define USE_HIP_FALLBACK_MODULE 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.fortran/interop-hip-nvidia-full.F90 b/libgomp/testsuite/libgomp.fortran/interop-hip-nvidia-full.F90 new file mode 100644 index 0000000..d29a689 --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/interop-hip-nvidia-full.F90 @@ -0,0 +1,12 @@ +! { dg-do run { target { offload_device_nvptx } } } +! { dg-do link { target { ! offload_device_nvptx } } } + +! { dg-require-effective-target gomp_hipfort_module } +! { dg-require-effective-target openacc_cudart } +! { dg-require-effective-target openacc_cuda } +! { dg-additional-options "-lcuda -lcudart" } + +#define HAVE_HIPFORT 1 +#define USE_CUDA_NAMES 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.fortran/interop-hip-nvidia-no-module.F90 b/libgomp/testsuite/libgomp.fortran/interop-hip-nvidia-no-module.F90 new file mode 100644 index 0000000..2063610 --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/interop-hip-nvidia-no-module.F90 @@ -0,0 +1,11 @@ +! { dg-do run { target { offload_device_nvptx } } } +! { dg-do link { target { ! offload_device_nvptx } } } + +! { dg-require-effective-target openacc_libcudart } +! { dg-require-effective-target openacc_libcuda } +! { dg-additional-options "-lcuda -lcudart" } + +#define USE_CUDA_NAMES 1 +#define USE_HIP_FALLBACK_MODULE 1 + +#include "interop-hip.h" diff --git a/libgomp/testsuite/libgomp.fortran/interop-hip.h b/libgomp/testsuite/libgomp.fortran/interop-hip.h new file mode 100644 index 0000000..753ccce --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/interop-hip.h @@ -0,0 +1,214 @@ +! Minimal check whether HIP works - by checking whether the API routines +! seem to work. This includes a fallback if hipfort is not available + +#ifndef HAVE_HIPFORT +#ifndef USE_HIP_FALLBACK_MODULE +#if USE_CUDA_NAMES +#warning "Using fallback implementation for module hipfort as HAVE_HIPFORT is undefined (for NVIDA/CUDA)" +#else +#warning "Using fallback implementation for module hipfort as HAVE_HIPFORT is undefined - assume AMD as USE_CUDA_NAMES is unset" +#endif +#endif +module hipfort ! Minimal implementation for the testsuite + implicit none + + enum, bind(c) + enumerator :: hipSuccess = 0 + enumerator :: hipErrorNotSupported = 801 + end enum + + enum, bind(c) + enumerator :: hipDeviceAttributeClockRate = 5 + enumerator :: hipDeviceAttributeMaxGridDimX = 29 + end enum + + interface + integer(kind(hipSuccess)) function hipDeviceGetAttribute (ip, attr, dev) & +#if USE_CUDA_NAMES + bind(c, name="cudaDeviceGetAttribute") +#else + bind(c, name="hipDeviceGetAttribute") +#endif + use iso_c_binding, only: c_ptr, c_int + import + implicit none + type(c_ptr), value :: ip + integer(kind(hipDeviceAttributeClockRate)), value :: attr + integer(c_int), value :: dev + end + + integer(kind(hipSuccess)) function hipCtxGetApiVersion (ctx, ip) & +#if USE_CUDA_NAMES + bind(c, name="cudaCtxGetApiVersion") +#else + bind(c, name="hipCtxGetApiVersion") +#endif + use iso_c_binding, only: c_ptr + import + implicit none + type(c_ptr), value :: ctx, ip + end + + integer(kind(hipSuccess)) function hipStreamQuery (stream) & +#if USE_CUDA_NAMES + bind(c, name="cudaStreamQuery") +#else + bind(c, name="hipStreamQuery") +#endif + use iso_c_binding, only: c_ptr + import + implicit none + type(c_ptr), value :: stream + end + + integer(kind(hipSuccess)) function hipStreamGetFlags (stream, flags) & +#if USE_CUDA_NAMES + bind(c, name="cudaStreamGetFlags") +#else + bind(c, name="hipStreamGetFlags") +#endif + use iso_c_binding, only: c_ptr + import + implicit none + type(c_ptr), value :: stream + type(c_ptr), value :: flags + end + end interface +end module +#endif + +program main + use iso_c_binding, only: c_ptr, c_int, c_loc + use omp_lib + use hipfort + implicit none (type, external) + +! Only supported since CUDA 12.8 - skip for better compatibility +! ! Manally implement hipStreamGetDevice as hipfort misses it +! ! -> https://github.com/ROCm/hipfort/issues/238 +! interface +! integer(kind(hipSuccess)) function my_hipStreamGetDevice(stream, dev) & +!#if USE_CUDA_NAMES +! bind(c, name="cudaStreamGetDevice") +!#else +! bind(c, name="hipStreamGetDevice") +!#endif +! use iso_c_binding, only: c_ptr, c_int +! import +! implicit none +! type(c_ptr), value :: stream +! integer(c_int) :: dev +! end +! end interface + + integer(c_int), target :: ivar + integer(omp_interop_rc_kind) :: res + integer(omp_interop_kind) :: obj + integer(omp_interop_fr_kind) :: fr + integer(kind(hipSuccess)) :: hip_err + integer(c_int) :: hip_dev, dev_stream + type(c_ptr) :: hip_ctx, hip_sm + + logical :: vendor_is_amd + + obj = omp_interop_none + + !$omp interop init(target, targetsync, prefer_type("hip") : obj) + + fr = omp_get_interop_int (obj, omp_ipr_fr_id, res) + if (res /= omp_irc_success) error stop 1 + if (fr /= omp_ifr_hip) error stop 1 + + ivar = omp_get_interop_int (obj, omp_ipr_vendor, res) + if (ivar == 1) then ! AMD + vendor_is_amd = .true. + else if (ivar == 11) then ! Nvidia + vendor_is_amd = .false. + else + error stop 1 ! Unknown + endif +#if USE_CUDA_NAMES + if (vendor_is_amd) error stop 1 +#else + if (.not. vendor_is_amd) error stop 1 +#endif + + ! Check whether the omp_ipr_device -> hipDevice_t yields a valid device. + + hip_dev = omp_get_interop_int (obj, omp_ipr_device, res) + if (res /= omp_irc_success) error stop 1 + +! AMD messed up in Fortran with the attribute handling, missing the +! translation table it has for C. +block + enum, bind(c) + enumerator :: cudaDevAttrClockRate = 13 + enumerator :: cudaDevAttrMaxGridDimX = 5 + end enum + + ! Assume a clock size is available and > 1 GHz; value is in kHz. + ! c_loc is completely bogus, but as AMD messed up the interface ... + ! Cf. https://github.com/ROCm/hipfort/issues/239 +if (vendor_is_amd) then + hip_err = hipDeviceGetAttribute (c_loc(ivar), hipDeviceAttributeClockRate, hip_dev) +else + hip_err = hipDeviceGetAttribute (c_loc(ivar), cudaDevAttrClockRate, hip_dev) +endif + if (hip_err /= hipSuccess) error stop 1 + if (ivar <= 1000000) error stop 1 ! in kHz + + ! Assume that the MaxGridDimX is available and > 1024 + ! c_loc is completely bogus, but as AMD messed up the interface ... + ! Cf. https://github.com/ROCm/hipfort/issues/239 +if (vendor_is_amd) then + hip_err = hipDeviceGetAttribute (c_loc(ivar), hipDeviceAttributeMaxGridDimX, hip_dev) +else + hip_err = hipDeviceGetAttribute (c_loc(ivar), cudaDevAttrMaxGridDimX, hip_dev) +endif + if (hip_err /= hipSuccess) error stop 1 + if (ivar <= 1024) error stop 1 +end block + + + ! Check whether the omp_ipr_device_context -> hipCtx_t yields a context. + + hip_ctx = omp_get_interop_ptr (obj, omp_ipr_device_context, res) + if (res /= omp_irc_success) error stop 1 + +! ! Assume API Version > 0 for Nvidia, hipErrorNotSupported for AMD. */ +! ivar = -99 +! ! AMD deprectated hipCtxGetApiVersion (in C/C++) +! hip_err = hipCtxGetApiVersion (hip_ctx, c_loc(ivar)) +! +! if (vendor_is_amd) then +! if (hip_err /= hipErrorNotSupported .or. ivar /= -99) error stop 1 +! else +! if (hip_err /= hipSuccess) error stop 1 +! if (ivar <= 0) error stop 1 +! end if + + + ! Check whether the omp_ipr_targetsync -> hipStream_t yields a stream. + + hip_sm = omp_get_interop_ptr (obj, omp_ipr_targetsync, res) + if (res /= omp_irc_success) error stop 1 + +! Skip as this is only in CUDA 12.8 +! dev_stream = 99 +! ! Not (yet) implemented: https://github.com/ROCm/hipfort/issues/238 +! ! hip_err = hipStreamGetDevice (hip_sm, dev_stream) +! hip_err = my_hipStreamGetDevice (hip_sm, dev_stream) +! if (hip_err /= hipSuccess) error stop 1 +! if (dev_stream /= hip_dev) error stop 1 + + ! Get flags of the stream + hip_err = hipStreamGetFlags (hip_sm, c_loc (ivar)) + if (hip_err /= hipSuccess) error stop 1 + ! Accept any value + + ! All jobs should have been completed (as there were none none) + hip_err = hipStreamQuery (hip_sm) + if (hip_err /= hipSuccess) error stop 1 + + !$omp interop destroy(obj) +end diff --git a/libgomp/testsuite/libgomp.fortran/map-alloc-comp-9-usm.f90 b/libgomp/testsuite/libgomp.fortran/map-alloc-comp-9-usm.f90 new file mode 100644 index 0000000..90378c0 --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/map-alloc-comp-9-usm.f90 @@ -0,0 +1,11 @@ +! { dg-additional-options "-cpp -DUSE_USM_REQUIREMENT=1 -Wno-openmp" } +! +! We silence the warning: +! Mapping of polymorphic list item '...' is unspecified behavior [-Wopenmp] +! +! Ensure that polymorphic mapping is diagnosed as undefined behavior +! Ensure that static access to polymorphic variables works + +! Run map-alloc-comp-9.f90 in unified-shared-memory mode + +#include "map-alloc-comp-9.f90" diff --git a/libgomp/testsuite/libgomp.fortran/map-alloc-comp-9.f90 b/libgomp/testsuite/libgomp.fortran/map-alloc-comp-9.f90 index 3cec392..26c73d7 100644 --- a/libgomp/testsuite/libgomp.fortran/map-alloc-comp-9.f90 +++ b/libgomp/testsuite/libgomp.fortran/map-alloc-comp-9.f90 @@ -1,8 +1,19 @@ +! { dg-additional-options "-cpp" } +! ! Ensure that polymorphic mapping is diagnosed as undefined behavior ! Ensure that static access to polymorphic variables works +! Some extended tests are only run with shared memory +! To enforce this (where possible) on the device side: +! #define USE_USM_REQUIREMENT +! which is done in map-alloc-comp-9-usm.f90 + subroutine test(case) implicit none(type, external) +#ifdef USE_USM_REQUIREMENT + !$omp requires unified_shared_memory +#endif + type t integer :: x(4) end type t @@ -73,10 +84,14 @@ var4%y2(2)%y%x%x = -7 * [1111,2222,3333,4444] var4%y2(2)%y%x2(1)%x = -8 * [1111,2222,3333,4444] var4%y2(2)%y%x2(2)%x = -9 * [1111,2222,3333,4444] +#ifdef USE_USM_REQUIREMENT +is_shared_mem = .true. +#else is_shared_mem = .false. !$omp target map(to: is_shared_mem) is_shared_mem = .true. !$omp end target +#endif if (case == 1) then ! implicit mapping @@ -532,6 +547,10 @@ end subroutine test program main use omp_lib implicit none(type, external) +#ifdef USE_USM_REQUIREMENT + !$omp requires unified_shared_memory +#endif + interface subroutine test(case) integer, value :: case diff --git a/libgomp/testsuite/libgomp.fortran/metadirective-1.f90 b/libgomp/testsuite/libgomp.fortran/metadirective-1.f90 index 7b3e09f..d6f4d5b 100644 --- a/libgomp/testsuite/libgomp.fortran/metadirective-1.f90 +++ b/libgomp/testsuite/libgomp.fortran/metadirective-1.f90 @@ -1,4 +1,5 @@ -! { dg-do run } +! { dg-do run { target { ! offload_target_nvptx } } } +! { dg-do compile { target offload_target_nvptx } } program test implicit none @@ -33,6 +34,10 @@ program test contains subroutine f (x, y, z) integer :: x(N), y(N), z(N) + ! The following fails as on the host the target side cannot be + ! resolved - and the 'teams' or not status affects how 'target' + ! is called. -> See PR118694, esp. comment 9. + ! Note also the dg-do compile above for offload_target_nvptx !$omp target map (to: x, y) map(from: z) block @@ -43,6 +48,7 @@ contains z(i) = x(i) * y(i) enddo end block + ! { dg-bogus "'target' construct with nested 'teams' construct contains directives outside of the 'teams' construct" "PR118694" { xfail offload_target_nvptx } .-9 } */ end subroutine subroutine g (x, y, z) integer :: x(N), y(N), z(N) @@ -56,6 +62,7 @@ contains z(i) = x(i) * y(i) enddo end block + ! { dg-bogus "'target' construct with nested 'teams' construct contains directives outside of the 'teams' construct" "PR118694" { xfail offload_target_nvptx } .-9 } */ !$omp end target end subroutine end program diff --git a/libgomp/testsuite/libgomp.fortran/omp_target_memset-2.f90 b/libgomp/testsuite/libgomp.fortran/omp_target_memset-2.f90 new file mode 100644 index 0000000..2641086 --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/omp_target_memset-2.f90 @@ -0,0 +1,67 @@ +! PR libgomp/120444 +! Async version + +use omp_lib +use iso_c_binding +implicit none (type, external) +integer(c_int) :: dev + +!$omp parallel do +do dev = omp_initial_device, omp_get_num_devices () +block + integer(c_int) :: i, val, start, tail + type(c_ptr) :: ptr, ptr2, tmpptr + integer(c_int8_t), pointer, contiguous :: fptr(:) + integer(c_intptr_t) :: intptr + integer(c_size_t), parameter :: count = 1024 + integer(omp_depend_kind) :: dep(1) + + ptr = omp_target_alloc (count, dev) + + !$omp depobj(dep(1)) depend(inout: ptr) + + ! Play also around with the alignment - as hsa_amd_memory_fill operates + ! on multiples of 4 bytes (c_int32_t) + + do start = 0, 31 + do tail = 0, 31 + val = iachar('0') + start + tail + + tmpptr = transfer (transfer (ptr, intptr) + start, tmpptr) + ptr2 = omp_target_memset_async (tmpptr, val, count - start - tail, dev, 0) + + if (.not. c_associated (tmpptr, ptr2)) stop 1 + + !$omp taskwait + + !$omp target device(dev) is_device_ptr(ptr) depend(depobj: dep(1)) nowait + do i = 1 + start, int(count, c_int) - start - tail + call c_f_pointer (ptr, fptr, [count]) + if (fptr(i) /= int (val, c_int8_t)) stop 2 + fptr(i) = fptr(i) + 2_c_int8_t + end do + !$omp end target + + ptr2 = omp_target_memset_async (tmpptr, val + 3, & + count - start - tail, dev, 1, dep) + + !$omp target device(dev) is_device_ptr(ptr) depend(depobj: dep(1)) nowait + do i = 1 + start, int(count, c_int) - start - tail + call c_f_pointer (ptr, fptr, [count]) + if (fptr(i) /= int (val + 3, c_int8_t)) stop 3 + fptr(i) = fptr(i) - 1_c_int8_t + end do + !$omp end target + + ptr2 = omp_target_memset_async (tmpptr, val - 3, & + count - start - tail, dev, 1, dep) + + !$omp taskwait depend (depobj: dep(1)) + end do + end do + + !$omp depobj(dep(1)) destroy + call omp_target_free (ptr, dev); +end block +end do +end diff --git a/libgomp/testsuite/libgomp.fortran/omp_target_memset.f90 b/libgomp/testsuite/libgomp.fortran/omp_target_memset.f90 new file mode 100644 index 0000000..1ee184a --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/omp_target_memset.f90 @@ -0,0 +1,39 @@ +! PR libgomp/120444 + +use omp_lib +use iso_c_binding +implicit none (type, external) + +integer(c_int) :: dev, i, val, start, tail +type(c_ptr) :: ptr, ptr2, tmpptr +integer(c_int8_t), pointer, contiguous :: fptr(:) +integer(c_intptr_t) :: intptr +integer(c_size_t), parameter :: count = 1024 + +do dev = omp_initial_device, omp_get_num_devices () + ptr = omp_target_alloc (count, dev) + + ! Play also around with the alignment - as hsa_amd_memory_fill operates + ! on multiples of 4 bytes (c_int32_t) + + do start = 0, 31 + do tail = 0, 31 + val = iachar('0') + start + tail + + tmpptr = transfer (transfer (ptr, intptr) + start, tmpptr) + ptr2 = omp_target_memset (tmpptr, val, count - start - tail, dev) + + if (.not. c_associated (tmpptr, ptr2)) stop 1 + + !$omp target device(dev) is_device_ptr(ptr) + do i = 1 + start, int(count, c_int) - start - tail + call c_f_pointer (ptr, fptr, [count]) + if (fptr(i) /= int (val, c_int8_t)) stop 2 + end do + !$omp end target + end do + end do + + call omp_target_free (ptr, dev); +end do +end diff --git a/libgomp/testsuite/libgomp.fortran/target-enter-data-8.f90 b/libgomp/testsuite/libgomp.fortran/target-enter-data-8.f90 new file mode 100644 index 0000000..c6d671c --- /dev/null +++ b/libgomp/testsuite/libgomp.fortran/target-enter-data-8.f90 @@ -0,0 +1,532 @@ +! { dg-additional-options "-cpp" } + +! FIXME: Some tests do not work yet. Those are for now in '#if 0' + +! Check that 'map(alloc:' properly works with +! - deferred-length character strings +! - arrays with array descriptors +! For those, the array descriptor / string length must be mapped with 'to:' + +program main +implicit none + +type t + integer :: ic(2:5), ic2 + character(len=11) :: ccstr(3:4), ccstr2 + character(len=11,kind=4) :: cc4str(3:7), cc4str2 + integer, pointer :: pc(:), pc2 + character(len=:), pointer :: pcstr(:), pcstr2 + character(len=:,kind=4), pointer :: pc4str(:), pc4str2 +end type t + +type(t) :: dt + +integer :: ii(5), ii2 +character(len=11) :: clstr(-1:1), clstr2 +character(len=11,kind=4) :: cl4str(0:3), cl4str2 +integer, pointer :: ip(:), ip2 +integer, allocatable :: ia(:), ia2 +character(len=:), pointer :: pstr(:), pstr2 +character(len=:), allocatable :: astr(:), astr2 +character(len=:,kind=4), pointer :: p4str(:), p4str2 +character(len=:,kind=4), allocatable :: a4str(:), a4str2 + + +allocate(dt%pc(5), dt%pc2) +allocate(character(len=2) :: dt%pcstr(2)) +allocate(character(len=4) :: dt%pcstr2) + +allocate(character(len=3,kind=4) :: dt%pc4str(2:3)) +allocate(character(len=5,kind=4) :: dt%pc4str2) + +allocate(ip(5), ip2, ia(8), ia2) +allocate(character(len=2) :: pstr(-2:0)) +allocate(character(len=4) :: pstr2) +allocate(character(len=6) :: astr(3:5)) +allocate(character(len=8) :: astr2) + +allocate(character(len=3,kind=4) :: p4str(2:4)) +allocate(character(len=5,kind=4) :: p4str2) +allocate(character(len=7,kind=4) :: a4str(-2:3)) +allocate(character(len=9,kind=4) :: a4str2) + + +! integer :: ic(2:5), ic2 + +!$omp target enter data map(alloc: dt%ic) +!$omp target map(alloc: dt%ic) + if (size(dt%ic) /= 4) error stop + if (lbound(dt%ic, 1) /= 2) error stop + if (ubound(dt%ic, 1) /= 5) error stop + dt%ic = [22, 33, 44, 55] +!$omp end target +!$omp target exit data map(from: dt%ic) +if (size(dt%ic) /= 4) error stop +if (lbound(dt%ic, 1) /= 2) error stop +if (ubound(dt%ic, 1) /= 5) error stop +if (any (dt%ic /= [22, 33, 44, 55])) error stop + +!$omp target enter data map(alloc: dt%ic2) +!$omp target map(alloc: dt%ic2) + dt%ic2 = 42 +!$omp end target +!$omp target exit data map(from: dt%ic2) +if (dt%ic2 /= 42) error stop + + +! character(len=11) :: ccstr(3:4), ccstr2 + +!$omp target enter data map(alloc: dt%ccstr) +!$omp target map(alloc: dt%ccstr) + if (len(dt%ccstr) /= 11) error stop + if (size(dt%ccstr) /= 2) error stop + if (lbound(dt%ccstr, 1) /= 3) error stop + if (ubound(dt%ccstr, 1) /= 4) error stop + dt%ccstr = ["12345678901", "abcdefghijk"] +!$omp end target +!$omp target exit data map(from: dt%ccstr) +if (len(dt%ccstr) /= 11) error stop +if (size(dt%ccstr) /= 2) error stop +if (lbound(dt%ccstr, 1) /= 3) error stop +if (ubound(dt%ccstr, 1) /= 4) error stop +if (any (dt%ccstr /= ["12345678901", "abcdefghijk"])) error stop + +!$omp target enter data map(alloc: dt%ccstr2) +!$omp target map(alloc: dt%ccstr2) + if (len(dt%ccstr2) /= 11) error stop + dt%ccstr2 = "ABCDEFGHIJK" +!$omp end target +!$omp target exit data map(from: dt%ccstr2) +if (len(dt%ccstr2) /= 11) error stop +if (dt%ccstr2 /= "ABCDEFGHIJK") error stop + + +! character(len=11,kind=4) :: cc4str(3:7), cc4str2 + +#if 0 +! Value check fails +!$omp target map(alloc: dt%cc4str) + if (len(dt%cc4str) /= 11) error stop + if (size(dt%cc4str) /= 5) error stop + if (lbound(dt%cc4str, 1) /= 3) error stop + if (ubound(dt%cc4str, 1) /= 7) error stop + dt%cc4str = [4_"12345678901", 4_"abcdefghijk", & + 4_"qerftcea6ds", 4_"a1f9g37ga4.", & + 4_"45ngwj56sj2"] +!$omp end target +!$omp target exit data map(from: dt%cc4str) +if (len(dt%cc4str) /= 11) error stop +if (size(dt%cc4str) /= 5) error stop +if (lbound(dt%cc4str, 1) /= 3) error stop +if (ubound(dt%cc4str, 1) /= 7) error stop +if (dt%cc4str(3) /= 4_"12345678901") error stop +if (dt%cc4str(4) /= 4_"abcdefghijk") error stop +if (dt%cc4str(5) /= 4_"qerftcea6ds") error stop +if (dt%cc4str(6) /= 4_"a1f9g37ga4.") error stop +if (dt%cc4str(7) /= 4_"45ngwj56sj2") error stop +#endif + +!$omp target enter data map(alloc: dt%cc4str2) +!$omp target map(alloc: dt%cc4str2) + if (len(dt%cc4str2) /= 11) error stop + dt%cc4str2 = 4_"ABCDEFGHIJK" +!$omp end target +!$omp target exit data map(from: dt%cc4str2) +if (len(dt%cc4str2) /= 11) error stop +if (dt%cc4str2 /= 4_"ABCDEFGHIJK") error stop + + +! integer, pointer :: pc(:), pc2 +! allocate(dt%pc(5), dt%pc2) + +!$omp target enter data map(alloc: dt%pc) +!$omp target map(alloc: dt%pc) + if (.not. associated(dt%pc)) error stop + if (size(dt%pc) /= 5) error stop + if (lbound(dt%pc, 1) /= 1) error stop + if (ubound(dt%pc, 1) /= 5) error stop + dt%pc = [11, 22, 33, 44, 55] +!$omp end target +!$omp target exit data map(from: dt%pc) +if (.not. associated(dt%pc)) error stop +if (size(dt%pc) /= 5) error stop +if (lbound(dt%pc, 1) /= 1) error stop +if (ubound(dt%pc, 1) /= 5) error stop +if (any (dt%pc /= [11, 22, 33, 44, 55])) error stop + +!$omp target enter data map(alloc: dt%pc2) +!$omp target map(alloc: dt%pc2) + if (.not. associated(dt%pc2)) error stop + dt%pc2 = 99 +!$omp end target +!$omp target exit data map(from: dt%pc2) +if (dt%pc2 /= 99) error stop +if (.not. associated(dt%pc2)) error stop + + +! character(len=:), pointer :: pcstr(:), pcstr2 +! allocate(character(len=2) :: dt%pcstr(2)) +! allocate(character(len=4) :: dt%pcstr2) + +!$omp target enter data map(alloc: dt%pcstr) +!$omp target map(alloc: dt%pcstr) + if (.not. associated(dt%pcstr)) error stop + if (len(dt%pcstr) /= 2) error stop + if (size(dt%pcstr) /= 2) error stop + if (lbound(dt%pcstr, 1) /= 1) error stop + if (ubound(dt%pcstr, 1) /= 2) error stop + dt%pcstr = ["01", "jk"] +!$omp end target +!$omp target exit data map(from: dt%pcstr) +if (.not. associated(dt%pcstr)) error stop +if (len(dt%pcstr) /= 2) error stop +if (size(dt%pcstr) /= 2) error stop +if (lbound(dt%pcstr, 1) /= 1) error stop +if (ubound(dt%pcstr, 1) /= 2) error stop +if (any (dt%pcstr /= ["01", "jk"])) error stop + + +!$omp target enter data map(alloc: dt%pcstr2) +!$omp target map(alloc: dt%pcstr2) + if (.not. associated(dt%pcstr2)) error stop + if (len(dt%pcstr2) /= 4) error stop + dt%pcstr2 = "HIJK" +!$omp end target +!$omp target exit data map(from: dt%pcstr2) +if (.not. associated(dt%pcstr2)) error stop +if (len(dt%pcstr2) /= 4) error stop +if (dt%pcstr2 /= "HIJK") error stop + + +! character(len=:,kind=4), pointer :: pc4str(:), pc4str2 +! allocate(character(len=3,kind=4) :: dt%pc4str(2:3)) +! allocate(character(len=5,kind=4) :: dt%pc4str2) + +!$omp target enter data map(alloc: dt%pc4str) +!$omp target map(alloc: dt%pc4str) + if (.not. associated(dt%pc4str)) error stop + if (len(dt%pc4str) /= 3) error stop + if (size(dt%pc4str) /= 2) error stop + if (lbound(dt%pc4str, 1) /= 2) error stop + if (ubound(dt%pc4str, 1) /= 3) error stop + dt%pc4str = [4_"456", 4_"tzu"] +!$omp end target +!$omp target exit data map(from: dt%pc4str) +if (.not. associated(dt%pc4str)) error stop +if (len(dt%pc4str) /= 3) error stop +if (size(dt%pc4str) /= 2) error stop +if (lbound(dt%pc4str, 1) /= 2) error stop +if (ubound(dt%pc4str, 1) /= 3) error stop +if (dt%pc4str(2) /= 4_"456") error stop +if (dt%pc4str(3) /= 4_"tzu") error stop + +!$omp target enter data map(alloc: dt%pc4str2) +!$omp target map(alloc: dt%pc4str2) + if (.not. associated(dt%pc4str2)) error stop + if (len(dt%pc4str2) /= 5) error stop + dt%pc4str2 = 4_"98765" +!$omp end target +!$omp target exit data map(from: dt%pc4str2) +if (.not. associated(dt%pc4str2)) error stop +if (len(dt%pc4str2) /= 5) error stop +if (dt%pc4str2 /= 4_"98765") error stop + + +! integer :: ii(5), ii2 + +!$omp target enter data map(alloc: ii) +!$omp target map(alloc: ii) + if (size(ii) /= 5) error stop + if (lbound(ii, 1) /= 1) error stop + if (ubound(ii, 1) /= 5) error stop + ii = [-1, -2, -3, -4, -5] +!$omp end target +!$omp target exit data map(from: ii) +if (size(ii) /= 5) error stop +if (lbound(ii, 1) /= 1) error stop +if (ubound(ii, 1) /= 5) error stop +if (any (ii /= [-1, -2, -3, -4, -5])) error stop + +!$omp target enter data map(alloc: ii2) +!$omp target map(alloc: ii2) + ii2 = -410 +!$omp end target +!$omp target exit data map(from: ii2) +if (ii2 /= -410) error stop + + +! character(len=11) :: clstr(-1:1), clstr2 + +!$omp target enter data map(alloc: clstr) +!$omp target map(alloc: clstr) + if (len(clstr) /= 11) error stop + if (size(clstr) /= 3) error stop + if (lbound(clstr, 1) /= -1) error stop + if (ubound(clstr, 1) /= 1) error stop + clstr = ["12345678901", "abcdefghijk", "ABCDEFGHIJK"] +!$omp end target +!$omp target exit data map(from: clstr) +if (len(clstr) /= 11) error stop +if (size(clstr) /= 3) error stop +if (lbound(clstr, 1) /= -1) error stop +if (ubound(clstr, 1) /= 1) error stop +if (any (clstr /= ["12345678901", "abcdefghijk", "ABCDEFGHIJK"])) error stop + +!$omp target enter data map(alloc: clstr2) +!$omp target map(alloc: clstr2) + if (len(clstr2) /= 11) error stop + clstr2 = "ABCDEFghijk" +!$omp end target +!$omp target exit data map(from: clstr2) +if (len(clstr2) /= 11) error stop +if (clstr2 /= "ABCDEFghijk") error stop + + +! character(len=11,kind=4) :: cl4str(0:3), cl4str2 + +!$omp target enter data map(alloc: cl4str) +!$omp target map(alloc: cl4str) + if (len(cl4str) /= 11) error stop + if (size(cl4str) /= 4) error stop + if (lbound(cl4str, 1) /= 0) error stop + if (ubound(cl4str, 1) /= 3) error stop + cl4str = [4_"12345678901", 4_"abcdefghijk", & + 4_"qerftcea6ds", 4_"a1f9g37ga4."] +!$omp end target +!$omp target exit data map(from: cl4str) +if (len(cl4str) /= 11) error stop +if (size(cl4str) /= 4) error stop +if (lbound(cl4str, 1) /= 0) error stop +if (ubound(cl4str, 1) /= 3) error stop +if (cl4str(0) /= 4_"12345678901") error stop +if (cl4str(1) /= 4_"abcdefghijk") error stop +if (cl4str(2) /= 4_"qerftcea6ds") error stop +if (cl4str(3) /= 4_"a1f9g37ga4.") error stop + +!$omp target enter data map(alloc: cl4str2) +!$omp target map(alloc: cl4str2) + if (len(cl4str2) /= 11) error stop + cl4str2 = 4_"ABCDEFGHIJK" +!$omp end target +!$omp target exit data map(from: cl4str2) +if (len(cl4str2) /= 11) error stop +if (cl4str2 /= 4_"ABCDEFGHIJK") error stop + + +! allocate(ip(5), ip2, ia(8), ia2) + +!$omp target enter data map(alloc: ip) +!$omp target map(alloc: ip) + if (.not. associated(ip)) error stop + if (size(ip) /= 5) error stop + if (lbound(ip, 1) /= 1) error stop + if (ubound(ip, 1) /= 5) error stop + ip = [11, 22, 33, 44, 55] +!$omp end target +!$omp target exit data map(from: ip) +if (.not. associated(ip)) error stop +if (size(ip) /= 5) error stop +if (lbound(ip, 1) /= 1) error stop +if (ubound(ip, 1) /= 5) error stop +if (any (ip /= [11, 22, 33, 44, 55])) error stop + +!$omp target enter data map(alloc: ip2) +!$omp target map(alloc: ip2) + if (.not. associated(ip2)) error stop + ip2 = 99 +!$omp end target +!$omp target exit data map(from: ip2) +if (ip2 /= 99) error stop +if (.not. associated(ip2)) error stop + + +! allocate(ip(5), ip2, ia(8), ia2) + +!$omp target enter data map(alloc: ia) +!$omp target map(alloc: ia) + if (.not. allocated(ia)) error stop + if (size(ia) /= 8) error stop + if (lbound(ia, 1) /= 1) error stop + if (ubound(ia, 1) /= 8) error stop + ia = [1,2,3,4,5,6,7,8] +!$omp end target +!$omp target exit data map(from: ia) +if (.not. allocated(ia)) error stop +if (size(ia) /= 8) error stop +if (lbound(ia, 1) /= 1) error stop +if (ubound(ia, 1) /= 8) error stop +if (any (ia /= [1,2,3,4,5,6,7,8])) error stop + +!$omp target enter data map(alloc: ia2) +!$omp target map(alloc: ia2) + if (.not. allocated(ia2)) error stop + ia2 = 102 +!$omp end target +!$omp target exit data map(from: ia2) +if (ia2 /= 102) error stop +if (.not. allocated(ia2)) error stop + + +! character(len=:), pointer :: pstr(:), pstr2 +! allocate(character(len=2) :: pstr(-2:0)) +! allocate(character(len=4) :: pstr2) + +!$omp target enter data map(alloc: pstr) +!$omp target map(alloc: pstr) + if (.not. associated(pstr)) error stop + if (len(pstr) /= 2) error stop + if (size(pstr) /= 3) error stop + if (lbound(pstr, 1) /= -2) error stop + if (ubound(pstr, 1) /= 0) error stop + pstr = ["01", "jk", "aq"] +!$omp end target +!$omp target exit data map(from: pstr) +if (.not. associated(pstr)) error stop +if (len(pstr) /= 2) error stop +if (size(pstr) /= 3) error stop +if (lbound(pstr, 1) /= -2) error stop +if (ubound(pstr, 1) /= 0) error stop +if (any (pstr /= ["01", "jk", "aq"])) error stop + +!$omp target enter data map(alloc: pstr2) +!$omp target map(alloc: pstr2) + if (.not. associated(pstr2)) error stop + if (len(pstr2) /= 4) error stop + pstr2 = "HIJK" +!$omp end target +!$omp target exit data map(from: pstr2) +if (.not. associated(pstr2)) error stop +if (len(pstr2) /= 4) error stop +if (pstr2 /= "HIJK") error stop + + +! character(len=:), allocatable :: astr(:), astr2 +! allocate(character(len=6) :: astr(3:5)) +! allocate(character(len=8) :: astr2) + + +!$omp target enter data map(alloc: astr) +!$omp target map(alloc: astr) + if (.not. allocated(astr)) error stop + if (len(astr) /= 6) error stop + if (size(astr) /= 3) error stop + if (lbound(astr, 1) /= 3) error stop + if (ubound(astr, 1) /= 5) error stop + astr = ["01db45", "jk$D%S", "zutg47"] +!$omp end target +!$omp target exit data map(from: astr) +if (.not. allocated(astr)) error stop +if (len(astr) /= 6) error stop +if (size(astr) /= 3) error stop +if (lbound(astr, 1) /= 3) error stop +if (ubound(astr, 1) /= 5) error stop +if (any (astr /= ["01db45", "jk$D%S", "zutg47"])) error stop + + +!$omp target enter data map(alloc: astr2) +!$omp target map(alloc: astr2) + if (.not. allocated(astr2)) error stop + if (len(astr2) /= 8) error stop + astr2 = "HIJKhijk" +!$omp end target +!$omp target exit data map(from: astr2) +if (.not. allocated(astr2)) error stop +if (len(astr2) /= 8) error stop +if (astr2 /= "HIJKhijk") error stop + + +! character(len=:,kind=4), pointer :: p4str(:), p4str2 +! allocate(character(len=3,kind=4) :: p4str(2:4)) +! allocate(character(len=5,kind=4) :: p4str2) + +! FAILS with value check + +!$omp target enter data map(alloc: p4str) +!$omp target map(alloc: p4str) + if (.not. associated(p4str)) error stop + if (len(p4str) /= 3) error stop + if (size(p4str) /= 3) error stop + if (lbound(p4str, 1) /= 2) error stop + if (ubound(p4str, 1) /= 4) error stop + p4str(:) = [4_"f85", 4_"8af", 4_"A%F"] +!$omp end target +!$omp target exit data map(from: p4str) +if (.not. associated(p4str)) error stop +if (len(p4str) /= 3) error stop +if (size(p4str) /= 3) error stop +if (lbound(p4str, 1) /= 2) error stop +if (ubound(p4str, 1) /= 4) error stop +if (p4str(2) /= 4_"f85") error stop +if (p4str(3) /= 4_"8af") error stop +if (p4str(4) /= 4_"A%F") error stop + +!$omp target enter data map(alloc: p4str2) +!$omp target map(alloc: p4str2) + if (.not. associated(p4str2)) error stop + if (len(p4str2) /= 5) error stop + p4str2 = 4_"9875a" +!$omp end target +!$omp target exit data map(from: p4str2) +if (.not. associated(p4str2)) error stop +if (len(p4str2) /= 5) error stop +if (p4str2 /= 4_"9875a") error stop + + +! character(len=:,kind=4), allocatable :: a4str(:), a4str2 +! allocate(character(len=7,kind=4) :: a4str(-2:3)) +! allocate(character(len=9,kind=4) :: a4str2) + +!$omp target enter data map(alloc: a4str) +!$omp target map(alloc: a4str) + if (.not. allocated(a4str)) error stop + if (len(a4str) /= 7) error stop + if (size(a4str) /= 6) error stop + if (lbound(a4str, 1) /= -2) error stop + if (ubound(a4str, 1) /= 3) error stop + ! See PR fortran/107508 why '(:)' is required + a4str(:) = [4_"sf456aq", 4_"3dtzu24", 4_"_4fh7sm", 4_"=ff85s7", 4_"j=8af4d", 4_".,A%Fsz"] +!$omp end target +!$omp target exit data map(from: a4str) +if (.not. allocated(a4str)) error stop +if (len(a4str) /= 7) error stop +if (size(a4str) /= 6) error stop +if (lbound(a4str, 1) /= -2) error stop +if (ubound(a4str, 1) /= 3) error stop +if (a4str(-2) /= 4_"sf456aq") error stop +if (a4str(-1) /= 4_"3dtzu24") error stop +if (a4str(0) /= 4_"_4fh7sm") error stop +if (a4str(1) /= 4_"=ff85s7") error stop +if (a4str(2) /= 4_"j=8af4d") error stop +if (a4str(3) /= 4_".,A%Fsz") error stop + +!$omp target enter data map(alloc: a4str2) +!$omp target map(alloc: a4str2) + if (.not. allocated(a4str2)) error stop + if (len(a4str2) /= 9) error stop + a4str2 = 4_"98765a23d" +!$omp end target +!$omp target exit data map(from: a4str2) +if (.not. allocated(a4str2)) error stop +if (len(a4str2) /= 9) error stop +if (a4str2 /= 4_"98765a23d") error stop + + +deallocate(dt%pc, dt%pc2) +deallocate(dt%pcstr) +deallocate(dt%pcstr2) + +deallocate(dt%pc4str) +deallocate(dt%pc4str2) + +deallocate(ip, ip2, ia, ia2) +deallocate(pstr) +deallocate(pstr2) +deallocate(astr) +deallocate(astr2) + +deallocate(p4str) +deallocate(p4str2) +deallocate(a4str) +deallocate(a4str2) + +end diff --git a/libgomp/testsuite/libgomp.oacc-c++/exceptions-bad_cast-1.C b/libgomp/testsuite/libgomp.oacc-c++/exceptions-bad_cast-1.C index 0545601..6957a6c 100644 --- a/libgomp/testsuite/libgomp.oacc-c++/exceptions-bad_cast-1.C +++ b/libgomp/testsuite/libgomp.oacc-c++/exceptions-bad_cast-1.C @@ -52,3 +52,6 @@ int main() PR119692. { dg-shouldfail {'std::bad_cast' exception} } */ +/* There are configurations where we 'WARNING: program timed out.' while in + 'dynamic_cast', see <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119692#c6>. + { dg-timeout 10 } ... to make sure that happens quickly. */ diff --git a/libgomp/testsuite/libgomp.oacc-c++/exceptions-bad_cast-2.C b/libgomp/testsuite/libgomp.oacc-c++/exceptions-bad_cast-2.C index 24399ef..0f84cf2 100644 --- a/libgomp/testsuite/libgomp.oacc-c++/exceptions-bad_cast-2.C +++ b/libgomp/testsuite/libgomp.oacc-c++/exceptions-bad_cast-2.C @@ -58,3 +58,6 @@ int main() For GCN, nvptx offload execution, there is no 'catch'ing; any exception is fatal. { dg-shouldfail {'std::bad_cast' exception} { ! openacc_host_selected } } */ +/* There are configurations where we 'WARNING: program timed out.' while in + 'dynamic_cast', see <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119692#c6>. + { dg-timeout 10 } ... to make sure that happens quickly. */ diff --git a/libgomp/testsuite/libgomp.oacc-c++/exceptions-throw-1.C b/libgomp/testsuite/libgomp.oacc-c++/exceptions-throw-1.C index f2ef751..08c5766 100644 --- a/libgomp/testsuite/libgomp.oacc-c++/exceptions-throw-1.C +++ b/libgomp/testsuite/libgomp.oacc-c++/exceptions-throw-1.C @@ -4,9 +4,6 @@ { dg-additional-options -fexceptions } */ /* { dg-additional-options -fdump-tree-optimized-raw } { dg-additional-options -foffload-options=-fdump-tree-optimized-raw } */ -/* { dg-bogus {Size expression must be absolute\.} PR119737 { target { openacc_radeon_accel_selected && __OPTIMIZE__ } xfail *-*-* } 0 } - { dg-ice PR119737 { openacc_radeon_accel_selected && __OPTIMIZE__ } } - { dg-excess-errors {'mkoffload' failure etc.} { xfail { openacc_radeon_accel_selected && __OPTIMIZE__ } } } */ /* See also '../libgomp.c++/target-exceptions-throw-1.C'. */ diff --git a/libgomp/testsuite/libgomp.oacc-c++/exceptions-throw-2.C b/libgomp/testsuite/libgomp.oacc-c++/exceptions-throw-2.C index f6dc970..a7408cd 100644 --- a/libgomp/testsuite/libgomp.oacc-c++/exceptions-throw-2.C +++ b/libgomp/testsuite/libgomp.oacc-c++/exceptions-throw-2.C @@ -6,9 +6,6 @@ { dg-additional-options -foffload-options=-fdump-tree-optimized-raw } */ /* { dg-bogus {undefined symbol: typeinfo name for MyException} PR119806 { target { openacc_radeon_accel_selected && { ! __OPTIMIZE__ } } xfail *-*-* } 0 } { dg-excess-errors {'mkoffload' failure etc.} { xfail { openacc_radeon_accel_selected && { ! __OPTIMIZE__ } } } } */ -/* { dg-bogus {Size expression must be absolute\.} PR119737 { target { openacc_radeon_accel_selected && __OPTIMIZE__ } xfail *-*-* } 0 } - { dg-ice PR119737 { openacc_radeon_accel_selected && __OPTIMIZE__ } } - { dg-excess-errors {'mkoffload' failures etc.} { xfail { openacc_radeon_accel_selected && __OPTIMIZE__ } } } */ /* { dg-bogus {Initial value type mismatch} PR119806 { target { openacc_nvidia_accel_selected && { ! __OPTIMIZE__ } } xfail *-*-* } 0 } { dg-excess-errors {'mkoffload' failure etc.} { xfail { openacc_nvidia_accel_selected && { ! __OPTIMIZE__ } } } } */ diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/abi-struct-1.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/abi-struct-1.c new file mode 100644 index 0000000..4b54171 --- /dev/null +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/abi-struct-1.c @@ -0,0 +1,125 @@ +/* Inspired by 'gcc.target/nvptx/abi-struct-arg.c', 'gcc.target/nvptx/abi-struct-ret.c'. */ + +/* See also '../libgomp.c-c++-common/target-abi-struct-1.c'. */ + +/* To exercise PR119835 (if optimizations enabled): disable inlining, so that + GIMPLE passes still see the functions that return aggregate types. */ +#pragma GCC optimize "-fno-inline" + +typedef struct {} empty; /* See 'gcc/doc/extend.texi', "Empty Structures". */ +typedef struct {char a;} schar; +typedef struct {short a;} sshort; +typedef struct {int a;} sint; +typedef struct {long long a;} slonglong; +typedef struct {int a, b[12];} sint_13; + +#pragma omp declare target + +#define M(T) ({T t; t.a = sizeof t; t;}) + +static __SIZE_TYPE__ empty_a; +#pragma acc declare create(empty_a) +#pragma acc routine +static empty rempty(void) +{ + return ({empty t; empty_a = sizeof t; t;}); +} + +#pragma acc routine +static schar rschar(void) +{ + return M(schar); +} + +#pragma acc routine +static sshort rsshort(void) +{ + return M(sshort); +} + +#pragma acc routine +static sint rsint(void) +{ + return M(sint); +} + +#pragma acc routine +static slonglong rslonglong(void) +{ + return M(slonglong); +} + +#pragma acc routine +static sint_13 rsint_13(void) +{ + return M(sint_13); +} + +#pragma acc routine +static void aempty(empty empty) +{ + (void) empty; + + __SIZE_TYPE__ empty_a_exp; +#ifndef __cplusplus + empty_a_exp = 0; +#else + empty_a_exp = sizeof (char); +#endif + if (empty_a != empty_a_exp) + __builtin_abort(); +} + +#pragma acc routine +static void aschar(schar schar) +{ + if (schar.a != sizeof (char)) + __builtin_abort(); +} + +#pragma acc routine +static void asshort(sshort sshort) +{ + if (sshort.a != sizeof (short)) + __builtin_abort(); +} + +#pragma acc routine +static void asint(sint sint) +{ + if (sint.a != sizeof (int)) + __builtin_abort(); +} + +#pragma acc routine +static void aslonglong(slonglong slonglong) +{ + if (slonglong.a != sizeof (long long)) + __builtin_abort(); +} + +#pragma acc routine +static void asint_13(sint_13 sint_13) +{ + if (sint_13.a != (sizeof (int) * 13)) + __builtin_abort(); +} + +#pragma omp end declare target + +int main() +{ +#pragma omp target +#pragma acc serial + /* { dg-bogus {using 'vector_length \(32\)', ignoring 1} {} { target openacc_nvidia_accel_selected xfail *-*-* } .-1 } */ + { + aempty(rempty()); + aschar(rschar()); + asshort(rsshort()); + asint(rsint()); + aslonglong(rslonglong()); + asint_13(rsint_13()); + } + + return 0; +} diff --git a/libgomp/testsuite/libgomp.oacc-c-c++-common/acc_memcpy_device-1.c b/libgomp/testsuite/libgomp.oacc-c-c++-common/acc_memcpy_device-1.c new file mode 100644 index 0000000..eda651d --- /dev/null +++ b/libgomp/testsuite/libgomp.oacc-c-c++-common/acc_memcpy_device-1.c @@ -0,0 +1,96 @@ +/* { dg-prune-output "using .vector_length \\(32\\)" } */ + +/* PR libgomp/93226 */ + +#include <stdlib.h> +#include <stdint.h> +#include <string.h> +#include <openacc.h> + +enum { N = 1024 }; + +static int D[N]; +#pragma acc declare device_resident(D) + +#pragma acc routine +intptr_t init_d() +{ + for (int i = 0; i < N; i++) + D[i] = 27*i; + return (intptr_t) &D[0]; +} + +int +main () +{ + int *a, *b, *e; + void *d_a, *d_b, *d_c, *d_d, *d_e, *d_f; + intptr_t intptr; + bool fail = false; + + a = (int *) malloc (N*sizeof (int)); + b = (int *) malloc (N*sizeof (int)); + e = (int *) malloc (N*sizeof (int)); + d_c = acc_malloc (N*sizeof (int)); + d_f = acc_malloc (N*sizeof (int)); + + memset (e, 0xff, N*sizeof (int)); + d_e = acc_copyin (e, N*sizeof (int)); + + #pragma acc serial copyout(intptr) + intptr = init_d (); + d_d = (void*) intptr; + acc_memcpy_device (d_c, d_d, N*sizeof (int)); + + #pragma acc serial copy(fail) deviceptr(d_c) firstprivate(intptr) + { + int *cc = (int *) d_c; + int *dd = (int *) intptr; + for (int i = 0; i < N; i++) + if (dd[i] != 27*i || cc[i] != 27*i) + { + fail = true; + __builtin_abort (); + } + } + if (fail) __builtin_abort (); + + for (int i = 0; i < N; i++) + a[i] = 11*i; + for (int i = 0; i < N; i++) + b[i] = 31*i; + + d_a = acc_copyin (a, N*sizeof (int)); + acc_copyin_async (b, N*sizeof (int), acc_async_noval); + + #pragma acc parallel deviceptr(d_c) async + { + int *cc = (int *) d_c; + #pragma acc loop + for (int i = 0; i < N; i++) + cc[i] = -17*i; + } + + acc_memcpy_device_async (d_d, d_a, N*sizeof (int), acc_async_noval); + acc_memcpy_device_async (d_f, d_c, N*sizeof (int), acc_async_noval); + acc_wait (acc_async_noval); + d_b = acc_deviceptr (b); + acc_memcpy_device_async (d_e, d_b, N*sizeof (int), acc_async_noval); + acc_wait (acc_async_noval); + + #pragma acc serial deviceptr(d_d, d_e, d_f) copy(fail) + { + int *dd = (int *) d_d; + int *ee = (int *) d_e; + int *ff = (int *) d_f; + for (int i = 0; i < N; i++) + if (dd[i] != 11*i + || ee[i] != 31*i + || ff[i] != -17*i) + { + fail = true; + __builtin_abort (); + } + } + if (fail) __builtin_abort (); +} diff --git a/libgomp/testsuite/libgomp.oacc-fortran/acc-attach-detach-1.f90 b/libgomp/testsuite/libgomp.oacc-fortran/acc-attach-detach-1.f90 new file mode 100644 index 0000000..15393b4 --- /dev/null +++ b/libgomp/testsuite/libgomp.oacc-fortran/acc-attach-detach-1.f90 @@ -0,0 +1,25 @@ +! { dg-do compile } +! { dg-additional-options "-fdump-tree-original" } + +use openacc +implicit none (type, external) +integer,pointer :: a, b(:) +integer,allocatable :: c, d(:) + +call acc_attach(a) ! ICE +call acc_attach_async(b, 4) +call acc_attach(c) + +call acc_detach(a) +call acc_detach_async(b, 4) +call acc_detach_finalize(c) +call acc_detach_finalize_async(d,7) +end + +! { dg-final { scan-tree-dump-times "acc_attach \\(&a\\);" 1 "original" } } +! { dg-final { scan-tree-dump-times "acc_attach_async \\(&\\(integer\\(kind=4\\)\\\[0:\\\] \\*\\) b.data, 4\\);" 1 "original" } } +! { dg-final { scan-tree-dump-times "acc_attach \\(&c\\);" 1 "original" } } +! { dg-final { scan-tree-dump-times "acc_detach \\(&a\\);" 1 "original" } } +! { dg-final { scan-tree-dump-times "acc_detach_async \\(&\\(integer\\(kind=4\\)\\\[0:\\\] \\*\\) b.data, 4\\);" 1 "original" } } +! { dg-final { scan-tree-dump-times "acc_detach_finalize \\(&c\\);" 1 "original" } } +! { dg-final { scan-tree-dump-times "acc_detach_finalize_async \\(&\\(integer\\(kind=4\\)\\\[0:\\\] \\* restrict\\) d.data, 7\\);" 1 "original" } } diff --git a/libgomp/testsuite/libgomp.oacc-fortran/acc-attach-detach-2.f90 b/libgomp/testsuite/libgomp.oacc-fortran/acc-attach-detach-2.f90 new file mode 100644 index 0000000..b2204ac --- /dev/null +++ b/libgomp/testsuite/libgomp.oacc-fortran/acc-attach-detach-2.f90 @@ -0,0 +1,62 @@ +! { dg-do run } + +use openacc +implicit none (type, external) +integer, target :: tgt_a, tgt_b(5) + +integer, pointer :: p1, p2(:) + +type t + integer,pointer :: a => null () + integer,pointer :: b(:) => null () + integer,allocatable :: c, d(:) +end type t + +type(t), target :: var + +tgt_a = 51 +tgt_b = [11,22,33,44,55] + +var%b => tgt_b +!$acc enter data copyin(var, tgt_a, tgt_b) +var%a => tgt_a + +call acc_attach(var%a) +call acc_attach(var%b) + +!$acc serial +! { dg-warning "using .vector_length \\(32\\)., ignoring 1" "" { target openacc_nvidia_accel_selected } .-1 } + if (var%a /= 51) stop 1 + if (any (var%b /= [11,22,33,44,55])) stop 2 +!$acc end serial + +call acc_detach(var%a) +call acc_detach(var%b) + +!$acc exit data delete(var, tgt_a, tgt_b) + +var%c = 9 +var%d = [1,2,3] + +p1 => var%c +p2 => var%d + +!$acc enter data copyin(p1, p2) +!$acc enter data copyin(var) +call acc_attach(var%c) +call acc_attach(var%d) + +!$acc serial +! { dg-warning "using .vector_length \\(32\\)., ignoring 1" "" { target openacc_nvidia_accel_selected } .-1 } + if (var%c /= 9) stop 3 + if (any (var%d /= [1,2,3])) stop 4 +!$acc end serial + +call acc_detach(var%c) +call acc_detach(var%d) + +!$acc exit data delete(var, p1, p2) + +deallocate(var%d) + +end diff --git a/libgomp/testsuite/libgomp.oacc-fortran/acc_memcpy_device-1.f90 b/libgomp/testsuite/libgomp.oacc-fortran/acc_memcpy_device-1.f90 new file mode 100644 index 0000000..8f3a8f0 --- /dev/null +++ b/libgomp/testsuite/libgomp.oacc-fortran/acc_memcpy_device-1.f90 @@ -0,0 +1,113 @@ +! { dg-prune-output "using .vector_length \\(32\\)" } + +! PR libgomp/93226 */ + +module m + use iso_c_binding + use openacc + implicit none (external, type) + + integer, parameter :: N = 1024 + + integer :: D(N) + !$acc declare device_resident(D) + +contains + + integer(c_intptr_t) function init_d() + !$acc routine + integer :: i + do i = 1, N + D(i) = 27*i + end do + init_d = loc(D) + end +end module + +program main + use m + implicit none (external, type) + + integer, allocatable, target :: a(:), b(:), e(:) + type(c_ptr) :: d_a, d_b, d_c, d_d, d_e, d_f + integer(c_intptr_t) intptr + integer :: i + logical fail + + fail = .false. + + allocate(a(N), b(N), e(N)) + d_c = acc_malloc (N*c_sizeof (i)) + d_f = acc_malloc (N*c_sizeof (i)) + + e = huge(e) + call acc_copyin (e, N*c_sizeof (i)); + d_e = acc_deviceptr (e); + + !$acc serial copyout(intptr) + intptr = init_d () + !$acc end serial + d_d = transfer(intptr, d_d) + call acc_memcpy_device (d_c, d_d, N*c_sizeof (i)) + + !$acc serial copy(fail) copy(a) deviceptr(d_c, d_d) firstprivate(intptr) + block + integer, pointer :: cc(:), dd(:) + call c_f_pointer (d_c, cc, [N]) + call c_f_pointer (d_d, dd, [N]) + a = cc + do i = 1, N + if (dd(i) /= 27*i .or. cc(i) /= 27*i) then + fail = .true. + stop 1 + end if + end do + end block + !$acc end serial + if (fail) error stop 1 + + do i = 1, N + a(i) = 11*i + b(i) = 31*i + end do + + call acc_copyin (a, N*c_sizeof (i)) + d_a = acc_deviceptr (a) + call acc_copyin_async (b, N*c_sizeof (i), acc_async_noval) + + !$acc parallel deviceptr(d_c) private(i) async + block + integer, pointer :: cc(:) + call c_f_pointer (d_c, cc, [N]) + !$acc loop + do i = 1, N + cc(i) = -17*i + end do + end block + !$acc end parallel + + call acc_memcpy_device_async (d_d, d_a, N*c_sizeof (i), acc_async_noval) + call acc_memcpy_device_async (d_f, d_c, N*c_sizeof (i), acc_async_noval) + call acc_wait (acc_async_noval) + d_b = acc_deviceptr (b) + call acc_memcpy_device_async (d_e, d_b, N*c_sizeof (i), acc_async_noval) + call acc_wait (acc_async_noval) + + !$acc serial deviceptr(d_d, d_e, d_f) private(i) copy(fail) + block + integer, pointer :: dd(:), ee(:), ff(:) + call c_f_pointer (d_d, dd, [N]) + call c_f_pointer (d_e, ee, [N]) + call c_f_pointer (d_f, ff, [N]) + do i = 1, N + if (dd(i) /= 11*i & + .or. ee(i) /= 31*i & + .or. ff(i) /= -17*i) then + fail = .true. + stop 2 + end if + end do + end block + !$acc end serial + if (fail) error stop 2 +end diff --git a/libgomp/testsuite/libgomp.oacc-fortran/lib-13.f90 b/libgomp/testsuite/libgomp.oacc-fortran/lib-13.f90 index deb2c28..f6bd27a 100644 --- a/libgomp/testsuite/libgomp.oacc-fortran/lib-13.f90 +++ b/libgomp/testsuite/libgomp.oacc-fortran/lib-13.f90 @@ -19,11 +19,10 @@ program main end do !$acc end parallel end do - !$acc end data call acc_wait_all_async (nprocs + 1) - call acc_wait (nprocs + 1) + !$acc end data if (acc_async_test (1) .neqv. .TRUE.) stop 1 if (acc_async_test (2) .neqv. .TRUE.) stop 2 |