/// Allocates memory buffer on the device associated with the given queue.
device_vector(const command_queue &q, size_t n) : n(n) {
if (n) {
q.context().set_current();
CUdeviceptr ptr;
cuda_check( cuMemAlloc(&ptr, n * sizeof(T)) );
buffer.reset(reinterpret_cast<char*>(static_cast<size_t>(ptr)), detail::deleter() );
}
}
/// Copies data from host memory to device.
void write(const command_queue &q, size_t offset, size_t size, const T *host,
bool blocking = false) const
{
(void)blocking;
if (size) {
q.context().set_current();
cuda_check( cuMemcpyHtoD(raw() + offset * sizeof(T), host, size * sizeof(T)) );
}
}
/// Copies data from device to host memory.
void read(const command_queue &q, size_t offset, size_t size, T *host,
bool blocking = false) const
{
(void)blocking;
if (size) {
q.context().set_current();
cuda_check( cuMemcpyDtoH(host, raw() + offset * sizeof(T), size * sizeof(T)) );
}
}
hard_event::hard_event(command_queue &q, cl_command_type command,
const ref_vector<event> &deps, action action) :
event(q.context(), deps, profile(q, action), [](event &ev){}),
_queue(q), _command(command), _fence(NULL) {
if (q.profiling_enabled())
_time_queued = timestamp::current(q);
q.sequence(*this);
trigger();
}
/// Constructor. Extracts a backend::kernel instance from backend::program.
kernel(const command_queue &queue,
const program &P,
const std::string &name,
std::function<size_t(size_t)> smem
)
: ctx(queue.context()), P(P), smem(0)
{
cuda_check( cuModuleGetFunction(&K, P.raw(), name.c_str()) );
config(queue, smem);
}
/// Constructor. Creates a backend::kernel instance from source.
kernel(const command_queue &queue,
const std::string &src, const std::string &name,
std::function<size_t(size_t)> smem,
const std::string &options = ""
)
: ctx(queue.context()), P(build_sources(queue, src, options)), smem(0)
{
cuda_check( cuModuleGetFunction(&K, P.raw(), name.c_str()) );
config(queue, smem);
}
/// Constructor. Extracts a backend::kernel instance from backend::program.
kernel(const command_queue &queue,
const program &P,
const std::string &name,
size_t smem_per_thread = 0
)
: ctx(queue.context()), P(P), smem(0)
{
cuda_check( cuModuleGetFunction(&K, P.raw(), name.c_str()) );
config(queue,
[smem_per_thread](size_t wgs){ return wgs * smem_per_thread; });
}
/// Constructor. Creates a backend::kernel instance from source.
kernel(const command_queue &queue,
const std::string &src,
const std::string &name,
size_t smem_per_thread = 0,
const std::string &options = ""
)
: ctx(queue.context()), P(build_sources(queue, src, options)), smem(0)
{
cuda_check( cuModuleGetFunction(&K, P.raw(), name.c_str()) );
config(queue,
[smem_per_thread](size_t wgs){ return wgs * smem_per_thread; });
}
device_vector(const command_queue &q, size_t n,
const H *host = 0, mem_flags flags = MEM_READ_WRITE)
: n(n)
{
(void)flags;
if (n) {
q.context().set_current();
CUdeviceptr ptr;
cuda_check( cuMemAlloc(&ptr, n * sizeof(T)) );
buffer.reset(reinterpret_cast<char*>(static_cast<size_t>(ptr)), detail::deleter() );
if (host) {
if (std::is_same<T, H>::value)
write(q, 0, n, reinterpret_cast<const T*>(host), true);
else
write(q, 0, n, std::vector<T>(host, host + n).data(), true);
}
}
}
/// Create and build a program from source string.
inline vex::backend::program build_sources(
const command_queue &queue, const std::string &source,
const std::string &options = ""
)
{
#ifdef VEXCL_SHOW_KERNELS
std::cout << source << std::endl;
#else
if (getenv("VEXCL_SHOW_KERNELS"))
std::cout << source << std::endl;
#endif
std::string compile_options = options + " " + get_compile_options(queue);
queue.context().set_current();
auto cc = queue.device().compute_capability();
std::ostringstream ccstr;
ccstr << std::get<0>(cc) << std::get<1>(cc);
sha1_hasher sha1;
sha1.process(source)
.process(queue.device().name())
.process(compile_options)
.process(ccstr.str())
;
std::string hash = static_cast<std::string>(sha1);
// Write source to a .cu file
std::string basename = program_binaries_path(hash, true) + "kernel";
std::string ptxfile = basename + ".ptx";
if ( !boost::filesystem::exists(ptxfile) ) {
std::string cufile = basename + ".cu";
{
std::ofstream f(cufile);
f << source;
}
// Compile the source to ptx.
std::ostringstream cmdline;
cmdline
<< "nvcc -ptx -O3"
<< " -arch=sm_" << std::get<0>(cc) << std::get<1>(cc)
<< " " << compile_options
<< " -o " << ptxfile << " " << cufile;
if (0 != system(cmdline.str().c_str()) ) {
#ifndef VEXCL_SHOW_KERNELS
std::cerr << source << std::endl;
#endif
vex::detail::print_backtrace();
throw std::runtime_error("nvcc invocation failed");
}
}
// Load the compiled ptx.
CUmodule prg;
cuda_check( cuModuleLoad(&prg, ptxfile.c_str()) );
return program(queue.context(), prg);
}
/// Create command queue on the same context and device as the given one.
inline command_queue duplicate_queue(const command_queue &q) {
return command_queue(q.context(), q.device(), q.flags());
}
/// Returns context for the given queue.
inline context get_context(const command_queue &q) {
return q.context();
}
/// Returns raw context id for the given queue.
inline context_id get_context_id(const command_queue &q) {
return q.context().raw();
}
/// Binds the specified CUDA context to the calling CPU thread.
inline void select_context(const command_queue &q) {
q.context().set_current();
}