bool add(key_t key, const add_t &val, val_t *_oval = 0) {
bits_t oval = data[key];
val_t nval = val_t(oval) + val;
while(true) {
bits_t noval = atomic.cas(&data[key], oval, nval.bits());
if(noval == oval) {
if(_oval)
*_oval = oval;
return true;
}
oval = noval;
nval = val_t(oval) + val;
}
return true;
}
static void get(backend::source_generator &src,
const ccsr_product<val_t, col_t, idx_t, T>&,
const backend::command_queue&, const std::string &prm_name,
detail::kernel_generator_state_ptr)
{
typedef decltype(val_t() * T()) res_t;
src.function<res_t>(prm_name + "_spmv")
.open("(")
.template parameter< global_ptr<const idx_t> >("idx")
.template parameter< global_ptr<const idx_t> >("row")
.template parameter< global_ptr<const col_t> >("col")
.template parameter< global_ptr<const val_t> >("val")
.template parameter< global_ptr<const T> >("vec")
.template parameter< size_t >("i")
.close(")").open("{");
src.new_line() << type_name<res_t>() << " sum = 0;";
src.new_line() << "for(size_t pos = idx[i], j = row[pos], end = row[pos+1]; j < end; ++j)";
src.open("{");
src.new_line() << "sum += val[j] * vec[i + col[j]];";
src.close("}");
src.new_line() << "return sum;";
src.close("}");
}
val_t get_val() const { return val_t(val); }