buffer::buffer(driver::Context const & context, std::string const & scalartype, unsigned int id, const tuple &shape, tuple const & strides) : array(context, scalartype, id), dim_(numgt1(shape))
{
//Attributes
attributes_["off"] = process("#name_off");
for(unsigned int i = 0 ; i < dim_ ; ++i){
std::string inc = "inc" + tools::to_string(i);
attributes_[inc] = process("#name_" + inc);
}
//Access
std::vector<std::string> args;
for(unsigned int i = 0 ; i < dim_ ; ++i)
args.push_back("x" + tools::to_string(i));
std::string off = "#off";
for(unsigned int i = 0 ; i < dim_ ; ++i)
{
std::string inc = "#inc" + tools::to_string(i);
off += " + (" + args[i] + ")*" + inc;
}
macros_.insert("at(" + tools::join(args, ",") + "): #pointer[" + off + "]");
//Broadcast
if(numgt1(shape)==0)
macros_.insert("at(i): at()");
if(dim_!=shape.size())
macros_.insert(make_broadcast(shape));
add_base("buffer");
add_load(strides[0]==1 && shape[0]>1);
}
placeholder::placeholder(driver::Context const & context, unsigned int level) : leaf(context, "int", "sforidx" + tools::to_string(level))
{
macros_.insert("at(): #name");
macros_.insert("at(i): #name");
macros_.insert("at(i,j): #name");
add_base("placebolder");
add_load(false);
}
host_scalar::host_scalar(driver::Context const & context, std::string const & scalartype, unsigned int id) : leaf(context, scalartype, id)
{
macros_.insert("at(): #name_value");
macros_.insert("at(i): #name_value");
macros_.insert("at(i,j): #name_value");
add_base("host_scalar");
add_load(false);
}
index_modifier::index_modifier(const std::string &scalartype, unsigned int id, size_t root, op_element op, expression_tree const & tree, symbols_table const & table) : array(tree.context(), scalartype, id), node(root, op, tree, table)
{
add_base("index_modifier");
add_load(false);
}
int main(int argc, char *argv[])
{
int load = 0;
int k = 3;
double epsilon = 0.001;
int maxsamples = 100;
int minparam = 11;
int pcount = 14;
int lcount = 1;
int clear_cache = 0;
int tod = 0;
double slope, intercept;
int i,p,bcount;
int cpubench = 0;
int compensate = 0;
int sleeptime = 10;
add_int_option("load", &load);
add_int_option("k", &k);
add_double_option("epsilon", &epsilon);
add_int_option("maxsamples", &maxsamples);
add_int_option("minparam", &minparam);
add_int_option("clear", &clear_cache);
add_int_option("pcount", &pcount);
add_int_option("lcount", &lcount);
add_int_option("cpubench", &cpubench);
add_int_option("tod", &tod);
add_int_option("compensate", &compensate);
add_int_option("sleeptime", &sleeptime);
parse_options(argc, argv, NULL);
show_options(stdout);
Mhz = mhz_full(1,sleeptime);
/* Do callibrations */
find_abs_performance(&intercept, &slope, clear_cache, 1, cpubench);
add_load(load, CACHE_LOAD);
p = minparam;
bcount = lcount;
for (i = 0; i < pcount; i++)
{
double cycs;
double expected = slope*p + intercept;
double error;
if (tod) {
if (cpubench)
cycs = fcyc_full_tod(cpufunct, p, clear_cache, k, epsilon, maxsamples, compensate);
else {
cycs = fcyc_full_tod(funct, p, clear_cache, k, epsilon, maxsamples, compensate);
}
} else {
if (cpubench)
cycs = fcyc_full(cpufunct, p, clear_cache,
k, epsilon, maxsamples, compensate);
else {
cycs = fcyc_full(funct, p, clear_cache,
k, epsilon, maxsamples, compensate);
}
}
printf("Iters= %d, Err= %0.4f, Cycs= %.0f, ms= %.3f",
has_converged(k, epsilon, maxsamples), err(k), cycs,
cycs / (1e3 * Mhz));
error = (cycs-expected)/expected;
printf(", Exp-cyc= %.1f, Exp-ms= %.3f, Actual-Err= %f\n", expected,
expected / (Mhz * 1e3), error);
p += minparam;
bcount--;
if (bcount == 0) {
bcount = lcount;
minparam += 2;
}
}
kill_loads();
return 0;
}