static void SetLocalMemorySize(cltune::Tuner &tuner, const size_t id, const Arguments<T> &args) {
if (V==1 || V==2) {
auto LocalMemorySize = [args] (std::vector<size_t> v) { return v[0]*GetBytes(args.precision); };
tuner.SetLocalMemoryUsage(id, LocalMemorySize, {"WGS"+std::to_string(V)});
}
else {
auto LocalMemorySize = [args] (std::vector<size_t> v) { return (v[0]*v[1] + v[1])*GetBytes(args.precision); };
tuner.SetLocalMemoryUsage(id, LocalMemorySize, {"WGS"+std::to_string(V), "WPT"+std::to_string(V)});
}
}
// Sets the constraints and local memory size
static void SetConstraints(cltune::Tuner &tuner, const size_t id) {
if (V==2 || V==3) {
auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
tuner.AddConstraint(id, MultipleOfX, {"WPT"+std::to_string(V), "VW"+std::to_string(V)});
}
if (V==3) {
auto LargerOrEqual = [] (std::vector<size_t> v) { return v[0] >= v[1]; };
tuner.AddConstraint(id, LargerOrEqual, {"WGS"+std::to_string(V), "WPT"+std::to_string(V)});
}
}
// Sets the kernel's arguments
static void SetArguments(cltune::Tuner &tuner, const Arguments<T> &args,
std::vector<T> &, std::vector<T> &,
std::vector<T> &a_mat, std::vector<T> &b_mat, std::vector<T> &,
std::vector<T> &) {
auto alpha_buffer = std::vector<T>{args.alpha};
tuner.AddArgumentScalar(static_cast<int>(args.m));
tuner.AddArgumentInput(a_mat);
tuner.AddArgumentOutput(b_mat);
tuner.AddArgumentInput(alpha_buffer);
}
// Sets the constraints
static void SetConstraints(cltune::Tuner &tuner, const size_t id) {
auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
auto MultipleOfXMulY = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]*v[2]); };
auto MultipleOfXMulYDivZ = [] (std::vector<size_t> v) { return IsMultiple(v[0], (v[1]*v[2])/v[3]); };
// Requirement for unrolling the WGD loop
tuner.AddConstraint(id, MultipleOfX, {"WGD", "KWID"});
// Required for integer MWID and NWID
tuner.AddConstraint(id, MultipleOfXMulY, {"WGD", "MDIMCD", "VWMD"});
tuner.AddConstraint(id, MultipleOfXMulY, {"WGD", "NDIMCD", "VWND"});
// Required for integer MWIAD and NWIBD
tuner.AddConstraint(id, MultipleOfXMulY, {"WGD", "MDIMAD", "VWMD"});
tuner.AddConstraint(id, MultipleOfXMulY, {"WGD", "NDIMBD", "VWND"});
// WGD has to be a multiple of KDIMAD = ((MDIMCD*NDIMCD)/(MDIMAD)) and KDIMBD = (...)
tuner.AddConstraint(id, MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "MDIMAD"});
tuner.AddConstraint(id, MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "NDIMBD"});
// Extra constraints for variation 1 to limit the set of options significantly
if (V==1) {
auto IsEqual = [] (std::vector<size_t> v) { return v[0] == v[1]; };
tuner.AddConstraint(id, IsEqual, {"MDIMCD", "MDIMAD"});
tuner.AddConstraint(id, IsEqual, {"NDIMCD", "NDIMBD"});
}
}
// Sets the tuning parameters and their possible values
static void SetParameters(cltune::Tuner &tuner, const size_t id) {
if (V==1) { // limited subset of tuning parameters - but explorable exhaustively
tuner.AddParameter(id, "WGD", {8, 16, 32});//64,128
tuner.AddParameter(id, "MDIMCD", {4, 8, 16});
tuner.AddParameter(id, "NDIMCD", {4, 8, 16});
tuner.AddParameter(id, "MDIMAD", {4, 8, 16});
tuner.AddParameter(id, "NDIMBD", {4, 8, 16});
tuner.AddParameter(id, "KWID", {1, 2, 4}); //1,4
tuner.AddParameter(id, "VWMD", {1, 2, 4});
tuner.AddParameter(id, "VWND", {1, 2, 4});
tuner.AddParameter(id, "PADA", {1});//0
tuner.AddParameter(id, "PADB", {1});//0
} // a lot more tuning parameters - has to be sampled randomly, too much to test all
else {
tuner.AddParameter(id, "WGD", {8, 16, 32, 64, 128});
tuner.AddParameter(id, "MDIMCD", {8, 16, 32});
tuner.AddParameter(id, "NDIMCD", {8, 16, 32});
tuner.AddParameter(id, "MDIMAD", {8, 16, 32});
tuner.AddParameter(id, "NDIMBD", {8, 16, 32});
tuner.AddParameter(id, "KWID", {2, 8, 16});
tuner.AddParameter(id, "VWMD", {1, 2, 4, 8});
tuner.AddParameter(id, "VWND", {1, 2, 4, 8});
tuner.AddParameter(id, "PADA", {0, 1});
tuner.AddParameter(id, "PADB", {0, 1});
}
}
// Sets the kernel's arguments
static void SetArguments(cltune::Tuner &tuner, const Arguments<T> &args,
std::vector<T> &, std::vector<T> &,
std::vector<T> &a_mat, std::vector<T> &b_mat, std::vector<T> &c_mat,
std::vector<T> &) {
tuner.AddArgumentScalar(static_cast<int>(args.m));
tuner.AddArgumentScalar(static_cast<int>(args.n));
tuner.AddArgumentScalar(static_cast<int>(args.k));
tuner.AddArgumentScalar(GetRealArg(args.alpha));
tuner.AddArgumentScalar(GetRealArg(args.beta));
tuner.AddArgumentInput(a_mat);
tuner.AddArgumentScalar(0); // a_offset
tuner.AddArgumentScalar(static_cast<int>(args.k)); // a_ld
tuner.AddArgumentInput(b_mat);
tuner.AddArgumentScalar(0); // b_offset
tuner.AddArgumentScalar(static_cast<int>(args.n)); // b_ld
tuner.AddArgumentOutput(c_mat);
tuner.AddArgumentScalar(0); // c_offset
tuner.AddArgumentScalar(static_cast<int>(args.n)); // c_ld
tuner.AddArgumentScalar(1); // c_do_transpose
tuner.AddArgumentScalar(0); // a_conjugate
tuner.AddArgumentScalar(0); // b_conjugate
}
// Sets the local memory size
static void SetLocalMemorySize(cltune::Tuner &tuner, const size_t id, const Arguments<T> &args) {
auto LocalMemorySize = [args] (std::vector<size_t> v) {
return ((v[0]*(v[0] + v[1]) + v[0]*(v[0] + v[2]))*GetBytes(args.precision));
};
tuner.SetLocalMemoryUsage(id, LocalMemorySize, {"WGD", "PADA", "PADB"});
}
// Sets the kernel's arguments
static void SetArguments(cltune::Tuner &tuner, const Arguments<T> &args,
std::vector<T> &x_vec, std::vector<T> &y_vec,
std::vector<T> &a_mat, std::vector<T> &, std::vector<T> &,
std::vector<T> &) {
auto a_rotated = (V==3) ? 1 : 0;
tuner.AddArgumentScalar(static_cast<int>(args.m));
tuner.AddArgumentScalar(static_cast<int>(args.n));
tuner.AddArgumentScalar(GetRealArg(args.alpha));
tuner.AddArgumentScalar(GetRealArg(args.beta));
tuner.AddArgumentScalar(static_cast<int>(a_rotated));
tuner.AddArgumentInput(a_mat);
tuner.AddArgumentScalar(0);
tuner.AddArgumentScalar(static_cast<int>(args.m));
tuner.AddArgumentInput(x_vec);
tuner.AddArgumentScalar(0);
tuner.AddArgumentScalar(1);
tuner.AddArgumentOutput(y_vec);
tuner.AddArgumentScalar(0);
tuner.AddArgumentScalar(1);
tuner.AddArgumentScalar(0); // Conjugate transpose
tuner.AddArgumentScalar(0); // Additional parameter
tuner.AddArgumentScalar(0); // Banded 'kl'
tuner.AddArgumentScalar(0); // Banded 'ku'
}
// Sets the tuning parameters and their possible values
static void SetParameters(cltune::Tuner &tuner, const size_t id) {
tuner.AddParameter(id, "TRA_DIM", {4, 8, 16, 32, 64});
tuner.AddParameter(id, "TRA_WPT", {1, 2, 4, 8, 16});
tuner.AddParameter(id, "TRA_PAD", {0, 1});
tuner.AddParameter(id, "TRA_SHUFFLE", {0, 1});
}