summaryrefslogtreecommitdiff
path: root/src/tuning/kernels/xgemv.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/tuning/kernels/xgemv.cpp')
-rw-r--r--src/tuning/kernels/xgemv.cpp46
1 files changed, 34 insertions, 12 deletions
diff --git a/src/tuning/kernels/xgemv.cpp b/src/tuning/kernels/xgemv.cpp
index d42155ae..7229602d 100644
--- a/src/tuning/kernels/xgemv.cpp
+++ b/src/tuning/kernels/xgemv.cpp
@@ -29,7 +29,7 @@ class TuneXgemv {
public:
// The representative kernel and the source code
- static std::string KernelFamily() { return "xgemv_"+std::to_string(V); }
+ static std::string KernelFamily() { return (V==1) ? "xgemv" : ((V==2) ? "xgemv_fast" : "xgemv_fast_rot"); }
static std::string KernelName() { return (V==1) ? "Xgemv" : ((V==2) ? "XgemvFast" : "XgemvFastRot"); }
static std::string GetSources() {
return
@@ -61,21 +61,42 @@ class TuneXgemv {
// Sets the tuning parameters and their possible values
static void SetParameters(cltune::Tuner &tuner, const size_t id) {
- tuner.AddParameter(id, "WGS"+std::to_string(V), {64, 128, 256});
- tuner.AddParameter(id, "WPT"+std::to_string(V), {1, 2, 4});
- if (V==2 || V==3) { tuner.AddParameter(id, "VW"+std::to_string(V), {1, 2, 4, 8}); }
+ if (V==1) {
+ tuner.AddParameter(id, "WGS"+std::to_string(V), {32, 64, 128, 256});
+ tuner.AddParameter(id, "WPT"+std::to_string(V), {1, 2, 4});
+ }
+ if (V==2) {
+ tuner.AddParameter(id, "WGS"+std::to_string(V), {16, 32, 64, 128, 256});
+ tuner.AddParameter(id, "WPT"+std::to_string(V), {1, 2, 4});
+ tuner.AddParameter(id, "VW"+std::to_string(V), {1, 2, 4, 8});
+ }
+ if (V==3) {
+ tuner.AddParameter(id, "WGS"+std::to_string(V), {16, 32, 64, 128});
+ tuner.AddParameter(id, "WPT"+std::to_string(V), {1, 2, 4, 8, 16, 32});
+ tuner.AddParameter(id, "VW"+std::to_string(V), {1, 2, 4, 8});
+ }
}
// Sets the constraints and local memory size
static void SetConstraints(cltune::Tuner &tuner, const size_t id) {
- auto MultipleOfX = [] (std::vector<size_t> v) { return IsMultiple(v[0], v[1]); };
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)});
+ }
}
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]*GetBytes(args.precision); };
- tuner.SetLocalMemoryUsage(id, LocalMemorySize, {"WGS"+std::to_string(V)});
+ 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 base thread configuration
@@ -89,20 +110,21 @@ class TuneXgemv {
static TransformVector MulLocal() { return {{"WGS"+std::to_string(V)}}; }
static TransformVector DivLocal() { return {}; }
static TransformVector MulGlobal() { return {}; }
- static TransformVector DivGlobal() { return {{"WPT"+std::to_string(V)}}; }
+ static TransformVector DivGlobal() {
+ if (V==1 || V==2) return {{"WPT"+std::to_string(V)}};
+ return {};
+ }
// 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 alpha_buffer = std::vector<T>{args.alpha};
- auto beta_buffer = std::vector<T>{args.beta};
auto a_rotated = (V==3) ? 1 : 0;
tuner.AddArgumentScalar(static_cast<int>(args.m));
tuner.AddArgumentScalar(static_cast<int>(args.n));
- tuner.AddArgumentInput(alpha_buffer);
- tuner.AddArgumentInput(beta_buffer);
+ tuner.AddArgumentScalar(GetRealArg(args.alpha));
+ tuner.AddArgumentScalar(GetRealArg(args.beta));
tuner.AddArgumentScalar(static_cast<int>(a_rotated));
tuner.AddArgumentInput(a_mat);
tuner.AddArgumentScalar(0);