summaryrefslogtreecommitdiff
path: root/src/tuning
diff options
context:
space:
mode:
Diffstat (limited to 'src/tuning')
-rw-r--r--src/tuning/kernels/copy_fast.cpp3
-rw-r--r--src/tuning/kernels/copy_pad.cpp3
-rw-r--r--src/tuning/kernels/transpose_fast.cpp3
-rw-r--r--src/tuning/kernels/transpose_pad.cpp3
-rw-r--r--src/tuning/kernels/xaxpy.cpp3
-rw-r--r--src/tuning/kernels/xgemm.cpp92
-rw-r--r--src/tuning/kernels/xgemv.cpp46
-rw-r--r--src/tuning/kernels/xger.cpp3
8 files changed, 103 insertions, 53 deletions
diff --git a/src/tuning/kernels/copy_fast.cpp b/src/tuning/kernels/copy_fast.cpp
index 34269bc7..78ded56e 100644
--- a/src/tuning/kernels/copy_fast.cpp
+++ b/src/tuning/kernels/copy_fast.cpp
@@ -86,11 +86,10 @@ class TuneCopy {
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);
+ tuner.AddArgumentScalar(GetRealArg(args.alpha));
}
// Describes how to compute the performance metrics
diff --git a/src/tuning/kernels/copy_pad.cpp b/src/tuning/kernels/copy_pad.cpp
index 1e0dccd3..90f5ea82 100644
--- a/src/tuning/kernels/copy_pad.cpp
+++ b/src/tuning/kernels/copy_pad.cpp
@@ -86,7 +86,6 @@ class TunePad {
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.AddArgumentScalar(static_cast<int>(args.n));
tuner.AddArgumentScalar(static_cast<int>(args.m));
@@ -97,7 +96,7 @@ class TunePad {
tuner.AddArgumentScalar(static_cast<int>(args.m));
tuner.AddArgumentScalar(0);
tuner.AddArgumentOutput(b_mat);
- tuner.AddArgumentInput(alpha_buffer);
+ tuner.AddArgumentScalar(GetRealArg(args.alpha));
tuner.AddArgumentScalar(0);
}
diff --git a/src/tuning/kernels/transpose_fast.cpp b/src/tuning/kernels/transpose_fast.cpp
index 7ac19cb6..10fa80cb 100644
--- a/src/tuning/kernels/transpose_fast.cpp
+++ b/src/tuning/kernels/transpose_fast.cpp
@@ -91,11 +91,10 @@ class TuneTranspose {
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);
+ tuner.AddArgumentScalar(GetRealArg(args.alpha));
}
// Describes how to compute the performance metrics
diff --git a/src/tuning/kernels/transpose_pad.cpp b/src/tuning/kernels/transpose_pad.cpp
index 63274415..507718eb 100644
--- a/src/tuning/kernels/transpose_pad.cpp
+++ b/src/tuning/kernels/transpose_pad.cpp
@@ -90,7 +90,6 @@ class TunePadTranspose {
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.AddArgumentScalar(static_cast<int>(args.n));
tuner.AddArgumentScalar(static_cast<int>(args.m));
@@ -101,7 +100,7 @@ class TunePadTranspose {
tuner.AddArgumentScalar(static_cast<int>(args.n));
tuner.AddArgumentScalar(0);
tuner.AddArgumentOutput(b_mat);
- tuner.AddArgumentInput(alpha_buffer);
+ tuner.AddArgumentScalar(GetRealArg(args.alpha));
tuner.AddArgumentScalar(0);
}
diff --git a/src/tuning/kernels/xaxpy.cpp b/src/tuning/kernels/xaxpy.cpp
index 88d12c1f..0033b3c6 100644
--- a/src/tuning/kernels/xaxpy.cpp
+++ b/src/tuning/kernels/xaxpy.cpp
@@ -89,9 +89,8 @@ class TuneXaxpy {
std::vector<T> &x_vec, std::vector<T> &y_vec,
std::vector<T> &, std::vector<T> &, std::vector<T> &,
std::vector<T> &) {
- auto alpha_buffer = std::vector<T>{args.alpha};
tuner.AddArgumentScalar(static_cast<int>(args.n));
- tuner.AddArgumentInput(alpha_buffer);
+ tuner.AddArgumentScalar(GetRealArg(args.alpha));
tuner.AddArgumentInput(x_vec);
tuner.AddArgumentOutput(y_vec);
}
diff --git a/src/tuning/kernels/xgemm.cpp b/src/tuning/kernels/xgemm.cpp
index 4b1efdef..4cb7fd00 100644
--- a/src/tuning/kernels/xgemm.cpp
+++ b/src/tuning/kernels/xgemm.cpp
@@ -7,7 +7,9 @@
// Author(s):
// Cedric Nugteren <www.cedricnugteren.nl>
//
-// This file uses the CLTune auto-tuner to tune the xgemm OpenCL kernels.
+// This file uses the CLTune auto-tuner to tune the xgemm OpenCL kernels. There are two variations:
+// - V==1: This tests some limited set of tuning parameters exhaustively.
+// - V==2: This tests a much larger set of tuning parameters by randomly sampling a subset.
//
// =================================================================================================
@@ -21,18 +23,19 @@ namespace clblast {
// =================================================================================================
// See comment at top of file for a description of the class
-template <typename T>
+template <typename T, int V>
class TuneXgemm {
public:
// The representative kernel and the source code
- static std::string KernelFamily() { return "xgemm"; }
+ static std::string KernelFamily() { return (V==1) ? "xgemm_1" : "xgemm_2"; }
static std::string KernelName() { return "Xgemm"; }
static std::string GetSources() {
return
#include "../src/kernels/common.opencl"
#include "../src/kernels/level3/xgemm_part1.opencl"
#include "../src/kernels/level3/xgemm_part2.opencl"
+ #include "../src/kernels/level3/xgemm_part3.opencl"
;
}
@@ -48,7 +51,7 @@ class TuneXgemm {
static size_t DefaultM() { return 1024; }
static size_t DefaultN() { return 1024; }
static size_t DefaultK() { return 1024; }
- static double DefaultFraction() { return 2048.0; }
+ static double DefaultFraction() { return (V==1) ? 1.0 : 512.0; } // test all or sample randomly
// Describes how to obtain the sizes of the buffers
static size_t GetSizeX(const Arguments<T> &) { return 1; } // N/A for this kernel
@@ -60,20 +63,38 @@ class TuneXgemm {
// Sets the tuning parameters and their possible values
static void SetParameters(cltune::Tuner &tuner, const size_t id) {
- tuner.AddParameter(id, "MWG", {16, 32, 64, 128});
- tuner.AddParameter(id, "NWG", {16, 32, 64, 128});
- tuner.AddParameter(id, "KWG", {16, 32});
- tuner.AddParameter(id, "MDIMC", {8, 16, 32});
- tuner.AddParameter(id, "NDIMC", {8, 16, 32});
- tuner.AddParameter(id, "MDIMA", {8, 16, 32});
- tuner.AddParameter(id, "NDIMB", {8, 16, 32});
- tuner.AddParameter(id, "KWI", {2, 8});
- tuner.AddParameter(id, "VWM", {1, 2, 4, 8});
- tuner.AddParameter(id, "VWN", {1, 2, 4, 8});
- tuner.AddParameter(id, "STRM", {0, 1});
- tuner.AddParameter(id, "STRN", {0, 1});
- tuner.AddParameter(id, "SA", {0, 1});
- tuner.AddParameter(id, "SB", {0, 1});
+ if (V==1) { // limited subset of tuning parameters - but explorable exhaustively
+ tuner.AddParameter(id, "MWG", {16, 32, 64});
+ tuner.AddParameter(id, "NWG", {16, 32, 64});
+ tuner.AddParameter(id, "KWG", {32});
+ tuner.AddParameter(id, "MDIMC", {8, 16, 32});
+ tuner.AddParameter(id, "NDIMC", {8, 16, 32});
+ tuner.AddParameter(id, "MDIMA", {8, 16, 32});
+ tuner.AddParameter(id, "NDIMB", {8, 16, 32});
+ tuner.AddParameter(id, "KWI", {2});
+ tuner.AddParameter(id, "VWM", {1, 2, 4});
+ tuner.AddParameter(id, "VWN", {1, 2, 4});
+ tuner.AddParameter(id, "STRM", {0});
+ tuner.AddParameter(id, "STRN", {0});
+ tuner.AddParameter(id, "SA", {0, 1});
+ tuner.AddParameter(id, "SB", {0, 1});
+ } // a lot more tuning parameters - has to be sampled randomly, too much to test all
+ else {
+ tuner.AddParameter(id, "MWG", {16, 32, 64, 128});
+ tuner.AddParameter(id, "NWG", {16, 32, 64, 128});
+ tuner.AddParameter(id, "KWG", {16, 32});
+ tuner.AddParameter(id, "MDIMC", {8, 16, 32});
+ tuner.AddParameter(id, "NDIMC", {8, 16, 32});
+ tuner.AddParameter(id, "MDIMA", {8, 16, 32});
+ tuner.AddParameter(id, "NDIMB", {8, 16, 32});
+ tuner.AddParameter(id, "KWI", {2});
+ tuner.AddParameter(id, "VWM", {1, 2, 4, 8});
+ tuner.AddParameter(id, "VWN", {1, 2, 4, 8});
+ tuner.AddParameter(id, "STRM", {0, 1});
+ tuner.AddParameter(id, "STRN", {0, 1});
+ tuner.AddParameter(id, "SA", {0, 1});
+ tuner.AddParameter(id, "SB", {0, 1});
+ }
}
// Sets the constraints
@@ -92,6 +113,14 @@ class TuneXgemm {
// KWG has to be a multiple of KDIMA = ((MDIMC*NDIMC)/(MDIMA)) and KDIMB = (...)
tuner.AddConstraint(id, MultipleOfXMulYDivZ, {"KWG", "MDIMC", "NDIMC", "MDIMA"});
tuner.AddConstraint(id, MultipleOfXMulYDivZ, {"KWG", "MDIMC", "NDIMC", "NDIMB"});
+
+ // 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, {"MDIMC", "MDIMA"});
+ tuner.AddConstraint(id, IsEqual, {"NDIMC", "NDIMB"});
+ tuner.AddConstraint(id, IsEqual, {"SA", "SB"});
+ }
}
// Sets the local memory size
@@ -121,13 +150,11 @@ class TuneXgemm {
std::vector<T> &, std::vector<T> &,
std::vector<T> &a_mat, std::vector<T> &b_mat, std::vector<T> &c_mat,
std::vector<T> &) {
- auto alpha_buffer = std::vector<T>{args.alpha};
- auto beta_buffer = std::vector<T>{args.beta};
tuner.AddArgumentScalar(static_cast<int>(args.m));
tuner.AddArgumentScalar(static_cast<int>(args.n));
tuner.AddArgumentScalar(static_cast<int>(args.k));
- tuner.AddArgumentInput(alpha_buffer);
- tuner.AddArgumentInput(beta_buffer);
+ tuner.AddArgumentScalar(GetRealArg(args.alpha));
+ tuner.AddArgumentScalar(GetRealArg(args.beta));
tuner.AddArgumentInput(a_mat);
tuner.AddArgumentInput(b_mat);
tuner.AddArgumentOutput(c_mat);
@@ -147,15 +174,22 @@ class TuneXgemm {
using float2 = clblast::float2;
using double2 = clblast::double2;
-// Main function (not within the clblast namespace)
-int main(int argc, char *argv[]) {
+// Function to tune a specific variation V (not within the clblast namespace)
+template <int V>
+void StartVariation(int argc, char *argv[]) {
switch(clblast::GetPrecision(argc, argv)) {
- case clblast::Precision::kHalf: clblast::Tuner<clblast::TuneXgemm<half>, half>(argc, argv); break;
- case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXgemm<float>, float>(argc, argv); break;
- case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXgemm<double>, double>(argc, argv); break;
- case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXgemm<float2>, float2>(argc, argv); break;
- case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXgemm<double2>, double2>(argc, argv); break;
+ case clblast::Precision::kHalf: clblast::Tuner<clblast::TuneXgemm<half,V>, half>(argc, argv); break;
+ case clblast::Precision::kSingle: clblast::Tuner<clblast::TuneXgemm<float,V>, float>(argc, argv); break;
+ case clblast::Precision::kDouble: clblast::Tuner<clblast::TuneXgemm<double,V>, double>(argc, argv); break;
+ case clblast::Precision::kComplexSingle: clblast::Tuner<clblast::TuneXgemm<float2,V>, float2>(argc, argv); break;
+ case clblast::Precision::kComplexDouble: clblast::Tuner<clblast::TuneXgemm<double2,V>, double2>(argc, argv); break;
}
+}
+
+// Main function (not within the clblast namespace)
+int main(int argc, char *argv[]) {
+ StartVariation<1>(argc, argv);
+ StartVariation<2>(argc, argv);
return 0;
}
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);
diff --git a/src/tuning/kernels/xger.cpp b/src/tuning/kernels/xger.cpp
index d2590c53..1fb5c531 100644
--- a/src/tuning/kernels/xger.cpp
+++ b/src/tuning/kernels/xger.cpp
@@ -85,10 +85,9 @@ class TuneXger {
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};
tuner.AddArgumentScalar(static_cast<int>(args.m));
tuner.AddArgumentScalar(static_cast<int>(args.n));
- tuner.AddArgumentInput(alpha_buffer);
+ tuner.AddArgumentScalar(GetRealArg(args.alpha));
tuner.AddArgumentInput(x_vec);
tuner.AddArgumentScalar(0); // x_offset
tuner.AddArgumentScalar(1); // x_increment
generated by cgit v1.2.3 (git 2.39.1) at 2024-08-01 08:04:42 +0000