1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
|
// =================================================================================================
// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This
// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max-
// width of 100 characters per line.
//
// Author(s):
// Cedric Nugteren <www.cedricnugteren.nl>
//
// This file uses the 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.
//
// =================================================================================================
#include <string>
#include <vector>
#include "utilities/utilities.hpp"
#include "tuning/tuning.hpp"
namespace clblast {
// =================================================================================================
// See comment at top of file for a description of the class
template <typename T, int V>
class TuneXgemm {
public:
// Settings for this kernel (default command-line arguments)
static TunerDefaults GetTunerDefaults() {
auto settings = TunerDefaults();
settings.options = {kArgM, kArgN, kArgK, kArgAlpha, kArgBeta, kArgFraction,
kArgHeuristicSelection, kArgPsoSwarmSize,
kArgPsoInfGlobal, kArgPsoInfLocal, kArgPsoInfRandom};
settings.default_m = 1024;
settings.default_n = 1024;
settings.default_k = 1024;
settings.default_fraction = (V==1) ? 1.0 : 512.0; // test all or sample randomly
settings.default_num_runs = 2;
return settings;
}
// Settings for this kernel (general)
static TunerSettings GetTunerSettings(const Arguments<T> &args) {
auto settings = TunerSettings();
// Identification of the kernel
settings.kernel_family = (V==1) ? "xgemm_1" : "xgemm_2";
settings.kernel_name = "Xgemm";
settings.sources =
#include "../src/kernels/level3/xgemm_part1.opencl"
#include "../src/kernels/level3/xgemm_part2.opencl"
#include "../src/kernels/level3/xgemm_part3.opencl"
;
// Buffer sizes
settings.size_a = args.m * args.k;
settings.size_b = args.n * args.k;
settings.size_c = args.m * args.n;
// Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5)
settings.inputs = {2, 3, 4};
settings.outputs = {4};
// Sets the base thread configuration
settings.global_size = {args.m, args.n};
settings.global_size_ref = settings.global_size;
settings.local_size = {1, 1};
settings.local_size_ref = {8, 8};
// Transforms the thread configuration based on the parameters
settings.mul_local = {{"MDIMC", "NDIMC"}};
settings.mul_global = {{"MDIMC", "NDIMC"}};
settings.div_global = {{"MWG", "NWG"}};
// Sets the tuning parameters and their possible values
if (V==1) { // limited subset of tuning parameters - but explorable exhaustively
settings.parameters = {
{"MWG", {16, 32, 64}},
{"NWG", {16, 32, 64}},
{"KWG", {32}},
{"MDIMC", {8, 16, 32}},
{"NDIMC", {8, 16, 32}},
{"MDIMA", {8, 16, 32}},
{"NDIMB", {8, 16, 32}},
{"KWI", {2}},
{"VWM", {1, 2, 4}},
{"VWN", {1, 2, 4}},
{"STRM", {0}},
{"STRN", {0}},
{"SA", {0, 1}},
{"SB", {0, 1}},
};
}
else { // a lot more tuning parameters - has to be sampled randomly, too much to test all
settings.parameters = {
{"MWG", {16, 32, 64, 128}},
{"NWG", {16, 32, 64, 128}},
{"KWG", {16, 32}},
{"MDIMC", {8, 16, 32}},
{"NDIMC", {8, 16, 32}},
{"MDIMA", {8, 16, 32}},
{"NDIMB", {8, 16, 32}},
{"KWI", {2}},
{"VWM", {1, 2, 4, 8}},
{"VWN", {1, 2, 4, 8}},
{"STRM", {0, 1}},
{"STRN", {0, 1}},
{"SA", {0, 1}},
{"SB", {0, 1}},
};
}
// Describes how to compute the performance metrics
settings.metric_amount = 2 * args.m * args.n * args.k;
settings.performance_unit = "GFLOPS";
return settings;
}
// Tests for valid arguments
static void TestValidArguments(const Arguments<T> &) { }
static std::vector<Constraint> SetConstraints() {
auto constraints = std::vector<Constraint>();
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 KWG loop
constraints.push_back({MultipleOfX, {"KWG", "KWI"}});
// Required for integer MWI and NWI
constraints.push_back({MultipleOfXMulY, {"MWG", "MDIMC", "VWM"}});
constraints.push_back({MultipleOfXMulY, {"NWG", "NDIMC", "VWN"}});
// Required for integer MWIA and NWIB
constraints.push_back({MultipleOfXMulY, {"MWG", "MDIMA", "VWM"}});
constraints.push_back({MultipleOfXMulY, {"NWG", "NDIMB", "VWN"}});
// KWG has to be a multiple of KDIMA = ((MDIMC*NDIMC)/(MDIMA)) and KDIMB = (...)
constraints.push_back({MultipleOfXMulYDivZ, {"KWG", "MDIMC", "NDIMC", "MDIMA"}});
constraints.push_back({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]; };
constraints.push_back({IsEqual, {"MDIMC", "MDIMA"}});
constraints.push_back({IsEqual, {"NDIMC", "NDIMB"}});
constraints.push_back({IsEqual, {"SA", "SB"}});
}
return constraints;
}
// Sets the kernel's arguments
static void SetArguments(Kernel &kernel, const Arguments<T> &args,
std::vector<Buffer<T>>& buffers) {
kernel.SetArgument(0, static_cast<int>(args.m));
kernel.SetArgument(1, static_cast<int>(args.n));
kernel.SetArgument(2, static_cast<int>(args.k));
kernel.SetArgument(3, GetRealArg(args.alpha));
kernel.SetArgument(4, GetRealArg(args.beta));
kernel.SetArgument(5, buffers[2]()); // 2 == A matrix
kernel.SetArgument(6, buffers[3]()); // 3 == B matrix
kernel.SetArgument(7, buffers[4]()); // 4 == C matrix
kernel.SetArgument(8, 0);
kernel.SetArgument(9, 0);
}
};
// =================================================================================================
} // namespace clblast
// Shortcuts to the clblast namespace
using half = clblast::half;
using float2 = clblast::float2;
using double2 = clblast::double2;
// Function to tune a specific variation V (not within the clblast namespace)
template <int V>
void StartVariation(int argc, char *argv[]) {
const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
switch(clblast::GetPrecision(command_line_args)) {
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;
}
// =================================================================================================
|
