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
|
// =================================================================================================
// 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 tunes the Xgemm routine at a high-level: choosing between the direct (single-kernel)
// and the in-direct (kernel plus pre/post-processing) methods.
//
// =================================================================================================
#include <exception>
#include <string>
#include <vector>
#include <iostream>
#include "utilities/utilities.hpp"
#include "test/test_utilities.hpp"
#include "tuning/routines/routine_tuner.hpp"
namespace clblast {
// =================================================================================================
template <typename T>
void RunGemmRoutine(const size_t value, const Queue& queue, const std::vector<Buffer<T>>& buffers) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = Gemm(Layout::kRowMajor, Transpose::kNo, Transpose::kNo,
value, value, value, ConstantOne<T>(),
buffers[0](), 0, value,
buffers[1](), 0, value, ConstantOne<T>(),
buffers[2](), 0, value,
&queue_plain, &event);
if (status != StatusCode::kSuccess) {
throw RuntimeError("Gemm failed with status " + ToString(status));
}
clWaitForEvents(1, &event);
clReleaseEvent(event);
}
template <typename T, size_t batch_count>
void RunGemmBatchedRoutine(const size_t value, const Queue& queue, const std::vector<Buffer<T>>& buffers) {
auto offsets = std::vector<size_t>(batch_count);
auto factors = std::vector<T>(batch_count);
for (auto i = size_t{0}; i < batch_count; ++i) {
offsets[i] = batch_count * value;
factors[i] = ConstantOne<T>();
}
auto queue_plain = queue();
auto event = cl_event{};
auto status = GemmBatched(Layout::kRowMajor, Transpose::kNo, Transpose::kNo,
value, value, value, factors.data(),
buffers[0](), offsets.data(), value,
buffers[1](), offsets.data(), value, factors.data(),
buffers[2](), offsets.data(), value, batch_count,
&queue_plain, &event);
if (status != StatusCode::kSuccess) {
throw RuntimeError("GemmBatched failed with status " + ToString(status));
}
clWaitForEvents(1, &event);
clReleaseEvent(event);
}
template <typename T, size_t batch_count>
void RunGemmStridedBatchedRoutine(const size_t value, const Queue& queue, const std::vector<Buffer<T>>& buffers) {
auto queue_plain = queue();
auto event = cl_event{};
auto status = GemmStridedBatched(Layout::kRowMajor, Transpose::kNo, Transpose::kNo,
value, value, value, ConstantOne<T>(),
buffers[0](), 0, value, value * value,
buffers[1](), 0, value, value * value, ConstantOne<T>(),
buffers[2](), 0, value, value * value, batch_count,
&queue_plain, &event);
if (status != StatusCode::kSuccess) {
throw RuntimeError("Gemm failed with status " + ToString(status));
}
clWaitForEvents(1, &event);
clReleaseEvent(event);
}
// =================================================================================================
template <typename T>
void TuneXgemm(int argc, char* argv[]) {
auto command_line_args = RetrieveCommandLineArguments(argc, argv);
auto help = std::string{"* Options given/available:\n"};
const auto platform_id = GetArgument(command_line_args, help, kArgPlatform, ConvertArgument(std::getenv("CLBLAST_PLATFORM"), size_t{0}));
const auto device_id = GetArgument(command_line_args, help, kArgDevice, ConvertArgument(std::getenv("CLBLAST_DEVICE"), size_t{0}));
const auto precision = GetArgument(command_line_args, help, kArgPrecision, Precision::kSingle);
const auto num_runs = GetArgument(command_line_args, help, kArgNumRuns, size_t{10});
fprintf(stdout, "%s\n", help.c_str());
// OpenCL initialisation
const auto platform = Platform(platform_id);
const auto device = Device(platform, device_id);
if (!PrecisionSupported<T>(device)) {
printf("* Unsupported precision, skipping this tuning run\n");
return;
}
const auto context = Context(device);
auto queue = Queue(context, device);
// Pre-load GEMM kernel tuning results if they exist
printf("* The GEMM routine tuner requires already tuned kernels\n");
printf(" Applying tuning results from disk if they exist...\n\n");
const auto kernel_names = {"xgemm_1", "xgemm_direct_1", "copy", "pad", "transpose", "padtranspose"};
for (const auto& kernel_name : kernel_names) {
const auto tuner_file_name = "clblast_" + std::string{kernel_name} + "_" +
ToString(static_cast<int>(precision)) + ".json";
printf("* Looking for tuning results in the current folder: '%s'\n", tuner_file_name.c_str());
if (std::ifstream(tuner_file_name)) { // Checks if the file exists on disk
OverrideParametersFromJSONFiles({tuner_file_name}, device(), precision);
}
else {
printf(" Not found: assuming the kernel '%s' is already tuned\n\n", kernel_name);
}
}
// Run the tuners for the XGEMM routines
TuneKernelSelection<T>(platform, device, context, queue, precision, RunGemmRoutine<T>,
64, 2048, 64, 1, num_runs,
"gemm", "GemmRoutine", "gemm_routine", "XGEMM_MIN_INDIRECT_SIZE");
//TuneKernelSelection<T>(platform, device, context, queue, precision, RunGemmBatchedRoutine<T, 30>,
// 16, 128, 32, 30, num_runs,
// "gemmbatched", "GemmRoutine", "gemm_routine_2", "XGEMMBATCHED_MIN_INDIRECT_SIZE");
//TuneKernelSelection<T>(platform, device, context, queue, precision, RunGemmStridedBatchedRoutine<T, 30>,
// 16, 128, 32, 30, num_runs,
// "gemmstridedbatched", "GemmRoutine", "gemm_routine_3", "XGEMMSTRIDEDBATCHED_MIN_INDIRECT_SIZE");
printf("* Completed tuning process\n");
printf("\n");
}
// =================================================================================================
} // namespace clblast
// Shortcuts to the clblast namespace
using half = clblast::half;
using float2 = clblast::float2;
using double2 = clblast::double2;
// Main function (not within the clblast namespace)
int main(int argc, char *argv[]) {
const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv);
switch(clblast::GetPrecision(command_line_args)) {
case clblast::Precision::kHalf: clblast::TuneXgemm<half>(argc, argv); break;
case clblast::Precision::kSingle: clblast::TuneXgemm<float>(argc, argv); break;
case clblast::Precision::kDouble: clblast::TuneXgemm<double>(argc, argv); break;
case clblast::Precision::kComplexSingle: clblast::TuneXgemm<float2>(argc, argv); break;
case clblast::Precision::kComplexDouble: clblast::TuneXgemm<double2>(argc, argv); break;
}
return 0;
}
// =================================================================================================
|
