Skip to content

Commit 06c0ddf

Browse files
mconcasmconcas
committed
Major refactoring of benchmark structure
1 parent afa8b77 commit 06c0ddf

5 files changed

Lines changed: 184 additions & 496 deletions

File tree

‎GPU/GPUbenchmark/CMakeLists.txt‎

Lines changed: 1 addition & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -9,8 +9,6 @@
99
# granted to it by virtue of its status as an Intergovernmental Organization
1010
# or submit itself to any jurisdiction.
1111

12-
set(HDRS_INSTALL ../Shared/Kernels.h)
13-
1412
if(CUDA_ENABLED)
1513
o2_add_executable(gpu-memory-benchmark-cuda
1614
SOURCES benchmark.cxx
@@ -55,4 +53,4 @@ if(HIP_ENABLED)
5553
endif()
5654
endif()
5755

58-
o2_add_test_root_macro(macro/showBenchmarks.C)
56+
o2_add_test_root_macro(macro/showBenchmarks.C)

‎GPU/GPUbenchmark/Shared/Kernels.h‎

Lines changed: 4 additions & 17 deletions
Original file line numberDiff line numberDiff line change
@@ -65,24 +65,11 @@ class GPUbenchmark final
6565
void printDevices(); // Dump info
6666

6767
// Initializations/Finalizations of tests. Not to be measured, in principle used for report
68-
void readInit();
69-
void readFinalize();
68+
void initTest(Test);
69+
void finalizeTest(Test);
7070

71-
void writeInit();
72-
void writeFinalize();
73-
74-
void copyInit();
75-
void copyFinalize();
76-
77-
// Kernel calling wrappers
78-
void readSequential(SplitLevel sl);
79-
void readConcurrent(SplitLevel sl, int nRegions = 2);
80-
81-
void writeSequential(SplitLevel sl);
82-
void writeConcurrent(SplitLevel sl, int nRegions = 2);
83-
84-
void copySequential(SplitLevel sl);
85-
void copyConcurrent(SplitLevel sl, int nRegions = 2);
71+
// Kernel calling wrapper
72+
void runTest(Test, Mode, KernelConfig);
8673

8774
private:
8875
gpuState<chunk_t> mState;

‎GPU/GPUbenchmark/Shared/Utils.h‎

Lines changed: 111 additions & 21 deletions
Original file line numberDiff line numberDiff line change
@@ -15,19 +15,24 @@
1515
#ifndef GPU_BENCHMARK_UTILS_H
1616
#define GPU_BENCHMARK_UTILS_H
1717

18+
#if defined(__HIPCC__)
19+
#include "hip/hip_runtime.h"
20+
#endif
21+
1822
#include <iostream>
1923
#include <iomanip>
2024
#include <typeinfo>
2125
#include <boost/program_options.hpp>
2226
#include <vector>
27+
#include <string>
2328
#include <TTree.h>
2429
#include <TFile.h>
2530

2631
#define KNRM "\x1B[0m"
2732
#define KRED "\x1B[31m"
2833
#define KGRN "\x1B[32m"
2934
#define KYEL "\x1B[33m"
30-
#define KBLU "\x1B[34m"
35+
#define configLU "\x1B[34m"
3136
#define KMAG "\x1B[35m"
3237
#define KCYN "\x1B[36m"
3338
#define KWHT "\x1B[37m"
@@ -48,16 +53,116 @@ enum class Test {
4853
Copy
4954
};
5055

56+
inline std::ostream& operator<<(std::ostream& os, Test test)
57+
{
58+
switch (test) {
59+
case Test::Read:
60+
os << "read";
61+
break;
62+
case Test::Write:
63+
os << "write";
64+
break;
65+
case Test::Copy:
66+
os << "copy";
67+
break;
68+
}
69+
return os;
70+
}
71+
5172
enum class Mode {
5273
Sequential,
5374
Concurrent
5475
};
5576

56-
enum class SplitLevel {
57-
Blocks,
58-
Threads
77+
inline std::ostream& operator<<(std::ostream& os, Mode mode)
78+
{
79+
switch (mode) {
80+
case Mode::Sequential:
81+
os << "sequential";
82+
break;
83+
case Mode::Concurrent:
84+
os << "concurrent";
85+
break;
86+
}
87+
return os;
88+
}
89+
90+
enum class KernelConfig {
91+
Single,
92+
Multi
5993
};
6094

95+
inline std::ostream& operator<<(std::ostream& os, KernelConfig config)
96+
{
97+
switch (config) {
98+
case KernelConfig::Single:
99+
os << "single";
100+
break;
101+
case KernelConfig::Multi:
102+
os << "multiple";
103+
break;
104+
}
105+
return os;
106+
}
107+
108+
template <class T>
109+
inline std::string getType()
110+
{
111+
if (typeid(T).name() == typeid(char).name()) {
112+
return std::string{"char"};
113+
}
114+
if (typeid(T).name() == typeid(size_t).name()) {
115+
return std::string{"unsigned_long"};
116+
}
117+
if (typeid(T).name() == typeid(int).name()) {
118+
return std::string{"int"};
119+
}
120+
if (typeid(T).name() == typeid(int4).name()) {
121+
return std::string{"int4"};
122+
}
123+
return std::string{"unknown"};
124+
}
125+
126+
inline std::string getTestName(Mode mode, Test test, KernelConfig blocks)
127+
{
128+
std::string tname;
129+
tname += (mode == Mode::Sequential) ? "seq_" : "conc_";
130+
tname += (test == Test::Read) ? "read_" : (test == Test::Write) ? "write_"
131+
: "copy_";
132+
tname += (blocks == KernelConfig::Single) ? "SB" : "MB";
133+
return tname;
134+
}
135+
136+
template <class chunk_t>
137+
inline chunk_t* getPartPtr(chunk_t* scratchPtr, float chunkReservedGB, int partNumber)
138+
{
139+
return reinterpret_cast<chunk_t*>(reinterpret_cast<char*>(scratchPtr) + static_cast<size_t>(GB * chunkReservedGB) * partNumber);
140+
}
141+
142+
inline float computeThroughput(Test test, float result, float chunkSizeGB, int ntests)
143+
{
144+
// https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html
145+
// Eff_bandwidth (GB/s) = (B_r + B_w) / (~1e9 * Time (s))
146+
147+
float throughput;
148+
switch (test) {
149+
case Test::Read: {
150+
throughput = 1e3 * chunkSizeGB * ntests / result;
151+
break;
152+
}
153+
case Test::Write: {
154+
throughput = 1e3 * chunkSizeGB * ntests / result;
155+
break;
156+
}
157+
case Test::Copy: {
158+
throughput = 2 * 1e3 * chunkSizeGB * ntests / result;
159+
break;
160+
}
161+
}
162+
163+
return throughput;
164+
}
165+
61166
namespace o2
62167
{
63168
namespace benchmark
@@ -69,7 +174,7 @@ struct benchmarkOpts {
69174
int deviceId = 0;
70175
std::vector<Test> tests = {Test::Read, Test::Write, Test::Copy};
71176
std::vector<Mode> modes = {Mode::Sequential, Mode::Concurrent};
72-
std::vector<SplitLevel> pools = {SplitLevel::Blocks, SplitLevel::Threads};
177+
std::vector<KernelConfig> pools = {KernelConfig::Single, KernelConfig::Multi};
73178
std::vector<std::string> dtypes = {"char", "int", "ulong"};
74179
float chunkReservedGB = 1.f;
75180
int nRegions = 2;
@@ -177,23 +282,8 @@ inline void ResultWriter::addBenchmarkEntry(const std::string bName, const std::
177282

178283
inline void ResultWriter::storeBenchmarkEntry(Test test, int chunk, float entry, float chunkSizeGB, int nLaunches)
179284
{
180-
// https://docs.nvidia.com/cuda/cuda-c-best-practices-guide/index.html
181-
// Eff_bandwidth (GB/s) = (B_r + B_w) / (~1e9 * Time (s))
182285
mTimeResults[chunk] = entry;
183-
switch (test) {
184-
case Test::Read: {
185-
mThroughputResults[chunk] = 1e3 * chunkSizeGB * nLaunches / entry;
186-
break;
187-
}
188-
case Test::Write: {
189-
mThroughputResults[chunk] = 1e3 * chunkSizeGB * nLaunches / entry;
190-
break;
191-
}
192-
case Test::Copy: {
193-
mThroughputResults[chunk] = 2 * 1e3 * chunkSizeGB * nLaunches / entry;
194-
break;
195-
}
196-
}
286+
mThroughputResults[chunk] = computeThroughput(test, entry, chunkSizeGB, nLaunches);
197287
}
198288

199289
inline void ResultWriter::snapshotBenchmark()

‎GPU/GPUbenchmark/benchmark.cxx‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -84,9 +84,9 @@ bool parseArgs(o2::benchmark::benchmarkOpts& conf, int argc, const char* argv[])
8484
conf.pools.clear();
8585
for (auto& pool : vm["pool"].as<std::vector<std::string>>()) {
8686
if (pool == "sb") {
87-
conf.pools.push_back(SplitLevel::Blocks);
87+
conf.pools.push_back(KernelConfig::Single);
8888
} else if (pool == "mb") {
89-
conf.pools.push_back(SplitLevel::Threads);
89+
conf.pools.push_back(KernelConfig::Multi);
9090
} else {
9191
std::cerr << "Unkonwn pool: " << pool << std::endl;
9292
exit(1);

0 commit comments

Comments
 (0)