Skip to content

Commit ac4e687

Browse files
OliverRietmannOliverRietmann
committed
Cleanup
1 parent c8caaf7 commit ac4e687

1 file changed

Lines changed: 80 additions & 80 deletions

File tree

‎GPU/GPUTracking/Base/GPUReconstructionCPU.cxx‎

Lines changed: 80 additions & 80 deletions
Original file line numberDiff line numberDiff line change
@@ -217,61 +217,62 @@ int32_t GPUReconstructionCPU::ExitDevice()
217217
}
218218

219219
namespace {
220-
221-
void write_header(std::ostream& stream) {
222-
if (stream.rdbuf() == std::cout.rdbuf()) {
223-
stream << "| | count | name | gpu (us) | cpu (us) | cpu/tot | tot (us) | GB/s | bytes | bytes/call |\n";
224-
stream << "|---|--------|-------------------------------------------|-----------|-----------|---------|-----------|-----------|---------------|---------------|\n";
225-
} else {
226-
stream << "type,count,name,gpu (us),cpu (us),cpu/total,total (us),GB/s,bytes,bytes/call\n";
227-
}
220+
void write_header(std::ostream& stream) {
221+
if (stream.rdbuf() == std::cout.rdbuf()) {
222+
stream << "| | count | name | gpu (us) | cpu (us) | cpu/tot | tot (us) | GB/s | bytes | bytes/call |\n";
223+
stream << "|---|--------|-------------------------------------------|-----------|-----------|---------|-----------|-----------|---------------|---------------|\n";
224+
} else {
225+
stream << "type,count,name,gpu (us),cpu (us),cpu/total,total (us),GB/s,bytes,bytes/call\n";
228226
}
227+
}
229228

230-
struct Row {
231-
char type = ' ';
232-
std::string name;
233-
uint32_t count = 0;
234-
double gpu_time = -1.0;
235-
double cpu_time = -1.0;
236-
double total_time = -1.0;
237-
uint32_t memSize = 0;
238-
uint32_t statNEvents;
239-
240-
void test(std::ostream& stream) {
241-
//static constexpr const char* fmt = "| {0:6} | {1:7} | {2:43} | {3:15.0f} | {4:15.0f} | {5:15.2f} | {6:15.0f} | {7:10.3f} | {8:10.0f} | {9:15.0f} |\n";
242-
double scale = 1000000.0 / statNEvents;
243-
stream << "| " << type << " | ";
244-
if (count != 0) stream << std::format("{:6} |", count);
245-
else stream << " |";
246-
stream << std::format(" {:41}", name) << " |" << std::format("{:10.0f}", gpu_time * scale) << " |";
247-
if (cpu_time != -1.0) stream << std::format("{:10.0f} |", cpu_time * scale);
248-
else stream << " |";
249-
if (cpu_time != -1.0 && total_time != -1.0) stream << std::format("{:8.2f} |", cpu_time / total_time);
250-
else stream << " |";
251-
if (total_time != -1.0) stream << std::format("{:10.0f} |", total_time * scale);
252-
else stream << " |";
253-
if (memSize != 0 && count != 0) stream << std::format("{:10.3f} |", memSize / gpu_time * 1e-9) << std::format("{:14} |", memSize / statNEvents) << std::format("{:14} |", memSize / statNEvents / count);
254-
else stream << " | | |";
255-
stream << std::endl;
256-
}
229+
struct Row {
230+
char type = ' ';
231+
std::string name;
232+
uint32_t count = 0;
233+
double gpu_time = -1.0;
234+
double cpu_time = -1.0;
235+
double total_time = -1.0;
236+
uint32_t memSize = 0;
237+
uint32_t statNEvents;
238+
239+
void writeToCout(std::ostream& stream) {
240+
double scale = 1000000.0 / statNEvents;
241+
stream << "| " << type << " | ";
242+
if (count != 0) stream << std::format("{:6} |", count);
243+
else stream << " |";
244+
stream << std::format(" {:42}|", name);
245+
if (gpu_time != -1.0) stream << std::format("{:10.0f} |", gpu_time * scale);
246+
else stream << " |";
247+
if (cpu_time != -1.0) stream << std::format("{:10.0f} |", cpu_time * scale);
248+
else stream << " |";
249+
if (cpu_time != -1.0 && total_time != -1.0) stream << std::format("{:8.2f} |", cpu_time / total_time);
250+
else stream << " |";
251+
if (total_time != -1.0) stream << std::format("{:10.0f} |", total_time * scale);
252+
else stream << " |";
253+
if (memSize != 0 && count != 0) stream << std::format("{:10.3f} |{:14} |{:14} |", memSize / gpu_time * 1e-9, memSize / statNEvents, memSize / statNEvents / count);
254+
else stream << " | | |";
255+
stream << std::endl;
256+
}
257257

258-
void write(std::ostream& stream) {
259-
double scale = 1000000.0 / statNEvents;
260-
//stream << std::format("{0},{1},{2},{3:.0f},{4:.0f},{5:.2f},{6:.0f},{7:.3f},{8},{9}\n", type, count, name, gpu_time * scale, cpu_time * scale, cpu_time / total_time, total_time * scale, memSize / gpu_time * 1e-9, memSize / statNEvents, memSize / statNEvents / count);
261-
stream << type << ",";
262-
if (count != 0) stream << count;
263-
stream << "," << name << "," << std::format("{:.0f}", gpu_time * scale) << ",";
264-
if (cpu_time != -1.0) stream << std::format("{:.0f}", cpu_time * scale);
265-
stream << ",";
266-
if (cpu_time != -1.0 && total_time != -1.0) stream << std::format("{:.2f}", cpu_time / total_time);
267-
stream << ",";
268-
if (total_time != -1.0) stream << std::format("{:.0f}", total_time * scale);
269-
stream << ",";
270-
if (memSize != 0 && count != 0) stream << std::format("{:.3f}", memSize / gpu_time * 1e-9) << "," << memSize / statNEvents << "," << memSize / statNEvents / count;
271-
else stream << ",,";
272-
stream << std::endl;
273-
}
274-
};
258+
void writeToFile(std::ostream& stream) {
259+
double scale = 1000000.0 / statNEvents;
260+
stream << type << ",";
261+
if (count != 0) stream << count;
262+
stream << "," << name << ",";
263+
if (gpu_time != -1.0) stream << std::format("{:.0f}", gpu_time * scale);
264+
stream << ",";
265+
if (cpu_time != -1.0) stream << std::format("{:.0f}", cpu_time * scale);
266+
stream << ",";
267+
if (cpu_time != -1.0 && total_time != -1.0) stream << std::format("{:.2f}", cpu_time / total_time);
268+
stream << ",";
269+
if (total_time != -1.0) stream << std::format("{:.0f}", total_time * scale);
270+
stream << ",";
271+
if (memSize != 0 && count != 0) stream << std::format("{:.3f},{},{}", memSize / gpu_time * 1e-9, memSize / statNEvents, memSize / statNEvents / count);
272+
else stream << ",,";
273+
stream << std::endl;
274+
}
275+
};
275276
}
276277

277278
int32_t GPUReconstructionCPU::RunChains()
@@ -318,7 +319,7 @@ int32_t GPUReconstructionCPU::RunChains()
318319
PrintMemoryOverview();
319320
}
320321

321-
mStatWallTime = (mTimerTotal.GetElapsedTime() * 1000000. / mStatNEvents);
322+
mStatWallTime = mTimerTotal.GetElapsedTime();
322323
std::string nEventReport;
323324
if (GetProcessingSettings().debugLevel >= 0 && mStatNEvents > 1) {
324325
nEventReport += " (avergage of " + std::to_string(mStatNEvents) + " runs)";
@@ -357,7 +358,6 @@ int32_t GPUReconstructionCPU::RunChains()
357358
int32_t stepNum = getRecoStepNum(mTimers[i]->step);
358359
kernelStepTimes[stepNum] += time;
359360
}
360-
char bandwidth[256] = "";
361361
Row task_row;
362362
task_row.type = 'K';
363363
task_row.name = mTimers[i]->name.c_str();
@@ -366,11 +366,11 @@ int32_t GPUReconstructionCPU::RunChains()
366366
task_row.statNEvents = mStatNEvents;
367367
if (mTimers[i]->memSize && mStatNEvents && time != 0.) {
368368
task_row.memSize = mTimers[i]->memSize;
369-
snprintf(bandwidth, 256, " (%8.3f GB/s - %'14zu bytes - %'14zu per call)", mTimers[i]->memSize / time * 1e-9, mTimers[i]->memSize / mStatNEvents, mTimers[i]->memSize / mStatNEvents / mTimers[i]->count);
370369
}
371-
if (benchmarkCSV.is_open()) task_row.write(benchmarkCSV);
372-
task_row.test(std::cout);
373-
//printf("Execution Time: Task (%c %8ux): %50s Time: %'10.0f us%s\n", type == 0 ? 'K' : 'C', mTimers[i]->count, mTimers[i]->name.c_str(), time * 1000000 / mStatNEvents, bandwidth);
370+
if (benchmarkCSV.is_open()) {
371+
task_row.writeToFile(benchmarkCSV);
372+
}
373+
task_row.writeToCout(std::cout);
374374
if (GetProcessingSettings().resetTimers) {
375375
mTimers[i]->count = 0;
376376
mTimers[i]->memSize = 0;
@@ -386,10 +386,10 @@ int32_t GPUReconstructionCPU::RunChains()
386386
reco_step_row.cpu_time = mTimersRecoSteps[i].timerCPU;
387387
reco_step_row.total_time = mTimersRecoSteps[i].timerTotal.GetElapsedTime();
388388
reco_step_row.statNEvents = mStatNEvents;
389-
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV);
390-
reco_step_row.test(std::cout);
391-
//printf("Execution Time: Step : %11s %38s Time: %'10.0f us %64s ( Total Time : %'14.0f us, CPU Time : %'14.0f us, %'7.2fx )\n", "Tasks",
392-
//gpudatatypes::RECO_STEP_NAMES[i], kernelStepTimes[i] * 1000000 / mStatNEvents, "", mTimersRecoSteps[i].timerTotal.GetElapsedTime() * 1000000 / mStatNEvents, mTimersRecoSteps[i].timerCPU * 1000000 / mStatNEvents, mTimersRecoSteps[i].timerCPU / mTimersRecoSteps[i].timerTotal.GetElapsedTime());
389+
if (benchmarkCSV.is_open()) {
390+
reco_step_row.writeToFile(benchmarkCSV);
391+
}
392+
reco_step_row.writeToCout(std::cout);
393393
}
394394
if (mTimersRecoSteps[i].bytesToGPU) {
395395
Row reco_step_row;
@@ -399,10 +399,10 @@ int32_t GPUReconstructionCPU::RunChains()
399399
reco_step_row.memSize = mTimersRecoSteps[i].bytesToGPU;
400400
reco_step_row.count = mTimersRecoSteps[i].countToGPU;
401401
reco_step_row.statNEvents = mStatNEvents;
402-
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV);
403-
reco_step_row.test(std::cout);
404-
//printf("Execution Time: Step (D %8ux): %11s %38s Time: %'10.0f us (%8.3f GB/s - %'14zu bytes - %'14zu per call)\n", mTimersRecoSteps[i].countToGPU, "DMA to GPU", gpudatatypes::RECO_STEP_NAMES[i], mTimersRecoSteps[i].timerToGPU.GetElapsedTime() * 1000000 / mStatNEvents,
405-
//mTimersRecoSteps[i].bytesToGPU / mTimersRecoSteps[i].timerToGPU.GetElapsedTime() * 1e-9, mTimersRecoSteps[i].bytesToGPU / mStatNEvents, mTimersRecoSteps[i].bytesToGPU / mTimersRecoSteps[i].countToGPU);
402+
if (benchmarkCSV.is_open()) {
403+
reco_step_row.writeToFile(benchmarkCSV);
404+
}
405+
reco_step_row.writeToCout(std::cout);
406406
}
407407
if (mTimersRecoSteps[i].bytesToHost) {
408408
Row reco_step_row;
@@ -412,10 +412,10 @@ int32_t GPUReconstructionCPU::RunChains()
412412
reco_step_row.memSize = mTimersRecoSteps[i].bytesToHost;
413413
reco_step_row.count = mTimersRecoSteps[i].countToHost;
414414
reco_step_row.statNEvents = mStatNEvents;
415-
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV);
416-
reco_step_row.test(std::cout);
417-
//printf("Execution Time: Step (D %8ux): %11s %38s Time: %'10.0f us (%8.3f GB/s - %'14zu bytes - %'14zu per call)\n", mTimersRecoSteps[i].countToHost, "DMA to Host", gpudatatypes::RECO_STEP_NAMES[i], mTimersRecoSteps[i].timerToHost.GetElapsedTime() * 1000000 / mStatNEvents,
418-
//mTimersRecoSteps[i].bytesToHost / mTimersRecoSteps[i].timerToHost.GetElapsedTime() * 1e-9, mTimersRecoSteps[i].bytesToHost / mStatNEvents, mTimersRecoSteps[i].bytesToHost / mTimersRecoSteps[i].countToHost);
415+
if (benchmarkCSV.is_open()) {
416+
reco_step_row.writeToFile(benchmarkCSV);
417+
}
418+
reco_step_row.writeToCout(std::cout);
419419
}
420420
if (GetProcessingSettings().resetTimers) {
421421
mTimersRecoSteps[i].bytesToGPU = mTimersRecoSteps[i].bytesToHost = 0;
@@ -433,26 +433,26 @@ int32_t GPUReconstructionCPU::RunChains()
433433
general_step_row.name = gpudatatypes::GENERAL_STEP_NAMES[i];
434434
general_step_row.gpu_time = mTimersGeneralSteps[i].GetElapsedTime();
435435
general_step_row.statNEvents = mStatNEvents;
436-
if (benchmarkCSV.is_open()) general_step_row.write(benchmarkCSV);
437-
general_step_row.test(std::cout);
438-
//printf("Execution Time: General Step : %50s Time: %'10.0f us\n", gpudatatypes::GENERAL_STEP_NAMES[i], mTimersGeneralSteps[i].GetElapsedTime() * 1000000 / mStatNEvents);
436+
if (benchmarkCSV.is_open()) {
437+
general_step_row.writeToFile(benchmarkCSV);
438+
}
439+
general_step_row.writeToCout(std::cout);
439440
}
440441
}
441442
Row wall_row;
442443
wall_row.name = "Wall";
443444
if (GetProcessingSettings().debugLevel >= 1) {
444445
wall_row.gpu_time = kernelTotal;
445-
mStatKernelTime = kernelTotal * 1000000 / mStatNEvents;
446-
//printf("Execution Time: Total : %50s Time: %'10.0f us%s\n", "Total Kernel", mStatKernelTime, nEventReport.c_str());
447446
}
448447
wall_row.cpu_time = mStatCPUTime;
449-
wall_row.total_time = mStatWallTime * mStatNEvents / 1000000;
448+
wall_row.total_time = mStatWallTime;
450449
wall_row.statNEvents = mStatNEvents;
451-
if (benchmarkCSV.is_open()) wall_row.write(benchmarkCSV);
452-
wall_row.test(std::cout);
453-
//printf("Execution Time: Total : %50s Time: %'10.0f us ( CPU Time : %'10.0f us, %7.2fx ) %s\n", "Total Wall", mStatWallTime, mStatCPUTime * 1000000 / mStatNEvents, mStatCPUTime / mTimerTotal.GetElapsedTime(), nEventReport.c_str());
450+
if (benchmarkCSV.is_open()) {
451+
wall_row.writeToFile(benchmarkCSV);
452+
}
453+
wall_row.writeToCout(std::cout);
454454
} else if (GetProcessingSettings().debugLevel >= 0) {
455-
GPUInfo("Total Wall Time: %10.0f us%s", mStatWallTime, nEventReport.c_str());
455+
GPUInfo("Total Wall Time: %10.0f us%s", mStatWallTime * 1000000 / mStatNEvents, nEventReport.c_str());
456456
}
457457
if (GetProcessingSettings().resetTimers) {
458458
mStatNEvents = 0;

0 commit comments

Comments
 (0)