Skip to content

Commit 45d289b

Browse files
OliverRietmannOliverRietmann
committed
Rename variables
1 parent ad03734 commit 45d289b

File tree

1 file changed

+35
-28
lines changed

1 file changed

+35
-28
lines changed

‎GPU/GPUTracking/Base/GPUReconstructionCPU.cxx‎

Lines changed: 35 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -228,45 +228,46 @@ namespace {
228228
}
229229

230230
struct Row {
231-
std::string name;
232231
char type = ' ';
233-
double kernel_time = -1.0;
232+
std::string name;
233+
uint32_t count = 0;
234+
double gpu_time = -1.0;
234235
double cpu_time = -1.0;
235236
double total_time = -1.0;
236237
uint32_t memSize = 0;
237-
uint32_t count = 0;
238+
uint32_t statNEvents;
238239

239-
void test(std::ostream& stream, uint32_t statNEvents) {
240+
void test(std::ostream& stream) {
240241
//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";
241242
double scale = 1000000.0 / statNEvents;
242243
stream << "| " << type << " | ";
243244
if (count != 0) stream << std::format("{:6} |", count);
244245
else stream << " |";
245-
stream << std::format(" {:41}", name) << " |" << std::format("{:10.0f}", kernel_time * scale) << " |";
246+
stream << std::format(" {:41}", name) << " |" << std::format("{:10.0f}", gpu_time * scale) << " |";
246247
if (cpu_time != -1.0) stream << std::format("{:10.0f} |", cpu_time * scale);
247248
else stream << " |";
248249
if (cpu_time != -1.0 && total_time != -1.0) stream << std::format("{:8.2f} |", cpu_time / total_time);
249250
else stream << " |";
250251
if (total_time != -1.0) stream << std::format("{:10.0f} |", total_time * scale);
251252
else stream << " |";
252-
if (memSize != 0 && count != 0) stream << std::format("{:10.3f} |", memSize / kernel_time * 1e-9) << std::format("{:14} |", memSize / statNEvents) << std::format("{:14} |", memSize / statNEvents / count);
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);
253254
else stream << " | | |";
254255
stream << std::endl;
255256
}
256257

257-
void write(std::ostream& stream, uint32_t statNEvents) {
258+
void write(std::ostream& stream) {
258259
double scale = 1000000.0 / statNEvents;
259-
//stream << std::format("{0},{1},{2},{3:.0f},{4:.0f},{5:.2f},{6:.0f},{7:.3f},{8},{9}\n", type, count, name, kernel_time * scale, cpu_time * scale, cpu_time / total_time, total_time * scale, memSize / kernel_time * 1e-9, memSize / statNEvents, memSize / statNEvents / count);
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);
260261
stream << type << ",";
261262
if (count != 0) stream << count;
262-
stream << "," << name << "," << std::format("{:.0f}", kernel_time * scale) << ",";
263+
stream << "," << name << "," << std::format("{:.0f}", gpu_time * scale) << ",";
263264
if (cpu_time != -1.0) stream << std::format("{:.0f}", cpu_time * scale);
264265
stream << ",";
265266
if (cpu_time != -1.0 && total_time != -1.0) stream << std::format("{:.2f}", cpu_time / total_time);
266267
stream << ",";
267268
if (total_time != -1.0) stream << std::format("{:.0f}", total_time * scale);
268269
stream << ",";
269-
if (memSize != 0 && count != 0) stream << std::format("{:.3f}", memSize / kernel_time * 1e-9) << "," << memSize / statNEvents << "," << memSize / statNEvents / count;
270+
if (memSize != 0 && count != 0) stream << std::format("{:.3f}", memSize / gpu_time * 1e-9) << "," << memSize / statNEvents << "," << memSize / statNEvents / count;
270271
else stream << ",,";
271272
stream << std::endl;
272273
}
@@ -360,14 +361,15 @@ int32_t GPUReconstructionCPU::RunChains()
360361
Row task_row;
361362
task_row.type = 'K';
362363
task_row.name = mTimers[i]->name.c_str();
363-
task_row.kernel_time = time;
364+
task_row.gpu_time = time;
364365
task_row.count = mTimers[i]->count;
366+
task_row.statNEvents = mStatNEvents;
365367
if (mTimers[i]->memSize && mStatNEvents && time != 0.) {
366368
task_row.memSize = mTimers[i]->memSize;
367369
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);
368370
}
369-
if (benchmarkCSV.is_open()) task_row.write(benchmarkCSV, mStatNEvents);
370-
task_row.test(std::cout, mStatNEvents);
371+
if (benchmarkCSV.is_open()) task_row.write(benchmarkCSV);
372+
task_row.test(std::cout);
371373
//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);
372374
if (GetProcessingSettings().resetTimers) {
373375
mTimers[i]->count = 0;
@@ -380,35 +382,38 @@ int32_t GPUReconstructionCPU::RunChains()
380382
if (kernelStepTimes[i] != 0. || mTimersRecoSteps[i].timerTotal.GetElapsedTime() != 0.) {
381383
Row reco_step_row;
382384
reco_step_row.name = std::string(gpudatatypes::RECO_STEP_NAMES[i]) + " (Tasks)";
383-
reco_step_row.kernel_time = kernelStepTimes[i];
385+
reco_step_row.gpu_time = kernelStepTimes[i];
384386
reco_step_row.cpu_time = mTimersRecoSteps[i].timerCPU;
385387
reco_step_row.total_time = mTimersRecoSteps[i].timerTotal.GetElapsedTime();
386-
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV, mStatNEvents);
387-
reco_step_row.test(std::cout, mStatNEvents);
388+
reco_step_row.statNEvents = mStatNEvents;
389+
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV);
390+
reco_step_row.test(std::cout);
388391
//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",
389392
//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());
390393
}
391394
if (mTimersRecoSteps[i].bytesToGPU) {
392395
Row reco_step_row;
393396
reco_step_row.type = 'D';
394397
reco_step_row.name = std::string(gpudatatypes::RECO_STEP_NAMES[i]) + " (DMA to GPU)";
395-
reco_step_row.kernel_time = mTimersRecoSteps[i].timerToGPU.GetElapsedTime();
398+
reco_step_row.gpu_time = mTimersRecoSteps[i].timerToGPU.GetElapsedTime();
396399
reco_step_row.memSize = mTimersRecoSteps[i].bytesToGPU;
397400
reco_step_row.count = mTimersRecoSteps[i].countToGPU;
398-
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV, mStatNEvents);
399-
reco_step_row.test(std::cout, mStatNEvents);
401+
reco_step_row.statNEvents = mStatNEvents;
402+
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV);
403+
reco_step_row.test(std::cout);
400404
//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,
401405
//mTimersRecoSteps[i].bytesToGPU / mTimersRecoSteps[i].timerToGPU.GetElapsedTime() * 1e-9, mTimersRecoSteps[i].bytesToGPU / mStatNEvents, mTimersRecoSteps[i].bytesToGPU / mTimersRecoSteps[i].countToGPU);
402406
}
403407
if (mTimersRecoSteps[i].bytesToHost) {
404408
Row reco_step_row;
405409
reco_step_row.type = 'D';
406410
reco_step_row.name = std::string(gpudatatypes::RECO_STEP_NAMES[i]) + " (DMA to Host)";
407-
reco_step_row.kernel_time = mTimersRecoSteps[i].timerToHost.GetElapsedTime();
411+
reco_step_row.gpu_time = mTimersRecoSteps[i].timerToHost.GetElapsedTime();
408412
reco_step_row.memSize = mTimersRecoSteps[i].bytesToHost;
409413
reco_step_row.count = mTimersRecoSteps[i].countToHost;
410-
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV, mStatNEvents);
411-
reco_step_row.test(std::cout, mStatNEvents);
414+
reco_step_row.statNEvents = mStatNEvents;
415+
if (benchmarkCSV.is_open()) reco_step_row.write(benchmarkCSV);
416+
reco_step_row.test(std::cout);
412417
//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,
413418
//mTimersRecoSteps[i].bytesToHost / mTimersRecoSteps[i].timerToHost.GetElapsedTime() * 1e-9, mTimersRecoSteps[i].bytesToHost / mStatNEvents, mTimersRecoSteps[i].bytesToHost / mTimersRecoSteps[i].countToHost);
414419
}
@@ -426,23 +431,25 @@ int32_t GPUReconstructionCPU::RunChains()
426431
if (mTimersGeneralSteps[i].GetElapsedTime() != 0.) {
427432
Row general_step_row;
428433
general_step_row.name = gpudatatypes::GENERAL_STEP_NAMES[i];
429-
general_step_row.kernel_time = mTimersGeneralSteps[i].GetElapsedTime();
430-
if (benchmarkCSV.is_open()) general_step_row.write(benchmarkCSV, mStatNEvents);
431-
general_step_row.test(std::cout, mStatNEvents);
434+
general_step_row.gpu_time = mTimersGeneralSteps[i].GetElapsedTime();
435+
general_step_row.statNEvents = mStatNEvents;
436+
if (benchmarkCSV.is_open()) general_step_row.write(benchmarkCSV);
437+
general_step_row.test(std::cout);
432438
//printf("Execution Time: General Step : %50s Time: %'10.0f us\n", gpudatatypes::GENERAL_STEP_NAMES[i], mTimersGeneralSteps[i].GetElapsedTime() * 1000000 / mStatNEvents);
433439
}
434440
}
435441
Row wall_row;
436442
wall_row.name = "Wall";
437443
if (GetProcessingSettings().debugLevel >= 1) {
438-
wall_row.kernel_time = kernelTotal;
444+
wall_row.gpu_time = kernelTotal;
439445
mStatKernelTime = kernelTotal * 1000000 / mStatNEvents;
440446
//printf("Execution Time: Total : %50s Time: %'10.0f us%s\n", "Total Kernel", mStatKernelTime, nEventReport.c_str());
441447
}
442448
wall_row.cpu_time = mStatCPUTime;
443449
wall_row.total_time = mStatWallTime * mStatNEvents / 1000000;
444-
if (benchmarkCSV.is_open()) wall_row.write(benchmarkCSV, mStatNEvents);
445-
wall_row.test(std::cout, mStatNEvents);
450+
wall_row.statNEvents = mStatNEvents;
451+
if (benchmarkCSV.is_open()) wall_row.write(benchmarkCSV);
452+
wall_row.test(std::cout);
446453
//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());
447454
} else if (GetProcessingSettings().debugLevel >= 0) {
448455
GPUInfo("Total Wall Time: %10.0f us%s", mStatWallTime, nEventReport.c_str());

0 commit comments

Comments
 (0)