TransferBench.cpp 57.3 KB
Newer Older
Gilbert Lee's avatar
Gilbert Lee committed
1
/*
gilbertlee-amd's avatar
gilbertlee-amd committed
2
Copyright (c) 2019-2023 Advanced Micro Devices, Inc. All rights reserved.
Gilbert Lee's avatar
Gilbert Lee committed
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.
*/

// This program measures simultaneous copy performance across multiple GPUs
// on the same node
#include <numa.h>
#include <numaif.h>
Gilbert Lee's avatar
Gilbert Lee committed
27
#include <random>
Gilbert Lee's avatar
Gilbert Lee committed
28
29
30
31
32
33
34
35
#include <stack>
#include <thread>

#include "TransferBench.hpp"
#include "GetClosestNumaNode.hpp"

int main(int argc, char **argv)
{
Gilbert Lee's avatar
Gilbert Lee committed
36
37
38
39
40
41
42
  // Check for NUMA library support
  if (numa_available() == -1)
  {
    printf("[ERROR] NUMA library not supported. Check to see if libnuma has been installed on this system\n");
    exit(1);
  }

Gilbert Lee's avatar
Gilbert Lee committed
43
44
45
46
47
48
49
50
51
52
53
54
  // Display usage instructions and detected topology
  if (argc <= 1)
  {
    int const outputToCsv = EnvVars::GetEnvVar("OUTPUT_TO_CSV", 0);
    if (!outputToCsv) DisplayUsage(argv[0]);
    DisplayTopology(outputToCsv);
    exit(0);
  }

  // Collect environment variables / display current run configuration
  EnvVars ev;

Gilbert Lee's avatar
Gilbert Lee committed
55
56
  // Determine number of bytes to run per Transfer
  size_t numBytesPerTransfer = argc > 2 ? atoll(argv[2]) : DEFAULT_BYTES_PER_TRANSFER;
Gilbert Lee's avatar
Gilbert Lee committed
57
58
59
60
61
62
  if (argc > 2)
  {
    // Adjust bytes if unit specified
    char units = argv[2][strlen(argv[2])-1];
    switch (units)
    {
Gilbert Lee's avatar
Gilbert Lee committed
63
64
65
    case 'K': case 'k': numBytesPerTransfer *= 1024; break;
    case 'M': case 'm': numBytesPerTransfer *= 1024*1024; break;
    case 'G': case 'g': numBytesPerTransfer *= 1024*1024*1024; break;
Gilbert Lee's avatar
Gilbert Lee committed
66
67
    }
  }
gilbertlee-amd's avatar
gilbertlee-amd committed
68
69
70
71
72
  if (numBytesPerTransfer % 4)
  {
    printf("[ERROR] numBytesPerTransfer (%lu) must be a multiple of 4\n", numBytesPerTransfer);
    exit(1);
  }
Gilbert Lee's avatar
Gilbert Lee committed
73

Gilbert Lee's avatar
Gilbert Lee committed
74
75
76
77
  // Check for preset tests
  // - Tests that sweep across possible sets of Transfers
  if (!strcmp(argv[1], "sweep") || !strcmp(argv[1], "rsweep"))
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
78
79
    int numGpuSubExecs = (argc > 3 ? atoi(argv[3]) : 4);
    int numCpuSubExecs = (argc > 4 ? atoi(argv[4]) : 4);
gilbertlee-amd's avatar
gilbertlee-amd committed
80

81
    ev.configMode = CFG_SWEEP;
gilbertlee-amd's avatar
gilbertlee-amd committed
82
    RunSweepPreset(ev, numBytesPerTransfer, numGpuSubExecs, numCpuSubExecs, !strcmp(argv[1], "rsweep"));
Gilbert Lee's avatar
Gilbert Lee committed
83
84
85
    exit(0);
  }
  // - Tests that benchmark peer-to-peer performance
gilbertlee-amd's avatar
gilbertlee-amd committed
86
  else if (!strcmp(argv[1], "p2p"))
Gilbert Lee's avatar
Gilbert Lee committed
87
  {
88
    ev.configMode = CFG_P2P;
gilbertlee-amd's avatar
gilbertlee-amd committed
89
    RunPeerToPeerBenchmarks(ev, numBytesPerTransfer / sizeof(float));
Gilbert Lee's avatar
Gilbert Lee committed
90
91
92
    exit(0);
  }

Gilbert Lee's avatar
Gilbert Lee committed
93
  // Check that Transfer configuration file can be opened
94
  ev.configMode = CFG_FILE;
Gilbert Lee's avatar
Gilbert Lee committed
95
96
97
  FILE* fp = fopen(argv[1], "r");
  if (!fp)
  {
Gilbert Lee's avatar
Gilbert Lee committed
98
    printf("[ERROR] Unable to open transfer configuration file: [%s]\n", argv[1]);
Gilbert Lee's avatar
Gilbert Lee committed
99
100
101
    exit(1);
  }

Gilbert Lee's avatar
Gilbert Lee committed
102
  // Print environment variables and CSV header
Gilbert Lee's avatar
Gilbert Lee committed
103
104
105
  ev.DisplayEnvVars();
  if (ev.outputToCsv)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
106
    printf("Test#,Transfer#,NumBytes,Src,Exe,Dst,CUs,BW(GB/s),Time(ms),SrcAddr,DstAddr\n");
Gilbert Lee's avatar
Gilbert Lee committed
107
108
109
110
111
112
113
114
115
  }

  int testNum = 0;
  char line[2048];
  while(fgets(line, 2048, fp))
  {
    // Check if line is a comment to be echoed to output (starts with ##)
    if (!ev.outputToCsv && line[0] == '#' && line[1] == '#') printf("%s", line);

Gilbert Lee's avatar
Gilbert Lee committed
116
117
118
119
    // Parse set of parallel Transfers to execute
    std::vector<Transfer> transfers;
    ParseTransfers(line, ev.numCpuDevices, ev.numGpuDevices, transfers);
    if (transfers.empty()) continue;
Gilbert Lee's avatar
Gilbert Lee committed
120

gilbertlee-amd's avatar
gilbertlee-amd committed
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
    // If the number of bytes is specified, use it
    if (numBytesPerTransfer != 0)
    {
      size_t N = numBytesPerTransfer / sizeof(float);
      ExecuteTransfers(ev, ++testNum, N, transfers);
    }
    else
    {
      // Otherwise generate a range of values
      for (int N = 256; N <= (1<<27); N *= 2)
      {
        int delta = std::max(32, N / ev.samplingFactor);
        int curr = N;
        while (curr < N * 2)
        {
          ExecuteTransfers(ev, ++testNum, N, transfers);
          curr += delta;
        }
      }
    }
Gilbert Lee's avatar
Gilbert Lee committed
141
142
  }
  fclose(fp);
Gilbert Lee's avatar
Gilbert Lee committed
143

Gilbert Lee's avatar
Gilbert Lee committed
144
145
  return 0;
}
Gilbert Lee's avatar
Gilbert Lee committed
146

Gilbert Lee's avatar
Gilbert Lee committed
147
void ExecuteTransfers(EnvVars const& ev,
gilbertlee-amd's avatar
gilbertlee-amd committed
148
149
150
151
                      int const testNum,
                      size_t const N,
                      std::vector<Transfer>& transfers,
                      bool verbose)
Gilbert Lee's avatar
Gilbert Lee committed
152
153
{
  int const initOffset = ev.byteOffset / sizeof(float);
Gilbert Lee's avatar
Gilbert Lee committed
154

Gilbert Lee's avatar
Gilbert Lee committed
155
156
  // Map transfers by executor
  TransferMap transferMap;
gilbertlee-amd's avatar
gilbertlee-amd committed
157
  for (Transfer& transfer : transfers)
Gilbert Lee's avatar
Gilbert Lee committed
158
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
159
    Executor executor(transfer.exeType, transfer.exeIndex);
Gilbert Lee's avatar
Gilbert Lee committed
160
    ExecutorInfo& executorInfo = transferMap[executor];
gilbertlee-amd's avatar
gilbertlee-amd committed
161
    executorInfo.transfers.push_back(&transfer);
Gilbert Lee's avatar
Gilbert Lee committed
162
  }
Gilbert Lee's avatar
Gilbert Lee committed
163

gilbertlee-amd's avatar
gilbertlee-amd committed
164
  // Loop over each executor and prepare sub-executors
gilbertlee-amd's avatar
gilbertlee-amd committed
165
  std::map<int, Transfer*> transferList;
Gilbert Lee's avatar
Gilbert Lee committed
166
167
168
  for (auto& exeInfoPair : transferMap)
  {
    Executor const& executor = exeInfoPair.first;
gilbertlee-amd's avatar
gilbertlee-amd committed
169
170
171
172
    ExecutorInfo& exeInfo    = exeInfoPair.second;
    ExeType const exeType    = executor.first;
    int     const exeIndex   = RemappedIndex(executor.second, IsCpuType(exeType));

Gilbert Lee's avatar
Gilbert Lee committed
173
    exeInfo.totalTime = 0.0;
gilbertlee-amd's avatar
gilbertlee-amd committed
174
    exeInfo.totalSubExecs = 0;
Gilbert Lee's avatar
Gilbert Lee committed
175
176

    // Loop over each transfer this executor is involved in
gilbertlee-amd's avatar
gilbertlee-amd committed
177
    for (Transfer* transfer : exeInfo.transfers)
Gilbert Lee's avatar
Gilbert Lee committed
178
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
179
180
181
182
183
184
      // Determine how many bytes to copy for this Transfer (use custom if pre-specified)
      transfer->numBytesActual = (transfer->numBytes ? transfer->numBytes : N * sizeof(float));

      // Allocate source memory
      transfer->srcMem.resize(transfer->numSrcs);
      for (int iSrc = 0; iSrc < transfer->numSrcs; ++iSrc)
Gilbert Lee's avatar
Gilbert Lee committed
185
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
186
187
188
        MemType const& srcType  = transfer->srcType[iSrc];
        int     const  srcIndex    = RemappedIndex(transfer->srcIndex[iSrc], IsCpuType(srcType));

Gilbert Lee's avatar
Gilbert Lee committed
189
        // Ensure executing GPU can access source memory
gilbertlee-amd's avatar
gilbertlee-amd committed
190
        if (IsGpuType(exeType) == MEM_GPU && IsGpuType(srcType) && srcIndex != exeIndex)
Gilbert Lee's avatar
Gilbert Lee committed
191
          EnablePeerAccess(exeIndex, srcIndex);
Gilbert Lee's avatar
Gilbert Lee committed
192

gilbertlee-amd's avatar
gilbertlee-amd committed
193
194
195
196
197
198
199
200
201
202
        AllocateMemory(srcType, srcIndex, transfer->numBytesActual + ev.byteOffset, (void**)&transfer->srcMem[iSrc]);
      }

      // Allocate destination memory
      transfer->dstMem.resize(transfer->numDsts);
      for (int iDst = 0; iDst < transfer->numDsts; ++iDst)
      {
        MemType const& dstType  = transfer->dstType[iDst];
        int     const  dstIndex    = RemappedIndex(transfer->dstIndex[iDst], IsCpuType(dstType));

Gilbert Lee's avatar
Gilbert Lee committed
203
        // Ensure executing GPU can access destination memory
gilbertlee-amd's avatar
gilbertlee-amd committed
204
        if (IsGpuType(exeType) == MEM_GPU && IsGpuType(dstType) && dstIndex != exeIndex)
Gilbert Lee's avatar
Gilbert Lee committed
205
206
          EnablePeerAccess(exeIndex, dstIndex);

gilbertlee-amd's avatar
gilbertlee-amd committed
207
208
        AllocateMemory(dstType, dstIndex, transfer->numBytesActual + ev.byteOffset, (void**)&transfer->dstMem[iDst]);
      }
Gilbert Lee's avatar
Gilbert Lee committed
209

gilbertlee-amd's avatar
gilbertlee-amd committed
210
      exeInfo.totalSubExecs += transfer->numSubExecs;
gilbertlee-amd's avatar
gilbertlee-amd committed
211
      transferList[transfer->transferIndex] = transfer;
Gilbert Lee's avatar
Gilbert Lee committed
212
213
    }

gilbertlee-amd's avatar
gilbertlee-amd committed
214
215
    // Prepare additional requirement for GPU-based executors
    if (IsGpuType(exeType))
Gilbert Lee's avatar
Gilbert Lee committed
216
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
217
218
219
220
221
222
      // Single-stream is only supported for GFX-based executors
      int const numStreamsToUse = (exeType == EXE_GPU_DMA || !ev.useSingleStream) ? exeInfo.transfers.size() : 1;
      exeInfo.streams.resize(numStreamsToUse);
      exeInfo.startEvents.resize(numStreamsToUse);
      exeInfo.stopEvents.resize(numStreamsToUse);
      for (int i = 0; i < numStreamsToUse; ++i)
Gilbert Lee's avatar
Gilbert Lee committed
223
      {
Gilbert Lee's avatar
Gilbert Lee committed
224
225
226
227
228
        HIP_CALL(hipSetDevice(exeIndex));
        HIP_CALL(hipStreamCreate(&exeInfo.streams[i]));
        HIP_CALL(hipEventCreate(&exeInfo.startEvents[i]));
        HIP_CALL(hipEventCreate(&exeInfo.stopEvents[i]));
      }
Gilbert Lee's avatar
Gilbert Lee committed
229

gilbertlee-amd's avatar
gilbertlee-amd committed
230
      if (exeType == EXE_GPU_GFX)
Gilbert Lee's avatar
Gilbert Lee committed
231
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
232
233
234
235
        // Allocate one contiguous chunk of GPU memory for threadblock parameters
        // This allows support for executing one transfer per stream, or all transfers in a single stream
        AllocateMemory(MEM_GPU, exeIndex, exeInfo.totalSubExecs * sizeof(SubExecParam),
                       (void**)&exeInfo.subExecParamGpu);
Gilbert Lee's avatar
Gilbert Lee committed
236
237
238
      }
    }
  }
Gilbert Lee's avatar
Gilbert Lee committed
239

gilbertlee-amd's avatar
gilbertlee-amd committed
240
241
242
243
  if (verbose && !ev.outputToCsv) printf("Test %d:\n", testNum);

  // Prepare input memory and block parameters for current N
  for (auto& exeInfoPair : transferMap)
Gilbert Lee's avatar
Gilbert Lee committed
244
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
245
246
    ExecutorInfo& exeInfo = exeInfoPair.second;
    exeInfo.totalBytes = 0;
Gilbert Lee's avatar
Gilbert Lee committed
247

gilbertlee-amd's avatar
gilbertlee-amd committed
248
249
    int transferOffset = 0;
    for (int i = 0; i < exeInfo.transfers.size(); ++i)
Gilbert Lee's avatar
Gilbert Lee committed
250
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
251
252
      // Prepare subarrays each threadblock works on and fill src memory with patterned data
      Transfer* transfer = exeInfo.transfers[i];
gilbertlee-amd's avatar
gilbertlee-amd committed
253
254
255
      transfer->PrepareSubExecParams(ev);
      transfer->PrepareSrc(ev);
      exeInfo.totalBytes += transfer->numBytesActual;
Gilbert Lee's avatar
Gilbert Lee committed
256

gilbertlee-amd's avatar
gilbertlee-amd committed
257
      // Copy block parameters to GPU for GPU executors
gilbertlee-amd's avatar
gilbertlee-amd committed
258
      if (transfer->exeType == EXE_GPU_GFX)
Gilbert Lee's avatar
Gilbert Lee committed
259
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
260
261
262
263
        exeInfo.transfers[i]->subExecParamGpuPtr = exeInfo.subExecParamGpu + transferOffset;
        HIP_CALL(hipMemcpy(&exeInfo.subExecParamGpu[transferOffset],
                           transfer->subExecParam.data(),
                           transfer->subExecParam.size() * sizeof(SubExecParam),
gilbertlee-amd's avatar
gilbertlee-amd committed
264
                           hipMemcpyHostToDevice));
gilbertlee-amd's avatar
gilbertlee-amd committed
265
266

        transferOffset += transfer->subExecParam.size();
Gilbert Lee's avatar
Gilbert Lee committed
267
268
      }
    }
gilbertlee-amd's avatar
gilbertlee-amd committed
269
  }
Gilbert Lee's avatar
Gilbert Lee committed
270

gilbertlee-amd's avatar
gilbertlee-amd committed
271
272
273
274
275
276
  // Launch kernels (warmup iterations are not counted)
  double totalCpuTime = 0;
  size_t numTimedIterations = 0;
  std::stack<std::thread> threads;
  for (int iteration = -ev.numWarmups; ; iteration++)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
277
    if (ev.numIterations > 0 && iteration    >= ev.numIterations) break;
gilbertlee-amd's avatar
gilbertlee-amd committed
278
    if (ev.numIterations < 0 && totalCpuTime > -ev.numIterations) break;
Gilbert Lee's avatar
Gilbert Lee committed
279

gilbertlee-amd's avatar
gilbertlee-amd committed
280
281
    // Pause before starting first timed iteration in interactive mode
    if (verbose && ev.useInteractive && iteration == 0)
Gilbert Lee's avatar
Gilbert Lee committed
282
    {
283
284
285
286
      printf("Memory prepared:\n");

      for (Transfer& transfer : transfers)
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
287
288
289
290
291
        printf("Transfer %03d:\n", transfer.transferIndex);
        for (int iSrc = 0; iSrc < transfer.numSrcs; ++iSrc)
          printf("  SRC %0d: %p\n", iSrc, transfer.srcMem[iSrc]);
        for (int iDst = 0; iDst < transfer.numDsts; ++iDst)
          printf("  DST %0d: %p\n", iDst, transfer.dstMem[iDst]);
292
      }
gilbertlee-amd's avatar
gilbertlee-amd committed
293
      printf("Hit <Enter> to continue: ");
Gilbert Lee's avatar
Gilbert Lee committed
294
295
296
      scanf("%*c");
      printf("\n");
    }
Gilbert Lee's avatar
Gilbert Lee committed
297

gilbertlee-amd's avatar
gilbertlee-amd committed
298
299
300
301
302
    // Start CPU timing for this iteration
    auto cpuStart = std::chrono::high_resolution_clock::now();

    // Execute all Transfers in parallel
    for (auto& exeInfoPair : transferMap)
303
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
304
      ExecutorInfo& exeInfo = exeInfoPair.second;
gilbertlee-amd's avatar
gilbertlee-amd committed
305
306
307
      ExeType       exeType = exeInfoPair.first.first;
      int const numTransfersToRun = (exeType == EXE_GPU_GFX && ev.useSingleStream) ? 1 : exeInfo.transfers.size();

gilbertlee-amd's avatar
gilbertlee-amd committed
308
309
      for (int i = 0; i < numTransfersToRun; ++i)
        threads.push(std::thread(RunTransfer, std::ref(ev), iteration, std::ref(exeInfo), i));
310
    }
Gilbert Lee's avatar
Gilbert Lee committed
311

gilbertlee-amd's avatar
gilbertlee-amd committed
312
313
314
315
316
317
318
    // Wait for all threads to finish
    int const numTransfers = threads.size();
    for (int i = 0; i < numTransfers; i++)
    {
      threads.top().join();
      threads.pop();
    }
Gilbert Lee's avatar
Gilbert Lee committed
319

gilbertlee-amd's avatar
gilbertlee-amd committed
320
321
322
323
324
    // Stop CPU timing for this iteration
    auto cpuDelta = std::chrono::high_resolution_clock::now() - cpuStart;
    double deltaSec = std::chrono::duration_cast<std::chrono::duration<double>>(cpuDelta).count();

    if (iteration >= 0)
Gilbert Lee's avatar
Gilbert Lee committed
325
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
326
327
328
329
      ++numTimedIterations;
      totalCpuTime += deltaSec;
    }
  }
Gilbert Lee's avatar
Gilbert Lee committed
330

gilbertlee-amd's avatar
gilbertlee-amd committed
331
332
333
334
335
336
337
  // Pause for interactive mode
  if (verbose && ev.useInteractive)
  {
    printf("Transfers complete. Hit <Enter> to continue: ");
    scanf("%*c");
    printf("\n");
  }
Gilbert Lee's avatar
Gilbert Lee committed
338

gilbertlee-amd's avatar
gilbertlee-amd committed
339
340
341
342
343
344
  // Validate that each transfer has transferred correctly
  size_t totalBytesTransferred = 0;
  int const numTransfers = transferList.size();
  for (auto transferPair : transferList)
  {
    Transfer* transfer = transferPair.second;
gilbertlee-amd's avatar
gilbertlee-amd committed
345
346
    transfer->ValidateDst(ev);
    totalBytesTransferred += transfer->numBytesActual;
gilbertlee-amd's avatar
gilbertlee-amd committed
347
  }
Gilbert Lee's avatar
Gilbert Lee committed
348

gilbertlee-amd's avatar
gilbertlee-amd committed
349
350
351
352
  // Report timings
  totalCpuTime = totalCpuTime / (1.0 * numTimedIterations) * 1000;
  double totalBandwidthGbs = (totalBytesTransferred / 1.0E6) / totalCpuTime;
  double maxGpuTime = 0;
Gilbert Lee's avatar
Gilbert Lee committed
353

gilbertlee-amd's avatar
gilbertlee-amd committed
354
355
356
357
  if (ev.useSingleStream)
  {
    for (auto& exeInfoPair : transferMap)
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
358
359
360
      ExecutorInfo  exeInfo  = exeInfoPair.second;
      ExeType const exeType  = exeInfoPair.first.first;
      int     const exeIndex = exeInfoPair.first.second;
Gilbert Lee's avatar
Gilbert Lee committed
361

gilbertlee-amd's avatar
gilbertlee-amd committed
362
363
      // Compute total time for non GPU executors
      if (exeType != EXE_GPU_GFX)
gilbertlee-amd's avatar
gilbertlee-amd committed
364
365
366
367
368
      {
        exeInfo.totalTime = 0;
        for (auto const& transfer : exeInfo.transfers)
          exeInfo.totalTime = std::max(exeInfo.totalTime, transfer->transferTime);
      }
369

gilbertlee-amd's avatar
gilbertlee-amd committed
370
371
372
      double exeDurationMsec = exeInfo.totalTime / (1.0 * numTimedIterations);
      double exeBandwidthGbs = (exeInfo.totalBytes / 1.0E9) / exeDurationMsec * 1000.0f;
      maxGpuTime = std::max(maxGpuTime, exeDurationMsec);
Gilbert Lee's avatar
Gilbert Lee committed
373

gilbertlee-amd's avatar
gilbertlee-amd committed
374
375
      if (verbose && !ev.outputToCsv)
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
376
377
        printf(" Executor: %3s %02d | %7.3f GB/s | %8.3f ms | %12lu bytes\n",
               ExeTypeName[exeType], exeIndex, exeBandwidthGbs, exeDurationMsec, exeInfo.totalBytes);
Gilbert Lee's avatar
Gilbert Lee committed
378
      }
gilbertlee-amd's avatar
gilbertlee-amd committed
379
380
381

      int totalCUs = 0;
      for (auto const& transfer : exeInfo.transfers)
Gilbert Lee's avatar
Gilbert Lee committed
382
      {
Gilbert Lee's avatar
Gilbert Lee committed
383
        double transferDurationMsec = transfer->transferTime / (1.0 * numTimedIterations);
gilbertlee-amd's avatar
gilbertlee-amd committed
384
385
        double transferBandwidthGbs = (transfer->numBytesActual / 1.0E9) / transferDurationMsec * 1000.0f;
        totalCUs += transfer->numSubExecs;
gilbertlee-amd's avatar
gilbertlee-amd committed
386
387

        if (!verbose) continue;
Gilbert Lee's avatar
Gilbert Lee committed
388
389
        if (!ev.outputToCsv)
        {
gilbertlee-amd's avatar
gilbertlee-amd committed
390
          printf("     Transfer %02d  | %7.3f GB/s | %8.3f ms | %12lu bytes | %s -> %s%02d:%03d -> %s\n",
Gilbert Lee's avatar
Gilbert Lee committed
391
                 transfer->transferIndex,
gilbertlee-amd's avatar
gilbertlee-amd committed
392
393
                 transferBandwidthGbs,
                 transferDurationMsec,
gilbertlee-amd's avatar
gilbertlee-amd committed
394
395
396
397
398
                 transfer->numBytesActual,
                 transfer->SrcToStr().c_str(),
                 ExeTypeName[transfer->exeType], transfer->exeIndex,
                 transfer->numSubExecs,
                 transfer->DstToStr().c_str());
Gilbert Lee's avatar
Gilbert Lee committed
399
400
401
        }
        else
        {
gilbertlee-amd's avatar
gilbertlee-amd committed
402
403
404
405
406
407
          printf("%d,%d,%lu,%s,%c%02d,%s,%d,%.3f,%.3f,%s,%s\n",
                 testNum, transfer->transferIndex, transfer->numBytesActual,
                 transfer->SrcToStr().c_str(),
                 MemTypeStr[transfer->exeType], transfer->exeIndex,
                 transfer->DstToStr().c_str(),
                 transfer->numSubExecs,
Gilbert Lee's avatar
Gilbert Lee committed
408
                 transferBandwidthGbs, transferDurationMsec,
gilbertlee-amd's avatar
gilbertlee-amd committed
409
410
                 PtrVectorToStr(transfer->srcMem, initOffset).c_str(),
                 PtrVectorToStr(transfer->dstMem, initOffset).c_str());
Gilbert Lee's avatar
Gilbert Lee committed
411
        }
Gilbert Lee's avatar
Gilbert Lee committed
412
      }
gilbertlee-amd's avatar
gilbertlee-amd committed
413
414
415

      if (verbose && ev.outputToCsv)
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
416
        printf("%d,ALL,%lu,ALL,%c%02d,ALL,%d,%.3f,%.3f,ALL,ALL\n",
gilbertlee-amd's avatar
gilbertlee-amd committed
417
               testNum, totalBytesTransferred,
gilbertlee-amd's avatar
gilbertlee-amd committed
418
               MemTypeStr[exeType], exeIndex, totalCUs,
gilbertlee-amd's avatar
gilbertlee-amd committed
419
420
               exeBandwidthGbs, exeDurationMsec);
      }
Gilbert Lee's avatar
Gilbert Lee committed
421
    }
gilbertlee-amd's avatar
gilbertlee-amd committed
422
423
424
425
426
427
428
  }
  else
  {
    for (auto const& transferPair : transferList)
    {
      Transfer* transfer = transferPair.second;
      double transferDurationMsec = transfer->transferTime / (1.0 * numTimedIterations);
gilbertlee-amd's avatar
gilbertlee-amd committed
429
      double transferBandwidthGbs = (transfer->numBytesActual / 1.0E9) / transferDurationMsec * 1000.0f;
gilbertlee-amd's avatar
gilbertlee-amd committed
430
431
432
433
      maxGpuTime = std::max(maxGpuTime, transferDurationMsec);
      if (!verbose) continue;
      if (!ev.outputToCsv)
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
434
        printf(" Transfer %02d      | %7.3f GB/s | %8.3f ms | %12lu bytes | %s -> %s%02d:%03d -> %s\n",
gilbertlee-amd's avatar
gilbertlee-amd committed
435
436
               transfer->transferIndex,
               transferBandwidthGbs, transferDurationMsec,
gilbertlee-amd's avatar
gilbertlee-amd committed
437
438
439
440
441
               transfer->numBytesActual,
               transfer->SrcToStr().c_str(),
               ExeTypeName[transfer->exeType], transfer->exeIndex,
               transfer->numSubExecs,
               transfer->DstToStr().c_str());
gilbertlee-amd's avatar
gilbertlee-amd committed
442
443
444
      }
      else
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
445
446
447
448
449
450
        printf("%d,%d,%lu,%s,%s%02d,%s,%d,%.3f,%.3f,%s,%s\n",
               testNum, transfer->transferIndex, transfer->numBytesActual,
               transfer->SrcToStr().c_str(),
               ExeTypeName[transfer->exeType], transfer->exeIndex,
               transfer->DstToStr().c_str(),
               transfer->numSubExecs,
gilbertlee-amd's avatar
gilbertlee-amd committed
451
               transferBandwidthGbs, transferDurationMsec,
gilbertlee-amd's avatar
gilbertlee-amd committed
452
453
               PtrVectorToStr(transfer->srcMem, initOffset).c_str(),
               PtrVectorToStr(transfer->dstMem, initOffset).c_str());
gilbertlee-amd's avatar
gilbertlee-amd committed
454
455
456
      }
    }
  }
Gilbert Lee's avatar
Gilbert Lee committed
457

gilbertlee-amd's avatar
gilbertlee-amd committed
458
459
460
  // Display aggregate statistics
  if (verbose)
  {
Gilbert Lee's avatar
Gilbert Lee committed
461
    if (!ev.outputToCsv)
Gilbert Lee's avatar
Gilbert Lee committed
462
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
463
      printf(" Aggregate (CPU)  | %7.3f GB/s | %8.3f ms | %12lu bytes | Overhead: %.3f ms\n",
464
             totalBandwidthGbs, totalCpuTime, totalBytesTransferred, totalCpuTime - maxGpuTime);
Gilbert Lee's avatar
Gilbert Lee committed
465
466
467
    }
    else
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
468
      printf("%d,ALL,%lu,ALL,ALL,ALL,ALL,%.3f,%.3f,ALL,ALL\n",
469
             testNum, totalBytesTransferred, totalBandwidthGbs, totalCpuTime);
Gilbert Lee's avatar
Gilbert Lee committed
470
471
    }
  }
Gilbert Lee's avatar
Gilbert Lee committed
472

Gilbert Lee's avatar
Gilbert Lee committed
473
474
475
  // Release GPU memory
  for (auto exeInfoPair : transferMap)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
476
477
478
479
    ExecutorInfo& exeInfo  = exeInfoPair.second;
    ExeType const exeType  = exeInfoPair.first.first;
    int     const exeIndex = RemappedIndex(exeInfoPair.first.second, IsCpuType(exeType));

Gilbert Lee's avatar
Gilbert Lee committed
480
481
    for (auto& transfer : exeInfo.transfers)
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
482
483
484
485
486
487
488
489
490
491
492
      for (int iSrc = 0; iSrc < transfer->numSrcs; ++iSrc)
      {
        MemType const& srcType = transfer->srcType[iSrc];
        DeallocateMemory(srcType, transfer->srcMem[iSrc], transfer->numBytesActual + ev.byteOffset);
      }
      for (int iDst = 0; iDst < transfer->numDsts; ++iDst)
      {
        MemType const& dstType = transfer->dstType[iDst];
        DeallocateMemory(dstType, transfer->dstMem[iDst], transfer->numBytesActual + ev.byteOffset);
      }
      transfer->subExecParam.clear();
Gilbert Lee's avatar
Gilbert Lee committed
493
494
    }

gilbertlee-amd's avatar
gilbertlee-amd committed
495
    if (IsGpuType(exeType))
Gilbert Lee's avatar
Gilbert Lee committed
496
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
497
498
      int const numStreams = (int)exeInfo.streams.size();
      for (int i = 0; i < numStreams; ++i)
Gilbert Lee's avatar
Gilbert Lee committed
499
      {
Gilbert Lee's avatar
Gilbert Lee committed
500
501
502
        HIP_CALL(hipEventDestroy(exeInfo.startEvents[i]));
        HIP_CALL(hipEventDestroy(exeInfo.stopEvents[i]));
        HIP_CALL(hipStreamDestroy(exeInfo.streams[i]));
Gilbert Lee's avatar
Gilbert Lee committed
503
      }
gilbertlee-amd's avatar
gilbertlee-amd committed
504
505
506
507
508

      if (exeType == EXE_GPU_GFX)
      {
        DeallocateMemory(MEM_GPU, exeInfo.subExecParamGpu);
      }
Gilbert Lee's avatar
Gilbert Lee committed
509
510
511
512
513
514
    }
  }
}

void DisplayUsage(char const* cmdName)
{
Gilbert Lee's avatar
Gilbert Lee committed
515
  printf("TransferBench v%s\n", TB_VERSION);
Gilbert Lee's avatar
Gilbert Lee committed
516
517
518
519
520
521
522
523
524
525
526
527
528
  printf("========================================\n");

  if (numa_available() == -1)
  {
    printf("[ERROR] NUMA library not supported. Check to see if libnuma has been installed on this system\n");
    exit(1);
  }
  int numGpuDevices;
  HIP_CALL(hipGetDeviceCount(&numGpuDevices));
  int const numCpuDevices = numa_num_configured_nodes();

  printf("Usage: %s config <N>\n", cmdName);
  printf("  config: Either:\n");
Gilbert Lee's avatar
Gilbert Lee committed
529
  printf("          - Filename of configFile containing Transfers to execute (see example.cfg for format)\n");
gilbertlee-amd's avatar
gilbertlee-amd committed
530
531
532
533
  printf("          - Name of preset config:\n");
  printf("              p2p          - Peer-to-peer benchmark tests\n");
  printf("              sweep/rsweep - Sweep/random sweep across possible sets of Transfers\n");
  printf("                             - 3rd/4th optional args for # GPU SubExecs / # CPU SubExecs per Transfer\n");
Gilbert Lee's avatar
Gilbert Lee committed
534
  printf("  N     : (Optional) Number of bytes to copy per Transfer.\n");
Gilbert Lee's avatar
Gilbert Lee committed
535
  printf("          If not specified, defaults to %lu bytes. Must be a multiple of 4 bytes\n",
Gilbert Lee's avatar
Gilbert Lee committed
536
         DEFAULT_BYTES_PER_TRANSFER);
Gilbert Lee's avatar
Gilbert Lee committed
537
538
539
540
541
542
543
  printf("          If 0 is specified, a range of Ns will be benchmarked\n");
  printf("          May append a suffix ('K', 'M', 'G') for kilobytes / megabytes / gigabytes\n");
  printf("\n");

  EnvVars::DisplayUsage();
}

gilbertlee-amd's avatar
gilbertlee-amd committed
544
int RemappedIndex(int const origIdx, bool const isCpuType)
Gilbert Lee's avatar
Gilbert Lee committed
545
{
546
547
  static std::vector<int> remappingCpu;
  static std::vector<int> remappingGpu;
Gilbert Lee's avatar
Gilbert Lee committed
548

549
550
551
552
553
554
555
556
  // Build CPU remapping on first use
  // Skip numa nodes that are not configured
  if (remappingCpu.empty())
  {
    for (int node = 0; node <= numa_max_node(); node++)
      if (numa_bitmask_isbitset(numa_get_mems_allowed(), node))
        remappingCpu.push_back(node);
  }
Gilbert Lee's avatar
Gilbert Lee committed
557

558
559
  // Build remappingGpu on first use
  if (remappingGpu.empty())
Gilbert Lee's avatar
Gilbert Lee committed
560
561
562
  {
    int numGpuDevices;
    HIP_CALL(hipGetDeviceCount(&numGpuDevices));
563
    remappingGpu.resize(numGpuDevices);
Gilbert Lee's avatar
Gilbert Lee committed
564
565
566
567

    int const usePcieIndexing = getenv("USE_PCIE_INDEX") ? atoi(getenv("USE_PCIE_INDEX")) : 0;
    if (!usePcieIndexing)
    {
568
      // For HIP-based indexing no remappingGpu is necessary
Gilbert Lee's avatar
Gilbert Lee committed
569
      for (int i = 0; i < numGpuDevices; ++i)
570
        remappingGpu[i] = i;
Gilbert Lee's avatar
Gilbert Lee committed
571
572
573
574
575
576
577
578
579
580
581
582
583
584
    }
    else
    {
      // Collect PCIe address for each GPU
      std::vector<std::pair<std::string, int>> mapping;
      char pciBusId[20];
      for (int i = 0; i < numGpuDevices; ++i)
      {
        HIP_CALL(hipDeviceGetPCIBusId(pciBusId, 20, i));
        mapping.push_back(std::make_pair(pciBusId, i));
      }
      // Sort GPUs by PCIe address then use that as mapping
      std::sort(mapping.begin(), mapping.end());
      for (int i = 0; i < numGpuDevices; ++i)
585
        remappingGpu[i] = mapping[i].second;
Gilbert Lee's avatar
Gilbert Lee committed
586
587
    }
  }
gilbertlee-amd's avatar
gilbertlee-amd committed
588
  return isCpuType ? remappingCpu[origIdx] : remappingGpu[origIdx];
Gilbert Lee's avatar
Gilbert Lee committed
589
590
591
592
}

void DisplayTopology(bool const outputToCsv)
{
593
  int numCpuDevices = numa_num_configured_nodes();
Gilbert Lee's avatar
Gilbert Lee committed
594
595
596
597
598
  int numGpuDevices;
  HIP_CALL(hipGetDeviceCount(&numGpuDevices));

  if (outputToCsv)
  {
599
    printf("NumCpus,%d\n", numCpuDevices);
Gilbert Lee's avatar
Gilbert Lee committed
600
    printf("NumGpus,%d\n", numGpuDevices);
601
602
603
  }
  else
  {
604
605
    printf("\nDetected topology: %d configured CPU NUMA node(s) [%d total]   %d GPU device(s)\n",
           numa_num_configured_nodes(), numa_max_node() + 1, numGpuDevices);
606
607
608
609
610
611
612
613
  }

  // Print out detected CPU topology
  if (outputToCsv)
  {
    printf("NUMA");
    for (int j = 0; j < numCpuDevices; j++)
      printf(",NUMA%02d", j);
614
    printf(",# CPUs,ClosestGPUs,ActualNode\n");
615
616
617
  }
  else
  {
618
    printf("            |");
619
    for (int j = 0; j < numCpuDevices; j++)
620
621
622
623
      printf("NUMA %02d|", j);
    printf(" #Cpus | Closest GPU(s)\n");

    printf("------------+");
624
    for (int j = 0; j <= numCpuDevices; j++)
625
626
      printf("-------+");
    printf("---------------\n");
627
628
629
630
  }

  for (int i = 0; i < numCpuDevices; i++)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
631
    int nodeI = RemappedIndex(i, true);
632
    printf("NUMA %02d (%02d)%s", i, nodeI, outputToCsv ? "," : "|");
633
634
    for (int j = 0; j < numCpuDevices; j++)
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
635
      int nodeJ = RemappedIndex(j, true);
636
      int numaDist = numa_distance(nodeI, nodeJ);
637
      if (outputToCsv)
gilbertlee-amd's avatar
gilbertlee-amd committed
638
        printf("%d,", numaDist);
639
      else
640
        printf(" %5d |", numaDist);
641
642
643
644
    }

    int numCpus = 0;
    for (int j = 0; j < numa_num_configured_cpus(); j++)
645
      if (numa_node_of_cpu(j) == nodeI) numCpus++;
646
647
648
    if (outputToCsv)
      printf("%d,", numCpus);
    else
649
      printf(" %5d | ", numCpus);
650
651
652
653

    bool isFirst = true;
    for (int j = 0; j < numGpuDevices; j++)
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
654
      if (GetClosestNumaNode(RemappedIndex(j, false)) == i)
655
656
      {
        if (isFirst) isFirst = false;
gilbertlee-amd's avatar
gilbertlee-amd committed
657
658
        else printf(",");
        printf("%d", j);
659
660
661
662
663
664
665
666
667
      }
    }
    printf("\n");
  }
  printf("\n");

  // Print out detected GPU topology
  if (outputToCsv)
  {
Gilbert Lee's avatar
Gilbert Lee committed
668
669
670
671
672
673
674
    printf("GPU");
    for (int j = 0; j < numGpuDevices; j++)
      printf(",GPU %02d", j);
    printf(",PCIe Bus ID,ClosestNUMA\n");
  }
  else
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
675
676
677
678
679
680
681
682
683
684
    printf("        |");
    for (int j = 0; j < numGpuDevices; j++)
    {
      hipDeviceProp_t prop;
      HIP_CALL(hipGetDeviceProperties(&prop, j));
      std::string fullName = prop.gcnArchName;
      std::string archName = fullName.substr(0, fullName.find(':'));
      printf(" %6s |", archName.c_str());
    }
    printf("\n");
Gilbert Lee's avatar
Gilbert Lee committed
685
686
687
    printf("        |");
    for (int j = 0; j < numGpuDevices; j++)
      printf(" GPU %02d |", j);
gilbertlee-amd's avatar
gilbertlee-amd committed
688
    printf(" PCIe Bus ID  | #CUs | Closest NUMA\n");
Gilbert Lee's avatar
Gilbert Lee committed
689
690
    for (int j = 0; j <= numGpuDevices; j++)
      printf("--------+");
gilbertlee-amd's avatar
gilbertlee-amd committed
691
    printf("--------------+------+-------------\n");
Gilbert Lee's avatar
Gilbert Lee committed
692
693
694
695
696
697
  }

  char pciBusId[20];

  for (int i = 0; i < numGpuDevices; i++)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
698
    int const deviceIdx = RemappedIndex(i, false);
Gilbert Lee's avatar
Gilbert Lee committed
699
700
701
702
703
704
705
706
707
708
709
710
711
    printf("%sGPU %02d%s", outputToCsv ? "" : " ", i, outputToCsv ? "," : " |");
    for (int j = 0; j < numGpuDevices; j++)
    {
      if (i == j)
      {
        if (outputToCsv)
          printf("-,");
        else
          printf("    -   |");
      }
      else
      {
        uint32_t linkType, hopCount;
gilbertlee-amd's avatar
gilbertlee-amd committed
712
713
        HIP_CALL(hipExtGetLinkTypeAndHopCount(deviceIdx,
                                              RemappedIndex(j, false),
Gilbert Lee's avatar
Gilbert Lee committed
714
715
716
717
718
719
720
721
722
723
724
                                              &linkType, &hopCount));
        printf("%s%s-%d%s",
               outputToCsv ? "" : " ",
               linkType == HSA_AMD_LINK_INFO_TYPE_HYPERTRANSPORT ? "  HT" :
               linkType == HSA_AMD_LINK_INFO_TYPE_QPI            ? " QPI" :
               linkType == HSA_AMD_LINK_INFO_TYPE_PCIE           ? "PCIE" :
               linkType == HSA_AMD_LINK_INFO_TYPE_INFINBAND      ? "INFB" :
               linkType == HSA_AMD_LINK_INFO_TYPE_XGMI           ? "XGMI" : "????",
               hopCount, outputToCsv ? "," : " |");
      }
    }
gilbertlee-amd's avatar
gilbertlee-amd committed
725
726
727
728
729
    HIP_CALL(hipDeviceGetPCIBusId(pciBusId, 20, deviceIdx));

    int numDeviceCUs = 0;
    HIP_CALL(hipDeviceGetAttribute(&numDeviceCUs, hipDeviceAttributeMultiprocessorCount, deviceIdx));

Gilbert Lee's avatar
Gilbert Lee committed
730
    if (outputToCsv)
gilbertlee-amd's avatar
gilbertlee-amd committed
731
      printf("%s,%d,%d\n", pciBusId, numDeviceCUs, GetClosestNumaNode(deviceIdx));
Gilbert Lee's avatar
Gilbert Lee committed
732
    else
gilbertlee-amd's avatar
gilbertlee-amd committed
733
      printf(" %11s | %4d | %d\n", pciBusId, numDeviceCUs, GetClosestNumaNode(deviceIdx));
Gilbert Lee's avatar
Gilbert Lee committed
734
735
736
  }
}

gilbertlee-amd's avatar
gilbertlee-amd committed
737
738
void ParseMemType(std::string const& token, int const numCpus, int const numGpus,
                  std::vector<MemType>& memTypes, std::vector<int>& memIndices)
Gilbert Lee's avatar
Gilbert Lee committed
739
740
{
  char typeChar;
gilbertlee-amd's avatar
gilbertlee-amd committed
741
742
  int offset = 0, devIndex, inc;
  bool found = false;
Gilbert Lee's avatar
Gilbert Lee committed
743

gilbertlee-amd's avatar
gilbertlee-amd committed
744
745
746
  memTypes.clear();
  memIndices.clear();
  while (sscanf(token.c_str() + offset, " %c %d%n", &typeChar, &devIndex, &inc) == 2)
Gilbert Lee's avatar
Gilbert Lee committed
747
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
748
749
750
751
    offset += inc;
    MemType memType = CharToMemType(typeChar);

    if (IsCpuType(memType) && (devIndex < 0 || devIndex >= numCpus))
Gilbert Lee's avatar
Gilbert Lee committed
752
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
753
      printf("[ERROR] CPU index must be between 0 and %d (instead of %d)\n", numCpus-1, devIndex);
Gilbert Lee's avatar
Gilbert Lee committed
754
755
      exit(1);
    }
gilbertlee-amd's avatar
gilbertlee-amd committed
756
    if (IsGpuType(memType) && (devIndex < 0 || devIndex >= numGpus))
Gilbert Lee's avatar
Gilbert Lee committed
757
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
758
      printf("[ERROR] GPU index must be between 0 and %d (instead of %d)\n", numGpus-1, devIndex);
Gilbert Lee's avatar
Gilbert Lee committed
759
760
      exit(1);
    }
gilbertlee-amd's avatar
gilbertlee-amd committed
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796

    found = true;
    if (memType != MEM_NULL)
    {
      memTypes.push_back(memType);
      memIndices.push_back(devIndex);
    }
  }
  if (!found)
  {
    printf("[ERROR] Unable to parse memory type token %s.  Expected one of %s followed by an index\n",
           token.c_str(), MemTypeStr);
    exit(1);
  }
}

void ParseExeType(std::string const& token, int const numCpus, int const numGpus,
                  ExeType &exeType, int& exeIndex)
{
  char typeChar;
  if (sscanf(token.c_str(), " %c%d", &typeChar, &exeIndex) != 2)
  {
    printf("[ERROR] Unable to parse valid executor token (%s).  Exepected one of %s followed by an index\n",
           token.c_str(), ExeTypeStr);
    exit(1);
  }
  exeType = CharToExeType(typeChar);

  if (IsCpuType(exeType) && (exeIndex < 0 || exeIndex >= numCpus))
  {
    printf("[ERROR] CPU index must be between 0 and %d (instead of %d)\n", numCpus-1, exeIndex);
    exit(1);
  }
  if (IsGpuType(exeType) && (exeIndex < 0 || exeIndex >= numGpus))
  {
    printf("[ERROR] GPU index must be between 0 and %d (instead of %d)\n", numGpus-1, exeIndex);
Gilbert Lee's avatar
Gilbert Lee committed
797
798
799
800
    exit(1);
  }
}

Gilbert Lee's avatar
Gilbert Lee committed
801
// Helper function to parse a list of Transfer definitions
Gilbert Lee's avatar
Gilbert Lee committed
802
void ParseTransfers(char* line, int numCpus, int numGpus, std::vector<Transfer>& transfers)
Gilbert Lee's avatar
Gilbert Lee committed
803
804
805
806
807
{
  // Replace any round brackets or '->' with spaces,
  for (int i = 1; line[i]; i++)
    if (line[i] == '(' || line[i] == ')' || line[i] == '-' || line[i] == '>' ) line[i] = ' ';

Gilbert Lee's avatar
Gilbert Lee committed
808
  transfers.clear();
Gilbert Lee's avatar
Gilbert Lee committed
809

Gilbert Lee's avatar
Gilbert Lee committed
810
  int numTransfers = 0;
Gilbert Lee's avatar
Gilbert Lee committed
811
  std::istringstream iss(line);
Gilbert Lee's avatar
Gilbert Lee committed
812
  iss >> numTransfers;
Gilbert Lee's avatar
Gilbert Lee committed
813
814
815
816
817
  if (iss.fail()) return;

  std::string exeMem;
  std::string srcMem;
  std::string dstMem;
Gilbert Lee's avatar
Gilbert Lee committed
818

gilbertlee-amd's avatar
gilbertlee-amd committed
819
  // If numTransfers < 0, read 5-tuple (srcMem, exeMem, dstMem, #CUs, #Bytes)
Gilbert Lee's avatar
Gilbert Lee committed
820
  // otherwise read triples (srcMem, exeMem, dstMem)
gilbertlee-amd's avatar
gilbertlee-amd committed
821
  bool const advancedMode = (numTransfers < 0);
Gilbert Lee's avatar
Gilbert Lee committed
822
823
  numTransfers = abs(numTransfers);

gilbertlee-amd's avatar
gilbertlee-amd committed
824
  int numSubExecs;
gilbertlee-amd's avatar
gilbertlee-amd committed
825
  if (!advancedMode)
Gilbert Lee's avatar
Gilbert Lee committed
826
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
827
828
    iss >> numSubExecs;
    if (numSubExecs <= 0 || iss.fail())
Gilbert Lee's avatar
Gilbert Lee committed
829
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
830
      printf("Parsing error: Number of blocks to use (%d) must be greater than 0\n", numSubExecs);
Gilbert Lee's avatar
Gilbert Lee committed
831
832
833
834
      exit(1);
    }
  }

gilbertlee-amd's avatar
gilbertlee-amd committed
835
  size_t numBytes = 0;
Gilbert Lee's avatar
Gilbert Lee committed
836
837
838
839
  for (int i = 0; i < numTransfers; i++)
  {
    Transfer transfer;
    transfer.transferIndex = i;
gilbertlee-amd's avatar
gilbertlee-amd committed
840
    transfer.numBytes = 0;
gilbertlee-amd's avatar
gilbertlee-amd committed
841
    transfer.numBytesActual = 0;
gilbertlee-amd's avatar
gilbertlee-amd committed
842
    if (!advancedMode)
Gilbert Lee's avatar
Gilbert Lee committed
843
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
844
845
846
847
848
849
850
851
852
853
      iss >> srcMem >> exeMem >> dstMem;
      if (iss.fail())
      {
        printf("Parsing error: Unable to read valid Transfer %d (SRC EXE DST) triplet\n", i+1);
        exit(1);
      }
    }
    else
    {
      std::string numBytesToken;
gilbertlee-amd's avatar
gilbertlee-amd committed
854
      iss >> srcMem >> exeMem >> dstMem >> numSubExecs >> numBytesToken;
gilbertlee-amd's avatar
gilbertlee-amd committed
855
856
857
858
859
860
861
862
863
864
865
      if (iss.fail())
      {
        printf("Parsing error: Unable to read valid Transfer %d (SRC EXE DST #CU #Bytes) tuple\n", i+1);
        exit(1);
      }
      if (sscanf(numBytesToken.c_str(), "%lu", &numBytes) != 1)
      {
        printf("Parsing error: '%s' is not a valid expression of numBytes for Transfer %d\n", numBytesToken.c_str(), i+1);
        exit(1);
      }
      char units = numBytesToken.back();
gilbertlee-amd's avatar
gilbertlee-amd committed
866
      switch (toupper(units))
gilbertlee-amd's avatar
gilbertlee-amd committed
867
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
868
869
870
      case 'K': numBytes *= 1024; break;
      case 'M': numBytes *= 1024*1024; break;
      case 'G': numBytes *= 1024*1024*1024; break;
gilbertlee-amd's avatar
gilbertlee-amd committed
871
      }
Gilbert Lee's avatar
Gilbert Lee committed
872
    }
Gilbert Lee's avatar
Gilbert Lee committed
873

gilbertlee-amd's avatar
gilbertlee-amd committed
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
    ParseMemType(srcMem, numCpus, numGpus, transfer.srcType, transfer.srcIndex);
    ParseMemType(dstMem, numCpus, numGpus, transfer.dstType, transfer.dstIndex);
    ParseExeType(exeMem, numCpus, numGpus, transfer.exeType, transfer.exeIndex);

    transfer.numSrcs = (int)transfer.srcType.size();
    transfer.numDsts = (int)transfer.dstType.size();
    if (transfer.numSrcs == 0 && transfer.numDsts == 0)
    {
      printf("[ERROR] Transfer must have at least one src or dst\n");
      exit(1);
    }

    if (transfer.exeType == EXE_GPU_DMA && (transfer.numSrcs > 1 || transfer.numDsts > 1))
    {
      printf("[ERROR] GPU DMA executor can only be used for single source / single dst Transfers\n");
      exit(1);
    }

    transfer.numSubExecs = numSubExecs;
gilbertlee-amd's avatar
gilbertlee-amd committed
893
    transfer.numBytes = numBytes;
Gilbert Lee's avatar
Gilbert Lee committed
894
    transfers.push_back(transfer);
Gilbert Lee's avatar
Gilbert Lee committed
895
896
897
898
899
900
901
902
903
904
905
906
907
  }
}

void EnablePeerAccess(int const deviceId, int const peerDeviceId)
{
  int canAccess;
  HIP_CALL(hipDeviceCanAccessPeer(&canAccess, deviceId, peerDeviceId));
  if (!canAccess)
  {
    printf("[ERROR] Unable to enable peer access from GPU devices %d to %d\n", peerDeviceId, deviceId);
    exit(1);
  }
  HIP_CALL(hipSetDevice(deviceId));
Gilbert Lee's avatar
Gilbert Lee committed
908
909
910
911
912
913
914
  hipError_t error = hipDeviceEnablePeerAccess(peerDeviceId, 0);
  if (error != hipSuccess && error != hipErrorPeerAccessAlreadyEnabled)
  {
    printf("[ERROR] Unable to enable peer to peer access from %d to %d (%s)\n",
           deviceId, peerDeviceId, hipGetErrorString(error));
    exit(1);
  }
Gilbert Lee's avatar
Gilbert Lee committed
915
916
917
918
919
920
921
922
923
924
}

void AllocateMemory(MemType memType, int devIndex, size_t numBytes, void** memPtr)
{
  if (numBytes == 0)
  {
    printf("[ERROR] Unable to allocate 0 bytes\n");
    exit(1);
  }

gilbertlee-amd's avatar
gilbertlee-amd committed
925
  if (IsCpuType(memType))
Gilbert Lee's avatar
Gilbert Lee committed
926
927
  {
    // Set numa policy prior to call to hipHostMalloc
928
    numa_set_preferred(devIndex);
Gilbert Lee's avatar
Gilbert Lee committed
929
930
931
932
933
934

    // Allocate host-pinned memory (should respect NUMA mem policy)
    if (memType == MEM_CPU_FINE)
    {
      HIP_CALL(hipHostMalloc((void **)memPtr, numBytes, hipHostMallocNumaUser));
    }
gilbertlee-amd's avatar
gilbertlee-amd committed
935
    else if (memType == MEM_CPU)
Gilbert Lee's avatar
Gilbert Lee committed
936
    {
937
938
939
940
941
      if (hipHostMalloc((void **)memPtr, numBytes, hipHostMallocNumaUser | hipHostMallocNonCoherent) != hipSuccess)
      {
        printf("[ERROR] Unable to allocate non-coherent host memory on NUMA node %d\n", devIndex);
        exit(1);
      }
Gilbert Lee's avatar
Gilbert Lee committed
942
    }
gilbertlee-amd's avatar
gilbertlee-amd committed
943
944
945
946
    else if (memType == MEM_CPU_UNPINNED)
    {
      *memPtr = numa_alloc_onnode(numBytes, devIndex);
    }
Gilbert Lee's avatar
Gilbert Lee committed
947
948

    // Check that the allocated pages are actually on the correct NUMA node
gilbertlee-amd's avatar
gilbertlee-amd committed
949
950
    memset(*memPtr, 0, numBytes);
    CheckPages((char*)*memPtr, numBytes, devIndex);
Gilbert Lee's avatar
Gilbert Lee committed
951
952

    // Reset to default numa mem policy
953
    numa_set_preferred(-1);
Gilbert Lee's avatar
Gilbert Lee committed
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
  }
  else if (memType == MEM_GPU)
  {
    // Allocate GPU memory on appropriate device
    HIP_CALL(hipSetDevice(devIndex));
    HIP_CALL(hipMalloc((void**)memPtr, numBytes));
  }
  else if (memType == MEM_GPU_FINE)
  {
    HIP_CALL(hipSetDevice(devIndex));
    HIP_CALL(hipExtMallocWithFlags((void**)memPtr, numBytes, hipDeviceMallocFinegrained));
  }
  else
  {
    printf("[ERROR] Unsupported memory type %d\n", memType);
    exit(1);
  }
}

gilbertlee-amd's avatar
gilbertlee-amd committed
973
void DeallocateMemory(MemType memType, void* memPtr, size_t const bytes)
Gilbert Lee's avatar
Gilbert Lee committed
974
975
976
977
978
{
  if (memType == MEM_CPU || memType == MEM_CPU_FINE)
  {
    HIP_CALL(hipHostFree(memPtr));
  }
gilbertlee-amd's avatar
gilbertlee-amd committed
979
980
981
982
  else if (memType == MEM_CPU_UNPINNED)
  {
    numa_free(memPtr, bytes);
  }
Gilbert Lee's avatar
Gilbert Lee committed
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
  else if (memType == MEM_GPU || memType == MEM_GPU_FINE)
  {
    HIP_CALL(hipFree(memPtr));
  }
}

void CheckPages(char* array, size_t numBytes, int targetId)
{
  unsigned long const pageSize = getpagesize();
  unsigned long const numPages = (numBytes + pageSize - 1) / pageSize;

  std::vector<void *> pages(numPages);
  std::vector<int> status(numPages);

  pages[0] = array;
  for (int i = 1; i < numPages; i++)
  {
    pages[i] = (char*)pages[i-1] + pageSize;
  }

  long const retCode = move_pages(0, numPages, pages.data(), NULL, status.data(), 0);
  if (retCode)
  {
    printf("[ERROR] Unable to collect page info\n");
    exit(1);
  }

  size_t mistakeCount = 0;
  for (int i = 0; i < numPages; i++)
  {
    if (status[i] < 0)
    {
      printf("[ERROR] Unexpected page status %d for page %d\n", status[i], i);
      exit(1);
    }
    if (status[i] != targetId) mistakeCount++;
  }
  if (mistakeCount > 0)
  {
    printf("[ERROR] %lu out of %lu pages for memory allocation were not on NUMA node %d\n", mistakeCount, numPages, targetId);
    printf("[ERROR] Ensure up-to-date ROCm is installed\n");
    exit(1);
  }
}

1028
void RunTransfer(EnvVars const& ev, int const iteration,
Gilbert Lee's avatar
Gilbert Lee committed
1029
                 ExecutorInfo& exeInfo, int const transferIdx)
Gilbert Lee's avatar
Gilbert Lee committed
1030
{
gilbertlee-amd's avatar
gilbertlee-amd committed
1031
  Transfer* transfer = exeInfo.transfers[transferIdx];
Gilbert Lee's avatar
Gilbert Lee committed
1032

gilbertlee-amd's avatar
gilbertlee-amd committed
1033
  if (transfer->exeType == EXE_GPU_GFX)
Gilbert Lee's avatar
Gilbert Lee committed
1034
1035
  {
    // Switch to executing GPU
gilbertlee-amd's avatar
gilbertlee-amd committed
1036
    int const exeIndex = RemappedIndex(transfer->exeIndex, false);
Gilbert Lee's avatar
Gilbert Lee committed
1037
1038
    HIP_CALL(hipSetDevice(exeIndex));

Gilbert Lee's avatar
Gilbert Lee committed
1039
1040
1041
    hipStream_t& stream     = exeInfo.streams[transferIdx];
    hipEvent_t&  startEvent = exeInfo.startEvents[transferIdx];
    hipEvent_t&  stopEvent  = exeInfo.stopEvents[transferIdx];
Gilbert Lee's avatar
Gilbert Lee committed
1042

gilbertlee-amd's avatar
gilbertlee-amd committed
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
    // Figure out how many threadblocks to use.
    // In single stream mode, all the threadblocks for this GPU are launched
    // Otherwise, just launch the threadblocks associated with this single Transfer
    int const numBlocksToRun = ev.useSingleStream ? exeInfo.totalSubExecs : transfer->numSubExecs;
    hipExtLaunchKernelGGL(GpuKernelTable[ev.gpuKernel],
                          dim3(numBlocksToRun, 1, 1),
                          dim3(BLOCKSIZE, 1, 1),
                          ev.sharedMemBytes, stream,
                          startEvent, stopEvent,
                          0, transfer->subExecParamGpuPtr);
Gilbert Lee's avatar
Gilbert Lee committed
1053
1054
1055

    // Synchronize per iteration, unless in single sync mode, in which case
    // synchronize during last warmup / last actual iteration
Gilbert Lee's avatar
Gilbert Lee committed
1056
    HIP_CALL(hipStreamSynchronize(stream));
Gilbert Lee's avatar
Gilbert Lee committed
1057
1058
1059
1060

    if (iteration >= 0)
    {
      // Record GPU timing
Gilbert Lee's avatar
Gilbert Lee committed
1061
1062
      float gpuDeltaMsec;
      HIP_CALL(hipEventElapsedTime(&gpuDeltaMsec, startEvent, stopEvent));
Gilbert Lee's avatar
Gilbert Lee committed
1063

Gilbert Lee's avatar
Gilbert Lee committed
1064
1065
      if (ev.useSingleStream)
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
1066
        // Figure out individual timings for Transfers that were all launched together
gilbertlee-amd's avatar
gilbertlee-amd committed
1067
        for (Transfer* currTransfer : exeInfo.transfers)
Gilbert Lee's avatar
Gilbert Lee committed
1068
        {
gilbertlee-amd's avatar
gilbertlee-amd committed
1069
1070
1071
          long long minStartCycle = currTransfer->subExecParamGpuPtr[0].startCycle;
          long long maxStopCycle  = currTransfer->subExecParamGpuPtr[0].stopCycle;
          for (int i = 1; i < currTransfer->numSubExecs; i++)
Gilbert Lee's avatar
Gilbert Lee committed
1072
          {
gilbertlee-amd's avatar
gilbertlee-amd committed
1073
1074
            minStartCycle = std::min(minStartCycle, currTransfer->subExecParamGpuPtr[i].startCycle);
            maxStopCycle  = std::max(maxStopCycle,  currTransfer->subExecParamGpuPtr[i].stopCycle);
Gilbert Lee's avatar
Gilbert Lee committed
1075
          }
Gilbert Lee's avatar
Gilbert Lee committed
1076
1077
          int const wallClockRate = GetWallClockRate(exeIndex);
          double iterationTimeMs = (maxStopCycle - minStartCycle) / (double)(wallClockRate);
gilbertlee-amd's avatar
gilbertlee-amd committed
1078
          currTransfer->transferTime += iterationTimeMs;
Gilbert Lee's avatar
Gilbert Lee committed
1079
        }
Gilbert Lee's avatar
Gilbert Lee committed
1080
1081
1082
1083
        exeInfo.totalTime += gpuDeltaMsec;
      }
      else
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
1084
        transfer->transferTime += gpuDeltaMsec;
Gilbert Lee's avatar
Gilbert Lee committed
1085
1086
1087
      }
    }
  }
gilbertlee-amd's avatar
gilbertlee-amd committed
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
  else if (transfer->exeType == EXE_GPU_DMA)
  {
    // Switch to executing GPU
    int const exeIndex = RemappedIndex(transfer->exeIndex, false);
    HIP_CALL(hipSetDevice(exeIndex));

    hipStream_t& stream     = exeInfo.streams[transferIdx];
    hipEvent_t&  startEvent = exeInfo.startEvents[transferIdx];
    hipEvent_t&  stopEvent  = exeInfo.stopEvents[transferIdx];

    HIP_CALL(hipEventRecord(startEvent, stream));
    if (transfer->numSrcs == 0 && transfer->numDsts == 1)
    {
      HIP_CALL(hipMemsetAsync(transfer->dstMem[0],
                              MEMSET_CHAR, transfer->numBytesActual, stream));
    }
    else if (transfer->numSrcs == 1 && transfer->numDsts == 1)
    {
      HIP_CALL(hipMemcpyAsync(transfer->dstMem[0], transfer->srcMem[0],
                              transfer->numBytesActual, hipMemcpyDefault,
                              stream));
    }
    HIP_CALL(hipEventRecord(stopEvent, stream));
    HIP_CALL(hipStreamSynchronize(stream));

    if (iteration >= 0)
    {
      // Record GPU timing
      float gpuDeltaMsec;
      HIP_CALL(hipEventElapsedTime(&gpuDeltaMsec, startEvent, stopEvent));
      transfer->transferTime += gpuDeltaMsec;
    }
  }
  else if (transfer->exeType == EXE_CPU) // CPU execution agent
Gilbert Lee's avatar
Gilbert Lee committed
1122
1123
  {
    // Force this thread and all child threads onto correct NUMA node
gilbertlee-amd's avatar
gilbertlee-amd committed
1124
    int const exeIndex = RemappedIndex(transfer->exeIndex, true);
1125
    if (numa_run_on_node(exeIndex))
Gilbert Lee's avatar
Gilbert Lee committed
1126
    {
1127
      printf("[ERROR] Unable to set CPU to NUMA node %d\n", exeIndex);
Gilbert Lee's avatar
Gilbert Lee committed
1128
1129
1130
1131
1132
1133
1134
      exit(1);
    }

    std::vector<std::thread> childThreads;

    auto cpuStart = std::chrono::high_resolution_clock::now();

gilbertlee-amd's avatar
gilbertlee-amd committed
1135
1136
1137
    // Launch each subExecutor in child-threads to perform memcopies
    for (int i = 0; i < transfer->numSubExecs; ++i)
      childThreads.push_back(std::thread(CpuReduceKernel, std::ref(transfer->subExecParam[i])));
Gilbert Lee's avatar
Gilbert Lee committed
1138
1139

    // Wait for child-threads to finish
gilbertlee-amd's avatar
gilbertlee-amd committed
1140
    for (int i = 0; i < transfer->numSubExecs; ++i)
Gilbert Lee's avatar
Gilbert Lee committed
1141
1142
1143
1144
1145
1146
      childThreads[i].join();

    auto cpuDelta = std::chrono::high_resolution_clock::now() - cpuStart;

    // Record time if not a warmup iteration
    if (iteration >= 0)
gilbertlee-amd's avatar
gilbertlee-amd committed
1147
      transfer->transferTime += (std::chrono::duration_cast<std::chrono::duration<double>>(cpuDelta).count() * 1000.0);
Gilbert Lee's avatar
Gilbert Lee committed
1148
1149
1150
  }
}

gilbertlee-amd's avatar
gilbertlee-amd committed
1151
void RunPeerToPeerBenchmarks(EnvVars const& ev, size_t N)
Gilbert Lee's avatar
Gilbert Lee committed
1152
{
gilbertlee-amd's avatar
gilbertlee-amd committed
1153
1154
  ev.DisplayP2PBenchmarkEnvVars();

Gilbert Lee's avatar
Gilbert Lee committed
1155
  // Collect the number of available CPUs/GPUs on this machine
gilbertlee-amd's avatar
gilbertlee-amd committed
1156
1157
  int const numCpus    = ev.numCpuDevices;
  int const numGpus    = ev.numGpuDevices;
Gilbert Lee's avatar
Gilbert Lee committed
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
  int const numDevices = numCpus + numGpus;

  // Enable peer to peer for each GPU
  for (int i = 0; i < numGpus; i++)
    for (int j = 0; j < numGpus; j++)
      if (i != j) EnablePeerAccess(i, j);

  // Perform unidirectional / bidirectional
  for (int isBidirectional = 0; isBidirectional <= 1; isBidirectional++)
  {
    // Print header
    if (!ev.outputToCsv)
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
1171
1172
1173
1174
1175
1176
1177
1178
      printf("%sdirectional copy peak bandwidth GB/s [%s read / %s write] (GPU-Executor: %s)\n", isBidirectional ? "Bi" : "Uni",
             ev.useRemoteRead ? "Remote" : "Local",
             ev.useRemoteRead ? "Local" : "Remote",
             ev.useDmaCopy    ? "DMA"   : "GFX");

      printf("%10s", "SRC\\DST");
      for (int i = 0; i < numCpus; i++) printf("%7s %02d", "CPU", i);
      for (int i = 0; i < numGpus; i++) printf("%7s %02d", "GPU", i);
Gilbert Lee's avatar
Gilbert Lee committed
1179
1180
1181
1182
1183
1184
      printf("\n");
    }

    // Loop over all possible src/dst pairs
    for (int src = 0; src < numDevices; src++)
    {
gilbertlee-amd's avatar
gilbertlee-amd committed
1185
1186
1187
      MemType const srcType  = (src < numCpus ? MEM_CPU : MEM_GPU);
      int     const srcIndex = (srcType == MEM_CPU ? src : src - numCpus);

Gilbert Lee's avatar
Gilbert Lee committed
1188
      if (!ev.outputToCsv)
gilbertlee-amd's avatar
gilbertlee-amd committed
1189
1190
        printf("%7s %02d", (srcType == MEM_CPU) ? "CPU" : "GPU", srcIndex);

Gilbert Lee's avatar
Gilbert Lee committed
1191
1192
      for (int dst = 0; dst < numDevices; dst++)
      {
gilbertlee-amd's avatar
gilbertlee-amd committed
1193
1194
1195
1196
        MemType const dstType  = (dst < numCpus ? MEM_CPU : MEM_GPU);
        int     const dstIndex = (dstType == MEM_CPU ? dst : dst - numCpus);

        double bandwidth = GetPeakBandwidth(ev, N, isBidirectional, srcType, srcIndex, dstType, dstIndex);
Gilbert Lee's avatar
Gilbert Lee committed
1197
1198
1199
1200
1201
1202
1203
1204
1205
        if (!ev.outputToCsv)
        {
          if (bandwidth == 0)
            printf("%10s", "N/A");
          else
            printf("%10.2f", bandwidth);
        }
        else
        {
gilbertlee-amd's avatar
gilbertlee-amd committed
1206
1207
1208
          printf("%s %02d,%s %02d,%s,%s,%s,%.2f,%lu\n",
                 srcType == MEM_CPU ? "CPU" : "GPU", srcIndex,
                 dstType == MEM_CPU ? "CPU" : "GPU", dstIndex,
Gilbert Lee's avatar
Gilbert Lee committed
1209
                 isBidirectional ? "bidirectional" : "unidirectional",
gilbertlee-amd's avatar
gilbertlee-amd committed
1210
1211
                 ev.useRemoteRead ? "Remote" : "Local",
                 ev.useDmaCopy ? "DMA" : "GFX",
Gilbert Lee's avatar
Gilbert Lee committed
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
                 bandwidth,
                 N * sizeof(float));
        }
        fflush(stdout);
      }
      if (!ev.outputToCsv) printf("\n");
    }
    if (!ev.outputToCsv) printf("\n");
  }
}

gilbertlee-amd's avatar
gilbertlee-amd committed
1223
double GetPeakBandwidth(EnvVars const& ev, size_t const N,
Gilbert Lee's avatar
Gilbert Lee committed
1224
                        int     const  isBidirectional,
gilbertlee-amd's avatar
gilbertlee-amd committed
1225
1226
                        MemType const  srcType, int const srcIndex,
                        MemType const  dstType, int const dstIndex)
Gilbert Lee's avatar
Gilbert Lee committed
1227
1228
{
  // Skip bidirectional on same device
gilbertlee-amd's avatar
gilbertlee-amd committed
1229
  if (isBidirectional && srcType == dstType && srcIndex == dstIndex) return 0.0f;
Gilbert Lee's avatar
Gilbert Lee committed
1230
1231
1232

  int const initOffset = ev.byteOffset / sizeof(float);

Gilbert Lee's avatar
Gilbert Lee committed
1233
  // Prepare Transfers
gilbertlee-amd's avatar
gilbertlee-amd committed
1234
  std::vector<Transfer> transfers(2);
gilbertlee-amd's avatar
gilbertlee-amd committed
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
  transfers[0].numBytes = transfers[1].numBytes = N * sizeof(float);

  // SRC -> DST
  transfers[0].numSrcs = transfers[0].numDsts = 1;
  transfers[0].srcType.push_back(srcType);
  transfers[0].dstType.push_back(dstType);
  transfers[0].srcIndex.push_back(srcIndex);
  transfers[0].dstIndex.push_back(dstIndex);

  // DST -> SRC
  transfers[1].numSrcs = transfers[1].numDsts = 1;
  transfers[1].srcType.push_back(dstType);
  transfers[1].dstType.push_back(srcType);
  transfers[1].srcIndex.push_back(dstIndex);
  transfers[1].dstIndex.push_back(srcIndex);
Gilbert Lee's avatar
Gilbert Lee committed
1250
1251

  // Either perform (local read + remote write), or (remote read + local write)
gilbertlee-amd's avatar
gilbertlee-amd committed
1252
1253
1254
1255
1256
1257
1258
1259
1260
  ExeType gpuExeType = ev.useDmaCopy ? EXE_GPU_DMA : EXE_GPU_GFX;
  transfers[0].exeType = IsGpuType(ev.useRemoteRead ? dstType : srcType) ? gpuExeType : EXE_CPU;
  transfers[1].exeType = IsGpuType(ev.useRemoteRead ? srcType : dstType) ? gpuExeType : EXE_CPU;
  transfers[0].exeIndex = (ev.useRemoteRead ? dstIndex : srcIndex);
  transfers[1].exeIndex = (ev.useRemoteRead ? srcIndex : dstIndex);
  transfers[0].numSubExecs = IsGpuType(transfers[0].exeType) ? ev.numGpuSubExecs : ev.numCpuSubExecs;
  transfers[1].numSubExecs = IsGpuType(transfers[0].exeType) ? ev.numGpuSubExecs : ev.numCpuSubExecs;

  // Remove (DST->SRC) if not bidirectional
gilbertlee-amd's avatar
gilbertlee-amd committed
1261
  transfers.resize(isBidirectional + 1);
Gilbert Lee's avatar
Gilbert Lee committed
1262

1263
1264
1265
  // Abort if executing on NUMA node with no CPUs
  for (int i = 0; i <= isBidirectional; i++)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
1266
    if (transfers[i].exeType == EXE_CPU && ev.numCpusPerNuma[transfers[i].exeIndex] == 0)
1267
1268
1269
      return 0;
  }

gilbertlee-amd's avatar
gilbertlee-amd committed
1270
  ExecuteTransfers(ev, 0, N, transfers, false);
Gilbert Lee's avatar
Gilbert Lee committed
1271
1272
1273
1274
1275

  // Collect aggregate bandwidth
  double totalBandwidth = 0;
  for (int i = 0; i <= isBidirectional; i++)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
1276
    double transferDurationMsec = transfers[i].transferTime / (1.0 * ev.numIterations);
gilbertlee-amd's avatar
gilbertlee-amd committed
1277
    double transferBandwidthGbs = (transfers[i].numBytesActual / 1.0E9) / transferDurationMsec * 1000.0f;
Gilbert Lee's avatar
Gilbert Lee committed
1278
    totalBandwidth += transferBandwidthGbs;
Gilbert Lee's avatar
Gilbert Lee committed
1279
1280
1281
1282
  }
  return totalBandwidth;
}

gilbertlee-amd's avatar
gilbertlee-amd committed
1283
void Transfer::PrepareSubExecParams(EnvVars const& ev)
Gilbert Lee's avatar
Gilbert Lee committed
1284
{
gilbertlee-amd's avatar
gilbertlee-amd committed
1285
1286
1287
1288
1289
1290
1291
  // Each subExecutor needs to know src/dst pointers and how many elements to transfer
  // Figure out the sub-array each subExecutor works on for this Transfer
  // - Partition N as evenly as possible, but try to keep subarray sizes as multiples of BLOCK_BYTES bytes,
  //   except the very last one, for alignment reasons
  size_t const N              = this->numBytesActual / sizeof(float);
  int    const initOffset     = ev.byteOffset / sizeof(float);
  int    const targetMultiple = ev.blockBytes / sizeof(float);
Gilbert Lee's avatar
Gilbert Lee committed
1292

gilbertlee-amd's avatar
gilbertlee-amd committed
1293
1294
1295
1296
1297
  // In some cases, there may not be enough data for all subExectors
  int const maxSubExecToUse = std::min((int)(N + targetMultiple - 1) / targetMultiple, this->numSubExecs);

  this->subExecParam.clear();
  this->subExecParam.resize(this->numSubExecs);
Gilbert Lee's avatar
Gilbert Lee committed
1298
1299

  size_t assigned = 0;
gilbertlee-amd's avatar
gilbertlee-amd committed
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
  for (int i = 0; i < this->numSubExecs; ++i)
  {
    int    const subExecLeft = std::max(0, maxSubExecToUse - i);
    size_t const leftover    = N - assigned;
    size_t const roundedN    = (leftover + targetMultiple - 1) / targetMultiple;

    SubExecParam& p = this->subExecParam[i];
    p.N             = subExecLeft ? std::min(leftover, ((roundedN / subExecLeft) * targetMultiple)) : 0;
    p.numSrcs       = this->numSrcs;
    p.numDsts       = this->numDsts;
    for (int iSrc = 0; iSrc < this->numSrcs; ++iSrc)
      p.src[iSrc] = this->srcMem[iSrc] + assigned + initOffset;
    for (int iDst = 0; iDst < this->numDsts; ++iDst)
      p.dst[iDst] = this->dstMem[iDst] + assigned + initOffset;

    if (ev.enableDebug)
    {
      printf("Transfer %02d SE:%02d: %10lu floats: %10lu to %10lu\n",
             this->transferIndex, i, p.N, assigned, assigned + p.N);
    }
Gilbert Lee's avatar
Gilbert Lee committed
1320

gilbertlee-amd's avatar
gilbertlee-amd committed
1321
1322
1323
    p.startCycle = 0;
    p.stopCycle  = 0;
    assigned += p.N;
Gilbert Lee's avatar
Gilbert Lee committed
1324
1325
  }

Gilbert Lee's avatar
Gilbert Lee committed
1326
  this->transferTime = 0.0;
Gilbert Lee's avatar
Gilbert Lee committed
1327
1328
}

gilbertlee-amd's avatar
gilbertlee-amd committed
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
void Transfer::PrepareReference(EnvVars const& ev, std::vector<float>& buffer, int bufferIdx)
{
  size_t N = buffer.size();
  if (bufferIdx >= 0)
  {
    size_t patternLen = ev.fillPattern.size();
    if (patternLen > 0)
    {
      for (size_t i = 0; i < N; ++i)
        buffer[i] = ev.fillPattern[i % patternLen];
    }
    else
    {
      for (size_t i = 0; i < N; ++i)
        buffer[i] = (i % 383 + 31) * (bufferIdx + 1);
    }
  }
  else // Destination buffer
  {
    if (this->numSrcs == 0)
    {
      // Note: 0x75757575 = 13323083.0
      memset(buffer.data(), MEMSET_CHAR, N * sizeof(float));
    }
    else
    {
      PrepareReference(ev, buffer, 0);

      if (this->numSrcs > 1)
      {
        std::vector<float> temp(N);
        for (int srcIdx = 1; srcIdx < this->numSrcs; ++srcIdx)
        {
          PrepareReference(ev, temp, srcIdx);
          for (int i = 0; i < N; ++i)
          {
            buffer[i] += temp[i];
          }
        }
      }
    }
  }
}

void Transfer::PrepareSrc(EnvVars const& ev)
{
  if (this->numSrcs == 0) return;
  size_t const N = this->numBytesActual / sizeof(float);
  int const initOffset = ev.byteOffset / sizeof(float);

  std::vector<float> reference(N);
  for (int srcIdx = 0; srcIdx < this->numSrcs; ++srcIdx)
  {
    //PrepareReference(ev, reference, srcIdx);
    PrepareReference(ev, reference, srcIdx);
    HIP_CALL(hipMemcpy(this->srcMem[srcIdx] + initOffset, reference.data(), this->numBytesActual, hipMemcpyDefault));
  }
}

void Transfer::ValidateDst(EnvVars const& ev)
{
  if (this->numDsts == 0) return;
  size_t const N = this->numBytesActual / sizeof(float);
  int const initOffset = ev.byteOffset / sizeof(float);

  std::vector<float> reference(N);
  PrepareReference(ev, reference, -1);

  std::vector<float> hostBuffer(N);
  for (int dstIdx = 0; dstIdx < this->numDsts; ++dstIdx)
  {
    float* output;
    if (IsCpuType(this->dstType[dstIdx]))
    {
      output = this->dstMem[dstIdx] + initOffset;
    }
    else
    {
      HIP_CALL(hipMemcpy(hostBuffer.data(), this->dstMem[dstIdx] + initOffset, this->numBytesActual, hipMemcpyDefault));
      output = hostBuffer.data();
    }

    for (size_t i = 0; i < N; ++i)
    {
      if (reference[i] != output[i])
      {
        printf("\n[ERROR] Destination array %d value at index %lu (%.3f) does not match expected value (%.3f)\n",
               dstIdx, i, output[i], reference[i]);
        printf("[ERROR] Failed Transfer details: #%d: %s -> [%c%d:%d] -> %s\n",
               this->transferIndex,
               this->SrcToStr().c_str(),
               ExeTypeStr[this->exeType], this->exeIndex,
               this->numSubExecs,
               this->DstToStr().c_str());
        exit(1);
      }
    }
  }
}

std::string Transfer::SrcToStr() const
{
  if (numSrcs == 0) return "N";
  std::stringstream ss;
  for (int i = 0; i < numSrcs; ++i)
    ss << MemTypeStr[srcType[i]] << srcIndex[i];
  return ss.str();
}

std::string Transfer::DstToStr() const
{
  if (numDsts == 0) return "N";
  std::stringstream ss;
  for (int i = 0; i < numDsts; ++i)
    ss << MemTypeStr[dstType[i]] << dstIndex[i];
  return ss.str();
}

Gilbert Lee's avatar
Gilbert Lee committed
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
// NOTE: This is a stop-gap solution until HIP provides wallclock values
int GetWallClockRate(int deviceId)
{
  static std::vector<int> wallClockPerDeviceMhz;

  if (wallClockPerDeviceMhz.size() == 0)
  {
    int numGpuDevices;
    HIP_CALL(hipGetDeviceCount(&numGpuDevices));
    wallClockPerDeviceMhz.resize(numGpuDevices);

    hipDeviceProp_t prop;
    for (int i = 0; i < numGpuDevices; i++)
    {
      HIP_CALL(hipGetDeviceProperties(&prop, i));
      int value = 25000;
      switch (prop.gcnArch)
      {
      case 906: case 910: value = 25000; break;
      default:
        printf("Unrecognized GCN arch %d\n", prop.gcnArch);
      }
      wallClockPerDeviceMhz[i] = value;
    }
  }
  return wallClockPerDeviceMhz[deviceId];
}
Gilbert Lee's avatar
Gilbert Lee committed
1474

gilbertlee-amd's avatar
gilbertlee-amd committed
1475
void RunSweepPreset(EnvVars const& ev, size_t const numBytesPerTransfer, int const numGpuSubExecs, int const numCpuSubExecs, bool const isRandom)
Gilbert Lee's avatar
Gilbert Lee committed
1476
1477
1478
1479
{
  ev.DisplaySweepEnvVars();

  // Compute how many possible Transfers are permitted (unique SRC/EXE/DST triplets)
gilbertlee-amd's avatar
gilbertlee-amd committed
1480
  std::vector<std::pair<ExeType, int>> exeList;
Gilbert Lee's avatar
Gilbert Lee committed
1481
1482
  for (auto exe : ev.sweepExe)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
1483
1484
    ExeType const exeType = CharToExeType(exe);
    if (IsGpuType(exeType))
Gilbert Lee's avatar
Gilbert Lee committed
1485
    {
1486
      for (int exeIndex = 0; exeIndex < ev.numGpuDevices; ++exeIndex)
gilbertlee-amd's avatar
gilbertlee-amd committed
1487
        exeList.push_back(std::make_pair(exeType, exeIndex));
Gilbert Lee's avatar
Gilbert Lee committed
1488
    }
gilbertlee-amd's avatar
gilbertlee-amd committed
1489
    else if (IsCpuType(exeType))
Gilbert Lee's avatar
Gilbert Lee committed
1490
    {
1491
1492
1493
1494
      for (int exeIndex = 0; exeIndex < ev.numCpuDevices; ++exeIndex)
      {
        // Skip NUMA nodes that have no CPUs (e.g. CXL)
        if (ev.numCpusPerNuma[exeIndex] == 0) continue;
gilbertlee-amd's avatar
gilbertlee-amd committed
1495
        exeList.push_back(std::make_pair(exeType, exeIndex));
1496
      }
Gilbert Lee's avatar
Gilbert Lee committed
1497
1498
    }
  }
1499
  int numExes = exeList.size();
Gilbert Lee's avatar
Gilbert Lee committed
1500
1501
1502
1503

  std::vector<std::pair<MemType, int>> srcList;
  for (auto src : ev.sweepSrc)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
1504
1505
    MemType const srcType = CharToMemType(src);
    int const numDevices = IsGpuType(srcType) ? ev.numGpuDevices : ev.numCpuDevices;
1506

Gilbert Lee's avatar
Gilbert Lee committed
1507
    for (int srcIndex = 0; srcIndex < numDevices; ++srcIndex)
gilbertlee-amd's avatar
gilbertlee-amd committed
1508
      srcList.push_back(std::make_pair(srcType, srcIndex));
Gilbert Lee's avatar
Gilbert Lee committed
1509
1510
1511
1512
1513
1514
1515
  }
  int numSrcs = srcList.size();


  std::vector<std::pair<MemType, int>> dstList;
  for (auto dst : ev.sweepDst)
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
1516
1517
    MemType const dstType = CharToMemType(dst);
    int const numDevices = IsGpuType(dstType) ? ev.numGpuDevices : ev.numCpuDevices;
Gilbert Lee's avatar
Gilbert Lee committed
1518
1519

    for (int dstIndex = 0; dstIndex < numDevices; ++dstIndex)
gilbertlee-amd's avatar
gilbertlee-amd committed
1520
      dstList.push_back(std::make_pair(dstType, dstIndex));
Gilbert Lee's avatar
Gilbert Lee committed
1521
1522
1523
  }
  int numDsts = dstList.size();

1524
1525
  // Build array of possibilities, respecting any additional restrictions (e.g. XGMI hop count)
  struct TransferInfo
Gilbert Lee's avatar
Gilbert Lee committed
1526
  {
gilbertlee-amd's avatar
gilbertlee-amd committed
1527
1528
1529
    MemType srcType; int srcIndex;
    ExeType exeType; int exeIndex;
    MemType dstType; int dstIndex;
1530
1531
1532
1533
1534
1535
1536
1537
  };

  // If either XGMI minimum is non-zero, or XGMI maximum is specified and non-zero then both links must be XGMI
  bool const useXgmiOnly = (ev.sweepXgmiMin > 0 || ev.sweepXgmiMax > 0);

  std::vector<TransferInfo> possibleTransfers;
  TransferInfo tinfo;
  for (int i = 0; i < numExes; ++i)
Gilbert Lee's avatar
Gilbert Lee committed
1538
  {
1539
1540
    // Skip CPU executors if XGMI link must be used
    if (useXgmiOnly && !IsGpuType(exeList[i].first)) continue;
gilbertlee-amd's avatar
gilbertlee-amd committed
1541
1542
    tinfo.exeType  = exeList[i].first;
    tinfo.exeIndex = exeList[i].second;
1543

gilbertlee-amd's avatar
gilbertlee-amd committed
1544
    bool isXgmiSrc  = false;
1545
1546
1547
1548
1549
1550
1551
1552
    int  numHopsSrc = 0;
    for (int j = 0; j < numSrcs; ++j)
    {
      if (IsGpuType(exeList[i].first) && IsGpuType(srcList[j].first))
      {
        if (exeList[i].second != srcList[j].second)
        {
          uint32_t exeToSrcLinkType, exeToSrcHopCount;
gilbertlee-amd's avatar
gilbertlee-amd committed
1553
1554
          HIP_CALL(hipExtGetLinkTypeAndHopCount(RemappedIndex(exeList[i].second, false),
                                                RemappedIndex(srcList[j].second, false),
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
                                                &exeToSrcLinkType,
                                                &exeToSrcHopCount));
          isXgmiSrc = (exeToSrcLinkType == HSA_AMD_LINK_INFO_TYPE_XGMI);
          if (isXgmiSrc) numHopsSrc = exeToSrcHopCount;
        }
        else
        {
          isXgmiSrc = true;
          numHopsSrc = 0;
        }

        // Skip this SRC if it is not XGMI but only XGMI links may be used
        if (useXgmiOnly && !isXgmiSrc) continue;

        // Skip this SRC if XGMI distance is already past limit
        if (ev.sweepXgmiMax >= 0 && isXgmiSrc && numHopsSrc > ev.sweepXgmiMax) continue;
      }
      else if (useXgmiOnly) continue;

gilbertlee-amd's avatar
gilbertlee-amd committed
1574
1575
      tinfo.srcType  = srcList[j].first;
      tinfo.srcIndex = srcList[j].second;
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585

      bool isXgmiDst = false;
      int  numHopsDst = 0;
      for (int k = 0; k < numDsts; ++k)
      {
        if (IsGpuType(exeList[i].first) && IsGpuType(dstList[k].first))
        {
          if (exeList[i].second != dstList[k].second)
          {
            uint32_t exeToDstLinkType, exeToDstHopCount;
gilbertlee-amd's avatar
gilbertlee-amd committed
1586
1587
            HIP_CALL(hipExtGetLinkTypeAndHopCount(RemappedIndex(exeList[i].second, false),
                                                  RemappedIndex(dstList[k].second, false),
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
                                                  &exeToDstLinkType,
                                                  &exeToDstHopCount));
            isXgmiDst = (exeToDstLinkType == HSA_AMD_LINK_INFO_TYPE_XGMI);
            if (isXgmiDst) numHopsDst = exeToDstHopCount;
          }
          else
          {
            isXgmiDst = true;
            numHopsDst = 0;
          }
        }

        // Skip this DST if it is not XGMI but only XGMI links may be used
        if (useXgmiOnly && !isXgmiDst) continue;

        // Skip this DST if total XGMI distance (SRC + DST) is less than min limit
        if (ev.sweepXgmiMin > 0 && (numHopsSrc + numHopsDst < ev.sweepXgmiMin)) continue;

        // Skip this DST if total XGMI distance (SRC + DST) is greater than max limit
        if (ev.sweepXgmiMax >= 0 && (numHopsSrc + numHopsDst) > ev.sweepXgmiMax) continue;

gilbertlee-amd's avatar
gilbertlee-amd committed
1609
1610
        tinfo.dstType  = dstList[k].first;
        tinfo.dstIndex = dstList[k].second;
1611
1612
1613
1614

        possibleTransfers.push_back(tinfo);
      }
    }
Gilbert Lee's avatar
Gilbert Lee committed
1615
1616
  }

1617
1618
1619
  int const numPossible = (int)possibleTransfers.size();
  int maxParallelTransfers = (ev.sweepMax == 0 ? numPossible : ev.sweepMax);

Gilbert Lee's avatar
Gilbert Lee committed
1620
1621
1622
1623
1624
1625
  if (ev.sweepMin > numPossible)
  {
    printf("No valid test configurations exist\n");
    return;
  }

1626
1627
1628
1629
1630
1631
  if (ev.outputToCsv)
  {
    printf("\nTest#,Transfer#,NumBytes,Src,Exe,Dst,CUs,BW(GB/s),Time(ms),"
           "ExeToSrcLinkType,ExeToDstLinkType,SrcAddr,DstAddr\n");
  }

Gilbert Lee's avatar
Gilbert Lee committed
1632
1633
  int numTestsRun = 0;
  int M = ev.sweepMin;
gilbertlee-amd's avatar
gilbertlee-amd committed
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
  std::uniform_int_distribution<int> randSize(1, numBytesPerTransfer / sizeof(float));
  std::uniform_int_distribution<int> distribution(ev.sweepMin, maxParallelTransfers);

  // Log sweep to configuration file
  FILE *fp = fopen("lastSweep.cfg", "w");
  if (!fp)
  {
    printf("[ERROR] Unable to open lastSweep.cfg.  Check permissions\n");
    exit(1);
  }

Gilbert Lee's avatar
Gilbert Lee committed
1645
1646
1647
1648
1649
1650
1651
1652
1653
  // Create bitmask of numPossible triplets, of which M will be chosen
  std::string bitmask(M, 1);  bitmask.resize(numPossible, 0);
  auto cpuStart = std::chrono::high_resolution_clock::now();
  while (1)
  {
    if (isRandom)
    {
      // Pick random number of simultaneous transfers to execute
      // NOTE: This currently skews distribution due to some #s having more possibilities than others
gilbertlee-amd's avatar
gilbertlee-amd committed
1654
      M = distribution(*ev.generator);
Gilbert Lee's avatar
Gilbert Lee committed
1655
1656
1657
1658

      // Generate a random bitmask
      for (int i = 0; i < numPossible; i++)
        bitmask[i] = (i < M) ? 1 : 0;
1659
      std::shuffle(bitmask.begin(), bitmask.end(), *ev.generator);
Gilbert Lee's avatar
Gilbert Lee committed
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
    }

    // Convert bitmask to list of Transfers
    std::vector<Transfer> transfers;
    for (int value = 0; value < numPossible; ++value)
    {
      if (bitmask[value])
      {
        // Convert integer value to (SRC->EXE->DST) triplet
        Transfer transfer;
gilbertlee-amd's avatar
gilbertlee-amd committed
1670
1671
1672
1673
1674
        transfer.numSrcs        = 1;
        transfer.numDsts        = 1;
        transfer.srcType        = {possibleTransfers[value].srcType};
        transfer.srcIndex       = {possibleTransfers[value].srcIndex};
        transfer.exeType        = possibleTransfers[value].exeType;
1675
        transfer.exeIndex       = possibleTransfers[value].exeIndex;
gilbertlee-amd's avatar
gilbertlee-amd committed
1676
1677
1678
        transfer.dstType        = {possibleTransfers[value].dstType};
        transfer.dstIndex       = {possibleTransfers[value].dstIndex};
        transfer.numSubExecs    = IsGpuType(transfer.exeType) ? numGpuSubExecs : numCpuSubExecs;
1679
        transfer.transferIndex  = transfers.size();
gilbertlee-amd's avatar
gilbertlee-amd committed
1680
        transfer.numBytes       = ev.sweepRandBytes ? randSize(*ev.generator) * sizeof(float) : 0;
Gilbert Lee's avatar
Gilbert Lee committed
1681
1682
1683
1684
        transfers.push_back(transfer);
      }
    }

gilbertlee-amd's avatar
gilbertlee-amd committed
1685
1686
    LogTransfers(fp, ++numTestsRun, transfers);
    ExecuteTransfers(ev, numTestsRun, numBytesPerTransfer / sizeof(float), transfers);
Gilbert Lee's avatar
Gilbert Lee committed
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717

    // Check for test limit
    if (numTestsRun == ev.sweepTestLimit)
    {
      printf("Test limit reached\n");
      break;
    }

    // Check for time limit
    auto cpuDelta = std::chrono::high_resolution_clock::now() - cpuStart;
    double totalCpuTime = std::chrono::duration_cast<std::chrono::duration<double>>(cpuDelta).count();
    if (ev.sweepTimeLimit && totalCpuTime > ev.sweepTimeLimit)
    {
      printf("Time limit exceeded\n");
      break;
    }

    // Increment bitmask if not random sweep
    if (!isRandom && !std::prev_permutation(bitmask.begin(), bitmask.end()))
    {
      M++;
      // Check for completion
      if (M > maxParallelTransfers)
      {
        printf("Sweep complete\n");
        break;
      }
      for (int i = 0; i < numPossible; i++)
        bitmask[i] = (i < M) ? 1 : 0;
    }
  }
gilbertlee-amd's avatar
gilbertlee-amd committed
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
  fclose(fp);
}

void LogTransfers(FILE *fp, int const testNum, std::vector<Transfer> const& transfers)
{
  fprintf(fp, "# Test %d\n", testNum);
  fprintf(fp, "%d", -1 * (int)transfers.size());
  for (auto const& transfer : transfers)
  {
    fprintf(fp, " (%c%d->%c%d->%c%d %d %lu)",
gilbertlee-amd's avatar
gilbertlee-amd committed
1728
1729
1730
1731
            MemTypeStr[transfer.srcType[0]], transfer.srcIndex[0],
            ExeTypeStr[transfer.exeType],    transfer.exeIndex,
            MemTypeStr[transfer.dstType[0]], transfer.dstIndex[0],
            transfer.numSubExecs,
gilbertlee-amd's avatar
gilbertlee-amd committed
1732
1733
1734
1735
            transfer.numBytes);
  }
  fprintf(fp, "\n");
  fflush(fp);
Gilbert Lee's avatar
Gilbert Lee committed
1736
}
gilbertlee-amd's avatar
gilbertlee-amd committed
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747

std::string PtrVectorToStr(std::vector<float*> const& strVector, int const initOffset)
{
  std::stringstream ss;
  for (int i = 0; i < strVector.size(); ++i)
  {
    if (i) ss << " ";
    ss << (strVector[i] + initOffset);
  }
  return ss.str();
}