AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Adding linkinfo and srcGPU to destGPU info

Browse Source 
[ROCm/rccl commit: 5a2f74b8d0]
This commit is contained in:
rohit pathania
2019-08-09 12:44:06 +05:30
parent 8645391260
commit 95162665c7
1 changed files with 25 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/tools/rccl-prim-test/rccl_prim_test.cpp
+25 -4
View File
@@ -234,6 +234,21 @@ bool cmdOptionExists(char** begin, char** end, const std::string& option) {
return std::find(begin, end, option) != end; return std::find(begin, end, option) != end;
} }
std::string link_type(uint32_t type) {
switch(type) {
case HSA_AMD_LINK_INFO_TYPE_QPI :
return "QPI";
case HSA_AMD_LINK_INFO_TYPE_INFINBAND:
return "INFINBAND";
case HSA_AMD_LINK_INFO_TYPE_XGMI:
return "XGMI";
case HSA_AMD_LINK_INFO_TYPE_PCIE:
return "PCIE";
default:
return "HYPERTRANSPORT";
}
}
int main(int argc,char* argv[]) int main(int argc,char* argv[])
{ {
if (cmdOptionExists(argv, argv + argc, "-h")) { if (cmdOptionExists(argv, argv + argc, "-h")) {
@@ -365,7 +380,7 @@ int main(int argc,char* argv[])
uint64_t opCount = 0; uint64_t opCount = 0;
for (int op = begin_op; op < end_op; op ++) { for (int op = begin_op; op < end_op; op ++) {
const char *OpsName[] = {"Copy", "Local Copy", "Double Copy", "Reduce", "ReduceCopy"}; const char *OpsName[] = {"Copy", "Local Copy", "Double Copy", "Reduce", "ReduceCopy"};
printf("Testing %s: \n", OpsName[op]); printf("[Testing %s]: \n", OpsName[op]);
// 2 warm up cycles // 2 warm up cycles
for (int i = 0; i < 2; i ++) { for (int i = 0; i < 2; i ++) {
for (int i = 0; i < nGpu; i ++) { for (int i = 0; i < nGpu; i ++) {
@@ -410,17 +425,23 @@ int main(int argc,char* argv[])
auto delta = std::chrono::high_resolution_clock::now() - start; auto delta = std::chrono::high_resolution_clock::now() - start;
double deltaSec = std::chrono::duration_cast<std::chrono::duration<double>>(delta).count(); double deltaSec = std::chrono::duration_cast<std::chrono::duration<double>>(delta).count();
std::cout<<"***GPU to GPU Transfer Profiling Data***"<<std::endl;
for (int i = 0; i < nGpu; i ++) { for (int i = 0; i < nGpu; i ++) {
HIPCHECK(hipMemcpyAsync(profiling_data[i], d_profiling_data[i], HIPCHECK(hipMemcpyAsync(profiling_data[i], d_profiling_data[i],
sizeof(struct profiling_data_t), hipMemcpyDeviceToHost, sizeof(struct profiling_data_t), hipMemcpyDeviceToHost,
stream[i])); stream[i]));
HIPCHECK(hipStreamSynchronize(stream[i])); HIPCHECK(hipStreamSynchronize(stream[i]));
#define RTC_CLOCK_FREQ 2.7E07 #define RTC_CLOCK_FREQ 2.7E07
int next_gpu = findNextGpu(ring_0, i, nGpu);
uint32_t linktype;
uint32_t hopcount;
HIPCHECK(hipExtGetLinkTypeAndHopCount(i, next_gpu , &linktype, &hopcount));
double t0 = (double)profiling_data[i]->write_cycles/((double)RTC_CLOCK_FREQ)/(double)workgroups; double t0 = (double)profiling_data[i]->write_cycles/((double)RTC_CLOCK_FREQ)/(double)workgroups;
fprintf(stderr, "GPU %d: time %.4fs bytes_transferred %lu kernel throughput %.2f GB/s\n", fprintf(stderr, "[GPU %d -> GPU %d][%s]:time %.4fs bytes_transferred %lu kernel throughput %.2f GB/s\n",
i, t0, profiling_data[i]->bytes_transferred, (double)profiling_data[i]->bytes_transferred/(t0*1.0E9)); i, next_gpu,link_type(linktype).c_str(),t0, profiling_data[i]->bytes_transferred, (double)profiling_data[i]->bytes_transferred/(t0*1.0E9));
} }
std::cout<<"***Application Level Transfer Profiling Data***"<<std::endl;
double speed = (double)(profiling_data[0]->bytes_transferred) / (deltaSec*1.0E9); double speed = (double)(profiling_data[0]->bytes_transferred) / (deltaSec*1.0E9);
printf("Transfered %lu bytes in %f s. Throughput %f GB/s\n", profiling_data[0]->bytes_transferred, deltaSec, speed); printf("Transfered %lu bytes in %f s. Throughput %f GB/s\n", profiling_data[0]->bytes_transferred, deltaSec, speed);
} }