[ON HOLD] Unit Tests for CPX mode #1319
Conversation
saurabhAMD
requested review from
wenkaidu,
gilbertlee-amd,
akolliasAMD,
edgargabriel,
PedramAlizadeh,
nusislam,
nileshnegi,
KawtharShafie,
AtlantaPepsi,
mberenjk,
corey-derochie-amd and
haripriya-amd
as code owners
| int getDeviceMode (bool *cpxMode){ | ||
| // Prepare parent->child pipe | ||
| int pipefd[2]; | ||
| if (pipe(pipefd) == -1) | ||
| { | ||
| ERROR("Unable to create parent->child pipe for getting the device mode\n"); | ||
| return TEST_FAIL; | ||
| } | ||
| pid_t pid = fork(); | ||
| if (0 == pid) | ||
| { | ||
| bool iscpxMode = false; | ||
| try { | ||
| std::string log = execCommand("rocm-smi --showcomputepartition"); | ||
| bool foundCPX = log.find("CPX") != std::string::npos; | ||
| if (foundCPX) { | ||
| iscpxMode = true; | ||
| } | ||
| } catch (const std::exception& e) { | ||
| std::cerr << "Error: " << e.what() << std::endl; | ||
| return 1; | ||
| } | ||
| if (write(pipefd[1], &iscpxMode, sizeof(iscpxMode)) != sizeof(iscpxMode)) return TEST_FAIL; | ||
| close(pipefd[0]); | ||
| close(pipefd[1]); | ||
| exit(EXIT_SUCCESS); | ||
| } | ||
| else { | ||
| int status; | ||
| if (read(pipefd[0], cpxMode, sizeof(*cpxMode)) != sizeof(*cpxMode)) return TEST_FAIL; | ||
| waitpid(pid, &status, 0); | ||
| assert(!status); | ||
| close(pipefd[0]); | ||
| close(pipefd[1]); | ||
| } | ||
| return TEST_SUCCESS; | ||
| return 0; | ||
| } |
There was a problem hiding this comment.
I'm curious why rocm-smi is run in a fork process?
There was a problem hiding this comment.
Because we cannot use HIP call prior to launching unless it is inside another child process. You can refer to line 207 for this comment. The same procedure is used for other features such as finding the architecture of the system as in line 211.
There was a problem hiding this comment.
In your case, I think popen already forks and creates a pipe internally, so you may not need your own fork. If so, getDeviceMode and getDevicePriority can be simplified.
| } | ||
| if(isGfx94) { | ||
| bool cpxMode = false; | ||
| getDeviceMode(&cpxMode); |
There was a problem hiding this comment.
The return value of getDeviceMode is never used, so if it can be simplified as I've suggested, you could rename it to isDeviceCpxMode and return bool.
corey-derochie-amd
changed the title
Details
Work item: "Internal"
What were the changes?
Enabling Unit Tests to run for CPX mode.
Why were the changes made?
Currently UT tests all possible numbers of GPUs up to 16. If same scheme is used to test CPX mode, UT will take forever to complete.
How was the outcome achieved?
Additional Documentation:
None.
Approval Checklist
Do not approve until these items are satisfied.