Lunderberg commented on code in PR #11173:
URL: https://github.com/apache/tvm/pull/11173#discussion_r865093976
##########
src/relay/backend/te_compiler.cc:
##########
@@ -952,25 +919,24 @@ backend::FunctionInfo UpdateMainWorkspaceSize(const
IRModule& mod, tec::TargetMa
Map<Target, Function> relay_primfuncs;
// Initialize all target workspaces to zero
- for (const auto& kv : targets) {
- auto tgt = kv.second;
- workspace_sizes.Set(tgt, 0);
+ for (const auto& target : config->primitive_targets) {
+ workspace_sizes.Set(target, 0);
}
for (const auto& dev_and_size : device_workspace) {
- auto tgt = tec::GetTargetFromInteger(dev_and_size.first, targets);
- workspace_sizes.Set(tgt, dev_and_size.second);
- relay_primfuncs.Set(tgt, func);
+ Target target = config->FindPrimitiveTargetOrFail(dev_and_size.first);
Review Comment:
Got it! So after the transitionary state, rather than needing to assume
that all targets that use a specific `device_type` will share the same resource
pool, the equivalent deduplication would be implicit in the virtual device
assignment?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
