Making InterIpuCopy non-allocating

Summary:
Making InterIpuCopy allocating can cause global exchange lowering to take significantly longer as we can get global exchanged with different source and destination layouts.
For the related regression ticket, the difference observed locally was 108 seconds vs 12 seconds.

Reviewers: #tensorflow, #framework_ip_review_-_any_oss_or_third-party_code_use_has_been_approved, samuelh

Reviewed By: #tensorflow, #framework_ip_review_-_any_oss_or_third-party_code_use_has_been_approved, samuelh

Maniphest Tasks: T65905

Differential Revision: https://phabricator.sourcevertex.net/D74105

Author: Alfie-Edwards

tensorflow/compiler/plugin/poplar/driver/passes/allocation_analysis.cc

@@ -195,8 +195,8 @@ Status FindStartPoints(AllocationLoopState& state, HloModule* module) {
 }
 
 // Only some inputs affect tile mapping of output, for this function
-// only add the user if next.instruction's tile mapping is related to
-// it's output
+// only add the user if the next instruction's tile mapping is related to
+// its output.
 Status AddIfMappingDependsOnOperand(std::vector<IndexedLocation>& to_visit,
                                     HloInstruction* user,
                                     const IndexedLocation& operand,

tensorflow/compiler/plugin/poplar/driver/tools/custom_ops/inter_ipu_copy.cc

@@ -29,7 +29,7 @@ absl::flat_hash_set<int64_t> HloInterIpuCopy::AllocatingIndices() const {
   return {};
 }
 
-bool HloInterIpuCopy::AllocatingOutput() const { return true; }
+bool HloInterIpuCopy::AllocatingOutput() const { return false; }
 
 absl::flat_hash_map<int64_t, int64_t> HloInterIpuCopy::LayoutDependencies()
     const {