NicolaLancellotti commented on code in PR #10959:
URL: https://github.com/apache/tvm/pull/10959#discussion_r863947886
##########
src/tir/contrib/ethosu/passes.cc:
##########
@@ -110,6 +110,98 @@ tvm::transform::Pass HoistAllocates() {
TVM_REGISTER_GLOBAL("tir.contrib.ethos-u.HoistAllocates").set_body_typed(HoistAllocates);
+/*!
+ * \brief Reorders copy and compute nodes in such a way that independent DMA
copies,
+ * and computes happen in parallel.
+ * Copies to buffers with local scope are not reordered, indeed they copy LUT
+ * into the SHRAM which already happens in parallel with copying weights into
+ * the weights encoder.
+ */
+class CopyComputeReorderingMutator : public StmtExprMutator {
+ public:
+ CopyComputeReorderingMutator(int max_copy_movements) :
_max_copy_movements{max_copy_movements} {}
+
+ PrimFunc operator()(PrimFunc main_func) {
+ if (_max_copy_movements > 0) {
+ auto n{main_func.CopyOnWrite()};
+ n->body = this->VisitStmt(main_func->body);
+ return GetRef<PrimFunc>(n);
+ }
+ return main_func;
+ }
+
+ private:
+ Stmt VisitStmt_(const SeqStmtNode* op) override {
+ if (op->size() <= 1) {
+ return StmtExprMutator::VisitStmt_(op);
+ }
+
+ auto seq_stmt{GetRef<SeqStmt>(op)};
+ std::vector<Stmt> new_seq(seq_stmt->size());
+ std::copy(seq_stmt->seq.begin(), seq_stmt->seq.end(), new_seq.begin());
+
+ for (size_t index{}; index < new_seq.size(); ++index) {
+ for (int offset{}; offset < _max_copy_movements; ++offset) {
Review Comment:
Done.
--
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]
