[codegen/reassociation] now recursively takes pointer arguments into account as well

This commit is contained in:
Philippe Tillet
2019-07-31 18:41:56 -07:00
parent f7bd976fc7
commit 3b92ddf7e6
5 changed files with 24 additions and 8 deletions

View File

@@ -130,7 +130,7 @@ public:
// create profile
triton::dnn::blocksparse::dot dot(N, params_.K, params_.segments, params_.C, "fp16", params_.bsize, params_.locks, params_.blocks, OP);
// blocksparse matmul
triton::dnn::base* op = dot.enqueue(stream, {&da, &db, &dc, &dlut}, triton::dnn::NO_TUNING);
triton::dnn::base* op = dot.enqueue(stream, {&da, &db, &dc, &dlut}, triton::dnn::FULL_TUNING);
triton::driver::buffer* locks_buffer = ((triton::dnn::blocksparse::dot*)op)->get_locks();
Tensor *tmp = nullptr;
TensorShape tmp_shapes;