graph: backend: remove fold_sum_scales pass

Author: xinyu-intel

src/graph/backend/dnnl/passes/transform.cpp

@@ -461,88 +461,6 @@ status_t fold_mul_scales(std::shared_ptr<subgraph_t> &sg) {
     return status::success;
 }
 
-status_t fold_sum_scales(std::shared_ptr<subgraph_t> &sg) {
-    std::set<op_t *> visited;
-    subgraph_rewriter_t rewriter(sg);
-
-    for (auto &cur_op : sg->get_ops()) {
-        if (!(cur_op->get_kind() == op_kind::dnnl_binary
-                    && static_cast<dnnl::algorithm>(
-                               cur_op->get_attr<int64_t>(op_attr::alg_kind))
-                            == dnnl::algorithm::binary_add)
-                || visited.count(cur_op.get()))
-            continue;
-
-        visited.insert(cur_op.get());
-        size_t mul_scale_op_offset = 2;
-        auto lhs_val = cur_op->get_input_value(0);
-        auto rhs_val = cur_op->get_input_value(1);
-
-        if (!lhs_val->has_producer() || !rhs_val->has_producer()) { continue; }
-        const auto &l_op = lhs_val->get_producer();
-        const auto &r_op = rhs_val->get_producer();
-
-        auto consumers = cur_op->get_output_values()[0]->get_consumers();
-        if (consumers.empty()
-                || consumers[0].get_op().get_kind()
-                        != op_kind::dnnl_mul_scales) {
-            continue;
-        }
-
-        if (l_op.get_kind() != op_kind::dnnl_mul_scales
-                || r_op.get_kind() != op_kind::dnnl_mul_scales) {
-            continue;
-        }
-        if (l_op.num_inputs() > 0 && l_op.get_input_value(0)->has_producer()
-                && l_op.get_input_value(0)->get_producer().get_kind()
-                        == op_kind::dnnl_reorder) {
-            mul_scale_op_offset = 1;
-        } else if (r_op.num_inputs() > 0
-                && r_op.get_input_value(0)->has_producer()
-                && r_op.get_input_value(0)->get_producer().get_kind()
-                        == op_kind::dnnl_reorder) {
-            mul_scale_op_offset = 0;
-        }
-
-        if (mul_scale_op_offset != 2
-                && ltw(lhs_val->get_logical_tensor()).vdims()
-                        == ltw(rhs_val->get_logical_tensor()).vdims()) {
-            auto in_val = cur_op->get_input_value(mul_scale_op_offset);
-            auto &mul_scale_op = in_val->get_producer();
-            auto scales = mul_scale_op.get_attr<std::vector<float>>(
-                    op_attr::scales);
-            assert(scales.size() == 1); // per tensor
-
-            auto tmp = mul_scale_op.get_input_value(0);
-            auto &add_zps_op = tmp->get_producer();
-            auto zps = add_zps_op.get_attr<std::vector<int64_t>>(op_attr::zps);
-            assert(scales.size() == zps.size());
-
-            auto out_val = cur_op->get_output_values()[0];
-            auto consumers = out_val->get_consumers();
-            auto &next_op = consumers[0].get_op();
-            // set sum post-ops' second input scale
-            float tmp_scale
-                    = next_op.get_attr<std::vector<float>>(op_attr::scales)[0];
-            scales[0] *= tmp_scale;
-            mul_scale_op.set_attr<std::vector<float>>(op_attr::scales, scales);
-
-            // update the output scales
-            auto other_val = cur_op->get_input_value(1 - mul_scale_op_offset);
-            auto &oscales_op = other_val->get_producer();
-            auto oscales
-                    = oscales_op.get_attr<std::vector<float>>(op_attr::scales);
-            for (auto &v : oscales)
-                v *= tmp_scale;
-            oscales_op.set_attr<std::vector<float>>(op_attr::scales, oscales);
-            rewriter.fuse_op_to_predecessor(next_op.shared_from_this());
-        }
-    }
-
-    rewriter.run();
-    return status::success;
-}
-
 // FIXME(xx) This pass works correctly only when all inputs/outputs scales/zps
 // are same, since we are simply ignoring the scales and zps. We can improve
 // this pass to support different per-tensor scale since oneDNN concat primitive