1/*******************************************************************************
2* Copyright 2022 Intel Corporation
3*
4* Licensed under the Apache License, Version 2.0 (the "License");
5* you may not use this file except in compliance with the License.
6* You may obtain a copy of the License at
7*
8* http://www.apache.org/licenses/LICENSE-2.0
9*
10* Unless required by applicable law or agreed to in writing, software
11* distributed under the License is distributed on an "AS IS" BASIS,
12* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13* See the License for the specific language governing permissions and
14* limitations under the License.
15*******************************************************************************/
16
17#ifndef GPU_JIT_PASS_HOIST_HPP
18#define GPU_JIT_PASS_HOIST_HPP
19
20#include "gpu/jit/ir/ir.hpp"
21
22namespace dnnl {
23namespace impl {
24namespace gpu {
25namespace jit {
26
27// Moves invariant expressions out of loops.
28stmt_t hoist_exprs(const stmt_t &s, ir_context_t &ir_ctx, int reserved_regs);
29
30// Moves boolean mask computation from send calls to the top of the statement
31// group corresponding to `label`. This is done to reduce GRF consumption and
32// to reuse masks between calls. A vector boolean mask is stored as u16 type
33// and converted to bool type right before the call. Transformation is limited
34// to the statement group corresponding to `label`.
35// If `split_by_and` is true then any ((A & B) & C) mask is split into A, B, C
36// sub-masks which are initialized independently. This allows reusing those
37// sub-masks for other masks.
38stmt_t hoist_send_masks(const stmt_t &s, ir_context_t &ir_ctx,
39 const stmt_label_t &label, bool split_by_and);
40
41} // namespace jit
42} // namespace gpu
43} // namespace impl
44} // namespace dnnl
45
46#endif
47