The U,V,W registers could be allocated to a selection vector directly. Then we can save some MOV instructions for the read_image functions.
0491433 GBE: optimize sample instruction.
backend/src/backend/gen_context.cpp | 20 ++++----------------
backend/src/backend/gen_encoder.cpp | 4 ++--
backend/src/backend/gen_encoder.hpp | 2 +-
backend/src/backend/gen_insn_selection.cpp | 26 ++++++++++++--------------
backend/src/backend/gen_insn_selection.hpp | 2 +-
backend/src/ocl_stdlib.tmpl.h | 12 ++++++------
6 files changed, 26 insertions(+), 40 deletions(-)
Upstream: cgit.freedesktop.org