diff options
author | Sviatoslav Peleshko <sviatoslav.peleshko@globallogic.com> | 2023-10-13 05:02:22 +0300 |
---|---|---|
committer | Marge Bot <emma+marge@anholt.net> | 2023-12-19 13:32:23 +0000 |
commit | 8f8cde4c6050d1e91101ec66e8982036da9d7700 (patch) | |
tree | 418cb14ef0b8b2ba2c27ffa3db02769336f3b47f | |
parent | a6459e0f7bcb662130c79764ccf54ccfa9151182 (diff) |
intel/fs: Don't optimize DW*1 MUL if it stores value to the accumulator
Fixes: a8b86459 ("i965/fs: Optimize a * 1.0 -> a.")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/9570
Signed-off-by: Sviatoslav Peleshko <sviatoslav.peleshko@globallogic.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25710>
-rw-r--r-- | src/intel/compiler/brw_fs.cpp | 23 |
1 files changed, 23 insertions, 0 deletions
diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp index 9dd660a1a81..ce43e53f2d3 100644 --- a/src/intel/compiler/brw_fs.cpp +++ b/src/intel/compiler/brw_fs.cpp @@ -2444,6 +2444,29 @@ fs_visitor::opt_algebraic() if (brw_reg_type_is_floating_point(inst->src[1].type)) break; + /* From the BDW PRM, Vol 2a, "mul - Multiply": + * + * "When multiplying integer datatypes, if src0 is DW and src1 + * is W, irrespective of the destination datatype, the + * accumulator maintains full 48-bit precision." + * ... + * "When multiplying integer data types, if one of the sources + * is a DW, the resulting full precision data is stored in + * the accumulator." + * + * There are also similar notes in earlier PRMs. + * + * The MOV instruction can copy the bits of the source, but it + * does not clear the higher bits of the accumulator. So, because + * we might use the full accumulator in the MUL/MACH macro, we + * shouldn't replace such MULs with MOVs. + */ + if ((brw_reg_type_to_size(inst->src[0].type) == 4 || + brw_reg_type_to_size(inst->src[1].type) == 4) && + (inst->dst.is_accumulator() || + inst->writes_accumulator_implicitly(devinfo))) + break; + /* a * 1.0 = a */ if (inst->src[1].is_one()) { inst->opcode = BRW_OPCODE_MOV; |