vector_alu: Simplify mbcnt pattern
This commit is contained in:
parent
b08db4603e
commit
a5574bdbf0
|
@ -925,25 +925,13 @@ void Translator::V_FFBL_B32(const GcnInst& inst) {
|
|||
void Translator::V_MBCNT_U32_B32(bool is_low, const GcnInst& inst) {
|
||||
const IR::U32 src0{GetSrc(inst.src[0])};
|
||||
const IR::U32 src1{GetSrc(inst.src[1])};
|
||||
const IR::U32 lane_id = ir.LaneId();
|
||||
|
||||
const auto [warp_half, mask_shift] = [&]() -> std::pair<IR::U32, IR::U32> {
|
||||
if (profile.subgroup_size == 32) {
|
||||
const IR::U32 warp_half = ir.BitwiseAnd(ir.WarpId(), ir.Imm32(1));
|
||||
return std::make_pair(warp_half, lane_id);
|
||||
}
|
||||
const IR::U32 warp_half = ir.ShiftRightLogical(lane_id, ir.Imm32(5));
|
||||
const IR::U32 mask_shift = ir.BitwiseAnd(lane_id, ir.Imm32(0x1F));
|
||||
return std::make_pair(warp_half, mask_shift);
|
||||
}();
|
||||
|
||||
const IR::U32 thread_mask = ir.ISub(ir.ShiftLeftLogical(ir.Imm32(1), mask_shift), ir.Imm32(1));
|
||||
const IR::U1 is_odd_warp = ir.INotEqual(warp_half, ir.Imm32(0));
|
||||
const IR::U32 mask = IR::U32{ir.Select(is_odd_warp, is_low ? ir.Imm32(~0U) : thread_mask,
|
||||
is_low ? thread_mask : ir.Imm32(0))};
|
||||
const IR::U32 masked_value = ir.BitwiseAnd(src0, mask);
|
||||
const IR::U32 result = ir.IAdd(src1, ir.BitCount(masked_value));
|
||||
SetDst(inst.dst[0], result);
|
||||
if (!is_low) {
|
||||
ASSERT(src0.IsImmediate() && src0.U32() == ~0U &&
|
||||
src1.IsImmediate() && src1.U32() == 0U);
|
||||
return;
|
||||
}
|
||||
ASSERT(src0.IsImmediate() && src0.U32() == ~0U);
|
||||
SetDst(inst.dst[0], ir.LaneId());
|
||||
}
|
||||
|
||||
} // namespace Shader::Gcn
|
||||
|
|
Loading…
Reference in New Issue