shader: keep pixel helper lanes out of the EXEC and VCC words

This commit is contained in:
nmzik
2026-10-26 06:53:55 +01:00
parent 025d1f7edf
commit 5429888d4c
20 changed files with 428 additions and 390 deletions
@@ -819,6 +819,13 @@ uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate) {
return ctx.Ballot(predicate);
}
uint32_t EmitLiveMask(EmitterState& state) {
const auto helper =
EmitNative<spv::OpLoad, IR::Type::U1>(state, state.helper_invocation_variable);
return EmitNative<spv::OpGroupNonUniformBallot, IR::Type::U32x4>(
state, ConstantU32(state, spv::ScopeSubgroup), EmitLogicalNotBool(state, helper));
}
uint32_t EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst) {
const auto ballot = ctx.Ballot(inst.Arg(1));
const auto lane = ctx.FirstLane(ballot);
@@ -230,6 +230,7 @@ uint32_t EmitDppUpdateU32(ValueEmitContext& ctx, const IR::Inst& in
uint32_t EmitWqmU64(EmitterState& state, uint32_t value);
uint32_t EmitLaneId(EmitterState& state);
uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate);
uint32_t EmitLiveMask(EmitterState& state);
uint32_t EmitConditionRef(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitReadFirstLane(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -146,6 +146,7 @@ struct EmitterState {
const IR::Block* current_block = nullptr;
uint32_t pixel_valid_mask_variable = 0;
uint32_t subgroup_local_invocation_id_variable = 0;
uint32_t helper_invocation_variable = 0;
uint32_t per_vertex_variable = 0;
uint32_t point_size_variable = 0;
uint32_t clip_distance_variable = 0;
@@ -1541,16 +1541,16 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
const auto index =
Binary(state, spv::OpShiftRightLogical, TypeU32(state), address, ConstantU32(state, 2));
const auto access = PrepareMemoryResourceAccess(state, mem);
const auto exec = ctx.Arg(inst, 1);
const auto ballot = ctx.Ballot(inst.Arg(1));
const auto low = state.builder.AllocateId();
const auto high = state.builder.AllocateId();
state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), low, ballot, 0);
state.builder.AddFunction(spv::OpCompositeExtract, TypeU32(state), high, ballot, 1);
// Elect and count from the guest EXEC words, which V_MBCNT also ranks lanes by.
const auto low = ctx.Arg(inst, 1);
const auto high = ctx.Arg(inst, 2);
const auto count = Binary(state, spv::OpIAdd, TypeU32(state),
Unary(state, spv::OpBitCount, TypeU32(state), low),
Unary(state, spv::OpBitCount, TypeU32(state), high));
const auto first = ctx.FirstLane(ballot);
const auto mask = state.builder.AllocateId();
state.builder.AddFunction(spv::OpCompositeConstruct, TypeU32Vector(state, 4), mask, low, high,
ConstantU32(state, 0), ConstantU32(state, 0));
const auto first = ctx.FirstLane(mask);
const auto source_lane =
state.lane_count == 2
? Binary(state, spv::OpBitwiseAnd, TypeU32(state), first, ConstantU32(state, 31))
@@ -1564,9 +1564,7 @@ uint32_t EmitAppendConsume(ValueEmitContext& ctx, const IR::Inst& inst) {
const auto condition = AndCondition(
state, is_first,
AndCondition(state,
state.lane_count == 2 ? Binary(state, spv::OpINotEqual, TypeBool(state), count,
ConstantU32(state, 0))
: exec,
Binary(state, spv::OpINotEqual, TypeBool(state), count, ConstantU32(state, 0)),
bounds));
const auto atomic = EmitValueOrZeroIfCondition(state, condition, [&]() {
const auto value = state.builder.AllocateId();
@@ -595,6 +595,13 @@ void DefineInputs(EmitterState& state) {
state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationFlat);
}
}
if (state.program.info.helper_invocation) {
const auto variable = DefineInterfaceVariable(state, TypeBool(state), spv::StorageClassInput,
"gl_HelperInvocation");
state.helper_invocation_variable = variable;
state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationBuiltIn,
spv::BuiltInHelperInvocation);
}
}
void DefineOutputs(EmitterState& state) {