shader: lower complete wave minima for sparse fragment subgroups

This commit is contained in:
nmzik
2026-10-14 18:03:48 +02:00
parent 892e70c8a1
commit 302240da70
12 changed files with 330 additions and 8 deletions
@@ -748,6 +748,18 @@ uint32_t EmitReadLane(ValueEmitContext& ctx, const IR::Inst& inst) {
return ctx.Shuffle(inst, 0, ctx.Arg(inst, 1));
}
uint32_t EmitSubgroupUMinU32(ValueEmitContext& ctx, const IR::Inst& inst) {
auto& state = ctx.state;
auto value = ctx.Arg(inst, 0);
if (ctx.other_half != nullptr) {
value = EmitUMin32(state, ctx.HalfArg(inst, 0, 0), ctx.HalfArg(inst, 0, 1));
}
const auto result = state.builder.AllocateId();
state.builder.AddFunction(spv::OpGroupNonUniformUMin, TypeU32(state), result,
ConstantU32(state, spv::ScopeSubgroup), spv::GroupOperationReduce, value);
return result;
}
uint32_t EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst) {
auto& state = ctx.state;
const auto hit = state.builder.AllocateId();
@@ -238,6 +238,7 @@ uint32_t EmitBallot(ValueEmitContext& ctx, IR::Value predicate);
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);
uint32_t EmitSubgroupUMinU32(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitWriteLane(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitPermlane16U32(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitGetAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -766,7 +766,7 @@ void DefineModule(EmitterState& state) {
}
if (state.lane_count == 2 || state.program.info.subgroup_barrier ||
state.program.info.subgroup_ballot || state.program.info.subgroup_shuffle ||
state.program.info.subgroup_local_invocation_id) {
state.program.info.subgroup_arithmetic || state.program.info.subgroup_local_invocation_id) {
state.builder.RequireCapability(spv::CapabilityGroupNonUniform);
}
if (state.lane_count == 2 || state.program.info.subgroup_ballot) {
@@ -775,6 +775,9 @@ void DefineModule(EmitterState& state) {
if (state.program.info.subgroup_shuffle) {
state.builder.RequireCapability(spv::CapabilityGroupNonUniformShuffle);
}
if (state.program.info.subgroup_arithmetic) {
state.builder.RequireCapability(spv::CapabilityGroupNonUniformArithmetic);
}
if (state.program.info.compute_derivatives && state.program.stage == ShaderType::Compute) {
state.builder.RequireCapability(spv::CapabilityComputeDerivativeGroupQuadsKHR);
state.builder.RequireExtension("SPV_KHR_compute_shader_derivatives");
@@ -226,8 +226,9 @@ void Invoke(Return (*emit)(Context&, Args...), ValueEmitContext& ctx, const IR::
void EmitDirectInstruction(ValueEmitContext& ctx, const IR::Inst& inst) {
if (ctx.half != 0 && (inst.GetOpcode() == IR::ValueOpcode::Ballot ||
inst.GetOpcode() == IR::ValueOpcode::ReadFirstLane)) {
// Both operations already combine both emulated halves into one whole-wave result.
inst.GetOpcode() == IR::ValueOpcode::ReadFirstLane ||
inst.GetOpcode() == IR::ValueOpcode::SubgroupUMinU32)) {
// These operations combine both emulated halves into one whole-wave result.
ctx.Define(inst, ctx.other_half->Result(inst));
return;
}