shader: preserve wave-wide vector-store completion before publication

This commit is contained in:
nmzik
2026-10-04 00:17:28 +04:00
parent 42a315f0e4
commit f58014ff0e
16 changed files with 144 additions and 32 deletions
@@ -263,6 +263,7 @@ Emitter::SpirvRequirements Emitter::AnalyzeProgramRequirements(const IR::Program
}
}
switch (inst.GetOpcode()) {
case IR::ValueOpcode::StoreCompletion: requirements.subgroup_barrier = true; break;
case IR::ValueOpcode::BvhIntersect: requirements.bvh = true; break;
case IR::ValueOpcode::Ballot: requirements.subgroup_ballot = true; break;
case IR::ValueOpcode::DppMoveU32:
@@ -574,6 +574,14 @@ uint32_t EmitIdentity(ValueEmitContext&, uint32_t value) {
void EmitVoid(ValueEmitContext&) {}
void EmitStoreCompletion(EmitterState& state) {
state.builder.AddFunction(
spv::OpControlBarrier, ConstantU32(state, spv::ScopeSubgroup),
ConstantU32(state, spv::ScopeDevice),
ConstantU32(state, spv::MemorySemanticsReleaseMask | spv::MemorySemanticsUniformMemoryMask |
spv::MemorySemanticsImageMemoryMask));
}
void EmitBarrier(EmitterState& state) {
const auto tessellation = state.program.stage == ShaderType::TessellationControl;
if (!tessellation && ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
@@ -188,11 +188,11 @@ void EmitVoid(ValueEmitContext&);
inline constexpr auto EmitReference = EmitVoid;
inline constexpr auto EmitReferenceU32 = EmitVoid;
inline constexpr auto EmitControlNop = EmitVoid;
inline constexpr auto EmitWaitcnt = EmitVoid;
inline constexpr auto EmitSendmsg = EmitVoid;
inline constexpr auto EmitTtraceData = EmitVoid;
inline constexpr auto EmitInstPrefetch = EmitVoid;
void EmitBarrier(EmitterState& state);
void EmitStoreCompletion(EmitterState& state);
void EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
uint32_t EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -61,6 +61,7 @@ const ImageDimensionInfo& ImageDimensionInfoFor(ImageDimension dimension);
struct SpirvRequirements {
bool bvh = false;
bool subgroup_ballot = false;
bool subgroup_barrier = false;
bool subgroup_shuffle = false;
bool subgroup_local_invocation_id = false;
bool compute_derivatives = false;
@@ -701,8 +701,9 @@ void DefineModule(EmitterState& state) {
if (state.requirements.image_gather_extended) {
state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
}
if (state.lane_count == 2 || state.requirements.subgroup_ballot ||
state.requirements.subgroup_shuffle || state.requirements.subgroup_local_invocation_id) {
if (state.lane_count == 2 || state.requirements.subgroup_barrier ||
state.requirements.subgroup_ballot || state.requirements.subgroup_shuffle ||
state.requirements.subgroup_local_invocation_id) {
state.builder.RequireCapability(spv::CapabilityGroupNonUniform);
}
if (state.lane_count == 2 || state.requirements.subgroup_ballot) {
@@ -315,6 +315,7 @@ void EmitBlock(ValueEmitContext& ctx, const IR::Block* block, EmitInstruction&&
auto& lane = half == 0 ? ctx : *ctx.other_half;
ctx.state.lane_half = half;
if (half == 0 || (inst.GetOpcode() != IR::ValueOpcode::Barrier &&
inst.GetOpcode() != IR::ValueOpcode::StoreCompletion &&
inst.GetOpcode() != IR::ValueOpcode::MeshAllocate)) {
emit_instruction(lane, inst);
}