raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 19:07:01 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: share SPIR-V atomic update lowering

Browse Source
This commit is contained in:
nmzik
2026-09-04 05:10:31 +02:00
parent d4cca82a22
commit 8596627ced
3 changed files with 79 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h
+45
View File
@@ -130,6 +130,7 @@ enum : uint32_t {
MemorySemanticsAcquireRelease = 0x00000008u,
MemorySemanticsUniformMemory = 0x00000040u,
MemorySemanticsWorkgroupMemory = 0x00000100u,
MemorySemanticsImageMemory = 0x00000800u,
};
enum : uint32_t {
@@ -671,6 +672,9 @@ uint32_t NormalizeFormatComponent(EmitterState& state, const Format::BufferForma
void EmitDeviceAtomicMemoryBarrier(EmitterState& state);
uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
bool max_value);
uint32_t EmitDsSwizzleTargetLane(EmitterState& state, uint32_t subid, uint32_t control);
uint32_t EmitSelectValueU32(EmitterState& state, uint32_t cond, uint32_t true_value,
@@ -786,6 +790,47 @@ uint32_t EmitValueOrZeroIfCondition(EmitterState& state, uint32_t condition, Fn&
std::forward<Fn>(fn));
}
template <typename Fn>
uint32_t AtomicUpdate(EmitterState& state, uint32_t pointer, IR::ResourceKind kind, Fn&& desired) {
const auto scope = kind == IR::ResourceKind::Lds ? ScopeWorkgroup : ScopeDevice;
const auto memory = [&] {
switch (kind) {
case IR::ResourceKind::Lds: return MemorySemanticsWorkgroupMemory;
case IR::ResourceKind::Image: return MemorySemanticsImageMemory;
default: return MemorySemanticsUniformMemory;
}
}();
const auto preheader = state.builder.AllocateId();
const auto header = state.builder.AllocateId();
const auto cont = state.builder.AllocateId();
const auto merge = state.builder.AllocateId();
const auto initial = state.builder.AllocateId();
const auto observed = state.builder.AllocateId();
const auto exchanged = state.builder.AllocateId();
state.builder.AddFunction({OpBranch, preheader});
EmitLabel(state, preheader);
state.builder.AddFunction({OpAtomicLoad, TypeU32(state), initial, pointer,
ConstantU32(state, scope), ConstantU32(state, MemorySemanticsNone)});
state.builder.AddFunction({OpBranch, header});
EmitLabel(state, header);
state.builder.AddFunction(
{OpPhi, TypeU32(state), observed, initial, preheader, exchanged, cont});
const auto next = desired(observed);
state.builder.AddFunction({OpAtomicCompareExchange, TypeU32(state), exchanged, pointer,
ConstantU32(state, scope), ConstantU32(state, MemorySemanticsNone),
ConstantU32(state, MemorySemanticsNone), next, observed});
const auto success = state.builder.AllocateId();
state.builder.AddFunction({OpIEqual, TypeBool(state), success, exchanged, observed});
state.builder.AddFunction({OpLoopMerge, merge, cont, LoopControlNone});
state.builder.AddFunction({OpBranchConditional, success, merge, cont});
EmitLabel(state, cont);
state.builder.AddFunction({OpBranch, header});
EmitLabel(state, merge);
state.builder.AddFunction({OpMemoryBarrier, ConstantU32(state, scope),
ConstantU32(state, MemorySemanticsAcquireRelease | memory)});
return observed;
}
} // namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter
#endif /* EMULATOR_INCLUDE_EMULATOR_GRAPHICS_SHADER_RECOMPILER_SPIRVEMITTER_INTERNAL_H_ */
src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemory.cpp
+1 -69
View File
@@ -481,41 +481,6 @@ uint32_t FormattedLoad(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Me
});
}
template <typename Fn>
uint32_t AtomicUpdate(EmitterState& state, uint32_t pointer, IR::ResourceKind kind, Fn&& desired) {
const auto scope = kind == IR::ResourceKind::Lds ? ScopeWorkgroup : ScopeDevice;
const auto memory = kind == IR::ResourceKind::Lds ? MemorySemanticsWorkgroupMemory
: MemorySemanticsUniformMemory;
const auto preheader = state.builder.AllocateId();
const auto header = state.builder.AllocateId();
const auto cont = state.builder.AllocateId();
const auto merge = state.builder.AllocateId();
const auto initial = state.builder.AllocateId();
const auto observed = state.builder.AllocateId();
const auto exchanged = state.builder.AllocateId();
state.builder.AddFunction({OpBranch, preheader});
EmitLabel(state, preheader);
state.builder.AddFunction({OpAtomicLoad, TypeU32(state), initial, pointer,
ConstantU32(state, scope), ConstantU32(state, MemorySemanticsNone)});
state.builder.AddFunction({OpBranch, header});
EmitLabel(state, header);
state.builder.AddFunction(
{OpPhi, TypeU32(state), observed, initial, preheader, exchanged, cont});
const auto next = desired(observed);
state.builder.AddFunction({OpAtomicCompareExchange, TypeU32(state), exchanged, pointer,
ConstantU32(state, scope), ConstantU32(state, MemorySemanticsNone),
ConstantU32(state, MemorySemanticsNone), next, observed});
const auto success = Binary(state, OpIEqual, TypeBool(state), exchanged, observed);
state.builder.AddFunction({OpLoopMerge, merge, cont, LoopControlNone});
state.builder.AddFunction({OpBranchConditional, success, merge, cont});
EmitLabel(state, cont);
state.builder.AddFunction({OpBranch, header});
EmitLabel(state, merge);
state.builder.AddFunction({OpMemoryBarrier, ConstantU32(state, scope),
ConstantU32(state, MemorySemanticsAcquireRelease | memory)});
return observed;
}
void StoreSubwordInBounds(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
const MemoryResourceAccess& resource, uint32_t address, uint32_t index,
uint32_t bits, uint32_t data) {
@@ -718,39 +683,6 @@ uint32_t EmitBufferAtomic64(ValueEmitContext& ctx, const IR::Inst& inst,
});
}
uint32_t FloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
bool max_value) {
struct OrderedBits {
uint32_t nan;
uint32_t zero;
uint32_t key;
};
const auto classify = [&](uint32_t bits) {
const auto value = Unary(state, OpBitcast, TypeF32(state), bits);
const auto cls = EmitClassifyF32(state, value);
const auto negative = Binary(
state, OpINotEqual, TypeBool(state),
Binary(state, OpBitwiseAnd, TypeU32(state), bits, ConstantU32(state, 0x80000000u)),
ConstantU32(state, 0));
const auto negative_key = Unary(state, OpNot, TypeU32(state), bits);
const auto positive_key =
Binary(state, OpBitwiseXor, TypeU32(state), bits, ConstantU32(state, 0x80000000u));
return OrderedBits {cls.nan, cls.zero,
Select(state, TypeU32(state), negative, negative_key, positive_key)};
};
const auto source_class = classify(source);
const auto old_class = classify(old);
const auto unordered =
Binary(state, OpLogicalOr, TypeBool(state),
Binary(state, OpLogicalOr, TypeBool(state), source_class.nan, old_class.nan),
Binary(state, OpLogicalAnd, TypeBool(state), source_class.zero, old_class.zero));
const auto ordered = Unary(state, OpLogicalNot, TypeBool(state), unordered);
const auto compare = Binary(state, max_value ? OpUGreaterThan : OpULessThan, TypeBool(state),
source_class.key, old_class.key);
return Select(state, TypeU32(state),
Binary(state, OpLogicalAnd, TypeBool(state), ordered, compare), source, old);
}
uint32_t FloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
bool max_value) {
return EmitValueOrZeroIfCondition(ctx.state, ctx.Arg(inst, inst.NumArgs() - 1), [&]() {
@@ -761,7 +693,7 @@ uint32_t FloatAtomic(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memo
EmitMemoryElementPointer(ctx.state, access.resource, access.index);
const auto source = ctx.Arg(inst, inst.NumArgs() - 2);
return AtomicUpdate(ctx.state, pointer, mem.kind, [&](uint32_t old) {
return FloatAtomicReplacement(ctx.state, old, source, max_value);
return EmitFloatAtomicReplacement(ctx.state, old, source, max_value);
});
});
});
src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemoryHelpers.cpp
+33
View File
@@ -325,6 +325,39 @@ void EmitDeviceAtomicMemoryBarrier(EmitterState& state) {
{OpMemoryBarrier, ConstantU32(state, ScopeDevice), ConstantU32(state, semantics)});
}
uint32_t EmitFloatAtomicReplacement(EmitterState& state, uint32_t old, uint32_t source,
bool max_value) {
struct OrderedBits {
uint32_t nan;
uint32_t zero;
uint32_t key;
};
const auto classify = [&](uint32_t bits) {
const auto value = EmitBitcastU32ToF32(state, bits);
const auto cls = EmitClassifyF32(state, value);
const auto negative = EmitCompareU32Constant(
state, OpINotEqual, EmitAndConstant(state, bits, 0x80000000u), 0u);
const auto negative_key = state.builder.AllocateId();
state.builder.AddFunction({OpNot, TypeU32(state), negative_key, bits});
const auto positive_key = EmitBinaryU32(
state, OpBitwiseXor, bits, ConstantU32(state, 0x80000000u));
return OrderedBits {
cls.nan, cls.zero,
EmitSelectValueU32(state, negative, negative_key, positive_key)};
};
const auto source_class = classify(source);
const auto old_class = classify(old);
const auto unordered = EmitLogicalOrBool(
state, EmitLogicalOrBool(state, source_class.nan, old_class.nan),
EmitLogicalAndBool(state, source_class.zero, old_class.zero));
const auto compare = state.builder.AllocateId();
state.builder.AddFunction({max_value ? OpUGreaterThan : OpULessThan, TypeBool(state), compare,
source_class.key, old_class.key});
return EmitSelectValueU32(
state, EmitLogicalAndBool(state, EmitLogicalNotBool(state, unordered), compare), source,
old);
}
uint32_t EmitDsSwizzleTargetLane(EmitterState& state, uint32_t subid, uint32_t control) {
if ((control & 0xc000u) == 0xc000u) {
const uint32_t mask = control & 0x1fu;