octo/yuzu-android
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

shader: Implement ATOM/S and RED

Browse Source
This commit is contained in:
ameerj
2021-04-11 02:07:02 -04:00
parent 479ca00071
commit 3db2b3effa
21 changed files with 1745 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

158
src/shader_recompiler/backend/spirv/emit_context.cpp
View File

@@ -15,6 +15,53 @@
namespace Shader::Backend::SPIRV {
namespace {
enum class CasFunctionType {
Increment,
Decrement,
FPAdd,
FPMin,
FPMax,
};
Id CasFunction(EmitContext& ctx, CasFunctionType function_type, Id value_type) {
const Id func_type{ctx.TypeFunction(value_type, value_type, value_type)};
const Id func{ctx.OpFunction(value_type, spv::FunctionControlMask::MaskNone, func_type)};
const Id op_a{ctx.OpFunctionParameter(value_type)};
const Id op_b{ctx.OpFunctionParameter(value_type)};
ctx.AddLabel();
Id result{};
switch (function_type) {
case CasFunctionType::Increment: {
const Id pred{ctx.OpUGreaterThanEqual(ctx.U1, op_a, op_b)};
const Id incr{ctx.OpIAdd(value_type, op_a, ctx.Constant(value_type, 1))};
result = ctx.OpSelect(value_type, pred, ctx.u32_zero_value, incr);
break;
}
case CasFunctionType::Decrement: {
const Id lhs{ctx.OpIEqual(ctx.U1, op_a, ctx.Constant(value_type, 0u))};
const Id rhs{ctx.OpUGreaterThan(ctx.U1, op_a, op_b)};
const Id pred{ctx.OpLogicalOr(ctx.U1, lhs, rhs)};
const Id decr{ctx.OpISub(value_type, op_a, ctx.Constant(value_type, 1))};
result = ctx.OpSelect(value_type, pred, op_b, decr);
break;
}
case CasFunctionType::FPAdd:
result = ctx.OpFAdd(value_type, op_a, op_b);
break;
case CasFunctionType::FPMin:
result = ctx.OpFMin(value_type, op_a, op_b);
break;
case CasFunctionType::FPMax:
result = ctx.OpFMax(value_type, op_a, op_b);
break;
default:
break;
}
ctx.OpReturnValue(result);
ctx.OpFunctionEnd();
return func;
}
Id ImageType(EmitContext& ctx, const TextureDescriptor& desc) {
const spv::ImageFormat format{spv::ImageFormat::Unknown};
const Id type{ctx.F32[1]};
@@ -196,6 +243,56 @@ Id EmitContext::Def(const IR::Value& value) {
}
}
Id EmitContext::CasLoop(Id function, CasPointerType pointer_type, Id value_type) {
const Id loop_header{OpLabel()};
const Id continue_block{OpLabel()};
const Id merge_block{OpLabel()};
const Id storage_type{pointer_type == CasPointerType::Shared ? shared_memory_u32_type
: storage_memory_u32};
const Id func_type{TypeFunction(value_type, U32[1], value_type, storage_type)};
const Id func{OpFunction(value_type, spv::FunctionControlMask::MaskNone, func_type)};
const Id index{OpFunctionParameter(U32[1])};
const Id op_b{OpFunctionParameter(value_type)};
const Id base{OpFunctionParameter(storage_type)};
AddLabel();
const Id one{Constant(U32[1], 1)};
OpBranch(loop_header);
AddLabel(loop_header);
OpLoopMerge(merge_block, continue_block, spv::LoopControlMask::MaskNone);
OpBranch(continue_block);
AddLabel(continue_block);
const Id word_pointer{pointer_type == CasPointerType::Shared
? OpAccessChain(shared_u32, base, index)
: OpAccessChain(storage_u32, base, u32_zero_value, index)};
if (value_type.value == F32[2].value) {
const Id u32_value{OpLoad(U32[1], word_pointer)};
const Id value{OpUnpackHalf2x16(F32[2], u32_value)};
const Id new_value{OpFunctionCall(value_type, function, value, op_b)};
const Id u32_new_value{OpPackHalf2x16(U32[1], new_value)};
const Id atomic_res{OpAtomicCompareExchange(U32[1], word_pointer, one, u32_zero_value,
u32_zero_value, u32_new_value, u32_value)};
const Id success{OpIEqual(U1, atomic_res, u32_value)};
OpBranchConditional(success, merge_block, loop_header);
AddLabel(merge_block);
OpReturnValue(OpUnpackHalf2x16(F32[2], atomic_res));
} else {
const Id value{OpLoad(U32[1], word_pointer)};
const Id new_value{OpBitcast(
U32[1], OpFunctionCall(value_type, function, OpBitcast(value_type, value), op_b))};
const Id atomic_res{OpAtomicCompareExchange(U32[1], word_pointer, one, u32_zero_value,
u32_zero_value, new_value, value)};
const Id success{OpIEqual(U1, atomic_res, value)};
OpBranchConditional(success, merge_block, loop_header);
AddLabel(merge_block);
OpReturnValue(OpBitcast(value_type, atomic_res));
}
OpFunctionEnd();
return func;
}
void EmitContext::DefineCommonTypes(const Info& info) {
void_id = TypeVoid();
@@ -300,9 +397,9 @@ void EmitContext::DefineSharedMemory(const IR::Program& program) {
}
const u32 num_elements{Common::DivCeil(program.shared_memory_size, 4U)};
const Id type{TypeArray(U32[1], Constant(U32[1], num_elements))};
const Id pointer_type{TypePointer(spv::StorageClass::Workgroup, type)};
shared_memory_u32_type = TypePointer(spv::StorageClass::Workgroup, type);
shared_u32 = TypePointer(spv::StorageClass::Workgroup, U32[1]);
shared_memory_u32 = AddGlobalVariable(pointer_type, spv::StorageClass::Workgroup);
shared_memory_u32 = AddGlobalVariable(shared_memory_u32_type, spv::StorageClass::Workgroup);
interfaces.push_back(shared_memory_u32);
const Id func_type{TypeFunction(void_id, U32[1], U32[1])};
@@ -346,6 +443,14 @@ void EmitContext::DefineSharedMemory(const IR::Program& program) {
if (program.info.uses_int16) {
shared_store_u16_func = make_function(16, 16);
}
if (program.info.uses_shared_increment) {
const Id inc_func{CasFunction(*this, CasFunctionType::Increment, U32[1])};
increment_cas_shared = CasLoop(inc_func, CasPointerType::Shared, U32[1]);
}
if (program.info.uses_shared_decrement) {
const Id dec_func{CasFunction(*this, CasFunctionType::Decrement, U32[1])};
decrement_cas_shared = CasLoop(dec_func, CasPointerType::Shared, U32[1]);
}
}
void EmitContext::DefineAttributeMemAccess(const Info& info) {
@@ -530,12 +635,12 @@ void EmitContext::DefineStorageBuffers(const Info& info, u32& binding) {
MemberName(struct_type, 0, "data");
MemberDecorate(struct_type, 0, spv::Decoration::Offset, 0U);
const Id storage_type{TypePointer(spv::StorageClass::StorageBuffer, struct_type)};
storage_memory_u32 = TypePointer(spv::StorageClass::StorageBuffer, struct_type);
storage_u32 = TypePointer(spv::StorageClass::StorageBuffer, U32[1]);
u32 index{};
for (const StorageBufferDescriptor& desc : info.storage_buffers_descriptors) {
const Id id{AddGlobalVariable(storage_type, spv::StorageClass::StorageBuffer)};
const Id id{AddGlobalVariable(storage_memory_u32, spv::StorageClass::StorageBuffer)};
Decorate(id, spv::Decoration::Binding, binding);
Decorate(id, spv::Decoration::DescriptorSet, 0U);
Name(id, fmt::format("ssbo{}", index));
@@ -546,6 +651,51 @@ void EmitContext::DefineStorageBuffers(const Info& info, u32& binding) {
index += desc.count;
binding += desc.count;
}
if (info.uses_global_increment) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id inc_func{CasFunction(*this, CasFunctionType::Increment, U32[1])};
increment_cas_ssbo = CasLoop(inc_func, CasPointerType::Ssbo, U32[1]);
}
if (info.uses_global_decrement) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id dec_func{CasFunction(*this, CasFunctionType::Decrement, U32[1])};
decrement_cas_ssbo = CasLoop(dec_func, CasPointerType::Ssbo, U32[1]);
}
if (info.uses_atomic_f32_add) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id add_func{CasFunction(*this, CasFunctionType::FPAdd, F32[1])};
f32_add_cas = CasLoop(add_func, CasPointerType::Ssbo, F32[1]);
}
if (info.uses_atomic_f16x2_add) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id add_func{CasFunction(*this, CasFunctionType::FPAdd, F16[2])};
f16x2_add_cas = CasLoop(add_func, CasPointerType::Ssbo, F16[2]);
}
if (info.uses_atomic_f16x2_min) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id func{CasFunction(*this, CasFunctionType::FPMin, F16[2])};
f16x2_min_cas = CasLoop(func, CasPointerType::Ssbo, F16[2]);
}
if (info.uses_atomic_f16x2_max) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id func{CasFunction(*this, CasFunctionType::FPMax, F16[2])};
f16x2_max_cas = CasLoop(func, CasPointerType::Ssbo, F16[2]);
}
if (info.uses_atomic_f32x2_add) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id add_func{CasFunction(*this, CasFunctionType::FPAdd, F32[2])};
f32x2_add_cas = CasLoop(add_func, CasPointerType::Ssbo, F32[2]);
}
if (info.uses_atomic_f32x2_min) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id func{CasFunction(*this, CasFunctionType::FPMin, F32[2])};
f32x2_min_cas = CasLoop(func, CasPointerType::Ssbo, F32[2]);
}
if (info.uses_atomic_f32x2_max) {
AddCapability(spv::Capability::VariablePointersStorageBuffer);
const Id func{CasFunction(*this, CasFunctionType::FPMax, F32[2])};
f32x2_max_cas = CasLoop(func, CasPointerType::Ssbo, F32[2]);
}
}
void EmitContext::DefineTextureBuffers(const Info& info, u32& binding) {

20
src/shader_recompiler/backend/spirv/emit_context.h
View File

@@ -94,6 +94,7 @@ public:
Id output_f32{};
Id storage_u32{};
Id storage_memory_u32{};
Id image_buffer_type{};
Id sampled_texture_buffer_type{};
@@ -136,9 +137,21 @@ public:
Id shared_memory_u32{};
Id shared_memory_u32x2{};
Id shared_memory_u32x4{};
Id shared_memory_u32_type{};
Id shared_store_u8_func{};
Id shared_store_u16_func{};
Id increment_cas_shared{};
Id increment_cas_ssbo{};
Id decrement_cas_shared{};
Id decrement_cas_ssbo{};
Id f32_add_cas{};
Id f16x2_add_cas{};
Id f16x2_min_cas{};
Id f16x2_max_cas{};
Id f32x2_add_cas{};
Id f32x2_min_cas{};
Id f32x2_max_cas{};
Id input_position{};
std::array<Id, 32> input_generics{};
@@ -153,6 +166,11 @@ public:
std::vector<Id> interfaces;
private:
enum class CasPointerType {
Shared,
Ssbo,
};
void DefineCommonTypes(const Info& info);
void DefineCommonConstants();
void DefineInterfaces(const Info& info);
@@ -171,6 +189,8 @@ private:
void DefineInputs(const Info& info);
void DefineOutputs(const Info& info);
[[nodiscard]] Id CasLoop(Id function, CasPointerType pointer_type, Id value_type);
};
} // namespace Shader::Backend::SPIRV

3
src/shader_recompiler/backend/spirv/emit_spirv.cpp
View File

@@ -238,6 +238,9 @@ void SetupCapabilities(const Profile& profile, const Info& info, EmitContext& ct
ctx.AddCapability(spv::Capability::SubgroupVoteKHR);
}
}
if (info.uses_64_bit_atomics && profile.support_int64_atomics) {
ctx.AddCapability(spv::Capability::Int64Atomics);
}
if (info.uses_typeless_image_reads && profile.support_typeless_image_loads) {
ctx.AddCapability(spv::Capability::StorageImageReadWithoutFormat);
}

95
src/shader_recompiler/backend/spirv/emit_spirv.h
View File

@@ -306,6 +306,101 @@ Id EmitUGreaterThan(EmitContext& ctx, Id lhs, Id rhs);
Id EmitINotEqual(EmitContext& ctx, Id lhs, Id rhs);
Id EmitSGreaterThanEqual(EmitContext& ctx, Id lhs, Id rhs);
Id EmitUGreaterThanEqual(EmitContext& ctx, Id lhs, Id rhs);
Id EmitSharedAtomicIAdd32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicSMin32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicUMin32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicSMax32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicUMax32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicInc32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicDec32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicAnd32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicOr32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicXor32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicExchange32(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitSharedAtomicExchange64(EmitContext& ctx, Id pointer_offset, Id value);
Id EmitStorageAtomicIAdd32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicSMin32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicUMin32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicSMax32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicUMax32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicInc32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicDec32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicAnd32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicOr32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicXor32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicExchange32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicIAdd64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicSMin64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicUMin64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicSMax64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicUMax64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicAnd64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicOr64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicXor64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicExchange64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicAddF32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicAddF16x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicAddF32x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicMinF16x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicMinF32x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicMaxF16x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitStorageAtomicMaxF32x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value);
Id EmitGlobalAtomicIAdd32(EmitContext& ctx);
Id EmitGlobalAtomicSMin32(EmitContext& ctx);
Id EmitGlobalAtomicUMin32(EmitContext& ctx);
Id EmitGlobalAtomicSMax32(EmitContext& ctx);
Id EmitGlobalAtomicUMax32(EmitContext& ctx);
Id EmitGlobalAtomicInc32(EmitContext& ctx);
Id EmitGlobalAtomicDec32(EmitContext& ctx);
Id EmitGlobalAtomicAnd32(EmitContext& ctx);
Id EmitGlobalAtomicOr32(EmitContext& ctx);
Id EmitGlobalAtomicXor32(EmitContext& ctx);
Id EmitGlobalAtomicExchange32(EmitContext& ctx);
Id EmitGlobalAtomicIAdd64(EmitContext& ctx);
Id EmitGlobalAtomicSMin64(EmitContext& ctx);
Id EmitGlobalAtomicUMin64(EmitContext& ctx);
Id EmitGlobalAtomicSMax64(EmitContext& ctx);
Id EmitGlobalAtomicUMax64(EmitContext& ctx);
Id EmitGlobalAtomicInc64(EmitContext& ctx);
Id EmitGlobalAtomicDec64(EmitContext& ctx);
Id EmitGlobalAtomicAnd64(EmitContext& ctx);
Id EmitGlobalAtomicOr64(EmitContext& ctx);
Id EmitGlobalAtomicXor64(EmitContext& ctx);
Id EmitGlobalAtomicExchange64(EmitContext& ctx);
Id EmitGlobalAtomicAddF32(EmitContext& ctx);
Id EmitGlobalAtomicAddF16x2(EmitContext& ctx);
Id EmitGlobalAtomicAddF32x2(EmitContext& ctx);
Id EmitGlobalAtomicMinF16x2(EmitContext& ctx);
Id EmitGlobalAtomicMinF32x2(EmitContext& ctx);
Id EmitGlobalAtomicMaxF16x2(EmitContext& ctx);
Id EmitGlobalAtomicMaxF32x2(EmitContext& ctx);
Id EmitLogicalOr(EmitContext& ctx, Id a, Id b);
Id EmitLogicalAnd(EmitContext& ctx, Id a, Id b);
Id EmitLogicalXor(EmitContext& ctx, Id a, Id b);

528
src/shader_recompiler/backend/spirv/emit_spirv_atomic.cpp Normal file
View File

@@ -0,0 +1,528 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "shader_recompiler/backend/spirv/emit_spirv.h"
namespace Shader::Backend::SPIRV {
namespace {
Id GetSharedPointer(EmitContext& ctx, Id offset, u32 index_offset = 0) {
const Id shift_id{ctx.Constant(ctx.U32[1], 2U)};
const Id shifted_value{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
const Id index{ctx.OpIAdd(ctx.U32[1], shifted_value, ctx.Constant(ctx.U32[1], index_offset))};
return ctx.profile.support_explicit_workgroup_layout
? ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, ctx.u32_zero_value, index)
: ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, index);
}
Id StorageIndex(EmitContext& ctx, const IR::Value& offset, size_t element_size) {
if (offset.IsImmediate()) {
const u32 imm_offset{static_cast<u32>(offset.U32() / element_size)};
return ctx.Constant(ctx.U32[1], imm_offset);
}
const u32 shift{static_cast<u32>(std::countr_zero(element_size))};
const Id index{ctx.Def(offset)};
if (shift == 0) {
return index;
}
const Id shift_id{ctx.Constant(ctx.U32[1], shift)};
return ctx.OpShiftRightLogical(ctx.U32[1], index, shift_id);
}
Id GetStoragePointer(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
u32 index_offset = 0) {
// TODO: Support reinterpreting bindings, guaranteed to be aligned
if (!binding.IsImmediate()) {
throw NotImplementedException("Dynamic storage buffer indexing");
}
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
const Id index{ctx.OpIAdd(ctx.U32[1], base_index, ctx.Constant(ctx.U32[1], index_offset))};
return ctx.OpAccessChain(ctx.storage_u32, ssbo, ctx.u32_zero_value, index);
}
std::pair<Id, Id> GetAtomicArgs(EmitContext& ctx) {
const Id scope{ctx.Constant(ctx.U32[1], static_cast<u32>(spv::Scope::Device))};
const Id semantics{ctx.u32_zero_value};
return {scope, semantics};
}
Id LoadU64(EmitContext& ctx, Id pointer_1, Id pointer_2) {
const Id value_1{ctx.OpLoad(ctx.U32[1], pointer_1)};
const Id value_2{ctx.OpLoad(ctx.U32[1], pointer_2)};
const Id original_composite{ctx.OpCompositeConstruct(ctx.U32[2], value_1, value_2)};
return ctx.OpBitcast(ctx.U64, original_composite);
}
void StoreResult(EmitContext& ctx, Id pointer_1, Id pointer_2, Id result) {
const Id composite{ctx.OpBitcast(ctx.U32[2], result)};
ctx.OpStore(pointer_1, ctx.OpCompositeExtract(ctx.U32[1], composite, 0));
ctx.OpStore(pointer_2, ctx.OpCompositeExtract(ctx.U32[1], composite, 1));
}
} // Anonymous namespace
Id EmitSharedAtomicIAdd32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicIAdd(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicSMin32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicSMin(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicUMin32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicUMin(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicSMax32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicSMax(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicUMax32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicUMax(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicInc32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id shift_id{ctx.Constant(ctx.U32[1], 2U)};
const Id index{ctx.OpShiftRightArithmetic(ctx.U32[1], pointer_offset, shift_id)};
return ctx.OpFunctionCall(ctx.U32[1], ctx.increment_cas_shared, index, value,
ctx.shared_memory_u32);
}
Id EmitSharedAtomicDec32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id shift_id{ctx.Constant(ctx.U32[1], 2U)};
const Id index{ctx.OpShiftRightArithmetic(ctx.U32[1], pointer_offset, shift_id)};
return ctx.OpFunctionCall(ctx.U32[1], ctx.decrement_cas_shared, index, value,
ctx.shared_memory_u32);
}
Id EmitSharedAtomicAnd32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicAnd(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicOr32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicOr(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicXor32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicXor(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicExchange32(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer{GetSharedPointer(ctx, pointer_offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicExchange(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitSharedAtomicExchange64(EmitContext& ctx, Id pointer_offset, Id value) {
const Id pointer_1{GetSharedPointer(ctx, pointer_offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicExchange(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetSharedPointer(ctx, pointer_offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
StoreResult(ctx, pointer_1, pointer_2, value);
return original_value;
}
Id EmitStorageAtomicIAdd32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicIAdd(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicSMin32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicSMin(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicUMin32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicUMin(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicSMax32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicSMax(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicUMax32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicUMax(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicInc32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
return ctx.OpFunctionCall(ctx.U32[1], ctx.increment_cas_ssbo, base_index, value, ssbo);
}
Id EmitStorageAtomicDec32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
return ctx.OpFunctionCall(ctx.U32[1], ctx.decrement_cas_ssbo, base_index, value, ssbo);
}
Id EmitStorageAtomicAnd32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicAnd(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicOr32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicOr(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicXor32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicXor(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicExchange32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer{GetStoragePointer(ctx, binding, offset)};
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicExchange(ctx.U32[1], pointer, scope, semantics, value);
}
Id EmitStorageAtomicIAdd64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicIAdd(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
const Id result{ctx.OpIAdd(ctx.U64, value, original_value)};
StoreResult(ctx, pointer_1, pointer_2, result);
return original_value;
}
Id EmitStorageAtomicSMin64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicSMin(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
const Id result{ctx.OpSMin(ctx.U64, value, original_value)};
StoreResult(ctx, pointer_1, pointer_2, result);
return original_value;
}
Id EmitStorageAtomicUMin64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicUMin(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
const Id result{ctx.OpUMin(ctx.U64, value, original_value)};
StoreResult(ctx, pointer_1, pointer_2, result);
return original_value;
}
Id EmitStorageAtomicSMax64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicSMax(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
const Id result{ctx.OpSMax(ctx.U64, value, original_value)};
StoreResult(ctx, pointer_1, pointer_2, result);
return original_value;
}
Id EmitStorageAtomicUMax64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicUMax(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
const Id result{ctx.OpUMax(ctx.U64, value, original_value)};
StoreResult(ctx, pointer_1, pointer_2, result);
return original_value;
}
Id EmitStorageAtomicAnd64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicAnd(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
const Id result{ctx.OpBitwiseAnd(ctx.U64, value, original_value)};
StoreResult(ctx, pointer_1, pointer_2, result);
return original_value;
}
Id EmitStorageAtomicOr64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicOr(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
const Id result{ctx.OpBitwiseOr(ctx.U64, value, original_value)};
StoreResult(ctx, pointer_1, pointer_2, result);
return original_value;
}
Id EmitStorageAtomicXor64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicXor(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
const Id result{ctx.OpBitwiseXor(ctx.U64, value, original_value)};
StoreResult(ctx, pointer_1, pointer_2, result);
return original_value;
}
Id EmitStorageAtomicExchange64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id pointer_1{GetStoragePointer(ctx, binding, offset)};
if (ctx.profile.support_int64_atomics) {
const auto [scope, semantics]{GetAtomicArgs(ctx)};
return ctx.OpAtomicExchange(ctx.U64, pointer_1, scope, semantics, value);
}
// LOG_WARNING(Render_Vulkan, "Int64 Atomics not supported, fallback to non-atomic");
const Id pointer_2{GetStoragePointer(ctx, binding, offset, 1)};
const Id original_value{LoadU64(ctx, pointer_1, pointer_2)};
StoreResult(ctx, pointer_1, pointer_2, value);
return original_value;
}
Id EmitStorageAtomicAddF32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
return ctx.OpFunctionCall(ctx.F32[1], ctx.f32_add_cas, base_index, value, ssbo);
}
Id EmitStorageAtomicAddF16x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
const Id result{ctx.OpFunctionCall(ctx.F16[2], ctx.f16x2_add_cas, base_index, value, ssbo)};
return ctx.OpBitcast(ctx.U32[1], result);
}
Id EmitStorageAtomicAddF32x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
const Id result{ctx.OpFunctionCall(ctx.F32[2], ctx.f32x2_add_cas, base_index, value, ssbo)};
return ctx.OpPackHalf2x16(ctx.U32[1], result);
}
Id EmitStorageAtomicMinF16x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
const Id result{ctx.OpFunctionCall(ctx.F16[2], ctx.f16x2_min_cas, base_index, value, ssbo)};
return ctx.OpBitcast(ctx.U32[1], result);
}
Id EmitStorageAtomicMinF32x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
const Id result{ctx.OpFunctionCall(ctx.F32[2], ctx.f32x2_min_cas, base_index, value, ssbo)};
return ctx.OpPackHalf2x16(ctx.U32[1], result);
}
Id EmitStorageAtomicMaxF16x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
const Id result{ctx.OpFunctionCall(ctx.F16[2], ctx.f16x2_max_cas, base_index, value, ssbo)};
return ctx.OpBitcast(ctx.U32[1], result);
}
Id EmitStorageAtomicMaxF32x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
Id value) {
const Id ssbo{ctx.ssbos[binding.U32()]};
const Id base_index{StorageIndex(ctx, offset, sizeof(u32))};
const Id result{ctx.OpFunctionCall(ctx.F32[2], ctx.f32x2_max_cas, base_index, value, ssbo)};
return ctx.OpPackHalf2x16(ctx.U32[1], result);
}
Id EmitGlobalAtomicIAdd32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicSMin32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicUMin32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicSMax32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicUMax32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicInc32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicDec32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicAnd32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicOr32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicXor32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicExchange32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicIAdd64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicSMin64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicUMin64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicSMax64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicUMax64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicInc64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicDec64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicAnd64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicOr64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicXor64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicExchange64(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicAddF32(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicAddF16x2(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicAddF32x2(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicMinF16x2(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicMinF32x2(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicMaxF16x2(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
Id EmitGlobalAtomicMaxF32x2(EmitContext&) {
throw NotImplementedException("SPIR-V Instruction");
}
} // namespace Shader::Backend::SPIRV