N-archive/suyu
1
0
Fork 0
mirror of https://git.suyu.dev/suyu/suyu.git synced 2024-11-30 02:32:46 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions

glsl: Revert ssbo aliasing. Storage Atomics impl

Browse source
This commit is contained in:
ameerj 2021-05-25 01:35:30 -04:00
parent e99d01ff53
commit 11ba190462
5 changed files with 134 additions and 75 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

42
src/shader_recompiler/backend/glsl/emit_context.cpp
View file

@ -59,27 +59,8 @@ void EmitContext::DefineStorageBuffers() {
}
u32 binding{};
for (const auto& desc : info.storage_buffers_descriptors) {
if (info.uses_s32_atomics) {
Add("layout(std430,binding={}) buffer ssbo_{}_s32{{int ssbo{}_s32[];}};", binding,
binding, desc.cbuf_index, desc.count);
}
if (True(info.used_storage_buffer_types & IR::Type::U32)) {
Add("layout(std430,binding={}) buffer ssbo_{}_u32{{uint ssbo{}_u32[];}};", binding,
binding, desc.cbuf_index, desc.count);
}
if (True(info.used_storage_buffer_types & IR::Type::F32)) {
Add("layout(std430,binding={}) buffer ssbo_{}_f32{{float ssbo{}_f32[];}};", binding,
binding, desc.cbuf_index, desc.count);
}
if (True(info.used_storage_buffer_types & IR::Type::U32x2)) {
Add("layout(std430,binding={}) buffer ssbo_{}_u32x2{{uvec2 ssbo{}_u32x2[];}};", binding,
binding, desc.cbuf_index, desc.count);
}
if (True(info.used_storage_buffer_types & IR::Type::U64) ||
True(info.used_storage_buffer_types & IR::Type::F64)) {
Add("layout(std430,binding={}) buffer ssbo_{}_u64{{uint64_t ssbo{}_u64[];}};", binding,
binding, desc.cbuf_index, desc.count);
}
Add("layout(std430,binding={}) buffer ssbo_{}{{uint ssbo{}[];}};", binding, binding,
desc.cbuf_index, desc.count);
++binding;
}
}
@ -92,6 +73,25 @@ void EmitContext::DefineHelperFunctions() {
code +=
"uint CasDecrement(uint op_a,uint op_b){return(op_a==0||op_a>op_b)?op_b:(op_a-1u);}\n";
}
if (info.uses_atomic_f32_add) {
code += "uint CasFloatAdd(uint op_a,uint op_b){return "
"floatBitsToUint(uintBitsToFloat(op_a)+uintBitsToFloat(op_b));}\n";
}
if (info.uses_atomic_f32x2_add) {
code += "uint CasFloatAdd32x2(uint op_a,uint op_b){return "
"packHalf2x16(unpackHalf2x16(op_a)+unpackHalf2x16(op_b));}\n";
}
if (info.uses_atomic_f32x2_min) {
code += "uint CasFloatMin32x2(uint op_a,uint op_b){return "
"packHalf2x16(min(unpackHalf2x16(op_a),unpackHalf2x16(op_b)));}\n";
}
if (info.uses_atomic_f32x2_max) {
code += "uint CasFloatMax32x2(uint op_a,uint op_b){return "
"packHalf2x16(max(unpackHalf2x16(op_a),unpackHalf2x16(op_b)));}\n";
}
// TODO: Track this usage
code += "uint CasMinS32(uint op_a,uint op_b){return uint(min(int(op_a),int(op_b)));}";
code += "uint CasMaxS32(uint op_a,uint op_b){return uint(max(int(op_a),int(op_b)));}";
}
} // namespace Shader::Backend::GLSL

139
src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
View file

@ -13,132 +13,193 @@
namespace Shader::Backend::GLSL {
namespace {
static constexpr std::string_view cas_loop{R"(
{} {};
uint {};
for (;;){{
{} old_value={};
uint old_value={};
{}=atomicCompSwap({},old_value,{}({},{}));
if ({}==old_value){{break;}}
}})"};
void CasFunction(EmitContext& ctx, IR::Inst& inst, std::string_view ssbo, std::string_view value,
std::string_view type, std::string_view function) {
void CasFunction(EmitContext& ctx, std::string_view ret, std::string_view ssbo,
std::string_view value, std::string_view function) {
ctx.Add(cas_loop.data(), ret, ssbo, ret, ssbo, function, ssbo, value, ret);
}
void CasFunctionInt32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value, std::string_view function) {
const auto ret{ctx.reg_alloc.Define(inst)};
ctx.Add(cas_loop.data(), type, ret, type, ssbo, ret, ssbo, function, ssbo, value, ret);
const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
CasFunction(ctx, ret, ssbo, value, function);
}
void CasFunctionF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value, std::string_view function) {
const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
const std::string u32_value{fmt::format("floatBitsToUint({})", value)};
const auto ret{ctx.reg_alloc.Define(inst)};
const auto ret_32{ret + "_u32"};
CasFunction(ctx, ret_32, ssbo, u32_value, function);
ctx.Add("float {}=uintBitsToFloat({});", ret, ret_32);
}
void CasFunctionF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value, std::string_view function) {
const std::string ssbo{fmt::format("ssbo{}[{}]", binding.U32(), offset.U32())};
const std::string u32_value{fmt::format("packHalf2x16({})", value)};
const auto ret{ctx.reg_alloc.Define(inst)};
CasFunction(ctx, ret, ssbo, u32_value, function);
}
} // namespace
void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicAdd(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU32("{}=atomicAdd(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicSMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddS32("{}=atomicMin(ssbo{}_s32[{}],int({}));", inst, binding.U32(), offset.U32(), value);
const std::string u32_value{fmt::format("uint({})", value)};
CasFunctionInt32(ctx, inst, binding, offset, u32_value, "CasMinS32");
}
void EmitStorageAtomicUMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicMin(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU32("{}=atomicMin(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicSMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddS32("{}=atomicMax(ssbo{}_s32[{}],int({}));", inst, binding.U32(), offset.U32(), value);
const std::string u32_value{fmt::format("uint({})", value)};
CasFunctionInt32(ctx, inst, binding, offset, u32_value, "CasMaxS32");
}
void EmitStorageAtomicUMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicMax(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU32("{}=atomicMax(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicInc32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
[[maybe_unused]] const IR::Value& offset, std::string_view value) {
// const auto ret{ctx.reg_alloc.Define(inst)};
// const auto type{"uint"};
// ctx.Add(cas_loop.data(), type, ret, type, ssbo, ret, ssbo, "CasIncrement", ssbo, value, ret);
const std::string ssbo{fmt::format("ssbo{}_u32[{}]", binding.U32(), offset.U32())};
CasFunction(ctx, inst, ssbo, value, "uint", "CasIncrement");
const IR::Value& offset, std::string_view value) {
CasFunctionInt32(ctx, inst, binding, offset, value, "CasIncrement");
}
void EmitStorageAtomicDec32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
const std::string ssbo{fmt::format("ssbo{}_u32[{}]", binding.U32(), offset.U32())};
CasFunction(ctx, inst, ssbo, value, "uint", "CasDecrement");
CasFunctionInt32(ctx, inst, binding, offset, value, "CasDecrement");
}
void EmitStorageAtomicAnd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicAnd(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU32("{}=atomicAnd(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicOr32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicOr(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU32("{}=atomicOr(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicXor32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicXor(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU32("{}=atomicXor(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicExchange32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicExchange(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU32("{}=atomicExchange(ssbo{}[{}],{});", inst, binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicIAdd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
// ctx.AddU64("{}=atomicAdd(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU64("{}=ssbo{}_u64[{}];", inst, binding.U32(), offset.U32());
ctx.Add("ssbo{}_u64[{}]+={};", binding.U32(), offset.U32(), value);
// LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1);
ctx.Add("ssbo{}[{}]+=unpackUint2x32({}).x;ssbo{}[{}]+=unpackUint2x32({}).y;", binding.U32(),
offset.U32(), value, binding.U32(), offset.U32() + 1, value);
}
void EmitStorageAtomicSMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddS64("{}=atomicMin(int64_t(ssbo{}_u64[{}]),int64_t({}));", inst, binding.U32(),
offset.U32(), value);
// LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddS64("{}=int64_t(ivec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1);
ctx.Add(R"(
for(int i=0;i<2;++i){{
ssbo{}[{}+i]=uint(min(int(ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));
}}
)",
binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicUMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU64("{}=atomicMin(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
// LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1);
ctx.Add(R"(
for(int i=0;i<2;++i){{
ssbo{}[{}+i]=min(ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]);
}}
)",
binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicSMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddS64("{}=atomicMax(int64_t(ssbo{}_u64[{}]),int64_t({}));", inst, binding.U32(),
offset.U32(), value);
// LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddS64("{}=int64_t(ivec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1);
ctx.Add(R"(
for(int i=0;i<2;++i){{
ssbo{}[{}+i]=uint(max(int(ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));
}}
)",
binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicUMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU64("{}=atomicMax(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
// LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1);
ctx.Add(R"(
for(int i=0;i<2;++i){{
ssbo{}[{}+i]=max(ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]);
}}
)",
binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
}
void EmitStorageAtomicAnd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU64("{}=atomicAnd(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU64("{}=uint64_t(uvec2(atomicAnd(ssbo{}[{}],unpackUint2x32({}).x),atomicAnd(ssbo{}[{}],"
"unpackUint2x32({}).y)));",
inst, binding.U32(), offset.U32(), value, binding.U32(), offset.U32() + 1, value);
}
void EmitStorageAtomicOr64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU64("{}=atomicOr(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU64("{}=uint64_t(uvec2(atomicOr(ssbo{}[{}],unpackUint2x32({}).x),atomicOr(ssbo{}[{}],"
"unpackUint2x32({}).y)));",
inst, binding.U32(), offset.U32(), value, binding.U32(), offset.U32() + 1, value);
}
void EmitStorageAtomicXor64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU64("{}=atomicXor(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU64("{}=uint64_t(uvec2(atomicXor(ssbo{}[{}],unpackUint2x32({}).x),atomicXor(ssbo{}[{}],"
"unpackUint2x32({}).y)));",
inst, binding.U32(), offset.U32(), value, binding.U32(), offset.U32() + 1, value);
}
void EmitStorageAtomicExchange64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddU64("{}=atomicExchange(ssbo{}_u64[{}],{});", inst, binding.U32(), offset.U32(), value);
ctx.AddU64("{}=uint64_t(uvec2(atomicExchange(ssbo{}[{}],unpackUint2x32({}).x),atomicExchange("
"ssbo{}[{}],unpackUint2x32({}).y)));",
inst, binding.U32(), offset.U32(), value, binding.U32(), offset.U32() + 1, value);
}
void EmitStorageAtomicAddF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) {
ctx.AddF32("{}=atomicAdd(ssbo{}_u32[{}],{});", inst, binding.U32(), offset.U32(), value);
CasFunctionF32(ctx, inst, binding, offset, value, "CasFloatAdd");
}
void EmitStorageAtomicAddF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
@ -152,7 +213,7 @@ void EmitStorageAtomicAddF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused
[[maybe_unused]] const IR::Value& binding,
[[maybe_unused]] const IR::Value& offset,
[[maybe_unused]] std::string_view value) {
throw NotImplementedException("GLSL Instrucion");
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatAdd32x2");
}
void EmitStorageAtomicMinF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
@ -166,7 +227,7 @@ void EmitStorageAtomicMinF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused
[[maybe_unused]] const IR::Value& binding,
[[maybe_unused]] const IR::Value& offset,
[[maybe_unused]] std::string_view value) {
throw NotImplementedException("GLSL Instrucion");
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMin32x2");
}
void EmitStorageAtomicMaxF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
@ -180,7 +241,7 @@ void EmitStorageAtomicMaxF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused
[[maybe_unused]] const IR::Value& binding,
[[maybe_unused]] const IR::Value& offset,
[[maybe_unused]] std::string_view value) {
throw NotImplementedException("GLSL Instrucion");
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMax32x2");
}
void EmitGlobalAtomicIAdd32(EmitContext&) {

3
src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
View file

@ -113,7 +113,8 @@ void EmitLoadStorageU16(EmitContext& ctx, const IR::Value& binding, const IR::Va
void EmitLoadStorageS16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
void EmitLoadStorage32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset);
void EmitLoadStorage64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
void EmitLoadStorage64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset);
void EmitLoadStorage128(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset);
void EmitWriteStorageU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
std::string_view value);

14
src/shader_recompiler/backend/glsl/emit_glsl_memory.cpp
View file

@ -34,12 +34,13 @@ void EmitLoadStorageS16([[maybe_unused]] EmitContext& ctx,
void EmitLoadStorage32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) {
ctx.AddU32("{}=ssbo{}_u32[{}];", inst, binding.U32(), offset.U32());
ctx.AddU32("{}=ssbo{}[{}];", inst, binding.U32(), offset.U32());
}
void EmitLoadStorage64([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] const IR::Value& binding,
[[maybe_unused]] const IR::Value& offset) {
throw NotImplementedException("GLSL Instrucion");
void EmitLoadStorage64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset) {
ctx.AddU32x2("{}=uvec2(ssbo{}[{}],ssbo{}[{}]);", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1);
}
void EmitLoadStorage128([[maybe_unused]] EmitContext& ctx,
@ -78,12 +79,13 @@ void EmitWriteStorageS16([[maybe_unused]] EmitContext& ctx,
void EmitWriteStorage32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
std::string_view value) {
ctx.Add("ssbo{}_u32[{}]={};", binding.U32(), offset.U32(), value);
ctx.Add("ssbo{}[{}]={};", binding.U32(), offset.U32(), value);
}
void EmitWriteStorage64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
std::string_view value) {
ctx.Add("ssbo{}_u32x2[{}]={};", binding.U32(), offset.U32(), value);
ctx.Add("ssbo{}[{}]={}.x;", binding.U32(), offset.U32(), value);
ctx.Add("ssbo{}[{}]={}.y;", binding.U32(), offset.U32() + 1, value);
}
void EmitWriteStorage128([[maybe_unused]] EmitContext& ctx,

9
src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp
View file

@ -687,7 +687,9 @@ void VisitUsages(Info& info, IR::Inst& inst) {
case IR::Opcode::LoadStorage32:
case IR::Opcode::WriteStorage32:
case IR::Opcode::StorageAtomicIAdd32:
case IR::Opcode::StorageAtomicSMin32:
case IR::Opcode::StorageAtomicUMin32:
case IR::Opcode::StorageAtomicSMax32:
case IR::Opcode::StorageAtomicUMax32:
case IR::Opcode::StorageAtomicAnd32:
case IR::Opcode::StorageAtomicOr32:
@ -712,13 +714,6 @@ void VisitUsages(Info& info, IR::Inst& inst) {
case IR::Opcode::SharedAtomicExchange64:
info.uses_int64_bit_atomics = true;
break;
case IR::Opcode::GlobalAtomicSMin32:
case IR::Opcode::GlobalAtomicSMax32:
case IR::Opcode::StorageAtomicSMin32:
case IR::Opcode::StorageAtomicSMax32:
info.used_storage_buffer_types |= IR::Type::U32;
info.uses_s32_atomics = true;
break;
case IR::Opcode::GlobalAtomicInc32:
case IR::Opcode::StorageAtomicInc32:
info.used_storage_buffer_types |= IR::Type::U32;