Exverge/suyu
3
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity

glsl: Fix atomic SSBO offsets

Browse source 
and implement misc getters
This commit is contained in:
ameerj 2021-05-31 14:17:00 -04:00
parent 19247ba4fa
commit 68d075d1e8
4 changed files with 74 additions and 67 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

133
src/shader_recompiler/backend/glsl/emit_glsl_atomic.cpp
View file

@ -27,16 +27,16 @@ void SharedCasFunction(EmitContext& ctx, IR::Inst& inst, std::string_view offset
void SsboCasFunction(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void SsboCasFunction(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value, std::string_view function) { const IR::Value& offset, std::string_view value, std::string_view function) {
const auto ret{ctx.var_alloc.Define(inst, GlslVarType::U32)}; const auto ret{ctx.var_alloc.Define(inst, GlslVarType::U32)};
const std::string ssbo{ const std::string ssbo{fmt::format("{}_ssbo{}[{}>>2]", ctx.stage_name, binding.U32(),
fmt::format("{}_ssbo{}[{}]", ctx.stage_name, binding.U32(), offset.U32())}; ctx.var_alloc.Consume(offset))};
ctx.Add(cas_loop.data(), ssbo, ret, ssbo, function, ssbo, value, ret); ctx.Add(cas_loop.data(), ssbo, ret, ssbo, function, ssbo, value, ret);
} }
void SsboCasFunctionF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void SsboCasFunctionF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value, const IR::Value& offset, std::string_view value,
std::string_view function) { std::string_view function) {
const std::string ssbo{ const std::string ssbo{fmt::format("{}_ssbo{}[{}>>2]", ctx.stage_name, binding.U32(),
fmt::format("{}_ssbo{}[{}]", ctx.stage_name, binding.U32(), offset.U32())}; ctx.var_alloc.Consume(offset))};
const auto ret{ctx.var_alloc.Define(inst, GlslVarType::U32)}; const auto ret{ctx.var_alloc.Define(inst, GlslVarType::U32)};
ctx.Add(cas_loop.data(), ssbo, ret, ssbo, function, ssbo, value, ret); ctx.Add(cas_loop.data(), ssbo, ret, ssbo, function, ssbo, value, ret);
ctx.AddF32("{}=uintBitsToFloat({});", inst, ret); ctx.AddF32("{}=uintBitsToFloat({});", inst, ret);
@ -111,8 +111,8 @@ void EmitSharedAtomicExchange64(EmitContext& ctx, IR::Inst& inst, std::string_vi
void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicAdd({}_ssbo{}[{}],{});", inst, ctx.stage_name, binding.U32(), offset.U32(), ctx.AddU32("{}=atomicAdd({}_ssbo{}[{}>>2],{});", inst, ctx.stage_name, binding.U32(),
value); ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicSMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicSMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -123,8 +123,8 @@ void EmitStorageAtomicSMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value&
void EmitStorageAtomicUMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicUMin32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicMin({}_ssbo{}[{}],{});", inst, ctx.stage_name, binding.U32(), offset.U32(), ctx.AddU32("{}=atomicMin({}_ssbo{}[{}>>2],{});", inst, ctx.stage_name, binding.U32(),
value); ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicSMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicSMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -135,8 +135,8 @@ void EmitStorageAtomicSMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value&
void EmitStorageAtomicUMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicUMax32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicMax({}_ssbo{}[{}],{});", inst, ctx.stage_name, binding.U32(), offset.U32(), ctx.AddU32("{}=atomicMax({}_ssbo{}[{}>>2],{});", inst, ctx.stage_name, binding.U32(),
value); ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicInc32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicInc32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -151,116 +151,123 @@ void EmitStorageAtomicDec32(EmitContext& ctx, IR::Inst& inst, const IR::Value& b
void EmitStorageAtomicAnd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicAnd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicAnd({}_ssbo{}[{}],{});", inst, ctx.stage_name, binding.U32(), offset.U32(), ctx.AddU32("{}=atomicAnd({}_ssbo{}[{}>>2],{});", inst, ctx.stage_name, binding.U32(),
value); ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicOr32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicOr32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicOr({}_ssbo{}[{}],{});", inst, ctx.stage_name, binding.U32(), offset.U32(), ctx.AddU32("{}=atomicOr({}_ssbo{}[{}>>2],{});", inst, ctx.stage_name, binding.U32(),
value); ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicXor32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicXor32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicXor({}_ssbo{}[{}],{});", inst, ctx.stage_name, binding.U32(), offset.U32(), ctx.AddU32("{}=atomicXor({}_ssbo{}[{}>>2],{});", inst, ctx.stage_name, binding.U32(),
value); ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicExchange32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicExchange32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU32("{}=atomicExchange({}_ssbo{}[{}],{});", inst, ctx.stage_name, binding.U32(), ctx.AddU32("{}=atomicExchange({}_ssbo{}[{}>>2],{});", inst, ctx.stage_name, binding.U32(),
offset.U32(), value); ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicIAdd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicIAdd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddU64("{}=packUint2x32(uvec2({}_ssbo{}[{}],{}_ssbo{}[{}]));", inst, ctx.stage_name, ctx.AddU64("{}=packUint2x32(uvec2({}_ssbo{}[{}>>2],{}_ssbo{}[({}>>2)+1]));", inst,
binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), offset.U32() + 1); ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
ctx.Add("{}_ssbo{}[{}]+=unpackUint2x32({}).x;{}_ssbo{}[{}]+=unpackUint2x32({}).y;", binding.U32(), ctx.var_alloc.Consume(offset));
ctx.stage_name, binding.U32(), offset.U32(), value, ctx.stage_name, binding.U32(), ctx.Add("{}_ssbo{}[{}>>2]+=unpackUint2x32({}).x;{}_ssbo{}[({}>>2)+1]+=unpackUint2x32({}).y;",
offset.U32() + 1, value); ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), value, ctx.stage_name,
binding.U32(), ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicSMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicSMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddS64("{}=packInt2x32(ivec2({}_ssbo{}[{}],{}_ssbo{}[{}]));", inst, ctx.stage_name, ctx.AddS64("{}=packInt2x32(ivec2({}_ssbo{}[{}>>2],{}_ssbo{}[({}>>2)+1]));", inst,
binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), offset.U32() + 1); ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
binding.U32(), ctx.var_alloc.Consume(offset));
ctx.Add("for(int i=0;i<2;++i){{ " ctx.Add("for(int i=0;i<2;++i){{ "
"{}_ssbo{}[{}+i]=uint(min(int({}_ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));}}", "{}_ssbo{}[({}>>2)+i]=uint(min(int({}_ssbo{}[({}>>2)+i]),unpackInt2x32(int64_t({}))[i])"
ctx.stage_name, binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), ");}}",
offset.U32(), value); ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
binding.U32(), ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicUMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicUMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddU64("{}=packUint2x32(uvec2({}_ssbo{}[{}],{}_ssbo{}[{}]));", inst, ctx.stage_name, ctx.AddU64("{}=packUint2x32(uvec2({}_ssbo{}[{}>>2],{}_ssbo{}[({}>>2)+1]));", inst,
binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), offset.U32() + 1); ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
binding.U32(), ctx.var_alloc.Consume(offset));
ctx.Add("for(int i=0;i<2;++i){{ " ctx.Add("for(int i=0;i<2;++i){{ "
"{}_ssbo{}[{}+i]=min({}_ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]);}}", "{}_ssbo{}[({}>>2)+i]=min({}_ssbo{}[({}>>2)+i],unpackUint2x32(uint64_t({}))[i]);}}",
ctx.stage_name, binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
offset.U32(), value); binding.U32(), ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicSMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicSMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddS64("{}=packInt2x32(ivec2({}_ssbo{}[{}],{}_ssbo{}[{}]));", inst, ctx.stage_name, ctx.AddS64("{}=packInt2x32(ivec2({}_ssbo{}[{}>>2],{}_ssbo{}[({}>>2)+1]));", inst,
binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), offset.U32() + 1); ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
binding.U32(), ctx.var_alloc.Consume(offset));
ctx.Add("for(int i=0;i<2;++i){{ " ctx.Add("for(int i=0;i<2;++i){{ "
"{}_ssbo{}[{}+i]=uint(max(int({}_ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));}}", "{}_ssbo{}[({}>>2)+i]=uint(max(int({}_ssbo{}[({}>>2)+i]),unpackInt2x32(int64_t({}))[i])"
ctx.stage_name, binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), ");}}",
offset.U32(), value); ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
binding.U32(), ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicUMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicUMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddU64("{}=packUint2x32(uvec2({}_ssbo{}[{}],{}_ssbo{}[{}]));", inst, ctx.stage_name,
binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), offset.U32() + 1); ctx.AddU64("{}=packUint2x32(uvec2({}_ssbo{}[{}>>2],{}_ssbo{}[({}>>2)+1]));", inst,
ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
binding.U32(), ctx.var_alloc.Consume(offset));
ctx.Add("for(int " ctx.Add("for(int "
"i=0;i<2;++i){{{}_ssbo{}[{}+i]=max({}_ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]);}}", "i=0;i<2;++i){{{}_ssbo{}[({}>>2)+i]=max({}_ssbo{}[({}>>2)+i],unpackUint2x32(uint64_t({}"
ctx.stage_name, binding.U32(), offset.U32(), ctx.stage_name, binding.U32(), "))[i]);}}",
offset.U32(), value); ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), ctx.stage_name,
binding.U32(), ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicAnd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicAnd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU64("{}=packUint2x32(uvec2(atomicAnd({}_ssbo{}[{}],unpackUint2x32({}).x),atomicAnd({}_" ctx.AddU64(
"ssbo{}[{}]," "{}=packUint2x32(uvec2(atomicAnd({}_ssbo{}[{}>>2],unpackUint2x32({}).x),atomicAnd({}_"
"unpackUint2x32({}).y)));", "ssbo{}[({}>>2)+1],unpackUint2x32({}).y)));",
inst, ctx.stage_name, binding.U32(), offset.U32(), value, ctx.stage_name, inst, ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), value, ctx.stage_name,
binding.U32(), offset.U32() + 1, value); binding.U32(), ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicOr64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicOr64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU64( ctx.AddU64("{}=packUint2x32(uvec2(atomicOr({}_ssbo{}[{}>>2],unpackUint2x32({}).x),atomicOr({}_"
"{}=packUint2x32(uvec2(atomicOr({}_ssbo{}[{}],unpackUint2x32({}).x),atomicOr({}_ssbo{}[{}]," "ssbo{}[({}>>2)+1],unpackUint2x32({}).y)));",
"unpackUint2x32({}).y)));", inst, ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), value,
inst, ctx.stage_name, binding.U32(), offset.U32(), value, ctx.stage_name, binding.U32(), ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), value);
offset.U32() + 1, value);
} }
void EmitStorageAtomicXor64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicXor64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU64("{}=packUint2x32(uvec2(atomicXor({}_ssbo{}[{}],unpackUint2x32({}).x),atomicXor({}_" ctx.AddU64(
"ssbo{}[{}]," "{}=packUint2x32(uvec2(atomicXor({}_ssbo{}[{}>>2],unpackUint2x32({}).x),atomicXor({}_"
"unpackUint2x32({}).y)));", "ssbo{}[({}>>2)+1],unpackUint2x32({}).y)));",
inst, ctx.stage_name, binding.U32(), offset.U32(), value, ctx.stage_name, inst, ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), value, ctx.stage_name,
binding.U32(), offset.U32() + 1, value); binding.U32(), ctx.var_alloc.Consume(offset), value);
} }
void EmitStorageAtomicExchange64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicExchange64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
const IR::Value& offset, std::string_view value) { const IR::Value& offset, std::string_view value) {
ctx.AddU64( ctx.AddU64("{}=packUint2x32(uvec2(atomicExchange({}_ssbo{}[{}>>2],unpackUint2x32({}).x),"
"{}=packUint2x32(uvec2(atomicExchange({}_ssbo{}[{}],unpackUint2x32({}).x),atomicExchange(" "atomicExchange({}_ssbo{}[({}>>2)+1],unpackUint2x32({}).y)));",
"{}_ssbo{}[{}],unpackUint2x32({}).y)));", inst, ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), value,
inst, ctx.stage_name, binding.U32(), offset.U32(), value, ctx.stage_name, binding.U32(), ctx.stage_name, binding.U32(), ctx.var_alloc.Consume(offset), value);
offset.U32() + 1, value);
} }
void EmitStorageAtomicAddF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicAddF32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,

2
src/shader_recompiler/backend/glsl/emit_glsl_instructions.h
View file

@ -92,7 +92,7 @@ void EmitWorkgroupId(EmitContext& ctx, IR::Inst& inst);
void EmitLocalInvocationId(EmitContext& ctx, IR::Inst& inst); void EmitLocalInvocationId(EmitContext& ctx, IR::Inst& inst);
void EmitInvocationId(EmitContext& ctx, IR::Inst& inst); void EmitInvocationId(EmitContext& ctx, IR::Inst& inst);
void EmitSampleId(EmitContext& ctx, IR::Inst& inst); void EmitSampleId(EmitContext& ctx, IR::Inst& inst);
void EmitIsHelperInvocation(EmitContext& ctx); void EmitIsHelperInvocation(EmitContext& ctx, IR::Inst& inst);
void EmitYDirection(EmitContext& ctx, IR::Inst& inst); void EmitYDirection(EmitContext& ctx, IR::Inst& inst);
void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, std::string_view word_offset); void EmitLoadLocal(EmitContext& ctx, IR::Inst& inst, std::string_view word_offset);
void EmitWriteLocal(EmitContext& ctx, std::string_view word_offset, std::string_view value); void EmitWriteLocal(EmitContext& ctx, std::string_view word_offset, std::string_view value);

4
src/shader_recompiler/backend/glsl/emit_glsl_not_implemented.cpp
View file

@ -211,8 +211,8 @@ void EmitSampleId(EmitContext& ctx, IR::Inst& inst) {
NotImplemented(); NotImplemented();
} }
void EmitIsHelperInvocation(EmitContext& ctx) { void EmitIsHelperInvocation(EmitContext& ctx, IR::Inst& inst) {
NotImplemented(); ctx.AddU1("{}=gl_HelperInvocation;", inst);
} }
void EmitYDirection(EmitContext& ctx, IR::Inst& inst) { void EmitYDirection(EmitContext& ctx, IR::Inst& inst) {

2
src/shader_recompiler/backend/glsl/emit_glsl_warp.cpp
View file

@ -38,7 +38,7 @@ std::string GetMaxThreadId(std::string_view thread_id, std::string_view clamp,
} // namespace } // namespace
void EmitLaneId([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst) { void EmitLaneId([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst) {
NotImplemented(); ctx.AddU32("{}=gl_SubGroupInvocationARB&31u;", inst);
} }
void EmitVoteAll(EmitContext& ctx, IR::Inst& inst, std::string_view pred) { void EmitVoteAll(EmitContext& ctx, IR::Inst& inst, std::string_view pred) {