glsl: F16x2 storage atomics

This commit is contained in:
ameerj 2021-05-25 01:52:02 -04:00
parent 11ba190462
commit 9cc1b8a873
7 changed files with 64 additions and 58 deletions

View file

@ -39,6 +39,10 @@ void EmitContext::SetupExtensions(std::string& header) {
if (info.uses_atomic_f16x2_add || info.uses_atomic_f16x2_min || info.uses_atomic_f16x2_max) { if (info.uses_atomic_f16x2_add || info.uses_atomic_f16x2_min || info.uses_atomic_f16x2_max) {
header += "#extension NV_shader_atomic_fp16_vector : enable\n"; header += "#extension NV_shader_atomic_fp16_vector : enable\n";
} }
if (info.uses_fp16) {
// TODO: AMD
header += "#extension GL_NV_gpu_shader5 : enable\n";
}
} }
void EmitContext::DefineConstantBuffers() { void EmitContext::DefineConstantBuffers() {
@ -89,6 +93,18 @@ void EmitContext::DefineHelperFunctions() {
code += "uint CasFloatMax32x2(uint op_a,uint op_b){return " code += "uint CasFloatMax32x2(uint op_a,uint op_b){return "
"packHalf2x16(max(unpackHalf2x16(op_a),unpackHalf2x16(op_b)));}\n"; "packHalf2x16(max(unpackHalf2x16(op_a),unpackHalf2x16(op_b)));}\n";
} }
if (info.uses_atomic_f16x2_add) {
code += "uint CasFloatAdd16x2(uint op_a,uint op_b){return "
"packFloat2x16(unpackFloat2x16(op_a)+unpackFloat2x16(op_b));}\n";
}
if (info.uses_atomic_f16x2_min) {
code += "uint CasFloatMin16x2(uint op_a,uint op_b){return "
"packFloat2x16(min(unpackFloat2x16(op_a),unpackFloat2x16(op_b)));}\n";
}
if (info.uses_atomic_f16x2_max) {
code += "uint CasFloatMax16x2(uint op_a,uint op_b){return "
"packFloat2x16(max(unpackFloat2x16(op_a),unpackFloat2x16(op_b)));}\n";
}
// TODO: Track this usage // TODO: Track this usage
code += "uint CasMinS32(uint op_a,uint op_b){return uint(min(int(op_a),int(op_b)));}"; 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)));}"; code += "uint CasMaxS32(uint op_a,uint op_b){return uint(max(int(op_a),int(op_b)));}";

View file

@ -43,6 +43,11 @@ public:
Add<Type::U1>(format_str, inst, args...); Add<Type::U1>(format_str, inst, args...);
} }
template <typename... Args>
void AddF16x2(const char* format_str, IR::Inst& inst, Args&&... args) {
Add<Type::F16x2>(format_str, inst, args...);
}
template <typename... Args> template <typename... Args>
void AddU32(const char* format_str, IR::Inst& inst, Args&&... args) { void AddU32(const char* format_str, IR::Inst& inst, Args&&... args) {
Add<Type::U32>(format_str, inst, args...); Add<Type::U32>(format_str, inst, args...);

View file

@ -12,8 +12,7 @@
namespace Shader::Backend::GLSL { namespace Shader::Backend::GLSL {
namespace { namespace {
static constexpr std::string_view cas_loop{R"( static constexpr std::string_view cas_loop{R"(uint {};
uint {};
for (;;){{ for (;;){{
uint old_value={}; uint old_value={};
{}=atomicCompSwap({},old_value,{}({},{})); {}=atomicCompSwap({},old_value,{}({},{}));
@ -49,6 +48,14 @@ void CasFunctionF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding
const auto ret{ctx.reg_alloc.Define(inst)}; const auto ret{ctx.reg_alloc.Define(inst)};
CasFunction(ctx, ret, ssbo, u32_value, function); CasFunction(ctx, ret, ssbo, u32_value, function);
} }
void CasFunctionF16x2(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("packFloat2x16({})", value)};
const auto ret{ctx.reg_alloc.Define(inst)};
CasFunction(ctx, ret, ssbo, u32_value, function);
}
} // namespace } // namespace
void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicIAdd32(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -122,11 +129,8 @@ void EmitStorageAtomicSMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value&
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddS64("{}=int64_t(ivec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(), ctx.AddS64("{}=int64_t(ivec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1); binding.U32(), offset.U32() + 1);
ctx.Add(R"( ctx.Add("for(int i=0;i<2;++i){{ "
for(int i=0;i<2;++i){{ "ssbo{}[{}+i]=uint(min(int(ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));}}",
ssbo{}[{}+i]=uint(min(int(ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));
}}
)",
binding.U32(), offset.U32(), binding.U32(), offset.U32(), value); binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
} }
@ -135,12 +139,9 @@ void EmitStorageAtomicUMin64(EmitContext& ctx, IR::Inst& inst, const IR::Value&
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(), ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1); binding.U32(), offset.U32() + 1);
ctx.Add(R"( ctx.Add(
for(int i=0;i<2;++i){{ "for(int i=0;i<2;++i){{ ssbo{}[{}+i]=min(ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]);}}",
ssbo{}[{}+i]=min(ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]); binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
}}
)",
binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
} }
void EmitStorageAtomicSMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicSMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -148,11 +149,8 @@ void EmitStorageAtomicSMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value&
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddS64("{}=int64_t(ivec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(), ctx.AddS64("{}=int64_t(ivec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1); binding.U32(), offset.U32() + 1);
ctx.Add(R"( ctx.Add("for(int i=0;i<2;++i){{ "
for(int i=0;i<2;++i){{ "ssbo{}[{}+i]=uint(max(int(ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));}}",
ssbo{}[{}+i]=uint(max(int(ssbo{}[{}+i]),unpackInt2x32(int64_t({}))[i]));
}}
)",
binding.U32(), offset.U32(), binding.U32(), offset.U32(), value); binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
} }
@ -161,12 +159,9 @@ void EmitStorageAtomicUMax64(EmitContext& ctx, IR::Inst& inst, const IR::Value&
// LOG_WARNING(..., "Op falling to non-atomic"); // LOG_WARNING(..., "Op falling to non-atomic");
ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(), ctx.AddU64("{}=uint64_t(uvec2(ssbo{}[{}],ssbo{}[{}]));", inst, binding.U32(), offset.U32(),
binding.U32(), offset.U32() + 1); binding.U32(), offset.U32() + 1);
ctx.Add(R"( ctx.Add(
for(int i=0;i<2;++i){{ "for(int i=0;i<2;++i){{ssbo{}[{}+i]=max(ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]);}}",
ssbo{}[{}+i]=max(ssbo{}[{}+i],unpackUint2x32(uint64_t({}))[i]); binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
}}
)",
binding.U32(), offset.U32(), binding.U32(), offset.U32(), value);
} }
void EmitStorageAtomicAnd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, void EmitStorageAtomicAnd64(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
@ -202,45 +197,33 @@ void EmitStorageAtomicAddF32(EmitContext& ctx, IR::Inst& inst, const IR::Value&
CasFunctionF32(ctx, inst, binding, offset, value, "CasFloatAdd"); CasFunctionF32(ctx, inst, binding, offset, value, "CasFloatAdd");
} }
void EmitStorageAtomicAddF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst, void EmitStorageAtomicAddF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
[[maybe_unused]] const IR::Value& binding, const IR::Value& offset, std::string_view value) {
[[maybe_unused]] const IR::Value& offset, CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatAdd16x2");
[[maybe_unused]] std::string_view value) {
throw NotImplementedException("GLSL Instrucion");
} }
void EmitStorageAtomicAddF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst, void EmitStorageAtomicAddF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
[[maybe_unused]] const IR::Value& binding, const IR::Value& offset, std::string_view value) {
[[maybe_unused]] const IR::Value& offset,
[[maybe_unused]] std::string_view value) {
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatAdd32x2"); CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatAdd32x2");
} }
void EmitStorageAtomicMinF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst, void EmitStorageAtomicMinF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
[[maybe_unused]] const IR::Value& binding, const IR::Value& offset, std::string_view value) {
[[maybe_unused]] const IR::Value& offset, CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatMin16x2");
[[maybe_unused]] std::string_view value) {
throw NotImplementedException("GLSL Instrucion");
} }
void EmitStorageAtomicMinF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst, void EmitStorageAtomicMinF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
[[maybe_unused]] const IR::Value& binding, const IR::Value& offset, std::string_view value) {
[[maybe_unused]] const IR::Value& offset,
[[maybe_unused]] std::string_view value) {
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMin32x2"); CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMin32x2");
} }
void EmitStorageAtomicMaxF16x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst, void EmitStorageAtomicMaxF16x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
[[maybe_unused]] const IR::Value& binding, const IR::Value& offset, std::string_view value) {
[[maybe_unused]] const IR::Value& offset, CasFunctionF16x2(ctx, inst, binding, offset, value, "CasFloatMax16x2");
[[maybe_unused]] std::string_view value) {
throw NotImplementedException("GLSL Instrucion");
} }
void EmitStorageAtomicMaxF32x2([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst, void EmitStorageAtomicMaxF32x2(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding,
[[maybe_unused]] const IR::Value& binding, const IR::Value& offset, std::string_view value) {
[[maybe_unused]] const IR::Value& offset,
[[maybe_unused]] std::string_view value) {
CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMax32x2"); CasFunctionF32x2(ctx, inst, binding, offset, value, "CasFloatMax32x2");
} }

View file

@ -62,13 +62,12 @@ void EmitUnpackUint2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value
ctx.AddU32x2("{}=unpackUint2x32({});", inst, value); ctx.AddU32x2("{}=unpackUint2x32({});", inst, value);
} }
void EmitPackFloat2x16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] std::string_view value) { void EmitPackFloat2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value) {
throw NotImplementedException("GLSL Instruction"); ctx.AddU32("{}=packFloat2x16({});", inst, value);
} }
void EmitUnpackFloat2x16([[maybe_unused]] EmitContext& ctx, void EmitUnpackFloat2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value) {
[[maybe_unused]] std::string_view value) { ctx.AddF16x2("{}=unpackFloat2x16({});", inst, value);
throw NotImplementedException("GLSL Instruction");
} }
void EmitPackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value) { void EmitPackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value) {

View file

@ -224,8 +224,8 @@ void EmitBitCastF32U32(EmitContext& ctx, IR::Inst& inst, std::string_view value)
void EmitBitCastF64U64(EmitContext& ctx, IR::Inst& inst, std::string_view value); void EmitBitCastF64U64(EmitContext& ctx, IR::Inst& inst, std::string_view value);
void EmitPackUint2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value); void EmitPackUint2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value);
void EmitUnpackUint2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value); void EmitUnpackUint2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value);
void EmitPackFloat2x16(EmitContext& ctx, std::string_view value); void EmitPackFloat2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value);
void EmitUnpackFloat2x16(EmitContext& ctx, std::string_view value); void EmitUnpackFloat2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value);
void EmitPackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value); void EmitPackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value);
void EmitUnpackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value); void EmitUnpackHalf2x16(EmitContext& ctx, IR::Inst& inst, std::string_view value);
void EmitPackDouble2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value); void EmitPackDouble2x32(EmitContext& ctx, IR::Inst& inst, std::string_view value);

View file

@ -113,6 +113,8 @@ std::string RegAlloc::GetType(Type type, u32 index) {
switch (type) { switch (type) {
case Type::U1: case Type::U1:
return "bool "; return "bool ";
case Type::F16x2:
return "f16vec2 ";
case Type::U32: case Type::U32:
return "uint "; return "uint ";
case Type::S32: case Type::S32:

View file

@ -18,6 +18,7 @@ enum class Type;
namespace Shader::Backend::GLSL { namespace Shader::Backend::GLSL {
enum class Type : u32 { enum class Type : u32 {
U1, U1,
F16x2,
S32, S32,
U32, U32,
F32, F32,