yuzu-emu
/
yuzu-mainline
Archived
1
0
Fork 0
Code

glasm: Initial GLASM fp64 support

This commit is contained in:
ReinUsesLisp 2021-05-09 18:03:01 -03:00 committed by ameerj
parent 9f851e3832
commit 4502595bc2
9 changed files with 152 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
src/shader_recompiler/backend/glasm/emit_context.h
View File

@ -29,6 +29,13 @@ public:
code += '\n';
}
template <typename... Args>
void LongAdd(const char* format_str, IR::Inst& inst, Args&&... args) {
code += fmt::format(format_str, reg_alloc.LongDefine(inst), std::forward<Args>(args)...);
// TODO: Remove this
code += '\n';
}
template <typename... Args>
void Add(const char* format_str, Args&&... args) {
code += fmt::format(format_str, std::forward<Args>(args)...);

17
src/shader_recompiler/backend/glasm/emit_glasm.cpp
View File

@ -42,7 +42,11 @@ template <bool scalar>
struct RegWrapper {
RegWrapper(EmitContext& ctx, Value value)
: reg_alloc{ctx.reg_alloc}, allocated{value.type != Type::Register} {
reg = allocated ? reg_alloc.AllocReg() : Register{value};
if (allocated) {
reg = value.type == Type::F64 ? reg_alloc.AllocLongReg() : reg_alloc.AllocReg();
} else {
reg = Register{value};
}
switch (value.type) {
case Type::Register:
break;
@ -55,6 +59,9 @@ struct RegWrapper {
case Type::F32:
ctx.Add("MOV.F {}.x,{};", reg, value.imm_f32);
break;
case Type::F64:
ctx.Add("MOV.F64 {}.x,{};", reg, value.imm_f64);
break;
}
}
~RegWrapper() {
@ -162,10 +169,12 @@ std::string EmitGLASM(const Profile&, IR::Program& program, Bindings&) {
for (size_t index = 0; index < ctx.reg_alloc.NumUsedRegisters(); ++index) {
header += fmt::format("R{},", index);
}
header += "RC;";
if (!program.info.storage_buffers_descriptors.empty()) {
header += "LONG TEMP LC;";
header += "RC;"
"LONG TEMP ";
for (size_t index = 0; index < ctx.reg_alloc.NumUsedLongRegisters(); ++index) {
header += fmt::format("D{},", index);
}
header += "DC;";
ctx.code.insert(0, header);
ctx.code += "END";
return ctx.code;

8
src/shader_recompiler/backend/glasm/emit_glasm_bitwise_conversion.cpp
View File

@ -72,4 +72,12 @@ void EmitUnpackHalf2x16(EmitContext& ctx, IR::Inst& inst, Register value) {
ctx.Add("UP2H {}.xy,{}.x;", inst, value);
}
void EmitPackDouble2x32(EmitContext& ctx, IR::Inst& inst, Register value) {
ctx.LongAdd("PK64 {}.x,{};", inst, value);
}
void EmitUnpackDouble2x32(EmitContext& ctx, IR::Inst& inst, Register value) {
ctx.Add("UP64 {}.xy,{}.x;", inst, value);
}
} // namespace Shader::Backend::GLASM

16
src/shader_recompiler/backend/glasm/emit_glasm_floating_point.cpp
View File

@ -10,7 +10,8 @@
namespace Shader::Backend::GLASM {
void EmitFPAbs16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] Register value) {
void EmitFPAbs16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
[[maybe_unused]] Register value) {
throw NotImplementedException("GLASM instruction");
}
@ -18,8 +19,8 @@ void EmitFPAbs32(EmitContext& ctx, IR::Inst& inst, ScalarF32 value) {
ctx.Add("MOV.F {}.x,|{}|;", inst, value);
}
void EmitFPAbs64([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] Register value) {
throw NotImplementedException("GLASM instruction");
void EmitFPAbs64(EmitContext& ctx, IR::Inst& inst, ScalarF64 value) {
ctx.LongAdd("MOV.F64 {}.x,|{}|;", inst, value);
}
void EmitFPAdd16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
@ -31,9 +32,8 @@ void EmitFPAdd32(EmitContext& ctx, IR::Inst& inst, ScalarF32 a, ScalarF32 b) {
ctx.Add("ADD.F {}.x,{},{};", inst, a, b);
}
void EmitFPAdd64([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
[[maybe_unused]] Register a, [[maybe_unused]] Register b) {
throw NotImplementedException("GLASM instruction");
void EmitFPAdd64(EmitContext& ctx, IR::Inst& inst, ScalarF64 a, ScalarF64 b) {
ctx.LongAdd("ADD.F64 {}.x,{},{};", inst, a, b);
}
void EmitFPFma16([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst& inst,
@ -94,8 +94,8 @@ void EmitFPNeg32(EmitContext& ctx, IR::Inst& inst, ScalarRegister value) {
ctx.Add("MOV.F {}.x,-{};", inst, value);
}
void EmitFPNeg64([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] Register value) {
throw NotImplementedException("GLASM instruction");
void EmitFPNeg64(EmitContext& ctx, IR::Inst& inst, Register value) {
ctx.LongAdd("MOV.F64 {}.x,-{};", inst, value);
}
void EmitFPSin([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] ScalarF32 value) {

12
src/shader_recompiler/backend/glasm/emit_glasm_instructions.h
View File

@ -202,20 +202,20 @@ void EmitPackFloat2x16(EmitContext& ctx, Register value);
void EmitUnpackFloat2x16(EmitContext& ctx, Register value);
void EmitPackHalf2x16(EmitContext& ctx, IR::Inst& inst, Register value);
void EmitUnpackHalf2x16(EmitContext& ctx, IR::Inst& inst, Register value);
void EmitPackDouble2x32(EmitContext& ctx, Register value);
void EmitUnpackDouble2x32(EmitContext& ctx, Register value);
void EmitPackDouble2x32(EmitContext& ctx, IR::Inst& inst, Register value);
void EmitUnpackDouble2x32(EmitContext& ctx, IR::Inst& inst, Register value);
void EmitGetZeroFromOp(EmitContext& ctx);
void EmitGetSignFromOp(EmitContext& ctx);
void EmitGetCarryFromOp(EmitContext& ctx);
void EmitGetOverflowFromOp(EmitContext& ctx);
void EmitGetSparseFromOp(EmitContext& ctx);
void EmitGetInBoundsFromOp(EmitContext& ctx);
void EmitFPAbs16(EmitContext& ctx, Register value);
void EmitFPAbs16(EmitContext& ctx, IR::Inst& inst, Register value);
void EmitFPAbs32(EmitContext& ctx, IR::Inst& inst, ScalarF32 value);
void EmitFPAbs64(EmitContext& ctx, Register value);
void EmitFPAbs64(EmitContext& ctx, IR::Inst& inst, ScalarF64 value);
void EmitFPAdd16(EmitContext& ctx, IR::Inst& inst, Register a, Register b);
void EmitFPAdd32(EmitContext& ctx, IR::Inst& inst, ScalarF32 a, ScalarF32 b);
void EmitFPAdd64(EmitContext& ctx, IR::Inst& inst, Register a, Register b);
void EmitFPAdd64(EmitContext& ctx, IR::Inst& inst, ScalarF64 a, ScalarF64 b);
void EmitFPFma16(EmitContext& ctx, IR::Inst& inst, Register a, Register b, Register c);
void EmitFPFma32(EmitContext& ctx, IR::Inst& inst, ScalarF32 a, ScalarF32 b, ScalarF32 c);
void EmitFPFma64(EmitContext& ctx, IR::Inst& inst, Register a, Register b, Register c);
@ -228,7 +228,7 @@ void EmitFPMul32(EmitContext& ctx, IR::Inst& inst, ScalarF32 a, ScalarF32 b);
void EmitFPMul64(EmitContext& ctx, IR::Inst& inst, Register a, Register b);
void EmitFPNeg16(EmitContext& ctx, Register value);
void EmitFPNeg32(EmitContext& ctx, IR::Inst& inst, ScalarRegister value);
void EmitFPNeg64(EmitContext& ctx, Register value);
void EmitFPNeg64(EmitContext& ctx, IR::Inst& inst, Register value);
void EmitFPSin(EmitContext& ctx, ScalarF32 value);
void EmitFPCos(EmitContext& ctx, ScalarF32 value);
void EmitFPExp2(EmitContext& ctx, ScalarF32 value);

10
src/shader_recompiler/backend/glasm/emit_glasm_memory.cpp
View File

@ -17,9 +17,9 @@ void StorageOp(EmitContext& ctx, const IR::Value& binding, ScalarU32 offset,
// address = c[binding].xy
// length = c[binding].z
const u32 sb_binding{binding.U32()};
ctx.Add("PK64.U LC,c[{}];" // pointer = address
"CVT.U64.U32 LC.z,{};" // offset = uint64_t(offset)
"ADD.U64 LC.x,LC.x,LC.z;" // pointer += offset
ctx.Add("PK64.U DC,c[{}];" // pointer = address
"CVT.U64.U32 DC.z,{};" // offset = uint64_t(offset)
"ADD.U64 DC.x,DC.x,DC.z;" // pointer += offset
"SLT.U.CC RC.x,{},c[{}].z;", // cc = offset < length
sb_binding, offset, offset, sb_binding);
if (else_expr.empty()) {
@ -32,13 +32,13 @@ void StorageOp(EmitContext& ctx, const IR::Value& binding, ScalarU32 offset,
template <typename ValueType>
void Store(EmitContext& ctx, const IR::Value& binding, ScalarU32 offset, ValueType value,
std::string_view size) {
StorageOp(ctx, binding, offset, fmt::format("STORE.{} {},LC.x;", size, value));
StorageOp(ctx, binding, offset, fmt::format("STORE.{} {},DC.x;", size, value));
}
void Load(EmitContext& ctx, IR::Inst& inst, const IR::Value& binding, ScalarU32 offset,
std::string_view size) {
const Register ret{ctx.reg_alloc.Define(inst)};
StorageOp(ctx, binding, offset, fmt::format("STORE.{} {},LC.x;", size, ret),
StorageOp(ctx, binding, offset, fmt::format("STORE.{} {},DC.x;", size, ret),
fmt::format("MOV.U {},{{0,0,0,0}};", ret));
}
} // Anonymous namespace

8
src/shader_recompiler/backend/glasm/emit_glasm_not_implemented.cpp
View File

@ -281,14 +281,6 @@ void EmitSelectF64(EmitContext& ctx, ScalarS32 cond, Register true_value, Regist
NotImplemented();
}
void EmitPackDouble2x32(EmitContext& ctx, Register value) {
NotImplemented();
}
void EmitUnpackDouble2x32(EmitContext& ctx, Register value) {
NotImplemented();
}
void EmitGetZeroFromOp(EmitContext& ctx) {
NotImplemented();
}

66
src/shader_recompiler/backend/glasm/reg_alloc.cpp
View File

@ -14,12 +14,11 @@
namespace Shader::Backend::GLASM {
Register RegAlloc::Define(IR::Inst& inst) {
const Id id{Alloc()};
inst.SetDefinition<Id>(id);
Register ret;
ret.type = Type::Register;
ret.id = id;
return ret;
return Define(inst, false);
}
Register RegAlloc::LongDefine(IR::Inst& inst) {
return Define(inst, true);
}
Value RegAlloc::Consume(const IR::Value& value) {
@ -40,6 +39,10 @@ Value RegAlloc::Consume(const IR::Value& value) {
ret.type = Type::F32;
ret.imm_f32 = value.F32();
break;
case IR::Type::F64:
ret.type = Type::F64;
ret.imm_f64 = value.F64();
break;
default:
throw NotImplementedException("Immediate type {}", value.Type());
}
@ -49,7 +52,14 @@ Value RegAlloc::Consume(const IR::Value& value) {
Register RegAlloc::AllocReg() {
Register ret;
ret.type = Type::Register;
ret.id = Alloc();
ret.id = Alloc(false);
return ret;
}
Register RegAlloc::AllocLongReg() {
Register ret;
ret.type = Type::Register;
ret.id = Alloc(true);
return ret;
}
@ -57,6 +67,15 @@ void RegAlloc::FreeReg(Register reg) {
Free(reg.id);
}
Register RegAlloc::Define(IR::Inst& inst, bool is_long) {
const Id id{Alloc(is_long)};
inst.SetDefinition<Id>(id);
Register ret;
ret.type = Type::Register;
ret.id = id;
return ret;
}
Value RegAlloc::Consume(IR::Inst& inst) {
const Id id{inst.Definition<Id>()};
inst.DestructiveRemoveUsage();
@ -69,18 +88,23 @@ Value RegAlloc::Consume(IR::Inst& inst) {
return ret;
}
Id RegAlloc::Alloc() {
for (size_t reg = 0; reg < NUM_REGS; ++reg) {
if (register_use[reg]) {
continue;
Id RegAlloc::Alloc(bool is_long) {
size_t& num_regs{is_long ? num_used_long_registers : num_used_registers};
std::bitset<NUM_REGS>& use{is_long ? long_register_use : register_use};
if (num_used_registers + num_used_long_registers < NUM_REGS) {
for (size_t reg = 0; reg < NUM_REGS; ++reg) {
if (use[reg]) {
continue;
}
num_regs = std::max(num_regs, reg + 1);
use[reg] = true;
Id ret{};
ret.index.Assign(static_cast<u32>(reg));
ret.is_long.Assign(is_long ? 1 : 0);
ret.is_spill.Assign(0);
ret.is_condition_code.Assign(0);
return ret;
}
num_used_registers = std::max(num_used_registers, reg + 1);
register_use[reg] = true;
Id ret{};
ret.index.Assign(static_cast<u32>(reg));
ret.is_spill.Assign(0);
ret.is_condition_code.Assign(0);
return ret;
}
throw NotImplementedException("Register spilling");
}
@ -89,7 +113,11 @@ void RegAlloc::Free(Id id) {
if (id.is_spill != 0) {
throw NotImplementedException("Free spill");
}
register_use[id.index] = false;
if (id.is_long != 0) {
long_register_use[id.index] = false;
} else {
register_use[id.index] = false;
}
}
} // namespace Shader::Backend::GLASM

63
src/shader_recompiler/backend/glasm/reg_alloc.h
View File

@ -27,12 +27,14 @@ enum class Type : u32 {
U32,
S32,
F32,
F64,
};
struct Id {
union {
u32 raw;
BitField<0, 30, u32> index;
BitField<0, 29, u32> index;
BitField<29, 1, u32> is_long;
BitField<30, 1, u32> is_spill;
BitField<31, 1, u32> is_condition_code;
};
@ -53,6 +55,7 @@ struct Value {
u32 imm_u32;
s32 imm_s32;
f32 imm_f32;
f64 imm_f64;
};
bool operator==(const Value& rhs) const noexcept {
@ -68,6 +71,8 @@ struct Value {
return imm_s32 == rhs.imm_s32;
case Type::F32:
return Common::BitCast<u32>(imm_f32) == Common::BitCast<u32>(rhs.imm_f32);
case Type::F64:
return Common::BitCast<u64>(imm_f64) == Common::BitCast<u64>(rhs.imm_f64);
}
return false;
}
@ -80,6 +85,7 @@ struct ScalarRegister : Value {};
struct ScalarU32 : Value {};
struct ScalarS32 : Value {};
struct ScalarF32 : Value {};
struct ScalarF64 : Value {};
class RegAlloc {
public:
@ -87,9 +93,13 @@ public:
Register Define(IR::Inst& inst);
Register LongDefine(IR::Inst& inst);
Value Consume(const IR::Value& value);
Register AllocReg();
[[nodiscard]] Register AllocReg();
[[nodiscard]] Register AllocLongReg();
void FreeReg(Register reg);
@ -97,19 +107,27 @@ public:
return num_used_registers;
}
[[nodiscard]] size_t NumUsedLongRegisters() const noexcept {
return num_used_long_registers;
}
private:
static constexpr size_t NUM_REGS = 4096;
static constexpr size_t NUM_ELEMENTS = 4;
Register Define(IR::Inst& inst, bool is_long);
Value Consume(IR::Inst& inst);
Id Alloc();
Id Alloc(bool is_long);
void Free(Id id);
EmitContext& ctx;
size_t num_used_registers{};
size_t num_used_long_registers{};
std::bitset<NUM_REGS> register_use{};
std::bitset<NUM_REGS> long_register_use{};
};
template <bool scalar, typename FormatContext>
@ -121,9 +139,17 @@ auto FormatTo(FormatContext& ctx, Id id) {
throw NotImplementedException("Spill emission");
}
if constexpr (scalar) {
return fmt::format_to(ctx.out(), "R{}.x", id.index.Value());
if (id.is_long != 0) {
return fmt::format_to(ctx.out(), "D{}.x", id.index.Value());
} else {
return fmt::format_to(ctx.out(), "R{}.x", id.index.Value());
}
} else {
return fmt::format_to(ctx.out(), "R{}", id.index.Value());
if (id.is_long != 0) {
return fmt::format_to(ctx.out(), "D{}", id.index.Value());
} else {
return fmt::format_to(ctx.out(), "R{}", id.index.Value());
}
}
}
@ -184,6 +210,8 @@ struct fmt::formatter<Shader::Backend::GLASM::ScalarU32> {
return fmt::format_to(ctx.out(), "{}", static_cast<u32>(value.imm_s32));
case Shader::Backend::GLASM::Type::F32:
return fmt::format_to(ctx.out(), "{}", Common::BitCast<u32>(value.imm_f32));
case Shader::Backend::GLASM::Type::F64:
break;
}
throw Shader::InvalidArgument("Invalid value type {}", value.type);
}
@ -205,6 +233,8 @@ struct fmt::formatter<Shader::Backend::GLASM::ScalarS32> {
return fmt::format_to(ctx.out(), "{}", value.imm_s32);
case Shader::Backend::GLASM::Type::F32:
return fmt::format_to(ctx.out(), "{}", Common::BitCast<s32>(value.imm_f32));
case Shader::Backend::GLASM::Type::F64:
break;
}
throw Shader::InvalidArgument("Invalid value type {}", value.type);
}
@ -226,6 +256,29 @@ struct fmt::formatter<Shader::Backend::GLASM::ScalarF32> {
return fmt::format_to(ctx.out(), "{}", Common::BitCast<s32>(value.imm_s32));
case Shader::Backend::GLASM::Type::F32:
return fmt::format_to(ctx.out(), "{}", value.imm_f32);
case Shader::Backend::GLASM::Type::F64:
break;
}
throw Shader::InvalidArgument("Invalid value type {}", value.type);
}
};
template <>
struct fmt::formatter<Shader::Backend::GLASM::ScalarF64> {
constexpr auto parse(format_parse_context& ctx) {
return ctx.begin();
}
template <typename FormatContext>
auto format(const Shader::Backend::GLASM::ScalarF64& value, FormatContext& ctx) {
switch (value.type) {
case Shader::Backend::GLASM::Type::Register:
return Shader::Backend::GLASM::FormatTo<true>(ctx, value.id);
case Shader::Backend::GLASM::Type::U32:
case Shader::Backend::GLASM::Type::S32:
case Shader::Backend::GLASM::Type::F32:
break;
case Shader::Backend::GLASM::Type::F64:
return format_to(ctx.out(), "{}", value.imm_f64);
}
throw Shader::InvalidArgument("Invalid value type {}", value.type);
}