N-archive/suyu
1
0
Fork 0
mirror of https://git.suyu.dev/suyu/suyu.git synced 2024-11-27 09:12:46 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions

shader: Add support for fp16 comparisons and misc fixes

Browse source
This commit is contained in:
ReinUsesLisp 2021-03-21 00:42:56 -03:00 committed by ameerj
parent 27fb97377e
commit a77e764726
11 changed files with 56 additions and 14 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/shader_recompiler/backend/spirv/emit_spirv.h
View file

@ -234,7 +234,9 @@ Id EmitFPOrdGreaterThanEqual64(EmitContext& ctx, Id lhs, Id rhs);
Id EmitFPUnordGreaterThanEqual16(EmitContext& ctx, Id lhs, Id rhs);
Id EmitFPUnordGreaterThanEqual32(EmitContext& ctx, Id lhs, Id rhs);
Id EmitFPUnordGreaterThanEqual64(EmitContext& ctx, Id lhs, Id rhs);
Id EmitFPIsNan16(EmitContext& ctx, Id value);
Id EmitFPIsNan32(EmitContext& ctx, Id value);
Id EmitFPIsNan64(EmitContext& ctx, Id value);
Id EmitIAdd32(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
void EmitIAdd64(EmitContext& ctx);
Id EmitISub32(EmitContext& ctx, Id a, Id b);

8
src/shader_recompiler/backend/spirv/emit_spirv_floating_point.cpp
View file

@ -346,8 +346,16 @@ Id EmitFPUnordGreaterThanEqual64(EmitContext& ctx, Id lhs, Id rhs) {
return ctx.OpFUnordGreaterThanEqual(ctx.U1, lhs, rhs);
}
Id EmitFPIsNan16(EmitContext& ctx, Id value) {
return ctx.OpIsNan(ctx.U1, value);
}
Id EmitFPIsNan32(EmitContext& ctx, Id value) {
return ctx.OpIsNan(ctx.U1, value);
}
Id EmitFPIsNan64(EmitContext& ctx, Id value) {
return ctx.OpIsNan(ctx.U1, value);
}
} // namespace Shader::Backend::SPIRV

21
src/shader_recompiler/frontend/ir/ir_emitter.cpp
View file

@ -895,15 +895,30 @@ U1 IREmitter::FPGreaterThanEqual(const F16F32F64& lhs, const F16F32F64& rhs, FpC
}
}
U1 IREmitter::FPIsNan(const F32& value) {
U1 IREmitter::FPIsNan(const F16F32F64& value) {
switch (value.Type()) {
case Type::F16:
return Inst<U1>(Opcode::FPIsNan16, value);
case Type::F32:
return Inst<U1>(Opcode::FPIsNan32, value);
case Type::F64:
return Inst<U1>(Opcode::FPIsNan64, value);
default:
ThrowInvalidType(value.Type());
}
}
U1 IREmitter::FPOrdered(const F32& lhs, const F32& rhs) {
U1 IREmitter::FPOrdered(const F16F32F64& lhs, const F16F32F64& rhs) {
if (lhs.Type() != rhs.Type()) {
throw InvalidArgument("Mismatching types {} and {}", lhs.Type(), rhs.Type());
}
return LogicalAnd(LogicalNot(FPIsNan(lhs)), LogicalNot(FPIsNan(rhs)));
}
U1 IREmitter::FPUnordered(const F32& lhs, const F32& rhs) {
U1 IREmitter::FPUnordered(const F16F32F64& lhs, const F16F32F64& rhs) {
if (lhs.Type() != rhs.Type()) {
throw InvalidArgument("Mismatching types {} and {}", lhs.Type(), rhs.Type());
}
return LogicalOr(FPIsNan(lhs), FPIsNan(rhs));
}

6
src/shader_recompiler/frontend/ir/ir_emitter.h
View file

@ -161,9 +161,9 @@ public:
FpControl control = {}, bool ordered = true);
[[nodiscard]] U1 FPGreaterThanEqual(const F16F32F64& lhs, const F16F32F64& rhs,
FpControl control = {}, bool ordered = true);
[[nodiscard]] U1 FPIsNan(const F32& value);
[[nodiscard]] U1 FPOrdered(const F32& lhs, const F32& rhs);
[[nodiscard]] U1 FPUnordered(const F32& lhs, const F32& rhs);
[[nodiscard]] U1 FPIsNan(const F16F32F64& value);
[[nodiscard]] U1 FPOrdered(const F16F32F64& lhs, const F16F32F64& rhs);
[[nodiscard]] U1 FPUnordered(const F16F32F64& lhs, const F16F32F64& rhs);
[[nodiscard]] F32F64 FPMax(const F32F64& lhs, const F32F64& rhs, FpControl control = {});
[[nodiscard]] F32F64 FPMin(const F32F64& lhs, const F32F64& rhs, FpControl control = {});

2
src/shader_recompiler/frontend/ir/opcodes.inc
View file

@ -236,7 +236,9 @@ OPCODE(FPOrdGreaterThanEqual64, U1, F64,
OPCODE(FPUnordGreaterThanEqual16, U1, F16, F16, )
OPCODE(FPUnordGreaterThanEqual32, U1, F32, F32, )
OPCODE(FPUnordGreaterThanEqual64, U1, F64, F64, )
OPCODE(FPIsNan16, U1, F16, )
OPCODE(FPIsNan32, U1, F32, )
OPCODE(FPIsNan64, U1, F64, )
// Integer operations
OPCODE(IAdd32, U32, U32, U32, )

3
src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_add.cpp
View file

@ -6,7 +6,6 @@
namespace Shader::Maxwell {
namespace {
void HADD2(TranslatorVisitor& v, u64 insn, Merge merge, bool ftz, bool sat, bool abs_a, bool neg_a,
Swizzle swizzle_a, bool abs_b, bool neg_b, Swizzle swizzle_b, const IR::U32& src_b) {
union {
@ -66,7 +65,7 @@ void HADD2(TranslatorVisitor& v, u64 insn, bool sat, bool abs_b, bool neg_b, Swi
HADD2(v, insn, hadd2.merge, hadd2.ftz != 0, sat, hadd2.abs_a != 0, hadd2.neg_a != 0,
hadd2.swizzle_a, abs_b, neg_b, swizzle_b, src_b);
}
} // namespace
} // Anonymous namespace
void TranslatorVisitor::HADD2_reg(u64 insn) {
union {

4
src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_fused_multiply_add.cpp
View file

@ -6,7 +6,6 @@
namespace Shader::Maxwell {
namespace {
void HFMA2(TranslatorVisitor& v, u64 insn, Merge merge, Swizzle swizzle_a, bool neg_b, bool neg_c,
Swizzle swizzle_b, Swizzle swizzle_c, const IR::U32& src_b, const IR::U32& src_c,
bool sat, HalfPrecision precision) {
@ -85,8 +84,7 @@ void HFMA2(TranslatorVisitor& v, u64 insn, bool neg_b, bool neg_c, Swizzle swizz
HFMA2(v, insn, hfma2.merge, hfma2.swizzle_a, neg_b, neg_c, swizzle_b, swizzle_c, src_b, src_c,
sat, precision);
}
} // namespace
} // Anonymous namespace
void TranslatorVisitor::HFMA2_reg(u64 insn) {
union {

3
src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_multiply.cpp
View file

@ -6,7 +6,6 @@
namespace Shader::Maxwell {
namespace {
void HMUL2(TranslatorVisitor& v, u64 insn, Merge merge, bool sat, bool abs_a, bool neg_a,
Swizzle swizzle_a, bool abs_b, bool neg_b, Swizzle swizzle_b, const IR::U32& src_b,
HalfPrecision precision) {
@ -79,7 +78,7 @@ void HMUL2(TranslatorVisitor& v, u64 insn, bool sat, bool abs_a, bool neg_a, boo
HMUL2(v, insn, hmul2.merge, sat, abs_a, neg_a, hmul2.swizzle_a, abs_b, neg_b, swizzle_b, src_b,
hmul2.precision);
}
} // namespace
} // Anonymous namespace
void TranslatorVisitor::HMUL2_reg(u64 insn) {
union {

1
src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_set.cpp
View file

@ -76,6 +76,7 @@ void TranslatorVisitor::HSET2_reg(u64 insn) {
BitField<35, 4, FPCompareOp> compare_op;
BitField<28, 2, Swizzle> swizzle_b;
} const hset2{insn};
HSET2(*this, insn, GetReg20(insn), hset2.bf != 0, hset2.ftz != 0, hset2.neg_b != 0,
hset2.abs_b != 0, hset2.compare_op, hset2.swizzle_b);
}

16
src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp
View file

@ -74,6 +74,9 @@ void VisitUsages(Info& info, IR::Inst& inst) {
case IR::Opcode::CompositeExtractF16x2:
case IR::Opcode::CompositeExtractF16x3:
case IR::Opcode::CompositeExtractF16x4:
case IR::Opcode::CompositeInsertF16x2:
case IR::Opcode::CompositeInsertF16x3:
case IR::Opcode::CompositeInsertF16x4:
case IR::Opcode::SelectF16:
case IR::Opcode::BitCastU16F16:
case IR::Opcode::BitCastF16U16:
@ -103,6 +106,19 @@ void VisitUsages(Info& info, IR::Inst& inst) {
case IR::Opcode::FPRoundEven16:
case IR::Opcode::FPSaturate16:
case IR::Opcode::FPTrunc16:
case IR::Opcode::FPOrdEqual16:
case IR::Opcode::FPUnordEqual16:
case IR::Opcode::FPOrdNotEqual16:
case IR::Opcode::FPUnordNotEqual16:
case IR::Opcode::FPOrdLessThan16:
case IR::Opcode::FPUnordLessThan16:
case IR::Opcode::FPOrdGreaterThan16:
case IR::Opcode::FPUnordGreaterThan16:
case IR::Opcode::FPOrdLessThanEqual16:
case IR::Opcode::FPUnordLessThanEqual16:
case IR::Opcode::FPOrdGreaterThanEqual16:
case IR::Opcode::FPUnordGreaterThanEqual16:
case IR::Opcode::FPIsNan16:
info.uses_fp16 = true;
break;
case IR::Opcode::FPAbs64:

2
src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp
View file

@ -74,6 +74,8 @@ IR::Opcode Replace(IR::Opcode op) {
return IR::Opcode::FPOrdGreaterThanEqual32;
case IR::Opcode::FPUnordGreaterThanEqual16:
return IR::Opcode::FPUnordGreaterThanEqual32;
case IR::Opcode::FPIsNan16:
return IR::Opcode::FPIsNan32;
case IR::Opcode::ConvertS16F16:
return IR::Opcode::ConvertS16F32;
case IR::Opcode::ConvertS32F16: