sneedmca/suyu
1
0
Fork 0
mirror of https://git.suyu.dev/suyu/suyu synced 2024-12-24 18:32:49 -06:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

shader: Fix floating point comparison for FP16

Browse source
This commit is contained in:
FernandoS27 2021-03-21 04:33:19 +01:00 committed by ameerj
parent e10d9c1b8e
commit 27fb97377e
5 changed files with 54 additions and 30 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/shader_recompiler/frontend/maxwell/translate/impl/common_funcs.cpp
View file

@ -72,7 +72,7 @@ bool IsCompareOpOrdered(FPCompareOp op) {
} }
} }
IR::U1 FloatingPointCompare(IR::IREmitter& ir, const IR::F32& operand_1, const IR::F32& operand_2, IR::U1 FloatingPointCompare(IR::IREmitter& ir, const IR::F16F32F64& operand_1, const IR::F16F32F64& operand_2,
FPCompareOp compare_op, IR::FpControl control) { FPCompareOp compare_op, IR::FpControl control) {
const bool ordered{IsCompareOpOrdered(compare_op)}; const bool ordered{IsCompareOpOrdered(compare_op)};
switch (compare_op) { switch (compare_op) {

4
src/shader_recompiler/frontend/maxwell/translate/impl/common_funcs.h
View file

@ -18,7 +18,7 @@ namespace Shader::Maxwell {
[[nodiscard]] bool IsCompareOpOrdered(FPCompareOp op); [[nodiscard]] bool IsCompareOpOrdered(FPCompareOp op);
[[nodiscard]] IR::U1 FloatingPointCompare(IR::IREmitter& ir, const IR::F32& operand_1, [[nodiscard]] IR::U1 FloatingPointCompare(IR::IREmitter& ir, const IR::F16F32F64& operand_1,
const IR::F32& operand_2, FPCompareOp compare_op, const IR::F16F32F64& operand_2, FPCompareOp compare_op,
IR::FpControl control = {}); IR::FpControl control = {});
} // namespace Shader::Maxwell } // namespace Shader::Maxwell

34
src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_set.cpp
View file

@ -22,8 +22,8 @@ void HSET2(TranslatorVisitor& v, u64 insn, const IR::U32& src_b, bool bf, bool f
auto [lhs_a, rhs_a]{Extract(v.ir, v.X(hset2.src_a_reg), hset2.swizzle_a)}; auto [lhs_a, rhs_a]{Extract(v.ir, v.X(hset2.src_a_reg), hset2.swizzle_a)};
auto [lhs_b, rhs_b]{Extract(v.ir, src_b, swizzle_b)}; auto [lhs_b, rhs_b]{Extract(v.ir, src_b, swizzle_b)};
// TODO: Implement FP16 FloatingPointCompare
//if (lhs_a.Type() != lhs_b.Type()) { if (lhs_a.Type() != lhs_b.Type()) {
if (lhs_a.Type() == IR::Type::F16) { if (lhs_a.Type() == IR::Type::F16) {
lhs_a = v.ir.FPConvert(32, lhs_a); lhs_a = v.ir.FPConvert(32, lhs_a);
rhs_a = v.ir.FPConvert(32, rhs_a); rhs_a = v.ir.FPConvert(32, rhs_a);
@ -32,7 +32,7 @@ void HSET2(TranslatorVisitor& v, u64 insn, const IR::U32& src_b, bool bf, bool f
lhs_b = v.ir.FPConvert(32, lhs_b); lhs_b = v.ir.FPConvert(32, lhs_b);
rhs_b = v.ir.FPConvert(32, rhs_b); rhs_b = v.ir.FPConvert(32, rhs_b);
} }
//} }
lhs_a = v.ir.FPAbsNeg(lhs_a, hset2.abs_a != 0, hset2.neg_a != 0); lhs_a = v.ir.FPAbsNeg(lhs_a, hset2.abs_a != 0, hset2.neg_a != 0);
rhs_a = v.ir.FPAbsNeg(rhs_a, hset2.abs_a != 0, hset2.neg_a != 0); rhs_a = v.ir.FPAbsNeg(rhs_a, hset2.abs_a != 0, hset2.neg_a != 0);
@ -94,22 +94,22 @@ void TranslatorVisitor::HSET2_cbuf(u64 insn) {
} }
void TranslatorVisitor::HSET2_imm(u64 insn) { void TranslatorVisitor::HSET2_imm(u64 insn) {
union { union {
u64 insn; u64 insn;
BitField<53, 1, u64> bf; BitField<53, 1, u64> bf;
BitField<54, 1, u64> ftz; BitField<54, 1, u64> ftz;
BitField<49, 4, FPCompareOp> compare_op; BitField<49, 4, FPCompareOp> compare_op;
BitField<56, 1, u64> neg_high; BitField<56, 1, u64> neg_high;
BitField<30, 9, u64> high; BitField<30, 9, u64> high;
BitField<29, 1, u64> neg_low; BitField<29, 1, u64> neg_low;
BitField<20, 9, u64> low; BitField<20, 9, u64> low;
} const hset2{insn}; } const hset2{insn};
const u32 imm{static_cast<u32>(hset2.low << 6) | ((hset2.neg_low != 0 ? 1 : 0) << 15) | const u32 imm{static_cast<u32>(hset2.low << 6) | ((hset2.neg_low != 0 ? 1 : 0) << 15) |
static_cast<u32>(hset2.high << 22) | ((hset2.neg_high != 0 ? 1 : 0) << 31)}; static_cast<u32>(hset2.high << 22) | ((hset2.neg_high != 0 ? 1 : 0) << 31)};
HSET2(*this, insn, ir.Imm32(imm), hset2.bf != 0, hset2.ftz != 0, false, false, HSET2(*this, insn, ir.Imm32(imm), hset2.bf != 0, hset2.ftz != 0, false, false, hset2.compare_op,
hset2.compare_op, Swizzle::H1_H0); Swizzle::H1_H0);
} }
} // namespace Shader::Maxwell } // namespace Shader::Maxwell

20
src/shader_recompiler/frontend/maxwell/translate/impl/half_floating_point_set_predicate.cpp
View file

@ -24,17 +24,17 @@ void HSETP2(TranslatorVisitor& v, u64 insn, const IR::U32& src_b, bool neg_b, bo
auto [lhs_a, rhs_a]{Extract(v.ir, v.X(hsetp2.src_a_reg), hsetp2.swizzle_a)}; auto [lhs_a, rhs_a]{Extract(v.ir, v.X(hsetp2.src_a_reg), hsetp2.swizzle_a)};
auto [lhs_b, rhs_b]{Extract(v.ir, src_b, swizzle_b)}; auto [lhs_b, rhs_b]{Extract(v.ir, src_b, swizzle_b)};
// TODO: Implement FP16 FloatingPointCompare
// if (lhs_a.Type() != lhs_b.Type()) { if (lhs_a.Type() != lhs_b.Type()) {
if (lhs_a.Type() == IR::Type::F16) { if (lhs_a.Type() == IR::Type::F16) {
lhs_a = v.ir.FPConvert(32, lhs_a); lhs_a = v.ir.FPConvert(32, lhs_a);
rhs_a = v.ir.FPConvert(32, rhs_a); rhs_a = v.ir.FPConvert(32, rhs_a);
}
if (lhs_b.Type() == IR::Type::F16) {
lhs_b = v.ir.FPConvert(32, lhs_b);
rhs_b = v.ir.FPConvert(32, rhs_b);
}
} }
if (lhs_b.Type() == IR::Type::F16) {
lhs_b = v.ir.FPConvert(32, lhs_b);
rhs_b = v.ir.FPConvert(32, rhs_b);
}
//}
lhs_a = v.ir.FPAbsNeg(lhs_a, hsetp2.abs_a != 0, hsetp2.neg_a != 0); lhs_a = v.ir.FPAbsNeg(lhs_a, hsetp2.abs_a != 0, hsetp2.neg_a != 0);
rhs_a = v.ir.FPAbsNeg(rhs_a, hsetp2.abs_a != 0, hsetp2.neg_a != 0); rhs_a = v.ir.FPAbsNeg(rhs_a, hsetp2.abs_a != 0, hsetp2.neg_a != 0);

24
src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp
View file

@ -50,6 +50,30 @@ IR::Opcode Replace(IR::Opcode op) {
return IR::Opcode::CompositeInsertF32x3; return IR::Opcode::CompositeInsertF32x3;
case IR::Opcode::CompositeInsertF16x4: case IR::Opcode::CompositeInsertF16x4:
return IR::Opcode::CompositeInsertF32x4; return IR::Opcode::CompositeInsertF32x4;
case IR::Opcode::FPOrdEqual16:
return IR::Opcode::FPOrdEqual32;
case IR::Opcode::FPUnordEqual16:
return IR::Opcode::FPUnordEqual32;
case IR::Opcode::FPOrdNotEqual16:
return IR::Opcode::FPOrdNotEqual32;
case IR::Opcode::FPUnordNotEqual16:
return IR::Opcode::FPUnordNotEqual32;
case IR::Opcode::FPOrdLessThan16:
return IR::Opcode::FPOrdLessThan32;
case IR::Opcode::FPUnordLessThan16:
return IR::Opcode::FPUnordLessThan32;
case IR::Opcode::FPOrdGreaterThan16:
return IR::Opcode::FPOrdGreaterThan32;
case IR::Opcode::FPUnordGreaterThan16:
return IR::Opcode::FPUnordGreaterThan32;
case IR::Opcode::FPOrdLessThanEqual16:
return IR::Opcode::FPOrdLessThanEqual32;
case IR::Opcode::FPUnordLessThanEqual16:
return IR::Opcode::FPUnordLessThanEqual32;
case IR::Opcode::FPOrdGreaterThanEqual16:
return IR::Opcode::FPOrdGreaterThanEqual32;
case IR::Opcode::FPUnordGreaterThanEqual16:
return IR::Opcode::FPUnordGreaterThanEqual32;
case IR::Opcode::ConvertS16F16: case IR::Opcode::ConvertS16F16:
return IR::Opcode::ConvertS16F32; return IR::Opcode::ConvertS16F32;
case IR::Opcode::ConvertS32F16: case IR::Opcode::ConvertS32F16: