shader: Implement I2I SAT
This commit is contained in:
parent
3c758d9b53
commit
dbc1e5cde7
|
@ -287,6 +287,8 @@ Id EmitSMin32(EmitContext& ctx, Id a, Id b);
|
||||||
Id EmitUMin32(EmitContext& ctx, Id a, Id b);
|
Id EmitUMin32(EmitContext& ctx, Id a, Id b);
|
||||||
Id EmitSMax32(EmitContext& ctx, Id a, Id b);
|
Id EmitSMax32(EmitContext& ctx, Id a, Id b);
|
||||||
Id EmitUMax32(EmitContext& ctx, Id a, Id b);
|
Id EmitUMax32(EmitContext& ctx, Id a, Id b);
|
||||||
|
Id EmitSClamp32(EmitContext& ctx, Id value, Id min, Id max);
|
||||||
|
Id EmitUClamp32(EmitContext& ctx, Id value, Id min, Id max);
|
||||||
Id EmitSLessThan(EmitContext& ctx, Id lhs, Id rhs);
|
Id EmitSLessThan(EmitContext& ctx, Id lhs, Id rhs);
|
||||||
Id EmitULessThan(EmitContext& ctx, Id lhs, Id rhs);
|
Id EmitULessThan(EmitContext& ctx, Id lhs, Id rhs);
|
||||||
Id EmitIEqual(EmitContext& ctx, Id lhs, Id rhs);
|
Id EmitIEqual(EmitContext& ctx, Id lhs, Id rhs);
|
||||||
|
|
|
@ -163,6 +163,14 @@ Id EmitUMax32(EmitContext& ctx, Id a, Id b) {
|
||||||
return ctx.OpUMax(ctx.U32[1], a, b);
|
return ctx.OpUMax(ctx.U32[1], a, b);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
Id EmitSClamp32(EmitContext& ctx, Id value, Id min, Id max) {
|
||||||
|
return ctx.OpSClamp(ctx.U32[1], value, min, max);
|
||||||
|
}
|
||||||
|
|
||||||
|
Id EmitUClamp32(EmitContext& ctx, Id value, Id min, Id max) {
|
||||||
|
return ctx.OpUClamp(ctx.U32[1], value, min, max);
|
||||||
|
}
|
||||||
|
|
||||||
Id EmitSLessThan(EmitContext& ctx, Id lhs, Id rhs) {
|
Id EmitSLessThan(EmitContext& ctx, Id lhs, Id rhs) {
|
||||||
return ctx.OpSLessThan(ctx.U1, lhs, rhs);
|
return ctx.OpSLessThan(ctx.U1, lhs, rhs);
|
||||||
}
|
}
|
||||||
|
|
|
@ -1183,6 +1183,14 @@ U32 IREmitter::IMax(const U32& a, const U32& b, bool is_signed) {
|
||||||
return is_signed ? SMax(a, b) : UMax(a, b);
|
return is_signed ? SMax(a, b) : UMax(a, b);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
U32 IREmitter::SClamp(const U32& value, const U32& min, const U32& max) {
|
||||||
|
return Inst<U32>(Opcode::SClamp32, value, min, max);
|
||||||
|
}
|
||||||
|
|
||||||
|
U32 IREmitter::UClamp(const U32& value, const U32& min, const U32& max) {
|
||||||
|
return Inst<U32>(Opcode::UClamp32, value, min, max);
|
||||||
|
}
|
||||||
|
|
||||||
U1 IREmitter::ILessThan(const U32& lhs, const U32& rhs, bool is_signed) {
|
U1 IREmitter::ILessThan(const U32& lhs, const U32& rhs, bool is_signed) {
|
||||||
return Inst<U1>(is_signed ? Opcode::SLessThan : Opcode::ULessThan, lhs, rhs);
|
return Inst<U1>(is_signed ? Opcode::SLessThan : Opcode::ULessThan, lhs, rhs);
|
||||||
}
|
}
|
||||||
|
|
|
@ -206,6 +206,8 @@ public:
|
||||||
[[nodiscard]] U32 SMax(const U32& a, const U32& b);
|
[[nodiscard]] U32 SMax(const U32& a, const U32& b);
|
||||||
[[nodiscard]] U32 UMax(const U32& a, const U32& b);
|
[[nodiscard]] U32 UMax(const U32& a, const U32& b);
|
||||||
[[nodiscard]] U32 IMax(const U32& a, const U32& b, bool is_signed);
|
[[nodiscard]] U32 IMax(const U32& a, const U32& b, bool is_signed);
|
||||||
|
[[nodiscard]] U32 SClamp(const U32& value, const U32& min, const U32& max);
|
||||||
|
[[nodiscard]] U32 UClamp(const U32& value, const U32& min, const U32& max);
|
||||||
|
|
||||||
[[nodiscard]] U1 ILessThan(const U32& lhs, const U32& rhs, bool is_signed);
|
[[nodiscard]] U1 ILessThan(const U32& lhs, const U32& rhs, bool is_signed);
|
||||||
[[nodiscard]] U1 IEqual(const U32U64& lhs, const U32U64& rhs);
|
[[nodiscard]] U1 IEqual(const U32U64& lhs, const U32U64& rhs);
|
||||||
|
|
|
@ -299,6 +299,8 @@ OPCODE(SMin32, U32, U32,
|
||||||
OPCODE(UMin32, U32, U32, U32, )
|
OPCODE(UMin32, U32, U32, U32, )
|
||||||
OPCODE(SMax32, U32, U32, U32, )
|
OPCODE(SMax32, U32, U32, U32, )
|
||||||
OPCODE(UMax32, U32, U32, U32, )
|
OPCODE(UMax32, U32, U32, U32, )
|
||||||
|
OPCODE(SClamp32, U32, U32, U32, U32, )
|
||||||
|
OPCODE(UClamp32, U32, U32, U32, U32, )
|
||||||
OPCODE(SLessThan, U1, U32, U32, )
|
OPCODE(SLessThan, U1, U32, U32, )
|
||||||
OPCODE(ULessThan, U1, U32, U32, )
|
OPCODE(ULessThan, U1, U32, U32, )
|
||||||
OPCODE(IEqual, U1, U32, U32, )
|
OPCODE(IEqual, U1, U32, U32, )
|
||||||
|
|
|
@ -30,16 +30,33 @@ enum class IntegerWidth : u64 {
|
||||||
[[nodiscard]] IR::U32 ConvertInteger(IR::IREmitter& ir, const IR::U32& src,
|
[[nodiscard]] IR::U32 ConvertInteger(IR::IREmitter& ir, const IR::U32& src,
|
||||||
IntegerWidth dst_width) {
|
IntegerWidth dst_width) {
|
||||||
const IR::U32 zero{ir.Imm32(0)};
|
const IR::U32 zero{ir.Imm32(0)};
|
||||||
|
const IR::U32 count{WidthSize(ir, dst_width)};
|
||||||
|
return ir.BitFieldExtract(src, zero, count, false);
|
||||||
|
}
|
||||||
|
|
||||||
|
[[nodiscard]] IR::U32 SaturateInteger(IR::IREmitter& ir, const IR::U32& src, IntegerWidth dst_width,
|
||||||
|
bool dst_signed, bool src_signed) {
|
||||||
|
IR::U32 min{};
|
||||||
|
IR::U32 max{};
|
||||||
|
const IR::U32 zero{ir.Imm32(0)};
|
||||||
switch (dst_width) {
|
switch (dst_width) {
|
||||||
case IntegerWidth::Byte:
|
case IntegerWidth::Byte:
|
||||||
return ir.BitFieldExtract(src, zero, ir.Imm32(8), false);
|
min = dst_signed && src_signed ? ir.Imm32(0xffffff80) : zero;
|
||||||
|
max = dst_signed ? ir.Imm32(0x7f) : ir.Imm32(0xff);
|
||||||
|
break;
|
||||||
case IntegerWidth::Short:
|
case IntegerWidth::Short:
|
||||||
return ir.BitFieldExtract(src, zero, ir.Imm32(16), false);
|
min = dst_signed && src_signed ? ir.Imm32(0xffff8000) : zero;
|
||||||
|
max = dst_signed ? ir.Imm32(0x7fff) : ir.Imm32(0xffff);
|
||||||
|
break;
|
||||||
case IntegerWidth::Word:
|
case IntegerWidth::Word:
|
||||||
return ir.BitFieldExtract(src, zero, ir.Imm32(32), false);
|
min = dst_signed && src_signed ? ir.Imm32(0x80000000) : zero;
|
||||||
|
max = dst_signed ? ir.Imm32(0x7fffffff) : ir.Imm32(0xffffffff);
|
||||||
|
break;
|
||||||
default:
|
default:
|
||||||
throw NotImplementedException("Invalid width {}", dst_width);
|
throw NotImplementedException("Invalid width {}", dst_width);
|
||||||
}
|
}
|
||||||
|
const IR::U32 value{!dst_signed && src_signed ? ir.SMax(zero, src) : src};
|
||||||
|
return dst_signed && src_signed ? ir.SClamp(value, min, max) : ir.UClamp(value, min, max);
|
||||||
}
|
}
|
||||||
|
|
||||||
void I2I(TranslatorVisitor& v, u64 insn, const IR::U32& src_a) {
|
void I2I(TranslatorVisitor& v, u64 insn, const IR::U32& src_a) {
|
||||||
|
@ -60,9 +77,6 @@ void I2I(TranslatorVisitor& v, u64 insn, const IR::U32& src_a) {
|
||||||
if (i2i.cc != 0) {
|
if (i2i.cc != 0) {
|
||||||
throw NotImplementedException("I2I CC");
|
throw NotImplementedException("I2I CC");
|
||||||
}
|
}
|
||||||
if (i2i.sat != 0) {
|
|
||||||
throw NotImplementedException("I2I SAT");
|
|
||||||
}
|
|
||||||
if (i2i.src_fmt == IntegerWidth::Short && (i2i.selector == 1 || i2i.selector == 3)) {
|
if (i2i.src_fmt == IntegerWidth::Short && (i2i.selector == 1 || i2i.selector == 3)) {
|
||||||
throw NotImplementedException("16-bit source format incompatible with selector {}",
|
throw NotImplementedException("16-bit source format incompatible with selector {}",
|
||||||
i2i.selector);
|
i2i.selector);
|
||||||
|
@ -75,15 +89,21 @@ void I2I(TranslatorVisitor& v, u64 insn, const IR::U32& src_a) {
|
||||||
const s32 selector{static_cast<s32>(i2i.selector)};
|
const s32 selector{static_cast<s32>(i2i.selector)};
|
||||||
const IR::U32 offset{v.ir.Imm32(selector * 8)};
|
const IR::U32 offset{v.ir.Imm32(selector * 8)};
|
||||||
const IR::U32 count{WidthSize(v.ir, i2i.src_fmt)};
|
const IR::U32 count{WidthSize(v.ir, i2i.src_fmt)};
|
||||||
IR::U32 src_values{v.ir.BitFieldExtract(src_a, offset, count, i2i.src_fmt_sign != 0)};
|
const bool src_signed{i2i.src_fmt_sign != 0};
|
||||||
if (i2i.abs) {
|
const bool dst_signed{i2i.dst_fmt_sign != 0};
|
||||||
|
const bool sat{i2i.sat != 0};
|
||||||
|
|
||||||
|
IR::U32 src_values{v.ir.BitFieldExtract(src_a, offset, count, src_signed)};
|
||||||
|
if (i2i.abs != 0) {
|
||||||
src_values = v.ir.IAbs(src_values);
|
src_values = v.ir.IAbs(src_values);
|
||||||
}
|
}
|
||||||
if (i2i.neg) {
|
if (i2i.neg != 0) {
|
||||||
src_values = v.ir.INeg(src_values);
|
src_values = v.ir.INeg(src_values);
|
||||||
}
|
}
|
||||||
|
const IR::U32 result{
|
||||||
|
sat ? SaturateInteger(v.ir, src_values, i2i.dst_fmt, dst_signed, src_signed)
|
||||||
|
: ConvertInteger(v.ir, src_values, i2i.dst_fmt)};
|
||||||
|
|
||||||
const IR::U32 result{ConvertInteger(v.ir, src_values, i2i.dst_fmt)};
|
|
||||||
v.X(i2i.dest_reg, result);
|
v.X(i2i.dest_reg, result);
|
||||||
}
|
}
|
||||||
} // Anonymous namespace
|
} // Anonymous namespace
|
||||||
|
|
Loading…
Reference in New Issue