raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 19:07:01 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: implement RDNA2 FP64 equality comparisons

Browse Source 
Co-authored-by: Senaxx <michelbeld@gmail.com>
Co-authored-by: Ali Almohaya <AliAlmohaya@gmail.com>
This commit is contained in:
nmzik
2026-10-01 09:48:26 +04:00
parent 64dcbe0ea1
commit 581dad7dd3
8 changed files with 125 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h
+1
View File
@@ -133,6 +133,7 @@ uint32_t EmitFloatCompare32(ValueEmitContext& ctx, const IR::Inst& inst) {
return EmitNative<opcode, IR::Type::U1>(ctx.state, operand(0), operand(1));
}
inline constexpr auto EmitFPOrdEqual32 = EmitFloatCompare32<spv::OpFOrdEqual>;
EMIT_NATIVE(FPOrdEqual64, OpFOrdEqual, U1, uint32_t, uint32_t)
inline constexpr auto EmitFPUnordEqual32 = EmitFloatCompare32<spv::OpFUnordEqual>;
inline constexpr auto EmitFPOrdNotEqual32 = EmitFloatCompare32<spv::OpFOrdNotEqual>;
inline constexpr auto EmitFPUnordNotEqual32 = EmitFloatCompare32<spv::OpFUnordNotEqual>;
src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.h
+1
View File
@@ -355,6 +355,7 @@ enum class Opcode {
V_CMP_NEQ_F32,
V_CMP_NLT_F32,
V_CMP_TRU_F32,
V_CMP_EQ_F64,
V_CMPX_LT_F32,
V_CMPX_EQ_F32,
V_CMPX_LE_F32,
src/graphics/shader/recompiler/frontend/decode/VectorAluOps.cpp
+3 -1
View File
@@ -228,6 +228,7 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
{0x19u, Opcode::V_CMPX_NGE_F32}, {0x1au, Opcode::V_CMPX_NLG_F32},
{0x1bu, Opcode::V_CMPX_NGT_F32}, {0x1cu, Opcode::V_CMPX_NLE_F32},
{0x1du, Opcode::V_CMPX_NEQ_F32}, {0x1eu, Opcode::V_CMPX_NLT_F32},
{0x22u, Opcode::V_CMP_EQ_F64, false},
{0x80u, Opcode::V_CMP_F_I32}, {0x81u, Opcode::V_CMP_LT_I32},
{0x82u, Opcode::V_CMP_EQ_I32}, {0x83u, Opcode::V_CMP_LE_I32},
{0x84u, Opcode::V_CMP_GT_I32}, {0x85u, Opcode::V_CMP_NE_I32},
@@ -829,6 +830,7 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
case Opcode::V_CMP_F_F32:
case Opcode::V_CMP_LT_F32:
case Opcode::V_CMP_EQ_F32:
case Opcode::V_CMP_EQ_F64:
case Opcode::V_CMP_LE_F32:
case Opcode::V_CMP_GT_F32:
case Opcode::V_CMP_LG_F32:
@@ -1139,7 +1141,7 @@ VopcSdwaFields DecodeVopcSdwaFields(uint32_t modifier) {
}
bool SupportsVopcSdwa(Opcode opcode) {
return opcode != Opcode::UNSUPPORTED;
return opcode != Opcode::UNSUPPORTED && opcode != Opcode::V_CMP_EQ_F64;
}
void DecodeVopcSdwa(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
src/graphics/shader/recompiler/frontend/translate/Compare.cpp
+5 -3
View File
@@ -40,11 +40,13 @@ void Translator::EmitInteger16Compare(const Decoder::Instruction& inst, IR::Valu
void Translator::EmitFloatCompare(const Decoder::Instruction& inst, IR::ValueOpcode opcode,
bool half, bool cmpx) {
const auto type = IR::ArgTypeOf(opcode, 0);
const auto lhs =
half ? IR::Value(ReadF16AsF32(inst.src0)) : ReadOperand(inst.src0, IR::Type::F32);
half ? IR::Value(ReadF16AsF32(inst.src0)) : ReadOperand(inst.src0, type);
const auto rhs =
half ? IR::Value(ReadF16AsF32(inst.src1)) : ReadOperand(inst.src1, IR::Type::F32);
const IR::FPCompareFlags flags{.flush_input_denorms = !half && flush_f32_inputs};
half ? IR::Value(ReadF16AsF32(inst.src1)) : ReadOperand(inst.src1, type);
const IR::FPCompareFlags flags{
.flush_input_denorms = !half && type == IR::Type::F32 && flush_f32_inputs};
EmitCompareResult(inst, IR::U1(ir.Emit(opcode, {lhs, rhs}, flags)), false, cmpx);
}
src/graphics/shader/recompiler/frontend/translate/Translate.cpp
+4
View File
@@ -889,6 +889,10 @@ void IncludeInstructionVectorRegisters(const Decoder::Instruction& inst, uint32_
case Decoder::Opcode::V_MUL_F64: include_vector(inst.src1, 2u); [[fallthrough]];
case Decoder::Opcode::V_RCP_F64: include_vector(inst.dst, 2u); [[fallthrough]];
case Decoder::Opcode::V_CVT_F32_F64: include_vector(inst.src0, 2u); break;
case Decoder::Opcode::V_CMP_EQ_F64:
include_vector(inst.src0, 2u);
include_vector(inst.src1, 2u);
break;
default: break;
}
if (inst.family == Decoder::Family::DS) {
src/graphics/shader/recompiler/frontend/translate/Vector.cpp
+3
View File
@@ -168,6 +168,9 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
case O::V_CMP_EQ_F32:
EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false, false);
return;
case O::V_CMP_EQ_F64:
EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual64, false, false);
return;
case O::V_CMPX_EQ_F32:
EmitFloatCompare(inst, IR::ValueOpcode::FPOrdEqual32, false, true);
return;
src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc
+1
View File
@@ -155,6 +155,7 @@ VALUE_OPCODE(LogicalAnd, U1, U1, U1)
VALUE_OPCODE(LogicalXor, U1, U1, U1)
VALUE_OPCODE(LogicalNot, U1, U1)
VALUE_OPCODE(FPOrdEqual32, U1, F32, F32)
VALUE_OPCODE(FPOrdEqual64, U1, F64, F64)
VALUE_OPCODE(FPUnordEqual32, U1, F32, F32)
VALUE_OPCODE(FPOrdNotEqual32, U1, F32, F32)
VALUE_OPCODE(FPUnordNotEqual32, U1, F32, F32)