raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-28 05:46:56 +01:00
Code Issues Packages Projects Releases Wiki Activity

shader: accept byte-selected CNDMASK SDWA operands

Browse Source
This commit is contained in:
nmzik
2026-08-21 01:49:03 +02:00
parent 770bdb9571
commit 61dd8a3031
2 changed files with 19 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/frontend/decode/VectorAluOps.cpp
+1 -4
View File
@@ -11,7 +11,6 @@ using Detail::OpcodeMap;
// architectural SDWA legality classes. // architectural SDWA legality classes.
enum class Vop2SdwaProfile { enum class Vop2SdwaProfile {
None, None,
Cndmask,
Float32, Float32,
Float16, Float16,
IntegerFullDestination, IntegerFullDestination,
@@ -29,7 +28,7 @@ struct Vop2OpcodeInfo {
}; };
constexpr Vop2OpcodeInfo VOP2_OPCODE_LIST[] = { constexpr Vop2OpcodeInfo VOP2_OPCODE_LIST[] = {
{0x01u, Opcode::V_CNDMASK_B32, Vop2SdwaProfile::Cndmask}, {0x01u, Opcode::V_CNDMASK_B32, Vop2SdwaProfile::IntegerPartialDestination},
{0x02u, Opcode::V_DOT2C_F32_F16}, {0x02u, Opcode::V_DOT2C_F32_F16},
{0x03u, Opcode::V_ADD_F32, Vop2SdwaProfile::Float32}, {0x03u, Opcode::V_ADD_F32, Vop2SdwaProfile::Float32},
{0x04u, Opcode::V_SUB_F32, Vop2SdwaProfile::Float32}, {0x04u, Opcode::V_SUB_F32, Vop2SdwaProfile::Float32},
@@ -929,8 +928,6 @@ constexpr uint32_t SdwaSelAll() {
constexpr Vop2SdwaRule VOP2_SDWA_RULES[] = { constexpr Vop2SdwaRule VOP2_SDWA_RULES[] = {
{}, {},
{SdwaSelWords() | SdwaSelFull(), SdwaSelWords() | SdwaSelFull(), SdwaSelWords() | SdwaSelFull(),
true, false},
{SdwaSelFull(), SdwaSelFull(), SdwaSelFull(), false, true}, {SdwaSelFull(), SdwaSelFull(), SdwaSelFull(), false, true},
{SdwaSelWords() | SdwaSelFull(), SdwaSelWords() | SdwaSelFull(), SdwaSelWords() | SdwaSelFull(), {SdwaSelWords() | SdwaSelFull(), SdwaSelWords() | SdwaSelFull(), SdwaSelWords() | SdwaSelFull(),
true, true}, true, true},
tests/ShaderRecompilerComputeTests.cpp
+18 -15
View File
@@ -14684,29 +14684,32 @@ TestCase Vop2SdwaCndmaskSourceModifier() {
O::BUFFER_STORE_DWORD, O::S_ENDPGM}}; O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
} }
TestCase Vop2SdwaCndmaskFullDestinationWithSubDwordSource() { TestCase Vop2SdwaCndmaskCapturedByte3Source() {
using O = ShaderOpcode; using O = ShaderOpcode;
std::vector<u32> code; std::vector<u32> code;
AppendVMovLiteral(&code, 0, 0xabcd1234u); AppendVMovLiteral(&code, 0, 0x55667788u);
AppendVMovLiteral(&code, 1, 0x55667788u); AppendBufferLoadDword(&code, 3, 30);
code.push_back(EncodeVopc(0xc0, Vgpr(0), 0)); // v_cmp_f_u32 code.push_back(EncodeVopc(0xc0, Vgpr(0), 0)); // v_cmp_f_u32
code.push_back(0x020e02f9u); code.push_back(0x023a00f9u);
code.push_back(0x06040600u); code.push_back(0x06030603u);
AppendStoreVgpr(&code, 7, 0); AppendStoreVgpr(&code, 29, 1);
code.push_back(EncodeVopc(0xc7, Vgpr(0), 0)); // v_cmp_t_u32 code.push_back(EncodeVopc(0xc7, Vgpr(0), 0)); // v_cmp_t_u32
code.push_back(0x020e02f9u); code.push_back(0x023a00f9u);
code.push_back(0x06040600u); code.push_back(0x06030603u);
AppendStoreVgpr(&code, 7, 1); AppendStoreVgpr(&code, 29, 2);
AppendEnd(&code); AppendEnd(&code);
TestCase test; TestCase test;
test.name = "Vop2SdwaCndmaskFullDestinationWithSubDwordSource"; test.name = "Vop2SdwaCndmaskCapturedByte3Source";
test.code = code; test.code = code;
test.expected = {0x00001234u, 0x55667788u}; test.initial = {0xabcd1234u, 0, 0};
test.opcodes = {O::V_MOV_B32, O::V_CMP_F_U32, O::V_CMP_T_U32, test.expected = {0xabcd1234u, 0x000000abu, 0x55667788u};
O::V_CNDMASK_B32, O::BUFFER_STORE_DWORD, O::S_ENDPGM}; test.opcodes = {O::V_MOV_B32, O::BUFFER_LOAD_DWORD, O::V_CMP_F_U32,
test.required_spirv = {"OpSelect"}; O::V_CMP_T_U32, O::V_CNDMASK_B32, O::BUFFER_STORE_DWORD,
O::S_ENDPGM};
test.decoded_counts = {{"V_CNDMASK_B32 v29, v3.sdwa(sel=3,sext=0), v0", 2}};
test.required_spirv = {"OpBitFieldUExtract", "OpSelect"};
return test; return test;
} }
@@ -19291,7 +19294,7 @@ std::vector<TestCase> MakeCases() {
AddCase(VectorCompareClassF32); AddCase(VectorCompareClassF32);
AddCase(VectorCompareF16Ops); AddCase(VectorCompareF16Ops);
AddCase(Vop2SdwaCndmaskSourceModifier); AddCase(Vop2SdwaCndmaskSourceModifier);
AddCase(Vop2SdwaCndmaskFullDestinationWithSubDwordSource); AddCase(Vop2SdwaCndmaskCapturedByte3Source);
AddCase(Vop3CndmaskUsesSgprMaskLaneBits); AddCase(Vop3CndmaskUsesSgprMaskLaneBits);
AddCase(Vop3CndmaskAllowsDataSourceModifier); AddCase(Vop3CndmaskAllowsDataSourceModifier);
AddCase(VectorCompareExecOps); AddCase(VectorCompareExecOps);