H3cJP/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-08-19 06:52:36 +00:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

shader: implement V_CNDMASK_B32

Browse Source
This commit is contained in:
nmzik
2026-08-05 02:59:28 +02:00
parent aa88d94755
commit 73da43a7be
3 changed files with 35 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/decompiler/VectorAluOps.cpp
+1 -1
View File
@@ -805,7 +805,7 @@ constexpr uint32_t SdwaSelAll() {
}
constexpr Vop2SdwaRule VOP2_SDWA_RULES[] = {
{Opcode::VCndmaskB32, SdwaSelWords(), SdwaSelWords() | SdwaSelFull(),
{Opcode::VCndmaskB32, SdwaSelWords() | SdwaSelFull(), SdwaSelWords() | SdwaSelFull(),
SdwaSelWords() | SdwaSelFull(), true, false},
{Opcode::VAddF32, SdwaSelFull(), SdwaSelFull(), SdwaSelFull(), false, true},
{Opcode::VSubF32, SdwaSelFull(), SdwaSelFull(), SdwaSelFull(), false, true},
tests/ShaderRecompilerComputeTests.cpp
+27
View File
@@ -11783,6 +11783,32 @@ TestCase Vop2SdwaCndmaskSourceModifier() {
{O::VMovB32, O::VCmpTU32, O::VCndmaskB32, O::BufferStoreDword, O::SEndpgm}};
}
TestCase Vop2SdwaCndmaskFullDestinationWithSubDwordSource() {
using O = ShaderOpcode;
std::vector<u32> code;
AppendVMovLiteral(&code, 0, 0xabcd1234u);
AppendVMovLiteral(&code, 1, 0x55667788u);
code.push_back(EncodeVopc(0xc0, Vgpr(0), 0)); // v_cmp_f_u32
code.push_back(0x020e02f9u);
code.push_back(0x06040600u);
AppendStoreVgpr(&code, 7, 0);
code.push_back(EncodeVopc(0xc7, Vgpr(0), 0)); // v_cmp_t_u32
code.push_back(0x020e02f9u);
code.push_back(0x06040600u);
AppendStoreVgpr(&code, 7, 1);
AppendEnd(&code);
TestCase test;
test.name = "Vop2SdwaCndmaskFullDestinationWithSubDwordSource";
test.code = code;
test.expected = {0x00001234u, 0x55667788u};
test.opcodes = {O::VMovB32, O::VCmpFU32, O::VCmpTU32, O::VCndmaskB32,
O::BufferStoreDword, O::SEndpgm};
test.required_spirv = {"OpBitFieldUExtract", "OpSelect"};
return test;
}
TestCase Vop3CndmaskUsesSgprMaskLaneBits() {
using O = ShaderOpcode;
@@ -15142,6 +15168,7 @@ std::vector<TestCase> MakeCases() {
AddCase(VectorCompareClassF32);
AddCase(VectorCompareF16Ops);
AddCase(Vop2SdwaCndmaskSourceModifier);
AddCase(Vop2SdwaCndmaskFullDestinationWithSubDwordSource);
AddCase(Vop3CndmaskUsesSgprMaskLaneBits);
AddCase(Vop3CndmaskAllowsDataSourceModifier);
AddCase(VectorCompareExecOps);
tests/shaderCfgTests.cpp
+7
View File
@@ -1170,6 +1170,8 @@ void TestNewShaderRecompilerMoreAluFamilies() {
EncodeVop2Sdwa(5, 6, 0, 4, 5),
0x025e6af9u,
0x16060635u, // v_cndmask_b32 v47, v53, -v53
0x020e02f9u,
0x06040600u, // v_cndmask_b32 v7, v0.lo, v1; full-width destination
0x100490f9u,
0x86860600u, // v_mul_f32 v2, s0, s72 (SDWA full)
EncodeVop2(0x35, 9, 249, 1),
@@ -1396,6 +1398,9 @@ void TestNewShaderRecompilerMoreAluFamilies() {
Check(Common::ContainsStr(result.decoded_dump, "v_cndmask_b32 v47, v53,") &&
Common::ContainsStr(result.decoded_dump, "v53.neg"),
"new decoder did not decode V_CNDMASK_B32 SDWA source modifier");
Check(Common::ContainsStr(result.decoded_dump, "v_cndmask_b32 v7, v0.sdwa(sel=4") &&
Common::ContainsStr(result.decoded_dump, "v1"),
"new decoder did not decode full-destination V_CNDMASK_B32 with SDWA source");
Check(Common::ContainsStr(result.decoded_dump, "v_add_f32 v125, v5.dpp"),
"new decoder did not decode VOP2 DPP source metadata");
Check(!Common::ContainsStr(result.decoded_dump, "VOP2 SDWA/DPP modifiers are not implemented"),
@@ -1700,6 +1705,8 @@ void TestNewShaderRecompilerMoreAluFamilies() {
Check(Common::ContainsStr(result.ir_dump, "SelectMaskF32Bits v47, vcc_lo, v53.neg, v53") &&
Common::ContainsStr(result.ir_dump, "v53.neg"),
"V_CNDMASK_B32 SDWA source modifier did not lower to float-bit select IR");
Check(Common::ContainsStr(result.ir_dump, "SelectMaskU32 v7, vcc_lo, v1, v0.sdwa(sel=4"),
"full-destination V_CNDMASK_B32 with SDWA source did not lower to integer select IR");
Check(Common::ContainsStr(result.ir_dump, "FAddF32 v125.dpp") &&
Common::ContainsStr(result.ir_dump, "v5.dpp"),
"VOP2 DPP source/destination did not lower to IR metadata");