raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-24 04:46:56 +02:00
Code Issues Packages Projects Releases Wiki Activity

Decode unsigned byte sources for SDWA half-float conversion

Browse Source
This commit is contained in:
nmzik
2026-09-21 11:13:21 +04:00
parent d792dfa885
commit cfb8f7b246
3 changed files with 47 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/frontend/decode/VectorAluOps.cpp
+12 -10
View File
@@ -527,8 +527,8 @@ constexpr Vop1SdwaRule VOP1_SDWA_RULES[] = {
{Opcode::V_CVT_F32_F16, SdwaSelWords() | SdwaSelFull(), 0, 0, true},
{Opcode::V_CVT_F16_F32, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelWords(),
SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), true},
{Opcode::V_CVT_F16_U16, SdwaSelWords() | SdwaSelFull(), SdwaSelWords(),
SdwaSelWords() | SdwaSelFull(), false},
{Opcode::V_CVT_F16_U16, SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), SdwaSelWords(),
SdwaSelBytes() | SdwaSelWords() | SdwaSelFull(), false},
{Opcode::V_CVT_U16_F16, SdwaSelWords() | SdwaSelFull(), SdwaSelWords(),
SdwaSelWords() | SdwaSelFull(), true},
{Opcode::V_CVT_F16_I16, SdwaSelWords() | SdwaSelFull(), SdwaSelWords(),
@@ -618,13 +618,14 @@ bool SupportsVop1Clamp(Opcode opcode) {
}
bool ValidateVop1Sdwa(Instruction& inst, uint32_t opcode, uint32_t modifier) {
const auto dst_sel = (modifier >> 8u) & 0x7u;
const auto dst_u = (modifier >> 11u) & 0x3u;
const auto clamp = (modifier >> 13u) & 0x1u;
const auto omod = (modifier >> 14u) & 0x3u;
const auto src0_sel = (modifier >> 16u) & 0x7u;
const auto src0_neg = (modifier >> 20u) & 0x1u;
const auto src0_abs = (modifier >> 21u) & 0x1u;
const auto dst_sel = (modifier >> 8u) & 0x7u;
const auto dst_u = (modifier >> 11u) & 0x3u;
const auto clamp = (modifier >> 13u) & 0x1u;
const auto omod = (modifier >> 14u) & 0x3u;
const auto src0_sel = (modifier >> 16u) & 0x7u;
const auto src0_sext = (modifier >> 19u) & 0x1u;
const auto src0_neg = (modifier >> 20u) & 0x1u;
const auto src0_abs = (modifier >> 21u) & 0x1u;
if (src0_sel > 6u || dst_sel > 6u) {
SetUnsupported(inst, Family::VOP1, opcode, "VOP1 SDWA selector is invalid");
@@ -640,7 +641,8 @@ bool ValidateVop1Sdwa(Instruction& inst, uint32_t opcode, uint32_t modifier) {
"VOP1 SDWA destination selector is not supported");
return false;
}
if (!IsVop1SdwaSourceSupported(inst.opcode, src0_sel, src0_neg != 0u, src0_abs != 0u)) {
if (!IsVop1SdwaSourceSupported(inst.opcode, src0_sel, src0_neg != 0u, src0_abs != 0u) ||
(inst.opcode == Opcode::V_CVT_F16_U16 && src0_sel <= 3u && src0_sext != 0u)) {
SetUnsupported(inst, Family::VOP1, opcode, "VOP1 SDWA source selector is not supported");
return false;
}
tests/ShaderRecompilerComputeTests.cpp
+28
View File
@@ -19624,6 +19624,33 @@ TestCase VectorMinMaxF16Ops() {
O::S_ENDPGM}};
}
TestCase VectorCvtF16U16CapturedByteSource() {
using O = ShaderOpcode;
constexpr std::array inputs = {0xff123456u, 0x80123456u, 0x00123456u};
std::vector<u32> code;
AppendVMovLiteral(&code, 16, 0xa5a55a5au);
for (u32 i = 0; i < inputs.size(); i++) {
AppendVMovU32(&code, 30, i * 4u);
AppendBufferLoadDword(&code, 18, 30);
// Captured shader 0x15082e211725205f: unsigned byte3 -> FP16 word0,
// preserving word1 of v16.
code.insert(code.end(), {0x7e20a0f9u, 0x00031412u});
AppendStoreVgpr(&code, 16, i);
}
AppendEnd(&code);
TestCase test;
test.name = "VectorCvtF16U16CapturedByteSource";
test.code = std::move(code);
test.initial.assign(inputs.begin(), inputs.end());
test.expected = {0xa5a55bf8u, 0xa5a55800u, 0xa5a50000u};
test.opcodes = {O::V_MOV_B32, O::BUFFER_LOAD_DWORD, O::V_CVT_F16_U16,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
test.required_spirv = {"OpBitFieldUExtract", "OpConvertUToF"};
return test;
}
TestCase VectorCvtU16F16Sdwa() {
using O = ShaderOpcode;
@@ -27458,6 +27485,7 @@ std::vector<TestCase> MakeCases() {
AddCase(CvtPkrtzF16F32SdwaAndOutputModifiers);
AddCase(PackedMinMaxF16NanAndSignedZeroEdges);
AddCase(VectorMinMaxF16Ops);
AddCase(VectorCvtF16U16CapturedByteSource);
AddCase(VectorCvtU16F16Sdwa);
AddCase(NativeAndSdwa16BitDestinationWrites);
AddCase(VectorMinMaxMed3F16Ops);
tests/shaderCfgTests.cpp
+7
View File
@@ -3466,6 +3466,13 @@ void TestNewShaderRecompilerCapturedVop1SdwaByteConvert() {
Check(SpirvContainsOpcode(result.spirv, 112),
"captured SDWA byte conversion did not emit OpConvertUToF");
CheckSpirvBinaryValidates(result.spirv);
const uint32_t byte_sext[] = {0x7e20a0f9u, 0x000b1412u};
ShaderRecompiler::Decoder::Instruction decoded;
ShaderRecompiler::Decoder::DecodeInstruction(byte_sext, 0, decoded);
Check(decoded.word_count == 2u &&
decoded.opcode == ShaderRecompiler::Decoder::Opcode::UNSUPPORTED,
"V_CVT_F16_U16 accepted unimplemented SDWA byte sign extension");
}
void TestNewShaderRecompilerVop1SdwaNotDestination() {