raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: support returning 64-bit LDS max and compare-exchange

Browse Source
This commit is contained in:
nmzik
2026-10-13 14:48:38 +02:00
parent e998f21e71
commit 9e41d8320a
12 changed files with 188 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+89 -1
View File
@@ -20100,6 +20100,8 @@ CoverageClass ClassifyOpcode(ShaderOpcode opcode,
case Opcode::FLAT_STORE_DWORDX4:
case Opcode::DS_ADD_U32:
case Opcode::DS_ADD_U64:
case Opcode::DS_MAX_RTN_U64:
case Opcode::DS_CMPST_RTN_B64:
case Opcode::DS_ADD_RTN_U32:
case Opcode::DS_SUB_U32:
case Opcode::DS_SUB_RTN_U32:
@@ -33184,6 +33186,90 @@ TestCase DsAtomic64Bounds(bool add) {
return test;
}
TestCase DsAtomicReturn64() {
using O = ShaderOpcode;
TestCase test;
test.name = "DsAtomicReturn64";
auto &code = test.code;
const auto pair = [&](u32 reg, uint64_t value) {
AppendVMovLiteral(&code, reg, static_cast<u32>(value));
AppendVMovLiteral(&code, reg + 1u, static_cast<u32>(value >> 32u));
};
const auto result = [&](u32 reg, uint64_t value) {
AppendStoreVgpr(&code, reg, static_cast<u32>(test.expected.size()));
test.expected.push_back(static_cast<u32>(value));
AppendStoreVgpr(&code, reg + 1u, static_cast<u32>(test.expected.size()));
test.expected.push_back(static_cast<u32>(value >> 32u));
};
AppendVMovU32(&code, 7, 0);
AppendVMovLiteral(&code, 6, 0x13579bdfu);
code.push_back(EncodeDs0(0x0d, 24));
code.push_back(EncodeDs1(0, 6, 7));
struct Case { uint64_t before, value; u32 address, offset; };
constexpr std::array cases{
Case{0x7fffffffffffffffull, 0x8000000000000000ull, 0, 8},
Case{0x8000000000000000ull, 0x7fffffffffffffffull, 0, 8},
Case{0x8000000000000001ull, 0x8000000000000002ull, 0x10007, 2},
Case{0xffffffffffffffffull, 0, 0, 8}};
for (const auto &item : cases) {
pair(2, item.before);
code.push_back(EncodeDs0(0x4d, 8));
code.push_back(EncodeDs1(0, 2, 7));
pair(4, item.value);
AppendVMovLiteral(&code, 0, item.address);
// CS23e96bf5029f62fd pc 0xec8: d9a00008 00000400. VDst aliases ADDR.
code.push_back(EncodeDs0(0x68, item.offset));
code.push_back(0x00000400u);
result(0, item.before);
code.push_back(EncodeDs0(0x76, 8));
code.push_back(EncodeDs1(8, 0, 7));
result(8, std::max(item.before, item.value));
}
struct CompareCase { uint64_t before, compare, replacement; };
constexpr std::array compare_cases{
CompareCase{0xfedcba9876543210ull, 0xfedcba9876543210ull, 0x0123456789abcdefull},
CompareCase{0xfedcba9876543210ull, 0x0edcba9876543210ull, UINT64_MAX},
CompareCase{0xfedcba9876543210ull, 0xfedcba9876543211ull, UINT64_MAX},
CompareCase{UINT64_MAX, UINT64_MAX, 0}};
AppendVMovU32(&code, 63, 0);
for (const auto &item : compare_cases) {
pair(4, item.before);
code.push_back(EncodeDs0(0x4d, 16));
code.push_back(EncodeDs1(0, 4, 7));
pair(0, item.compare);
pair(2, item.replacement);
// pc 0x1048: DATA compares, DATA2 replaces, VDst aliases DATA2.
code.insert(code.end(), {0xd9c00010u, 0x0202003fu});
result(2, item.before);
code.push_back(EncodeDs0(0x76, 16));
code.push_back(EncodeDs1(8, 0, 7));
result(8, item.before == item.compare ? item.replacement : item.before);
}
pair(4, UINT64_MAX);
AppendVMovU32(&code, 0, 24); // One DWORD remains: ignore the entire atomic.
code.push_back(EncodeDs0(0x68));
code.push_back(EncodeDs1(2, 4, 0));
result(2, 0);
pair(2, 0xabcdef0112345678ull);
AppendVMovU32(&code, 0, 0);
code.push_back(EncodeSMovB32(126, InlineU32(0)));
code.insert(code.end(), {0xd9a00008u, 0x02000400u});
code.push_back(EncodeSMovB32(126, InlineU32(1)));
result(2, 0xabcdef0112345678ull); // Inactive lanes retain both destination DWORDs.
code.push_back(EncodeDs0(0x36, 24));
code.push_back(EncodeDs1(6, 0, 7));
AppendStoreVgpr(&code, 6, static_cast<u32>(test.expected.size()));
test.expected.push_back(0x13579bdfu);
AppendEnd(&code);
test.compute_info.lds_size_dwords = 7;
test.opcodes = {O::V_MOV_B32, O::S_MOV_B32, O::DS_WRITE_B32, O::DS_WRITE_B64,
O::DS_MAX_RTN_U64, O::DS_CMPST_RTN_B64, O::DS_READ_B32, O::DS_READ_B64,
O::BUFFER_STORE_DWORD, O::S_ENDPGM};
test.required_spirv = {"OpAtomicUMax %ulong", "OpAtomicCompareExchange %ulong", "WorkgroupMemoryExplicitLayoutKHR", "Aliased"};
test.forbidden_spirv = {"OpAtomicLoad", "OpLoopMerge"};
return test;
}
TestCase DsAtomic64Contention(bool add, u32 wave_size) {
using O = ShaderOpcode;
TestCase test;
@@ -39204,6 +39290,7 @@ std::vector<TestCase> MakeCases() {
cases.push_back(DsMaskedOrContention(wave_size));
}
AddCase(DsAtomicReturnVariants);
AddCase(DsAtomicReturn64);
for (bool add : {false, true}) {
cases.push_back(DsAtomic64Bounds(add));
for (u32 wave_size : {32, 64}) {
@@ -44656,7 +44743,7 @@ int main(int argc, char **argv) {
RunCase(&vulkan, DsWideLdsPartialBounds());
RunCase(&vulkan, DsAtomic64Bounds(false));
RunCase(&vulkan, DsAtomic64Bounds(true));
for (auto test : {DsWideLdsPartialBounds(), DsAtomic64Bounds(false), DsAtomic64Bounds(true)}) {
for (auto test : {DsWideLdsPartialBounds(), DsAtomic64Bounds(false), DsAtomic64Bounds(true), DsAtomicReturn64()}) {
test.compute_info.lds_storage = true;
std::erase(test.required_spirv, "WorkgroupMemoryExplicitLayoutKHR");
test.required_spirv.push_back("StorageBuffer");
@@ -44700,6 +44787,7 @@ int main(int argc, char **argv) {
RunCase(&vulkan, DsAtomic64Contention(add, wave_size));
}
}
RunCase(&vulkan, DsAtomicReturn64());
RunCase(&vulkan, BufferAtomicAndX2GlcAndExec());
RunCase(&vulkan, ImageAtomicSignedMinMax<false>());
RunCase(&vulkan, ImageAtomicSignedMinMax<true>());