raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-25 12:26:51 +01:00
Code Issues Packages Projects Releases Wiki Activity

shader: decode S_FLBIT_I32 and S_FLBIT_I32_I64 with a 4-op signed leading-bit search

Browse Source
This commit is contained in:
nmzik
2026-10-23 13:22:50 +02:00
parent 3688412dca
commit ecc30a1f2d
6 changed files with 71 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+39
View File
@@ -23119,6 +23119,43 @@ TestCase ScalarFlbitI32B32Gpu() {
O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
}
TestCase ScalarFlbitSignedGpu() {
using O = ShaderOpcode;
// RDNA2's S_FLBIT_I32 examples; 64-bit signs come from the high word.
constexpr u32 values32[] = {0x00000000u, 0x0000ccccu, 0xffff3333u,
0x7fffffffu, 0x80000000u, 0xffffffffu};
constexpr uint64_t values64[] = {0x0000000000000000ull, 0xffffffffffffffffull,
0x0000000100000000ull, 0xffffffff7fffffffull,
0xc000000000000000ull, 0x0000000000008000ull};
std::vector<u32> code;
u32 out = 0;
for (const auto value : values32) {
AppendSMovLiteral(&code, 0, value);
code.push_back(EncodeSop1(0x17, 20, 0));
AppendStoreSgpr(&code, 20, out++);
}
for (const auto value : values64) {
AppendSMovLiteral(&code, 0, static_cast<u32>(value));
AppendSMovLiteral(&code, 1, static_cast<u32>(value >> 32u));
code.push_back(EncodeSop1(0x18, 20, 0));
AppendStoreSgpr(&code, 20, out++);
}
// A signed 64-bit source sign-extends a literal.
code.push_back(EncodeSop1(0x18, 20, 255));
code.push_back(0x80000000u);
AppendStoreSgpr(&code, 20, out);
AppendEnd(&code);
return {"ScalarFlbitSignedGpu",
code,
{},
{0xffffffffu, 16, 16, 1, 1, 0xffffffffu, 0xffffffffu, 0xffffffffu,
31, 32, 2, 48, 33},
{O::S_MOV_B32, O::S_FLBIT_I32, O::S_FLBIT_I32_I64, O::V_MOV_B32,
O::BUFFER_STORE_DWORD, O::S_ENDPGM}};
}
TestCase ScalarSaveExecOps() {
using O = ShaderOpcode;
@@ -43766,6 +43803,7 @@ std::vector<TestCase> MakeCases() {
AddCase(ScalarQuadmaskB64);
AddCase(ScalarFlbitI32B64Gpu);
AddCase(ScalarFlbitI32B32Gpu);
AddCase(ScalarFlbitSignedGpu);
AddCase(ScalarSaveExecOps);
AddCase(ScalarOrn2SaveexecUsesSourceOrNotExec);
cases.push_back(ScalarOrn2SaveexecB32(32, 32));
@@ -50070,6 +50108,7 @@ int main(int argc, char **argv) {
RunCase(&vulkan, VectorCompareF16Ops());
RunCase(&vulkan, ScalarBrevB64OperandsAndMasks(32));
RunCase(&vulkan, ScalarBrevB64OperandsAndMasks(64));
RunCase(&vulkan, ScalarFlbitSignedGpu());
for (bool add : {false, true}) {
RunCase(&vulkan, DsAtomic64Bounds(add));
for (u32 wave_size : {32, 64}) {