raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 02:47:02 +02:00
Code Issues Packages Projects Releases Wiki Activity

shader: allow reverse left shifts to use shared SDWA partial writes

Browse Source
This commit is contained in:
nmzik
2026-10-14 05:20:23 +02:00
parent 26d326d1b0
commit 280b87adec
3 changed files with 71 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+50
View File
@@ -23215,6 +23215,52 @@ TestCase Vop2SdwaBitwiseByteDestinations(u32 wave_size) {
return test;
}
TestCase Vop2SdwaLshlrevByteDestinations(u32 wave_size) {
using O = ShaderOpcode;
constexpr u32 sentinel = 0xa1b2c3d4u;
TestCase test;
test.name = wave_size == 64 ? "Vop2SdwaLshlrevByteDestinations64" : "Vop2SdwaLshlrevByteDestinations32";
test.initial = {0x44668122u};
test.expected = test.initial;
auto &code = test.code;
const auto store = [&](u32 reg, u32 expected) {
AppendStoreVgpr(&code, reg, static_cast<u32>(test.expected.size()));
test.expected.push_back(expected);
};
AppendBufferLoadDword(&code, 4, 30);
for (const u32 count : {6u, 7u, 39u}) {
const u32 byte = count == 6u ? 0x40u : 0x80u;
for (u32 selector = 0; selector < 4u; ++selector) {
for (u32 unused = 0; unused < 3u; ++unused) {
AppendVMovLiteral(&code, 5, sentinel);
code.push_back(EncodeVop2(0x1a, 5, 249, 4));
// count7/byte1/preserve is captured 340a08f9 01861187 at PC4e98.
code.push_back(EncodeVop2Sdwa(InlineU32(count), selector, unused, 6, 1,
0, 0, 0, 0, 0, 0, 1));
const u32 shift = selector * 8u;
const u32 fill = unused == 1u && byte == 0x80u ? 0xffffff00u : 0u;
store(5, ((byte | fill) << shift) |
(unused == 2u ? sentinel & ~(0xffu << shift) : 0u));
}
}
}
code.insert(code.end(), {EncodeVop2(0x1a, 4, 249, 4), 0x01861187u});
store(4, 0x44668022u); // Source1 aliases the preserved destination.
AppendVMovLiteral(&code, 5, sentinel);
code.push_back(EncodeSMovB32(126, InlineU32(0)));
code.insert(code.end(), {0x340a08f9u, 0x01861187u});
code.push_back(EncodeSMovB32(126, InlineU32(1)));
store(5, sentinel);
AppendEnd(&code);
test.initial.resize(test.expected.size());
test.opcodes = {O::BUFFER_LOAD_DWORD, O::V_MOV_B32, O::S_MOV_B32,
O::V_LSHLREV_B32, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
test.required_spirv = {"OpShiftLeftLogical", "OpBitFieldInsert", "OpBitFieldSExtract"};
test.compute_info.wave_size = wave_size;
test.has_compute_info = true;
return test;
}
TestCase Vop1SdwaMovByteDestinations() {
using O = ShaderOpcode;
@@ -39393,6 +39439,8 @@ std::vector<TestCase> MakeCases() {
AddCase(Vop1SdwaNotPartialSourcesAndDestinations);
cases.push_back(Vop2SdwaBitwiseByteDestinations(32));
cases.push_back(Vop2SdwaBitwiseByteDestinations(64));
cases.push_back(Vop2SdwaLshlrevByteDestinations(32));
cases.push_back(Vop2SdwaLshlrevByteDestinations(64));
AddCase(Vop1SdwaMovByteDestinations);
cases.push_back(Vop1SdwaMovByteSources(32));
cases.push_back(Vop1SdwaMovByteSources(64));
@@ -45801,6 +45849,8 @@ int main(int argc, char **argv) {
VulkanHarness vulkan;
RunCase(&vulkan, Vop2SdwaBitwiseByteDestinations(32));
RunCase(&vulkan, Vop2SdwaBitwiseByteDestinations(64));
RunCase(&vulkan, Vop2SdwaLshlrevByteDestinations(32));
RunCase(&vulkan, Vop2SdwaLshlrevByteDestinations(64));
RunCase(&vulkan, Vop1SdwaNotPartialSourcesAndDestinations());
return 0;
}