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: DPP8 moves

Browse Source 
Co-authored-by: foufouadi <attar.foued@gmail.com>
Co-authored-by: fxpw <fxpw@outlook.com>
This commit is contained in:
nmzik
2026-09-18 00:11:16 +04:00
parent 165234b159
commit bfc8cd0e0b
10 changed files with 127 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+68
View File
@@ -19697,6 +19697,62 @@ TestCase VectorPermlane16FetchInactiveFi() {
return test;
}
TestCase VectorDpp8Captured(bool masked_exec) {
using O = ShaderOpcode;
constexpr u32 sentinel = 0xaaaaaaaau;
const u32 masks[] = {masked_exec ? 0x0f0f0f0fu : 0xffffffffu,
masked_exec ? 0xf0f0f0f0u : 0xffffffffu};
std::vector<u32> code;
AppendVMovU32(&code, 5, 100);
code.push_back(EncodeVop2(0x25, 5, Vgpr(0), 5));
AppendVMovLiteral(&code, 2, sentinel);
AppendVMovLiteral(&code, 3, sentinel);
AppendSMovLiteral(&code, 126, masks[0]);
AppendSMovLiteral(&code, 127, masks[1]);
// Captured DPP8 selectors broadcast each group's upper/lower four lanes.
code.push_back(EncodeVop1(0x01, 2, 233));
code.push_back(0xfacfac05u);
code.push_back(EncodeVop1(0x01, 3, 233));
code.push_back(0x68868805u);
AppendSMovLiteral(&code, 126, 0xffffffffu);
AppendSMovLiteral(&code, 127, 0xffffffffu);
code.push_back(EncodeVop2(0x1a, 6, InlineU32(2), 0));
AppendBufferStoreDword(&code, 2, 6);
AppendVMovU32(&code, 7, 64u * sizeof(u32));
code.push_back(EncodeVop2(0x25, 6, Vgpr(7), 6));
AppendBufferStoreDword(&code, 3, 6);
AppendEnd(&code);
TestCase test;
test.name = masked_exec ? "VectorDpp8CapturedMaskedExec"
: "VectorDpp8Captured";
test.code = std::move(code);
test.expected.resize(128, sentinel);
for (u32 lane = 0; lane < 64; ++lane) {
const auto mask = masks[lane / 32u];
if ((mask & (1u << (lane % 32u))) == 0) {
continue;
}
const u32 sources[] = {(lane & ~7u) | 4u | (lane & 3u),
(lane & ~7u) | (lane & 3u)};
for (u32 permutation = 0; permutation < 2; ++permutation) {
const auto source = sources[permutation];
test.expected[permutation * 64u + lane] =
(mask & (1u << (source % 32u))) != 0 ? 100u + source : 0u;
}
}
test.opcodes = {O::V_MOV_B32, O::V_ADD_NC_U32, O::S_MOV_B32,
O::V_LSHLREV_B32, O::BUFFER_STORE_DWORD, O::S_ENDPGM};
test.decoded_counts = {{".dpp8(", 2}};
test.compute_info.wave_size = 64;
test.compute_info.threads_num[0] = 64;
test.compute_info.threads_num[1] = 1;
test.compute_info.threads_num[2] = 1;
test.compute_info.thread_ids_num = 1;
test.has_compute_info = true;
return test;
}
TestCase VectorDppQuadPermuteReverse() {
using O = ShaderOpcode;
@@ -26871,6 +26927,8 @@ std::vector<TestCase> MakeCases() {
AddCase(VectorPermlanex16);
AddCase(VectorPermlane16FetchInactiveZero);
AddCase(VectorPermlane16FetchInactiveFi);
cases.push_back(VectorDpp8Captured(false));
cases.push_back(VectorDpp8Captured(true));
AddCase(VectorDppQuadPermuteReverse);
AddCase(VectorDppRowXmask);
AddCase(VectorDppBankMaskPreservesDestination);
@@ -31570,6 +31628,16 @@ int main(int argc, char **argv) {
std::setvbuf(stdout, nullptr, _IONBF, 0);
EnsureConfigInitialized();
CheckLeastRecentlyUsedCacheOrdering();
if (argc == 2 && std::strcmp(argv[1], "--dpp-only") == 0) {
VulkanHarness vulkan;
RunCase(&vulkan, VectorDpp8Captured(false));
RunCase(&vulkan, VectorDpp8Captured(true));
RunCase(&vulkan, VectorDppQuadPermuteReverse());
RunCase(&vulkan, VectorDppRowXmask());
RunCase(&vulkan, VectorDppBankMaskPreservesDestination());
RunCase(&vulkan, VectorDppBoundsControlZeroPreservesDestination());
return 0;
}
if (argc == 2 && std::strcmp(argv[1], "--buffer-snorm-store-only") == 0) {
VulkanHarness vulkan;
RunCase(&vulkan, BufferStoreFormatXyzwSnorm16CapturedSkinningVectors());