gremling/KytyPS5
1
0 Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-30 01:01:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

shader: initialize local vertex count from active lanes

Browse Source
This commit is contained in:
Çağatay Güçlü authored and nmzik committed 2026-10-29 05:52:04 +01:00
1 parent 143f640706
commit 63207739bc
2 files changed
+78 -7
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/graphics/shader/recompiler/frontend/translate/Translate.cpp
+10 -7
View File
@@ -1165,6 +1165,14 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
: IR::U32(IR::Value(0u));
entry_ir.SetExecLo(initial_lo);
entry_ir.SetExecHi(initial_hi);
const auto active_lane_count = [&] {
auto count = IR::U32(entry_ir.Emit(IR::ValueOpcode::BitCount32, {initial_lo}));
if (options.wave_size == 64u) {
count = entry_ir.IAdd(count, IR::U32(entry_ir.Emit(
IR::ValueOpcode::BitCount32, {initial_hi})));
}
return count;
};
if (options.stage == ShaderType::Compute) {
const auto* cs = options.input_info.compute;
const auto thread_ids =
@@ -1302,7 +1310,7 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
builtin(IR::StageInputKind::WorkgroupId, 1));
}
} else if (options.stage == ShaderType::Local) {
entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));
entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), active_lane_count());
entry_ir.SetVectorReg(static_cast<IR::VectorReg>(2),
builtin(IR::StageInputKind::VertexIndex));
entry_ir.SetVectorReg(static_cast<IR::VectorReg>(3), IR::U32(IR::Value(0u)));
@@ -1380,15 +1388,10 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
} else if (options.stage == ShaderType::Vertex) {
// Vulkan owns primitive assembly; each vertex subgroup is one NGG wave.
// ESVertCount describes input vertices, independently of lane ordinals.
auto vertex_count = IR::U32(entry_ir.Emit(IR::ValueOpcode::BitCount32, {initial_lo}));
if (options.wave_size == 64u) {
vertex_count = entry_ir.IAdd(vertex_count, IR::U32(entry_ir.Emit(
IR::ValueOpcode::BitCount32, {initial_hi})));
}
entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(2),
IR::U32(IR::Value(options.wave_size << 12u)));
entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3),
entry_ir.BitwiseOr(vertex_count, IR::U32(IR::Value(1u << 28u))));
entry_ir.BitwiseOr(active_lane_count(), IR::U32(IR::Value(1u << 28u))));
// Vulkan adds an indirect record's bases to the indices; the hardware passes them
// only in the user SGPRs the packet patches. SH_NOP leaves those SGPRs unchanged;
// it does not make the record's bases zero.