raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-28 13:56:59 +01:00
Code Issues Packages Projects Releases Wiki Activity

graphics: draw patched-offset indirect draws from GPU-read records

Browse Source
This commit is contained in:
nmzik
2026-10-27 09:31:50 +01:00
parent 09753bf8ff
commit a5b3673a39
14 changed files with 118 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/shader/recompiler/backend/spirv/spirvEmitterFlow.cpp
+2 -1
View File
@@ -63,7 +63,8 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
}
if (kind == IR::StageInputKind::VertexIndex || kind == IR::StageInputKind::InstanceIndex ||
kind == IR::StageInputKind::InvocationId || kind == IR::StageInputKind::PrimitiveId ||
kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId) {
kind == IR::StageInputKind::Layer || kind == IR::StageInputKind::SampleId ||
kind == IR::StageInputKind::BaseVertex || kind == IR::StageInputKind::BaseInstance) {
const auto value = state.builder.AllocateId();
const auto bits = state.builder.AllocateId();
state.builder.AddFunction(spv::OpLoad, TypeI32(state), value, variable);
src/graphics/shader/recompiler/backend/spirv/spirvEmitterModule.cpp
+9 -1
View File
@@ -476,6 +476,8 @@ uint32_t BuiltInForInput(IR::StageInputKind kind) {
case IR::StageInputKind::LocalInvocationId: return spv::BuiltInLocalInvocationId;
case IR::StageInputKind::LocalInvocationIndex: return spv::BuiltInLocalInvocationIndex;
case IR::StageInputKind::GlobalInvocationId: return spv::BuiltInGlobalInvocationId;
case IR::StageInputKind::BaseVertex: return spv::BuiltInBaseVertex;
case IR::StageInputKind::BaseInstance: return spv::BuiltInBaseInstance;
default: return UINT32_MAX;
}
}
@@ -526,7 +528,9 @@ void DefineInputs(EmitterState& state) {
case IR::StageInputKind::PrimitiveId:
case IR::StageInputKind::InstanceIndex:
case IR::StageInputKind::Layer:
case IR::StageInputKind::SampleId: type = TypeI32(state); break;
case IR::StageInputKind::SampleId:
case IR::StageInputKind::BaseVertex:
case IR::StageInputKind::BaseInstance: type = TypeI32(state); break;
case IR::StageInputKind::WorkgroupId:
case IR::StageInputKind::NumWorkgroups:
case IR::StageInputKind::LocalInvocationId:
@@ -802,6 +806,10 @@ void DefineModule(EmitterState& state) {
if (InputVariableForKind(state, IR::StageInputKind::SampleId) != 0) {
state.builder.RequireCapability(spv::CapabilitySampleRateShading);
}
if (InputVariableForKind(state, IR::StageInputKind::BaseVertex) != 0 ||
InputVariableForKind(state, IR::StageInputKind::BaseInstance) != 0) {
state.builder.RequireCapability(spv::CapabilityDrawParameters);
}
if (state.program.info.image_gather_extended) {
state.builder.RequireCapability(spv::CapabilityImageGatherExtended);
}
src/graphics/shader/recompiler/frontend/translate/Translate.cpp
+19 -2
View File
@@ -1389,10 +1389,27 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
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))));
// Vulkan adds an indirect record's bases to the indices; the hardware passes them
// only in the user SGPRs the packet patches.
const auto& offsets = options.input_info.vertex->indirect_offsets;
const auto index = [&](uint8_t slot, IR::StageInputKind kind,
IR::StageInputKind base_kind) {
if (offsets == ShaderIndirectOffsets {}) {
return builtin(kind);
}
const auto base = builtin(base_kind);
if (slot < options.user_data_count) {
entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(options.user_data_base + slot),
base);
}
return entry_ir.ISub(builtin(kind), base);
};
entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
builtin(IR::StageInputKind::VertexIndex));
index(offsets.vertex, IR::StageInputKind::VertexIndex,
IR::StageInputKind::BaseVertex));
entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
builtin(IR::StageInputKind::InstanceIndex));
index(offsets.instance, IR::StageInputKind::InstanceIndex,
IR::StageInputKind::BaseInstance));
}
}
const bool flush_f32_inputs = options.stage == ShaderType::Compute &&
src/graphics/shader/recompiler/ir/ShaderIR.h
+2
View File
@@ -251,6 +251,8 @@ enum class StageInputKind {
LocalInvocationId,
LocalInvocationIndex,
GlobalInvocationId,
BaseVertex,
BaseInstance,
Parameter,
};
src/graphics/shader/recompiler/ir/passes/ShaderInfoCollection.cpp
+4
View File
@@ -175,6 +175,8 @@ void AddBuiltinInput(ShaderInfo& info, StageInputKind kind) {
case StageInputKind::GlobalInvocationId:
AddInput(info, kind, 0, 3, "gl_GlobalInvocationID");
break;
case StageInputKind::BaseVertex: AddInput(info, kind, 0, 1, "gl_BaseVertex"); break;
case StageInputKind::BaseInstance: AddInput(info, kind, 0, 1, "gl_BaseInstance"); break;
case StageInputKind::PackedAncillary:
case StageInputKind::Parameter: break;
}
@@ -444,6 +446,8 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
case StageInputKind::InstanceIndex:
case StageInputKind::FrontFacing:
case StageInputKind::LocalInvocationIndex:
case StageInputKind::BaseVertex:
case StageInputKind::BaseInstance:
if (component != 0u) {
return Fail("typed scalar builtin component is out of range");
}