shader: leave the start instance out of mesh and NGG instance ID VGPRs

Co-authored-by: Senaxx <michelbeld@gmail.com>

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -1161,7 +1161,6 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		if (mesh_active) {
 			draw_data = {draw.index_count,
 			             static_cast<uint32_t>(emit.vertex_offset),
-			             emit.first_instance,
 			             index_source.guest_element_size,
 			             static_cast<uint32_t>(index_source.address),
 			             static_cast<uint32_t>(index_source.address >> 32u)};

src/graphics/shader/recompiler/frontend/translate/Translate.cpp

@@ -1244,12 +1244,12 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			    entry_ir.BitwiseOr(wave_info, entry_ir.BitwiseOr(entry_ir.ShiftLeftLogical(
 			                                                         primitive_count, u32(8)),
 			                                                     vertex_count)));
+			// Instance IDs exclude the start instance, which shaders add from user data.
 			if (mesh.fast_launch) {
 				// Fast launch broadcasts the subgroup's base vertex and instance to every lane.
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5), entry_ir.IAdd(draw(1), group));
-				entry_ir.SetVectorReg(
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(6),
-				    static_cast<IR::VectorReg>(6),
+				                      builtin(IR::StageInputKind::WorkgroupId, 1));
-				    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
 			} else {
 				// GS adjacency addresses local ES records in LDS. Fans retain the draw's
 				// center in every subgroup; strip winding follows the global primitive.
@@ -1277,14 +1277,14 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				                                         entry_ir.ShiftLeftLogical(second, u32(18))));
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(1),
 				                      entry_ir.ShiftLeftLogical(third, u32(2)));
-				const auto index_bytes  = draw(3);
+				const auto index_bytes  = draw(2);
 				const auto indexed      = entry_ir.INotEqual(index_bytes, u32(0));
-				const auto index_low    = draw(4);
+				const auto index_low    = draw(3);
 				const auto byte_offset  = entry_ir.IAdd(entry_ir.BitwiseAnd(index_low, u32(3)),
 				                                           entry_ir.IMul(input_vertex, index_bytes));
 				const auto index_resource = entry_ir.Emit(
 				    IR::ValueOpcode::GetAddressResource,
-				    {entry_ir.BitwiseAnd(index_low, u32(~3u)), draw(5)});
+				    {entry_ir.BitwiseAnd(index_low, u32(~3u)), draw(4)});
 				const auto memory_index = static_cast<uint32_t>(result.memory_info.size());
 				result.memory_info.push_back({.kind = IR::ResourceKind::Global});
 				const auto packed_index = entry_ir.Emit(
@@ -1298,9 +1298,8 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 				     entry_ir.IMul(index_bytes, u32(8))}));
 				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(5),
 				                      entry_ir.IAdd(draw(1), entry_ir.Select(indexed, index, input_vertex)));
-				entry_ir.SetVectorReg(
+				entry_ir.SetVectorReg(static_cast<IR::VectorReg>(8),
-				    static_cast<IR::VectorReg>(8),
+				                      builtin(IR::StageInputKind::WorkgroupId, 1));
-				    entry_ir.IAdd(draw(2), builtin(IR::StageInputKind::WorkgroupId, 1)));
 			}
 		} else if (options.stage == ShaderType::Local) {
 			entry_ir.SetScalarReg(static_cast<IR::ScalarReg>(3), IR::U32(IR::Value(64u)));

src/graphics/shader/recompiler/ir/ShaderIR.h

@@ -365,7 +365,7 @@ static_assert(static_cast<uint32_t>(DescriptorBindingKind::Count) == 58u);
 
 struct PushData {
 	static constexpr uint32_t DwordCount = 32;
-	static constexpr uint32_t MeshDrawDwordCount = 6;
+	static constexpr uint32_t MeshDrawDwordCount = 5;
 	static constexpr uint32_t NoStart    = UINT32_MAX;
 	std::array<uint32_t, DwordCount> dwords {};
 

tests/shaderCfgTests.cpp

@@ -11787,7 +11787,7 @@ void TestMeshInputAssembly() {
     options.user_data_count = 0;
     options.input_info.vertex = &input;
     auto program = Frontend::TranslateProgram(decoded, graph, options);
-    const uint32_t draw[] = {test.count, test.base_vertex, 7, test.width,
+    const uint32_t draw[] = {test.count, test.base_vertex, test.width,
                              test.address_low, 0x12};
     Inst *load = nullptr;
     for (auto &inst : *program.blocks.front()) {
@@ -11832,7 +11832,7 @@ void TestMeshInputAssembly() {
     Check(sgprs[2] == test.group_info && sgprs[3] == test.wave_info &&
               vgprs[0] == ((test.first << 2) | (test.second << 18)) &&
               vgprs[1] == test.third * 4 && vgprs[5] == test.vertex_id &&
-              vgprs[test.fast_launch ? 6 : 8] == 9u,
+              vgprs[test.fast_launch ? 6 : 8] == 2u,
           "mesh prolog changed input assembly, wave counts, vertex ID, or instance ID");
   }
 }