graphics: fix vertex and instance IDs for unpatched indirect draws

src/graphics/shader/recompiler/ShaderRecompiler.cpp

@@ -814,7 +814,7 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	const auto& cfg = *selected_cfg;
 	Frontend::EmbeddedFetchPlan embedded_fetch;
 	if ((options.stage == ShaderType::Vertex || options.stage == ShaderType::Local) &&
-	    options.input_info.vertex->fetch_embedded) {
+	    options.input_info.vertex->has_dynamic_vertex_fetch) {
 		embedded_fetch = DetectEmbeddedVertexFetch(
 		    decoded, *options.input_info.vertex, options.user_data_base,
 		    static_cast<uint32_t>(options.user_data.size()), options.wave_size);

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

@@ -1390,11 +1390,12 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 			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.
+			// only in the user SGPRs the packet patches. SH_NOP leaves those SGPRs unchanged;
+			// it does not make the record's bases zero.
 			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 {}) {
+				if (!options.input_info.vertex->indirect_draw) {
 					return builtin(kind);
 				}
 				const auto base = builtin(base_kind);

src/graphics/shader/shader.cpp

@@ -291,10 +291,9 @@ static bool ShaderGetStaticVertexInputInfo(uint64_t shader_addr, const HW::UserS
 	const auto& metadata = data.vertex;
 
 	if (metadata.vertex_attrib_reg >= 0) {
-		info.fetch_external   = false;
-		info.fetch_embedded   = true;
-		info.fetch_attrib_reg = metadata.vertex_attrib_reg;
-		info.fetch_buffer_reg = metadata.vertex_buffer_reg;
+		info.has_dynamic_vertex_fetch = true;
+		info.fetch_attrib_reg         = metadata.vertex_attrib_reg;
+		info.fetch_buffer_reg         = metadata.vertex_buffer_reg;
 
 		const auto* attrib = reinterpret_cast<const uint32_t*>(
 		    static_cast<uint64_t>(user_sgpr.value[metadata.vertex_attrib_reg]) |
@@ -413,7 +412,7 @@ static void ShaderGetStaticInputInfoCS(const HW::ComputeShaderInfo& regs,
 void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t>& key) {
 	EXIT_IF(info.resources_num < 0 || info.resources_num > ShaderVertexInputInfo::RES_MAX);
 	key.clear();
-	key.push_back(static_cast<uint32_t>(info.fetch_embedded));
+	key.push_back(static_cast<uint32_t>(info.has_dynamic_vertex_fetch));
 	key.push_back(static_cast<uint32_t>(info.fetch_attrib_reg));
 	key.push_back(static_cast<uint32_t>(info.fetch_buffer_reg));
 	key.push_back(info.resources_num);
@@ -422,7 +421,8 @@ void BuildStageStaticKey(const ShaderVertexInputInfo& info, std::vector<uint32_t
 	key.push_back(info.lds_size_dwords);
 	key.push_back(info.pa_cl_vs_out_cntl);
 	key.push_back(info.user_clip_planes);
-	key.push_back(info.indirect_offsets.vertex | (uint32_t {info.indirect_offsets.instance} << 8u));
+	key.push_back(info.indirect_offsets.vertex | (uint32_t {info.indirect_offsets.instance} << 8u) |
+	              (uint32_t {info.indirect_draw} << 16u));
 	key.push_back(static_cast<uint32_t>(info.clip_space.enabled));
 	if (info.clip_space.enabled) {
 		for (const float value: info.clip_space.scale) {
@@ -693,9 +693,7 @@ void ShaderDbgDumpInputInfo(const ShaderVertexInputInfo& info) {
 
 	LOGF("ShaderDbgDumpInputInfo()\n");
 
-	LOGF("\t fetch_external = %s\n"
-	     "\t fetch_embedded = %s\n",
-	     info.fetch_external ? "true" : "false", info.fetch_embedded ? "true" : "false");
+	LOGF("\t has_dynamic_vertex_fetch = %s\n", info.has_dynamic_vertex_fetch ? "true" : "false");
 
 	for (int i = 0; i < info.resources_num; i++) {
 		LOGF("\t input %d\n", i);

src/graphics/shader/shader.h

@@ -125,8 +125,7 @@ struct ShaderTessellationInputInfo {
 };
 
 // GS user-data slots an indirect draw packet writes its record's base vertex and first instance
-// to. Like the hardware index VGPRs, the vertex shader's indices exclude both bases once either
-// slot is set.
+// to. A packet can leave both slots unused even when its record contains nonzero offsets.
 struct ShaderIndirectOffsets {
 	static constexpr uint8_t None = UINT8_MAX;
 
@@ -159,8 +158,10 @@ struct ShaderVertexInputInfo {
 	ShaderMeshInputInfo      mesh;
 	ShaderTessellationInputInfo tess;
 	ShaderIndirectOffsets   indirect_offsets;
-	bool                    fetch_external      = false;
-	bool                    fetch_embedded      = false;
+	// GPU-read records add Vulkan bases even when the packet patches no user SGPRs.
+	bool                    indirect_draw       = false;
+	// Attribute-table metadata enables dynamic fetch detection; loads are matched later.
+	bool                    has_dynamic_vertex_fetch = false;
 };
 
 struct ShaderComputeInputInfo: ShaderWorkgroupInputInfo {