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

src/graphics/host_gpu/renderer/pipeline/pipelineCache.cpp

@@ -687,6 +687,9 @@ void PipelineCache::GetGraphicsPrograms(const CommandBuffer& command, DrawRender
 		vertex_params[0] = PrepareProgram(vertex_regs, context, user_config, vertex_info[0]);
 	}
 	const bool mesh_active = vertex_info[0].logical_stage == ShaderType::Mesh;
+	if (!tess_active && !mesh_active) {
+		vertex_info[0].indirect_offsets = state.indirect_offsets;
+	}
 	if (object_primitive_id) {
 		// CxObjectIdMode: register source plus a 32-bit pipelined primitive-ID addition.
 		EXIT_NOT_IMPLEMENTED(sh.m_paClObjprimIdCntl != 2u || mesh_active || tess_active ||

src/graphics/host_gpu/renderer/pipeline/pipelineCache.h

@@ -137,6 +137,8 @@ struct DrawRenderState {
 	bool                                 ps_active                                = false;
 	std::array<ShaderVertexInputInfo, 3> vertex_info;
 	ShaderPixelInputInfo                 ps_input_info;
+	// Input: the user-data slots a GPU-read indirect draw patches.
+	ShaderIndirectOffsets                indirect_offsets;
 
 	[[nodiscard]] std::span<const ShaderVertexInputInfo> VertexStages() const {
 		return {vertex_info.data(), vertex_info[1].stage ? 3u : 1u};

src/graphics/host_gpu/renderer/render.h

@@ -45,9 +45,10 @@ enum class DrawOffsetSource : uint8_t {
 
 // Nonzero address: the GPU reads count Vulkan indirect draw records, stride bytes apart.
 struct DrawIndirectArgs {
-	uint64_t address = 0;
-	uint32_t count   = 0;
-	uint32_t stride  = 0;
+	uint64_t              address = 0;
+	uint32_t              count   = 0;
+	uint32_t              stride  = 0;
+	ShaderIndirectOffsets offsets;
 };
 
 struct DrawIndexArgs {

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -1250,8 +1250,9 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	}
 }
 
-// Vulkan offsets for a prepared draw, or nullopt when its indirect records need CPU counts, as
-// mesh and quad-list draws do. Indirect records carry their own offsets.
+// Vulkan offsets for a prepared draw, or nullopt when its indirect records need the CPU: mesh
+// and quad-list draws for counts, tessellation draws for patched SGPRs. Indirect records carry
+// their own offsets.
 static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&        ucfg,
                                                    const ShaderVertexInputInfo& vs,
                                                    DrawOffsetSource             source,
@@ -1260,7 +1261,8 @@ static std::optional<DrawEmitInfo> ResolveDrawEmit(const HW::UserConfig&
                                                    const DrawIndirectArgs&      indirect) {
 	if (indirect.address != 0) {
 		if (vs.stage.module->info.stage == ShaderType::Mesh ||
-		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy) {
+		    ucfg.GetPrimType() == Prospero::PrimitiveType::kQuadListLegacy ||
+		    vs.indirect_offsets != indirect.offsets) {
 			return std::nullopt;
 		}
 		return DrawEmitInfo {.indirect = indirect};
@@ -1324,7 +1326,7 @@ bool RenderExecutor::DrawIndex(uint64_t submit_id, CommandBuffer& buffer,
 
 	hw_check(buffer);
 
-	DrawRenderState state {};
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets};
 	if (!GetDrawTopology(ucfg, state.topology)) {
 		return true;
 	}
@@ -1429,7 +1431,7 @@ bool RenderExecutor::DrawAuto(uint64_t submit_id, CommandBuffer& buffer, const D
 	const DrawCallInfo draw {CommandBufferDebugOp::DrawIndexAuto,
 	                         args.vertex_count, args.instance_count, args.first_instance};
 
-	DrawRenderState state {};
+	DrawRenderState state {.indirect_offsets = args.indirect.offsets};
 	if (!GetDrawTopology(ucfg, state.topology)) {
 		ResetBindings();
 		return true;