def _compile(ext_func): """ This is the main wrapper that accepts an external relay function and runs all the passes to lower it down to command stream Parameters ---------- ext_func : tvm.relay.function.Function The partitioned relay function Returns ------- cs : str An hex string of the bytes of command stream encoded_constants : str An hex string of the bytes that includes concat'd encoded weights, encoded biases and scales. scratch_size : int The size of the scratch buffer needed. """ mod = tvm.IRModule() mod["main"] = ext_func mod = LegalizeEthosU()(mod) mod = relay.transform.InferType()(mod) # We are currently using copy_constants scheduler In the long run, # this should be a single intelligent and a composite scheduler # that can perform scheduling based on user inputs such as # scratch memory size. tir_mod, params = lower_to_tir(mod["main"], copy_constants()) cmms, encoded_constants, scratch_size = tir_to_cs_translator.translate( tir_mod, params) return cmms, encoded_constants, scratch_size
def relay_to_tir(mod: tvm.ir.IRModule) -> tvm.ir.IRModule: """ This is the hook for python-based lowering of a Relay module which lowers NPU external functions to TIR. Parameters ---------- mod : tvm.ir.IRModule This is the Relay module. Returns ------- mod : tvm.ir.IRModule The Relay module with scheduled NPU external functions. """ mod = OutlineCompilerFunctions("ethos-u")(mod) mod = LegalizeEthosU()(mod) mod = LUTsOptimizer()(mod) mod = relay.transform.InferType()(mod) mod = IdentityOptimizer()(mod) mod = LayoutOptimizer()(mod) mod = relay.transform.InferType()(mod) device_contexts = { gv: "ethos-u" for gv, _ in filter(lambda x: util.is_npu_func(x[1]), mod.functions.items()) } mod = mod.with_attr("device_contexts", device_contexts) # Use the cascader if it is enabled for the U55 accelerator, otherwise use copy_constants # scheduler if util.is_cascader_enabled(): if util.get_accelerator_config() == "ethos-u65-256": raise ValueError( "Cascading is not supported for the U65 accelerator") workspace_memory_pools = mod.attrs["workspace_memory_pools"] if not workspace_memory_pools: raise ValueError( "Workspace memory pool needs to be provided for the U55 cascader" ) if len(workspace_memory_pools.pools) != 1: raise ValueError( "Exactly one workspace pool needs to be provided for the U55 cascader" ) memory_pressure = _calculate_memory_pressure(mod) sram = extract_memory_info(workspace_memory_pools.pools[0], memory_pressure) tir_mod = LowerToTIR( _ethos_u55_cascader(sram, util.is_striping_enabled()))(mod) else: tir_mod = LowerToTIR(copy_constants())(mod) return tir_mod
def relay_to_tir_func(ext_func: relay.Function) -> tvm.tir.PrimFunc: """ This is the hook for python-based lowering of relay function that gets offloaded to the microNPU. Parameters ---------- ext_func : relay.Function This is the partitioned relay function Returns ------- primfunc : tir.PrimFunc This returns the scheduled PrimFunc """ assert len(ext_func.params) == 1 input_size = util.calculate_size_bytes(ext_func.params[0]) output_size = util.calculate_size_bytes(ext_func.body) mod = tvm.IRModule() mod["main"] = ext_func mod = LegalizeEthosU()(mod) mod = LUTsOptimizer()(mod) mod = LayoutOptimizer()(mod) mod = relay.transform.InferType()(mod) # We are currently using copy_constants scheduler In the long run, # this should be a single intelligent and a composite scheduler # that can perform scheduling based on user inputs such as # scratch memory size. tir_mod, const_dict = lower_to_tir(mod["main"], copy_constants()) for idx in const_dict.keys(): const_dict[idx] = tvm.nd.array(const_dict[idx]) primfunc = tir_mod["main"] primfunc = primfunc.with_attr("global_symbol", ext_func.attrs["global_symbol"]) primfunc = primfunc.with_attr("ethos-u.constants", const_dict) primfunc = primfunc.with_attr("ethos-u.input_size", input_size) primfunc = primfunc.with_attr("ethos-u.output_size", output_size) return primfunc
def relay_to_tir(mod: tvm.ir.IRModule) -> tvm.ir.IRModule: """ This is the hook for python-based lowering of a Relay module which lowers NPU external functions to TIR. Parameters ---------- mod : tvm.ir.IRModule This is the Relay module. Returns ------- mod : tvm.ir.IRModule The Relay module with scheduled NPU external functions. """ mod = OutlineCompilerFunctions("ethos-u")(mod) mod = LegalizeEthosU()(mod) mod = LUTsOptimizer()(mod) mod = relay.transform.InferType()(mod) mod = IdentityOptimizer()(mod) mod = LayoutOptimizer()(mod) mod = relay.transform.InferType()(mod) device_contexts = { gv: "ethos-u" for gv, _ in filter(lambda x: util.is_npu_func(x[1]), mod.functions.items()) } mod = mod.with_attr("device_contexts", device_contexts) # We are currently using copy_constants scheduler In the long run, # this should be a single intelligent and a composite scheduler # that can perform scheduling based on user inputs such as # scratch memory size. mod = LowerToTIR(copy_constants)(mod) return mod