From fc2713e5a501441875470e2976dd83d82eaadcbe Mon Sep 17 00:00:00 2001 From: Wei Chen Date: Wed, 9 Oct 2019 17:47:04 -0700 Subject: [PATCH] [Relay][VM] Fix constant folding issue in VM compiler (#4077) * [Relay][VM] Fix constant folding issue in VM compiler 1. allow pass params when compile a module 2. enhance profiler robustness * remove dead code * fix lint * add get_params * fix test * don't pass params back * remove get_params * docs * move compile function to api * compile clashes with builtin name * fix compilation error * remove dead code --- python/tvm/relay/backend/profiler_vm.py | 89 +++++++------- python/tvm/relay/backend/vm.py | 134 +++++++++++++--------- src/relay/backend/vm/compiler.cc | 56 ++++++++- src/relay/backend/vm/compiler.h | 30 ++++- src/runtime/vm/profiler/vm.cc | 5 + tests/python/relay/test_vm.py | 9 +- tests/python/relay/test_vm_serialization.py | 10 +- tests/python/unittest/test_runtime_vm_profiler.py | 4 +- 8 files changed, 213 insertions(+), 124 deletions(-) diff --git a/python/tvm/relay/backend/profiler_vm.py b/python/tvm/relay/backend/profiler_vm.py index 3adbeca..8ae3161 100644 --- a/python/tvm/relay/backend/profiler_vm.py +++ b/python/tvm/relay/backend/profiler_vm.py @@ -14,33 +14,53 @@ # KIND, either express or implied. See the License for the # specific language governing permissions and limitations # under the License. -# pylint: disable=no-else-return, unidiomatic-typecheck, undefined-variable, invalid-name +# pylint: disable=no-else-return, unidiomatic-typecheck, undefined-variable, invalid-name, redefined-builtin """ The Relay Virtual Machine profiler. Provides extra APIs for profiling vm execution. """ -import tvm from . import vm, _vm -def _update_target(target): - target = target if target else tvm.target.current_target() - if target is None: - raise ValueError("Target is not set in env or passed as argument.") +def compile(mod, target=None, target_host=None, params=None): + """ + Parameters + ---------- + mod : relay.Module + The Relay module to build. - tgts = {} - if isinstance(target, (str, tvm.target.Target)): - dev_type = tvm.expr.IntImm("int32", tvm.nd.context(str(target)).device_type) - tgts[dev_type] = tvm.target.create(target) - elif isinstance(target, dict): - for dev, tgt in target.items(): - dev_type = tvm.expr.IntImm("int32", tvm.nd.context(dev).device_type) - tgts[dev_type] = tvm.target.create(tgt) - else: - raise TypeError("target is expected to be str, tvm.target.Target, " + - "or dict of str to str/tvm.target.Target, but received " + - "{}".format(type(target))) - return tgts + target : str, :any:`tvm.target.Target`, or dict of str(i.e. + device/context name) to str/tvm.target.Target, optional + For heterogeneous compilation, it is a dictionary indicating context + to target mapping. For homogeneous compilation, it is a build target. + + target_host : str or :any:`tvm.target.Target`, optional + Host compilation target, if target is device. + When TVM compiles device specific program such as CUDA, + we also need host(CPU) side code to interact with the driver + to setup the dimensions and parameters correctly. + target_host is used to specify the host side codegen target. + By default, llvm is used if it is enabled, + otherwise a stackvm intepreter is used. + + params : dict of str to NDArray + Input parameters to the graph that do not change + during inference time. Used for constant folding. + + Returns + ------- + vm : VirtualMachineProfiler + The profile VM runtime. + """ + compiler = VMCompilerProfiler() + target = compiler.update_target(target) + target_host = compiler.update_target_host(target, target_host) + if params: + compiler.set_params(params) + tophub_context = compiler.tophub_context(target) + with tophub_context: + compiler._compile(mod, target, target_host) + return VirtualMachineProfiler(compiler._get_vm()) class VMCompilerProfiler(vm.VMCompiler): """Build Relay module to run on VM runtime.""" @@ -49,36 +69,7 @@ class VMCompilerProfiler(vm.VMCompiler): self.mod = _vm._VMCompilerProfiler() self._compile = self.mod["compile"] self._get_vm = self.mod["get_vm"] - - def compile(self, mod, target=None, target_host=None): - """ - Parameters - ---------- - mod : relay.Module - The Relay module to build. - - target : str, :any:`tvm.target.Target`, or dict of str(i.e. - device/context name) to str/tvm.target.Target, optional - For heterogeneous compilation, it is a dictionary indicating context - to target mapping. For homogeneous compilation, it is a build target. - - target_host : str or :any:`tvm.target.Target`, optional - Host compilation target, if target is device. - When TVM compiles device specific program such as CUDA, - we also need host(CPU) side code to interact with the driver - to setup the dimensions and parameters correctly. - target_host is used to specify the host side codegen target. - By default, llvm is used if it is enabled, - otherwise a stackvm intepreter is used. - - Returns - ------- - vm : VirtualMachineProfiler - The profile VM runtime. - """ - target = _update_target(target) - self._compile(mod, target, target_host) - return VirtualMachineProfiler(self._get_vm()) + self._set_params_func = self.mod["set_params"] class VirtualMachineProfiler(vm.VirtualMachine): """Relay profile VM runtime.""" diff --git a/python/tvm/relay/backend/vm.py b/python/tvm/relay/backend/vm.py index a6cb91c..e54629d 100644 --- a/python/tvm/relay/backend/vm.py +++ b/python/tvm/relay/backend/vm.py @@ -14,7 +14,7 @@ # KIND, either express or implied. See the License for the # specific language governing permissions and limitations # under the License. -# pylint: disable=no-else-return, unidiomatic-typecheck, undefined-variable, invalid-name +# pylint: disable=no-else-return, unidiomatic-typecheck, undefined-variable, invalid-name, redefined-builtin """ The Relay Virtual Machine. @@ -25,30 +25,11 @@ import numpy as np import tvm from tvm import autotvm from tvm._ffi.runtime_ctypes import TVMByteArray +from tvm.relay import expr as _expr from . import _vm from . import vmobj as _obj from .interpreter import Executor - -def _update_target(target): - target = target if target else tvm.target.current_target() - if target is None: - raise ValueError("Target is not set in env or passed as argument.") - - tgts = {} - if isinstance(target, (str, tvm.target.Target)): - dev_type = tvm.expr.IntImm("int32", tvm.nd.context(str(target)).device_type) - tgts[dev_type] = tvm.target.create(target) - elif isinstance(target, dict): - for dev, tgt in target.items(): - dev_type = tvm.expr.IntImm("int32", tvm.nd.context(dev).device_type) - tgts[dev_type] = tvm.target.create(tgt) - else: - raise TypeError("target is expected to be str, tvm.target.Target, " + - "or dict of str to str/tvm.target.Target, but received " + - "{}".format(type(target))) - return tgts - def _convert(arg, cargs): if isinstance(arg, (np.ndarray, tvm.nd.NDArray)): cargs.append(_obj.tensor_object(arg)) @@ -144,40 +125,85 @@ class VirtualMachine(object): return self.mod +def compile(mod, target=None, target_host=None, params=None): + """ + Parameters + ---------- + mod : relay.Module + The Relay module to build. + + target : str, :any:`tvm.target.Target`, or dict of str(i.e. + device/context name) to str/tvm.target.Target, optional + For heterogeneous compilation, it is a dictionary indicating context + to target mapping. For homogeneous compilation, it is a build target. + + target_host : str or :any:`tvm.target.Target`, optional + Host compilation target, if target is device. + When TVM compiles device specific program such as CUDA, + we also need host(CPU) side code to interact with the driver + to setup the dimensions and parameters correctly. + target_host is used to specify the host side codegen target. + By default, llvm is used if it is enabled, + otherwise a stackvm intepreter is used. + + params : dict of str to NDArray + Input parameters to the graph that do not change + during inference time. Used for constant folding. + + Returns + ------- + vm : VirtualMachine + The VM runtime. + """ + compiler = VMCompiler() + + target = compiler.update_target(target) + target_host = compiler.update_target_host(target, target_host) + if params: + compiler.set_params(params) + tophub_context = compiler.tophub_context(target) + with tophub_context: + compiler._compile(mod, target, target_host) + return VirtualMachine(compiler._get_vm()) + class VMCompiler(object): """Build Relay module to run on VM runtime.""" def __init__(self): self.mod = _vm._VMCompiler() self._compile = self.mod["compile"] self._get_vm = self.mod["get_vm"] + self._set_params_func = self.mod["set_params"] + + def set_params(self, params): + """Set constant parameters for the model""" + inputs = {} + for name, param in params.items(): + if isinstance(param, np.ndarray): + param = _nd.array(param) + inputs[name] = _expr.const(param) + self._set_params_func(inputs) + + def update_target(self, target): + """Update target""" + target = target if target else tvm.target.current_target() + if target is None: + raise ValueError("Target is not set in env or passed as argument.") + tgts = {} + if isinstance(target, (str, tvm.target.Target)): + dev_type = tvm.expr.IntImm("int32", tvm.nd.context(str(target)).device_type) + tgts[dev_type] = tvm.target.create(target) + elif isinstance(target, dict): + for dev, tgt in target.items(): + dev_type = tvm.expr.IntImm("int32", tvm.nd.context(dev).device_type) + tgts[dev_type] = tvm.target.create(tgt) + else: + raise TypeError("target is expected to be str, tvm.target.Target, " + + "or dict of str to str/tvm.target.Target, but received " + + "{}".format(type(target))) + return tgts - def compile(self, mod, target=None, target_host=None): - """ - Parameters - ---------- - mod : relay.Module - The Relay module to build. - - target : str, :any:`tvm.target.Target`, or dict of str(i.e. - device/context name) to str/tvm.target.Target, optional - For heterogeneous compilation, it is a dictionary indicating context - to target mapping. For homogeneous compilation, it is a build target. - - target_host : str or :any:`tvm.target.Target`, optional - Host compilation target, if target is device. - When TVM compiles device specific program such as CUDA, - we also need host(CPU) side code to interact with the driver - to setup the dimensions and parameters correctly. - target_host is used to specify the host side codegen target. - By default, llvm is used if it is enabled, - otherwise a stackvm intepreter is used. - - Returns - ------- - vm : VirtualMachine - The VM runtime. - """ - target = _update_target(target) + def update_target_host(self, target, target_host): + """Update target host""" target_host = None if target_host == "" else target_host if not target_host: for device_type, tgt in target.items(): @@ -186,19 +212,16 @@ class VMCompiler(object): break if not target_host: target_host = "llvm" if tvm.module.enabled("llvm") else "stackvm" - target_host = tvm.target.create(target_host) + return tvm.target.create(target_host) + def tophub_context(self, target): # If current dispatch context is fallback context (the default root context), # then load pre-tuned parameters from TopHub if isinstance(autotvm.DispatchContext.current, autotvm.FallbackContext): tophub_context = autotvm.tophub.context(list(target.values())) else: tophub_context = autotvm.util.EmptyContext() - - with tophub_context: - self._compile(mod, target, target_host) - return VirtualMachine(self._get_vm()) - + return tophub_context class VMExecutor(Executor): """ @@ -226,8 +249,7 @@ class VMExecutor(Executor): self.mod = mod self.ctx = ctx self.target = target - compiler = VMCompiler() - self.vm = compiler.compile(mod, target) + self.vm = compile(mod, target) self.vm.init(ctx) def _make_executor(self, expr=None): diff --git a/src/relay/backend/vm/compiler.cc b/src/relay/backend/vm/compiler.cc index 013d8b0..00d4fb4 100644 --- a/src/relay/backend/vm/compiler.cc +++ b/src/relay/backend/vm/compiler.cc @@ -780,23 +780,73 @@ PackedFunc VMCompiler::GetFunction(const std::string& name, if (name == "compile") { return PackedFunc([sptr_to_self, this](TVMArgs args, TVMRetValue* rv) { CHECK_EQ(args.num_args, 3); - this->Compile(args[0], args[1], args[2]); + Module mod = args[0]; + this->Compile(mod, args[1], args[2]); }); } else if (name == "get_vm") { return PackedFunc([sptr_to_self, this](TVMArgs args, TVMRetValue* rv) { *rv = runtime::Module(vm_); }); + } else if (name == "set_params") { + return PackedFunc([sptr_to_self, this](TVMArgs args, TVMRetValue* rv) { + Map params = args[0]; + for (const auto& kv : params) { + this->SetParam(kv.first, kv.second->data); + } + }); } else { LOG(FATAL) << "Unknown packed function: " << name; return PackedFunc([sptr_to_self, name](TVMArgs args, TVMRetValue* rv) {}); } } -void VMCompiler::Compile(const Module& mod_ref, +void VMCompiler::SetParam(const std::string& name, runtime::NDArray data_in) { + params_[name] = data_in; +} + +relay::Function VMCompiler::BindParamsByName( + relay::Function func, + const std::unordered_map& params) { + std::unordered_map name_dict; + std::unordered_set repeat_var; + for (auto arg : func->params) { + const auto &name = arg->name_hint(); + if (name_dict.count(name)) { + repeat_var.insert(arg); + } else { + name_dict[name] = arg; + } + } + std::unordered_map bind_dict; + for (auto &kv : params) { + if (name_dict.count(kv.first) == 0) { + continue; + } + auto arg = name_dict.at(kv.first); + if (repeat_var.count(arg)) { + LOG(FATAL) << "Multiple args in the function have name " << kv.first; + } + bind_dict[arg] = ConstantNode::make(kv.second); + } + Expr bound_expr = relay::Bind(func, bind_dict); + Function ret = Downcast(bound_expr); + CHECK(ret.defined()) + << "The returning type is expected to be a Relay Function." + << "\n"; + return ret; +} + + +void VMCompiler::Compile(Module mod, const TargetsMap& targets, const tvm::Target& target_host) { CHECK_EQ(targets.size(), 1) << "Currently VM compiler doesn't support heterogeneous compilation"; + if (params_.size()) { + auto f = BindParamsByName(mod->Lookup("main"), params_); + auto gvar = mod->GetGlobalVar("main"); + mod->Add(gvar, f); + } InitVM(); targets_ = targets; @@ -804,7 +854,7 @@ void VMCompiler::Compile(const Module& mod_ref, // Run some optimizations first, this code should // be moved to pass manager. - context_.module = OptimizeModule(mod_ref, targets_); + context_.module = OptimizeModule(mod, targets_); // Populate the global map. // diff --git a/src/relay/backend/vm/compiler.h b/src/relay/backend/vm/compiler.h index 14a5035..dff1ef7 100644 --- a/src/relay/backend/vm/compiler.h +++ b/src/relay/backend/vm/compiler.h @@ -100,11 +100,37 @@ class VMCompiler : public runtime::ModuleNode { vm_ = std::make_shared(); } - void Compile(const Module& mod_ref, + /*! + * \brief Set the parameters + * + * \param name name of parameter + * \param data_in input DLTensor + */ + void SetParam(const std::string& name, runtime::NDArray data_in); + + /*! + * \brief Compile functions in a Module + * + * \param mod Relay Module + * \param targets For heterogeneous compilation, it is a dictionary indicating context + to target mapping. For homogeneous compilation, it is a build target. + * \param target_host Host compilation target, if target is device. + */ + void Compile(Module mod, const TargetsMap& targets, const tvm::Target& target_host); protected: + /*! + * \brief Bind params to function by using name + * \param func Relay function + * \param params params dict + * \return relay::Function + */ + relay::Function BindParamsByName( + relay::Function func, + const std::unordered_map& params); + Module OptimizeModule(const Module& mod, const TargetsMap& targets); void PopulateGlobalMap(); @@ -120,6 +146,8 @@ class VMCompiler : public runtime::ModuleNode { VMCompilerContext context_; /*! \brief Compiled virtual machine. */ std::shared_ptr vm_; + /*! \brief parameters */ + std::unordered_map params_; }; } // namespace vm diff --git a/src/runtime/vm/profiler/vm.cc b/src/runtime/vm/profiler/vm.cc index 1d3ac83..5f59f6e 100644 --- a/src/runtime/vm/profiler/vm.cc +++ b/src/runtime/vm/profiler/vm.cc @@ -98,6 +98,11 @@ void VirtualMachineDebug::InvokePacked(Index packed_index, Index output_size, const std::vector& args) { auto ctx = VirtualMachine::GetParamsContext(); + // warmup + VirtualMachine::InvokePacked(packed_index, func, arg_count, output_size, + args); + TVMSynchronize(ctx.device_type, ctx.device_id, nullptr); + auto op_begin = std::chrono::high_resolution_clock::now(); VirtualMachine::InvokePacked(packed_index, func, arg_count, output_size, args); diff --git a/tests/python/relay/test_vm.py b/tests/python/relay/test_vm.py index f60c533..f643f8a 100644 --- a/tests/python/relay/test_vm.py +++ b/tests/python/relay/test_vm.py @@ -47,15 +47,13 @@ def veval(f, *args, ctx=tvm.cpu(), target="llvm"): if isinstance(f, relay.Expr): mod = relay.Module() mod["main"] = f - compiler = relay.vm.VMCompiler() - vm = compiler.compile(mod, target) + vm = relay.vm.compile(mod, target) vm.init(tvm.cpu()) return vm.invoke("main", *args) else: assert isinstance(f, relay.Module), "expected expression or module" mod = f - compiler = relay.vm.VMCompiler() - vm = compiler.compile(mod, target) + vm = relay.vm.compile(mod, target) vm.init(tvm.cpu()) ret = vm.invoke("main", *args) return ret @@ -582,8 +580,7 @@ def test_set_params(): b = relay.var('b', shape=(6,)) y = relay.nn.bias_add(relay.nn.dense(x, w), b) mod["main"] = relay.Function([x, w, b], y) - compiler = relay.vm.VMCompiler() - vm = compiler.compile(mod, 'llvm') + vm = relay.vm.compile(mod, 'llvm') vm.init(tvm.cpu()) x_np = np.random.uniform(size=(10, 5)).astype('float32') diff --git a/tests/python/relay/test_vm_serialization.py b/tests/python/relay/test_vm_serialization.py index a32ec27..3a317fc 100644 --- a/tests/python/relay/test_vm_serialization.py +++ b/tests/python/relay/test_vm_serialization.py @@ -28,18 +28,16 @@ from tvm.relay.prelude import Prelude from tvm.contrib import util from tvm.relay import testing -def create_vm(f, ctx=tvm.cpu(), target="llvm"): +def create_vm(f, ctx=tvm.cpu(), target="llvm", params=None): if isinstance(f, relay.Expr): mod = relay.Module() mod["main"] = f - compiler = relay.vm.VMCompiler() - vm = compiler.compile(mod, target) + vm = _vm.compile(mod, target=target, params=params) vm.init(ctx) return vm else: assert isinstance(f, relay.Module), "expected mod as relay.Module" - compiler = relay.vm.VMCompiler() - vm = compiler.compile(f, target) + vm = _vm.compile(f, target=target, params=params) vm.init(ctx) return vm @@ -61,7 +59,7 @@ def run_network(mod, return result.asnumpy().astype(dtype) def get_serialized_output(mod, data, params, target, ctx, dtype='float32'): - vm = create_vm(mod, ctx, target) + vm = create_vm(mod, ctx, target, params=params) ser = serializer.Serializer(vm) code, lib = ser.serialize() deser = deserializer.Deserializer(code, lib) diff --git a/tests/python/unittest/test_runtime_vm_profiler.py b/tests/python/unittest/test_runtime_vm_profiler.py index 4281ccc..b5ce0ec 100644 --- a/tests/python/unittest/test_runtime_vm_profiler.py +++ b/tests/python/unittest/test_runtime_vm_profiler.py @@ -22,13 +22,11 @@ import pytest from tvm import relay from tvm.relay.testing import resnet -@pytest.mark.skip def test_basic(): mod, params = resnet.get_workload() - compiler = relay.profiler_vm.VMCompilerProfiler() target = 'llvm' ctx = tvm.cpu() - vm = compiler.compile(mod, target) + vm = relay.profiler_vm.compile(mod, target) vm.init(ctx) vm.load_params(params) -- 2.7.4