[JIT] assertion failed in emitFloatCompare

[clover2123]

When I run a sample tfjs-model (imagenet) within aarch64 platform, it failed in the following part.

In emitFloatCompare method https://github.com/Samsung/walrus/blob/6f000941bc76de44ccc69a42466652bd3c637473/src/jit/FloatMathInl.h#L441-L444

I found that for a certain condition, instr->next() is CodeLabel which leads to assertion fail at instr->next()->asInstruction() because it seems that CodeLabel is not a instruction group.

I briefly share the issue here first.

[zherczeg]

It looks like I forgot this case. A compare cannot be the last instruction of a code block, because it always ends with a return. I silently included labels into this rule, which is not considered as an instruction by the compiler. We just need to put the code block into an isInstruction() check.

[zherczeg]

https://github.com/Samsung/walrus/blob/6f000941bc76de44ccc69a42466652bd3c637473/src/jit/IntMath32Inl.h#L1001 We did it correctly for math

[clover2123]

@zherczeg Thanks! I fixed it adding the code into a new if (instr->next()->isInstruction()) block, but another problem has occurred.

Now, it fails on the below check code: https://github.com/zherczeg/sljit/blob/44cee7aa5d6732de8f5eba892d52c734f1d0fe2c/sljit_src/sljitLir.c#L2700

Could you guess the reason? Or I'll give you a detailed description to reproduce this error. (I tested on aarch64 environment because current walrus supports SIMD JIT only for arm platforms)

[zherczeg]

These are usually caused by passing wrong parameters to the compiler. I would need a bit of backtrace where it is called (a sljit_emit_simd_lane_mov call somewhere from src/jit) and what was the arguments. I can check a detailed report as well.

[clover2123]

Following is the whole backtrace when the assertion fails I run a sample model by lwnode-escargot-walrus with your comment updated

#1  0x0000fffff71faaac in __GI_abort () at abort.c:79
#2  0x0000fffff6fe8f64 in check_sljit_emit_simd_lane_mov (compiler=0xaaaaac9b4640, type=541697, freg=1, lane_index=1, srcdst=127, srcdstw=0) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/third_party/walrus/src/jit/../../third_party/sljit/sljit_src/sljitLir.c:2700
#3  0x0000fffff6ff3984 in sljit_emit_simd_lane_mov (compiler=0xaaaaac9b4640, type=541697, freg=1, lane_index=1, srcdst=127, srcdstw=0) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/third_party/walrus/src/jit/../../third_party/sljit/sljit_src/sljitNativeARM_64.c:2818
#4  0x0000fffff7002560 in Walrus::emitExtractLaneSIMD (compiler=0xaaaaac9b4640, instr=0xaaaaac8f6da0) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/third_party/walrus/src/jit/SimdInl.h:91
#5  0x0000fffff7004678 in Walrus::JITCompiler::compile (this=0xffffffffde10) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/third_party/walrus/src/jit/Backend.cpp:812
#6  0x0000fffff7014104 in Walrus::Module::jitCompile (this=0xaaaaabe836f0, functions=0x0, functionsLength=0, verboseLevel=1) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/third_party/walrus/src/jit/ByteCodeParser.cpp:1289
#7  0x0000fffff7037d54 in Walrus::WASMParser::parseBinary (store=0xaaaaabe327b0, filename="", data=0xfffff4394010 "", len=424594) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/third_party/walrus/src/parser/WASMParser.cpp:2510
#8  0x0000fffff707d378 in wasm_module_new (store=0xaaaaabe32810, binary=0xffffffffe2a8) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/third_party/walrus/src/api/wasm.cpp:796
#9  0x0000fffff7bfed84 in Escargot::WASMOperations::compileModule (state=..., thisValue=..., argc=1, argv=0xffffffffe490, newTarget=...) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/wasm/WASMOperations.cpp:237
#10 0x0000fffff7b6132c in Escargot::NativeFunctionObject::processNativeFunctionCall<false, true> (this=0x126e610, state=..., receiverSrc=..., argc=1, argv=0xffffffffe490, newTarget=...) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/runtime/FunctionObjectInlines.h:301
#11 0x0000fffff7b60ec4 in Escargot::NativeFunctionObject::call (this=0x126e610, state=..., thisValue=..., argc=1, argv=0xffffffffe490) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/runtime/NativeFunctionObject.cpp:78
#12 0x0000fffff7b6bc88 in Escargot::Object::call (state=..., callee=..., thisValue=..., argc=1, argv=0xffffffffe490) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/runtime/Object.cpp:1280
#13 0x0000fffff7b57e88 in Escargot::PromiseReactionJob::<lambda()>::operator()(void) const (__closure=0xffffffffe618) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/runtime/Job.cpp:70
#14 0x0000fffff7b597e8 in std::_Function_handler<Escargot::Value(), Escargot::PromiseReactionJob::run()::<lambda()> >::_M_invoke(const std::_Any_data &) (__functor=...) at /usr/include/c++/9/bits/std_function.h:285
#15 0x0000fffff7ba61c4 in std::function<Escargot::Value ()>::operator()() const (this=0xffffffffe618) at /usr/include/c++/9/bits/std_function.h:688
#16 0x0000fffff7ba460c in Escargot::SandBox::run(std::function<Escargot::Value ()> const&) (this=0xffffffffe5b0, scriptRunner=...) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/runtime/SandBox.cpp:123
#17 0x0000fffff7b58138 in Escargot::PromiseReactionJob::run (this=0x17bda10) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/runtime/Job.cpp:86
#18 0x0000fffff7be4640 in Escargot::VMInstance::executePendingJob (this=0x57070) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/runtime/VMInstance.cpp:743
#19 0x0000fffff788d0a8 in Escargot::VMInstanceRef::executePendingJob (this=0x57070) at /home/ubuntu/hwpark/lwnode_wasm_jit/deps/escargot/src/api/EscargotPublic.cpp:1194
#20 0x0000aaaaab451528 in v8::MicrotasksScope::PerformCheckpoint (v8_isolate=0x4ed90) at ../../../src/api-environment.cc:2532
#21 0x0000aaaaab768d70 in node::InternalCallbackScope::Close (this=0xffffffffec20) at ../../../deps/node/src/api/callback.cc:114
#22 0x0000aaaaab768b28 in node::InternalCallbackScope::~InternalCallbackScope (this=0xffffffffec20, __in_chrg=<optimized out>) at ../../../deps/node/src/api/callback.cc:77
#23 0x0000aaaaab4d0c84 in node::StartExecution(node::Environment*, std::function<v8::MaybeLocal<v8::Value> (node::StartExecutionCallbackInfo const&)>) (env=0xaaaaabe48e20, cb=...) at ../../../deps/node/src/node.cc:449
#24 0x0000aaaaab76df68 in node::LoadEnvironment(node::Environment*, std::function<v8::MaybeLocal<v8::Value> (node::StartExecutionCallbackInfo const&)>, std::unique_ptr<node::InspectorParentHandle, std::default_delete<node::InspectorParentHandle> >) (env=0xaaaaabe48e20, cb=...,
    removeme=std::unique_ptr<struct node::InspectorParentHandle> = {...}) at ../../../deps/node/src/api/environment.cc:488
#25 0x0000aaaaab76dea4 in node::LoadEnvironment (env=0xaaaaabe48e20) at ../../../deps/node/src/api/environment.cc:478
#26 0x0000aaaaab4cf0c4 in LWNode::LWNodeMainRunner::Run (this=0xffffffffee50, nodeMainInstance=...) at ../../../deps/node/src/node_main_lw_runner-inl.h:127
#27 0x0000aaaaab4d3108 in node::Start (argc=3, argv=0xfffffffff1f8) at ../../../deps/node/src/node.cc:1107
#28 0x0000aaaaab3ae360 in main (argc=3, argv=0xfffffffff1f8) at ../../../deps/node/src/node_main_lw.cc:83

[clover2123]

I share the procedure to build and reproduce the error below.

• Environment : aarch64
• lwnode engine : https://github.com/Samsung/lwnode main branch

• Build lwnode 1) Do git clone and git submodule update (recursive) of lwnode 2) Open escargot.gyp file and add the following flag in escargot_configs '-DESCARGOT_WASM=ON', 3) fix parseBinary function defined in deps/escargot/third_party/walrus/src/parser/WASMParser.cpp as below to force JIT for every module

  std::pair<Optional<Module*>, std::string> WASMParser::parseBinary(Store* store, const std::string& filename, const uint8_t* data, size_t len)
  {
  wabt::WASMBinaryReader delegate;
  
  std::string error = ReadWasmBinary(filename, data, len, &delegate);
  if (error.length()) {
      return std::make_pair(nullptr, error);
  }
  
  Module* module = new Module(store, delegate.parsingResult());
  module->jitCompile(nullptr, 0, 1);
  return std::make_pair(module, std::string());
  }

  3) Build debug mode

  ./configure.py --debug
  ninja -C out/linux/Debug lwnode

  4) Run a sample AI model by lwnode with WASM backend (run with wasm argument)

  out/linux/Debug/lwnode sample.js wasm

  If run the sample code without argument wasm, it will be run in JS backend mode

• sample code (sample.js)

  
  let tf;

async function useWasm() { tf = require('@tensorflow/tfjs'); const { getThreadsCount, setThreadsCount, } = require('@tensorflow/tfjs-backend-wasm'); await tf.setBackend('wasm'); setThreadsCount(2); console.log('step 0 - set backend', tf.getBackend(), getThreadsCount()); }

async function useNode() { tf = require('@tensorflow/tfjs-node'); console.log('step 0 - set backend', 'node adapter'); }

async function run() { try { console.time('tf');

            if (process.argv[2] === 'wasm') {
                    await useWasm();
            } else if(process.argv[2] === 'node') {
                    useNode();
            } else {
                    tf = require('@tensorflow/tfjs');
            }

            const img = tf.ones([1, 224, 224, 3]).toFloat();

            console.log('step 1 - download model');
            const model = await tf.loadGraphModel(
                    'https://tfhub.dev/google/tfjs-model/imagenet/mobilenet_v2_130_224/classification/3/default/1',
                    { fromTFHub: true },
            );

            console.log('step 2 - start prediction');
            const prediction = await model.predict(img);
            console.log(prediction);
            console.timeEnd('tf');
    } catch (error) {
            console.log('error', error);
    }

}

run();

[clover2123]

Oops, I forgot to add the package.json file below To run the above sample code, you need to install some modules which are defined in package.json file.

• place the following package.json file on the same directory of sample.js package.json

  {
      "name": "example",
      "version": "1.0.0",
      "description": "",
      "main": "sample.js",
      "author": "",
      "license": "ISC",
      "dependencies": {
              "@tensorflow/tfjs-backend-cpu": "^4.0.0",
              "@tensorflow/tfjs-backend-wasm": "^4.0.0",
              "@tensorflow/tfjs-node": "^4.0.0"
      }
  }

• install required packages

  npm install

[zherczeg]

Thank you for all the info. I was able to set up the project on x86. I hoped the error will be present there because working with ARM-64 hw is a bit difficult. Fortunately it worked (after applying PR 179).

It looks like this bug is a typo: https://github.com/Samsung/walrus/blob/main/src/jit/ByteCodeParser.cpp#L594 Instruction* instr = compiler->append(byteCode, Instruction::ExtractLaneSIMD, opcode, 1, 1);

The extract lane has 1 input and 1 output argument, not two input and zero output.

I got this afterwards:

step 2 - start prediction
Tensor {
  kept: false,
  isDisposedInternal: false,
  shape: [ 1, 1001 ],
  dtype: 'float32',
  size: 1001,
  strides: [ 1001 ],
  dataId: { id: 423 },
  id: 212,
  rankType: '2',
  scopeId: 1
}
tf: 53.845s

I hope this is correct.

[zherczeg]

Two bugs are fixed in #185

[clover2123]

@zherczeg Thank you for the quick patch!