TRT INT8 unsupported hardware fix

src/operator/subgraph/tensorrt/onnx_to_tensorrt.cc

@@ -143,6 +143,7 @@ std::tuple<unique_ptr<nvinfer1::ICudaEngine>,
       builder_config->setInt8Calibrator(calibrator);
     } else {
       LOG(WARNING) << "TensorRT can't use int8 on this platform";
+      calibrator->setDone();
       calibrator = nullptr;
     }
   }
@@ -177,6 +178,7 @@ std::tuple<unique_ptr<nvinfer1::ICudaEngine>,
       trt_builder->setInt8Calibrator(calibrator);
     } else {
       LOG(WARNING) << "TensorRT can't use int8 on this platform";
+      calibrator->setDone();
       calibrator = nullptr;
     }
   }

src/operator/subgraph/tensorrt/tensorrt_int8_calibrator.cc

@@ -118,6 +118,10 @@ void TRTInt8Calibrator::writeCalibrationCache(const void* ptr,
           << " length=" << length;
 }
 
+void TRTInt8Calibrator::setDone() {
+  done_ = true;
+}
+
 void TRTInt8Calibrator::waitAndSetDone() {
   std::unique_lock<std::mutex> lk(mutex_);
   cv_.wait(lk, [&]{ return (!batch_is_set_ && !calib_running_) || done_; });

src/operator/subgraph/tensorrt/tensorrt_int8_calibrator.h

@@ -75,6 +75,8 @@ struct TRTInt8Calibrator : public nvinfer1::IInt8EntropyCalibrator2 {
   // TODO(spanev): determine if we need to serialize it
   const std::string& getCalibrationTableAsString() { return calibration_table_; }
 
+  void setDone();
+
   void waitAndSetDone();
 
   bool isCacheEmpty();

tests/python/tensorrt/test_tensorrt.py

@@ -143,12 +143,6 @@ def get_top1(logits):
 
 def test_tensorrt_symbol_int8():
     ctx = mx.gpu(0)
-    cuda_arch = get_cuda_compute_capability(ctx)
-    cuda_arch_min = 70
-    if cuda_arch < cuda_arch_min:
-        print('Bypassing test_tensorrt_symbol_int8 on cuda arch {}, need arch >= {}).'.format(
-              cuda_arch, cuda_arch_min))
-        return
 
     # INT8 engine output are not lossless, so we don't expect numerical uniformity,
     # but we have to compare the TOP1 metric