address PR comments (#3312)

* address PR comments

* PR comments

* PR comments

* disable logging

* typo

Co-authored-by: Ethan Tao <ettao@microsoft.com>

include/onnxruntime/core/framework/data_types.h

@@ -55,20 +55,20 @@ union MLFloat16 {
   MLFloat16() : val(0) {}
 
   // Taken from https://stackoverflow.com/a/60047308/12627730
-  float as_float(uint32_t x) const {
+  float AsFloat(uint32_t x) const {
     float out = 0.0f;
     std::memcpy(&out, &x, sizeof(x));
     return out;
   }
 
   // Taken from https://stackoverflow.com/a/60047308/12627730
-  uint32_t as_uint(float x) const {
+  uint32_t AsUint(float x) const {
     uint32_t out = 0;
     std::memcpy(&out, &x, sizeof(x));
     return out;
   }
 
-  float half_to_float(const uint16_t x) const {
+  float HalfToFloat(const uint16_t x) const {
     uint16_t half = x;
     if (endian::native == endian::big) {
       // Taken from https://stackoverflow.com/a/2182184/12627730
@@ -81,7 +81,7 @@ union MLFloat16 {
     const uint32_t e = (half & 0x7C00) >> 10;  // exponent
     const uint32_t m = (half & 0x03FF) << 13;  // mantissa
     // evil log2 bit hack to count leading zeros in denormalized format
-    const uint32_t v = as_uint(static_cast<float>(m)) >> 23;
+    const uint32_t v = AsUint(static_cast<float>(m)) >> 23;
     uint32_t full = (half & 0x8000) << 16 | (e != 0) * ((e + 112) << 23 | m) |
                     ((e == 0) & (m != 0)) * ((v - 37) << 23 | ((m << (150 - v)) & 0x007FE000));  // sign : normalized : denormalized
 
@@ -93,11 +93,11 @@ union MLFloat16 {
              ((full << 24) & 0xff000000);  // byte 0 to byte 3
     }
 
-    return as_float(full);
+    return AsFloat(full);
   }
 
   operator float() const {
-    return half_to_float(val);
+    return HalfToFloat(val);
   }
 };
 

include/onnxruntime/core/framework/execution_provider.h

@@ -97,7 +97,7 @@ class IExecutionProvider {
   virtual std::shared_ptr<KernelRegistry> GetKernelRegistry() const;
 
   /**
-     Get the device id of current excution provider
+     Get the device id of current execution provider
   */
   virtual int GetDeviceId() const { return -1; };
 

onnxruntime/contrib_ops/cpu/transpose_matmul.cc

@@ -25,9 +25,7 @@ TransposeMatMul::TransposeMatMul(const OpKernelInfo& info)
 }
 
 Status TransposeMatMul::Compute(OpKernelContext* context) const {
-  auto* context_internal = dynamic_cast<OpKernelContextInternal*>(context);
-  ORT_ENFORCE(context_internal, "Failed to get OpKernelContextInternal instance.");
-  concurrency::ThreadPool* thread_pool = context_internal->GetOperatorThreadPool();
+  concurrency::ThreadPool* thread_pool = context->GetOperatorThreadPool();
 
   const Tensor* A = context->Input<Tensor>(0);
   const Tensor* B = context->Input<Tensor>(1);
@@ -40,9 +38,6 @@ Status TransposeMatMul::Compute(OpKernelContext* context) const {
   ORT_RETURN_IF_ERROR(helper.Compute(A->Shape(), B->Shape(), trans_a, trans_b));
 
   Tensor* Y = context->Output(0, helper.OutputShape());
-  if (Y->DataType() != DataTypeImpl::GetType<float>()) {
-    ORT_NOT_IMPLEMENTED("TransposeMatMul CPU operator only supports float inputs.");
-  }
 
   const size_t num_offsets = helper.OutputOffsets().size();
   for (size_t i = 0; i < num_offsets; ++i) {

onnxruntime/core/framework/execution_frame.cc

@@ -222,8 +222,8 @@ ExecutionFrame::ExecutionFrame(const std::vector<int>& feed_mlvalue_idxs, const
                              : nullptr;
           buffers_[mem_patterns_->locations[i]] = BufferUniquePtr(buffer, alloc);
 
-          // comment out following line to see the size of activation
-          // printf("Allocated memory for activations, size: %zu\n", mem_patterns_->patterns[i].PeakSize());
+          // log size of activation. Keep it commented out for now to avoid log flooding.
+          // VLOGS(session_state_.Logger(), 1) << "Allocated memory for activations, size: " << mem_patterns_->patterns[i].PeakSize();
         }
       }
     }
@@ -343,8 +343,11 @@ Status ExecutionFrame::AllocateMLValueTensorPreAllocateBuffer(OrtValue& ort_valu
 
     // be generous and use the buffer if it's large enough. log a warning though as it indicates a bad model
     if (buffer_num_elements >= required_num_elements) {
-      // Todo: View Operator is reusing the buffer bigger than the required size. Disabling warming message for now.
-      //LOGS_DEFAULT(WARNING) << message;
+      // View Operator is reusing the buffer bigger than the required size. 
+      // Disabling warning message for now. The op is in the process of being deprecated.
+#ifndef ENABLE_TRAINING
+      LOGS(session_state_.Logger(), WARNING) << message;
+#endif
     } else {
       return ORT_MAKE_STATUS(ONNXRUNTIME, FAIL, message);
     }

onnxruntime/test/providers/cpu/controlflow/scan_test.cc

@@ -275,9 +275,7 @@ static void RunTest_v8(const std::string test_name, int64_t batch_size, int64_t
   Model model(test_name, false, DefaultLoggingManager().DefaultLogger());
   auto& graph = model.MainGraph();
   auto status = CreateSubgraph(graph, options, options.add_bad_shape ? failure_message : "");
-  if (!status.IsOK()) {
-    return;
-  }
+  ASSERT_STATUS_OK(status);
   auto& proto = graph.ToGraphProto();
 
   ScanOpTester test{8};