DPCT1097

Intel® DPC++ Compatibility Tool Developer Guide and Reference

Download PDF

ID 768918

Date 6/24/2024

Version

Public

A newer version of this document is available. Customers should click here to go to the newest version.

DPCT1097

Message

The function <backward function name> may require the workspace used to save intermediate results from function <forward function name>. By default, a workspace from engine_ext is selected according to the source data pointer, but this may be incorrect and cause a workspace data race. You may need to rewrite this code.

Detailed Help

You can manually pass a dnnl::memory object generated from the forward function to the backward function.

For example, this original CUDA* code:

void test(cudnnHandle_t handle, cudnnTensorDescriptor_t dataTensor,
          cudnnTensorDescriptor_t outTensor,
          cudnnTensorDescriptor_t diffdataTensor,
          cudnnTensorDescriptor_t diffoutTensor, float *data, float *out,
          float *diffdata, float *diffout, float alpha, float beta,
          cudnnLRNDescriptor_t desc) {
  ...
  cudnnLRNCrossChannelForward(handle, desc, CUDNN_LRN_CROSS_CHANNEL_DIM1,
                              &alpha, dataTensor, data, &beta, outTensor, out);
  ...
  cudnnLRNCrossChannelBackward(handle, desc, CUDNN_LRN_CROSS_CHANNEL_DIM1,
                               &alpha, outTensor, out, diffoutTensor, diffout,
                               dataTensor, data, &beta, diffdataTensor,
                               diffdata);
  ...
}

results in the following migrated SYCL* code:

void test(dpct::dnnl::engine_ext handle, dpct::dnnl::memory_desc_ext dataTensor,
          dpct::dnnl::memory_desc_ext outTensor,
          dpct::dnnl::memory_desc_ext diffdataTensor,
          dpct::dnnl::memory_desc_ext diffoutTensor, float *data, float *out,
          float *diffdata, float *diffout, float alpha, float beta,
          dpct::dnnl::lrn_desc desc) {
  ...
  handle.async_lrn_forward(desc, alpha, dataTensor, data, beta, outTensor, out);
  ...
  /*
  DPCT1097:0: The function "async_lrn_backward" may require the workspace used
  to save intermediate results from function "async_lrn_forward". By default, a
  workspace from engine_ext is selected according to the source data pointer,
  but this may be incorrect and cause a workspace data race. You may need to
  rewrite this code.
  */
  handle.async_lrn_backward(desc, alpha, outTensor, out, diffoutTensor, diffout,
                            dataTensor, data, beta, diffdataTensor, diffdata);
  ...
}

which is manually adjusted to:

void test(dpct::dnnl::engine_ext handle, dpct::dnnl::memory_desc_ext dataTensor,
          dpct::dnnl::memory_desc_ext outTensor,
          dpct::dnnl::memory_desc_ext diffdataTensor,
          dpct::dnnl::memory_desc_ext diffoutTensor, float *data, float *out,
          float *diffdata, float *diffout, float alpha, float beta,
          dpct::dnnl::lrn_desc desc) {
  ...
  dnnl::memory workspace;
  handle.async_lrn_forward(desc, alpha, dataTensor, data, beta, outTensor, out,
                           &workspace);
  ...
  handle.async_lrn_backward(desc, alpha, outTensor, out, diffoutTensor, diffout,
                            dataTensor, data, beta, diffdataTensor, diffdata,
                            &workspace);
  ...
}

Suggestions to Fix

You may need to adjust the original code.

Select Your Language

Using Intel.com Search

Quick Links

Recent Searches

Advanced Search

Only search in

Intel® DPC++ Compatibility Tool Developer Guide and Reference

DPCT1097

Message

Detailed Help

Suggestions to Fix