The following PR/commit introduces a build error on the ROCm platform https://github.com/tensorflow/tensorflow/pull/38802 The error is caused by a call to the `CsrgemmBufferSize` routine which only exists on the CUDA side. The call to it was not guarded by the same #if block that guards the function declaration + definition. Adding the missing #if block fixes the issue. This PR also adds some explicit `GOOGLE_CUDA &&` and `|| TENSORFLOW_USE_ROCM` conditions to some `#if` to make things clear. |
||
---|---|---|
.. | ||
add_op.cc | ||
BUILD | ||
conj_op.cc | ||
csr_sparse_matrix_to_dense_op.cc | ||
csr_sparse_matrix_to_sparse_tensor_op.cc | ||
dense_to_csr_sparse_matrix_op.cc | ||
kernels_gpu.cu.cc | ||
kernels_test.cc | ||
kernels.cc | ||
kernels.h | ||
mat_mul_op.cc | ||
mul_op.cc | ||
nnz_op.cc | ||
softmax_op.cc | ||
sparse_cholesky_op.cc | ||
sparse_mat_mul_op.cc | ||
sparse_matrix_components_op.cc | ||
sparse_matrix.cc | ||
sparse_matrix.h | ||
sparse_ordering_amd_op.cc | ||
sparse_tensor_to_csr_sparse_matrix_op.cc | ||
transpose_op.cc | ||
transpose_op.h | ||
zeros_op.cc | ||
zeros_op.h |