Modifications to support building w/ CUDA 10.2, NCCL for 10.2 & Intel…

… MPI libs

Build string: bazel build --config=mkl --config=v2 --config=opt //tensorflow/tools/pip_package:build_pip_package

tensorflow/lite/experimental/ruy/platform.h

@@ -53,12 +53,14 @@ limitations under the License.
 // compilation.
 //
 // TODO(b/138433137) Select AVX-512 at runtime rather than via compile options.
-#if defined(__AVX512F__) && defined(__AVX512DQ__) && defined(__AVX512CD__) && \
-    defined(__AVX512BW__) && defined(__AVX512VL__)
-#define RUY_DONOTUSEDIRECTLY_AVX512 1
-#else
+/* #if defined(__AVX512F__) && defined(__AVX512DQ__) && defined(__AVX512CD__) && \ */
+/*     defined(__AVX512BW__) && defined(__AVX512VL__) */
+/* #define RUY_DONOTUSEDIRECTLY_AVX512 1 */
+/* #else */
+/* #define RUY_DONOTUSEDIRECTLY_AVX512 0 */
+/* #endif */
+
 #define RUY_DONOTUSEDIRECTLY_AVX512 0
-#endif
 
 // Detect APPLE
 #ifdef __APPLE__

third_party/gpus/find_cuda_config.py

@@ -176,6 +176,7 @@ def _header_paths():
       "include/*-linux-gnu",
       "extras/CUPTI/include",
       "include/cuda/CUPTI",
+      "targets/x86_64-linux/include",
   ]
 
 

third_party/mpi/mpi.bzl

@@ -2,7 +2,7 @@
 #based on the configuration options return one or the other
 
 def mpi_hdr():
-    MPI_LIB_IS_OPENMPI = True
+    MPI_LIB_IS_OPENMPI = False
     hdrs = []
     if MPI_LIB_IS_OPENMPI:
         hdrs = ["mpi.h", "mpi_portable_platform.h"]  #When using OpenMPI

third_party/nccl/build_defs.bzl.tpl

@@ -113,7 +113,7 @@ def _device_link_impl(ctx):
             "--cmdline=--compile-only",
             "--link",
             "--compress-all",
-            "--bin2c-path=%s" % bin2c.dirname,
+## Removing for CUDA 10.2 compatibility "--bin2c-path=%s" % bin2c.dirname,
             "--create=%s" % tmp_fatbin.path,
             "--embedded-fatbin=%s" % fatbin_h.path,
         ] + images,