[XLA] [StreamExecutor] Tune GEMMs when possible.

cublas 8 adds the cublasGemmEx function, which lets you specify an
explicit "algorithm" for the computation.  This functions as an opaque
tuning hint to cublas.

This patch adds support for cublasGemmEx to StreamExecutor, and wires up
XLA's GemmThunk to use the new function.

This patch does not add GEMM autotuning support in TensorFlow proper,
only XLA.
Change: 149068961

1 files changed, 17 insertions, 0 deletions

diff --git a/tensorflow/stream_executor/blas.cc b/tensorflow/stream_executor/blas.cc
index 239e3fce01..a59a1dda71 100644
--- a/tensorflow/stream_executor/blas.cc
+++ b/tensorflow/stream_executor/blas.cc
@@ -67,6 +67,23 @@ string SideString(Side s) {
   }
 }
 
+string ComputationTypeString(ComputationType ty) {
+  switch (ty) {
+    case ComputationType::kF16:
+      return "f16";
+    case ComputationType::kF32:
+      return "f32";
+    case ComputationType::kF64:
+      return "f64";
+    case ComputationType::kComplexF32:
+      return "complex f32";
+    case ComputationType::kComplexF64:
+      return "complex f64";
+    default:
+      LOG(FATAL) << "Unknown ComputationType " << static_cast<int32>(ty);
+  }
+}
+
 }  // namespace blas
 }  // namespace gputools
 }  // namespace perftools