From ae07801dd8d295657f28b006e1e4999edf835052 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <eugene.zhulenev@gmail.com>
Date: Wed, 18 Dec 2019 20:07:00 +0000
Subject: Tensor block evaluation cost model

---
 unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

(limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h')

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
index 620c8741c..3408f90d1 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBroadcasting.h
@@ -620,12 +620,10 @@ struct TensorEvaluator<const TensorBroadcastingOp<Broadcast, ArgType>, Device>
   internal::TensorBlockResourceRequirements getResourceRequirements() const {
     // TODO(wuke): Targeting L1 size is 30% faster than targeting L{-1} on large
     // tensors. But this might need further tuning.
-    const size_t target_block_size = numext::maxi<size_t>(
-        1, m_device.firstLevelCacheSize() / sizeof(Scalar));
-
+    const size_t target_size = m_device.firstLevelCacheSize();
     return internal::TensorBlockResourceRequirements::merge(
-        {internal::TensorBlockShapeType::kSkewedInnerDims, target_block_size},
-        m_impl.getResourceRequirements());
+        m_impl.getResourceRequirements(),
+        internal::TensorBlockResourceRequirements::skewed<Scalar>(target_size));
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlock
-- 
cgit v1.2.3