From 383d1cc2ed76d1757a45cd0b2d6559dee7e2ee1b Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Fri, 20 Nov 2015 11:09:46 -0800
Subject: Added proper support for fast 64bit integer division on CUDA

---
 unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h | 31 +++++++++--------------
 1 file changed, 12 insertions(+), 19 deletions(-)

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

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
index 058fb2c42..81c661269 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorIntDiv.h
@@ -61,13 +61,8 @@ namespace {
 
   template <typename T>
   struct DividerTraits {
-#if defined(__SIZEOF_INT128__) && !defined(__CUDACC__)
     typedef typename UnsignedTraits<T>::type type;
     static const int N = sizeof(T) * 8;
-#else
-    typedef uint32_t type;
-    static const int N = 32;
-#endif
   };
 
   template <typename T>
@@ -79,40 +74,38 @@ namespace {
 #endif
   }
 
+  template <typename T>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
 #if defined(__CUDA_ARCH__)
- template <typename T>
- EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
     return __umul64hi(a, b);
- }
-#else
-  template <typename T>
-  EIGEN_ALWAYS_INLINE uint64_t muluh(const uint64_t a, const T b) {
-#if defined(__SIZEOF_INT128__) && !defined(__CUDACC__)
+#elif defined(__SIZEOF_INT128__)
     __uint128_t v = static_cast<__uint128_t>(a) * static_cast<__uint128_t>(b);
     return static_cast<uint64_t>(v >> 64);
 #else
-    EIGEN_STATIC_ASSERT(sizeof(T) == 4, YOU_MADE_A_PROGRAMMING_MISTAKE);
-    return (a * b) >> 32;
+    return (TensorUInt128<static_val<0>, uint64_t>(a) * TensorUInt128<static_val<0>, uint64_t>(b)).upper();
 #endif
   }
-#endif
 
   template <int N, typename T>
   struct DividerHelper {
-    static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t computeMultiplier (const int log_div, const T divider) {
+    static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint32_t computeMultiplier(const int log_div, const T divider) {
       EIGEN_STATIC_ASSERT(N == 32, YOU_MADE_A_PROGRAMMING_MISTAKE);
       return static_cast<uint32_t>((static_cast<uint64_t>(1) << (N+log_div)) / divider - (static_cast<uint64_t>(1) << N) + 1);
     }
   };
 
-#if defined(__SIZEOF_INT128__) && !defined(__CUDACC__)
   template <typename T>
   struct DividerHelper<64, T> {
-    static EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
+    static EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE uint64_t computeMultiplier(const int log_div, const T divider) {
+#if defined(__SIZEOF_INT128__) && !defined(__CUDA_ARCH__)
       return static_cast<uint64_t>((static_cast<__uint128_t>(1) << (64+log_div)) / static_cast<__uint128_t>(divider) - (static_cast<__uint128_t>(1) << 64) + 1);
+#else
+      const uint64_t shift = 1ULL << log_div;
+      TensorUInt128<uint64_t, uint64_t> result = (TensorUInt128<uint64_t, static_val<0> >(shift, 0) / TensorUInt128<static_val<0>, uint64_t>(divider) - TensorUInt128<static_val<1>, static_val<0> >(1, 0) + TensorUInt128<static_val<0>, static_val<1> >(1));
+      return static_cast<uint64_t>(result);
+#endif
     }
   };
-#endif
 }
 
 
-- 
cgit v1.2.3