From 5bcd92f2974d94e8add31816d3b9d48a42289500 Mon Sep 17 00:00:00 2001
From: Ivan Shapovalov <intelfx@intelfx.name>
Date: Sat, 26 Nov 2016 20:53:42 +0300
Subject: Routine, Cache: generalize, reduce amount of copying in fast path

Implement a generalized Cache<K, V>. Two variants are provided: the
first one is based on std::map, using C++14-specific transparent
std::less<> and generalized std::map::find() to allow searching by tuple
of references. The second one is based on std::vector and O(n) lookup,
but remains C++11-compliant.
---
 src/routines/level3/xgemm.cpp  | 18 ++++++------------
 src/routines/level3/xhemm.cpp  |  3 +--
 src/routines/level3/xhemm.hpp  |  1 +
 src/routines/level3/xher2k.cpp | 17 +++++++----------
 src/routines/level3/xherk.cpp  | 13 +++++--------
 src/routines/level3/xsymm.cpp  | 15 +++++++--------
 src/routines/level3/xsymm.hpp  |  1 +
 src/routines/level3/xsyr2k.cpp | 13 +++++--------
 src/routines/level3/xsyrk.cpp  | 11 ++++-------
 src/routines/level3/xtrmm.cpp  |  3 +--
 src/routines/level3/xtrmm.hpp  |  1 +
 11 files changed, 39 insertions(+), 57 deletions(-)

(limited to 'src/routines/level3')

diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp
index 0015b629..7bd388c1 100644
--- a/src/routines/level3/xgemm.cpp
+++ b/src/routines/level3/xgemm.cpp
@@ -150,9 +150,6 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
   const auto c_one_i = (c_want_rotated) ? n_ceiled : m_ceiled;
   const auto c_two_i = (c_want_rotated) ? m_ceiled : n_ceiled;
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   // Determines whether or not temporary matrices are needed
   auto a_no_temp = a_one == a_one_i && a_two == a_two_i && a_ld == a_one && a_offset == 0 &&
                    a_do_transpose == false && a_conjugate == false;
@@ -178,7 +175,7 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
                            a_one, a_two, a_ld, a_offset, a_buffer,
                            a_one_i, a_two_i, a_one_i, 0, a_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, a_do_transpose, a_conjugate);
     eventWaitList.push_back(eventProcessA);
   }
@@ -189,7 +186,7 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
                            b_one, b_two, b_ld, b_offset, b_buffer,
                            b_one_i, b_two_i, b_one_i, 0, b_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, b_do_transpose, b_conjugate);
     eventWaitList.push_back(eventProcessB);
   }
@@ -200,13 +197,13 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
                            c_one, c_two, c_ld, c_offset, c_buffer,
                            c_one_i, c_two_i, c_one_i, 0, c_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, c_do_transpose, false);
     eventWaitList.push_back(eventProcessC);
   }
 
   // Retrieves the Xgemm kernel from the compiled binary
-  auto kernel = Kernel(program, "Xgemm");
+  auto kernel = Kernel(program_, "Xgemm");
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(m_ceiled));
@@ -236,7 +233,7 @@ void Xgemm<T>::GemmIndirect(const size_t m, const size_t n, const size_t k,
     PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
                            c_one_i, c_two_i, c_one_i, 0, c_temp,
                            c_one, c_two, c_ld, c_offset, c_buffer,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            false, c_do_transpose, false);
   }
 }
@@ -255,13 +252,10 @@ void Xgemm<T>::GemmDirect(const size_t m, const size_t n, const size_t k,
                           const bool a_do_transpose, const bool b_do_transpose, const bool c_do_transpose,
                           const bool a_conjugate, const bool b_conjugate) {
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   // Retrieves the proper XgemmDirect kernel from the compiled binary
   const auto name = (a_do_transpose) ? (b_do_transpose ? "XgemmDirectTT" : "XgemmDirectTN") :
                                        (b_do_transpose ? "XgemmDirectNT" : "XgemmDirectNN");
-  auto kernel = Kernel(program, name);
+  auto kernel = Kernel(program_, name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(m));
diff --git a/src/routines/level3/xhemm.cpp b/src/routines/level3/xhemm.cpp
index e5b1502a..8629f3de 100644
--- a/src/routines/level3/xhemm.cpp
+++ b/src/routines/level3/xhemm.cpp
@@ -58,8 +58,7 @@ void Xhemm<T>::DoHemm(const Layout layout, const Side side, const Triangle trian
 
   // Creates a general matrix from the hermitian matrix to be able to run the regular Xgemm
   // routine afterwards
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the arguments for the hermitian-to-squared kernel
   kernel.SetArgument(0, static_cast<int>(k));
diff --git a/src/routines/level3/xhemm.hpp b/src/routines/level3/xhemm.hpp
index 2385706e..7c011915 100644
--- a/src/routines/level3/xhemm.hpp
+++ b/src/routines/level3/xhemm.hpp
@@ -30,6 +30,7 @@ class Xhemm: public Xgemm<T> {
   using Xgemm<T>::queue_;
   using Xgemm<T>::context_;
   using Xgemm<T>::device_;
+  using Xgemm<T>::program_;
   using Xgemm<T>::db_;
   using Xgemm<T>::DoGemm;
 
diff --git a/src/routines/level3/xher2k.cpp b/src/routines/level3/xher2k.cpp
index ee3bb8b8..2aed2781 100644
--- a/src/routines/level3/xher2k.cpp
+++ b/src/routines/level3/xher2k.cpp
@@ -81,9 +81,6 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
   // Decides which kernel to run: the upper-triangular or lower-triangular version
   auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower";
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   // Determines whether or not temporary matrices are needed
   auto a1_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
                     ab_rotated == false && ab_conjugate == false;
@@ -116,7 +113,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA1.pointer(), emptyEventList,
                            ab_one, ab_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a1_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, ab_conjugate);
     eventWaitList.push_back(eventProcessA1);
   }
@@ -125,7 +122,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA2.pointer(), emptyEventList,
                            ab_one, ab_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a2_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, !ab_conjugate);
     eventWaitList.push_back(eventProcessA2);
   }
@@ -134,7 +131,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB1.pointer(), emptyEventList,
                            ab_one, ab_two, b_ld, b_offset, b_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, b1_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, ab_conjugate);
     eventWaitList.push_back(eventProcessB1);
   }
@@ -143,7 +140,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB2.pointer(), emptyEventList,
                            ab_one, ab_two, b_ld, b_offset, b_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, b2_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, !ab_conjugate);
     eventWaitList.push_back(eventProcessB2);
   }
@@ -154,12 +151,12 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
   PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
                          n, n, c_ld, c_offset, c_buffer,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          true, c_rotated, false);
   eventWaitList.push_back(eventProcessC);
 
   // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(n_ceiled));
@@ -201,7 +198,7 @@ void Xher2k<T,U>::DoHer2k(const Layout layout, const Triangle triangle, const Tr
   PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
                          n, n, c_ld, c_offset, c_buffer,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          false, c_rotated, false, upper, lower, true);
 }
 
diff --git a/src/routines/level3/xherk.cpp b/src/routines/level3/xherk.cpp
index ae8e9324..d982859e 100644
--- a/src/routines/level3/xherk.cpp
+++ b/src/routines/level3/xherk.cpp
@@ -79,9 +79,6 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
   // Decides which kernel to run: the upper-triangular or lower-triangular version
   auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower";
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   // Determines whether or not temporary matrices are needed
   auto a_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
                    a_rotated == false && a_conjugate == false;
@@ -109,7 +106,7 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
                            a_one, a_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, a_rotated, a_conjugate);
     eventWaitList.push_back(eventProcessA);
   }
@@ -118,7 +115,7 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
                            a_one, a_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, b_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, a_rotated, b_conjugate);
     eventWaitList.push_back(eventProcessB);
   }
@@ -129,12 +126,12 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
   PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
                          n, n, c_ld, c_offset, c_buffer,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          true, c_rotated, false);
   eventWaitList.push_back(eventProcessC);
 
   // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(n_ceiled));
@@ -163,7 +160,7 @@ void Xherk<T,U>::DoHerk(const Layout layout, const Triangle triangle, const Tran
   PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
                          n, n, c_ld, c_offset, c_buffer,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          false, c_rotated, false, upper, lower, true);
 }
 
diff --git a/src/routines/level3/xsymm.cpp b/src/routines/level3/xsymm.cpp
index d7f771d1..969edfc8 100644
--- a/src/routines/level3/xsymm.cpp
+++ b/src/routines/level3/xsymm.cpp
@@ -30,12 +30,12 @@ Xsymm<T>::Xsymm(Queue &queue, EventPointer event, const std::string &name):
 // The main routine
 template <typename T>
 void Xsymm<T>::DoSymm(const Layout layout, const Side side, const Triangle triangle,
-                            const size_t m, const size_t n,
-                            const T alpha,
-                            const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
-                            const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld,
-                            const T beta,
-                            const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld) {
+                      const size_t m, const size_t n,
+                      const T alpha,
+                      const Buffer<T> &a_buffer, const size_t a_offset, const size_t a_ld,
+                      const Buffer<T> &b_buffer, const size_t b_offset, const size_t b_ld,
+                      const T beta,
+                      const Buffer<T> &c_buffer, const size_t c_offset, const size_t c_ld) {
 
   // Makes sure all dimensions are larger than zero
   if ((m == 0) || (n == 0) ) { throw BLASError(StatusCode::kInvalidDimension); }
@@ -58,8 +58,7 @@ void Xsymm<T>::DoSymm(const Layout layout, const Side side, const Triangle trian
 
   // Creates a general matrix from the symmetric matrix to be able to run the regular Xgemm
   // routine afterwards
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the arguments for the symmetric-to-squared kernel
   kernel.SetArgument(0, static_cast<int>(k));
diff --git a/src/routines/level3/xsymm.hpp b/src/routines/level3/xsymm.hpp
index ee965364..7a584560 100644
--- a/src/routines/level3/xsymm.hpp
+++ b/src/routines/level3/xsymm.hpp
@@ -32,6 +32,7 @@ class Xsymm: public Xgemm<T> {
   using Xgemm<T>::queue_;
   using Xgemm<T>::context_;
   using Xgemm<T>::device_;
+  using Xgemm<T>::program_;
   using Xgemm<T>::db_;
   using Xgemm<T>::DoGemm;
 
diff --git a/src/routines/level3/xsyr2k.cpp b/src/routines/level3/xsyr2k.cpp
index cb0e0461..fdef43dc 100644
--- a/src/routines/level3/xsyr2k.cpp
+++ b/src/routines/level3/xsyr2k.cpp
@@ -77,9 +77,6 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
   // Decides which kernel to run: the upper-triangular or lower-triangular version
   auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower";
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   // Determines whether or not temporary matrices are needed
   auto a_no_temp = ab_one == n_ceiled && ab_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
                    ab_rotated == false;
@@ -103,7 +100,7 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
                            ab_one, ab_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, false);
     eventWaitList.push_back(eventProcessA);
   }
@@ -112,7 +109,7 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessB.pointer(), emptyEventList,
                            ab_one, ab_two, b_ld, b_offset, b_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, b_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, ab_rotated, false);
     eventWaitList.push_back(eventProcessB);
   }
@@ -123,12 +120,12 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
   PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
                          n, n, c_ld, c_offset, c_buffer,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          true, c_rotated, false);
   eventWaitList.push_back(eventProcessC);
 
   // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(n_ceiled));
@@ -168,7 +165,7 @@ void Xsyr2k<T>::DoSyr2k(const Layout layout, const Triangle triangle, const Tran
   PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
                          n, n, c_ld, c_offset, c_buffer,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          false, c_rotated, false, upper, lower, false);
 }
 
diff --git a/src/routines/level3/xsyrk.cpp b/src/routines/level3/xsyrk.cpp
index bd6c4b25..9588c28c 100644
--- a/src/routines/level3/xsyrk.cpp
+++ b/src/routines/level3/xsyrk.cpp
@@ -74,9 +74,6 @@ void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transp
   // Decides which kernel to run: the upper-triangular or lower-triangular version
   auto kernel_name = (triangle == Triangle::kUpper) ? "XgemmUpper" : "XgemmLower";
 
-  // Loads the program from the database
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-
   // Determines whether or not temporary matrices are needed
   auto a_no_temp = a_one == n_ceiled && a_two == k_ceiled && a_ld == n_ceiled && a_offset == 0 &&
                    a_rotated == false;
@@ -97,7 +94,7 @@ void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transp
     PadCopyTransposeMatrix(queue_, device_, db_, eventProcessA.pointer(), emptyEventList,
                            a_one, a_two, a_ld, a_offset, a_buffer,
                            n_ceiled, k_ceiled, n_ceiled, 0, a_temp,
-                           ConstantOne<T>(), program,
+                           ConstantOne<T>(), program_,
                            true, a_rotated, false);
     eventWaitList.push_back(eventProcessA);
   }
@@ -108,12 +105,12 @@ void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transp
   PadCopyTransposeMatrix(queue_, device_, db_, eventProcessC.pointer(), emptyEventList,
                          n, n, c_ld, c_offset, c_buffer,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          true, c_rotated, false);
   eventWaitList.push_back(eventProcessC);
 
   // Retrieves the XgemmUpper or XgemmLower kernel from the compiled binary
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the kernel arguments
   kernel.SetArgument(0, static_cast<int>(n_ceiled));
@@ -142,7 +139,7 @@ void Xsyrk<T>::DoSyrk(const Layout layout, const Triangle triangle, const Transp
   PadCopyTransposeMatrix(queue_, device_, db_, event_, eventWaitList,
                          n_ceiled, n_ceiled, n_ceiled, 0, c_temp,
                          n, n, c_ld, c_offset, c_buffer,
-                         ConstantOne<T>(), program,
+                         ConstantOne<T>(), program_,
                          false, c_rotated, false, upper, lower, false);
 }
 
diff --git a/src/routines/level3/xtrmm.cpp b/src/routines/level3/xtrmm.cpp
index ed810e72..02c295ac 100644
--- a/src/routines/level3/xtrmm.cpp
+++ b/src/routines/level3/xtrmm.cpp
@@ -70,8 +70,7 @@ void Xtrmm<T>::DoTrmm(const Layout layout, const Side side, const Triangle trian
 
   // Creates a general matrix from the triangular matrix to be able to run the regular Xgemm
   // routine afterwards
-  const auto program = GetProgramFromCache(context_, PrecisionValue<T>(), routine_name_);
-  auto kernel = Kernel(program, kernel_name);
+  auto kernel = Kernel(program_, kernel_name);
 
   // Sets the arguments for the triangular-to-squared kernel
   kernel.SetArgument(0, static_cast<int>(k));
diff --git a/src/routines/level3/xtrmm.hpp b/src/routines/level3/xtrmm.hpp
index 967bf132..e77b7214 100644
--- a/src/routines/level3/xtrmm.hpp
+++ b/src/routines/level3/xtrmm.hpp
@@ -31,6 +31,7 @@ class Xtrmm: public Xgemm<T> {
   using Xgemm<T>::queue_;
   using Xgemm<T>::context_;
   using Xgemm<T>::device_;
+  using Xgemm<T>::program_;
   using Xgemm<T>::db_;
   using Xgemm<T>::DoGemm;
 
-- 
cgit v1.2.3