UnitTests: add missing fences to fix UVM issue on Kepler

fixes kokkos#51
mndevec · Sep 20, 2017 · 7654277 · 7654277
1 parent cd5d9fb
commit 7654277
Show file tree

Hide file tree

Showing 24 changed files with 90 additions and 1 deletion.
diff --git a/unit_test/Makefile b/unit_test/Makefile
@@ -61,6 +61,7 @@ TEST_HEADERS += $(wildcard $(KOKKOSKERNELS_SRC_PATH)/unit_test/blas/*.hpp)
 TEST_HEADERS += $(wildcard $(KOKKOSKERNELS_SRC_PATH)/unit_test/sparse/*.hpp)
 TEST_HEADERS += $(wildcard $(KOKKOSKERNELS_SRC_PATH)/unit_test/graph/*.hpp)
 TEST_HEADERS += $(wildcard $(KOKKOSKERNELS_SRC_PATH)/unit_test/common/*.hpp)
+TEST_HEADERS += $(wildcard $(KOKKOSKERNELS_SRC_PATH)/unit_test/batched/*.hpp)
 
 
 

diff --git a/unit_test/batched/Test_Batched_SerialGemm.hpp b/unit_test/batched/Test_Batched_SerialGemm.hpp
@@ -80,6 +80,8 @@ namespace Test {
  Kokkos::fill_random(b0, random, value_type(1.0));
  Kokkos::fill_random(c0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(a1, a0);
  Kokkos::deep_copy(b1, b0);
  Kokkos::deep_copy(c1, c0);
@@ -90,6 +92,8 @@ namespace Test {
  Functor_TestBatchedSerialGemm<DeviceType,ViewType,ScalarType,
  ParamTagType,AlgoTagType>(alpha, a1, b1, beta, c1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror c0_host = Kokkos::create_mirror_view(c0);
  typename ViewType::HostMirror c1_host = Kokkos::create_mirror_view(c1);

diff --git a/unit_test/batched/Test_Batched_SerialGemv.hpp b/unit_test/batched/Test_Batched_SerialGemv.hpp
@@ -78,6 +78,8 @@ namespace Test {
  Kokkos::fill_random(b0, random, value_type(1.0));
  Kokkos::fill_random(c0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(a1, a0);
  Kokkos::deep_copy(b1, b0);
  Kokkos::deep_copy(c1, c0);
@@ -88,6 +90,8 @@ namespace Test {
  Functor_TestBatchedSerialGemv<DeviceType,ViewType,ScalarType,
  ParamTagType,AlgoTagType>(alpha, a1, b1, beta, c1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror c0_host = Kokkos::create_mirror_view(c0);
  typename ViewType::HostMirror c1_host = Kokkos::create_mirror_view(c1);

diff --git a/unit_test/batched/Test_Batched_SerialLU.hpp b/unit_test/batched/Test_Batched_SerialLU.hpp
@@ -56,11 +56,15 @@ namespace Test {
  Kokkos::Random_XorShift64_Pool<typename DeviceType::execution_space> random(13718);
  Kokkos::fill_random(a0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(a1, a0);
 
  Functor_TestBatchedSerialLU<DeviceType,ViewType,Algo::LU::Unblocked>(a0).run();
  Functor_TestBatchedSerialLU<DeviceType,ViewType,AlgoTagType>(a1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror a0_host = Kokkos::create_mirror_view(a0);
  typename ViewType::HostMirror a1_host = Kokkos::create_mirror_view(a1);

diff --git a/unit_test/batched/Test_Batched_SerialMatUtil.hpp b/unit_test/batched/Test_Batched_SerialMatUtil.hpp
@@ -94,12 +94,17 @@ namespace Test {
 
  Kokkos::Random_XorShift64_Pool<typename DeviceType::execution_space> random(13718);
  Kokkos::fill_random(a, random, value_type(1.0));
+
+ Kokkos::fence();
+
  Kokkos::deep_copy(b, a);
 
  /// test body
  Functor_TestBatchedSerialMatUtil<DeviceType,ViewType,ScalarType,NaiveTag, TestID>(alpha, a).run();
  Functor_TestBatchedSerialMatUtil<DeviceType,ViewType,ScalarType,KokkosKernelTag,TestID>(alpha, b).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror a_host = Kokkos::create_mirror_view(a);
  typename ViewType::HostMirror b_host = Kokkos::create_mirror_view(b);

diff --git a/unit_test/batched/Test_Batched_SerialTrsm.hpp b/unit_test/batched/Test_Batched_SerialTrsm.hpp
@@ -82,6 +82,8 @@ namespace Test {
  Kokkos::fill_random(a0, random, value_type(1.0));
  Kokkos::fill_random(b0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(a1, a0);
  Kokkos::deep_copy(b1, b0);
 
@@ -90,6 +92,8 @@ namespace Test {
  Functor_TestBatchedSerialTrsm<DeviceType,ViewType,ScalarType,
  ParamTagType,AlgoTagType>(alpha, a1, b1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror b0_host = Kokkos::create_mirror_view(b0);
  typename ViewType::HostMirror b1_host = Kokkos::create_mirror_view(b1);

diff --git a/unit_test/batched/Test_Batched_SerialTrsv.hpp b/unit_test/batched/Test_Batched_SerialTrsv.hpp
@@ -77,6 +77,8 @@ namespace Test {
  Kokkos::fill_random(a0, random, value_type(1.0));
  Kokkos::fill_random(b0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b0, 1.0);
 
  Kokkos::deep_copy(a1, a0);
@@ -87,6 +89,8 @@ namespace Test {
  Functor_TestBatchedSerialTrsv<DeviceType,ViewType,ScalarType,
  ParamTagType,AlgoTagType>(alpha, a1, b1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror b0_host = Kokkos::create_mirror_view(b0);
  typename ViewType::HostMirror b1_host = Kokkos::create_mirror_view(b1);

diff --git a/unit_test/batched/Test_Batched_TeamGemm.hpp b/unit_test/batched/Test_Batched_TeamGemm.hpp
@@ -86,6 +86,8 @@ namespace Test {
  Kokkos::fill_random(b0, random, value_type(1.0));
  Kokkos::fill_random(c0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(a1, a0);
  Kokkos::deep_copy(b1, b0);
  Kokkos::deep_copy(c1, c0);
@@ -96,6 +98,8 @@ namespace Test {
  Functor_TestBatchedTeamGemm<DeviceType,ViewType,ScalarType,
  ParamTagType,AlgoTagType>(alpha, a1, b1, beta, c1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror c0_host = Kokkos::create_mirror_view(c0);
  typename ViewType::HostMirror c1_host = Kokkos::create_mirror_view(c1);

diff --git a/unit_test/batched/Test_Batched_TeamGemv.hpp b/unit_test/batched/Test_Batched_TeamGemv.hpp
@@ -84,6 +84,8 @@ namespace Test {
  Kokkos::fill_random(b0, random, value_type(1.0));
  Kokkos::fill_random(c0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(a1, a0);
  Kokkos::deep_copy(b1, b0);
  Kokkos::deep_copy(c1, c0);
@@ -94,6 +96,8 @@ namespace Test {
  Functor_TestBatchedTeamGemv<DeviceType,ViewType,ScalarType,
  ParamTagType,AlgoTagType>(alpha, a1, b1, beta, c1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror c0_host = Kokkos::create_mirror_view(c0);
  typename ViewType::HostMirror c1_host = Kokkos::create_mirror_view(c1);

diff --git a/unit_test/batched/Test_Batched_TeamLU.hpp b/unit_test/batched/Test_Batched_TeamLU.hpp
@@ -63,11 +63,15 @@ namespace Test {
  Kokkos::Random_XorShift64_Pool<typename DeviceType::execution_space> random(13718);
  Kokkos::fill_random(a0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(a1, a0);
 
  Functor_TestBatchedTeamLU<DeviceType,ViewType,Algo::LU::Unblocked>(a0).run();
  Functor_TestBatchedTeamLU<DeviceType,ViewType,AlgoTagType>(a1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror a0_host = Kokkos::create_mirror_view(a0);
  typename ViewType::HostMirror a1_host = Kokkos::create_mirror_view(a1);

diff --git a/unit_test/batched/Test_Batched_TeamMatUtil.hpp b/unit_test/batched/Test_Batched_TeamMatUtil.hpp
@@ -99,12 +99,17 @@ namespace Test {
 
  Kokkos::Random_XorShift64_Pool<typename DeviceType::execution_space> random(13718);
  Kokkos::fill_random(a, random, value_type(1.0));
+
+ Kokkos::fence();
+
  Kokkos::deep_copy(b, a);
 
  /// test body
  Functor_TestBatchedTeamMatUtil<DeviceType,ViewType,ScalarType,NaiveTag, TestID>(alpha, a).run();
  Functor_TestBatchedTeamMatUtil<DeviceType,ViewType,ScalarType,KokkosKernelTag,TestID>(alpha, b).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror a_host = Kokkos::create_mirror_view(a);
  typename ViewType::HostMirror b_host = Kokkos::create_mirror_view(b);

diff --git a/unit_test/batched/Test_Batched_TeamTrsm.hpp b/unit_test/batched/Test_Batched_TeamTrsm.hpp
@@ -88,6 +88,8 @@ namespace Test {
  Kokkos::fill_random(a0, random, value_type(1.0));
  Kokkos::fill_random(b0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(a1, a0);
  Kokkos::deep_copy(b1, b0);
 
@@ -96,6 +98,8 @@ namespace Test {
  Functor_TestBatchedTeamTrsm<DeviceType,ViewType,ScalarType,
  ParamTagType,AlgoTagType>(alpha, a1, b1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror b0_host = Kokkos::create_mirror_view(b0);
  typename ViewType::HostMirror b1_host = Kokkos::create_mirror_view(b1);

diff --git a/unit_test/batched/Test_Batched_TeamTrsv.hpp b/unit_test/batched/Test_Batched_TeamTrsv.hpp
@@ -83,6 +83,8 @@ namespace Test {
  Kokkos::fill_random(a0, random, value_type(1.0));
  Kokkos::fill_random(b0, random, value_type(1.0));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b0, 1.0);
 
  Kokkos::deep_copy(a1, a0);
@@ -93,6 +95,8 @@ namespace Test {
  Functor_TestBatchedTeamTrsv<DeviceType,ViewType,ScalarType,
  ParamTagType,AlgoTagType>(alpha, a1, b1).run();
 
+ Kokkos::fence();
+
  /// for comparison send it to host
  typename ViewType::HostMirror b0_host = Kokkos::create_mirror_view(b0);
  typename ViewType::HostMirror b1_host = Kokkos::create_mirror_view(b1);

diff --git a/unit_test/blas/Test_Blas1_asum.hpp b/unit_test/blas/Test_Blas1_asum.hpp
@@ -28,7 +28,7 @@ namespace Test {
  Kokkos::Random_XorShift64_Pool<typename Device::execution_space> rand_pool(13718);
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
-
+ Kokkos::fence();
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;

diff --git a/unit_test/blas/Test_Blas1_dot.hpp b/unit_test/blas/Test_Blas1_dot.hpp
@@ -38,6 +38,8 @@ namespace Test {
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
  Kokkos::fill_random(b_b,rand_pool,ScalarB(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
  Kokkos::deep_copy(h_b_b,b_b);
 
@@ -90,6 +92,8 @@ namespace Test {
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
  Kokkos::fill_random(b_b,rand_pool,ScalarB(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
  Kokkos::deep_copy(h_b_b,b_b);
 

diff --git a/unit_test/blas/Test_Blas1_nrm1.hpp b/unit_test/blas/Test_Blas1_nrm1.hpp
@@ -29,6 +29,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;
@@ -68,6 +70,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;

diff --git a/unit_test/blas/Test_Blas1_nrm2.hpp b/unit_test/blas/Test_Blas1_nrm2.hpp
@@ -29,6 +29,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;
@@ -69,6 +71,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;

diff --git a/unit_test/blas/Test_Blas1_nrm2_squared.hpp b/unit_test/blas/Test_Blas1_nrm2_squared.hpp
@@ -29,6 +29,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;
@@ -68,6 +70,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;

diff --git a/unit_test/blas/Test_Blas1_nrminf.hpp b/unit_test/blas/Test_Blas1_nrminf.hpp
@@ -29,6 +29,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;
@@ -70,6 +72,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;

diff --git a/unit_test/blas/Test_Blas1_reciprocal.hpp b/unit_test/blas/Test_Blas1_reciprocal.hpp
@@ -45,6 +45,8 @@ namespace Test {
  Kokkos::fill_random(b_x,rand_pool,ScalarA(10));
  Kokkos::fill_random(b_y,rand_pool,ScalarB(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b_org_y,b_y);
 
  Kokkos::deep_copy(h_b_x,b_x);
@@ -95,6 +97,8 @@ namespace Test {
  Kokkos::fill_random(b_x,rand_pool,ScalarA(10));
  Kokkos::fill_random(b_y,rand_pool,ScalarB(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b_org_y,b_y);
 
  Kokkos::deep_copy(h_b_x,b_x);

diff --git a/unit_test/blas/Test_Blas1_scal.hpp b/unit_test/blas/Test_Blas1_scal.hpp
@@ -46,6 +46,8 @@ namespace Test {
  Kokkos::fill_random(b_x,rand_pool,ScalarA(10));
  Kokkos::fill_random(b_y,rand_pool,ScalarB(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b_org_y,b_y);
 
  Kokkos::deep_copy(h_b_x,b_x);
@@ -95,6 +97,8 @@ namespace Test {
  Kokkos::fill_random(b_x,rand_pool,ScalarA(10));
  Kokkos::fill_random(b_y,rand_pool,ScalarB(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b_org_y,b_y);
 
  Kokkos::deep_copy(h_b_x,b_x);

diff --git a/unit_test/blas/Test_Blas1_sum.hpp b/unit_test/blas/Test_Blas1_sum.hpp
@@ -28,6 +28,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;
@@ -67,6 +69,8 @@ namespace Test {
 
  Kokkos::fill_random(b_a,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(h_b_a,b_a);
 
  typename ViewTypeA::const_type c_a = a;

diff --git a/unit_test/blas/Test_Blas1_update.hpp b/unit_test/blas/Test_Blas1_update.hpp
@@ -58,6 +58,8 @@ namespace Test {
  Kokkos::fill_random(b_y,rand_pool,ScalarB(10));
  Kokkos::fill_random(b_z,rand_pool,ScalarC(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b_org_z,b_z);
 
  Kokkos::deep_copy(h_b_x,b_x);
@@ -121,6 +123,8 @@ namespace Test {
  Kokkos::fill_random(b_y,rand_pool,ScalarB(10));
  Kokkos::fill_random(b_z,rand_pool,ScalarC(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b_org_z,b_z);
 
  Kokkos::deep_copy(h_b_x,b_x);

diff --git a/unit_test/blas/Test_Blas2_gemv.hpp b/unit_test/blas/Test_Blas2_gemv.hpp
@@ -56,6 +56,8 @@ namespace Test {
  Kokkos::fill_random(b_y,rand_pool,ScalarY(10));
  Kokkos::fill_random(b_A,rand_pool,ScalarA(10));
 
+ Kokkos::fence();
+
  Kokkos::deep_copy(b_org_y,b_y);
 
  Kokkos::deep_copy(h_b_x,b_x);