Merge pull request #210 from NexGenAnalytics/gemv_kokkos_fixes

[ready] gemv kokkos: fix tolerances in test and constraint in impl

Author: mhoemmen

tests/kokkos-based/overwriting_matrix_vector_product.cpp

@@ -51,7 +51,7 @@ void kokkos_blas_overwriting_gemv_impl(A_t A, x_t x, y_t y)
     // check A and y
     std::size_t count=0;
     for (std::size_t i=0; i<extent0; ++i){
-      EXPECT_FLOAT_EQ(y(i), y_gold(i));
+      EXPECT_NEAR(y(i), y_gold(i), 1e-2);
       for (std::size_t j=0; j<extent1; ++j){
 	EXPECT_FLOAT_EQ(A(i,j), A_preKernel[count++]);
       }
@@ -67,7 +67,7 @@ void kokkos_blas_overwriting_gemv_impl(A_t A, x_t x, y_t y)
     // check A and y
     std::size_t count=0;
     for (std::size_t i=0; i<extent0; ++i){
-      EXPECT_DOUBLE_EQ(y(i), y_gold(i));
+      EXPECT_NEAR(y(i), y_gold(i), 1e-9);
       for (std::size_t j=0; j<extent1; ++j){
 	EXPECT_DOUBLE_EQ(A(i,j), A_preKernel[count++]);
       }
@@ -84,8 +84,8 @@ void kokkos_blas_overwriting_gemv_impl(A_t A, x_t x, y_t y)
     // check A and y
     std::size_t count=0;
     for (std::size_t i=0; i<extent0; ++i){
-      EXPECT_DOUBLE_EQ(y(i).real(), y_gold(i).real());
-      EXPECT_DOUBLE_EQ(y(i).imag(), y_gold(i).imag());
+      EXPECT_NEAR(y(i).real(), y_gold(i).real(), 1e-9);
+      EXPECT_NEAR(y(i).imag(), y_gold(i).imag(), 1e-9);
 
       for (std::size_t j=0; j<extent1; ++j){
 	EXPECT_DOUBLE_EQ(A(i,j).real(), A_preKernel[count].real());

tests/kokkos-based/updating_matrix_vector_product.cpp

@@ -52,7 +52,7 @@ void kokkos_blas_updating_gemv_impl(A_t A, x_t x, y_t y, z_t z)
     // check A, z, y
     std::size_t count=0;
     for (std::size_t i=0; i<extent0; ++i){
-      EXPECT_FLOAT_EQ(z(i), z_gold(i));
+      EXPECT_NEAR(z(i), z_gold(i), 1e-2);
       EXPECT_FLOAT_EQ(y(i), y_preKernel[i]);
       for (std::size_t j=0; j<extent1; ++j){
 	EXPECT_FLOAT_EQ(A(i,j), A_preKernel[count++]);
@@ -69,8 +69,8 @@ void kokkos_blas_updating_gemv_impl(A_t A, x_t x, y_t y, z_t z)
     // check A, y, z
     std::size_t count=0;
     for (std::size_t i=0; i<extent0; ++i){
-      EXPECT_FLOAT_EQ(z(i), z_gold(i));
-      EXPECT_FLOAT_EQ(y(i), y_preKernel[i]);
+      EXPECT_NEAR(z(i), z_gold(i), 1e-9);
+      EXPECT_DOUBLE_EQ(y(i), y_preKernel[i]);
       for (std::size_t j=0; j<extent1; ++j){
 	EXPECT_DOUBLE_EQ(A(i,j), A_preKernel[count++]);
       }
@@ -87,8 +87,8 @@ void kokkos_blas_updating_gemv_impl(A_t A, x_t x, y_t y, z_t z)
     // check A, y, z
     std::size_t count=0;
     for (std::size_t i=0; i<extent0; ++i){
-      EXPECT_DOUBLE_EQ(z(i).real(), z_gold(i).real());
-      EXPECT_DOUBLE_EQ(z(i).imag(), z_gold(i).imag());
+      EXPECT_NEAR(z(i).real(), z_gold(i).real(), 1e-9);
+      EXPECT_NEAR(z(i).imag(), z_gold(i).imag(), 1e-9);
 
       EXPECT_DOUBLE_EQ(y(i).real(), y_preKernel[i].real());
       EXPECT_DOUBLE_EQ(y(i).imag(), y_preKernel[i].imag());

tpl-implementations/include/experimental/__p1673_bits/kokkos-kernels/blas2_gemv_kk.hpp

@@ -15,12 +15,13 @@ T get_scaling_factor(Accessor /*a*/, T /*v*/) {
 
 template<class Accessor, class S, class T>
 auto get_scaling_factor(std::experimental::linalg::accessor_scaled<Accessor,S> a,
-			T /*v*/)
+      T /*v*/)
 {
   return T(a.scale_factor());
 }
 
-} //end anon namespace
+} //end gemv_impl namespace
+
 
 //
 // overwriting gemv: y = Ax
@@ -193,7 +194,10 @@ template<//class ExecSpace,
          class ElementType_z,
          std::experimental::extents<>::size_type ext_z,
          class Layout_z>
-void matrix_vector_product(kokkos_exec<>,
+std::enable_if_t<
+  !std::is_same_v< Accessor_y, std::experimental::default_accessor<ElementType_y>>
+  >
+matrix_vector_product(kokkos_exec<>,
 			   std::experimental::mdspan<
 			    ElementType_A,
 			    std::experimental::extents<numRows_A, numCols_A>,
@@ -235,8 +239,8 @@ void matrix_vector_product(kokkos_exec<>,
   using yv_t = typename decltype(y)::value_type;
   using zv_t = typename decltype(z)::value_type;
 
-  // z = alpha1*y
-  auto alpha1 = gemv_impl::get_scaling_factor(y.accessor(), yv_t{});
+  // z = alpha1*y, we know from constraint y has non default accessor
+  auto alpha1 = y.accessor().scale_factor();
   constexpr zv_t zero(0);
   KokkosBlas::axpby(alpha1, y_view, zero, z_view);