Removed conditional per-device Permutation and SliceInfo setting from NDArray and SlicedNDArray (55ded6ad) · Commits · TNL / tnl-dev

src/Benchmarks/NDArray/tnl-benchmark-ndarray-boundary.h

+0 −11

Original line number	Diff line number	Diff line
		@@ -85,7 +85,6 @@ void benchmark_1D( Benchmark& benchmark, index_type size = 500000000 )
		NDArray< value_type,
		SizesHolder< index_type, 0 >,
		std::make_index_sequence< 1 >,
		std::make_index_sequence< 1 >,
		Device > a, b;
		a.setSizes( size );
		b.setSizes( size );
		@@ -113,7 +112,6 @@ void benchmark_2D( Benchmark& benchmark, index_type size = 22333 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0 >,
		std::make_index_sequence< 2 >,
		std::make_index_sequence< 2 >,
		Device > a, b;
		a.setSizes( size, size );
		b.setSizes( size, size );
		@@ -141,7 +139,6 @@ void benchmark_3D( Benchmark& benchmark, index_type size = 800 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0 >,
		std::make_index_sequence< 3 >,
		std::make_index_sequence< 3 >,
		Device > a, b;
		a.setSizes( size, size, size );
		b.setSizes( size, size, size );
		@@ -170,7 +167,6 @@ void benchmark_3D( Benchmark& benchmark, index_type size = 800 )
		// NDArray< value_type,
		// SizesHolder< index_type, 0, 0, 0, 0 >,
		// std::make_index_sequence< 4 >,
		// std::make_index_sequence< 4 >,
		// Device > a, b;
		// a.setSizes( size, size, size, size );
		// b.setSizes( size, size, size, size );
		@@ -198,7 +194,6 @@ void benchmark_3D( Benchmark& benchmark, index_type size = 800 )
		// NDArray< value_type,
		// SizesHolder< index_type, 0, 0, 0, 0, 0 >,
		// std::make_index_sequence< 5 >,
		// std::make_index_sequence< 5 >,
		// Device > a, b;
		// a.setSizes( size, size, size, size, size );
		// b.setSizes( size, size, size, size, size );
		@@ -226,7 +221,6 @@ void benchmark_3D( Benchmark& benchmark, index_type size = 800 )
		// NDArray< value_type,
		// SizesHolder< index_type, 0, 0, 0, 0, 0, 0 >,
		// std::make_index_sequence< 6 >,
		// std::make_index_sequence< 6 >,
		// Device > a, b;
		// a.setSizes( size, size, size, size, size, size );
		// b.setSizes( size, size, size, size, size, size );
		@@ -255,7 +249,6 @@ void benchmark_2D_perm( Benchmark& benchmark, index_type size = 22333 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0 >,
		std::index_sequence< 1, 0 >,
		std::index_sequence< 1, 0 >,
		Device > a, b;
		a.setSizes( size, size );
		b.setSizes( size, size );
		@@ -283,7 +276,6 @@ void benchmark_3D_perm( Benchmark& benchmark, index_type size = 800 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0 >,
		std::index_sequence< 2, 1, 0 >,
		std::index_sequence< 2, 1, 0 >,
		Device > a, b;
		a.setSizes( size, size, size );
		b.setSizes( size, size, size );
		@@ -312,7 +304,6 @@ void benchmark_3D_perm( Benchmark& benchmark, index_type size = 800 )
		// NDArray< value_type,
		// SizesHolder< index_type, 0, 0, 0, 0 >,
		// std::index_sequence< 3, 2, 1, 0 >,
		// std::index_sequence< 3, 2, 1, 0 >,
		// Device > a, b;
		// a.setSizes( size, size, size, size );
		// b.setSizes( size, size, size, size );
		@@ -340,7 +331,6 @@ void benchmark_3D_perm( Benchmark& benchmark, index_type size = 800 )
		// NDArray< value_type,
		// SizesHolder< index_type, 0, 0, 0, 0, 0 >,
		// std::index_sequence< 4, 3, 2, 1, 0 >,
		// std::index_sequence< 4, 3, 2, 1, 0 >,
		// Device > a, b;
		// a.setSizes( size, size, size, size, size );
		// b.setSizes( size, size, size, size, size );
		@@ -368,7 +358,6 @@ void benchmark_3D_perm( Benchmark& benchmark, index_type size = 800 )
		// NDArray< value_type,
		// SizesHolder< index_type, 0, 0, 0, 0, 0, 0 >,
		// std::index_sequence< 5, 4, 3, 2, 1, 0 >,
		// std::index_sequence< 5, 4, 3, 2, 1, 0 >,
		// Device > a, b;
		// a.setSizes( size, size, size, size, size, size );
		// b.setSizes( size, size, size, size, size, size );

src/Benchmarks/NDArray/tnl-benchmark-ndarray.h

+0 −11

Original line number	Diff line number	Diff line
		@@ -117,7 +117,6 @@ void benchmark_1D( Benchmark& benchmark, index_type size = 500000000 )
		NDArray< value_type,
		SizesHolder< index_type, 0 >,
		std::make_index_sequence< 1 >,
		std::make_index_sequence< 1 >,
		Device > a, b;
		a.setSizes( size );
		b.setSizes( size );
		@@ -141,7 +140,6 @@ void benchmark_2D( Benchmark& benchmark, index_type size = 22333 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0 >,
		std::make_index_sequence< 2 >,
		std::make_index_sequence< 2 >,
		Device > a, b;
		a.setSizes( size, size );
		b.setSizes( size, size );
		@@ -165,7 +163,6 @@ void benchmark_3D( Benchmark& benchmark, index_type size = 800 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0 >,
		std::make_index_sequence< 3 >,
		std::make_index_sequence< 3 >,
		Device > a, b;
		a.setSizes( size, size, size );
		b.setSizes( size, size, size );
		@@ -189,7 +186,6 @@ void benchmark_4D( Benchmark& benchmark, index_type size = 150 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0, 0 >,
		std::make_index_sequence< 4 >,
		std::make_index_sequence< 4 >,
		Device > a, b;
		a.setSizes( size, size, size, size );
		b.setSizes( size, size, size, size );
		@@ -213,7 +209,6 @@ void benchmark_5D( Benchmark& benchmark, index_type size = 56 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0, 0, 0 >,
		std::make_index_sequence< 5 >,
		std::make_index_sequence< 5 >,
		Device > a, b;
		a.setSizes( size, size, size, size, size );
		b.setSizes( size, size, size, size, size );
		@@ -237,7 +232,6 @@ void benchmark_6D( Benchmark& benchmark, index_type size = 28 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0, 0, 0, 0 >,
		std::make_index_sequence< 6 >,
		std::make_index_sequence< 6 >,
		Device > a, b;
		a.setSizes( size, size, size, size, size, size );
		b.setSizes( size, size, size, size, size, size );
		@@ -262,7 +256,6 @@ void benchmark_2D_perm( Benchmark& benchmark, index_type size = 22333 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0 >,
		std::index_sequence< 1, 0 >,
		std::index_sequence< 1, 0 >,
		Device > a, b;
		a.setSizes( size, size );
		b.setSizes( size, size );
		@@ -286,7 +279,6 @@ void benchmark_3D_perm( Benchmark& benchmark, index_type size = 800 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0 >,
		std::index_sequence< 2, 1, 0 >,
		std::index_sequence< 2, 1, 0 >,
		Device > a, b;
		a.setSizes( size, size, size );
		b.setSizes( size, size, size );
		@@ -310,7 +302,6 @@ void benchmark_4D_perm( Benchmark& benchmark, index_type size = 150 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0, 0 >,
		std::index_sequence< 3, 2, 1, 0 >,
		std::index_sequence< 3, 2, 1, 0 >,
		Device > a, b;
		a.setSizes( size, size, size, size );
		b.setSizes( size, size, size, size );
		@@ -334,7 +325,6 @@ void benchmark_5D_perm( Benchmark& benchmark, index_type size = 56 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0, 0, 0 >,
		std::index_sequence< 4, 3, 2, 1, 0 >,
		std::index_sequence< 4, 3, 2, 1, 0 >,
		Device > a, b;
		a.setSizes( size, size, size, size, size );
		b.setSizes( size, size, size, size, size );
		@@ -358,7 +348,6 @@ void benchmark_6D_perm( Benchmark& benchmark, index_type size = 28 )
		NDArray< value_type,
		SizesHolder< index_type, 0, 0, 0, 0, 0, 0 >,
		std::index_sequence< 5, 4, 3, 2, 1, 0 >,
		std::index_sequence< 5, 4, 3, 2, 1, 0 >,
		Device > a, b;
		a.setSizes( size, size, size, size, size, size );
		b.setSizes( size, size, size, size, size, size );

src/TNL/Containers/NDArray.h

+9 −28

Original line number	Diff line number	Diff line
		@@ -324,23 +324,18 @@ protected:

		template< typename Value,
		typename SizesHolder,
		typename PermutationHost = std::make_index_sequence< SizesHolder::getDimension() >, // identity by default
		typename PermutationCuda = std::make_index_sequence< SizesHolder::getDimension() >, // identity by default
		typename Permutation = std::make_index_sequence< SizesHolder::getDimension() >, // identity by default
		typename Device = Devices::Host,
		typename Index = typename SizesHolder::IndexType >
		class NDArray
		: public NDArrayStorage< Array< Value, Device, Index >,
		SizesHolder,
		typename std::conditional< std::is_same< Device, Devices::Host >::value,
		PermutationHost,
		PermutationCuda >::type,
		Permutation,
		__ndarray_impl::NDArrayBase< SliceInfo< 0, 0 > > >
		{
		using Base = NDArrayStorage< Array< Value, Device, Index >,
		SizesHolder,
		typename std::conditional< std::is_same< Device, Devices::Host >::value,
		PermutationHost,
		PermutationCuda >::type,
		Permutation,
		__ndarray_impl::NDArrayBase< SliceInfo< 0, 0 > > >;

		public:
		@@ -403,34 +398,20 @@ public:

		template< typename Value,
		typename SizesHolder,
		typename PermutationHost = std::make_index_sequence< SizesHolder::getDimension() >, // identity by default
		typename SliceInfoHost = SliceInfo<>, // no slicing by default
		typename PermutationCuda = std::make_index_sequence< SizesHolder::getDimension() >, // identity by default
		typename SliceInfoCuda = SliceInfo<>, // no slicing by default
		typename Permutation = std::make_index_sequence< SizesHolder::getDimension() >, // identity by default
		typename SliceInfo = SliceInfo<>, // no slicing by default
		typename Device = Devices::Host,
		typename Index = typename SizesHolder::IndexType >
		class SlicedNDArray
		: public NDArrayStorage< Array< Value, Device, Index >,
		SizesHolder,
		typename std::conditional< std::is_same< Device, Devices::Host >::value,
		PermutationHost,
		PermutationCuda >::type,
		__ndarray_impl::SlicedNDArrayBase<
		typename std::conditional< std::is_same< Device, Devices::Host >::value,
		SliceInfoHost,
		SliceInfoCuda >::type >
		>
		Permutation,
		__ndarray_impl::SlicedNDArrayBase< SliceInfo > >
		{
		using Base = NDArrayStorage< Array< Value, Device, Index >,
		SizesHolder,
		typename std::conditional< std::is_same< Device, Devices::Host >::value,
		PermutationHost,
		PermutationCuda >::type,
		__ndarray_impl::SlicedNDArrayBase<
		typename std::conditional< std::is_same< Device, Devices::Host >::value,
		SliceInfoHost,
		SliceInfoCuda >::type >
		>;
		Permutation,
		__ndarray_impl::SlicedNDArrayBase< SliceInfo > >;

		public:
		// inherit all assignment operators

src/TNL/Containers/ndarray/SynchronizerBuffers.h

+0 −1

Original line number	Diff line number	Diff line
		@@ -29,7 +29,6 @@ struct SynchronizerBuffersLayer
		using NDArrayType = NDArray< typename DistributedNDArray::ValueType,
		typename DistributedNDArray::SizesHolderType,
		typename DistributedNDArray::PermutationType,
		typename DistributedNDArray::PermutationType,
		typename DistributedNDArray::DeviceType >;
		NDArrayType left_send_buffer, left_recv_buffer, right_send_buffer, right_recv_buffer;
		typename DistributedNDArray::LocalBeginsType left_send_offsets, left_recv_offsets, right_send_offsets, right_recv_offsets;

src/UnitTests/Containers/ndarray/DistributedNDArrayOverlaps_1D_test.h

+0 −4

Original line number	Diff line number	Diff line
		@@ -71,7 +71,6 @@ using DistributedNDArrayTypes = ::testing::Types<
		DistributedNDArray< NDArray< double,
		SizesHolder< int, 0 >,
		std::index_sequence< 0 >,
		std::index_sequence< 0 >,
		Devices::Host >,
		Communicators::MpiCommunicator,
		std::index_sequence< 2 > >
		@@ -79,7 +78,6 @@ using DistributedNDArrayTypes = ::testing::Types<
		// DistributedNDArray< NDArray< double,
		// SizesHolder< int, 0 >,
		// std::index_sequence< 0 >,
		// std::index_sequence< 0 >,
		// Devices::Host >,
		// Communicators::NoDistrCommunicator,
		// std::index_sequence< 2 > >
		@@ -88,7 +86,6 @@ using DistributedNDArrayTypes = ::testing::Types<
		DistributedNDArray< NDArray< double,
		SizesHolder< int, 0 >,
		std::index_sequence< 0 >,
		std::index_sequence< 0 >,
		Devices::Cuda >,
		Communicators::MpiCommunicator,
		std::index_sequence< 2 > >
		@@ -96,7 +93,6 @@ using DistributedNDArrayTypes = ::testing::Types<
		// DistributedNDArray< NDArray< double,
		// SizesHolder< int, 0 >,
		// std::index_sequence< 0 >,
		// std::index_sequence< 0 >,
		// Devices::Cuda >,
		// Communicators::NoDistrCommunicator,
		// std::index_sequence< 2 > >