CUTLASS 2.1 (#83)

CUTLASS 2.1 contributes:
- BLAS-style host-side API added to CUTLASS Library
- Planar Complex GEMM kernels targeting Volta and Turing Tensor Cores
- Minor enhancements and bug fixes
This commit is contained in:
Andrew Kerr
2020-04-07 13:51:25 -07:00
committed by GitHub
parent 7c0cd26d13
commit 96dab34ad9
196 changed files with 20653 additions and 1995 deletions
@@ -0,0 +1,172 @@
/***************************************************************************************************
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
*
* Redistribution and use in source and binary forms, with or without modification, are permitted
* provided that the following conditions are met:
* * Redistributions of source code must retain the above copyright notice, this list of
* conditions and the following disclaimer.
* * Redistributions in binary form must reproduce the above copyright notice, this list of
* conditions and the following disclaimer in the documentation and/or other materials
* provided with the distribution.
* * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
* to endorse or promote products derived from this software without specific prior written
* permission.
*
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
* FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
* STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
**************************************************************************************************/
/*!
\file
\brief The universal GEMM accommodates serial reductions, parallel reductions, batched strided, and
batched array variants.
*/
#pragma once
#include "cutlass/cutlass.h"
#include "cutlass/gemm/device/gemm_universal_base.h"
////////////////////////////////////////////////////////////////////////////////
namespace cutlass {
namespace gemm {
namespace device {
/////////////////////////////////////////////////////////////////////////////////////////////////
template <typename GemmKernel_>
class GemmUniversalAdapter {
public:
using GemmKernel = GemmKernel_;
static_assert(std::is_same<typename GemmKernel::LayoutC, cutlass::layout::RowMajor>::value,
"Universal adapter expects the kernel to be row-major and transposes its arguments.");
using ThreadblockShape = typename GemmKernel::Mma::Shape;
using WarpShape = typename GemmKernel::WarpShape;
using InstructionShape = typename GemmKernel::InstructionShape;
using OperatorClass = typename GemmKernel::OperatorClass;
using ArchTag = typename GemmKernel::ArchTag;
// Type, layout, and complex transform deliberately exchanged with B
using ElementA = typename GemmKernel::ElementB;
using LayoutA = typename layout::LayoutTranspose<typename GemmKernel::LayoutB>::type;
using TensorRefA = TensorRef<ElementA const, LayoutA>;
static ComplexTransform const kTransformA = GemmKernel::kTransformB;
static int const kAlignmentA = GemmKernel::kAlignmentA;
// Type, layout, and complex transform deliberately exchanged with A
using ElementB = typename GemmKernel::ElementA;
using LayoutB = typename layout::LayoutTranspose<typename GemmKernel::LayoutA>::type;
using TensorRefB = TensorRef<ElementB const, LayoutB>;
static ComplexTransform const kTransformB = GemmKernel::kTransformA;
static int const kAlignmentB = GemmKernel::kAlignmentB;
using ElementC = typename GemmKernel::ElementC;
using LayoutC = cutlass::layout::ColumnMajor;
using TensorRefC = TensorRef<ElementC const, LayoutC>;
using TensorRefD = TensorRef<ElementC, LayoutC>;
static int const kAlignmentC = GemmKernel::kAlignmentC;
using ElementAccumulator = typename GemmKernel::Mma::Policy::Operator::ElementC;
static int const kStages = GemmKernel::Mma::kStages;
using EpilogueOutputOp = typename GemmKernel::EpilogueOutputOp;
using ThreadblockSwizzle = typename GemmKernel::ThreadblockSwizzle;
using Operator = typename GemmKernel::Operator;
using UnderlyingOperator = GemmUniversalBase<GemmKernel>;
using Arguments = typename UnderlyingOperator::Arguments;
private:
UnderlyingOperator underlying_operator_;
public:
/// Constructs the GEMM.
GemmUniversalAdapter() { }
/// Helper to construct a transposed equivalent for the underying GEMM operator
static Arguments to_underlying_arguments(Arguments const &args) {
return args.transposed_problem();
}
/// Determines whether the GEMM can execute the given problem.
static Status can_implement(Arguments const &args) {
return UnderlyingOperator::can_implement(to_underlying_arguments(args));
}
/// Gets the workspace size
static size_t get_workspace_size(Arguments const &args) {
return UnderlyingOperator::get_workspace_size(to_underlying_arguments(args));
}
/// Computes the grid shape
static dim3 get_grid_shape(Arguments const &args) {
return UnderlyingOperator::get_grid_shape(to_underlying_arguments(args));
}
/// Computes the maximum number of active blocks per multiprocessor
static int maximum_active_blocks(int smem_capacity = -1) {
return UnderlyingOperator::maximum_active_blocks(smem_capacity);
}
/// Initializes GEMM state from arguments.
Status initialize(Arguments const &args, void *workspace = nullptr, cudaStream_t stream = nullptr) {
return underlying_operator_.initialize(to_underlying_arguments(args), workspace, stream);
}
/// Lightweight update given a subset of arguments
Status update(Arguments const &args, void *workspace = nullptr) {
return underlying_operator_.update(to_underlying_arguments(args), workspace);
}
/// Runs the kernel using initialized state.
Status run(cudaStream_t stream = nullptr) {
return underlying_operator_.run(stream);
}
/// Runs the kernel using initialized state.
Status operator()(cudaStream_t stream = nullptr) {
return run(stream);
}
/// Runs the kernel using initialized state.
Status operator()(
Arguments const &args,
void *workspace = nullptr,
cudaStream_t stream = nullptr) {
Status status = initialize(args, workspace, stream);
if (status == Status::kSuccess) {
status = run(stream);
}
return status;
}
};
/////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace device
} // namespace gemm
} // namespace cutlass
/////////////////////////////////////////////////////////////////////////////////////////////////