CUTLASS 2.0 (#62)

CUTLASS 2.0

Substantially refactored for

- Better performance, particularly for native Turing Tensor Cores
- Robust and durable templates spanning the design space
- Encapsulated functionality embodying modern C++11 programming techniques
- Optimized containers and data types for efficient, generic, portable device code

Updates to:
- Quick start guide
- Documentation
- Utilities
- CUTLASS Profiler

Native Turing Tensor Cores
- Efficient GEMM kernels targeting Turing Tensor Cores
- Mixed-precision floating point, 8-bit integer, 4-bit integer, and binarized operands

Coverage of existing CUTLASS functionality:
- GEMM kernels targeting CUDA and Tensor Cores in NVIDIA GPUs
- Volta Tensor Cores through native mma.sync and through WMMA API
- Optimizations such as parallel reductions, threadblock rasterization, and intra-threadblock reductions
- Batched GEMM operations
- Complex-valued GEMMs

Note: this commit and all that follow require a host compiler supporting C++11 or greater.
This commit is contained in:
Andrew Kerr
2019-11-19 16:55:34 -08:00
committed by GitHub
parent b5cab177a9
commit fb335f6a5f
5434 changed files with 599799 additions and 250176 deletions

View File

@@ -0,0 +1,182 @@
/***************************************************************************************************
* Copyright (c) 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 Defines a class for using integer types smaller than one byte in host or
device code.
*/
#pragma once
#include <cstdint>
#include "cutlass/platform/platform.h"
namespace cutlass {
///////////////////////////////////////////////////////////////////////////////////////////////////
/// 4-bit signed integer type
template <int Bits, bool Signed = true>
struct integer_subbyte {
/// Number of bits
static int const kBits = Bits;
/// Whether type is signed
static bool const kSigned = Signed;
/// External type
using T = typename std::conditional<kSigned, int, unsigned>::type;
/// Storage type
using Storage = uint8_t;
/// Bitmask used to truncate from larger integers
static Storage const kMask = Storage((1 << kBits) - 1);
//
// Data members
//
Storage storage;
//
// Methods
//
/// No operation
CUTLASS_HOST_DEVICE
integer_subbyte() { }
/// Conversion from integer type
CUTLASS_HOST_DEVICE
integer_subbyte(int value)
: storage(reinterpret_cast<Storage const &>(value) & kMask) {}
CUTLASS_HOST_DEVICE
integer_subbyte(unsigned value)
: storage(reinterpret_cast<Storage const &>(value) & kMask) {}
/// Conversion from double
CUTLASS_HOST_DEVICE
integer_subbyte(double value) {
T tmp = (T)value;
storage = reinterpret_cast<Storage const &>(tmp) & kMask;
}
///
CUTLASS_HOST_DEVICE
operator T() const {
if (kSigned) {
// Sign extend
if (storage & Storage(1 << (kBits - 1))) {
return T(storage) | ~T(kMask);
}
}
return T(storage);
}
/// Equality
CUTLASS_HOST_DEVICE
bool operator==(integer_subbyte const &rhs) const {
return storage == rhs.storage;
}
/// Inequality
CUTLASS_HOST_DEVICE
bool operator!=(integer_subbyte const &rhs) const {
return storage != rhs.storage;
}
/// Less than or equal
CUTLASS_HOST_DEVICE
bool operator<=(integer_subbyte const &rhs) const {
if (kSigned) {
if (storage & (1 << (kBits - 1))) {
return !(rhs.storage < storage);
}
}
return storage < rhs.storage;
}
/// Less than
CUTLASS_HOST_DEVICE
bool operator<(integer_subbyte const &rhs) const {
if (kSigned) {
if (storage & (1 << (kBits - 1))) {
return !(rhs.storage <= storage);
}
}
return storage < rhs.storage;
}
/// Greater than or equal
CUTLASS_HOST_DEVICE
bool operator>=(integer_subbyte const &rhs) const {
return !(*this < rhs);
}
/// Greater than
CUTLASS_HOST_DEVICE
bool operator>(integer_subbyte const &rhs) const {
return !(rhs < *this);
}
};
///////////////////////////////////////////////////////////////////////////////////////////////////
/// 1-bit Unsigned integer type
using uint1b_t = integer_subbyte<1, false>;
/// 4-bit Integer type
using int4b_t = integer_subbyte<4, true>;
/// 4-bit Unsigned integer type
using uint4b_t = integer_subbyte<4, false>;
///////////////////////////////////////////////////////////////////////////////////////////////////
/// Defines the size of an element in bits - specialized for uint1b_t
template <>
struct sizeof_bits<uint1b_t> {
static int const value = 1;
};
/// Defines the size of an element in bits - specialized for int4b_t
template <>
struct sizeof_bits<int4b_t> {
static int const value = 4;
};
/// Defines the size of an element in bits - specialized for uint4b_t
template <>
struct sizeof_bits<uint4b_t> {
static int const value = 4;
};
///////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace cutlass