leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

CUTLASS 2.0 (#62)

Browse Source 
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
Download Patch File Download Diff File Expand all files Collapse all files
media/docs/code_organization.md
+215
View File
@@ -0,0 +1,215 @@
![ALT](/media/images/gemm-hierarchy-with-epilogue-no-labels.png "CUTLASS Code Organization")
[README](/README.md#documentation) > **Code Organization**
# CUTLASS Code Organization
This document describes the layout of the CUTLASS repository. The main components are:
* **CUTLASS Template Library** - CUDA Templates for Linear Algebra Subroutines and Solvers (header only)
* **CUTLASS Utilities** - Additional templates
* **CUTLASS Instance Library** - instantiations of CUTLASS templates covering the design space
* **CUTLASS Profiler** - CUTLASS Library, Profiler, and Utilities
* **Examples** - SDK examples of CUTLASS Template Library and components
* **Media** - supporting documentation and media content
* **Tests** - test components for CUTLASS Template Library and tools
## CUTLASS Template Library
CUDA Templates for Linear Algebra Subroutines and Solvers is a library of CUDA C++ template classes for
performing efficient matrix computations on NVIDIA GPUs.
Like NVIDIA CUB, the components of CUTLASS are organized hierarchically based on the scope of cooperative
elements. For example, warp-level GEMM components perform a matrix multiply collectively by the
set of threads within a warp. The following figure illustrates each layer.
Components are designed to be usable by client applications accessing functionailty at each scope.
CUTLASS Templates are implemented by header files in the following directory structure:
```
include/ # Top-level include directory. Client applications should target this path.
cutlass/ # CUDA Templates for Linear Algebra Subroutines and Solvers - headers only
arch/ # direct exposure of architecture features (including instruction-level GEMMs)
*
gemm/ # code specialized for general matrix product computations
thread/ # thread-level operators
warp/ # warp-level operators
threadblock/ # CTA-level operators
kernel/ # CUDA kernel entry points
device/ # launches kernel(s) over a full device
* # scope-agnostic components and basic vocabular type definitions for GEMM
layout/ # layout definitions for matrices, tensors, and other mathematical objects in memory
*
reduction/ # bandwidth-limited reduction kernels that do not fit the "gemm" models
thread/ # thread-level operators
warp/ # warp-level operators
threadblock/ # CTA-level operators
kernel/ # CUDA kernel entry points
device/ # launches kernel(s) over a full device
* # scope-agnostic components and basic vocabular type definitions
transform/ # code specialized for layout, type, and domain transformations
thread/ # thread-level operators
warp/ # warp-level operators
threadblock/ # CTA-level operators
kernel/ # CUDA kernel entry points
device/ # launches kernel(s) over a full device
* # scope-agnostic components and basic vocabulary type definitions
util/ # miscellaneous CUTLASS components
*
* # core vocabulary types and fundamental arithmetic operators
```
See [Programming Guidelines](/media/docs/programming_guidelines.md) for further details about
conventions and design patterns used throughout CUTLASS.
## Tools
The `tools/` directory contains clients of the CUTLASS Template library and includes the following.
## CUTLASS Instance Library
The CUTLASS Instance Library contains instantiations of the above CUTLASS templates covering supported configurations,
data types, block structure, and tile sizes. These instantiations are procedurally generated using a set of
scripts to span the design space.
```
tools/
library/ # static/dynamic library containing all kernel instantiations of interest
# (with some build-level filter switches to compile specific subsets, perhaps by architecture)
include/
cutlass/
library/ # header files for CUTLASS Deliverables Library (in cutlass::library:: namespace)
library.h # defines enums and structs to describe the tiled structure of operator instances
manifest.h # collection of all instances
scripts/ # scripts to procedurally generate CUTLASS template instances
gemm_operations.py
library.py
generator.py # entry point of procedural generation scripts - invoked by cmake
manifest.py
src/
```
## Examples
To demonstrate CUTLASS components, several SDK examples are implemented in `examples/`.
When CMake is executed, the CUTLASS Instance Library generator scripts are executed to construct a set of
instantiations in `build/tools/library/generated/`.
The CUTLASS Profiler is designed to initialize the CUTLASS Instance Library and execute all operations contained therein.
### CUTLASS Utilities
`tools/util/` defines a companion library of headers and sources that support the CUTLASS test programs, examples, and other client applications. Its structure is as follows:
```
tools/
util/
include/
cutlass/
util/ # CUTLASS Utility companion library
reference/ # reference implementation of CUTLASS operators - minimal consideration for performance
detail/
*
device/ # device-side reference implementations of CUTLASS operators
thread/
kernel/
*
host/ # host-side reference implementations of CUTLASS operators
*
*
```
[More details about CUTLASS Utilities may be found here.](media/docs/utilities.md)
### CUTLASS Profiler
This is application constructs an execution environment for evaluating the functionality and performance of
CUTLASS components. It is implemented in
```
tools/
profiler/
```
and may be built as follows.
```
$ make cutlass_profiler -j
```
[Further details about the CUTLASS Profiler are described here.](media/docs/profiler.md)
## Examples
To demonstrate CUTLASS components, several SDK examples are implemented in `examples/`.
## Media
This directory contains documentation, images, and performance result data which accompanies the CUTLASS library and components.
## Tests
Test programs for CUTLASS. Tests are organized hierarchically, mirroring the organization of source files.
```
test/ # unit tests for CUTLASS Template Library
unit/
arch/
core/
gemm/
device/
kernel/
thread/
threadblock/
warp/
reduction/
kernel/
thread/
transform/
threadblock/
*
```
Tests can be built and run at the top level scope by invoking `make test_unit` or by building
and explicitly executing each individual target, e.g. `cutlass_test_unit_gemm_device`.
Tests are configured to specify appropriate GTest filter strings to avoid running except on
architectures where they are expected to pass. Thus, no tests should fail. The actual number
of tests run may vary over time as more are added.
# Copyright
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.
```