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

Cutlass 1.3 Release (#42)

Browse Source 
CUTLASS 1.3 Release
- Efficient GEMM kernel targeting Volta Tensor Cores via mma.sync instruction added in CUDA 10.1.
This commit is contained in:
Andrew Kerr
2019-03-20 10:49:17 -07:00
committed by GitHub
parent 19a9d64e3c
commit 877bdcace6
256 changed files with 16930 additions and 802 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/00_basic_gemm/CMakeLists.txt
+1 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
# 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:
examples/00_basic_gemm/basic_gemm.cu
+3 -1
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
* 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:
@@ -57,6 +57,8 @@
// Defines cutlass::gemm::SgemmTraits, the structural components for single-precision GEMM
#include "cutlass/gemm/sgemm_traits.h"
#pragma warning( disable : 4503)
///////////////////////////////////////////////////////////////////////////////////////////////////
//
// This function defines a CUTLASS GEMM kernel instantiation, constructs its parameters object,
examples/01_tensor_view/CMakeLists.txt
+1 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
# 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:
examples/01_tensor_view/tensor_view.cu
+1 -1
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
* 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:
examples/02_cutlass_utilities/CMakeLists.txt
+1 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
# 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:
examples/02_cutlass_utilities/cutlass_utilities.cu
+1 -1
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
* 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:
examples/03_strided_batched_gemm/CMakeLists.txt
+1 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
# 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:
examples/03_strided_batched_gemm/strided_batched_gemm.cu
+3 -1
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
* 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:
@@ -29,6 +29,8 @@
#include "cutlass/gemm/gemm.h"
#include "cutlass/gemm/sgemm_traits.h"
#pragma warning( disable : 4503)
/*
This example demonstrates how to use cutlass to compute a batched strided gemm.
In this example, both A and B matrix are non-transpose and column major matrix
examples/04_tile_iterator/CMakeLists.txt
+1 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
# 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:
examples/04_tile_iterator/tile_iterator.cu
+1 -1
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
* 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:
examples/05_wmma_gemm/CMakeLists.txt
+1 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
# 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:
examples/05_wmma_gemm/wmma_gemm.cu
+10 -8
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
* 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:
@@ -47,8 +47,10 @@
// CUTLASS includes needed for WMMA GEMM kernel
#include "cutlass/wmma_matrix.h"
#pragma warning( disable : 4503)
// This example works only when this MACRO is defined in "cutlass/wmma_matrix.h"
#ifdef CUTLASS_USE_SUBBYTE_WMMA
#ifdef CUTLASS_USE_INT_WMMA
// Defines cutlass::gemm::Gemm, the generic Gemm computation template class.
#include "cutlass/gemm/gemm.h"
@@ -273,7 +275,7 @@ cudaError_t TestCutlassGemm(int M, int N, int K, int alpha, int beta) {
// Passed error check
return cudaSuccess;
}
#endif // defined CUTLASS_USE_SUBBYTE_WMMA
#endif // defined CUTLASS_USE_INT_WMMA
///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -285,7 +287,7 @@ cudaError_t TestCutlassGemm(int M, int N, int K, int alpha, int beta) {
//
int main(int argc, const char *arg[]) {
#ifdef CUTLASS_USE_SUBBYTE_WMMA
#ifdef CUTLASS_USE_INT_WMMA
// Properties of CUDA device
cudaDeviceProp device_properties;
@@ -299,8 +301,8 @@ int main(int argc, const char *arg[]) {
return -1;
}
if ((device_properties.major * 10 + device_properties.minor) < 75) {
std::cerr << "This example needs to run on a Turing device." << std::endl;
if ((device_properties.major * 10 + device_properties.minor) < 72) {
std::cerr << "This example needs to run on a device which has at least 7.2 compute capability." << std::endl;
return -1;
}
@@ -344,9 +346,9 @@ int main(int argc, const char *arg[]) {
return result == cudaSuccess ? 0 : -1;
#else
std::cerr << "CUTLASS WMMA GEMM targeting Turing Tensor Cores features requires CUDA 10." << std::endl;
std::cerr << "CUTLASS WMMA GEMM targeting Turing Tensor Cores features requires compute capability 7.2." << std::endl;
return -1;
#endif // defined CUTLASS_USE_SUBBYTE_WMMA
#endif // defined CUTLASS_USE_INT_WMMA
}
///////////////////////////////////////////////////////////////////////////////////////////////////
examples/06_splitK_gemm/CMakeLists.txt
+1 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
# 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:
examples/06_splitK_gemm/splitK_gemm.cu
+2 -6
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
* 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:
@@ -84,11 +84,7 @@ cudaError_t cutlass_splitK_sgemm_nn(float const *A,
typename deviceGemm::Params deviceGemmParams(m, n, k);
// query if workspace is needed. the workspace size is sizeof(accumulateType) * M * N * splits_count
int workspace_size = deviceGemmParams.required_workspace_memory_in_byte();
if (workspace_size <= 0) {
std::cerr << "splitK workspace_size is smaller than 0" << std::endl;
return cudaErrorInvalidValue;
}
size_t workspace_size = deviceGemmParams.required_workspace_memory_in_byte();
// allocate workspace memory
float *workspace_ptr;
examples/CMakeLists.txt
+2 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2018, NVIDIA CORPORATION. All rights reserved.
# 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:
@@ -27,3 +27,4 @@ add_subdirectory(03_strided_batched_gemm)
add_subdirectory(04_tile_iterator)
add_subdirectory(05_wmma_gemm)
add_subdirectory(06_splitK_gemm)