leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 24 Packages Projects Releases Wiki Activity

CUTLASS 2.2 (#96)

Browse Source 
Adds support for NVIDIA Ampere Architecture features. CUDA 11 Toolkit recommended.
This commit is contained in:
Andrew Kerr
2020-06-08 16:17:35 -07:00
committed by GitHub
parent e33d90b361
commit 86931fef85
584 changed files with 51080 additions and 3373 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/07_volta_tensorop_gemm/CMakeLists.txt
+1 -1
View File
@@ -1,4 +1,4 @@
# Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
# Copyright (c) 2017-2020, 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/07_volta_tensorop_gemm/volta_tensorop_gemm.cu
+27 -19
View File
@@ -1,5 +1,5 @@
/***************************************************************************************************
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2017-2020, 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:
@@ -156,7 +156,7 @@ using ShapeMMAWarp = cutlass::gemm::GemmShape<64, 64, 32>; // <- warp tile M =
using ShapeMMAOp = cutlass::gemm::GemmShape<8, 8, 4>; // <- MMA Op tile M = 8, N = 8, K = 4
// This code section describes how threadblocks are scheduled on GPU
using SwizzleThreadBlock = cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle; // <- ??
using SwizzleThreadBlock = cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>; // <- ??
// This code section describes ?
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<
@@ -188,15 +188,7 @@ using Gemm = cutlass::gemm::device::Gemm<ElementInputA,
SwizzleThreadBlock,
NumStages>;
int main() {
// Volta Tensor Core operations exposed with mma.sync are first available in CUDA 10.1.
//
// CUTLASS must be compiled with CUDA 10.1 Toolkit to run these examples.
if (!(__CUDACC_VER_MAJOR__ > 10 || (__CUDACC_VER_MAJOR__ == 10 && __CUDACC_VER_MINOR__ >= 1))) {
std::cerr << "Volta Tensor Core operations must be compiled with CUDA 10.1 Toolkit or later." << std::endl;
return -1;
}
int run() {
cudaDeviceProp props;
@@ -223,7 +215,7 @@ int main() {
cutlass::HostTensor<ElementInputA, LayoutInputA> tensor_a(
problem_size.mk()); // <- Create matrix A with dimensions M x K
cutlass::HostTensor<ElementInputB, LayoutInputB> tensor_b(
problem_size.nk()); // <- Create matrix B with dimensions N x K
problem_size.kn()); // <- Create matrix B with dimensions K x N
cutlass::HostTensor<ElementOutput, LayoutOutput> tensor_c(
problem_size.mn()); // <- Create matrix C with dimensions M x N
cutlass::HostTensor<ElementOutput, LayoutOutput> tensor_d(
@@ -326,12 +318,28 @@ int main() {
tensor_ref_d.sync_host();
// Check if output from CUTLASS kernel and reference kernel are equal or not
std::cout << (cutlass::reference::host::TensorEquals(tensor_d.host_view(),
tensor_ref_d.host_view())
? "Passed"
: "Failed")
<< std::endl;
bool passed = cutlass::reference::host::TensorEquals(
tensor_d.host_view(),
tensor_ref_d.host_view());
CUTLASS_CHECK(status);
return 0;
std::cout << (passed ? "Passed" : "Failed") << std::endl;
return (passed ? 0 : -1);
}
int main() {
// Volta Tensor Core operations exposed with mma.sync are first available in CUDA 10.1.
//
// CUTLASS must be compiled with CUDA 10.1 Toolkit to run these examples.
if (!(__CUDACC_VER_MAJOR__ > 10 || (__CUDACC_VER_MAJOR__ == 10 && __CUDACC_VER_MINOR__ >= 1))) {
std::cerr << "Volta Tensor Core operations must be compiled with CUDA 10.1 Toolkit or later." << std::endl;
// Returning zero when built on older Toolkits so tests pass. The actions of this SDK example are no-op.
return 0;
}
else {
return run();
}
}