leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 27 Packages Projects Releases Wiki Activity

Updates for CUTLASS 3.5.0 (#1468)

Browse Source
This commit is contained in:
Vijay Thakkar
2024-04-11 21:33:40 -04:00
committed by GitHub
parent a40e08e9d5
commit 7d49e6c7e2
171 changed files with 7526 additions and 1888 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/unit/conv/device/conv2d_testbed.h
+8 -8
View File
@@ -153,7 +153,6 @@ public:
else if (dist_kind == cutlass::Distribution::Identity) {
cutlass::reference::host::TensorFillIdentity(view);
}
else if (dist_kind == cutlass::Distribution::Gaussian) {
@@ -489,7 +488,8 @@ public:
fname << "error_Conv2d_ImplicitGemm_device_"
<< (split_k_mode == cutlass::conv::SplitKMode::kSerial ? "serial_reduction_" : "parallel_reduction_")
<< (Conv2d::kConvolutionalOperator == cutlass::conv::Operator::kFprop ? "fprop_" :
(Conv2d::kConvolutionalOperator == cutlass::conv::Operator::kDgrad ? "dgrad_" : "wgrad_"))
(Conv2d::kConvolutionalOperator == cutlass::conv::Operator::kDgrad ? "dgrad_" :
(Conv2d::kConvolutionalOperator == cutlass::conv::Operator::kDeconv ? "deconv_" : "wgrad_")))
<< ss_problem_size_text.str()
<< Conv2d::ThreadblockShape::kM << "x"
<< Conv2d::ThreadblockShape::kN << "x"
@@ -635,8 +635,8 @@ bool TestAllConv2d(
//
// CUTLASS DGRAD's *unity* stride specialization only support stride {1, 1}
if ((ImplicitGemm::kConvolutionalOperator ==
cutlass::conv::Operator::kDgrad) &&
if ((ImplicitGemm::kConvolutionalOperator == cutlass::conv::Operator::kDgrad ||
ImplicitGemm::kConvolutionalOperator == cutlass::conv::Operator::kDeconv) &&
(ImplicitGemm::UnderlyingKernel::Mma::IteratorA::kStrideSupport ==
cutlass::conv::StrideSupport::kUnity)) {
if (!((conv_problem.stride_h == 1) && (conv_problem.stride_w == 1))) {
@@ -663,8 +663,8 @@ bool TestAllConv2d(
// CUTLASS DGRAD's *strided* stride specialization supports all stride {stride_h, stride_w}
// Although strided dgrad works for all stride combinations, we are only going
// to run strided dgrad for non-unity strides
if ((ImplicitGemm::kConvolutionalOperator ==
cutlass::conv::Operator::kDgrad) &&
if ((ImplicitGemm::kConvolutionalOperator == cutlass::conv::Operator::kDgrad ||
ImplicitGemm::kConvolutionalOperator == cutlass::conv::Operator::kDeconv) &&
(ImplicitGemm::UnderlyingKernel::Mma::IteratorA::kStrideSupport ==
cutlass::conv::StrideSupport::kStrided)) {
if (((conv_problem.stride_h == 1) && (conv_problem.stride_w == 1))) {
@@ -718,8 +718,8 @@ bool TestAllConv2d(
}
// CUTLASS DGRAD's *strided* specialization does not support split-k mode
if ((ImplicitGemm::kConvolutionalOperator ==
cutlass::conv::Operator::kDgrad) &&
if ((ImplicitGemm::kConvolutionalOperator == cutlass::conv::Operator::kDgrad ||
ImplicitGemm::kConvolutionalOperator == cutlass::conv::Operator::kDeconv) &&
(ImplicitGemm::UnderlyingKernel::Mma::IteratorA::kStrideSupport ==
cutlass::conv::StrideSupport::kStrided)) {