leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 30 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_with_broadcast_testbed.h
+11 -10
View File
@@ -404,9 +404,9 @@ public:
// compute tensor Z and tensor T
for (int n = 0; n < problem_size.N; ++n) {
for (int p = 0; p < problem_size.P; ++p) {
for (int q = 0; q < problem_size.Q; ++q) {
for (int k = 0; k < problem_size.K; ++k) {
for (int p = 0; p < (kConvolutionalOperator == cutlass::conv::Operator::kFprop ? problem_size.P : problem_size.H); ++p) {
for (int q = 0; q < (kConvolutionalOperator == cutlass::conv::Operator::kFprop ? problem_size.Q : problem_size.W); ++q) {
for (int k = 0; k < (kConvolutionalOperator == cutlass::conv::Operator::kFprop ? problem_size.K : problem_size.C); ++k) {
ElementZ z{};
ElementT t{};
@@ -449,7 +449,8 @@ 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_")))
<< "nhwc_"
<< problem_size.N << "x"
<< problem_size.H << "x"
@@ -602,8 +603,8 @@ bool TestAllConv2dWithBroadcast(
//
// 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))) {
@@ -613,8 +614,8 @@ bool TestAllConv2dWithBroadcast(
#if 0 // relax restrictions on analytic strided dgrad
// CUTLASS DGRAD's *strided* specialization only support stride >= {2, 2}
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))) {
@@ -650,8 +651,8 @@ bool TestAllConv2dWithBroadcast(
}
// 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)) {