leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 25 Packages Projects Releases Wiki Activity

3.6.0 update (#2005)

Browse Source 
* 3.6.0 update

* doc and swap stuff

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Yujia Zhai
2024-12-25 01:34:40 -05:00
committed by GitHub
co-authored by yuzhai Haicheng Wu
parent e1cd8c7866
commit 3d261a5974
258 changed files with 10863 additions and 3883 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/unit/conv/device_3x/conv_problem_sizes.hpp
+31
View File
@@ -1159,6 +1159,37 @@ std::vector<cutlass::conv::ConvProblemShape<cutlass::conv::Operator::kDgrad, 1>>
get_conv_problem_vector<1, cutlass::conv::Operator::kDgrad, true>() {
using ProblemShape = cutlass::conv::ConvProblemShape<cutlass::conv::Operator::kDgrad, 1>;
std::vector<ProblemShape> problem_shapes;
// Test TMA truncation
problem_shapes.push_back({
cutlass::conv::Mode::kCrossCorrelation,
{1, 512, 64}, // nqk
{64, 1, 64}, // ksc
{0}, // padding lower (pad_w)
{0}, // padding upper (pad_w)
{2}, // stride (stride_w)
{1}, // dilation (dilation_w)
1 // group
});
problem_shapes.push_back({
cutlass::conv::Mode::kCrossCorrelation,
{1, 1024, 64}, // nqk
{64, 1, 64}, // ksc
{0}, // padding lower (pad_w)
{0}, // padding upper (pad_w)
{4}, // stride (stride_w)
{1}, // dilation (dilation_w)
1 // group
});
problem_shapes.push_back({
cutlass::conv::Mode::kCrossCorrelation,
{1, 2048, 64}, // nqk
{64, 1, 64}, // ksc
{0}, // padding lower (pad_w)
{0}, // padding upper (pad_w)
{8}, // stride (stride_w)
{1}, // dilation (dilation_w)
1 // group
});
// non-packed input/output strides.
// stride divides dilation
// asymmetric padding
test/unit/conv/device_3x/testbed_conv.hpp
+10 -2
View File
@@ -336,10 +336,17 @@ struct ConvTestbed {
// Scale
if constexpr (cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::ScaledGELU_taylor<ElementCompute>> ||
cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::ScaledGELU<ElementCompute>>) {
cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::ScaledGELU<ElementCompute>> ||
cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::ScaledSiLu<ElementCompute>> ||
cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::ScaledHardSwish<ElementCompute>> ) {
fusion_args.activation.scale = ElementCompute{1};
}
// LeakyRelu
if constexpr (cute::is_same_v<ActivationFunctor, cutlass::epilogue::thread::LeakyReLU<ElementCompute>> ) {
fusion_args.activation.leaky_alpha = ElementCompute{0};
}
cutlass::Status status = cutlass::Status::kInvalid;
status = conv_op.can_implement(args);
@@ -617,8 +624,9 @@ bool TestAllConv(double alpha = 1.0, double beta = 0.0, float epsilon = 0.0f
for (DecompositionMode decomp_mode : decomposition_modes) {
std::vector problem_splits = {Splits{1}};
if constexpr (UsesStreamKScheduler) {
if (decomp_mode == DecompositionMode::Heuristic || decomp_mode == DecompositionMode::SplitK) {
if (decomp_mode == DecompositionMode::SplitK) {
problem_splits.push_back(Splits{2});
problem_splits.push_back(Splits{4});
}
}
for (auto splits : problem_splits) {