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

v3.9 (#2185)

Browse Source 
* v3.8 update x

* fix blackwell gg

* doc change

* doc change

* doc change

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
Co-authored-by: Haicheng Wu <57973641+hwu36@users.noreply.github.com>
This commit is contained in:
Yujia Zhai
2025-03-21 01:52:23 -04:00
committed by GitHub
co-authored by yuzhai Haicheng Wu Haicheng Wu
parent 8c4d1dc47d
commit 62750a2b75
334 changed files with 91517 additions and 2656 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/unit/gemm/device/gemm_testbed_3x_ptr_array.hpp
+11 -11
View File
@@ -552,10 +552,10 @@ struct HostCollectiveMainloop<cutlass::gemm::KernelPtrArrayTmaWarpSpecializedBlo
static constexpr int SFVecSize = Gemm::GemmKernel::CollectiveMainloop::SFVecSize;
using ElementSF = typename Gemm::GemmKernel::ElementSF;
using Sm100BlkScaledConfig = typename Gemm::GemmKernel::CollectiveMainloop::Sm100BlkScaledConfig;
using Blk_MN = typename Sm100BlkScaledConfig::Blk_MN;
using Blk_SF = typename Sm100BlkScaledConfig::Blk_SF;
using SfAtom = typename Sm100BlkScaledConfig::SfAtom;
using Sm1xxBlkScaledConfig = typename Gemm::GemmKernel::CollectiveMainloop::Sm1xxBlkScaledConfig;
using Blk_MN = typename Sm1xxBlkScaledConfig::Blk_MN;
using Blk_SF = typename Sm1xxBlkScaledConfig::Blk_SF;
using SfAtom = typename Sm1xxBlkScaledConfig::SfAtom;
using LayoutSFA = typename Gemm::GemmKernel::CollectiveMainloop::LayoutSFA;
using InternalLayoutSFA = typename Gemm::GemmKernel::CollectiveMainloop::InternalLayoutSFA;
using LayoutSFB = typename Gemm::GemmKernel::CollectiveMainloop::LayoutSFB;
@@ -662,8 +662,8 @@ struct HostCollectiveMainloop<cutlass::gemm::KernelPtrArrayTmaWarpSpecializedBlo
auto k_blks = cutlass::ceil_div(K, size<1>(shape(SfAtom{})));
auto m_blks = cutlass::ceil_div(M, Blk_MN{});
auto n_blks = cutlass::ceil_div(N, Blk_MN{});
layout_sfa_host.push_back(Sm100BlkScaledConfig::tile_atom_to_shape_SFA(cute::make_shape(M, N, K, 1)));
layout_sfb_host.push_back(Sm100BlkScaledConfig::tile_atom_to_shape_SFB(cute::make_shape(M, N, K, 1)));
layout_sfa_host.push_back(Sm1xxBlkScaledConfig::tile_atom_to_shape_SFA(cute::make_shape(M, N, K, 1)));
layout_sfb_host.push_back(Sm1xxBlkScaledConfig::tile_atom_to_shape_SFB(cute::make_shape(M, N, K, 1)));
// 2.x host tensor does not natively contain a batch stride or coord, so we spoof if by folding it into the outer mode
auto sfa_coord = cutlass::make_Coord(m_blks * Blk_MN{}, k_blks * Blk_SF{});
@@ -1110,12 +1110,12 @@ struct HostCollectiveEpilogue {
static constexpr SfStrategy SfGenStrategy = (!IsBlockScaleSupported) ? SfStrategy::None : SfStrategy::SfDGen;
static constexpr int32_t SFD_VectorSize = IsBlockScaleSupported ? FusionOp::SFVecSize : 1;
using ElementSFD = non_void_t<cute::remove_pointer_t<typename FusionOp::ElementBlockScaleFactor>, ElementD>;
using Sm100BlockScaledOutputConfig = cutlass::detail::Sm100BlockScaledOutputConfig<
using Sm1xxBlockScaledOutputConfig= cutlass::detail::Sm1xxBlockScaledOutputConfig<
SFD_VectorSize
>;
using Blk_MN = typename Sm100BlockScaledOutputConfig::Blk_MN;
using Blk_SF = typename Sm100BlockScaledOutputConfig::Blk_SF;
using OutputSFAtom = typename Sm100BlockScaledOutputConfig::SfAtom;
using Blk_MN = typename Sm1xxBlockScaledOutputConfig::Blk_MN;
using Blk_SF = typename Sm1xxBlockScaledOutputConfig::Blk_SF;
using OutputSFAtom = typename Sm1xxBlockScaledOutputConfig::SfAtom;
std::vector<cutlass::HostTensor<ElementSFD, LayoutTagD>> tensors_SFD;
std::vector<cutlass::HostTensor<ElementSFD, LayoutTagD>> references_SFD;
cutlass::DeviceAllocation<ElementSFD *> device_tensors_SFD;
@@ -1711,7 +1711,7 @@ struct HostCollectiveEpilogue {
auto SfD = [&](){
if constexpr (IsBlockScaleSupported) {
auto tensor = make_tensor(detail::make_iterator(references_SFD[batch].host_data()),
Sm100BlockScaledOutputConfig::tile_atom_to_shape_SFD(problem_shape_MNKL));
Sm1xxBlockScaledOutputConfig::tile_atom_to_shape_SFD(problem_shape_MNKL));
return tensor;
}
else {