kInstrK Class — pytorch Architecture
Architecture documentation for the kInstrK class in default_warp_iterator_from_smem.h from the pytorch codebase.
Entity Profile
Source Code
aten/src/ATen/native/transformers/cuda/mem_eff_attention/iterators/default_warp_iterator_from_smem.h lines 59–76
template <typename RegularWarpIterator, typename Policy, int kInstrK>
struct DefaultWarpIteratorAFromSharedMemory<
cutlass::gemm::GemmShape<32, 32, 32>,
cutlass::gemm::GemmShape<16, 8, kInstrK>,
RegularWarpIterator,
Policy,
typename platform::enable_if<(
sizeof_bits<typename RegularWarpIterator::Element>::value == 16 &&
Policy::Operator::Policy::OpDelta::kRow == 1)>::type> {
using OpDelta = typename Policy::Operator::Policy::OpDelta;
using WarpShape = cutlass::MatrixShape<32, 32>;
using InstructionShape = cutlass::gemm::GemmShape<16, 8, kInstrK>;
using WarpIterator = cutlass::gemm::warp::WarpIteratorFromSmem<
cutlass::gemm::Operand::kA,
typename RegularWarpIterator::Element,
cutlass::MatrixShape<InstructionShape::kM, InstructionShape::kK>>;
};
Source
Analyze Your Own Codebase
Get architecture documentation, dependency graphs, and domain analysis for your codebase in minutes.
Try Supermodel Free