CUTLASS 2.4 (Implicit GEMM convolution) (#147)
CUTLASS 2.4 (Implicit GEMM Convolution) Co-authored-by: Manish Gupta <manigupta@nvidia.com>, Haicheng Wu <haichengw@nvidia.com>, Dustyn Blasig <dblasig@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
This commit is contained in:
Manish Gupta
co-authored by
Manish Gupta <manigupta@nvidia.com>, Haicheng Wu <haichengw@nvidia.com>, Dustyn Blasig <dblasig@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
Manish Gupta <manigupta@nvidia.com>, Haicheng Wu <haichengw@nvidia.com>, Dustyn Blasig <dblasig@nvidia.com>, Andrew Kerr <akerr@nvidia.com>
parent
c2b80ad4e4
commit
6615010cd0
@@ -367,6 +367,52 @@ struct DefaultInterleavedEpilogueTensorOp {
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines sensible defaults for epilogues for TensorOps which uses
|
||||
/// intereleaved output layout. For this case, shared memory is not needed.
|
||||
template <typename Shape_, typename WarpMmaTensorOp_, int PartitionsK,
|
||||
typename OutputOp_, int ElementsPerAccess, int InterleavedK,
|
||||
bool IsBetaZero = false, bool isSplitK = false>
|
||||
struct DefaultInterleavedConvEpilogue {
|
||||
using Shape = Shape_;
|
||||
using WarpMmaTensorOp = WarpMmaTensorOp_;
|
||||
static int const kPartitionsK = PartitionsK;
|
||||
using OutputOp = OutputOp_;
|
||||
static int const kElementsPerAccess = ElementsPerAccess;
|
||||
|
||||
using ElementOutput = typename OutputOp::ElementOutput;
|
||||
using ElementAccumulator = typename WarpMmaTensorOp::ElementC;
|
||||
|
||||
//
|
||||
// Thread map
|
||||
//
|
||||
using OutputTileThreadMap = typename cutlass::epilogue::threadblock::
|
||||
DefaultInterleavedConvThreadMapTensorOp<
|
||||
Shape, typename WarpMmaTensorOp::Shape, kPartitionsK, ElementOutput,
|
||||
kElementsPerAccess, InterleavedK>::Type;
|
||||
|
||||
using OutputTileIterator =
|
||||
cutlass::epilogue::threadblock::InterleavedConvPredicatedTileIterator<
|
||||
OutputTileThreadMap, ElementOutput, InterleavedK>;
|
||||
|
||||
using AccumulatorFragmentIterator =
|
||||
cutlass::epilogue::warp::FragmentIteratorTensorOp<
|
||||
typename WarpMmaTensorOp::Shape,
|
||||
typename WarpMmaTensorOp::Policy::Operator::Shape,
|
||||
typename WarpMmaTensorOp::Policy::Operator::ElementC,
|
||||
typename WarpMmaTensorOp::Policy::Operator::FragmentC,
|
||||
// can reuse the gemm version here to do element selection
|
||||
layout::ColumnMajorInterleaved<InterleavedK>>;
|
||||
|
||||
//
|
||||
// Define the epilogue
|
||||
//
|
||||
using Epilogue = cutlass::epilogue::threadblock::InterleavedEpilogue<
|
||||
Shape, WarpMmaTensorOp, kPartitionsK, OutputTileIterator,
|
||||
AccumulatorFragmentIterator, OutputOp, InterleavedK, IsBetaZero>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace threadblock
|
||||
} // namespace epilogue
|
||||
} // namespace cutlass
|
||||
|
||||
@@ -144,6 +144,55 @@ struct DefaultInterleavedThreadMapTensorOp {
|
||||
Detail::kThreads, kElementsPerAccess, sizeof_bits<Element>::value>;
|
||||
};
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Defines the optimal thread map for TensorOp accumulator layouts
|
||||
template <typename ThreadblockShape_, typename WarpShape_, int PartitionsK,
|
||||
typename Element_, int ElementsPerAccess, int InterleavedK>
|
||||
struct DefaultInterleavedConvThreadMapTensorOp {
|
||||
using ThreadblockShape = ThreadblockShape_;
|
||||
using WarpShape = WarpShape_;
|
||||
static int const kPartitionsK = PartitionsK;
|
||||
using Element = Element_;
|
||||
static int const kElementsPerAccess = ElementsPerAccess;
|
||||
static int const kInterleavedK = InterleavedK;
|
||||
|
||||
//
|
||||
// Definitions
|
||||
//
|
||||
|
||||
struct Detail {
|
||||
/// Tensor Operations fundamentally perform operations on 8 rows
|
||||
static int const kTensorOpRows = 8;
|
||||
static int const kWarpSize = 32;
|
||||
|
||||
static_assert(!(ThreadblockShape::kM % WarpShape::kM) &&
|
||||
!(ThreadblockShape::kN % WarpShape::kN),
|
||||
"Divisibility");
|
||||
|
||||
/// Number of warps
|
||||
using WarpCount =
|
||||
gemm::GemmShape<ThreadblockShape::kM / WarpShape::kM,
|
||||
ThreadblockShape::kN / WarpShape::kN, kPartitionsK>;
|
||||
|
||||
/// Number of participating threads
|
||||
static int const kThreads = WarpCount::kCount * kWarpSize;
|
||||
};
|
||||
|
||||
//
|
||||
// ThreadMap
|
||||
//
|
||||
|
||||
/// ThreadMap to be used by epilogue::MaskedTileIterator satisfying concept
|
||||
/// InterleavedOutputTileThreadMap
|
||||
using Type = InterleavedConvOutputTileThreadMap<
|
||||
MatrixShape<Detail::WarpCount::kM, Detail::WarpCount::kN>,
|
||||
MatrixShape<WarpShape::kM / Detail::kTensorOpRows,
|
||||
WarpShape::kN / InterleavedK>,
|
||||
Detail::kThreads, kElementsPerAccess, sizeof_bits<Element>::value>;
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace threadblock
|
||||
|
||||
@@ -0,0 +1,92 @@
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
#include "cutlass/conv/convolution.h"
|
||||
#include "cutlass/conv/conv2d_problem_size.h"
|
||||
#include "cutlass/conv/conv3d_problem_size.h"
|
||||
#include "cutlass/layout/tensor.h"
|
||||
#include "cutlass/layout/matrix.h"
|
||||
#include "cutlass/tensor_ref.h"
|
||||
|
||||
namespace cutlass {
|
||||
namespace epilogue {
|
||||
namespace threadblock {
|
||||
|
||||
template<
|
||||
typename TensorLayout_, ///! The original output tensor layout
|
||||
typename OutputIteratorLayout_, ///! Layout used by epilogue output iterator
|
||||
typename TensorRef_, ///! Input tensor to epilogue output iterator
|
||||
conv::Operator ConvOperator, ///! Convolutional operator (Fprop, Dgrad, Wgrad)
|
||||
typename ConvProblemSize_ ///! Convolutional operator on 2D or 3D problem
|
||||
>
|
||||
struct ConvOutputIteratorParameter {
|
||||
|
||||
using TensorLayout = TensorLayout_;
|
||||
using OutputIteratorLayout = OutputIteratorLayout_;
|
||||
using OutputTensorCoord = typename OutputIteratorLayout::TensorCoord;
|
||||
using TensorRef = TensorRef_;
|
||||
static conv::Operator const kConvolutionalOperator = ConvOperator;
|
||||
using ConvProblemSize = ConvProblemSize_;
|
||||
|
||||
/// Wgrad stride idx for implicit gemm algorithm
|
||||
// Conv2d row-major matrix (KxRSC)
|
||||
// Conv3d row-major matrix (KxTRSC)
|
||||
static int const kWgradStrideIdx =
|
||||
platform::is_same<TensorLayout, layout::TensorNHWC>::value ? 2 : 3;
|
||||
|
||||
/// This chooses the appropriate stride element of the C tensor.
|
||||
static int const kTensorStrideIdx =
|
||||
(kConvolutionalOperator == conv::Operator::kWgrad ? kWgradStrideIdx : 0);
|
||||
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
static OutputIteratorLayout layout(const TensorRef & ref) {
|
||||
return ref.stride(kTensorStrideIdx);
|
||||
}
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
static OutputTensorCoord extent(ConvProblemSize problem_size) {
|
||||
return conv::implicit_gemm_problem_size(kConvolutionalOperator, problem_size).mn();
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
|
||||
|
||||
template <
|
||||
int InterleavedK,
|
||||
typename TensorRef_,
|
||||
conv::Operator ConvOperator,
|
||||
typename ConvProblemSize_
|
||||
>
|
||||
struct ConvOutputIteratorParameter<
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
layout::TensorNCxHWx<InterleavedK>,
|
||||
TensorRef_,
|
||||
ConvOperator,
|
||||
ConvProblemSize_>
|
||||
{
|
||||
|
||||
using TensorLayout = typename layout::TensorNCxHWx<InterleavedK>;
|
||||
using OutputIteratorLayout = typename layout::TensorNCxHWx<InterleavedK>;
|
||||
using OutputTensorCoord = typename OutputIteratorLayout::TensorCoord;
|
||||
using TensorRef = TensorRef_;
|
||||
static conv::Operator const kConvolutionalOperator = ConvOperator;
|
||||
using ConvProblemSize = ConvProblemSize_;
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
static OutputIteratorLayout layout(const TensorRef & ref) {
|
||||
return ref.stride();
|
||||
}
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
static OutputTensorCoord extent(ConvProblemSize problem_size) {
|
||||
return problem_size.output_extent();
|
||||
}
|
||||
|
||||
};
|
||||
|
||||
} // namespace threadblock
|
||||
} // namespace epilogue
|
||||
} // namespace cutlass
|
||||
@@ -488,6 +488,68 @@ struct InterleavedOutputTileThreadMap {
|
||||
}
|
||||
};
|
||||
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Template metaprogram for partitioning a 4D interleaved layout across warps
|
||||
/// to achieve several performance objectives:
|
||||
///
|
||||
/// - coalesced memory accesses in units of 64 Byte lines
|
||||
/// - minimal address arithmetic
|
||||
/// - minimal predicate calculations
|
||||
///
|
||||
template <typename WarpCount_, typename Iterations_, int Threads,
|
||||
int ElementsPerAccess, int ElementSize>
|
||||
struct InterleavedConvOutputTileThreadMap {
|
||||
using WarpCount = WarpCount_;
|
||||
|
||||
static int const kWarpSize = 32;
|
||||
static int const kThreads = Threads;
|
||||
static int const kWarpCount = kThreads / kWarpSize;
|
||||
|
||||
static int const kElementsPerAccess = ElementsPerAccess;
|
||||
static int const kElementSize = ElementSize;
|
||||
|
||||
//
|
||||
// Metaprogram computation
|
||||
//
|
||||
|
||||
struct Detail {};
|
||||
|
||||
//
|
||||
// Output
|
||||
//
|
||||
|
||||
using Iterations = Iterations_;
|
||||
|
||||
using Delta = MatrixShape<kWarpSize / 4, 4 * kElementsPerAccess>;
|
||||
|
||||
/// Initial offset function
|
||||
CUTLASS_HOST_DEVICE
|
||||
static MatrixCoord initial_offset(int thread_idx) {
|
||||
int warp_idx = thread_idx / kWarpSize;
|
||||
int lane_idx = thread_idx % kWarpSize;
|
||||
|
||||
// Compute warp location
|
||||
MatrixCoord warp_footprint{
|
||||
Delta::kRow * Iterations::kRow,
|
||||
Delta::kColumn * Iterations::kColumn,
|
||||
};
|
||||
|
||||
MatrixCoord warp_offset{warp_idx % WarpCount::kRow,
|
||||
warp_idx / WarpCount::kRow};
|
||||
|
||||
// Compute per-lane offset
|
||||
MatrixCoord thread_offset_in_warp{lane_idx / 4,
|
||||
(lane_idx % 4) * kElementsPerAccess};
|
||||
|
||||
MatrixCoord thread_offset_in_threadblock_tile =
|
||||
warp_footprint * warp_offset + thread_offset_in_warp;
|
||||
|
||||
return thread_offset_in_threadblock_tile;
|
||||
}
|
||||
};
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace threadblock
|
||||
|
||||
@@ -43,6 +43,7 @@
|
||||
#include "cutlass/epilogue/threadblock/output_tile_thread_map.h"
|
||||
#include "cutlass/arch/arch.h"
|
||||
#include "cutlass/arch/memory.h"
|
||||
#include "cutlass/epilogue/threadblock/predicated_tile_iterator_params.h"
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
@@ -102,68 +103,20 @@ public:
|
||||
// Parameters struct
|
||||
//
|
||||
|
||||
struct Params {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
LongIndex stride; ///< stride in bytes between rows
|
||||
|
||||
LongIndex increment_row; ///< increment quantity (in bytes) to advance when moving between rows
|
||||
LongIndex increment_group; ///< increment quantity (in bytes) to advance when moving to the next group
|
||||
LongIndex increment_cluster; ///< increment quantity (in bytes) to advance when moving to the next cluster
|
||||
|
||||
LongIndex advance_row; ///< amount to add to move to the next 'row' position
|
||||
LongIndex advance_group; ///< amount to add to move to the next 'group' position
|
||||
LongIndex advance_cluster; ///< amount to add to move to the next 'cluster' position
|
||||
LongIndex advance_tile; ///< amount to add to move to the next 'tile'
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
/// Uses a non-template class
|
||||
struct Params : PredicatedTileIteratorParams {
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Status initialize(Index stride_) {
|
||||
|
||||
stride = LongIndex(stride_);
|
||||
|
||||
increment_row = stride * ThreadMap::Delta::kRow;
|
||||
|
||||
increment_group = stride * ThreadMap::Delta::kGroup
|
||||
- stride * ThreadMap::Delta::kRow * (ThreadMap::Iterations::kRow - 1);
|
||||
|
||||
increment_cluster = stride * ThreadMap::Delta::kCluster
|
||||
- stride * ThreadMap::Delta::kGroup * (ThreadMap::Iterations::kGroup - 1)
|
||||
- stride * ThreadMap::Delta::kRow * (ThreadMap::Iterations::kRow - 1);
|
||||
|
||||
advance_row = stride * ThreadMap::Shape::kRow;
|
||||
|
||||
advance_group = stride * (ThreadMap::Shape::kGroup - 1) * ThreadMap::Shape::kRow * ThreadMap::Count::kRow;
|
||||
|
||||
advance_cluster =
|
||||
stride *
|
||||
ThreadMap::Count::kGroup * ThreadMap::Shape::kGroup * ThreadMap::Count::kRow * ThreadMap::Shape::kRow;;
|
||||
|
||||
advance_tile =
|
||||
stride *
|
||||
ThreadMap::Shape::kGroup *
|
||||
ThreadMap::Shape::kRow *
|
||||
ThreadMap::Shape::kCluster *
|
||||
ThreadMap::Shape::kTile;
|
||||
|
||||
return Status::kSuccess;
|
||||
}
|
||||
Params() { }
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params() {
|
||||
initialize(0);
|
||||
}
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params(Layout const &layout) {
|
||||
|
||||
initialize(layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess);
|
||||
Params(Layout const &layout):
|
||||
PredicatedTileIteratorParams(
|
||||
layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess,
|
||||
make_OutputTileThreadMapDesc<ThreadMap>()
|
||||
)
|
||||
{
|
||||
|
||||
}
|
||||
};
|
||||
|
||||
@@ -207,7 +160,7 @@ private:
|
||||
//
|
||||
|
||||
/// Parameters structure containing reference and precomputed state.
|
||||
Params params_;
|
||||
PredicatedTileIteratorParams params_;
|
||||
|
||||
/// Byte-level pointer
|
||||
uint8_t *byte_pointer_;
|
||||
@@ -239,12 +192,13 @@ public:
|
||||
/// Constructor
|
||||
CUTLASS_DEVICE
|
||||
PredicatedTileIterator(
|
||||
Params const & params,
|
||||
PredicatedTileIteratorParams const & params,
|
||||
Element *pointer,
|
||||
TensorCoord extent,
|
||||
int thread_idx,
|
||||
TensorCoord threadblock_offset = TensorCoord()
|
||||
): params_(params)
|
||||
):
|
||||
params_(params)
|
||||
{
|
||||
|
||||
TensorCoord thread_offset = ThreadMap::initial_offset(thread_idx) + threadblock_offset;
|
||||
@@ -745,6 +699,309 @@ public:
|
||||
};
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Tile iterator used to load output tile from shared memory in epilogue.
|
||||
///
|
||||
/// Satisfies: ReadableTileIterator | InterleavedMaskedTileIterator | ForwardTileIterator
|
||||
///
|
||||
template <
|
||||
typename ThreadMap_, ///< Thread map (conept: OutputTileThreadMap)
|
||||
typename Element_, ///< Element data type
|
||||
int InterleavedN ///< Number of Interleaved N
|
||||
>
|
||||
class InterleavedConvPredicatedTileIterator {
|
||||
public:
|
||||
using ThreadMap = ThreadMap_;
|
||||
|
||||
using Element = Element_;
|
||||
|
||||
using Layout = layout::TensorNCxHWx<InterleavedN>;
|
||||
using TensorRef = TensorRef<Element, Layout>;
|
||||
using ConstTensorRef = typename TensorRef::ConstTensorRef;
|
||||
|
||||
using Index = typename Layout::Index;
|
||||
using LongIndex = typename Layout::LongIndex;
|
||||
using TensorCoord = Tensor4DCoord;
|
||||
|
||||
static int const kElementsPerAccess = ThreadMap::kElementsPerAccess;
|
||||
static int const kThreads = ThreadMap::kThreads;
|
||||
static int const kIterations = ThreadMap::Iterations::kCount;
|
||||
|
||||
/// Fragment object
|
||||
using Fragment = Array<Element, ThreadMap::kElementsPerAccess>;
|
||||
|
||||
/// Memory access size
|
||||
using AccessType = AlignedArray<Element, ThreadMap::kElementsPerAccess>;
|
||||
|
||||
//
|
||||
// Parameters struct
|
||||
//
|
||||
|
||||
struct Params {
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
LongIndex stride_col; ///< stride in bytes between columns
|
||||
LongIndex stride_row; ///< stride in bytes between rows
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Status initialize(typename Layout::Stride stride_) {
|
||||
stride_col = stride_[1];
|
||||
stride_row = stride_[2];
|
||||
|
||||
return Status::kSuccess;
|
||||
}
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params() {
|
||||
initialize(cutlass::make_Coord(0, 0, 0));
|
||||
}
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Params(Layout const &layout) {
|
||||
|
||||
initialize(layout.stride());
|
||||
}
|
||||
};
|
||||
|
||||
/// Mask object
|
||||
struct Mask {
|
||||
static int const kCount =
|
||||
(ThreadMap::Iterations::kRow < 8) ? 8 : ThreadMap::Iterations::kRow;
|
||||
|
||||
/// Predicate state
|
||||
bool predicates[kCount];
|
||||
|
||||
//
|
||||
// Mask
|
||||
//
|
||||
CUTLASS_HOST_DEVICE
|
||||
Mask() {
|
||||
enable();
|
||||
}
|
||||
|
||||
///< Efficiently disables all accesses guarded by mask
|
||||
CUTLASS_HOST_DEVICE void clear() {
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < kCount; ++i) {
|
||||
predicates[i] = false;
|
||||
}
|
||||
}
|
||||
|
||||
///< CUTLASS_HOST_DEVICE enables all accesses guarded by mask
|
||||
CUTLASS_DEVICE void enable() {
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int i = 0; i < kCount; ++i) {
|
||||
predicates[i] = true;
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
private:
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
/// Parameters structure containing reference and precomputed state.
|
||||
Params params_;
|
||||
|
||||
/// Byte-level pointer
|
||||
uint8_t *byte_pointer_;
|
||||
|
||||
/// Array of boolean values to contain steady-state predicates
|
||||
Mask mask_;
|
||||
|
||||
/// Extent of the matrix tile in columns
|
||||
Index extent_col_;
|
||||
|
||||
/// Extent of the matrix tile in rows
|
||||
Index extent_row_;
|
||||
|
||||
/// Extent of the matrix tile in pq
|
||||
Index extent_pq_;
|
||||
|
||||
/// A thread's starting row position (assuming steady-state predicates have
|
||||
/// been computed)
|
||||
Index thread_start_row_;
|
||||
|
||||
/// A thread's starting column position (assuming steady-state predicates have
|
||||
/// been computed)
|
||||
Index thread_start_col_;
|
||||
|
||||
/// Internal iteration counter
|
||||
LongIndex iteration_row_;
|
||||
LongIndex iteration_col_;
|
||||
|
||||
uint32_t pq_mul_;
|
||||
|
||||
uint32_t pq_shr_;
|
||||
|
||||
private:
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
public:
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
/// Constructor
|
||||
CUTLASS_DEVICE
|
||||
InterleavedConvPredicatedTileIterator(
|
||||
Params const & params,
|
||||
Element *pointer,
|
||||
TensorCoord extent,
|
||||
int thread_idx,
|
||||
MatrixCoord threadblock_offset
|
||||
):
|
||||
params_(params) {
|
||||
MatrixCoord thread_offset = ThreadMap::initial_offset(thread_idx) + threadblock_offset;
|
||||
|
||||
extent_col_ = extent.c();
|
||||
extent_pq_ = extent.h() * extent.w();
|
||||
extent_row_ = extent.n() * extent_pq_;
|
||||
|
||||
find_divisor(pq_mul_, pq_shr_, extent_pq_);
|
||||
|
||||
thread_start_row_ = thread_offset.row();
|
||||
thread_start_col_ = thread_offset.column();
|
||||
|
||||
// Initialize predicates
|
||||
CUTLASS_PRAGMA_UNROLL
|
||||
for (int r = 0; r < ThreadMap::Iterations::kRow; ++r) {
|
||||
mask_.predicates[r] =
|
||||
((thread_offset.row() + ThreadMap::Delta::kRow * r) < extent_row_);
|
||||
}
|
||||
|
||||
// Initialize pointer
|
||||
byte_pointer_ = reinterpret_cast<uint8_t *>(pointer) +
|
||||
((thread_start_col_ / InterleavedN) * params_.stride_col +
|
||||
(thread_start_col_ % InterleavedN)) *
|
||||
sizeof_bits<Element>::value / 8;
|
||||
|
||||
// Initialize internal state counter
|
||||
iteration_row_ = iteration_col_ = 0;
|
||||
}
|
||||
|
||||
/// Adds a pointer offset in units of Element
|
||||
CUTLASS_HOST_DEVICE
|
||||
void add_pointer_offset(LongIndex pointer_offset) {
|
||||
byte_pointer_ += pointer_offset * sizeof_bits<Element>::value / 8;
|
||||
}
|
||||
|
||||
/// Loads a fragment from memory
|
||||
CUTLASS_DEVICE
|
||||
void load(Fragment &frag) {
|
||||
|
||||
int col_offset = iteration_col_ * ThreadMap::Delta::kColumn;
|
||||
bool col_guard = ((thread_start_col_ + col_offset) < extent_col_);
|
||||
bool guard = col_guard && mask_.predicates[iteration_row_];
|
||||
|
||||
int n, pq_rem;
|
||||
|
||||
fast_divmod(n, pq_rem,
|
||||
thread_start_row_ + iteration_row_ * ThreadMap::Delta::kRow,
|
||||
extent_pq_, pq_mul_, pq_shr_);
|
||||
|
||||
uint8_t *byte_pointer =
|
||||
byte_pointer_ + (n * params_.stride_row + pq_rem * InterleavedN) *
|
||||
sizeof_bits<Element>::value / 8;
|
||||
AccessType *frag_ptr = reinterpret_cast<AccessType *>(&frag);
|
||||
AccessType const *memory_pointer =
|
||||
reinterpret_cast<AccessType const *>(byte_pointer);
|
||||
|
||||
cutlass::arch::global_load<
|
||||
AccessType,
|
||||
sizeof(AccessType)
|
||||
>(
|
||||
*frag_ptr,
|
||||
(void *)memory_pointer,
|
||||
guard);
|
||||
}
|
||||
|
||||
/// Stores a fragment to memory
|
||||
CUTLASS_DEVICE
|
||||
void store(Fragment const &frag) {
|
||||
|
||||
int col_offset = iteration_col_ * ThreadMap::Delta::kColumn;
|
||||
bool col_guard = ((thread_start_col_ + col_offset) < extent_col_);
|
||||
bool guard = col_guard && mask_.predicates[iteration_row_];
|
||||
|
||||
int n, pq_rem;
|
||||
|
||||
fast_divmod(n, pq_rem,
|
||||
thread_start_row_ + iteration_row_ * ThreadMap::Delta::kRow,
|
||||
extent_pq_, pq_mul_, pq_shr_);
|
||||
|
||||
uint8_t *byte_pointer =
|
||||
byte_pointer_ + (n * params_.stride_row + pq_rem * InterleavedN) *
|
||||
sizeof_bits<Element>::value / 8;
|
||||
AccessType const *frag_ptr = reinterpret_cast<AccessType const *>(&frag);
|
||||
AccessType *memory_pointer = reinterpret_cast<AccessType *>(byte_pointer);
|
||||
|
||||
if (guard) {
|
||||
*memory_pointer = *frag_ptr;
|
||||
}
|
||||
}
|
||||
|
||||
/// Overrides the internal iteration index
|
||||
CUTLASS_HOST_DEVICE
|
||||
void set_iteration_index(int iteration) {
|
||||
iteration_row_ = iteration % ThreadMap::Iterations::kRow;
|
||||
iteration_col_ = iteration / ThreadMap::Iterations::kRow;
|
||||
}
|
||||
|
||||
/// Advances to the next position to load or store
|
||||
CUTLASS_HOST_DEVICE
|
||||
InterleavedConvPredicatedTileIterator &operator++() {
|
||||
|
||||
++iteration_row_;
|
||||
|
||||
if (iteration_row_ == ThreadMap::Iterations::kRow) {
|
||||
|
||||
iteration_row_ = 0;
|
||||
++iteration_col_;
|
||||
byte_pointer_ += params_.stride_col;
|
||||
|
||||
if (iteration_col_ == ThreadMap::Iterations::kColumn) {
|
||||
iteration_col_ = 0;
|
||||
}
|
||||
}
|
||||
|
||||
return *this;
|
||||
}
|
||||
|
||||
///< Efficiently disables all accesses guarded by mask
|
||||
CUTLASS_DEVICE void clear_mask() {
|
||||
mask_.clear();
|
||||
}
|
||||
|
||||
///< Efficiently enables all accesses guarded by mask
|
||||
CUTLASS_DEVICE void enable_mask() {
|
||||
mask_.enable();
|
||||
}
|
||||
|
||||
///< Sets the mask
|
||||
CUTLASS_DEVICE void get_mask(Mask &mask) {
|
||||
return mask_;
|
||||
}
|
||||
|
||||
///< Sets the mask
|
||||
CUTLASS_DEVICE void set_mask(Mask const &mask) {
|
||||
mask_ = mask;
|
||||
}
|
||||
};
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace threadblock
|
||||
|
||||
@@ -0,0 +1,227 @@
|
||||
/***************************************************************************************************
|
||||
* Copyright (c) 2017-2020, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* Redistribution and use in source and binary forms, with or without modification, are permitted
|
||||
* provided that the following conditions are met:
|
||||
* * Redistributions of source code must retain the above copyright notice, this list of
|
||||
* conditions and the following disclaimer.
|
||||
* * Redistributions in binary form must reproduce the above copyright notice, this list of
|
||||
* conditions and the following disclaimer in the documentation and/or other materials
|
||||
* provided with the distribution.
|
||||
* * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
|
||||
* to endorse or promote products derived from this software without specific prior written
|
||||
* permission.
|
||||
*
|
||||
* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
|
||||
* IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
|
||||
* FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
|
||||
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
|
||||
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
|
||||
* OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
|
||||
* STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
||||
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
||||
*
|
||||
**************************************************************************************************/
|
||||
/*! \file
|
||||
\brief
|
||||
*/
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "cutlass/cutlass.h"
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
namespace cutlass {
|
||||
namespace epilogue {
|
||||
namespace threadblock {
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
struct OutputTileShapeDesc {
|
||||
|
||||
int column;
|
||||
int row;
|
||||
int group;
|
||||
int cluster;
|
||||
int tile;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
/// Default ctor
|
||||
CUTLASS_HOST_DEVICE
|
||||
OutputTileShapeDesc(): column(0), row(0), group(0), cluster(0), tile(0) { }
|
||||
|
||||
/// Ctor
|
||||
CUTLASS_HOST_DEVICE
|
||||
OutputTileShapeDesc(
|
||||
int column_,
|
||||
int row_,
|
||||
int group_,
|
||||
int cluster_,
|
||||
int tile_
|
||||
):
|
||||
column(column_),
|
||||
row(row_),
|
||||
group(group_),
|
||||
cluster(cluster_),
|
||||
tile(tile_) { }
|
||||
|
||||
/// Total number of points in the 5D space
|
||||
CUTLASS_HOST_DEVICE
|
||||
int count() const {
|
||||
return column * row * group * cluster * tile;
|
||||
}
|
||||
};
|
||||
|
||||
/// Helper template to construct an OutputTileShapeDesc from a OutputTileShape template.
|
||||
template <typename Shape>
|
||||
CUTLASS_HOST_DEVICE
|
||||
OutputTileShapeDesc make_OutputTileShapeDesc() {
|
||||
return OutputTileShapeDesc(
|
||||
Shape::kColumn,
|
||||
Shape::kRow,
|
||||
Shape::kGroup,
|
||||
Shape::kCluster,
|
||||
Shape::kTile
|
||||
);
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
/// Thread map description
|
||||
struct OutputTileThreadMapDesc {
|
||||
|
||||
int threads;
|
||||
int elements_per_access;
|
||||
OutputTileShapeDesc shape;
|
||||
OutputTileShapeDesc iterations;
|
||||
OutputTileShapeDesc delta;
|
||||
OutputTileShapeDesc count;
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
OutputTileThreadMapDesc() { }
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
OutputTileThreadMapDesc(
|
||||
int threads_,
|
||||
int elements_per_access_,
|
||||
OutputTileShapeDesc shape_,
|
||||
OutputTileShapeDesc iterations_,
|
||||
OutputTileShapeDesc delta_,
|
||||
OutputTileShapeDesc count_
|
||||
):
|
||||
threads(threads_),
|
||||
elements_per_access(elements_per_access_),
|
||||
shape(shape_),
|
||||
iterations(iterations_),
|
||||
delta(delta_),
|
||||
count(count_) { }
|
||||
};
|
||||
|
||||
/// Helper template to construct an OutputTileShapeDesc from a OutputTileThreadMap template.
|
||||
template <typename ThreadMap>
|
||||
CUTLASS_HOST_DEVICE
|
||||
OutputTileThreadMapDesc make_OutputTileThreadMapDesc() {
|
||||
return OutputTileThreadMapDesc(
|
||||
ThreadMap::kThreads,
|
||||
ThreadMap::kElementsPerAccess,
|
||||
make_OutputTileShapeDesc<typename ThreadMap::Shape>(),
|
||||
make_OutputTileShapeDesc<typename ThreadMap::Iterations>(),
|
||||
make_OutputTileShapeDesc<typename ThreadMap::Delta>(),
|
||||
make_OutputTileShapeDesc<typename ThreadMap::Count>()
|
||||
);
|
||||
}
|
||||
|
||||
/////////////////////////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
//
|
||||
// Parameters struct
|
||||
//
|
||||
|
||||
struct PredicatedTileIteratorParams {
|
||||
|
||||
using Index = int32_t;
|
||||
using LongIndex = int64_t;
|
||||
|
||||
//
|
||||
// Data members
|
||||
//
|
||||
|
||||
LongIndex stride; ///< stride in bytes between rows
|
||||
|
||||
LongIndex increment_row; ///< increment quantity (in bytes) to advance when moving between rows
|
||||
LongIndex increment_group; ///< increment quantity (in bytes) to advance when moving to the next group
|
||||
LongIndex increment_cluster; ///< increment quantity (in bytes) to advance when moving to the next cluster
|
||||
|
||||
LongIndex advance_row; ///< amount to add to move to the next 'row' position
|
||||
LongIndex advance_group; ///< amount to add to move to the next 'group' position
|
||||
LongIndex advance_cluster; ///< amount to add to move to the next 'cluster' position
|
||||
LongIndex advance_tile; ///< amount to add to move to the next 'tile'
|
||||
|
||||
//
|
||||
// Methods
|
||||
//
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
Status initialize(Index stride_, OutputTileThreadMapDesc thread_map) {
|
||||
|
||||
stride = LongIndex(stride_);
|
||||
|
||||
increment_row = stride * thread_map.delta.row;
|
||||
|
||||
increment_group = stride * thread_map.delta.group
|
||||
- stride * thread_map.delta.row * (thread_map.iterations.row - 1);
|
||||
|
||||
increment_cluster = stride * thread_map.delta.cluster
|
||||
- stride * thread_map.delta.group * (thread_map.iterations.group - 1)
|
||||
- stride * thread_map.delta.row * (thread_map.iterations.row - 1);
|
||||
|
||||
advance_row = stride * thread_map.shape.row;
|
||||
|
||||
advance_group =
|
||||
stride *
|
||||
(thread_map.shape.group - 1) * thread_map.shape.row * thread_map.count.row;
|
||||
|
||||
advance_cluster =
|
||||
stride *
|
||||
thread_map.count.group *
|
||||
thread_map.shape.group *
|
||||
thread_map.count.row *
|
||||
thread_map.shape.row;
|
||||
|
||||
advance_tile =
|
||||
stride *
|
||||
thread_map.shape.group *
|
||||
thread_map.shape.row *
|
||||
thread_map.shape.cluster *
|
||||
thread_map.shape.tile;
|
||||
|
||||
return Status::kSuccess;
|
||||
}
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
PredicatedTileIteratorParams() {
|
||||
initialize(0, OutputTileThreadMapDesc());
|
||||
}
|
||||
|
||||
CUTLASS_HOST_DEVICE
|
||||
PredicatedTileIteratorParams(Index stride, OutputTileThreadMapDesc thread_map) {
|
||||
|
||||
initialize(stride, thread_map);
|
||||
}
|
||||
};
|
||||
|
||||
///////////////////////////////////////////////////////////////////////////////
|
||||
|
||||
} // namespace threadblock
|
||||
} // namespace epilogue
|
||||
} // namespace cutlass
|
||||
|
||||
////////////////////////////////////////////////////////////////////////////////
|
||||
Reference in New Issue
Block a user