310 lines
8.9 KiB
C++
310 lines
8.9 KiB
C++
/***************************************************************************************************
|
|
* Copyright (c) 2017 - 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
|
|
* SPDX-License-Identifier: BSD-3-Clause
|
|
*
|
|
* Redistribution and use in source and binary forms, with or without
|
|
* modification, are permitted provided that the following conditions are met:
|
|
*
|
|
* 1. Redistributions of source code must retain the above copyright notice, this
|
|
* list of conditions and the following disclaimer.
|
|
*
|
|
* 2. 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.
|
|
*
|
|
* 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (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 PredicatedTileIteratorPredicates.
|
|
|
|
PredicatedTileIteratorPredicates enables both upper and lower bounds for predicates.
|
|
|
|
*/
|
|
|
|
#pragma once
|
|
|
|
#include "cutlass/cutlass.h"
|
|
#include "cutlass/numeric_types.h"
|
|
#include "cutlass/array.h"
|
|
#include "cutlass/layout/matrix.h"
|
|
#include "cutlass/layout/tensor.h"
|
|
#include "cutlass/matrix_shape.h"
|
|
#include "cutlass/tensor_ref.h"
|
|
#include "cutlass/transform/pitch_linear_thread_map.h"
|
|
#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"
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace cutlass {
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
namespace epilogue {
|
|
namespace threadblock {
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|
|
|
|
/// Tile iterator predicates used to bound computations in epilogue.
|
|
///
|
|
/// Satisfies: ReadableTileIterator | PredicatedTileIterator | ForwardTileIterator
|
|
///
|
|
template <
|
|
typename ThreadMap_, ///< Thread map (conept: OutputTileThreadMap)
|
|
typename Element_ ///< Element data type
|
|
>
|
|
class PredicatedTileIteratorPredicates {
|
|
public:
|
|
using ThreadMap = ThreadMap_;
|
|
using Shape = typename ThreadMap::Shape;
|
|
|
|
using Element = Element_;
|
|
|
|
using Layout = layout::RowMajor;
|
|
using TensorRef = TensorRef<Element, Layout>;
|
|
using ConstTensorRef = typename TensorRef::ConstTensorRef;
|
|
|
|
using Index = typename Layout::Index;
|
|
using LongIndex = typename Layout::LongIndex;
|
|
using TensorCoord = MatrixCoord;
|
|
|
|
static int const kElementsPerAccess = ThreadMap::kElementsPerAccess;
|
|
static int const kThreads = ThreadMap::kThreads;
|
|
static int const kIterations = ThreadMap::Count::kTile;
|
|
|
|
static_assert( ThreadMap::Iterations::kRow > 0,"ThreadMap::Iterations::kRow must be > 0");
|
|
static_assert( ThreadMap::Iterations::kGroup > 0,"ThreadMap::Iterations::kGroup must be > 0");
|
|
static_assert( ThreadMap::Iterations::kCluster > 0,"ThreadMap::Iterations::kCluster must be > 0");
|
|
static_assert( ThreadMap::Iterations::kColumn > 0,"ThreadMap::Iterations::kColumn must be > 0");
|
|
|
|
/// Fragment object
|
|
using Fragment = Array<
|
|
Element,
|
|
ThreadMap::Iterations::kColumn *
|
|
ThreadMap::Iterations::kRow *
|
|
ThreadMap::Iterations::kGroup *
|
|
ThreadMap::Iterations::kCluster * ThreadMap::kElementsPerAccess>;
|
|
|
|
/// Memory access size
|
|
using AccessType = AlignedArray<Element, ThreadMap::kElementsPerAccess>;
|
|
|
|
//
|
|
// Parameters struct
|
|
//
|
|
|
|
/// Uses a non-template class
|
|
struct Params : PredicatedTileIteratorParams {
|
|
|
|
CUTLASS_HOST_DEVICE
|
|
Params() { }
|
|
|
|
CUTLASS_HOST_DEVICE
|
|
Params(Layout const &layout):
|
|
PredicatedTileIteratorParams(
|
|
layout.stride(0) * int(sizeof(AccessType)) / kElementsPerAccess,
|
|
make_OutputTileThreadMapDesc<ThreadMap>()
|
|
)
|
|
{
|
|
|
|
}
|
|
};
|
|
|
|
/// Mask object
|
|
struct Mask {
|
|
|
|
static int const kCount = ThreadMap::Iterations::kColumn;
|
|
|
|
/// 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.
|
|
PredicatedTileIteratorParams params_;
|
|
|
|
/// Array of boolean values to contain steady-state predicates
|
|
Mask mask_;
|
|
|
|
/// Extent of the matrix tile in rows
|
|
Index lower_extent_row_;
|
|
Index upper_extent_row_;
|
|
|
|
/// A thread's starting row position (assuming steady-state predicates have been computed)
|
|
Index thread_start_row_;
|
|
|
|
/// Internal state counter
|
|
int state_[3];
|
|
|
|
//
|
|
// Static asserts about internal strides
|
|
//
|
|
|
|
static_assert(sizeof(lower_extent_row_) == 4, "Expected 32b extents");
|
|
static_assert(sizeof(upper_extent_row_) == 4, "Expected 32b extents");
|
|
static_assert(sizeof(thread_start_row_) == 4, "Expected 32b extents");
|
|
static_assert(sizeof(PredicatedTileIteratorParams::stride) == 8, "Expected 64b strides");
|
|
|
|
private:
|
|
|
|
//
|
|
// Methods
|
|
//
|
|
|
|
public:
|
|
|
|
//
|
|
// Methods
|
|
//
|
|
|
|
/// Constructor
|
|
CUTLASS_DEVICE
|
|
PredicatedTileIteratorPredicates(
|
|
PredicatedTileIteratorParams const & params,
|
|
TensorCoord lower_extent,
|
|
TensorCoord upper_extent,
|
|
int thread_idx,
|
|
TensorCoord threadblock_offset = TensorCoord()
|
|
):
|
|
params_(params)
|
|
{
|
|
|
|
TensorCoord thread_offset = ThreadMap::initial_offset(thread_idx) + threadblock_offset;
|
|
|
|
lower_extent_row_ = lower_extent.row();
|
|
upper_extent_row_ = upper_extent.row();
|
|
thread_start_row_ = thread_offset.row();
|
|
|
|
// Initialize predicates
|
|
CUTLASS_PRAGMA_UNROLL
|
|
for (int c = 0; c < ThreadMap::Iterations::kColumn; ++c) {
|
|
|
|
mask_.predicates[c] = ((thread_offset.column()
|
|
+ ThreadMap::Delta::kColumn * c) < upper_extent.column()) &&
|
|
((thread_offset.column() + ThreadMap::Delta::kColumn * c) >= lower_extent.column());
|
|
}
|
|
|
|
// Initialize internal state counter
|
|
state_[0] = state_[1] = state_[2] = 0;
|
|
}
|
|
|
|
/// Advances to the next position to load or store
|
|
CUTLASS_HOST_DEVICE
|
|
PredicatedTileIteratorPredicates &operator++() {
|
|
|
|
++state_[0];
|
|
thread_start_row_ += ThreadMap::Shape::kRow;
|
|
|
|
if (state_[0] == ThreadMap::Count::kRow) {
|
|
|
|
state_[0] = 0;
|
|
++state_[1];
|
|
|
|
thread_start_row_ += (ThreadMap::Shape::kGroup - 1) *
|
|
ThreadMap::Shape::kRow * ThreadMap::Count::kRow;
|
|
|
|
if (state_[1] == ThreadMap::Count::kGroup) {
|
|
|
|
state_[1] = 0;
|
|
++state_[2];
|
|
|
|
thread_start_row_ += ThreadMap::Count::kGroup *
|
|
ThreadMap::Shape::kGroup * ThreadMap::Count::kRow * ThreadMap::Shape::kRow;
|
|
|
|
if (state_[2] == ThreadMap::Count::kCluster) {
|
|
state_[2] = 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();
|
|
}
|
|
|
|
///< Gets the mask
|
|
CUTLASS_DEVICE void get_mask(Mask &mask) {
|
|
mask = mask_;
|
|
}
|
|
|
|
///< Sets the mask
|
|
CUTLASS_DEVICE void set_mask(Mask const &mask) {
|
|
mask_ = mask;
|
|
}
|
|
|
|
///< Gets lower_extent_row_
|
|
CUTLASS_DEVICE Index get_lower_extent_row() {
|
|
return lower_extent_row_;
|
|
}
|
|
|
|
///< Gets upper_extent_row_
|
|
CUTLASS_DEVICE Index get_upper_extent_row() {
|
|
return upper_extent_row_;
|
|
}
|
|
|
|
///< Gets thread_start_row_
|
|
CUTLASS_DEVICE Index get_thread_start_row() {
|
|
return thread_start_row_;
|
|
}
|
|
};
|
|
|
|
///////////////////////////////////////////////////////////////////////////////
|
|
|
|
} // namespace threadblock
|
|
} // namespace epilogue
|
|
} // namespace cutlass
|
|
|
|
////////////////////////////////////////////////////////////////////////////////
|