Spaces:
Sleeping
Sleeping
File size: 5,895 Bytes
e45d058 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 |
/***************************************************************************************************
* Copyright (c) 2024 - 2024 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.
*
**************************************************************************************************/
#pragma once
#include <cute/config.hpp>
#include <cute/tensor.hpp>
#include <cute/atom/copy_atom.hpp>
namespace cute
{
//
// Prefetch global tensors into L2
//
template <uint32_t NumThreads, uint32_t FetchBytes = 64,
class GEngine, class GLayout>
CUTE_HOST_DEVICE
void
cooperative_prefetch(uint32_t const& tid,
Tensor<GEngine, GLayout> const& src)
{
static_assert(is_gmem<GEngine>::value, "Expected global tensor for prefetch");
constexpr int V = decltype(max_common_vector(src, src))::value;
if constexpr (V > 1) {
// L2 sector is 32B, default fetch granularity is 64B
using VecType = conditional_t<(V * sizeof_bits_v<typename GEngine::value_type>) < (FetchBytes * 8),
ArrayEngine<typename GEngine::value_type, V>,
uint8_t[FetchBytes] >;
Tensor src_v = recast<VecType const>(src);
CUTE_UNROLL
for (int i = tid; i < size(src_v); i += NumThreads) {
prefetch(raw_pointer_cast(&src_v(i)));
}
} else {
CUTE_UNROLL
for (int i = tid; i < size(src); i += NumThreads) {
prefetch(raw_pointer_cast(&src(i)));
}
}
}
template <class GEngine, class GLayout>
CUTE_HOST_DEVICE
void
prefetch(Tensor<GEngine, GLayout> const& src)
{
return cooperative_prefetch<1>(0, src);
}
// Prefetch with copy atom
namespace detail {
template <class CopyOp, class = void>
constexpr bool has_prefetch = false;
template <class CopyOp>
constexpr bool has_prefetch<CopyOp, void_t<typename CopyOp::PREFETCH>> = true;
template <class CopyOp, class = void>
constexpr bool is_prefetch = false;
template <class CopyOp>
constexpr bool is_prefetch<CopyOp, void_t<typename CopyOp::PREFETCH>> = is_same_v<CopyOp, typename CopyOp::PREFETCH>;
} // end namespace detail
template <class CopyOp, class... CT_Args, class... CA_Args,
class GEngine, class GLayout>
CUTE_HOST_DEVICE
void
prefetch(Copy_Atom<Copy_Traits<CopyOp, CT_Args...>, CA_Args...> const& atom,
Tensor<GEngine, GLayout> const& src)
{
if constexpr (detail::has_prefetch<CopyOp>) {
using Prefetch_Traits = Copy_Traits<typename CopyOp::PREFETCH, CT_Args...>;
using Prefetch_Atom = Copy_Atom<Prefetch_Traits, CA_Args...>;
Prefetch_Atom prefetch_atom{atom};
auto& dst = const_cast<Tensor<GEngine, GLayout>&>(src); // dst is ignored for prefetch atoms
return copy(prefetch_atom, src, dst);
} else {
return prefetch(src);
}
}
#if defined(CUTE_COPY_ATOM_TMA_SM90_ENABLED)
template <class... CT_Args,
class SrcEngine, class SrcLayout>
CUTE_HOST_DEVICE
void
prefetch(Copy_Traits<SM90_BULK_COPY_AUTO, CT_Args...> const& atom,
Tensor<SrcEngine, SrcLayout> const& src)
{
using SrcType = typename SrcEngine::value_type;
static_assert(is_gmem<SrcEngine>::value, "Expected global tensor for L2 prefetch");
auto tiler = max_common_layout(src, src);
constexpr int vec_elem = decltype(size(tiler))::value;
constexpr int vec_bits = vec_elem * sizeof_bits_v<SrcType>;
static_assert(vec_bits >= 128, "Expected at least 128-bits for BLKCP");
// Construct a new concrete Atom of the vector size
auto bulk_atom = Copy_Atom<Copy_Traits<SM90_BULK_COPY_G2S, Int<vec_bits>>, SrcType>{};
return prefetch(bulk_atom, logical_divide(src, tiler));
}
// Backwards-compat. Throw out any extra Copy_Atom args.
template <class... CT_Args, class... CA_Args,
class SrcEngine, class SrcLayout>
CUTE_HOST_DEVICE
void
prefetch(Copy_Atom<Copy_Traits<SM90_BULK_COPY_AUTO, CT_Args...>, CA_Args...> const& atom,
Tensor<SrcEngine, SrcLayout> const& src)
{
return prefetch(static_cast<Copy_Traits<SM90_BULK_COPY_AUTO, CT_Args...> const&>(atom), src);
}
#endif // #if defined(CUTE_COPY_ATOM_TMA_SM90_ENABLED)
} // end namespace cute
|