mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 14:59:17 +00:00
[Ck tile] support rmsnorm and related fusion (#1605)
* Add reduce2d new api
* Prevent user use cross warp reduction
* Fix bug of std caculation
* Add rmsnorm2d
* Add rmsnorm small example
* Remove static assert to prevent compile fail
* Add script to test performance and correctness
* Add missing cmake change
* refine naming
* refine example of rmsnorm
* Fix bug of rmsnorm
* Refine naming
* Fix cmake
* clang format
* Refine pipeline name
* Add add_rmsnorm2d_rdquant kernel
* Add reduce op
* host verification
* Fix bug of one pass pipeline
* Refine tile size
* Add two pass pipeline
* Rename two pass to three pass
* Fix bug of kSaveX == false
* Add instance library
* Add test script
* Fix bug of x verification
* Add save_x to trait
* Add README
* Move reduce2d into reduce folder
* Fix bug of welford when number of m warp > 1
* remove reduncant comment
* 1. move 06_rmsnorm2d to 10_rmsnorm2d
2. move 07_add_rmsnorm2d_rdquant to 11_add_rmsnorm2d_rdquant
* clang format and add missing header
* Add host validation of add + layernorm2d + rsquant
* Revert "Add host validation of add + layernorm2d + rsquant"
This reverts commit 936cb45797.
* Remove deprecated flag
This commit is contained in:
rocking
@@ -0,0 +1,239 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/common.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// host side args
|
||||
struct AddRmsnorm2dRdquantFwdHostArgs
|
||||
{
|
||||
const void* p_a;
|
||||
const void* p_b;
|
||||
const void* p_gamma;
|
||||
|
||||
void* p_x;
|
||||
void* p_yscale;
|
||||
void* p_qy;
|
||||
|
||||
float epsilon;
|
||||
|
||||
index_t m;
|
||||
index_t n;
|
||||
index_t stride; // row_stride
|
||||
};
|
||||
|
||||
// TODO: Extract some type to wrapper class
|
||||
template <typename Pipeline_>
|
||||
struct AddRmsnorm2dRdquantFwd
|
||||
{
|
||||
using Pipeline = remove_cvref_t<Pipeline_>;
|
||||
using Problem = typename Pipeline::Problem;
|
||||
|
||||
using ADataType = remove_cvref_t<typename Problem::ADataType>;
|
||||
using BDataType = remove_cvref_t<typename Problem::BDataType>;
|
||||
using GammaDataType = remove_cvref_t<typename Problem::GammaDataType>;
|
||||
using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
|
||||
using XDataType = remove_cvref_t<typename Problem::XDataType>;
|
||||
using YScaleDataType = remove_cvref_t<typename Problem::YScaleDataType>;
|
||||
using QYDataType = remove_cvref_t<typename Problem::QYDataType>;
|
||||
|
||||
static constexpr bool kSaveX = Problem::kSaveX;
|
||||
|
||||
static constexpr index_t Block_M = Problem::BlockShape::Block_M;
|
||||
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
|
||||
static constexpr bool kPadM = false; // always no need to pad along M
|
||||
static constexpr bool kPadN = Problem::kPadN;
|
||||
static constexpr bool kThreePass = Problem::kThreePass;
|
||||
|
||||
static constexpr index_t ThreadPerWarp_N = Problem::BlockShape::ThreadPerWarp_N;
|
||||
static constexpr index_t Vector_N = Problem::BlockShape::Vector_N;
|
||||
static constexpr index_t Repeat_N = Problem::BlockShape::Repeat_N;
|
||||
|
||||
static constexpr auto I0 = number<0>{};
|
||||
static constexpr auto I1 = number<1>{};
|
||||
|
||||
struct Kargs
|
||||
{
|
||||
const void* p_a;
|
||||
const void* p_b;
|
||||
const void* p_gamma;
|
||||
|
||||
void* p_x;
|
||||
void* p_yscale;
|
||||
void* p_qy;
|
||||
|
||||
float epsilon;
|
||||
|
||||
index_t m;
|
||||
index_t n;
|
||||
index_t stride; // row_stride
|
||||
};
|
||||
using Hargs = AddRmsnorm2dRdquantFwdHostArgs;
|
||||
|
||||
CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
|
||||
{
|
||||
return Kargs{hargs.p_a,
|
||||
hargs.p_b,
|
||||
hargs.p_gamma,
|
||||
hargs.p_x,
|
||||
hargs.p_yscale,
|
||||
hargs.p_qy,
|
||||
hargs.epsilon,
|
||||
hargs.m,
|
||||
hargs.n,
|
||||
hargs.stride};
|
||||
}
|
||||
|
||||
CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
|
||||
{
|
||||
return integer_divide_ceil(hargs.m, Block_M);
|
||||
}
|
||||
|
||||
CK_TILE_HOST static constexpr auto BlockSize() { return Problem::BlockShape::BlockSize; }
|
||||
|
||||
// clang-format off
|
||||
template <typename T> struct t2s;
|
||||
template <> struct t2s<float> { static constexpr const char * name = "fp32"; };
|
||||
template <> struct t2s<ck_tile::fp16_t> { static constexpr const char * name = "fp16"; };
|
||||
template <> struct t2s<ck_tile::bf16_t> { static constexpr const char * name = "bf16"; };
|
||||
template <> struct t2s<ck_tile::fp8_t> { static constexpr const char * name = "fp8"; };
|
||||
template <> struct t2s<ck_tile::bf8_t> { static constexpr const char * name = "bf8"; };
|
||||
// clang-format on
|
||||
|
||||
// in byte
|
||||
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize() { return Pipeline::GetSmemSize(); }
|
||||
|
||||
CK_TILE_HOST static std::string GetName()
|
||||
{
|
||||
// clang-format off
|
||||
using S_ = typename Problem::BlockShape;
|
||||
auto surfix = [&] () {
|
||||
std::string n;
|
||||
if (kPadN) n += "_pn";
|
||||
if (kSaveX) n += "_x";
|
||||
if (kThreePass) n += "_2p";
|
||||
return n; }();
|
||||
|
||||
#define _SS_ std::string
|
||||
#define _TS_ std::to_string
|
||||
return _SS_("add_rmsnorm2d_rdquant_fwd_") + _SS_(t2s<XDataType>::name) + "_" +
|
||||
_TS_(S_::Block_M) + "x" + _TS_(S_::Block_N) + "_" + _TS_(S_::WarpPerBlock_M) + "x" + _TS_(S_::WarpPerBlock_N) + "_" +
|
||||
_TS_(S_::Warp_M) + "x" + _TS_(S_::Warp_N) + "_" + _TS_(S_::Vector_M) + "x" + _TS_(S_::Vector_N) + "_" +
|
||||
_SS_(Pipeline::name) + surfix;
|
||||
#undef _SS_
|
||||
#undef _TS_
|
||||
// clang-format on
|
||||
}
|
||||
|
||||
CK_TILE_DEVICE void operator()(Kargs kargs) const
|
||||
{
|
||||
const auto iM = get_block_id() * Block_M;
|
||||
|
||||
const auto a_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<const ADataType*>(kargs.p_a),
|
||||
make_tuple(kargs.m, kargs.n),
|
||||
make_tuple(kargs.stride, 1),
|
||||
number<Vector_N>{},
|
||||
number<1>{});
|
||||
|
||||
const auto tmp2_ = pad_tensor_view(
|
||||
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
|
||||
return make_tile_window(
|
||||
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
|
||||
}();
|
||||
|
||||
const auto b_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<const BDataType*>(kargs.p_b),
|
||||
make_tuple(kargs.m, kargs.n),
|
||||
make_tuple(kargs.stride, 1),
|
||||
number<Vector_N>{},
|
||||
number<1>{});
|
||||
|
||||
const auto tmp2_ = pad_tensor_view(
|
||||
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
|
||||
return make_tile_window(
|
||||
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
|
||||
}();
|
||||
|
||||
const auto gamma_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<const GammaDataType*>(kargs.p_gamma),
|
||||
make_tuple(kargs.n),
|
||||
make_tuple(1),
|
||||
number<Vector_N>{},
|
||||
number<1>{});
|
||||
|
||||
const auto tmp2_ =
|
||||
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<kPadM>{});
|
||||
|
||||
return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
|
||||
}();
|
||||
|
||||
auto x_window = [&]() {
|
||||
if constexpr(kSaveX)
|
||||
{
|
||||
const auto tmp2_ = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<XDataType*>(kargs.p_x),
|
||||
make_tuple(kargs.m, kargs.n),
|
||||
make_tuple(kargs.stride, 1),
|
||||
number<Vector_N>{},
|
||||
number<1>{});
|
||||
|
||||
return pad_tensor_view(tmp_,
|
||||
make_tuple(number<Block_M>{}, number<Block_N>{}),
|
||||
sequence<kPadM, kPadN>{});
|
||||
}();
|
||||
return make_tile_window(
|
||||
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
|
||||
}
|
||||
else
|
||||
return make_null_tile_window(make_tuple(number<Block_M>{}, number<Block_N>{}));
|
||||
}();
|
||||
|
||||
auto yscale_window = [&]() {
|
||||
auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<YScaleDataType*>(kargs.p_yscale),
|
||||
make_tuple(kargs.m),
|
||||
make_tuple(1),
|
||||
number<1>{});
|
||||
|
||||
auto tmp2_ = pad_tensor_view(tmp_, make_tuple(number<Block_M>{}), sequence<kPadM>{});
|
||||
return make_tile_window(tmp2_, make_tuple(number<Block_M>{}), {iM});
|
||||
}();
|
||||
|
||||
auto qy_window = [&]() {
|
||||
auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<QYDataType*>(kargs.p_qy),
|
||||
make_tuple(kargs.m, kargs.n),
|
||||
make_tuple(kargs.stride, 1),
|
||||
number<Vector_N>{},
|
||||
number<1>{});
|
||||
|
||||
auto tmp2_ = pad_tensor_view(
|
||||
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
|
||||
return make_tile_window(
|
||||
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
|
||||
}();
|
||||
|
||||
__shared__ char smem[GetSmemSize()];
|
||||
|
||||
Pipeline{}(a_window,
|
||||
b_window,
|
||||
gamma_window,
|
||||
x_window,
|
||||
yscale_window,
|
||||
qy_window,
|
||||
static_cast<const ComputeDataType>(kargs.epsilon),
|
||||
kargs.n,
|
||||
smem);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,78 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
/*
|
||||
// clang-format off
|
||||
|
||||
4-level descriptor: BlockTile-> WarpPerBlock-> WarpTile-> Vector
|
||||
|
||||
Block_N (Warp_N * WarpPerBlock_N * Repeat_N )
|
||||
+<----------------------< Repeat_N(2)>--------------------->+
|
||||
| |
|
||||
+<-- <WarpPerBlock_N(2)> -->+
|
||||
Warp_N
|
||||
+--------------+--------------+--------------+--------------+----+----------------+
|
||||
Warp_M | wrap_0 | wrap_1 | | ^ ^
|
||||
+--------------+--------------+ | <WarpPerBlock_M(2)> |
|
||||
| wrap_2 | wrap_3 | | v
|
||||
+--------------+--------------+--------------+--------------+----+ Block_M
|
||||
| | |
|
||||
+ + |
|
||||
| | | v
|
||||
+--------------+--------------+--------------+--------------+ +
|
||||
|
||||
each Warp-tile (e.g 16 thrd per row)
|
||||
|
||||
Vector_N (contiguous pixels each thrd holds along N, or vector size)
|
||||
+-----------+-----------+-----------+-----------+-----------+
|
||||
| thrd_0 | thrd_1 | thrd_2 | thrd_3 | ... Vector_M
|
||||
+-----------+-----------+-----------+-----------+-----------+
|
||||
| thrd_16 | thrd_17 | thrd_18 | thrd_19 | ...
|
||||
+-----------+-----------+-----------+-----------+-----------+
|
||||
// clang-format on
|
||||
*/
|
||||
template <typename BlockTile_, // block size, seq<M, N>
|
||||
typename WarpPerBlock_, // num warps along seq<M, N>
|
||||
typename WarpTile_, // warp size, seq<M, N>
|
||||
typename Vector_, // contiguous pixels(vector size) along seq<M, N>
|
||||
index_t BlockSize_ =
|
||||
warpSize* reduce_on_sequence(WarpPerBlock_{}, multiplies{}, number<1>{})>
|
||||
struct AddRmsnorm2dRdquantShape
|
||||
{
|
||||
// block size
|
||||
static constexpr index_t Block_M = BlockTile_::at(number<0>{});
|
||||
static constexpr index_t Block_N = BlockTile_::at(number<1>{});
|
||||
|
||||
// num warps along seq<M, N>, within each block
|
||||
static constexpr index_t WarpPerBlock_M = WarpPerBlock_::at(number<0>{});
|
||||
static constexpr index_t WarpPerBlock_N = WarpPerBlock_::at(number<1>{});
|
||||
|
||||
// warp size
|
||||
static constexpr index_t Warp_M = WarpTile_::at(number<0>{});
|
||||
static constexpr index_t Warp_N = WarpTile_::at(number<1>{});
|
||||
|
||||
static_assert(Block_M % (WarpPerBlock_M * Warp_M) == 0);
|
||||
static_assert(Block_N % (WarpPerBlock_N * Warp_N) == 0);
|
||||
// repeat of each thread along seq<M, N>
|
||||
static constexpr index_t Repeat_M = Block_M / (WarpPerBlock_M * Warp_M);
|
||||
static constexpr index_t Repeat_N = Block_N / (WarpPerBlock_N * Warp_N);
|
||||
|
||||
// vector size along seq<M, N>
|
||||
static constexpr index_t Vector_M = Vector_::at(number<0>{});
|
||||
static constexpr index_t Vector_N = Vector_::at(number<1>{});
|
||||
|
||||
static_assert(Warp_M % Vector_M == 0);
|
||||
static_assert(Warp_N % Vector_N == 0);
|
||||
// num of threads along seq<M, N>, within each warp
|
||||
static constexpr index_t ThreadPerWarp_M = Warp_M / Vector_M;
|
||||
static constexpr index_t ThreadPerWarp_N = Warp_N / Vector_N;
|
||||
|
||||
static constexpr index_t BlockSize = BlockSize_;
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,94 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/reduce/block/block_reduce2d_problem.hpp"
|
||||
#include "ck_tile/ops/reduce/block/block_reduce2d.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
struct AddRmsnorm2dRdquantFwdPipelineDefaultPolicy
|
||||
{
|
||||
template <typename Problem>
|
||||
CK_TILE_DEVICE static constexpr auto MakeABXBlockTileDistribution()
|
||||
{
|
||||
using S = typename Problem::BlockShape;
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<
|
||||
sequence<>,
|
||||
tuple<sequence<S::Repeat_M, S::WarpPerBlock_M, S::ThreadPerWarp_M, S::Vector_M>,
|
||||
sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
|
||||
tuple<sequence<1, 2>, sequence<1, 2>>,
|
||||
tuple<sequence<1, 1>, sequence<2, 2>>,
|
||||
sequence<1, 1, 2, 2>,
|
||||
sequence<0, 3, 0, 3>>{});
|
||||
}
|
||||
template <typename Problem>
|
||||
CK_TILE_DEVICE static constexpr auto MakeGammaBlockTileDistribution()
|
||||
{
|
||||
using S = typename Problem::BlockShape;
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<
|
||||
sequence<S::WarpPerBlock_M, S::ThreadPerWarp_M>,
|
||||
tuple<sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
|
||||
tuple<sequence<0, 1>, sequence<0, 1>>,
|
||||
tuple<sequence<0, 1>, sequence<1, 2>>,
|
||||
sequence<1, 1>,
|
||||
sequence<0, 3>>{});
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2d()
|
||||
{
|
||||
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
|
||||
typename Problem::ComputeDataType,
|
||||
typename Problem::BlockShape>;
|
||||
return BlockReduce2d<P_>{};
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dSync()
|
||||
{
|
||||
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
|
||||
typename Problem::ComputeDataType,
|
||||
typename Problem::BlockShape>;
|
||||
return BlockReduce2dSync<P_>{};
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dCrossWarpSync()
|
||||
{
|
||||
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
|
||||
typename Problem::ComputeDataType,
|
||||
typename Problem::BlockShape>;
|
||||
return BlockReduce2dCrossWarpSync<P_>{};
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
|
||||
{
|
||||
if constexpr(Problem::kNeedCrossWarpSync)
|
||||
{
|
||||
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
|
||||
typename Problem::ComputeDataType,
|
||||
typename Problem::BlockShape>;
|
||||
|
||||
using block_reduce2d = BlockReduce2d<P_>;
|
||||
using x_block_tile =
|
||||
decltype(make_static_distributed_tensor<typename Problem::ComputeDataType>(
|
||||
MakeABXBlockTileDistribution<Problem>()));
|
||||
using y_block_tile = decltype(block_reduce2d::template MakeYBlockTile<x_block_tile>());
|
||||
|
||||
return GetBlockReduce2dCrossWarpSync<Problem>().template GetSmemSize<y_block_tile>();
|
||||
}
|
||||
else
|
||||
{
|
||||
return 1; // zero size arrays are an extension
|
||||
}
|
||||
}
|
||||
};
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,142 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
|
||||
#include <string>
|
||||
#include <type_traits>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename Problem_, typename Policy_ = AddRmsnorm2dRdquantFwdPipelineDefaultPolicy>
|
||||
struct AddRmsnorm2dRdquantFwdPipelineOnePass
|
||||
{
|
||||
using Problem = ck_tile::remove_cvref_t<Problem_>;
|
||||
using Policy = ck_tile::remove_cvref_t<Policy_>;
|
||||
|
||||
using ADataType = ck_tile::remove_cvref_t<typename Problem::ADataType>;
|
||||
using BDataType = ck_tile::remove_cvref_t<typename Problem::BDataType>;
|
||||
using GammaDataType = ck_tile::remove_cvref_t<typename Problem::GammaDataType>;
|
||||
using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
|
||||
using XDataType = ck_tile::remove_cvref_t<typename Problem::XDataType>;
|
||||
using YScaleDataType = ck_tile::remove_cvref_t<typename Problem::YScaleDataType>;
|
||||
using QYDataType = ck_tile::remove_cvref_t<typename Problem::QYDataType>;
|
||||
|
||||
static constexpr bool kHasGamma = !std::is_same_v<GammaDataType, ck_tile::null_type>;
|
||||
static constexpr bool kSaveX = Problem::kSaveX;
|
||||
|
||||
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
|
||||
static constexpr bool kPadM = false; // TODO - BlockAddRmsnorm2dRdquantFwdProblem::kPadM
|
||||
static constexpr bool kPadN = Problem::kPadN;
|
||||
|
||||
static constexpr const char* name = []() {
|
||||
if constexpr(kNeedCrossWarpSync)
|
||||
return "bpr_op"; // block per row
|
||||
else
|
||||
return "wpr_op"; // warp per row
|
||||
}();
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
|
||||
{
|
||||
return Policy::template GetSmemSize<Problem>();
|
||||
}
|
||||
|
||||
template <typename AWindow,
|
||||
typename BWindow,
|
||||
typename GammaWindow,
|
||||
typename XWindow,
|
||||
typename YScaleWindow,
|
||||
typename QYWindow>
|
||||
CK_TILE_DEVICE auto operator()(const AWindow& a_window_,
|
||||
const BWindow& b_window_,
|
||||
const GammaWindow& gamma_window_,
|
||||
XWindow& x_window,
|
||||
YScaleWindow& yscale_window,
|
||||
QYWindow& qy_window,
|
||||
ComputeDataType epsilon,
|
||||
ck_tile::index_t row_size,
|
||||
void* smem) const
|
||||
{
|
||||
const auto a_window =
|
||||
make_tile_window(a_window_, Policy::template MakeABXBlockTileDistribution<Problem>());
|
||||
const auto b_window =
|
||||
make_tile_window(b_window_, Policy::template MakeABXBlockTileDistribution<Problem>());
|
||||
const auto gamma_window = make_tile_window(
|
||||
gamma_window_, Policy::template MakeGammaBlockTileDistribution<Problem>());
|
||||
|
||||
auto reduce_square_sum_func = ReduceOp::SquareAdd{};
|
||||
auto reduce_sum_func = ReduceOp::Add{};
|
||||
auto reduce_absmax_func = ReduceOp::AbsMax{};
|
||||
auto reduce_max_func = ReduceOp::Max{};
|
||||
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
|
||||
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
|
||||
auto block_reduce2d_cross_warp_sync =
|
||||
Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
|
||||
|
||||
const auto a = load_tile(a_window);
|
||||
const auto b = load_tile(b_window);
|
||||
const auto gamma = load_tile(gamma_window);
|
||||
|
||||
auto x = tile_elementwise_in(
|
||||
[&](const auto& a_, const auto& b_) {
|
||||
return type_convert<ComputeDataType>(a_) + type_convert<ComputeDataType>(b_);
|
||||
},
|
||||
a,
|
||||
b);
|
||||
|
||||
if constexpr(kSaveX)
|
||||
store_tile(x_window, cast_tile<XDataType>(x));
|
||||
|
||||
// compute mean square, each-thread->cross-lane->cross-warp
|
||||
auto square_sum = block_reduce2d(
|
||||
x, reduce_square_sum_func.GetIdentityValue<ComputeDataType>(), reduce_square_sum_func);
|
||||
block_reduce2d_sync(square_sum, reduce_sum_func);
|
||||
block_reduce2d_cross_warp_sync(square_sum, smem, reduce_sum_func);
|
||||
|
||||
auto inv_rms = tile_elementwise_in(
|
||||
[&](const auto& v_) {
|
||||
return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ / row_size + epsilon));
|
||||
},
|
||||
square_sum);
|
||||
|
||||
// rmsnorm computation
|
||||
auto y = make_static_distributed_tensor<ComputeDataType>(x.get_tile_distribution());
|
||||
sweep_tile(y, [&, inv_rms_ = inv_rms](auto idx) {
|
||||
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
|
||||
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
|
||||
|
||||
const auto gamma_ = type_convert<ComputeDataType>(gamma[j_idx]);
|
||||
|
||||
const auto x_ = type_convert<ComputeDataType>(x[idx]);
|
||||
auto y_ = x_ * inv_rms_[i_idx] * gamma_;
|
||||
|
||||
y(idx) = type_convert<ComputeDataType>(y_);
|
||||
});
|
||||
|
||||
// compute absmax, each-thread->cross-lane->cross-warp
|
||||
auto absmax = block_reduce2d(
|
||||
y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
|
||||
block_reduce2d_sync(absmax, reduce_max_func);
|
||||
block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
|
||||
|
||||
// ex: yscale = absmax / 127 if int8
|
||||
auto yscale = tile_elementwise_in(
|
||||
[&](const auto& v_) {
|
||||
return v_ / type_convert<ComputeDataType>(numeric<QYDataType>::max());
|
||||
},
|
||||
absmax);
|
||||
store_tile(yscale_window, cast_tile<YScaleDataType>(yscale));
|
||||
|
||||
// quantize y to qy
|
||||
auto qy = make_static_distributed_tensor<QYDataType>(y.get_tile_distribution());
|
||||
sweep_tile(qy, [&, yscale_ = yscale](auto idx) {
|
||||
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
|
||||
auto qy_ = y[idx] / yscale_[i_idx];
|
||||
qy(idx) = saturates<QYDataType>{}(qy_);
|
||||
});
|
||||
store_tile(qy_window, qy);
|
||||
}
|
||||
};
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,41 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core/utility/type_traits.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// X = A + B, Y = Rmsnorm2d(X), QY = RowwiseDynamicQuant(Y) = SaturateCast(Y / YScale)
|
||||
template <typename ADataType_,
|
||||
typename BDataType_,
|
||||
typename GammaDataType_,
|
||||
typename ComputeDataType_,
|
||||
typename XDataType_,
|
||||
typename YScaleDataType_,
|
||||
typename QYDataType_,
|
||||
typename BlockShape_,
|
||||
bool kPadN_,
|
||||
bool kSaveX_,
|
||||
bool kThreePass_>
|
||||
struct AddRmsnorm2dRdquantFwdPipelineProblem
|
||||
{
|
||||
using ADataType = remove_cvref_t<ADataType_>;
|
||||
using BDataType = remove_cvref_t<BDataType_>;
|
||||
using GammaDataType = remove_cvref_t<GammaDataType_>;
|
||||
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
|
||||
using XDataType = remove_cvref_t<XDataType_>;
|
||||
using YScaleDataType = remove_cvref_t<YScaleDataType_>;
|
||||
using QYDataType = remove_cvref_t<QYDataType_>;
|
||||
using BlockShape = remove_cvref_t<BlockShape_>;
|
||||
|
||||
static constexpr bool kNeedCrossLaneSync = BlockShape::ThreadPerWarp_N > 1;
|
||||
static constexpr bool kNeedCrossWarpSync = BlockShape::WarpPerBlock_N > 1;
|
||||
|
||||
static constexpr bool kPadN = kPadN_;
|
||||
static constexpr bool kSaveX = kSaveX_;
|
||||
static constexpr bool kThreePass = kThreePass_;
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,266 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
|
||||
#include <string>
|
||||
#include <type_traits>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename Problem_, typename Policy_ = AddRmsnorm2dRdquantFwdPipelineDefaultPolicy>
|
||||
struct AddRmsnorm2dRdquantFwdPipelineThreePass
|
||||
{
|
||||
using Problem = ck_tile::remove_cvref_t<Problem_>;
|
||||
using Policy = ck_tile::remove_cvref_t<Policy_>;
|
||||
|
||||
using ADataType = ck_tile::remove_cvref_t<typename Problem::ADataType>;
|
||||
using BDataType = ck_tile::remove_cvref_t<typename Problem::BDataType>;
|
||||
using GammaDataType = ck_tile::remove_cvref_t<typename Problem::GammaDataType>;
|
||||
using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
|
||||
using XDataType = ck_tile::remove_cvref_t<typename Problem::XDataType>;
|
||||
using YScaleDataType = ck_tile::remove_cvref_t<typename Problem::YScaleDataType>;
|
||||
using QYDataType = ck_tile::remove_cvref_t<typename Problem::QYDataType>;
|
||||
|
||||
static constexpr bool kHasGamma = !std::is_same_v<GammaDataType, ck_tile::null_type>;
|
||||
static constexpr bool kSaveX = Problem::kSaveX;
|
||||
|
||||
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
|
||||
static constexpr bool kPadM = false; // TODO - BlockAddRmsnorm2dRdquantFwdProblem::kPadM
|
||||
static constexpr bool kPadN = Problem::kPadN;
|
||||
|
||||
static constexpr const char* name = []() {
|
||||
if constexpr(kNeedCrossWarpSync)
|
||||
return "bpr_tp"; // block per row
|
||||
else
|
||||
return "wpr_tp"; // warp per row
|
||||
}();
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
|
||||
{
|
||||
return Policy::template GetSmemSize<Problem>();
|
||||
}
|
||||
|
||||
template <typename AWindow,
|
||||
typename BWindow,
|
||||
typename GammaWindow,
|
||||
typename XWindow,
|
||||
typename YScaleWindow,
|
||||
typename QYWindow>
|
||||
CK_TILE_DEVICE auto operator()(const AWindow& a_window_,
|
||||
const BWindow& b_window_,
|
||||
const GammaWindow& gamma_window_,
|
||||
XWindow& x_window_,
|
||||
YScaleWindow& yscale_window,
|
||||
QYWindow& qy_window,
|
||||
ComputeDataType epsilon,
|
||||
ck_tile::index_t row_size,
|
||||
void* smem) const
|
||||
{
|
||||
auto a_window =
|
||||
make_tile_window(a_window_, Policy::template MakeABXBlockTileDistribution<Problem>());
|
||||
auto b_window =
|
||||
make_tile_window(b_window_, Policy::template MakeABXBlockTileDistribution<Problem>());
|
||||
auto x_window = [&]() {
|
||||
if constexpr(kSaveX)
|
||||
return make_tile_window(x_window_,
|
||||
Policy::template MakeABXBlockTileDistribution<Problem>());
|
||||
else
|
||||
return x_window_;
|
||||
}();
|
||||
auto gamma_window = make_tile_window(
|
||||
gamma_window_, Policy::template MakeGammaBlockTileDistribution<Problem>());
|
||||
|
||||
auto reduce_square_sum_func = ReduceOp::SquareAdd{};
|
||||
auto reduce_sum_func = ReduceOp::Add{};
|
||||
auto reduce_absmax_func = ReduceOp::AbsMax{};
|
||||
auto reduce_max_func = ReduceOp::Max{};
|
||||
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
|
||||
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
|
||||
auto block_reduce2d_cross_warp_sync =
|
||||
Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
|
||||
|
||||
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
|
||||
index_t num_n_tile_iteration =
|
||||
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
|
||||
|
||||
using XTensorType = decltype(cast_tile<ComputeDataType>(load_tile(a_window)));
|
||||
auto square_sum = block_reduce2d.template MakeYBlockTile<XTensorType>();
|
||||
set_tile(square_sum, reduce_square_sum_func.GetIdentityValue<ComputeDataType>());
|
||||
|
||||
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
|
||||
{
|
||||
const auto a = load_tile(a_window);
|
||||
const auto b = load_tile(b_window);
|
||||
|
||||
auto x = tile_elementwise_in(
|
||||
[&](const auto& a_, const auto& b_) {
|
||||
return type_convert<ComputeDataType>(a_) + type_convert<ComputeDataType>(b_);
|
||||
},
|
||||
a,
|
||||
b);
|
||||
|
||||
if constexpr(kSaveX)
|
||||
store_tile(x_window, cast_tile<XDataType>(x));
|
||||
|
||||
block_reduce2d(x, square_sum, reduce_square_sum_func);
|
||||
move_tile_window(x_window, {0, Block_N});
|
||||
move_tile_window(a_window, {0, Block_N});
|
||||
move_tile_window(b_window, {0, Block_N});
|
||||
}
|
||||
|
||||
block_reduce2d_sync(square_sum, reduce_sum_func);
|
||||
block_reduce2d_cross_warp_sync(square_sum, smem, reduce_sum_func);
|
||||
|
||||
auto inv_rms = tile_elementwise_in(
|
||||
[&](const auto& v_) {
|
||||
return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ / row_size + epsilon));
|
||||
},
|
||||
square_sum);
|
||||
|
||||
// reverse read x to reuse cache
|
||||
ck_tile::index_t stride_to_right_most_window =
|
||||
row_size % Block_N == 0 ? row_size - Block_N : row_size - row_size % Block_N;
|
||||
|
||||
if constexpr(kSaveX)
|
||||
move_tile_window(x_window, {0, -Block_N});
|
||||
else
|
||||
{
|
||||
move_tile_window(a_window, {0, -Block_N});
|
||||
move_tile_window(b_window, {0, -Block_N});
|
||||
}
|
||||
move_tile_window(gamma_window, {stride_to_right_most_window});
|
||||
|
||||
using YTensorType = XTensorType;
|
||||
auto absmax = block_reduce2d.template MakeYBlockTile<YTensorType>();
|
||||
set_tile(absmax, reduce_absmax_func.GetIdentityValue<ComputeDataType>());
|
||||
|
||||
// rmsnorm computation + absmax(threadwise reduce)
|
||||
if constexpr(kSaveX)
|
||||
__syncthreads();
|
||||
|
||||
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
|
||||
{
|
||||
auto x = [&]() {
|
||||
if constexpr(kSaveX)
|
||||
{
|
||||
return load_tile(x_window);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto a = load_tile(a_window);
|
||||
const auto b = load_tile(b_window);
|
||||
return tile_elementwise_in(
|
||||
[&](const auto& a_, const auto& b_) {
|
||||
return type_convert<ComputeDataType>(a_) +
|
||||
type_convert<ComputeDataType>(b_);
|
||||
},
|
||||
a,
|
||||
b);
|
||||
}
|
||||
}();
|
||||
|
||||
auto gamma = load_tile(gamma_window);
|
||||
auto y = make_static_distributed_tensor<ComputeDataType>(x.get_tile_distribution());
|
||||
|
||||
sweep_tile(y, [&](auto idx) {
|
||||
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
|
||||
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
|
||||
|
||||
const auto gamma_ = type_convert<ComputeDataType>(gamma[j_idx]);
|
||||
|
||||
const auto x_ = type_convert<ComputeDataType>(x[idx]);
|
||||
auto y_ = x_ * inv_rms[i_idx] * gamma_;
|
||||
|
||||
y(idx) = type_convert<ComputeDataType>(y_);
|
||||
});
|
||||
|
||||
block_reduce2d(y, absmax, reduce_absmax_func);
|
||||
|
||||
if constexpr(kSaveX)
|
||||
move_tile_window(x_window, {0, -Block_N});
|
||||
else
|
||||
{
|
||||
move_tile_window(a_window, {0, -Block_N});
|
||||
move_tile_window(b_window, {0, -Block_N});
|
||||
}
|
||||
move_tile_window(gamma_window, {-Block_N});
|
||||
}
|
||||
|
||||
// compute absmax, cross-lane->cross-warp
|
||||
block_reduce2d_sync(absmax, reduce_max_func);
|
||||
block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
|
||||
|
||||
// ex: yscale = absmax / 127 if int8
|
||||
auto yscale = tile_elementwise_in(
|
||||
[&](const auto& v_) {
|
||||
return v_ / type_convert<ComputeDataType>(numeric<QYDataType>::max());
|
||||
},
|
||||
absmax);
|
||||
store_tile(yscale_window, cast_tile<YScaleDataType>(yscale));
|
||||
|
||||
// quantize y to qy
|
||||
// recompute rmsnorm, try to save y in the future
|
||||
if constexpr(kSaveX)
|
||||
move_tile_window(x_window, {0, Block_N});
|
||||
else
|
||||
{
|
||||
move_tile_window(a_window, {0, Block_N});
|
||||
move_tile_window(b_window, {0, Block_N});
|
||||
}
|
||||
move_tile_window(gamma_window, {Block_N});
|
||||
|
||||
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
|
||||
{
|
||||
auto x = [&]() {
|
||||
if constexpr(kSaveX)
|
||||
{
|
||||
return load_tile(x_window);
|
||||
}
|
||||
else
|
||||
{
|
||||
const auto a = load_tile(a_window);
|
||||
const auto b = load_tile(b_window);
|
||||
return tile_elementwise_in(
|
||||
[&](const auto& a_, const auto& b_) {
|
||||
return type_convert<ComputeDataType>(a_) +
|
||||
type_convert<ComputeDataType>(b_);
|
||||
},
|
||||
a,
|
||||
b);
|
||||
}
|
||||
}();
|
||||
|
||||
auto gamma = load_tile(gamma_window);
|
||||
auto y = make_static_distributed_tensor<ComputeDataType>(x.get_tile_distribution());
|
||||
auto qy = make_static_distributed_tensor<QYDataType>(y.get_tile_distribution());
|
||||
|
||||
sweep_tile(y, [&](auto idx) {
|
||||
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
|
||||
constexpr auto j_idx = make_tuple(idx[number<1>{}]);
|
||||
|
||||
const auto gamma_ = type_convert<ComputeDataType>(gamma[j_idx]);
|
||||
|
||||
const auto x_ = type_convert<ComputeDataType>(x[idx]);
|
||||
auto y_ = x_ * inv_rms[i_idx] * gamma_;
|
||||
auto qy_ = y_ / yscale[i_idx];
|
||||
qy(idx) = saturates<QYDataType>{}(qy_);
|
||||
});
|
||||
|
||||
store_tile(qy_window, qy);
|
||||
|
||||
if constexpr(kSaveX)
|
||||
move_tile_window(x_window, {0, Block_N});
|
||||
else
|
||||
{
|
||||
move_tile_window(a_window, {0, Block_N});
|
||||
move_tile_window(b_window, {0, Block_N});
|
||||
}
|
||||
move_tile_window(gamma_window, {Block_N});
|
||||
move_tile_window(qy_window, {0, Block_N});
|
||||
}
|
||||
}
|
||||
};
|
||||
} // namespace ck_tile
|
||||
Reference in New Issue
Block a user