ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-03 13:11:25 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK_TILE] Add Various Fusion Functions to RMSNorm (#1802)

Browse Source 
* Add shortcut to RMSNorm

* Modify test for adding shortcut for RMSNorm

* Add fused parameter into tests

* 1. Add YDataType. 2. rmsnorm2d_fwd_traits_ from rmsnorm2d_fwd.hpp to rmsnorm2d_fwd_api.cpp and rmsnorm2d_fwd_instance_common.hpp

* 1. Supports various stride and percisions.

* Add support of Epilogue

* Add fuse and epilogue support to rmsnorm ref

* Modify rmsnorm example

* Refactor tests/examples

* Bug fix for newly added tests/examples

* Bug fix for new tests 2

* Modify smoke test scripts

remove dbg code

* Supports non-smooth dyanmic quant

* Update Rmsnorm2dFwd::GetName()

* rename xscale and prec_sx to smoothscale and prec_sm

Bug fix after rename

Remove files

* change example_rmsnorm2d_fwd.cpp

* update performance calculator

* Fix issue in two-pass when fuse add is enabled

* Remove comment of beta

---------

Co-authored-by: rocking <ChunYu.Lai@amd.com>
This commit is contained in:
ruanjm
2025-01-15 10:23:48 +08:00
committed by GitHub
parent c0b90f130f
commit 04dd314883
58 changed files with 1823 additions and 1045 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
include/ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp
View File

@@ -45,7 +45,7 @@ struct Rmsnorm2dFwdPipelineDefaultPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2d()
{
using P_ = BlockReduce2dProblem<typename Problem::XDataType,
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
return BlockReduce2d<P_>{};
@@ -54,7 +54,7 @@ struct Rmsnorm2dFwdPipelineDefaultPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dSync()
{
using P_ = BlockReduce2dProblem<typename Problem::XDataType,
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
return BlockReduce2dSync<P_>{};
@@ -63,7 +63,7 @@ struct Rmsnorm2dFwdPipelineDefaultPolicy
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dCrossWarpSync()
{
using P_ = BlockReduce2dProblem<typename Problem::XDataType,
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
return BlockReduce2dCrossWarpSync<P_>{};
@@ -74,13 +74,13 @@ struct Rmsnorm2dFwdPipelineDefaultPolicy
{
if constexpr(Problem::kNeedCrossWarpSync)
{
using P_ = BlockReduce2dProblem<typename Problem::XDataType,
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::XDataType>(
decltype(make_static_distributed_tensor<typename Problem::ComputeDataType>(
MakeXBlockTileDistribution<Problem>()));
using y_block_tile = decltype(block_reduce2d::template MakeYBlockTile<x_block_tile>());

81
include/ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_one_pass.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -22,12 +22,17 @@ struct Rmsnorm2dFwdPipelineOnePass
using YDataType = ck_tile::remove_cvref_t<typename Problem::YDataType>;
using InvRmsDataType = ck_tile::remove_cvref_t<typename Problem::InvRmsDataType>;
using XResidualDataType = XDataType;
using YResidualDataType = XDataType;
static constexpr bool kHasGamma = !std::is_same_v<GammaDataType, ck_tile::null_type>;
static constexpr bool kSaveInvRms = Problem::kSaveInvRms;
static constexpr bool kSaveInvRms = Problem::Traits::kSaveInvRms;
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockRmsnorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
@@ -41,19 +46,36 @@ struct Rmsnorm2dFwdPipelineOnePass
return Policy::template GetSmemSize<Problem>();
}
template <typename XWindow, typename GammaWindow, typename YWindow, typename InvRmsWindow>
template <typename XWindow,
typename XResidualWindow,
typename GammaWindow,
typename YWindow,
typename YResidualWindow,
typename InvRmsWindow,
typename SmoothScaleWindow,
typename YScaleWindow,
typename Epilogue>
CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
const XResidualWindow& x_residual_window_,
const GammaWindow& gamma_window_,
YWindow& y_window,
YWindow& y_window_,
const YResidualWindow& y_residual_window_,
InvRmsWindow& inv_rms_window,
const SmoothScaleWindow& sm_scale_window_,
YScaleWindow& y_scale_window_,
ComputeDataType epsilon,
ck_tile::index_t row_size,
void* smem) const
void* smem,
Epilogue) const
{
const auto x_window =
make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
const auto gamma_window = make_tile_window(
gamma_window_, Policy::template MakeGammaBlockTileDistribution<Problem>());
const auto x_residual_window = make_tile_window(
x_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto y_residual_window = make_tile_window(
y_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto reduce_square_sum_func = ReduceOp::SquareAdd{};
auto reduce_sum_func = ReduceOp::Add{};
@@ -62,13 +84,31 @@ struct Rmsnorm2dFwdPipelineOnePass
auto block_reduce2d_cross_warp_sync =
Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
const auto x = load_tile(x_window);
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
// load gamma (TODO: support no gamma?)
const auto gamma = load_tile(gamma_window);
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Rmsnorm2dFusedAddEnum::PRE_ADD ||
kFusedAdd == Rmsnorm2dFusedAddEnum::PRE_ADD_STORE)
{
sweep_tile(x_resi, [&](auto idx) {
// compute x = x_resi + x
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
if constexpr(kFusedAdd == Rmsnorm2dFusedAddEnum::PRE_ADD_STORE)
{
store_tile(y_residual_window, cast_tile<YResidualDataType>(acc));
}
}
// 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);
auto square_sum = block_reduce2d(acc,
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);
@@ -83,19 +123,30 @@ struct Rmsnorm2dFwdPipelineOnePass
store_tile(inv_rms_window, cast_tile<InvRmsDataType>(inv_rms));
// rmsnorm computation
auto y = make_static_distributed_tensor<YDataType>(x.get_tile_distribution());
sweep_tile(y, [&, inv_rms_ = inv_rms](auto idx) {
auto rmsn = make_static_distributed_tensor<ComputeDataType>(x.get_tile_distribution());
sweep_tile(rmsn, [&, 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_;
auto rmsn_ = acc[idx] * inv_rms_[i_idx] * gamma_;
y(idx) = type_convert<YDataType>(y_);
rmsn(idx) = rmsn_;
});
store_tile(y_window, y);
if constexpr(kFusedQuant == Rmsnorm2dFusedQuantEnum::SMOOTH_DYNAMIC_QUANT)
{
Epilogue{}(y_window_, sm_scale_window_, y_scale_window_, rmsn, smem);
}
else if constexpr(kFusedQuant == Rmsnorm2dFusedQuantEnum::DYNAMIC_QUANT)
{
Epilogue{}(y_window_, y_scale_window_, rmsn, smem);
}
else
{
Epilogue{}(y_window_, rmsn);
}
}
};
} // namespace ck_tile

26
include/ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_problem.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -12,25 +12,25 @@ template <typename XDataType_,
typename ComputeDataType_,
typename YDataType_,
typename InvRmsDataType_,
typename SmoothScaleDataType_,
typename YScaleDataType_,
typename BlockShape_,
bool kPadN_,
bool kSaveInvRms_,
bool kTwoPass_>
typename Traits_>
struct Rmsnorm2dFwdPipelineProblem
{
using XDataType = remove_cvref_t<XDataType_>;
using GammaDataType = remove_cvref_t<GammaDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
using YDataType = remove_cvref_t<YDataType_>;
using InvRmsDataType = remove_cvref_t<InvRmsDataType_>;
using BlockShape = remove_cvref_t<BlockShape_>;
using XDataType = remove_cvref_t<XDataType_>;
using GammaDataType = remove_cvref_t<GammaDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
using YDataType = remove_cvref_t<YDataType_>;
using InvRmsDataType = remove_cvref_t<InvRmsDataType_>;
using SmoothScaleDataType = remove_cvref_t<SmoothScaleDataType_>;
using YScaleDataType = remove_cvref_t<YScaleDataType_>;
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 kSaveInvRms = kSaveInvRms_;
static constexpr bool kTwoPass = kTwoPass_;
using Traits = remove_cvref_t<Traits_>;
};
} // namespace ck_tile

91
include/ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_two_pass.hpp
View File

@@ -1,5 +1,5 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -22,12 +22,17 @@ struct Rmsnorm2dFwdPipelineTwoPass
using YDataType = ck_tile::remove_cvref_t<typename Problem::YDataType>;
using InvRmsDataType = ck_tile::remove_cvref_t<typename Problem::InvRmsDataType>;
using XResidualDataType = XDataType;
using YResidualDataType = XDataType;
static constexpr bool kHasGamma = !std::is_same_v<GammaDataType, ck_tile::null_type>;
static constexpr bool kSaveInvRms = Problem::kSaveInvRms;
static constexpr bool kSaveInvRms = Problem::Traits::kSaveInvRms;
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockRmsnorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
@@ -41,19 +46,36 @@ struct Rmsnorm2dFwdPipelineTwoPass
return Policy::template GetSmemSize<Problem>();
}
template <typename XWindow, typename GammaWindow, typename YWindow, typename InvRmsWindow>
template <typename XWindow,
typename XResidualWindow,
typename GammaWindow,
typename YWindow,
typename YResidualWindow,
typename InvRmsWindow,
typename SmoothScaleWindow,
typename YScaleWindow,
typename Epilogue>
CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
const XResidualWindow& x_residual_window_,
const GammaWindow& gamma_window_,
YWindow& y_window,
const YResidualWindow& y_residual_window_,
InvRmsWindow& inv_rms_window,
const SmoothScaleWindow& /*sm_scale_window_*/,
YScaleWindow& /*y_scale_window*/,
ComputeDataType epsilon,
ck_tile::index_t row_size,
void* smem) const
void* smem,
Epilogue) const
{
auto x_window =
make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto gamma_window = make_tile_window(
gamma_window_, Policy::template MakeGammaBlockTileDistribution<Problem>());
auto x_residual_window = make_tile_window(
x_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto y_residual_window = make_tile_window(
y_residual_window_, Policy::template MakeXBlockTileDistribution<Problem>());
// Problem::BlockShape
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
@@ -67,15 +89,34 @@ struct Rmsnorm2dFwdPipelineTwoPass
auto block_reduce2d_cross_warp_sync =
Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
using XTensorType = decltype(load_tile(x_window));
auto square_sum = block_reduce2d.template MakeYBlockTile<XTensorType>();
using ComputeTensorType = decltype(cast_tile<ComputeDataType>(load_tile(x_window)));
auto square_sum = block_reduce2d.template MakeYBlockTile<ComputeTensorType>();
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 x = load_tile(x_window);
block_reduce2d(x, square_sum, reduce_square_sum_func);
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
move_tile_window(x_window, {0, Block_N});
move_tile_window(x_residual_window, {0, Block_N});
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Rmsnorm2dFusedAddEnum::PRE_ADD ||
kFusedAdd == Rmsnorm2dFusedAddEnum::PRE_ADD_STORE)
{
sweep_tile(x_resi, [&](auto idx) {
// compute x = x_resi + x
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
if constexpr(kFusedAdd == Rmsnorm2dFusedAddEnum::PRE_ADD_STORE)
{
store_tile(y_residual_window, cast_tile<YResidualDataType>(acc));
move_tile_window(y_residual_window, {0, Block_N});
}
}
block_reduce2d(acc, square_sum, reduce_square_sum_func);
}
block_reduce2d_sync(square_sum, reduce_sum_func);
@@ -96,33 +137,47 @@ struct Rmsnorm2dFwdPipelineTwoPass
row_size % Block_N == 0 ? row_size - Block_N : row_size - row_size % Block_N;
move_tile_window(x_window, {0, -Block_N});
move_tile_window(x_residual_window, {0, -Block_N});
move_tile_window(gamma_window, {stride_to_right_most_window});
move_tile_window(y_window, {0, stride_to_right_most_window});
// rmsnorm computation
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
const auto x = load_tile(x_window);
// load gamma/beta (TODO: support no gamma/beta?)
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Rmsnorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Rmsnorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
// compute x = x_resi + x
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
}
// load gamma (TODO: support no gamma?)
const auto gamma = load_tile(gamma_window);
auto y = make_static_distributed_tensor<YDataType>(x.get_tile_distribution());
sweep_tile(y, [&, inv_rms_ = inv_rms](auto idx) {
// rmsnorm computation
auto rmsn = make_static_distributed_tensor<ComputeDataType>(x.get_tile_distribution());
sweep_tile(rmsn, [&, 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_;
auto rmsn_ = acc(idx) * inv_rms_[i_idx] * gamma_;
y(idx) = type_convert<YDataType>(y_);
rmsn(idx) = rmsn_;
});
store_tile(y_window, y);
static_assert(kFusedQuant == Rmsnorm2dFusedQuantEnum::NO_SWEEP);
Epilogue{}(y_window, rmsn);
move_tile_window(x_window, {0, -Block_N});
move_tile_window(x_residual_window, {0, -Block_N});
move_tile_window(gamma_window, {-Block_N});
move_tile_window(y_window, {0, -Block_N});
}

54
include/ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_traits.hpp Normal file
View File

@@ -0,0 +1,54 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core/utility/type_traits.hpp"
namespace ck_tile {
enum class Rmsnorm2dFusedAddEnum
{
NO_ADD = 0,
// fused add before RMSNorm and store result to global
PRE_ADD_STORE = 1,
// fused add before RMSNorm, but not store result
PRE_ADD = 2,
};
// clang-format off
template<Rmsnorm2dFusedAddEnum> struct Rmsnorm2dFusedAddEnumName;
template<> struct Rmsnorm2dFusedAddEnumName<Rmsnorm2dFusedAddEnum::NO_ADD> { static constexpr const char * name = "no"; };
template<> struct Rmsnorm2dFusedAddEnumName<Rmsnorm2dFusedAddEnum::PRE_ADD_STORE> { static constexpr const char * name = "pras"; };
template<> struct Rmsnorm2dFusedAddEnumName<Rmsnorm2dFusedAddEnum::PRE_ADD> { static constexpr const char * name = "pra"; };
// clang-format on
enum class Rmsnorm2dFusedQuantEnum
{
NO_SWEEP = 0,
SMOOTH_DYNAMIC_QUANT = 1, // smooth oulier + rowwise quant, need input x-scale and store y_scale
DYNAMIC_QUANT = 2, // rowwise quant, store out a y-scale
};
// clang-format off
template<Rmsnorm2dFusedQuantEnum> struct Rmsnorm2dFusedQuantEnumName;
template<> struct Rmsnorm2dFusedQuantEnumName<Rmsnorm2dFusedQuantEnum::NO_SWEEP> { static constexpr const char * name = "no"; };
template<> struct Rmsnorm2dFusedQuantEnumName<Rmsnorm2dFusedQuantEnum::DYNAMIC_QUANT> { static constexpr const char * name = "dqt"; };
template<> struct Rmsnorm2dFusedQuantEnumName<Rmsnorm2dFusedQuantEnum::SMOOTH_DYNAMIC_QUANT> { static constexpr const char * name = "smdqt"; };
// clang-format on
template <bool kPadN_,
bool kSaveInvRms_,
bool kTwoPass_,
Rmsnorm2dFusedAddEnum kFusedAdd_,
Rmsnorm2dFusedQuantEnum kFusedQuant_>
struct Rmsnorm2dFwdTraits
{
static constexpr bool kPadN = kPadN_;
static constexpr bool kSaveInvRms = kSaveInvRms_;
static constexpr bool kTwoPass = kTwoPass_;
static constexpr Rmsnorm2dFusedAddEnum kFusedAdd = kFusedAdd_;
static constexpr Rmsnorm2dFusedQuantEnum kFusedQuant = kFusedQuant_;
};
} // namespace ck_tile