ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +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

38
include/ck_tile/ops/layernorm2d/kernel/layernorm2d_fwd_kernel.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
@@ -14,7 +14,7 @@ struct Layernorm2dFwdHostArgs
{
const void* p_x; // [m ,n], input, fp16/bf16
const void* p_x_residual; // [m ,n], shortcut input, prec same as input, nullptr if not used
const void* p_x_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used
const void* p_sm_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used
const void* p_x_bias; // [1, n], bias, prec same as input
const void* p_gamma; // [1, n], gamma, prec same as input
const void* p_beta; // [1, n], beta, prec same as input
@@ -43,16 +43,16 @@ struct Layernorm2dFwd
using Epilogue = remove_cvref_t<Epilogue_>;
using Problem = typename Pipeline::Problem;
using XDataType = remove_cvref_t<typename Problem::XDataType>;
using XBiasDataType = remove_cvref_t<typename Problem::XBiasDataType>;
using GammaDataType = remove_cvref_t<typename Problem::GammaDataType>;
using BetaDataType = remove_cvref_t<typename Problem::BetaDataType>;
using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
using YDataType = remove_cvref_t<typename Problem::YDataType>;
using MeanDataType = remove_cvref_t<typename Problem::MeanDataType>;
using InvStdDataType = remove_cvref_t<typename Problem::InvStdDataType>;
using XScaleDataType = remove_cvref_t<typename Problem::XScaleDataType>;
using YScaleDataType = remove_cvref_t<typename Problem::YScaleDataType>;
using XDataType = remove_cvref_t<typename Problem::XDataType>;
using XBiasDataType = remove_cvref_t<typename Problem::XBiasDataType>;
using GammaDataType = remove_cvref_t<typename Problem::GammaDataType>;
using BetaDataType = remove_cvref_t<typename Problem::BetaDataType>;
using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
using YDataType = remove_cvref_t<typename Problem::YDataType>;
using MeanDataType = remove_cvref_t<typename Problem::MeanDataType>;
using InvStdDataType = remove_cvref_t<typename Problem::InvStdDataType>;
using SmoothScaleDataType = remove_cvref_t<typename Problem::SmoothScaleDataType>;
using YScaleDataType = remove_cvref_t<typename Problem::YScaleDataType>;
// for simplicity, shortcut input/output type is same as X
using XResidualDataType = XDataType;
@@ -84,7 +84,7 @@ struct Layernorm2dFwd
{
const void* p_x; // [m ,n], input, fp16/bf16
const void* p_x_residual; // [m ,n], shortcut input, prec same as input, nullptr if not used
const void* p_x_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used
const void* p_sm_scale; // [1 ,n], smooth scale input, fp32, nullptr if not used
const void* p_x_bias; // [1, n], bias, prec same as input
const void* p_gamma; // [1, n], gamma, prec same as input
const void* p_beta; // [1, n], beta, prec same as input
@@ -111,7 +111,7 @@ struct Layernorm2dFwd
{
return Kargs{hargs.p_x,
hargs.p_x_residual,
hargs.p_x_scale,
hargs.p_sm_scale,
hargs.p_x_bias,
hargs.p_gamma,
hargs.p_beta,
@@ -171,7 +171,7 @@ struct Layernorm2dFwd
base_str += _SS_("_") + _SS_(t2s<YDataType>::name);
}
if (kFusedQuant == Layernorm2dFusedQuantEnum::SMOOTH_DYNAMIC_QUANT) {
base_str += _SS_("_sx") + _SS_(t2s<XScaleDataType>::name);
base_str += _SS_("_sx") + _SS_(t2s<SmoothScaleDataType>::name);
base_str += _SS_("_sy") + _SS_(t2s<YScaleDataType>::name);
}
if (kFusedQuant == Layernorm2dFusedQuantEnum::DYNAMIC_QUANT) {
@@ -356,18 +356,18 @@ struct Layernorm2dFwd
return make_null_tile_window(make_tuple(number<Block_M>{}));
}();
auto x_scale_window = [&]() {
auto sm_scale_window = [&]() {
if constexpr(kFusedQuant == Layernorm2dFusedQuantEnum::SMOOTH_DYNAMIC_QUANT)
{
const auto win_ = [&]() {
const auto tmp_0_ = make_naive_tensor_view_packed<address_space_enum::global>(
static_cast<const XScaleDataType*>(kargs.p_x_scale),
static_cast<const SmoothScaleDataType*>(kargs.p_sm_scale),
make_tuple(kargs.n),
number<Vector_N>{});
return pad_tensor_view(tmp_0_,
make_tuple(number<Block_N>{}),
sequence<false>{}); // x_scale no need pad
sequence<false>{}); // sm_scale no need pad
}();
return make_tile_window(win_, make_tuple(number<Block_N>{}), {0});
}
@@ -405,7 +405,7 @@ struct Layernorm2dFwd
y_residual_window,
mean_window,
inv_std_window,
x_scale_window,
sm_scale_window,
y_scale_window,
static_cast<const ComputeDataType>(kargs.epsilon),
kargs.n,