Skip to content

GitLab

Unverified Commit 04dd3148 authored by ruanjm's avatar ruanjm Committed by GitHub
Browse files

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



* 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: default avatarrocking <ChunYu.Lai@amd.com>
parent c0b90f13
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 825 additions and 416 deletions
example/ck_tile/02_layernorm2d/README.md
View file @ 04dd3148
......@@ -59,7 +59,7 @@ args:
-kname print kernel name or not (default:1)
-prec_i input precision (default:fp16)
-prec_o output precision, set auto will be the same as input (default:auto)
-prec_sx output quant scale type, set auto will be the same as input. used when fquant=1 (default:auto)
-prec_sm output quant scale type, set auto will be the same as input. used when fquant=1 (default:auto)
-prec_sy output quant scale type, set auto will be the same as input. used when fquant=1 or 2 (default:auto)
-fadd fused-add, 0:no fused add, 1:preadd+store, 2:preadd only (default:0)
-fquant fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant (default:0)
......@@ -69,7 +69,7 @@ args:
```
## limitations
Note that `fquant=2`, `fadd=2`, `prec_sx/prec_sy` other than `fp32` are not by default generated. Though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, `N>8192` case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet. If need suport `N>8192` and `fused+residual+store`, you can use this example together with `12_smoothquant`, to construct layernorm+residual, and smoothquant, 2 kernels for this purpose.
Note that `fquant=2`, `fadd=2`, `prec_sm/prec_sy` other than `fp32` are not by default generated. Though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, `N>8192` case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet. If need suport `N>8192` and `fused+residual+store`, you can use this example together with `12_smoothquant`, to construct layernorm+residual, and smoothquant, 2 kernels for this purpose.
```
# some case
......
example/ck_tile/02_layernorm2d/generate.py
View file @ 04dd3148
# 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.
# generate kernel instances to speed up compilation
import argparse
......@@ -52,7 +52,7 @@ class layernorm_fwd_codegen:
// this is used to pattern-match internl kernel implementation, not to instantiate kernel
template <typename XDataType_,
typename YDataType_,
typename XScaleDataType_,
typename SmoothScaleDataType_,
typename YScaleDataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
......@@ -71,7 +71,7 @@ struct layernorm2d_fwd_traits_
{
using XDataType = ck_tile::remove_cvref_t<XDataType_>;
using YDataType = ck_tile::remove_cvref_t<YDataType_>;
using XScaleDataType = ck_tile::remove_cvref_t<XScaleDataType_>;
using SmoothScaleDataType = ck_tile::remove_cvref_t<SmoothScaleDataType_>;
using YScaleDataType = ck_tile::remove_cvref_t<YScaleDataType_>;
static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
......@@ -135,7 +135,7 @@ struct layernorm2d_fwd_traits_
template <typename XDataType_,
typename YDataType_,
typename XScaleDataType_,
typename SmoothScaleDataType_,
typename YScaleDataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
......@@ -152,7 +152,7 @@ template <typename XDataType_,
int kFusedQuant_>
using traits_ = layernorm2d_fwd_traits_<XDataType_,
YDataType_,
XScaleDataType_,
SmoothScaleDataType_,
YScaleDataType_,
Repeat_M_,
Repeat_N_,
......@@ -170,7 +170,7 @@ using traits_ = layernorm2d_fwd_traits_<XDataType_,
"""
API_COMMON_HEADER = """
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "layernorm2d_fwd.hpp"
......@@ -189,9 +189,9 @@ float layernorm2d_fwd_(const S& s, A a)
{{
using XDataType = typename Traits_::XDataType;
using YDataType = typename Traits_::YDataType;
using XScaleDataType = typename Traits_::XScaleDataType;
using SmoothScaleDataType = typename Traits_::SmoothScaleDataType;
using YScaleDataType = typename Traits_::YScaleDataType;
using ComputeDataType = typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::ComputeDataType;
using ComputeDataType = typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::ComputeDataType;
using PipelineTraits = ck_tile::Layernorm2dFwdTraits<Traits_::kPadN,
Traits_::kSaveMeanInvStd,
......@@ -202,16 +202,16 @@ float layernorm2d_fwd_(const S& s, A a)
static_cast<ck_tile::Layernorm2dFusedAddEnum>(Traits_::kFusedAdd),
static_cast<ck_tile::Layernorm2dFusedQuantEnum>(Traits_::kFusedQuant)>;
using PipelineProblem = ck_tile::Layernorm2dFwdPipelineProblem<
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::XDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::XBiasDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::GammaDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::BetaDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::ComputeDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::YDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::MeanDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::InvStdDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::XScaleDataType,
typename LayerNormTypeConfig<XDataType, YDataType, XScaleDataType, YScaleDataType>::YScaleDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::XDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::XBiasDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::GammaDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::BetaDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::ComputeDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::YDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::MeanDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::InvStdDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::SmoothScaleDataType,
typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::YScaleDataType,
typename Traits_::Shape,
PipelineTraits>;
......@@ -224,7 +224,7 @@ float layernorm2d_fwd_(const S& s, A a)
static constexpr bool UseSmoothInputScale = Traits_::kFusedQuant == 1;
static constexpr bool UseRawStore = sizeof(YDataType) == 4;
using DynamicQuantEpilogueProblem = ck_tile::DynamicQuantEpilogueProblem<ComputeDataType, XScaleDataType, YScaleDataType, YDataType, typename Traits_::Shape,
using DynamicQuantEpilogueProblem = ck_tile::DynamicQuantEpilogueProblem<ComputeDataType, SmoothScaleDataType, YScaleDataType, YDataType, typename Traits_::Shape,
ck_tile::DynamicQuantEpilogueTraits<false, Traits_::kPadN, UseSmoothInputScale, UseRawStore, true/*max3*/>>;
using DynamicQuantEpilogue = ck_tile::DynamicQuantEpilogue<DynamicQuantEpilogueProblem>;
......@@ -249,7 +249,7 @@ float layernorm2d_fwd_(const S& s, A a)
API_BASE = """
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "layernorm2d_fwd.hpp"
......@@ -285,7 +285,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
INSTANCE_BASE = """
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "layernorm2d_fwd_api_common.hpp"
......@@ -374,7 +374,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
class h_traits:
F_XDataType : str
F_YDataType : str
F_XScaleDataType : str
F_SmoothScaleDataType : str
F_YScaleDataType : str
F_Repeat_M : int
F_Repeat_N : int
......@@ -392,7 +392,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
@property
def trait_name(self) ->str:
t_ = f'{DATA_TYPE_MAP[self.F_XDataType]}, {DATA_TYPE_MAP[self.F_YDataType]}, {DATA_TYPE_MAP[self.F_XScaleDataType]}, {DATA_TYPE_MAP[self.F_YScaleDataType]}, {self.F_Repeat_M:2}, {self.F_Repeat_N:2}, {self.F_ThreadPerBlock_M:2}, {self.F_ThreadPerBlock_N:4}'
t_ = f'{DATA_TYPE_MAP[self.F_XDataType]}, {DATA_TYPE_MAP[self.F_YDataType]}, {DATA_TYPE_MAP[self.F_SmoothScaleDataType]}, {DATA_TYPE_MAP[self.F_YScaleDataType]}, {self.F_Repeat_M:2}, {self.F_Repeat_N:2}, {self.F_ThreadPerBlock_M:2}, {self.F_ThreadPerBlock_N:4}'
t_ += f', {self.F_Vector_N:2}, {BOOL_MAP(self.F_kPadN):5}, {BOOL_MAP(self.F_kSaveMeanInvStd_):5}, {BOOL_MAP(self.F_kFastFDiv_):5}, {BOOL_MAP(self.F_kWelford_):5}'
t_ += f', {BOOL_MAP(self.F_kTwoPass_):5}, {self.F_kXbias:4}, {self.F_kFusedAdd:4}, {self.F_kFusedQuant:4}'
return t_
......@@ -477,8 +477,8 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
if ins.F_kFusedQuant == 0:
_sweep_cond = 't.fused_quant == {f_fused_sweep}'.format(f_fused_sweep = ins.F_kFusedQuant)
elif ins.F_kFusedQuant == 1:
_sweep_cond = 't.fused_quant == {f_fused_sweep} && (t.prec_sx == \"{f_sx_type}\" && t.prec_sy == \"{f_sy_type}\")'.format(
f_fused_sweep = ins.F_kFusedQuant, f_sx_type=ins.F_XScaleDataType, f_sy_type=ins.F_YScaleDataType)
_sweep_cond = 't.fused_quant == {f_fused_sweep} && (t.prec_sm == \"{f_sx_type}\" && t.prec_sy == \"{f_sy_type}\")'.format(
f_fused_sweep = ins.F_kFusedQuant, f_sx_type=ins.F_SmoothScaleDataType, f_sy_type=ins.F_YScaleDataType)
elif ins.F_kFusedQuant == 2:
_sweep_cond = 't.fused_quant == {f_fused_sweep} && (t.prec_sy == \"{f_sy_type}\")'.format(
f_fused_sweep = ins.F_kFusedQuant, f_sy_type=ins.F_YScaleDataType)
......@@ -572,7 +572,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
current_n = hs[0].F_Repeat_N * hs[0].F_ThreadPerBlock_N * hs[0].F_Vector_N
for dtype, scale_type, xbias, fused_add, fused_quant in itertools.product(dtype_list, scale_list, xbias_list, fused_add_list, fused_sweep_list):
prec_i, prec_o = dtype.split(',')
scale_x, scale_y = scale_type.split(',')
scale_sm, scale_y = scale_type.split(',')
if prec_o in dynamic_quant_out_dtype and fused_quant != 1:
continue # skip non dynamic quant case
if fused_quant == 1 and hs_key == 'big':
......@@ -582,8 +582,8 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
h_ = copy.copy(chs_) # copy the base instance out
h_.F_XDataType = prec_i
h_.F_YDataType = prec_o
h_.F_XScaleDataType = scale_y
h_.F_YScaleDataType = scale_x
h_.F_SmoothScaleDataType = scale_sm
h_.F_YScaleDataType = scale_y
h_.F_kXbias = xbias
h_.F_kFusedAdd = fused_add
h_.F_kFusedQuant = fused_quant
......
example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp
View file @ 04dd3148
......@@ -35,7 +35,7 @@ auto create_args(int argc, char* argv[])
.insert("kname", "1", "print kernel name or not")
.insert("prec_i", "fp16", "input precision")
.insert("prec_o", "auto", "output precision, set auto will be the same as input")
.insert("prec_sx",
.insert("prec_sm",
"auto",
"output quant scale type, set auto will use fp32. used when fquant=1")
.insert("prec_sy",
......@@ -53,7 +53,7 @@ auto create_args(int argc, char* argv[])
template <typename InDataType,
typename OutDataType,
typename XScaleDataType,
typename SmoothScaleDataType,
typename YScaleDataType,
bool SaveMeanVar>
bool run(const ck_tile::ArgParser& arg_parser)
......@@ -75,15 +75,15 @@ bool run(const ck_tile::ArgParser& arg_parser)
float epsilon = arg_parser.get_float("e");
std::string prec_i = arg_parser.get_str("prec_i");
std::string prec_o = arg_parser.get_str("prec_o");
std::string prec_sx = arg_parser.get_str("prec_sx");
std::string prec_sm = arg_parser.get_str("prec_sm");
std::string prec_sy = arg_parser.get_str("prec_sy");
if(prec_o == "auto")
{
prec_o = prec_i;
}
if(prec_sx == "auto")
if(prec_sm == "auto")
{
prec_sx = "fp32";
prec_sm = "fp32";
}
if(prec_sy == "auto")
{
......@@ -105,7 +105,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
assert(x_stride >= n);
using TypeConfig = LayerNormTypeConfig<InDataType, OutDataType, XScaleDataType, YScaleDataType>;
using TypeConfig =
LayerNormTypeConfig<InDataType, OutDataType, SmoothScaleDataType, YScaleDataType>;
using XDataType = typename TypeConfig::XDataType;
using YDataType = typename TypeConfig::YDataType;
......@@ -139,12 +140,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::HostTensor<YScaleDataType> y_scale_host_ref({m});
ck_tile::HostTensor<YScaleDataType> y_scale_host_dev({m});
ck_tile::HostTensor<XScaleDataType> x_scale_host({n});
ck_tile::HostTensor<XScaleDataType> x_scale_host_dev({n});
ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host({n});
ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host_dev({n});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XResidualDataType>{-.5f, .5f}(x_residual_host);
ck_tile::FillUniformDistribution<XScaleDataType>{-1.f, 1.f}(x_scale_host);
ck_tile::FillUniformDistribution<SmoothScaleDataType>{-1.f, 1.f}(sm_scale_host);
ck_tile::FillUniformDistribution<XBiasDataType>{-.5f, .5f}(x_bias_host);
ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
ck_tile::FillUniformDistribution<BetaDataType>{-.5f, .5f}(beta_host);
......@@ -155,7 +156,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::DeviceMem beta_buf(beta_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem y_buf(y_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem y_scale_buf(y_scale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem x_scale_buf(x_scale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem sm_scale_buf(sm_scale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem x_residual_buf(x_residual_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem y_residual_buf(y_residual_host.get_element_space_size_in_bytes());
......@@ -165,7 +166,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
gamma_buf.ToDevice(gamma_host.data());
beta_buf.ToDevice(beta_host.data());
x_residual_buf.ToDevice(x_residual_host.data());
x_scale_buf.ToDevice(x_scale_host.data());
sm_scale_buf.ToDevice(sm_scale_host.data());
auto prec_str = [&]() {
auto base_str = prec_i;
......@@ -186,11 +187,11 @@ bool run(const ck_tile::ArgParser& arg_parser)
<< ", yr_stride:" << yr_stride << std::flush;
layernorm2d_fwd_traits traits{
prec_i, prec_o, prec_sx, prec_sy, SaveMeanVar, xbias, fused_add, fused_quant};
prec_i, prec_o, prec_sm, prec_sy, SaveMeanVar, xbias, fused_add, fused_quant};
layernorm2d_fwd_args args{x_buf.GetDeviceBuffer(),
fused_add != 0 ? x_residual_buf.GetDeviceBuffer() : nullptr,
fused_quant == 1 ? x_scale_buf.GetDeviceBuffer() : nullptr,
fused_quant == 1 ? sm_scale_buf.GetDeviceBuffer() : nullptr,
x_bias_buf.GetDeviceBuffer(),
gamma_buf.GetDeviceBuffer(),
beta_buf.GetDeviceBuffer(),
......@@ -279,8 +280,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
for(int n_ = 0; n_ < N_; n_++)
{
// input smooth outlier
acc_(m_, n_) =
acc_(m_, n_) * ck_tile::type_convert<ComputeDataType>(x_scale_host(n_));
acc_(m_, n_) = acc_(m_, n_) *
ck_tile::type_convert<ComputeDataType>(sm_scale_host(n_));
}
}
ComputeDataType absmax = static_cast<ComputeDataType>(0);
......@@ -402,16 +403,16 @@ int main(int argc, char* argv[])
std::string prec_i = arg_parser.get_str("prec_i");
std::string prec_o = arg_parser.get_str("prec_o");
std::string prec_sx = arg_parser.get_str("prec_sx");
std::string prec_sm = arg_parser.get_str("prec_sm");
std::string prec_sy = arg_parser.get_str("prec_sy");
if(prec_o == "auto")
{
prec_o = prec_i;
}
if(prec_sx == "auto")
if(prec_sm == "auto")
{
prec_sx = "fp32";
prec_sm = "fp32";
}
if(prec_sy == "auto")
{
......@@ -420,33 +421,33 @@ int main(int argc, char* argv[])
int save_mv = arg_parser.get_int("save_mv");
// no dynamic quant case
if(prec_i == "fp16" && prec_o == "fp16" && prec_sx == "fp32" && prec_sy == "fp32" && save_mv)
if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" && save_mv)
{
return run<ck_tile::half_t, ck_tile::half_t, float, float, true>(arg_parser) ? 0 : -2;
}
else if(prec_i == "fp16" && prec_o == "fp16" && prec_sx == "fp32" && prec_sy == "fp32" &&
else if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" &&
!save_mv)
{
return run<ck_tile::half_t, ck_tile::half_t, float, float, false>(arg_parser) ? 0 : -2;
}
else if(prec_i == "bf16" && prec_o == "bf16" && prec_sx == "fp32" && prec_sy == "fp32" &&
else if(prec_i == "bf16" && prec_o == "bf16" && prec_sm == "fp32" && prec_sy == "fp32" &&
save_mv)
{
return run<ck_tile::bf16_t, ck_tile::bf16_t, float, float, true>(arg_parser) ? 0 : -2;
}
else if(prec_i == "bf16" && prec_o == "bf16" && prec_sx == "fp32" && prec_sy == "fp32" &&
else if(prec_i == "bf16" && prec_o == "bf16" && prec_sm == "fp32" && prec_sy == "fp32" &&
!save_mv)
{
return run<ck_tile::bf16_t, ck_tile::bf16_t, float, float, true>(arg_parser) ? 0 : -2;
}
// dynamic quant case, only in inference
else if(prec_i == "fp16" && prec_o == "int8" && prec_sx == "fp32" && prec_sy == "fp32" &&
else if(prec_i == "fp16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
!save_mv)
{
return run<ck_tile::half_t, ck_tile::int8_t, float, float, false>(arg_parser) ? 0 : -2;
}
else if(prec_i == "bf16" && prec_o == "int8" && prec_sx == "fp32" && prec_sy == "fp32" &&
else if(prec_i == "bf16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
!save_mv)
{
return run<ck_tile::bf16_t, ck_tile::int8_t, float, float, false>(arg_parser) ? 0 : -2;
......
example/ck_tile/02_layernorm2d/layernorm2d_fwd.hpp
View file @ 04dd3148
// 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
......@@ -8,37 +8,40 @@
#include "ck_tile/ops/layernorm2d.hpp"
#include <string>
template <typename InType, typename OutType, typename XScaleDataType_, typename YScaleDataType_>
template <typename InType,
typename OutType,
typename SmoothSScaleDataType_,
typename YScaleDataType_>
struct LayerNormTypeConfig;
template <typename OutType, typename XScaleDataType_, typename YScaleDataType_>
struct LayerNormTypeConfig<ck_tile::half_t, OutType, XScaleDataType_, YScaleDataType_>
template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
struct LayerNormTypeConfig<ck_tile::half_t, OutType, SmoothScaleDataType_, YScaleDataType_>
{
using XDataType = ck_tile::half_t;
using YDataType = OutType;
using XBiasDataType = ck_tile::half_t;
using GammaDataType = ck_tile::half_t;
using BetaDataType = ck_tile::half_t;
using MeanDataType = ck_tile::half_t;
using InvStdDataType = ck_tile::half_t;
using ComputeDataType = float;
using XScaleDataType = XScaleDataType_;
using YScaleDataType = YScaleDataType_;
using XDataType = ck_tile::half_t;
using YDataType = OutType;
using XBiasDataType = ck_tile::half_t;
using GammaDataType = ck_tile::half_t;
using BetaDataType = ck_tile::half_t;
using MeanDataType = ck_tile::half_t;
using InvStdDataType = ck_tile::half_t;
using ComputeDataType = float;
using SmoothScaleDataType = SmoothScaleDataType_;
using YScaleDataType = YScaleDataType_;
};
template <typename OutType, typename XScaleDataType_, typename YScaleDataType_>
struct LayerNormTypeConfig<ck_tile::bf16_t, OutType, XScaleDataType_, YScaleDataType_>
template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
struct LayerNormTypeConfig<ck_tile::bf16_t, OutType, SmoothScaleDataType_, YScaleDataType_>
{
using XDataType = ck_tile::bf16_t;
using YDataType = OutType;
using XBiasDataType = ck_tile::bf16_t;
using GammaDataType = ck_tile::bf16_t;
using BetaDataType = ck_tile::bf16_t;
using MeanDataType = ck_tile::bf16_t;
using InvStdDataType = ck_tile::bf16_t;
using ComputeDataType = float;
using XScaleDataType = XScaleDataType_;
using YScaleDataType = YScaleDataType_;
using XDataType = ck_tile::bf16_t;
using YDataType = OutType;
using XBiasDataType = ck_tile::bf16_t;
using GammaDataType = ck_tile::bf16_t;
using BetaDataType = ck_tile::bf16_t;
using MeanDataType = ck_tile::bf16_t;
using InvStdDataType = ck_tile::bf16_t;
using ComputeDataType = float;
using SmoothScaleDataType = SmoothScaleDataType_;
using YScaleDataType = YScaleDataType_;
};
// runtime args
......@@ -52,10 +55,10 @@ struct layernorm2d_fwd_traits
std::string prec_i; // input precision
std::string prec_o; // output precision
// if fused_quant == 1, need set prec_sx/prec_sy to proper string, otherwise can set
// if fused_quant == 1, need set prec_sm/prec_sy to proper string, otherwise can set
// arbitrary(will skip check) if fused_quant == 2, need set prec_sy to proper string, otherwise
// can set arbitrary(will skip check)
std::string prec_sx; // x-scale, used for [1*N] input smooth quant
std::string prec_sm; // x-scale, used for [1*N] input smooth quant
std::string prec_sy; // y-scale, used for [M*1] output for next layer
bool save_mean_var; //
......
example/ck_tile/10_rmsnorm2d/CMakeLists.txt
View file @ 04dd3148
set(RMSNORM2D_FWD_KNOWN_APIS "fwd;bwd")
set(RMSNORM2D_FWD_ENABLE_APIS "fwd" CACHE STRING
"semicolon-separated list of APIs to generate (${RMSNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
if(RMSNORM2D_FWD_ENABLE_APIS STREQUAL "all")
set(RMSNORM2D_FWD_ENABLE_APIS ${RMSNORM2D_FWD_KNOWN_APIS})
endif()
# generate a list of kernels, but not actually emit files at config sta
execute_process(
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api ${RMSNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --list_blobs
RESULT_VARIABLE ret
)
if(ret AND NOT ret EQUAL 0)
message( FATAL_ERROR "Fail to generate kernels via Python. ${ret}")
endif()
file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/rmsnorm2d_fwd_blobs.txt RMSNORM2D_FWD_GEN_BLOBS)
add_custom_command(
OUTPUT ${RMSNORM2D_FWD_GEN_BLOBS}
COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
--api ${RMSNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --gen_blobs
)
set(TILE_RMSNORM2D_FWD "tile_rmsnorm2d_fwd")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message("adding ${TILE_RMSNORM2D_FWD}")
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_executable(${TILE_RMSNORM2D_FWD} EXCLUDE_FROM_ALL rmsnorm2d_fwd.cpp)
target_include_directories(${TILE_RMSNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${TILE_RMSNORM2D_FWD} PRIVATE ${INSTANCE_SRCS})
target_sources(${TILE_RMSNORM2D_FWD} PRIVATE ${RMSNORM2D_FWD_GEN_BLOBS})
set(TILE_RMSNORM2D_FWD_COMPILE_OPTIONS)
......
example/ck_tile/10_rmsnorm2d/example_rmsnorm2d_fwd.cpp
View file @ 04dd3148
#include "ck_tile/host.hpp"
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/epilogue.hpp"
#include "ck_tile/ops/rmsnorm2d.hpp"
#include <cstring>
......@@ -36,10 +37,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
assert(stride >= n);
using XDataType = DataType;
using YDataType = DataType;
using GammaDataType = DataType;
using InvRmsDataType = ck_tile::null_type;
using XDataType = DataType;
using YDataType = DataType;
using GammaDataType = DataType;
using InvRmsDataType = ck_tile::null_type;
using SmoothScaleDataType = ck_tile::null_type;
using YScaleDataType = ck_tile::null_type;
using ComputeDataType = float;
......@@ -68,30 +71,49 @@ bool run(const ck_tile::ArgParser& arg_parser)
using BlockTile = ck_tile::sequence<2, 128>;
using WarpTile = ck_tile::sequence<1, 64>;
using Vector = ck_tile::sequence<1, 1>;
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
using PipelineTraits =
ck_tile::Rmsnorm2dFwdTraits<true, // kPadN
false, // kSaveInvRms
kTwoPass,
ck_tile::Rmsnorm2dFusedAddEnum::NO_ADD, // fuse add
ck_tile::Rmsnorm2dFusedQuantEnum::NO_SWEEP>; // fuse quant
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
using Problem = ck_tile::Rmsnorm2dFwdPipelineProblem<XDataType,
GammaDataType,
ComputeDataType,
YDataType,
InvRmsDataType,
SmoothScaleDataType,
YScaleDataType,
Shape,
true, // kPadN
false, // kSaveInvRms
kTwoPass>;
PipelineTraits>;
using OnePassPipeline = ck_tile::Rmsnorm2dFwdPipelineOnePass<Problem>;
using TwoPassPipeline = ck_tile::Rmsnorm2dFwdPipelineTwoPass<Problem>;
using Pipeline = std::conditional_t<kTwoPass, TwoPassPipeline, OnePassPipeline>;
using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline>;
using Default2DEpilogueProblem = ck_tile::
Default2DEpilogueProblem<ComputeDataType, YDataType, false, PipelineTraits::kPadN, false>;
using Default2DEpilogue = ck_tile::Default2DEpilogue<Default2DEpilogueProblem>;
using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline, Default2DEpilogue>;
ck_tile::Rmsnorm2dFwdHostArgs args{x_buf.GetDeviceBuffer(),
nullptr,
nullptr,
gamma_buf.GetDeviceBuffer(),
y_buf.GetDeviceBuffer(),
nullptr,
nullptr,
nullptr,
epsilon,
m,
n,
stride,
stride,
stride,
stride};
auto kargs = Kernel::MakeKargs(args);
......
example/ck_tile/10_rmsnorm2d/generate.py 0 → 100644
View file @ 04dd3148
This diff is collapsed.
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_api.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "rmsnorm2d_fwd.hpp"
template <typename DataType_,
ck_tile::index_t Repeat_M_, // each thread repeat along M
ck_tile::index_t Repeat_N_, // each thread repeat along N
ck_tile::index_t ThreadPerBlock_M_, // num threads along M
ck_tile::index_t ThreadPerBlock_N_, // num threads along N
ck_tile::index_t Vector_N_, // vector size along N
bool kPadN_,
bool kSaveInvRms_,
bool kTwoPass_>
using trait_ = rmsnorm2d_fwd_traits_<DataType_,
Repeat_M_,
Repeat_N_,
ThreadPerBlock_M_,
ThreadPerBlock_N_,
Vector_N_,
kPadN_,
kSaveInvRms_,
kTwoPass_>;
template <typename data_type>
float rmsnorm2d_fwd_b16_(rmsnorm2d_fwd_traits /*t*/,
rmsnorm2d_fwd_args a,
const ck_tile::stream_config& s)
{
float r = -1;
// clang-format off
// rm rn tm tn vn pd rms 2p
if(a.n <= 64) {
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 128) {
if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 256) {
if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 512) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 4, 64, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 4, 64, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 8, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 768) {
if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 4, 64, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 6, 4, 64, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1,12, 4, 64, 1, true, false, false>>(s, a);
}
else if(a.n <= 1024) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 2, 128, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 2, 128, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 2, 128, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 1536) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 4, 64, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 2, 128, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 1, 256, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 6, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 2048) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 1, 1, 256, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 8, 1, 256, 1, true, false, false>>(s, a);
}
else if(a.n <= 3072) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 1, 128, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 6, 1, 256, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 3, 1, 1024, 1, true, false, false>>(s, a);
}
else if(a.n <= 4096) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 8, true, false, false>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 4, true, false, false>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 1024, 2, true, false, false>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 1024, 1, true, false, false>>(s, a);
}
else if(a.n > 4096) {
if (a.n % 8 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 256, 8, true, false, true>>(s, a);
else if (a.n % 4 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 256, 4, true, false, true>>(s, a);
else if (a.n % 2 == 0)
r = rmsnorm2d_fwd_<trait_<data_type, 1, 2, 1, 1024, 2, true, false, true>>(s, a);
else
r = rmsnorm2d_fwd_<trait_<data_type, 1, 4, 1, 1024, 1, true, false, true>>(s, a);
}
return r;
// clang-format on
}
float rmsnorm2d_fwd(rmsnorm2d_fwd_traits t, rmsnorm2d_fwd_args a, const ck_tile::stream_config& s)
{
if(t.data_type.compare("fp16") == 0)
{
return rmsnorm2d_fwd_b16_<ck_tile::fp16_t>(t, a, s);
}
else if(t.data_type.compare("bf16") == 0)
{
return rmsnorm2d_fwd_b16_<ck_tile::bf16_t>(t, a, s);
}
else
throw std::runtime_error("Without supported instances!");
}
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n1024_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
#if 0
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 8, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 16, 4, 64, 1, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 4, true , false, false>>(const S&, A);
#endif
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 2, 128, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 2, 128, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 2, 128, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n1536_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 2, 128, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n2048_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 1, 256, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n256_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n3072_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 128, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n4096_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n4096_tp_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 256, 8, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 256, 4, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 1, 1024, 2, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 1, 1024, 1, true, false, true>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n512_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 8, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 4, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 8, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n64_n128_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 1, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 1, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 2, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n768_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 3, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 6, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::bf16_t, 1, 12, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n1024_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
#if 0
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 8, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 16, 4, 64, 1, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 1, 256, 4, true , false, false>>(const S&, A);
#endif
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 2, 128, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 2, 128, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 2, 128, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n1536_instance.cpp deleted 100644 → 0
View file @ c0b90f13
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "rmsnorm2d_fwd_instance_common.hpp"
// clang-format off
// rm rn tm tn vn pd rms 2p
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 4, 64, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 2, 128, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment