Skip to content

GitLab

"...targets/git@developer.sourcefind.cn:gaoqiong/migraphx.git" did not exist on "fa52b516e073a72ae6c88a6bc0c1a1b97b07a2fe"
Commit 1b616990 authored by aska-0096's avatar aska-0096
Browse files

Merge branch 'develop' of https://github.com/ROCm/composable_kernel into update_cka8w8_uc

parents af30d6b6 800cf897
Hide whitespace changes
Inline Side-by-side
Showing with 330 additions and 435 deletions
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_bf16_n256_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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 @ af30d6b6
// 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 @ af30d6b6
// 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 @ af30d6b6
// 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 @ af30d6b6
// 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 @ af30d6b6
// 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 @ af30d6b6
// 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 @ af30d6b6
// 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 @ af30d6b6
// 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
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n2048_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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, 1, 1, 256, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 1, 256, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n256_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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, 1, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n3072_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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, 1, 128, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 1, 256, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 3, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n4096_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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, 2, 1, 256, 8, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n4096_tp_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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, 2, 1, 256, 8, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 256, 4, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 1, 1024, 2, true, false, true>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 1, 1024, 1, true, false, true>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n512_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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, 1, 4, 64, 8, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 4, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 8, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n64_n128_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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, 1, 4, 64, 1, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 1, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 2, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_fp16_n768_instance.cpp deleted 100644 → 0
View file @ af30d6b6
// 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, 4, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 6, 4, 64, 2, true , false, false>>(const S&, A);
template float rmsnorm2d_fwd_<trait_<ck_tile::fp16_t, 1, 12, 4, 64, 1, true , false, false>>(const S&, A);
// clang-format on
example/ck_tile/10_rmsnorm2d/instances/rmsnorm2d_fwd_instance_common.hpp deleted 100644 → 0
View file @ af30d6b6
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <ck_tile/core.hpp>
#include "rmsnorm2d_fwd.hpp"
#include <iostream>
#pragma once
using S = ck_tile::stream_config;
using A = rmsnorm2d_fwd_args;
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 Traits_>
float rmsnorm2d_fwd_(const S& s, A a)
{
using DataType = typename Traits_::DataType;
using PipelineProblem =
ck_tile::Rmsnorm2dFwdPipelineProblem<typename RmsnormTypeConfig<DataType>::XDataType,
typename RmsnormTypeConfig<DataType>::GammaDataType,
typename RmsnormTypeConfig<DataType>::ComputeDataType,
typename RmsnormTypeConfig<DataType>::YDataType,
typename RmsnormTypeConfig<DataType>::InvRmsDataType,
typename Traits_::Shape,
Traits_::kPadN,
Traits_::kSaveInvRms,
Traits_::kTwoPass>;
using OnePassPipeline = ck_tile::Rmsnorm2dFwdPipelineOnePass<PipelineProblem>;
using TwoPassPipeline = ck_tile::Rmsnorm2dFwdPipelineTwoPass<PipelineProblem>;
using Pipeline = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline>;
const dim3 grids = Kernel::GridSize(a);
constexpr dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto kargs = Kernel::MakeKargs(a);
if(s.log_level_ > 0)
std::cout << ", " << Kernel::GetName() << std::flush;
return ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
}
example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.cpp
View file @ 1b616990
...@@ -19,17 +19,37 @@ auto get_elimit<ck_tile::bf16_t>() ...@@ -19,17 +19,37 @@ auto get_elimit<ck_tile::bf16_t>()
return ck_tile::make_tuple(rtol, atol); return ck_tile::make_tuple(rtol, atol);
} }
template <>
auto get_elimit<ck_tile::int8_t>()
{
double rtol = 1e-02;
double atol = 1.0;
return ck_tile::make_tuple(rtol, atol);
}
auto create_args(int argc, char* argv[]) auto create_args(int argc, char* argv[])
{ {
ck_tile::ArgParser arg_parser; ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension") arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension") .insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n") .insert("x_stride", "-1", "x row_stride, if -1 then equal to n")
.insert("xr_stride", "-1", "x residule row_stride, if -1 then equal to n")
.insert("y_stride", "-1", "y row_stride, if -1 then equal to n")
.insert("yr_stride", "-1", "y residule row_stride, if -1 then equal to n")
.insert("e", "1e-5", "epsilon") .insert("e", "1e-5", "epsilon")
.insert("save_rms", "0", "save rms(invrms) or not. set to 1 in training case") .insert("save_rms", "0", "save rms(invrms) or not. set to 1 in training case")
.insert("v", "1", "cpu validation or not") .insert("v", "1", "cpu validation or not")
.insert("kname", "1", "print kernel name or not") .insert("kname", "1", "print kernel name or not")
.insert("prec", "fp16", "precision") .insert("prec_i", "fp16", "input precision")
.insert("prec_o", "auto", "output precision, set auto will be the same as input")
.insert("prec_sm",
"auto",
"output quant scale type, set auto will use fp32. used when fquant=1")
.insert("prec_sy",
"auto",
"output quant scale type, set auto will use fp32. used when fquant=1 or 2")
.insert("fadd", "0", "fused-add, 0:no fused add, 1:preadd+store, 2:preadd only")
.insert("fquant", "0", "fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant")
.insert("warmup", "5", "cold iter") .insert("warmup", "5", "cold iter")
.insert("repeat", "20", "hot iter"); .insert("repeat", "20", "hot iter");
...@@ -37,28 +57,70 @@ auto create_args(int argc, char* argv[]) ...@@ -37,28 +57,70 @@ auto create_args(int argc, char* argv[])
return std::make_tuple(result, arg_parser); return std::make_tuple(result, arg_parser);
} }
template <typename DataType, bool SaveRms> template <typename InDataType,
typename OutDataType,
typename SmoothScaleDataType,
typename YScaleDataType,
bool SaveRms>
bool run(const ck_tile::ArgParser& arg_parser) bool run(const ck_tile::ArgParser& arg_parser)
{ {
ck_tile::index_t m = arg_parser.get_int("m"); ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n"); ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride"); float epsilon = arg_parser.get_float("e");
if(stride < 0) int kname = arg_parser.get_int("kname");
stride = n; int do_validation = arg_parser.get_int("v");
float epsilon = arg_parser.get_float("e"); int fused_add = arg_parser.get_int("fadd");
std::string data_type = arg_parser.get_str("prec"); int fused_quant = arg_parser.get_int("fquant");
int kname = arg_parser.get_int("kname"); int warmup = arg_parser.get_int("warmup");
int do_validation = arg_parser.get_int("v"); int repeat = arg_parser.get_int("repeat");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat"); ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
if(x_stride < 0)
assert(stride >= n); x_stride = n;
ck_tile::index_t xr_stride = arg_parser.get_int("xr_stride");
using TypeConfig = RmsnormTypeConfig<DataType>; if(xr_stride < 0)
xr_stride = n;
using XDataType = typename TypeConfig::XDataType; ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
using YDataType = typename TypeConfig::YDataType; if(y_stride < 0)
using GammaDataType = typename TypeConfig::GammaDataType; y_stride = n;
ck_tile::index_t yr_stride = arg_parser.get_int("yr_stride");
if(yr_stride < 0)
yr_stride = n;
assert(x_stride >= n);
std::string prec_i = arg_parser.get_str("prec_i");
std::string prec_o = arg_parser.get_str("prec_o");
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_sm == "auto")
{
prec_sm = "fp32";
}
if(prec_sy == "auto")
{
prec_sy = "fp32";
}
if((fused_quant == 1 || fused_quant == 2) && prec_o != "int8" && prec_o != "fp8")
{
std::cout
<< "if fused_quant is 1 or 2, only support \"-prec_o=int8\" or \"-prec_o=fp8\" cases."
<< std::endl;
return false;
}
using TypeConfig =
RmsnormTypeConfig<InDataType, OutDataType, SmoothScaleDataType, YScaleDataType>;
using XDataType = typename TypeConfig::XDataType;
using YDataType = typename TypeConfig::YDataType;
using GammaDataType = typename TypeConfig::GammaDataType;
using XResidualDataType = XDataType;
using YResidualDataType = XDataType;
using InvRmsDataType = using InvRmsDataType =
std::conditional_t<SaveRms, typename TypeConfig::InvRmsDataType, ck_tile::null_type>; std::conditional_t<SaveRms, typename TypeConfig::InvRmsDataType, ck_tile::null_type>;
...@@ -66,43 +128,84 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -66,43 +128,84 @@ bool run(const ck_tile::ArgParser& arg_parser)
using ComputeDataType = typename TypeConfig::ComputeDataType; using ComputeDataType = typename TypeConfig::ComputeDataType;
// host verify // host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1}); ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
ck_tile::HostTensor<GammaDataType> gamma_host({n}); ck_tile::HostTensor<GammaDataType> gamma_host({n});
ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host({n});
ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host_dev({n});
ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {stride, 1}); ck_tile::HostTensor<XResidualDataType> x_residual_host({m, n}, {xr_stride, 1});
ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {stride, 1}); ck_tile::HostTensor<YResidualDataType> y_residual_host({m, n}, {yr_stride, 1});
ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {y_stride, 1});
ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {y_stride, 1});
ck_tile::HostTensor<YScaleDataType> y_scale_host_ref({m});
ck_tile::HostTensor<YScaleDataType> y_scale_host_dev({m});
ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m}); ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host); ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XResidualDataType>{-.5f, .5f}(x_residual_host);
ck_tile::FillUniformDistribution<SmoothScaleDataType>{-1.f, 1.f}(sm_scale_host);
ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host); ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes()); ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem gamma_buf(gamma_host.get_element_space_size_in_bytes()); ck_tile::DeviceMem gamma_buf(gamma_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_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 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());
x_buf.ToDevice(x_host.data()); x_buf.ToDevice(x_host.data());
gamma_buf.ToDevice(gamma_host.data()); gamma_buf.ToDevice(gamma_host.data());
x_residual_buf.ToDevice(x_residual_host.data());
sm_scale_buf.ToDevice(sm_scale_host.data());
auto prec_str = [&]() {
auto base_str = prec_i;
if(prec_i != prec_o)
{
base_str += "|" + prec_o;
}
if(fused_quant == 1)
{
base_str += std::string("(") + prec_sy + ")";
}
return base_str;
}();
std::cout << "[" << data_type << "]" std::cout << "[" << prec_str << "]"
<< " m:" << m << ", n:" << n << ", stride:" << stride << std::flush; << " m:" << m << ", n:" << n << ", x_stride:" << x_stride
<< ", xr_stride:" << xr_stride << ", y_stride:" << y_stride
<< ", yr_stride:" << yr_stride << std::flush;
rmsnorm2d_fwd_traits traits{data_type, SaveRms}; rmsnorm2d_fwd_traits traits{prec_i, prec_o, prec_sm, prec_sy, SaveRms, fused_add, fused_quant};
rmsnorm2d_fwd_args args{x_buf.GetDeviceBuffer(), rmsnorm2d_fwd_args args{x_buf.GetDeviceBuffer(),
fused_add != 0 ? x_residual_buf.GetDeviceBuffer() : nullptr,
fused_quant == 1 ? sm_scale_buf.GetDeviceBuffer() : nullptr,
gamma_buf.GetDeviceBuffer(), gamma_buf.GetDeviceBuffer(),
y_buf.GetDeviceBuffer(), y_buf.GetDeviceBuffer(),
nullptr, fused_add == 1 ? y_residual_buf.GetDeviceBuffer() : nullptr,
fused_quant != 0 ? y_scale_buf.GetDeviceBuffer() : nullptr,
nullptr, // p_invRms, unsupported yet
epsilon, epsilon,
m, m,
n, n,
stride}; x_stride, // x row_stride
xr_stride, // x residule row stride
y_stride, // y row stride
yr_stride}; // y residule row stride
float ave_time = rmsnorm2d_fwd( float ave_time = rmsnorm2d_fwd(
traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat}); traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
std::size_t num_byte = std::size_t num_byte =
sizeof(XDataType) * m * n + sizeof(GammaDataType) * n + sizeof(YDataType) * m * n; sizeof(XDataType) * m * n + sizeof(GammaDataType) * n + sizeof(YDataType) * m * n;
num_byte += SaveRms ? sizeof(InvRmsDataType) * m * n : 0;
num_byte += fused_add ? sizeof(XResidualDataType) * m * n : 0;
num_byte += ((fused_quant == 1) || (fused_quant == 2)) ? sizeof(YScaleDataType) * m : 0;
num_byte += (fused_quant == 1) ? sizeof(SmoothScaleDataType) * n : 0;
float gb_per_sec = num_byte / 1.E6 / ave_time; float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush; std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush;
...@@ -112,38 +215,135 @@ bool run(const ck_tile::ArgParser& arg_parser) ...@@ -112,38 +215,135 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(do_validation) if(do_validation)
{ {
// reference // reference
ck_tile::reference_rmsnorm2d_fwd<XDataType, if(fused_add != 0)
GammaDataType, {
ComputeDataType, // fused pre_add/pre_add_store
YDataType, // TODO we accumulate directly to x_host for simplcity here...
InvRmsDataType>( std::transform(x_host.mData.cbegin(),
x_host, gamma_host, y_host_ref, invRms_host_ref, epsilon); x_host.mData.cend(),
x_residual_host.mData.cbegin(),
x_host.mData.begin(),
[](auto x_, auto r_) {
auto o_ = ck_tile::type_convert<ComputeDataType>(x_) +
ck_tile::type_convert<ComputeDataType>(r_);
return ck_tile::type_convert<XDataType>(o_);
});
}
if(fused_quant != 0)
{
auto dquant_functor = [&](int m_, auto& o_, auto& acc_) {
int N_ = acc_.mDesc.get_lengths()[1];
if(fused_quant == 1)
{
for(int n_ = 0; n_ < N_; n_++)
{
// input smooth outlier
acc_(m_, n_) = acc_(m_, n_) *
ck_tile::type_convert<ComputeDataType>(sm_scale_host(n_));
}
}
ComputeDataType absmax = static_cast<ComputeDataType>(0);
for(int n_ = 0; n_ < N_; n_++)
{
const auto a = ck_tile::abs(acc_(m_, n_));
absmax = a > absmax ? a : absmax;
}
// printf("cpu:absmax:%f\n", absmax);
constexpr ComputeDataType kMaxY =
std::is_same<YDataType, ck_tile::fp8_t>::value ? 240.0
: std::is_same<YDataType, ck_tile::int8_t>::value ? 127.0
: 0.0;
ComputeDataType y_scale = absmax / kMaxY;
y_scale_host_ref(m_) = ck_tile::type_convert<YScaleDataType>(y_scale);
for(int n_ = 0; n_ < N_; n_++)
{
o_(m_, n_) = ck_tile::type_convert<YDataType>(acc_(m_, n_) / y_scale);
}
};
ck_tile::reference_rmsnorm2d_fwd<XDataType,
GammaDataType,
ComputeDataType,
YDataType,
InvRmsDataType>(
x_host, gamma_host, y_host_ref, invRms_host_ref, epsilon, dquant_functor);
}
else
{
ck_tile::reference_rmsnorm2d_fwd<XDataType,
GammaDataType,
ComputeDataType,
YDataType,
InvRmsDataType>(
x_host, gamma_host, y_host_ref, invRms_host_ref, epsilon);
}
y_buf.FromDevice(y_host_dev.data()); y_buf.FromDevice(y_host_dev.data());
auto [rtol, atol] = get_elimit<DataType>(); ck_tile::HostTensor<YResidualDataType> y_residual_host_dev({m, n}, {yr_stride, 1});
if(stride == n) if(fused_add == 1)
{
y_residual_buf.FromDevice(y_residual_host_dev.data());
}
auto [rtol, atol] = get_elimit<YDataType>();
if(x_stride == n)
{ {
pass = ck_tile::check_err( pass = ck_tile::check_err(
y_host_dev, y_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol); y_host_dev, y_host_ref, std::string("\nOUT Error: Incorrect results!"), rtol, atol);
if(fused_add == 1)
{
pass &= ck_tile::check_err(y_residual_host_dev,
x_host,
std::string("\nADD Error: Incorrect results!"),
rtol,
atol);
}
} }
else else
{ {
for(int i_r = 0; i_r < m; i_r++) for(int i_r = 0; i_r < m; i_r++)
{ {
std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * stride, std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * y_stride,
y_host_dev.begin() + i_r * stride + n); y_host_dev.begin() + i_r * y_stride + n);
std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * stride, std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * y_stride,
y_host_ref.begin() + i_r * stride + n); y_host_ref.begin() + i_r * y_stride + n);
pass &= ck_tile::check_err(y_host_dev_row, pass &= ck_tile::check_err(y_host_dev_row,
y_host_ref_row, y_host_ref_row,
std::string("OUT[") + std::to_string(i_r) + std::string("\nOUT[") + std::to_string(i_r) +
std::string("] Error: Incorrect results!"), std::string("] Error: Incorrect results!"),
rtol, rtol,
atol); atol);
if(fused_add == 1)
{
std::vector<YResidualDataType> y_residual_host_dev_row(
y_residual_host_dev.begin() + i_r * yr_stride,
y_residual_host_dev.begin() + i_r * yr_stride + n);
std::vector<YResidualDataType> y_residual_host_ref_row(
x_host.begin() + i_r * yr_stride, x_host.begin() + i_r * yr_stride + n);
pass &= ck_tile::check_err(y_residual_host_dev_row,
y_residual_host_ref_row,
std::string("\nADD[") + std::to_string(i_r) +
std::string("] Error: Incorrect results!"),
rtol,
atol);
}
} }
} }
if(fused_quant == 1)
{
y_scale_buf.FromDevice(y_scale_host_dev.data());
pass &= ck_tile::check_err(y_scale_host_dev,
y_scale_host_ref,
std::string("\nSCALE Error: Incorrect results!"),
rtol,
atol);
}
std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl; std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
} }
...@@ -156,23 +356,65 @@ int main(int argc, char* argv[]) ...@@ -156,23 +356,65 @@ int main(int argc, char* argv[])
if(!result) if(!result)
return -1; return -1;
const std::string data_type = arg_parser.get_str("prec"); std::string prec_i = arg_parser.get_str("prec_i");
int save_rms = arg_parser.get_int("save_rms"); std::string prec_o = arg_parser.get_str("prec_o");
if(data_type == "fp16" && save_rms) 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_sm == "auto")
{
prec_sm = "fp32";
}
if(prec_sy == "auto")
{
prec_sy = "fp32";
}
int save_rms = arg_parser.get_int("save_rms");
if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" && save_rms)
{
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_sm == "fp32" && prec_sy == "fp32" &&
!save_rms)
{
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_sm == "fp32" && prec_sy == "fp32" &&
save_rms)
{
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_sm == "fp32" && prec_sy == "fp32" &&
!save_rms)
{
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_sm == "fp32" && prec_sy == "fp32" &&
!save_rms)
{ {
return run<ck_tile::half_t, true>(arg_parser) ? 0 : -2; return run<ck_tile::half_t, ck_tile::int8_t, float, float, true>(arg_parser) ? 0 : -2;
} }
else if(data_type == "fp16" && !save_rms) else if(prec_i == "bf16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
!save_rms)
{ {
return run<ck_tile::half_t, false>(arg_parser) ? 0 : -2; return run<ck_tile::bf16_t, ck_tile::int8_t, float, float, true>(arg_parser) ? 0 : -2;
} }
else if(data_type == "bf16" && save_rms) else if(prec_i == "fp16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
!save_rms)
{ {
return run<ck_tile::bf16_t, true>(arg_parser) ? 0 : -2; return run<ck_tile::half_t, ck_tile::fp8_t, float, float, false>(arg_parser) ? 0 : -2;
} }
else if(data_type == "bf16" && !save_rms) else if(prec_i == "bf16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
!save_rms)
{ {
return run<ck_tile::bf16_t, true>(arg_parser) ? 0 : -2; return run<ck_tile::bf16_t, ck_tile::fp8_t, float, float, false>(arg_parser) ? 0 : -2;
} }
return -3; return -3;
......
example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.hpp
View file @ 1b616990
// SPDX-License-Identifier: MIT // 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 #pragma once
...@@ -8,27 +8,34 @@ ...@@ -8,27 +8,34 @@
#include "ck_tile/ops/rmsnorm2d.hpp" #include "ck_tile/ops/rmsnorm2d.hpp"
#include <string> #include <string>
template <typename DataType> template <typename InType,
typename OutType,
typename SmoothScaleDataType_,
typename YScaleDataType_>
struct RmsnormTypeConfig; struct RmsnormTypeConfig;
template <> template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
struct RmsnormTypeConfig<ck_tile::half_t> struct RmsnormTypeConfig<ck_tile::half_t, OutType, SmoothScaleDataType_, YScaleDataType_>
{ {
using XDataType = ck_tile::half_t; using XDataType = ck_tile::half_t;
using YDataType = ck_tile::half_t; using YDataType = OutType;
using GammaDataType = ck_tile::half_t; using GammaDataType = ck_tile::half_t;
using InvRmsDataType = ck_tile::half_t; using InvRmsDataType = ck_tile::half_t;
using ComputeDataType = float; using ComputeDataType = float;
using SmoothScaleDataType = SmoothScaleDataType_;
using YScaleDataType = YScaleDataType_;
}; };
template <> template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
struct RmsnormTypeConfig<ck_tile::bf16_t> struct RmsnormTypeConfig<ck_tile::bf16_t, OutType, SmoothScaleDataType_, YScaleDataType_>
{ {
using XDataType = ck_tile::bf16_t; using XDataType = ck_tile::bf16_t;
using YDataType = ck_tile::bf16_t; using YDataType = OutType;
using GammaDataType = ck_tile::bf16_t; using GammaDataType = ck_tile::bf16_t;
using InvRmsDataType = ck_tile::bf16_t; using InvRmsDataType = ck_tile::bf16_t;
using ComputeDataType = float; using ComputeDataType = float;
using SmoothScaleDataType = SmoothScaleDataType_;
using YScaleDataType = YScaleDataType_;
}; };
// runtime args // runtime args
...@@ -36,82 +43,24 @@ struct rmsnorm2d_fwd_args : public ck_tile::Rmsnorm2dFwdHostArgs ...@@ -36,82 +43,24 @@ struct rmsnorm2d_fwd_args : public ck_tile::Rmsnorm2dFwdHostArgs
{ {
}; };
// this is used to pattern-match internl kernel implementation, not to instantiate kernel
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_>
struct rmsnorm2d_fwd_traits_
{
using DataType = ck_tile::remove_cvref_t<DataType_>;
static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0);
static constexpr ck_tile::index_t total_warps =
(ThreadPerBlock_M_ * ThreadPerBlock_N_) / warpSize;
// num of warps along m
static constexpr ck_tile::index_t BlockWarps_M = []() {
if constexpr(is_warp_per_row)
{
static_assert(warpSize % ThreadPerBlock_N_ == 0);
return total_warps * (warpSize / ThreadPerBlock_N_);
}
else
{
// static_assert(warpSize % ThreadPerBlock_M_ == 0);
return total_warps / (ThreadPerBlock_N_ / warpSize);
}
}();
// num of warps along n
static constexpr ck_tile::index_t BlockWarps_N = []() {
if constexpr(is_warp_per_row)
{
static_assert(warpSize % ThreadPerBlock_N_ == 0);
return 1;
}
else
{
static_assert(ThreadPerBlock_N_ % warpSize == 0);
return ThreadPerBlock_N_ / warpSize;
}
}();
static constexpr ck_tile::index_t Repeat_M = Repeat_M_;
static constexpr ck_tile::index_t Repeat_N = Repeat_N_;
static constexpr ck_tile::index_t Block_M = Repeat_M_ * ThreadPerBlock_M_;
static constexpr ck_tile::index_t Block_N = Repeat_N_ * ThreadPerBlock_N_ * Vector_N_;
static constexpr ck_tile::index_t Warp_M = ThreadPerBlock_M_ / BlockWarps_M;
static constexpr ck_tile::index_t Warp_N = ThreadPerBlock_N_ / BlockWarps_N * Vector_N_;
using BlockTile = ck_tile::sequence<Block_M, Block_N>;
using BlockWarps = ck_tile::sequence<BlockWarps_M, BlockWarps_N>;
using WarpTile = ck_tile::sequence<Warp_M, Warp_N>;
using Vector = ck_tile::sequence<1, Vector_N_>;
using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
static constexpr bool kPadN = kPadN_;
static constexpr bool kSaveInvRms = kSaveInvRms_;
static constexpr bool kTwoPass = kTwoPass_;
};
template <typename Traits_> template <typename Traits_>
float rmsnorm2d_fwd_(const ck_tile::stream_config& s, rmsnorm2d_fwd_args a); float rmsnorm2d_fwd_(const ck_tile::stream_config& s, rmsnorm2d_fwd_args a);
// This is the public API, will be generated by script // This is the public API, will be generated by script
struct rmsnorm2d_fwd_traits struct rmsnorm2d_fwd_traits
{ {
std::string data_type; std::string prec_i; // input precision
std::string prec_o; // output precision
// 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_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_rms; bool save_rms;
int fused_add; // 0:no-add, 1:pre-add-store, 2:pre-add
int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
}; };
float rmsnorm2d_fwd(rmsnorm2d_fwd_traits, rmsnorm2d_fwd_args, const ck_tile::stream_config&); float rmsnorm2d_fwd(rmsnorm2d_fwd_traits, rmsnorm2d_fwd_args, const ck_tile::stream_config&);
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