mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-07-11 17:51:40 +00:00
local base version
This commit is contained in:
AMD-dteng
@@ -25,6 +25,7 @@
|
||||
#include "ck_tile/host/reference/reference_gemm.hpp"
|
||||
#include "ck_tile/host/reference/reference_im2col.hpp"
|
||||
#include "ck_tile/host/reference/reference_layernorm2d_fwd.hpp"
|
||||
#include "ck_tile/host/reference/reference_layernorm2d_bwd.hpp"
|
||||
#include "ck_tile/host/reference/reference_moe_sorting.hpp"
|
||||
#include "ck_tile/host/reference/reference_permute.hpp"
|
||||
#include "ck_tile/host/reference/reference_reduce.hpp"
|
||||
|
||||
86
include/ck_tile/host/reference/reference_layernorm2d_bwd.hpp
Normal file
86
include/ck_tile/host/reference/reference_layernorm2d_bwd.hpp
Normal file
@@ -0,0 +1,86 @@
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename XDataType,
|
||||
typename GammaDataType,
|
||||
typename BetaDataType,
|
||||
typename ComputeDataType,
|
||||
typename YDataType,
|
||||
typename MeanDataType,
|
||||
typename InvStdDataType>
|
||||
CK_TILE_HOST void reference_layernorm2d_bwd_gamma_part(const HostTensor<XDataType>& x_m_n,
|
||||
const HostTensor<YDataType>& dy_m_n,
|
||||
const HostTensor<GammaDataType>& gamma_n,
|
||||
const HostTensor<MeanDataType>& mean_m,
|
||||
const HostTensor<InvStdDataType>& inv_std_m,
|
||||
HostTensor<GammaDataType>& dgamma_mpart_n,
|
||||
HostTensor<BetaDataType>& dbeta_mpart_n,
|
||||
HostTensor<XDataType>& dx_m_n,
|
||||
|
||||
//tmp
|
||||
HostTensor<ComputeDataType>& ds_m,
|
||||
HostTensor<ComputeDataType>& db_m)
|
||||
{
|
||||
|
||||
const auto MN = x_m_n.mDesc.get_lengths();
|
||||
const int M = MN[0];
|
||||
const int N = MN[1];
|
||||
const int PartM = dgamma_mpart_n.mDesc.get_lengths()[0];
|
||||
const int MLoop = (M + PartM - 1) / PartM;
|
||||
printf("\ndteng print---M=%d,N=%d,PartM=%d,MLoop=%d\n",M,N,PartM,MLoop);
|
||||
auto f = [&](auto m) {
|
||||
const int m_offset = m * MLoop;
|
||||
//calculate dgamma, dbeta
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
ComputeDataType gamma_acc = 0;
|
||||
ComputeDataType beta_acc = 0;
|
||||
for(int inner_m = 0; inner_m < MLoop && m_offset + inner_m < M; inner_m++)
|
||||
{
|
||||
const ComputeDataType mean = ck_tile::type_convert<ComputeDataType>(mean_m(m_offset + inner_m));
|
||||
const ComputeDataType inv_std = ck_tile::type_convert<ComputeDataType>(inv_std_m(m_offset + inner_m));
|
||||
const ComputeDataType x = ck_tile::type_convert<ComputeDataType>(x_m_n(m_offset + inner_m, n));
|
||||
const ComputeDataType dy = ck_tile::type_convert<ComputeDataType>(dy_m_n(m_offset + inner_m, n));
|
||||
gamma_acc += dy * (x - mean) * inv_std;
|
||||
beta_acc += dy;
|
||||
}
|
||||
|
||||
dgamma_mpart_n(m, n) = ck_tile::type_convert<GammaDataType>(gamma_acc);
|
||||
dbeta_mpart_n(m, n) = ck_tile::type_convert<BetaDataType>(beta_acc);
|
||||
}
|
||||
|
||||
//calculate dx
|
||||
for(int inner_m = 0; inner_m < MLoop && m_offset + inner_m < M; inner_m++)
|
||||
{
|
||||
ComputeDataType ds = 0;
|
||||
ComputeDataType db = 0;
|
||||
const ComputeDataType mean = ck_tile::type_convert<ComputeDataType>(mean_m(m_offset + inner_m));
|
||||
const ComputeDataType inv_std = ck_tile::type_convert<ComputeDataType>(inv_std_m(m_offset + inner_m));
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const ComputeDataType dy = ck_tile::type_convert<ComputeDataType>(dy_m_n(m_offset + inner_m, n));
|
||||
const ComputeDataType x = ck_tile::type_convert<ComputeDataType>(x_m_n(m_offset + inner_m, n));
|
||||
const ComputeDataType gamma = ck_tile::type_convert<ComputeDataType>(gamma_n(n));
|
||||
ds += dy * gamma * x;
|
||||
db += dy * gamma;
|
||||
}
|
||||
ComputeDataType b = (db * mean - ds) * inv_std * inv_std * inv_std / N;
|
||||
ComputeDataType c = -b * mean - db * inv_std / N;
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const ComputeDataType dy = ck_tile::type_convert<ComputeDataType>(dy_m_n(m_offset + inner_m, n));
|
||||
const ComputeDataType x = ck_tile::type_convert<ComputeDataType>(x_m_n(m_offset + inner_m, n));
|
||||
const ComputeDataType gamma = ck_tile::type_convert<ComputeDataType>(gamma_n(n));
|
||||
dx_m_n(m_offset + inner_m, n) = ck_tile::type_convert<XDataType>(dy * gamma * inv_std + b * x + c);
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f, PartM)(std::thread::hardware_concurrency());
|
||||
}
|
||||
} // namespace ck_tile
|
||||
@@ -10,4 +10,10 @@
|
||||
#include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_two_pass.hpp"
|
||||
#include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_traits.hpp"
|
||||
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
|
||||
|
||||
#include "ck_tile/ops/layernorm2d/kernel/layernorm2d_bwd_gamma_beta_kernel.hpp"
|
||||
#include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_default_policy.hpp"
|
||||
#include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_gamma_beta.hpp"
|
||||
#include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_problem.hpp"
|
||||
|
||||
#include "ck_tile/ops/common/tensor_layout.hpp"
|
||||
|
||||
@@ -0,0 +1,244 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/common.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
// host side args
|
||||
struct Layernorm2dBwdGammaBetaHostArgs
|
||||
{
|
||||
const void* p_x;
|
||||
const void* p_dY;
|
||||
const void* p_gamma;
|
||||
const void* p_mean;
|
||||
const void* p_invStd;
|
||||
|
||||
void* p_dGamma;
|
||||
void* p_dBeta;
|
||||
void* p_dX;
|
||||
|
||||
index_t m;
|
||||
index_t n;
|
||||
index_t stride; // row_stride
|
||||
};
|
||||
|
||||
// TODO: Extract some type to wrapper class
|
||||
template <typename Pipeline_>
|
||||
struct Layernorm2dBwdGammaBeta
|
||||
{
|
||||
using Pipeline = remove_cvref_t<Pipeline_>;
|
||||
using Problem = typename Pipeline::Problem;
|
||||
|
||||
using XDataType = remove_cvref_t<typename Problem::XDataType>;
|
||||
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>;
|
||||
|
||||
static constexpr index_t Block_M = Problem::BlockShape::Block_M;
|
||||
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
|
||||
static constexpr bool kPadM = false; // always no need to pad along M
|
||||
static constexpr bool kPadN = Problem::kPadN;
|
||||
|
||||
static constexpr index_t ThreadPerWarp_N = Problem::BlockShape::ThreadPerWarp_N;
|
||||
|
||||
static constexpr auto I0 = number<0>{};
|
||||
static constexpr auto I1 = number<1>{};
|
||||
|
||||
struct Kargs
|
||||
{
|
||||
const void* p_x;
|
||||
const void* p_dY;
|
||||
const void* p_gamma;
|
||||
const void* p_mean;
|
||||
const void* p_invStd;
|
||||
|
||||
void* p_dGamma;
|
||||
void* p_dBeta;
|
||||
void* p_dX;
|
||||
|
||||
index_t m;
|
||||
index_t n;
|
||||
index_t stride; // row_stride
|
||||
};
|
||||
using Hargs = Layernorm2dBwdGammaBetaHostArgs;
|
||||
|
||||
CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
|
||||
{
|
||||
return Kargs{hargs.p_x,
|
||||
hargs.p_dY,
|
||||
hargs.p_gamma,
|
||||
hargs.p_mean,
|
||||
hargs.p_invStd,
|
||||
hargs.p_dGamma,
|
||||
hargs.p_dBeta,
|
||||
hargs.p_dX,
|
||||
hargs.m,
|
||||
hargs.n,
|
||||
hargs.stride};
|
||||
}
|
||||
|
||||
CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
|
||||
{
|
||||
return (hargs.m + Block_M - 1) / Block_M;
|
||||
}
|
||||
|
||||
CK_TILE_HOST static constexpr auto BlockSize() { return Problem::BlockShape::BlockSize; }
|
||||
|
||||
// clang-format off
|
||||
template <typename T> struct t2s;
|
||||
template <> struct t2s<float> { static constexpr const char * name = "fp32"; };
|
||||
template <> struct t2s<ck_tile::fp16_t> { static constexpr const char * name = "fp16"; };
|
||||
template <> struct t2s<ck_tile::bf16_t> { static constexpr const char * name = "bf16"; };
|
||||
template <> struct t2s<ck_tile::fp8_t> { static constexpr const char * name = "fp8"; };
|
||||
template <> struct t2s<ck_tile::bf8_t> { static constexpr const char * name = "bf8"; };
|
||||
// clang-format on
|
||||
|
||||
// in byte
|
||||
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize() { return Pipeline::GetSmemSize(); }
|
||||
|
||||
CK_TILE_HOST static std::string GetName()
|
||||
{
|
||||
// clang-format off
|
||||
using S_ = typename Problem::BlockShape;
|
||||
auto surfix = [&] () {
|
||||
std::string n;
|
||||
if (kPadN) n += "_pn";
|
||||
return n; }();
|
||||
|
||||
#define _SS_ std::string
|
||||
#define _TS_ std::to_string
|
||||
return _SS_("layernorm2d_bwd_") + _SS_(t2s<XDataType>::name) + "_" +
|
||||
_TS_(S_::Block_M) + "x" + _TS_(S_::Block_N) + "_" + _TS_(S_::WarpPerBlock_M) + "x" + _TS_(S_::WarpPerBlock_N) + "_" +
|
||||
_TS_(S_::Warp_M) + "x" + _TS_(S_::Warp_N) + "_" + _TS_(S_::Vector_M) + "x" + _TS_(1) + "_" +
|
||||
_SS_(Pipeline::name) + surfix;
|
||||
#undef _SS_
|
||||
#undef _TS_
|
||||
// clang-format on
|
||||
}
|
||||
|
||||
CK_TILE_DEVICE void operator()(Kargs kargs) const
|
||||
{
|
||||
const auto block_id = get_block_id();
|
||||
const auto iM = block_id * Block_M;
|
||||
|
||||
const auto x_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<const XDataType*>(kargs.p_x),
|
||||
make_tuple(kargs.m, kargs.n),
|
||||
make_tuple(kargs.stride, 1));
|
||||
|
||||
// NOTE: we don't do any pad in this kernel for loading, assume that inside kernel will
|
||||
// check the max count dynamically
|
||||
const auto tmp2_ = pad_tensor_view(
|
||||
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<false, false>{});
|
||||
return make_tile_window(
|
||||
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
|
||||
}();
|
||||
|
||||
const auto dy_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<const YDataType*>(kargs.p_dY),
|
||||
make_tuple(kargs.m, kargs.n),
|
||||
make_tuple(kargs.stride, 1));
|
||||
|
||||
// NOTE: we don't do any pad in this kernel for loading, assume that inside kernel will
|
||||
// check the max count dynamically
|
||||
const auto tmp2_ = pad_tensor_view(
|
||||
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<false, false>{});
|
||||
return make_tile_window(
|
||||
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
|
||||
}();
|
||||
|
||||
const auto gamma_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<const MeanDataType*>(kargs.p_gamma),
|
||||
make_tuple(kargs.n),
|
||||
make_tuple(1));
|
||||
|
||||
const auto tmp2_ =
|
||||
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<false>{});
|
||||
|
||||
return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
|
||||
}();
|
||||
|
||||
const auto mean_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<const MeanDataType*>(kargs.p_mean),
|
||||
make_tuple(kargs.m),
|
||||
make_tuple(1));
|
||||
|
||||
const auto tmp2_ =
|
||||
pad_tensor_view(tmp_, make_tuple(number<Block_M>{}), sequence<false>{});
|
||||
|
||||
return make_tile_window(tmp2_, make_tuple(number<Block_M>{}), {iM});
|
||||
}();
|
||||
|
||||
const auto invstd_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<const MeanDataType*>(kargs.p_invStd),
|
||||
make_tuple(kargs.m),
|
||||
make_tuple(1));
|
||||
|
||||
const auto tmp2_ =
|
||||
pad_tensor_view(tmp_, make_tuple(number<Block_M>{}), sequence<false>{});
|
||||
|
||||
return make_tile_window(tmp2_, make_tuple(number<Block_M>{}), {iM});
|
||||
}();
|
||||
|
||||
auto dgamma_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<GammaDataType*>(kargs.p_dGamma),
|
||||
make_tuple(gridDim.x, kargs.n),
|
||||
make_tuple(kargs.n, 1));
|
||||
|
||||
const auto tmp2_ =
|
||||
pad_tensor_view(tmp_, make_tuple(number<1>{}, number<Block_N>{}), sequence<false, kPadN>{});
|
||||
|
||||
return make_tile_window(tmp2_, make_tuple(number<1>{}, number<Block_N>{}), {block_id, 0});
|
||||
}();
|
||||
|
||||
auto dbeta_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<BetaDataType*>(kargs.p_dBeta),
|
||||
make_tuple(gridDim.x, kargs.n),
|
||||
make_tuple(kargs.n, 1));
|
||||
|
||||
const auto tmp2_ =
|
||||
pad_tensor_view(tmp_, make_tuple(number<1>{}, number<Block_N>{}), sequence<false, kPadN>{});
|
||||
return make_tile_window(tmp2_, make_tuple(number<1>{}, number<Block_N>{}), {block_id, 0});
|
||||
}();
|
||||
|
||||
auto dx_window = [&]() {
|
||||
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
|
||||
static_cast<XDataType*>(kargs.p_dX),
|
||||
make_tuple(kargs.m, kargs.n),
|
||||
make_tuple(kargs.stride, 1));
|
||||
|
||||
const auto tmp2_ =
|
||||
pad_tensor_view(tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<false, false>{});
|
||||
return make_tile_window(tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
|
||||
}();
|
||||
|
||||
__shared__ char smem[GetSmemSize()];
|
||||
|
||||
Pipeline{}(x_window,
|
||||
dy_window,
|
||||
gamma_window,
|
||||
mean_window,
|
||||
invstd_window,
|
||||
dgamma_window,
|
||||
dbeta_window,
|
||||
dx_window,
|
||||
kargs.n,
|
||||
smem);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,79 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
struct Layernorm2dBwdGammaBetaPipelineDefaultPolicy
|
||||
{
|
||||
template <typename Problem>
|
||||
CK_TILE_DEVICE static constexpr auto MakeXBlockTileDistribution()
|
||||
{
|
||||
using S = typename Problem::BlockShape;
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<
|
||||
sequence<>,
|
||||
tuple<sequence<S::Repeat_M, S::WarpPerBlock_M, S::ThreadPerWarp_M>,
|
||||
sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N>>,
|
||||
tuple<sequence<1, 2>, sequence<1, 2>>,
|
||||
tuple<sequence<1, 1>, sequence<2, 2>>,
|
||||
sequence<1, 2>,
|
||||
sequence<0, 0>>{});
|
||||
}
|
||||
template <typename Problem>
|
||||
CK_TILE_DEVICE static constexpr auto MakeMeanBlockTileDistribution()
|
||||
{
|
||||
using S = typename Problem::BlockShape;
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<
|
||||
sequence<S::WarpPerBlock_N, S::ThreadPerWarp_N>,
|
||||
tuple<sequence<S::Repeat_M, S::WarpPerBlock_M, S::ThreadPerWarp_M>>,
|
||||
tuple<sequence<1, 0>, sequence<1, 0>>,
|
||||
tuple<sequence<1, 0>, sequence<2, 1>>,
|
||||
sequence<1>,
|
||||
sequence<0>>{});
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_DEVICE static constexpr auto MakeDGammaBetaBlockTileDistribution()
|
||||
{
|
||||
using S = typename Problem::BlockShape;
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<
|
||||
sequence<>,
|
||||
tuple<sequence<S::WarpPerBlock_M, S::ThreadPerWarp_M>,
|
||||
sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N>>,
|
||||
tuple<sequence<1, 2>, sequence<1, 2>>,
|
||||
tuple<sequence<0, 1>, sequence<1, 2>>,
|
||||
sequence<2>,
|
||||
sequence<0>>{});
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_DEVICE static constexpr auto MakeGammaBetaBlockTileDistribution()
|
||||
{
|
||||
using S = typename Problem::BlockShape;
|
||||
|
||||
return make_static_tile_distribution(
|
||||
tile_distribution_encoding<
|
||||
sequence<S::WarpPerBlock_M, S::ThreadPerWarp_M>,
|
||||
tuple<sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
|
||||
tuple<sequence<0, 1>, sequence<0, 1>>,
|
||||
tuple<sequence<0, 1>, sequence<1, 2>>,
|
||||
sequence<1, 1>,
|
||||
sequence<0, 3>>{});
|
||||
}
|
||||
|
||||
template <typename Problem>
|
||||
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
|
||||
{
|
||||
return 1;
|
||||
}
|
||||
};
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,132 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/ops/layernorm2d/pipeline/layernorm2d_bwd_pipeline_default_policy.hpp"
|
||||
#include <string>
|
||||
#include <type_traits>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename Problem_, typename Policy_ = Layernorm2dBwdGammaBetaPipelineDefaultPolicy>
|
||||
struct Layernorm2dBwdGammaBetaPipeline
|
||||
{
|
||||
using Problem = ck_tile::remove_cvref_t<Problem_>;
|
||||
using Policy = ck_tile::remove_cvref_t<Policy_>;
|
||||
|
||||
using XDataType = ck_tile::remove_cvref_t<typename Problem::XDataType>;
|
||||
using GammaDataType = ck_tile::remove_cvref_t<typename Problem::GammaDataType>;
|
||||
using BetaDataType = ck_tile::remove_cvref_t<typename Problem::BetaDataType>;
|
||||
using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
|
||||
using YDataType = ck_tile::remove_cvref_t<typename Problem::YDataType>;
|
||||
using MeanDataType = ck_tile::remove_cvref_t<typename Problem::MeanDataType>;
|
||||
using InvStdDataType = ck_tile::remove_cvref_t<typename Problem::InvStdDataType>;
|
||||
|
||||
static constexpr bool kPadM = false;
|
||||
static constexpr bool kPadN = Problem::kPadN;
|
||||
|
||||
static constexpr const char* name = []() {
|
||||
return "bwd_gamma_beta";
|
||||
}();
|
||||
|
||||
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
|
||||
{
|
||||
return Policy::template GetSmemSize<Problem>();
|
||||
}
|
||||
template <typename XWindow,
|
||||
typename GammaWindow,
|
||||
typename MeanWindow,
|
||||
typename InvStdWindow,
|
||||
typename DGammaWindow,
|
||||
typename DBetaWindow,
|
||||
typename DXWindow>
|
||||
CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
|
||||
const XWindow& dy_window_,
|
||||
const GammaWindow& gamma_window_,
|
||||
const MeanWindow& mean_window_,
|
||||
const InvStdWindow& inv_std_window_,
|
||||
DGammaWindow& dgamma_window_,
|
||||
DBetaWindow& dbeta_window_,
|
||||
DXWindow& dx_window_,
|
||||
ck_tile::index_t row_size,
|
||||
void* smem) const
|
||||
{
|
||||
(void)row_size;
|
||||
(void)smem;
|
||||
|
||||
auto gamma_beta_dist = Policy::template MakeGammaBetaBlockTileDistribution<Problem>();
|
||||
auto dgamma_beta_dist = Policy::template MakeDGammaBetaBlockTileDistribution<Problem>();
|
||||
auto mean_dist = Policy::template MakeMeanBlockTileDistribution<Problem>();
|
||||
auto x_dist = Policy::template MakeXBlockTileDistribution<Problem>();
|
||||
|
||||
const auto x_window = make_tile_window(x_window_, x_dist);
|
||||
const auto dy_window = make_tile_window(dy_window_, x_dist);
|
||||
const auto gamma_window = make_tile_window(gamma_window_, gamma_beta_dist); //TO CHECK
|
||||
const auto mean_window = make_tile_window(mean_window_, mean_dist);
|
||||
const auto inv_std_window = make_tile_window(inv_std_window_, mean_dist);
|
||||
const auto x_tile = load_tile(x_window);
|
||||
const auto dy_tile = load_tile(dy_window);
|
||||
const auto gamma_tile = load_tile(gamma_window);
|
||||
const auto mean_tile = load_tile(mean_window);
|
||||
const auto inv_std_tile = load_tile(inv_std_window);
|
||||
|
||||
auto dgamma_window = make_tile_window(dgamma_window_, dgamma_beta_dist);
|
||||
auto dbeta_window = make_tile_window(dbeta_window_, dgamma_beta_dist);
|
||||
auto dx_window = make_tile_window(dx_window_, x_dist);
|
||||
auto dgamma_tile = make_static_distributed_tensor<GammaDataType>(dgamma_beta_dist);
|
||||
auto dbeta_tile = make_static_distributed_tensor<BetaDataType>(dgamma_beta_dist);
|
||||
auto dx_tile = make_static_distributed_tensor<XDataType>(x_dist);
|
||||
auto dgamma = cast_tile<ComputeDataType>(dgamma_tile);
|
||||
auto dbeta = cast_tile<ComputeDataType>(dbeta_tile);
|
||||
auto dx = cast_tile<XDataType>(dx_tile);
|
||||
|
||||
(void)dx_window;
|
||||
(void)dx;
|
||||
(void)gamma_tile;
|
||||
|
||||
sweep_tile(x_tile, [&](auto idx) {
|
||||
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
|
||||
//constexpr auto j_idx = make_tuple(idx[number<1>{}]);
|
||||
constexpr auto gb_idx = make_tuple(number<0>{}, idx[number<1>{}]);
|
||||
// auto &gamma = gamma_tile(gb_idx);
|
||||
// auto &beta = beta_tile(gb_idx);
|
||||
const auto x = type_convert<ComputeDataType>(x_tile[idx]);
|
||||
const auto dy = type_convert<ComputeDataType>(dy_tile[idx]);
|
||||
const auto mean = type_convert<ComputeDataType>(mean_tile[i_idx]);
|
||||
const auto inv_std = type_convert<ComputeDataType>(inv_std_tile[i_idx]);
|
||||
// beta += type_convert<BetaDataType>(dy);
|
||||
// gamma += type_convert<GammaDataType>(dy * (x - mean) * inv_std);
|
||||
dbeta(gb_idx) += dy;
|
||||
dgamma(gb_idx) += dy * (x - mean) * inv_std;
|
||||
// index_t tid = (threadIdx.y * blockDim.x) + threadIdx.x;
|
||||
// if(blockIdx.x < 3 && blockIdx.y == 0 && tid < 3) {
|
||||
// printf("bid %d tid %d count %d gb %f %f\n",blockIdx.x, tid, count, type_convert<float>(g), type_convert<float>(b));
|
||||
// }
|
||||
});
|
||||
store_tile(dbeta_window, cast_tile<BetaDataType>(dbeta));
|
||||
store_tile(dgamma_window, cast_tile<GammaDataType>(dgamma));
|
||||
// store_tile(gamma_window, gamma_tile);
|
||||
// store_tile(beta_window, beta_tile);
|
||||
|
||||
|
||||
// auto ds = cast_tile<ComputeDataType>(mean_tile);
|
||||
// auto db = cast_tile<ComputeDataType>(mean_tile);
|
||||
// //calculate dx
|
||||
// sweep_tile(x_tile, [&](auto idx)) {
|
||||
// constexpr auto i_idx = make_tuple(idx[number<0>{}]);
|
||||
// constexpr auto j_idx = make_tuple(idx[number<1>{}]);
|
||||
|
||||
// const auto x = type_convert<ComputeDataType>(x_tile[idx]);
|
||||
// const auto dy = type_convert<ComputeDataType>(dy_tile[idx]);
|
||||
// const auto gamma = type_convert<ComputeDataType>(gamma_tile[j_idx]);
|
||||
// // const auto mean = type_convert<ComputeDataType>(mean_tile[i_idx]);
|
||||
// // const auto inv_std = type_convert<ComputeDataType>(inv_std_tile[i_idx]);
|
||||
// ds[i_idx] += dy * gamma * x;
|
||||
// db[i_idx] += dy * gamma;
|
||||
// }
|
||||
|
||||
}
|
||||
};
|
||||
} // namespace ck_tile
|
||||
@@ -0,0 +1,33 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core/utility/type_traits.hpp"
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename XDataType_,
|
||||
typename GammaDataType_,
|
||||
typename BetaDataType_,
|
||||
typename ComputeDataType_,
|
||||
typename YDataType_,
|
||||
typename MeanDataType_,
|
||||
typename InvStdDataType_,
|
||||
typename BlockShape_,
|
||||
bool kPadN_>
|
||||
struct Layernorm2dBwdGammaBetaPipelineProblem
|
||||
{
|
||||
using XDataType = remove_cvref_t<XDataType_>;
|
||||
using GammaDataType = remove_cvref_t<GammaDataType_>;
|
||||
using BetaDataType = remove_cvref_t<BetaDataType_>;
|
||||
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
|
||||
using YDataType = remove_cvref_t<YDataType_>;
|
||||
using MeanDataType = remove_cvref_t<MeanDataType_>;
|
||||
using InvStdDataType = remove_cvref_t<InvStdDataType_>;
|
||||
using BlockShape = remove_cvref_t<BlockShape_>;
|
||||
|
||||
static constexpr bool kPadN = kPadN_;
|
||||
};
|
||||
|
||||
} // namespace ck_tile
|
||||
Reference in New Issue
Block a user