Skip to content

GitLab

"include/vscode:/vscode.git/clone" did not exist on "26fa4782ca5c733e2dcde14e69bf8f3752429a2d"
Commit d6b0e59e authored by rocking's avatar rocking
Browse files

Add smoothquant op

parent f4228e0a
Show whitespace changes
Inline Side-by-side
Showing with 874 additions and 0 deletions
example/ck_tile/12_smoothquant/CMakeLists.txt 0 → 100644
View file @ d6b0e59e
set(EXAMPLE_SMOOTHQUANT "tile_example_smoothquant")
# 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 example ${EXAMPLE_SMOOTHQUANT}")
add_executable(${EXAMPLE_SMOOTHQUANT} EXCLUDE_FROM_ALL example_smoothquant.cpp)
target_include_directories(${EXAMPLE_SMOOTHQUANT} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
set(EXAMPLE_SMOOTHQUANT_COMPILE_OPTIONS)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND EXAMPLE_SMOOTHQUANT_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
target_compile_options(${EXAMPLE_SMOOTHQUANT} PRIVATE ${EXAMPLE_SMOOTHQUANT_COMPILE_OPTIONS})
# TODO: we have to turn off this global prop, otherwise the progress bar generated
# by cmake will print too many files, execvp: /bin/sh: Argument list too long
# however, this property may affect global
# TODO: consider codegen a makefile by us
set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)
\ No newline at end of file
example/ck_tile/12_smoothquant/example_smoothquant.cpp 0 → 100644
View file @ d6b0e59e
#include "ck_tile/host.hpp"
#include "ck_tile/core.hpp"
#include "ck_tile/host/kernel_launch.hpp"
#include "ck_tile/ops/smoothquant.hpp"
#include <cstring>
// different threshold for different dtype
template <typename DataType>
auto get_elimit()
{
double rtol = 1e-4;
double atol = 1e-4;
return ck_tile::make_tuple(rtol, atol);
}
template <>
auto get_elimit<ck_tile::bf16_t>()
{
double rtol = 1e-2;
double atol = 1e-2;
return ck_tile::make_tuple(rtol, atol);
}
template <>
auto get_elimit<ck_tile::int8_t>()
{
// due to rounding, int8 quantization might have 1 abs error
double rtol = 1;
double atol = 1;
return ck_tile::make_tuple(rtol, atol);
}
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
arg_parser.insert("m", "3328", "m dimension")
.insert("n", "4096", "n dimension")
.insert("stride", "-1", "stride per row, if -1 then equal to n")
.insert("e", "1e-5", "epsilon")
.insert("v", "1", "cpu validation or not")
.insert("prec", "fp16", "precision")
.insert("warmup", "0", "cold iter")
.insert("repeat", "1", "hot iter");
bool result = arg_parser.parse(argc, argv);
return std::make_tuple(result, arg_parser);
}
template <typename DataType>
bool run(const ck_tile::ArgParser& arg_parser)
{
ck_tile::index_t m = arg_parser.get_int("m");
ck_tile::index_t n = arg_parser.get_int("n");
ck_tile::index_t stride = arg_parser.get_int("stride");
if(stride < 0)
stride = n;
std::string data_type = arg_parser.get_str("prec");
int do_validation = arg_parser.get_int("v");
int warmup = arg_parser.get_int("warmup");
int repeat = arg_parser.get_int("repeat");
assert(stride >= n);
using XDataType = DataType;
using XScaleDataType = DataType;
using YScaleDataType = DataType;
using QYDataType = ck_tile::int8_t;
using ComputeDataType = float;
// host verify
ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1});
ck_tile::HostTensor<XScaleDataType> xscale_host({n});
ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1});
ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1});
ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
ck_tile::FillUniformDistribution<XScaleDataType>{1e-3, .5f}(xscale_host);
ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem xscale_buf(xscale_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
x_buf.ToDevice(x_host.data());
xscale_buf.ToDevice(xscale_host.data());
constexpr bool kTwoPass = false;
using BlockWarps = ck_tile::sequence<2, 2>;
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::SmoothquantShape<BlockTile, BlockWarps, WarpTile, Vector>;
using Problem = ck_tile::SmoothquantPipelineProblem<XDataType,
XScaleDataType,
ComputeDataType,
YScaleDataType,
QYDataType,
Shape,
true,
kTwoPass>;
using OnePassPipeline = ck_tile::SmoothquantPipelineOnePass<Problem>;
using TwoPassPipeline = ck_tile::SmoothquantPipelineTwoPass<Problem>;
using Pipeline = std::conditional_t<kTwoPass, TwoPassPipeline, OnePassPipeline>;
using Kernel = ck_tile::Smoothquant<Pipeline>;
ck_tile::SmoothquantHostArgs args{x_buf.GetDeviceBuffer(),
xscale_buf.GetDeviceBuffer(),
yscale_buf.GetDeviceBuffer(),
qy_buf.GetDeviceBuffer(),
m,
n,
stride};
auto kargs = Kernel::MakeKargs(args);
const dim3 grids = Kernel::GridSize(args);
constexpr dim3 blocks = Kernel::BlockSize();
constexpr ck_tile::index_t kBlockPerCu = 1;
auto s = ck_tile::stream_config{nullptr, true, 0, warmup, repeat};
ck_tile::launch_kernel(
s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
bool pass = true;
if(do_validation)
{
using YDataType = ComputeDataType;
ck_tile::HostTensor<ComputeDataType> y_host({m, n}, {stride, 1});
// smooth outlier
{
auto f = [&](auto n_) {
auto v_xscale = ck_tile::type_convert<ComputeDataType>(xscale_host(n_));
for(int m_ = 0; m_ < m; ++m_)
{
auto v_x = ck_tile::type_convert<ComputeDataType>(x_host(m_, n_));
y_host(m_, n_) = v_x * v_xscale;
}
};
ck_tile::make_ParallelTensorFunctor(f, xscale_host.get_element_space_size())(
std::thread::hardware_concurrency());
}
// yscale
{
ck_tile::HostTensor<YDataType> y_rowwise_amax_host({m});
using ReduceAmax = ck_tile::ReduceOp::AbsMax;
ck_tile::reference_reduce<ComputeDataType, ComputeDataType, YDataType>(
y_host, y_rowwise_amax_host, ReduceAmax{});
auto op = [](const auto& v0) {
return v0 /
ck_tile::type_convert<ComputeDataType>(ck_tile::numeric<QYDataType>::max());
};
ck_tile::reference_unary_elementwise<YDataType, YScaleDataType, ComputeDataType>(
y_rowwise_amax_host, yscale_host_ref, op);
yscale_buf.FromDevice(yscale_host_dev.mData.data());
auto [rtol, atol] = get_elimit<YScaleDataType>();
pass &= ck_tile::check_err(yscale_host_dev,
yscale_host_ref,
std::string("yscale Error: Incorrect results!"),
rtol,
atol);
}
// rowwise quantization
{
ck_tile::reference_rowwise_quantization2d<YDataType, YScaleDataType, QYDataType>(
y_host, yscale_host_ref, qy_host_ref);
qy_buf.FromDevice(qy_host_dev.data());
auto [rtol, atol] = get_elimit<QYDataType>();
if(stride == n)
{
pass = ck_tile::check_err(qy_host_dev,
qy_host_ref,
std::string("qy Error: Incorrect results!"),
rtol,
atol);
}
else
{
for(int i_r = 0; i_r < m; i_r++)
{
std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
qy_host_dev.begin() + i_r * stride + n);
std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
qy_host_ref.begin() + i_r * stride + n);
pass &= ck_tile::check_err(qy_host_dev_row,
qy_host_ref_row,
std::string("qy[") + std::to_string(i_r) +
std::string("] Error: Incorrect results!"),
rtol,
atol);
}
}
}
std::cout << "[" << data_type << "]"
<< " m:" << m << ", n:" << n << ", stride:" << stride
<< ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
}
return pass;
}
int main(int argc, char* argv[])
{
auto [result, arg_parser] = create_args(argc, argv);
if(!result)
return -1;
const std::string data_type = arg_parser.get_str("prec");
if(data_type == "fp16")
{
return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
}
return -3;
}
example/ck_tile/CMakeLists.txt
View file @ d6b0e59e
...@@ -11,3 +11,4 @@ add_subdirectory(06_permute) ...@@ -11,3 +11,4 @@ add_subdirectory(06_permute)
add_subdirectory(09_topk_softmax) add_subdirectory(09_topk_softmax)
add_subdirectory(10_rmsnorm2d) add_subdirectory(10_rmsnorm2d)
add_subdirectory(11_add_rmsnorm2d_rdquant) add_subdirectory(11_add_rmsnorm2d_rdquant)
add_subdirectory(12_smoothquant)
include/ck_tile/ops/smoothquant.hpp 0 → 100644
View file @ d6b0e59e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/ops/smoothquant/kernel/smoothquant_kernel.hpp"
#include "ck_tile/ops/smoothquant/kernel/smoothquant_shape.hpp"
#include "ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_default_policy.hpp"
#include "ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_one_pass.hpp"
#include "ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp"
#include "ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_problem.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
include/ck_tile/ops/smoothquant/kernel/smoothquant_kernel.hpp 0 → 100644
View file @ d6b0e59e
// 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 SmoothquantHostArgs
{
const void* p_x;
const void* p_xscale;
void* p_yscale;
void* p_qy;
index_t m;
index_t n;
index_t stride; // row_stride
};
// TODO: Extract some type to wrapper class
template <typename Pipeline_>
struct Smoothquant
{
using Pipeline = remove_cvref_t<Pipeline_>;
using Problem = typename Pipeline::Problem;
using XDataType = remove_cvref_t<typename Problem::XDataType>;
using XScaleDataType = remove_cvref_t<typename Problem::XScaleDataType>;
using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
using YScaleDataType = remove_cvref_t<typename Problem::YScaleDataType>;
using QYDataType = remove_cvref_t<typename Problem::QYDataType>;
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 bool kTwoPass = Problem::kTwoPass;
static constexpr index_t ThreadPerWarp_N = Problem::BlockShape::ThreadPerWarp_N;
static constexpr index_t Vector_N = Problem::BlockShape::Vector_N;
static constexpr index_t Repeat_N = Problem::BlockShape::Repeat_N;
static constexpr auto I0 = number<0>{};
static constexpr auto I1 = number<1>{};
struct Kargs
{
const void* p_x;
const void* p_xscale;
void* p_yscale;
void* p_qy;
index_t m;
index_t n;
index_t stride; // row_stride
};
using Hargs = SmoothquantHostArgs;
CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
{
return Kargs{
hargs.p_x, hargs.p_xscale, hargs.p_yscale, hargs.p_qy, 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";
if (kTwoPass) n += "_2p";
return n; }();
#define _SS_ std::string
#define _TS_ std::to_string
return _SS_("smoothquant_fwd_") + _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_(S_::Vector_N) + "_" +
_SS_(Pipeline::name) + surfix;
#undef _SS_
#undef _TS_
// clang-format on
}
CK_TILE_DEVICE void operator()(Kargs kargs) const
{
const auto iM = get_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),
number<Vector_N>{},
number<1>{});
const auto tmp2_ = pad_tensor_view(
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
return make_tile_window(
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
}();
const auto xscale_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XScaleDataType*>(kargs.p_xscale),
make_tuple(kargs.n),
make_tuple(1),
number<Vector_N>{},
number<1>{});
const auto tmp2_ =
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<kPadN>{});
return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
}();
auto yscale_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<YScaleDataType*>(kargs.p_yscale),
make_tuple(kargs.m),
make_tuple(1),
number<1>{});
const auto tmp2_ =
pad_tensor_view(tmp_, make_tuple(number<Block_M>{}), sequence<kPadM>{});
return make_tile_window(tmp2_, make_tuple(number<Block_M>{}), {iM});
}();
auto qy_window = [&]() {
auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<QYDataType*>(kargs.p_qy),
make_tuple(kargs.m, kargs.n),
make_tuple(kargs.stride, 1),
number<Vector_N>{},
number<1>{});
auto tmp2_ = pad_tensor_view(
tmp_, make_tuple(number<Block_M>{}, number<Block_N>{}), sequence<kPadM, kPadN>{});
return make_tile_window(
tmp2_, make_tuple(number<Block_M>{}, number<Block_N>{}), {iM, 0});
}();
__shared__ char smem[GetSmemSize()];
Pipeline{}(x_window, xscale_window, yscale_window, qy_window, kargs.n, smem);
}
};
} // namespace ck_tile
include/ck_tile/ops/smoothquant/kernel/smoothquant_shape.hpp 0 → 100644
View file @ d6b0e59e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/core.hpp"
namespace ck_tile {
/*
// clang-format off
4-level descriptor: BlockTile-> WarpPerBlock-> WarpTile-> Vector
Block_N (Warp_N * WarpPerBlock_N * Repeat_N )
+<----------------------< Repeat_N(2)>--------------------->+
| |
+<-- <WarpPerBlock_N(2)> -->+
Warp_N
+--------------+--------------+--------------+--------------+----+----------------+
Warp_M | wrap_0 | wrap_1 | | ^ ^
+--------------+--------------+ | <WarpPerBlock_M(2)> |
| wrap_2 | wrap_3 | | v
+--------------+--------------+--------------+--------------+----+ Block_M
| | |
+ + |
| | | v
+--------------+--------------+--------------+--------------+ +
each Warp-tile (e.g 16 thrd per row)
Vector_N (contiguous pixels each thrd holds along N, or vector size)
+-----------+-----------+-----------+-----------+-----------+
| thrd_0 | thrd_1 | thrd_2 | thrd_3 | ... Vector_M
+-----------+-----------+-----------+-----------+-----------+
| thrd_16 | thrd_17 | thrd_18 | thrd_19 | ...
+-----------+-----------+-----------+-----------+-----------+
// clang-format on
*/
template <typename BlockTile_, // block size, seq<M, N>
typename WarpPerBlock_, // num warps along seq<M, N>
typename WarpTile_, // warp size, seq<M, N>
typename Vector_, // contiguous pixels(vector size) along seq<M, N>
index_t BlockSize_ =
warpSize * reduce_on_sequence(WarpPerBlock_{}, multiplies{}, number<1>{})>
struct SmoothquantShape
{
// block size
static constexpr index_t Block_M = BlockTile_::at(number<0>{});
static constexpr index_t Block_N = BlockTile_::at(number<1>{});
// num warps along seq<M, N>, within each block
static constexpr index_t WarpPerBlock_M = WarpPerBlock_::at(number<0>{});
static constexpr index_t WarpPerBlock_N = WarpPerBlock_::at(number<1>{});
// warp size
static constexpr index_t Warp_M = WarpTile_::at(number<0>{});
static constexpr index_t Warp_N = WarpTile_::at(number<1>{});
static_assert(Block_M % (WarpPerBlock_M * Warp_M) == 0);
static_assert(Block_N % (WarpPerBlock_N * Warp_N) == 0);
// repeat of each thread along seq<M, N>
static constexpr index_t Repeat_M = Block_M / (WarpPerBlock_M * Warp_M);
static constexpr index_t Repeat_N = Block_N / (WarpPerBlock_N * Warp_N);
// vector size along seq<M, N>
static constexpr index_t Vector_M = Vector_::at(number<0>{});
static constexpr index_t Vector_N = Vector_::at(number<1>{});
static_assert(Warp_M % Vector_M == 0);
static_assert(Warp_N % Vector_N == 0);
// num of threads along seq<M, N>, within each warp
static constexpr index_t ThreadPerWarp_M = Warp_M / Vector_M;
static constexpr index_t ThreadPerWarp_N = Warp_N / Vector_N;
static constexpr index_t BlockSize = BlockSize_;
};
} // namespace ck_tile
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_default_policy.hpp 0 → 100644
View file @ d6b0e59e
// 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/reduce/block/block_reduce2d_problem.hpp"
#include "ck_tile/ops/reduce/block/block_reduce2d.hpp"
namespace ck_tile {
struct SmoothquantPipelineDefaultPolicy
{
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, S::Vector_M>,
sequence<S::Repeat_N, S::WarpPerBlock_N, S::ThreadPerWarp_N, S::Vector_N>>,
tuple<sequence<1, 2>, sequence<1, 2>>,
tuple<sequence<1, 1>, sequence<2, 2>>,
sequence<1, 1, 2, 2>,
sequence<0, 3, 0, 3>>{});
}
template <typename Problem>
CK_TILE_DEVICE static constexpr auto MakeXScaleBlockTileDistribution()
{
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 auto GetBlockReduce2d()
{
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
return BlockReduce2d<P_>{};
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dSync()
{
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
return BlockReduce2dSync<P_>{};
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr auto GetBlockReduce2dCrossWarpSync()
{
using P_ = BlockReduce2dProblem<typename Problem::ComputeDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
return BlockReduce2dCrossWarpSync<P_>{};
}
template <typename Problem>
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
if constexpr(Problem::kNeedCrossWarpSync)
{
using P_ = BlockReduce2dProblem<typename Problem::XDataType,
typename Problem::ComputeDataType,
typename Problem::BlockShape>;
using block_reduce2d = BlockReduce2d<P_>;
using x_block_tile =
decltype(make_static_distributed_tensor<typename Problem::XDataType>(
MakeXBlockTileDistribution<Problem>()));
using y_block_tile = decltype(block_reduce2d::template MakeYBlockTile<x_block_tile>());
return GetBlockReduce2dCrossWarpSync<Problem>().template GetSmemSize<y_block_tile>();
}
else
{
return 1; // zero size arrays are an extension
}
}
};
} // namespace ck_tile
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_one_pass.hpp 0 → 100644
View file @ d6b0e59e
// 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/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
template <typename Problem_, typename Policy_ = SmoothquantPipelineDefaultPolicy>
struct SmoothquantPipelineOnePass
{
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 XScaleDataType = ck_tile::remove_cvref_t<typename Problem::XScaleDataType>;
using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
using QYDataType = ck_tile::remove_cvref_t<typename Problem::QYDataType>;
using YScaleDataType = ck_tile::remove_cvref_t<typename Problem::YScaleDataType>;
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockSmoothquantProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
return "bpr_op"; // block per row
else
return "wpr_op"; // warp per row
}();
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
}
template <typename XWindow, typename XScaleWindow, typename QYWindow, typename YScaleWindow>
CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
const XScaleWindow& xscale_window_,
YScaleWindow& yscale_window,
QYWindow& qy_window,
ck_tile::index_t,
void* smem) const
{
auto x_window =
make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto xscale_window = make_tile_window(
xscale_window_, Policy::template MakeXScaleBlockTileDistribution<Problem>());
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
auto block_reduce2d_cross_warp_sync =
Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
const auto x = load_tile(x_window);
const auto xscale = load_tile(xscale_window);
auto y = tile_elementwise_in(
[&](const auto& a, const auto& b) {
return type_convert<ComputeDataType>(a) * type_convert<ComputeDataType>(b);
},
x,
xscale);
// compute absmax, cross-lane->cross-warp
auto absmax = block_reduce2d(
y, reduce_absmax_func.GetIdentityValue<ComputeDataType>(), reduce_absmax_func);
block_reduce2d_sync(absmax, reduce_max_func);
block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
// ex: yscale = absmax / 127 if int8
auto yscale = tile_elementwise_in(
[&](const auto& v_) {
return v_ / type_convert<ComputeDataType>(numeric<QYDataType>::max());
},
absmax);
store_tile(yscale_window, cast_tile<YScaleDataType>(yscale));
// quantize y to qy
auto qy = make_static_distributed_tensor<QYDataType>(y.get_tile_distribution());
sweep_tile(qy, [&](auto idx) {
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
auto qy_ = y[idx] / yscale[i_idx];
qy(idx) = saturates<QYDataType>{}(qy_);
});
store_tile(qy_window, qy);
}
};
} // namespace ck_tile
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_problem.hpp 0 → 100644
View file @ d6b0e59e
// 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 {
// Y = X * XScale, QY = RowwiseDynamicQuant(Y) = SaturateCast(Y / YScale)
template <typename XDataType_,
typename XScaleDataType_,
typename ComputeDataType_,
typename YScaleDataType_,
typename QYDataType_,
typename BlockShape_,
bool kPadN_,
bool kTwoPass_>
struct SmoothquantPipelineProblem
{
using XDataType = remove_cvref_t<XDataType_>;
using XScaleDataType = remove_cvref_t<XScaleDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
using YScaleDataType = remove_cvref_t<YScaleDataType_>;
using QYDataType = remove_cvref_t<QYDataType_>;
using BlockShape = remove_cvref_t<BlockShape_>;
static constexpr bool kNeedCrossLaneSync = BlockShape::ThreadPerWarp_N > 1;
static constexpr bool kNeedCrossWarpSync = BlockShape::WarpPerBlock_N > 1;
static constexpr bool kPadN = kPadN_;
static constexpr bool kTwoPass = kTwoPass_;
};
} // namespace ck_tile
include/ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp 0 → 100644
View file @ d6b0e59e
// 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/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
#include <string>
#include <type_traits>
namespace ck_tile {
template <typename Problem_, typename Policy_ = SmoothquantPipelineDefaultPolicy>
struct SmoothquantPipelineTwoPass
{
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 XScaleDataType = ck_tile::remove_cvref_t<typename Problem::XScaleDataType>;
using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
using QYDataType = ck_tile::remove_cvref_t<typename Problem::QYDataType>;
using YScaleDataType = ck_tile::remove_cvref_t<typename Problem::YScaleDataType>;
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockSmoothquantProblem::kPadM
static constexpr bool kPadN = Problem::kPadN;
static constexpr const char* name = []() {
if constexpr(kNeedCrossWarpSync)
return "bpr_tp"; // block per row
else
return "wpr_tp"; // warp per row
}();
CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
{
return Policy::template GetSmemSize<Problem>();
}
template <typename XWindow, typename XScaleWindow, typename QYWindow, typename YScaleWindow>
CK_TILE_DEVICE auto operator()(const XWindow& x_window_,
const XScaleWindow& xscale_window_,
YScaleWindow& yscale_window,
QYWindow& qy_window,
ck_tile::index_t row_size,
void* smem) const
{
auto x_window =
make_tile_window(x_window_, Policy::template MakeXBlockTileDistribution<Problem>());
auto xscale_window = make_tile_window(
xscale_window_, Policy::template MakeXScaleBlockTileDistribution<Problem>());
static constexpr index_t Block_N = Problem::BlockShape::Block_N;
index_t num_n_tile_iteration =
__builtin_amdgcn_readfirstlane(integer_divide_ceil(row_size, Block_N));
auto reduce_absmax_func = ReduceOp::AbsMax{};
auto reduce_max_func = ReduceOp::Max{};
auto block_reduce2d = Policy::template GetBlockReduce2d<Problem>();
auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
auto block_reduce2d_cross_warp_sync =
Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
using XTensorType = decltype(cast_tile<ComputeDataType>(load_tile(x_window)));
auto absmax = block_reduce2d.template MakeYBlockTile<XTensorType>();
set_tile(absmax, reduce_absmax_func.GetIdentityValue<ComputeDataType>());
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
const auto x = load_tile(x_window);
const auto xscale = load_tile(xscale_window);
const auto y = tile_elementwise_in(
[&](const auto& a, const auto& b) {
return type_convert<ComputeDataType>(a) * type_convert<ComputeDataType>(b);
},
x,
xscale);
block_reduce2d(y, absmax, reduce_absmax_func);
move_tile_window(x_window, {0, Block_N});
move_tile_window(xscale_window, {Block_N});
}
// compute absmax, cross-lane->cross-warp
block_reduce2d_sync(absmax, reduce_max_func);
block_reduce2d_cross_warp_sync(absmax, smem, reduce_max_func);
// ex: yscale = absmax / 127 if int8
auto yscale = tile_elementwise_in(
[&](const auto& v_) {
return v_ / type_convert<ComputeDataType>(numeric<QYDataType>::max());
},
absmax);
store_tile(yscale_window, cast_tile<YScaleDataType>(yscale));
// reverse read x to reuse cache
ck_tile::index_t stride_to_right_most_window =
row_size % Block_N == 0 ? row_size - Block_N : row_size - row_size % Block_N;
move_tile_window(x_window, {0, -Block_N});
move_tile_window(xscale_window, {-Block_N});
move_tile_window(qy_window, {0, stride_to_right_most_window});
// recompute y and quantize y to qy
for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
{
const auto x = load_tile(x_window);
const auto xscale = load_tile(xscale_window);
const auto y = tile_elementwise_in(
[&](const auto& a, const auto& b) {
return type_convert<ComputeDataType>(a) * type_convert<ComputeDataType>(b);
},
x,
xscale);
auto qy = make_static_distributed_tensor<QYDataType>(y.get_tile_distribution());
sweep_tile(qy, [&](auto idx) {
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
auto qy_ = y[idx] / yscale[i_idx];
qy(idx) = saturates<QYDataType>{}(qy_);
});
store_tile(qy_window, qy);
move_tile_window(x_window, {0, -Block_N});
move_tile_window(xscale_window, {0, -Block_N});
move_tile_window(qy_window, {0, -Block_N});
}
}
};
} // namespace ck_tile
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