Skip to content

GitLab

Unverified Commit 9533a172 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge branch 'develop' into codegen-enable-hiprtc

parents c2cf0733 50ee4267
Hide whitespace changes
Inline Side-by-side
Showing with 1132 additions and 131 deletions
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_two_pass.hpp
View file @ 9533a172
......@@ -35,6 +35,7 @@ struct Layernorm2dFwdPipelineTwoPass
static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
static constexpr bool kPadM = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
static constexpr bool kPadN = Problem::Traits::kPadN;
static constexpr bool kFastFDiv = Problem::Traits::kFastFDiv;
static constexpr auto kFusedAdd = Problem::Traits::kFusedAdd;
static constexpr auto kFusedQuant = Problem::Traits::kFusedQuant;
......@@ -106,7 +107,7 @@ struct Layernorm2dFwdPipelineTwoPass
auto block_welford_cross_warp_sync =
Policy::template GetBlockWelfordCrossWarpSync<Problem>();
using XTensorType = decltype(load_tile(x_window));
using XTensorType = decltype(cast_tile<ComputeDataType>(load_tile(x_window)));
auto mean = block_welford.template MakeMeanVarBlockTile<XTensorType>();
auto var = block_welford.template MakeMeanVarBlockTile<XTensorType>();
......@@ -117,34 +118,42 @@ struct Layernorm2dFwdPipelineTwoPass
move_tile_window(x_window, {0, Block_N});
move_tile_window(x_residual_window, {0, Block_N});
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
// compute x = x_resi + x
x(idx) = type_convert<YResidualDataType>(x_resi(idx)) +
type_convert<YResidualDataType>(x(idx));
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE)
{
store_tile(y_residual_window, x);
store_tile(y_residual_window, cast_tile<YResidualDataType>(acc));
move_tile_window(y_residual_window, {0, Block_N});
}
}
block_welford(x, mean, var, cur_count, max_count);
block_welford(acc, mean, var, cur_count, max_count);
}
block_welford_sync(mean, var, cur_count);
block_welford_cross_warp_sync(mean, var, cur_count, smem);
block_tile_welford_post_scale_var(var, cur_count);
block_tile_welford_post_scale_var(var, cur_count, constant<kFastFDiv>{});
// compute inv-std
auto inv_std = tile_elementwise_in(
[&](const auto& v_) {
return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ + epsilon));
if(kFastFDiv && std::is_same_v<ComputeDataType, float>)
{
return type_convert<ComputeDataType>(1.0f) *
__builtin_amdgcn_rcpf(sqrt(v_ + epsilon));
}
else
{
return type_convert<ComputeDataType>(1.0f) / sqrt(v_ + epsilon);
}
},
var);
if constexpr(kSaveMean)
store_tile(mean_window, cast_tile<MeanDataType>(mean));
if constexpr(kSaveInvStd)
......@@ -165,20 +174,21 @@ struct Layernorm2dFwdPipelineTwoPass
{
auto x = load_tile(x_window);
auto x_resi = load_tile(x_residual_window);
auto acc = cast_tile<ComputeDataType>(x);
if constexpr(kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD_STORE ||
kFusedAdd == Layernorm2dFusedAddEnum::PRE_ADD)
{
sweep_tile(x_resi, [&](auto idx) {
// compute x = x_resi + x
x(idx) = type_convert<YResidualDataType>(x_resi(idx)) +
type_convert<YResidualDataType>(x(idx));
acc(idx) = type_convert<ComputeDataType>(x_resi(idx)) + acc(idx);
});
}
// load gamma/beta (TODO: support no gamma/beta?)
const auto gamma = load_tile(gamma_window);
const auto beta = load_tile(beta_window);
auto ln = make_static_distributed_tensor<ComputeDataType>(x.get_tile_distribution());
auto ln = make_static_distributed_tensor<ComputeDataType>(acc.get_tile_distribution());
sweep_tile(ln, [&, mean_ = mean](auto idx) {
constexpr auto i_idx = make_tuple(idx[number<0>{}]);
......@@ -187,8 +197,7 @@ struct Layernorm2dFwdPipelineTwoPass
const auto gamma_ = type_convert<ComputeDataType>(gamma[j_idx]);
const auto beta_ = type_convert<ComputeDataType>(beta[j_idx]);
const auto x_ = type_convert<ComputeDataType>(x[idx]);
auto ln_ = (x_ - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
auto ln_ = (acc(idx) - mean_[i_idx]) * inv_std[i_idx] * gamma_ + beta_;
ln(idx) = ln_;
});
......
include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_traits.hpp
View file @ 9533a172
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -39,6 +39,7 @@ template<> struct Layernorm2dFusedQuantEnumName<Layernorm2dFusedQuantEnum::SMOOT
template <bool kPadN_,
bool kSaveMeanInvStd_,
bool kFastFDiv_,
bool kTwoPass_,
Layernorm2dFusedAddEnum kFusedAdd_,
Layernorm2dFusedQuantEnum kFusedQuant_>
......@@ -46,6 +47,7 @@ struct Layernorm2dFwdTraits
{
static constexpr bool kPadN = kPadN_;
static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
static constexpr bool kFastFDiv = kFastFDiv_;
static constexpr bool kTwoPass = kTwoPass_;
static constexpr Layernorm2dFusedAddEnum kFusedAdd = kFusedAdd_;
static constexpr Layernorm2dFusedQuantEnum kFusedQuant = kFusedQuant_;
......
include/ck_tile/ops/reduce/block/block_reduce2d.hpp
View file @ 9533a172
......@@ -29,7 +29,8 @@ struct BlockReduce2d
sweep_tile<XDistributedTensor_>(
[&](auto... idx_) {
constexpr auto idx_0 = make_tuple(make_tuple(idx_[number<0>{}]...)[number<0>{}]);
y_tensor(idx_0) = reduce_func(y_tensor(idx_0), x_tensor[idx_]...);
y_tensor(idx_0) = reduce_func(
y_tensor(idx_0), ck_tile::type_convert<ComputeDataType>(x_tensor[idx_])...);
},
ReducePacksPerXDim{});
#if 0
......
include/ck_tile/ops/rmsnorm2d.hpp
View file @ 9533a172
......@@ -4,7 +4,6 @@
#pragma once
#include "ck_tile/ops/rmsnorm2d/kernel/rmsnorm2d_fwd_kernel.hpp"
#include "ck_tile/ops/rmsnorm2d/kernel/rmsnorm2d_fwd_shape.hpp"
#include "ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp"
#include "ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_one_pass.hpp"
#include "ck_tile/ops/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_problem.hpp"
......
include/ck_tile/ops/rmsnorm2d/kernel/rmsnorm2d_fwd_kernel.hpp
View file @ 9533a172
......@@ -11,11 +11,11 @@ namespace ck_tile {
// host side args
struct Rmsnorm2dFwdHostArgs
{
const void* p_x;
const void* p_gamma;
const void* p_x; // [m ,n], input, fp16/bf16
const void* p_gamma; // [1, n], gamma, prec same as input
void* p_y;
void* p_invRms;
void* p_y; // [m, n], output, fp16/bf16
void* p_invRms; // [m, 1], output inv-rms, prec same as input, nullptr if not used
float epsilon;
......@@ -83,7 +83,7 @@ struct Rmsnorm2dFwd
CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
{
return (hargs.m + Block_M - 1) / Block_M;
return dim3(integer_divide_ceil(hargs.m, Block_M));
}
CK_TILE_HOST static constexpr auto BlockSize() { return Problem::BlockShape::BlockSize; }
......@@ -149,7 +149,7 @@ struct Rmsnorm2dFwd
number<1>{});
const auto tmp2_ =
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<kPadM>{});
pad_tensor_view(tmp_, make_tuple(number<Block_N>{}), sequence<kPadN>{});
return make_tile_window(tmp2_, make_tuple(number<Block_N>{}), {0});
}();
......
include/ck_tile/ops/rmsnorm2d/kernel/rmsnorm2d_fwd_shape.hpp deleted 100644 → 0
View file @ c2cf0733
// 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 {
/*
// 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 Rmsnorm2dShape
{
// 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/rmsnorm2d/pipeline/rmsnorm2d_fwd_pipeline_default_policy.hpp
View file @ 9533a172
......@@ -26,6 +26,7 @@ struct Rmsnorm2dFwdPipelineDefaultPolicy
sequence<1, 1, 2, 2>,
sequence<0, 3, 0, 3>>{});
}
template <typename Problem>
CK_TILE_DEVICE static constexpr auto MakeGammaBlockTileDistribution()
{
......
include/ck_tile/ops/smoothquant.hpp 0 → 100644
View file @ 9533a172
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck_tile/ops/smoothquant/kernel/moe_smoothquant_kernel.hpp"
#include "ck_tile/ops/smoothquant/kernel/smoothquant_kernel.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_problem.hpp"
#include "ck_tile/ops/smoothquant/pipeline/smoothquant_pipeline_two_pass.hpp"
#include "ck_tile/ops/common/generic_2d_block_shape.hpp"
#include "ck_tile/ops/common/tensor_layout.hpp"
include/ck_tile/ops/smoothquant/kernel/moe_smoothquant_kernel.hpp 0 → 100644
View file @ 9533a172
// 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 MoeSmoothquantHostArgs
{
const void* p_x; // [tokens ,hidden_size], input, fp16/bf16
const void* p_xscale; // [experts, hidden_size], input, columnwise scale, fp32
const void* p_topk_ids; // [tokens, topk]
void* p_yscale; // [topk * tokens, 1], output, rowwise quant scale
void* p_qy; // [topk * tokens, hidden_size], output
index_t tokens;
index_t hidden_size;
index_t experts;
index_t topk;
index_t x_stride; // input x row stride
index_t y_stride; // output y stride(stride for topk)
};
// TODO: Extract some type to wrapper class
template <typename Pipeline_>
struct MoeSmoothquant
{
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>{};
static_assert(Problem::BlockShape::Repeat_M == 1);
struct Kargs
{
const void* p_x; // [tokens ,hidden_size], input, fp16/bf16
const void* p_xscale; // [experts, hidden_size], input, columnwise scale, fp32
const void* p_topk_ids; // [tokens, topk]
void* p_yscale; // [topk, tokens, 1], output, rowwise quant scale
void* p_qy; // [topk, tokens, hidden_size], output
index_t tokens;
index_t hidden_size;
index_t experts;
index_t topk;
index_t x_stride; // input x row stride
index_t y_stride; // output y stride(stride for topk)
};
using Hargs = MoeSmoothquantHostArgs;
CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs)
{
return Kargs{hargs.p_x,
hargs.p_xscale,
hargs.p_topk_ids,
hargs.p_yscale,
hargs.p_qy,
hargs.tokens,
hargs.hidden_size,
hargs.experts,
hargs.topk,
hargs.x_stride,
hargs.y_stride};
}
CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs)
{
return dim3(hargs.topk, integer_divide_ceil(hargs.tokens, Block_M), 1);
}
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_("moe_smoothquant_") + _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 index_t i_topk = blockIdx.x;
const index_t i_token = blockIdx.y * Block_M;
const index_t i_token_in_thrd =
__builtin_amdgcn_readfirstlane(threadIdx.x / Problem::BlockShape::ThreadPerBlock_N);
const index_t i_expert = reinterpret_cast<const index_t*>(
kargs.p_topk_ids)[(i_token + i_token_in_thrd) * kargs.topk + i_topk];
// [tokens ,hidden_size]
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.tokens, kargs.hidden_size),
make_tuple(kargs.x_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>{}), {i_token, 0});
}();
// [experts, hidden_size],
const auto xscale_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<const XScaleDataType*>(kargs.p_xscale) + i_expert * kargs.hidden_size,
make_tuple(kargs.hidden_size),
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});
}();
// [topk, tokens]
auto yscale_window = [&]() {
const auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<YScaleDataType*>(kargs.p_yscale) + i_topk * kargs.tokens,
make_tuple(kargs.tokens),
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>{}), {i_token});
}();
// [topk, tokens, hidden_size]
auto qy_window = [&]() {
auto tmp_ = make_naive_tensor_view<address_space_enum::global>(
static_cast<QYDataType*>(kargs.p_qy) + i_topk * kargs.tokens * kargs.y_stride,
make_tuple(kargs.tokens, kargs.hidden_size),
make_tuple(kargs.y_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>{}), {i_token, 0});
}();
__shared__ char smem[GetSmemSize()];
Pipeline{}(x_window, xscale_window, yscale_window, qy_window, kargs.hidden_size, smem);
}
};
} // namespace ck_tile
include/ck_tile/ops/smoothquant/kernel/smoothquant_kernel.hpp 0 → 100644
View file @ 9533a172
// 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; // [m ,n], input, fp16/bf16
const void* p_xscale; // [1, n], input, columnwise scale, fp32
void* p_yscale; // [m, 1], output, rowwise quant scale (amax / 127) of (p_x * p_xscale)
void* p_qy; // [m, n], output, p_x * p_xscale / p_yscale
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 dim3(integer_divide_ceil(hargs.m, 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/pipeline/smoothquant_pipeline_default_policy.hpp 0 → 100644
View file @ 9533a172
// 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 @ 9533a172
// 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 bool UseMax3 = true; // TODO - Move to trait
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_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
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 = [&]() {
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
{
return block_reduce2d(y,
reduce_absmax_func.GetIdentityValue<ComputeDataType>(),
reduce_absmax3_func,
sequence<1, 2>{});
}
else
{
return 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 @ 9533a172
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, 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 @ 9533a172
// 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 bool UseMax3 = true; // TODO - Move to trait
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_absmax3_func = [](auto acc_, auto v_0_, auto v_1_) {
float rtn;
asm volatile("v_max3_f32 %0, %1, abs(%2), abs(%3)"
: "=v"(rtn)
: "v"(acc_), "v"(v_0_), "v"(v_1_));
return rtn;
};
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);
constexpr auto x_size_per_row =
x.get_tile_distribution().get_ys_to_d_descriptor().get_lengths().at(number<1>{});
if constexpr(UseMax3 && std::is_same_v<ComputeDataType, float> &&
x_size_per_row % 2 == 0)
block_reduce2d(y, absmax, reduce_absmax3_func, sequence<1, 2>{});
else
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
include/ck_tile/ops/welford/block/block_welford.hpp
View file @ 9533a172
......@@ -11,9 +11,10 @@ namespace ck_tile {
template <typename Problem_, typename Policy_ = void>
struct BlockWelford
{
using Problem = remove_cvref_t<Problem_>;
using XDataType = typename Problem::XDataType;
using ComputeDataType = typename Problem::ComputeDataType;
using Problem = remove_cvref_t<Problem_>;
using XDataType = typename Problem::XDataType;
using ComputeDataType = typename Problem::ComputeDataType;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
CK_TILE_DEVICE constexpr BlockWelford() {}
......@@ -46,8 +47,11 @@ struct BlockWelford
auto x = ck_tile::type_convert<ComputeDataType>(x_tensor[in_dstr_idx]);
welford_update(
mean_tensor(out_dstr_idx), var_tensor(out_dstr_idx), x, cur_count_);
welford_update(mean_tensor(out_dstr_idx),
var_tensor(out_dstr_idx),
x,
cur_count_,
constant<kFastFDiv>{});
});
}
});
......@@ -89,7 +93,8 @@ struct BlockWelford
template <typename Problem_, typename Policy_ = void>
struct BlockWelfordSync
{
using Problem = remove_cvref_t<Problem_>;
using Problem = remove_cvref_t<Problem_>;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
template <typename MeanDistributedTensor_, typename VarDistributedTensor_>
CK_TILE_DEVICE void
......@@ -157,7 +162,8 @@ struct BlockWelfordSync
v_local_count,
v_remote_mean,
v_remote_var,
v_remote_count);
v_remote_count,
constant<kFastFDiv>{});
});
}
});
......@@ -173,8 +179,9 @@ struct BlockWelfordSync
template <typename Problem_, typename Policy_ = void>
struct BlockWelfordCrossWarpSync
{
using Problem = remove_cvref_t<Problem_>;
using BlockShape = typename Problem::BlockShape;
using Problem = remove_cvref_t<Problem_>;
using BlockShape = typename Problem::BlockShape;
static constexpr bool kFastFDiv = Problem::kFastFDiv;
template <typename MeanDistributedTensor_>
CK_TILE_DEVICE static constexpr index_t GetReduceWarps()
......@@ -304,7 +311,8 @@ struct BlockWelfordCrossWarpSync
v_local_count,
v_remote_mean,
v_remote_var,
v_remote_count);
v_remote_count,
constant<kFastFDiv>{});
});
mean_tensor.get_thread_buffer()(i_0) = v_local_mean;
......@@ -351,12 +359,23 @@ CK_TILE_DEVICE constexpr index_t block_tile_welford_calculate_max_count(int row_
}
// Note: this function must be called after all the computation
template <typename VarDistributedTensor_>
template <typename VarDistributedTensor_, bool FastFdiv_ = false>
CK_TILE_DEVICE constexpr void block_tile_welford_post_scale_var(VarDistributedTensor_& var_tensor,
int count)
int count,
bool_constant<FastFdiv_> = {})
{
using DataType = typename VarDistributedTensor_::DataType;
tile_elementwise_inout([&count](auto& x) { x = x / type_convert<DataType>(count); },
var_tensor);
tile_elementwise_inout(
[&count](auto& x) {
if(FastFdiv_ && std::is_same_v<DataType, float>)
{
x = x * __builtin_amdgcn_rcpf(type_convert<DataType>(count));
}
else
{
x = x / type_convert<DataType>(count);
}
},
var_tensor);
}
} // namespace ck_tile
include/ck_tile/ops/welford/block/block_welford_problem.hpp
View file @ 9533a172
......@@ -7,12 +7,13 @@
namespace ck_tile {
template <typename XDataType_, typename ComputeDataType_, typename BlockShape_>
template <typename XDataType_, typename ComputeDataType_, typename BlockShape_, bool kFastFDiv_>
struct BlockWelfordProblem
{
using XDataType = remove_cvref_t<XDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
using BlockShape = remove_cvref_t<BlockShape_>;
using XDataType = remove_cvref_t<XDataType_>;
using ComputeDataType = remove_cvref_t<ComputeDataType_>;
using BlockShape = remove_cvref_t<BlockShape_>;
static constexpr bool kFastFDiv = kFastFDiv_;
};
} // namespace ck_tile
include/ck_tile/ops/welford/thread/thread_welford.hpp
View file @ 9533a172
......@@ -7,25 +7,46 @@
namespace ck_tile {
template <typename T>
CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count)
template <typename T, bool kFastFDiv = false>
CK_TILE_DEVICE void welford_update(T& mean, T& var, T x, int count, bool_constant<kFastFDiv> = {})
{
// TODO: check nan? maybe no
T delta = x - mean;
mean += delta / count;
if(kFastFDiv && std::is_same_v<T, float>)
{
mean += delta * __builtin_amdgcn_rcpf(count);
}
else
{
mean += delta / count;
}
T delta2 = x - mean;
var += delta * delta2;
}
template <typename T>
CK_TILE_DEVICE static void
welford_merge(T& mean_a, T& var_a, int& count_a, T mean_b, T var_b, int count_b)
template <typename T, bool kFastFDiv = false>
CK_TILE_DEVICE static void welford_merge(T& mean_a,
T& var_a,
int& count_a,
T mean_b,
T var_b,
int count_b,
bool_constant<kFastFDiv> = {})
{
int count = count_a + count_b;
T count_ = type_convert<T>(count);
T count_a_ = type_convert<T>(count_a);
T count_b_ = type_convert<T>(count_b);
T count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_;
int count = count_a + count_b;
T count_ = type_convert<T>(count);
T count_a_ = type_convert<T>(count_a);
T count_b_ = type_convert<T>(count_b);
T count_b_over_count;
if(kFastFDiv && std::is_same_v<T, float>)
{
count_b_over_count =
count == 0 ? type_convert<T>(0) : count_b_ * __builtin_amdgcn_rcpf(count_);
}
else
{
count_b_over_count = count == 0 ? type_convert<T>(0) : count_b_ / count_;
}
T delta = mean_b - mean_a;
mean_a += delta * count_b_over_count;
......
include/ck_tile/remod.py
View file @ 9533a172
from datetime import datetime
import pathlib
from pathlib import Path
import subprocess
......@@ -8,8 +9,8 @@ NS = 'ck_tile'
OPS = 'ops'
OPS_COMMON = 'common' # common header will be duplicated into ops/* other module
HEADER_COMMON = """// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
HEADER_COMMON = f"""// SPDX-License-Identifier: MIT
// Copyright (c) 2018-{datetime.now().year}, Advanced Micro Devices, Inc. All rights reserved.\n
"""
# aa/bb/cc/file.hpp -> (aa, bb, cc, file.hpp)
......
library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
View file @ 9533a172
......@@ -180,6 +180,58 @@ struct DeviceOperationInstanceFactory<
}
}
#endif
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, ck::bhalf_t> && is_same_v<BDataType, ck::bhalf_t> &&
is_same_v<CDataType, ck::bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_wmma_bf16_bf16_bf16_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_wmma_bf16_bf16_bf16_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_wmma_bf16_bf16_bf16_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_wmma_bf16_bf16_bf16_km_nk_mn_instances(op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_INT8
if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
is_same_v<CDataType, int8_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_wmma_int8_int8_int8_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_wmma_int8_int8_int8_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_wmma_int8_int8_int8_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_wmma_int8_int8_int8_km_nk_mn_instances(op_ptrs);
}
}
#endif
#endif
#ifdef CK_USE_XDL
......
library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_batched.hpp 0 → 100644
View file @ 9533a172
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_xdl_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_BF16
void add_device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmV2MultiD<Row,
Col,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmV2MultiD<Row,
Col,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmV2MultiD<Row,
Col,
Empty_Tuple,
Row,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP8
void add_device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_default_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmV2MultiD<Row,
Col,
Empty_Tuple,
Row,
F8,
F8,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmV2MultiD<Row,
Col,
Empty_Tuple,
Row,
F8,
F8,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances(
std::vector<std::unique_ptr<DeviceBatchedGemmV2MultiD<Row,
Col,
Empty_Tuple,
Row,
F8,
F8,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
template <typename ADataType,
typename BDataType,
typename DsDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename DsLayout,
typename CLayout>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatchedGemmV2MultiD<
ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
DsDataType,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
{
using DeviceOp = DeviceBatchedGemmV2MultiD<ALayout,
BLayout,
DsLayout,
CLayout,
ADataType,
BDataType,
DsDataType,
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_default_instances(
op_ptrs);
add_device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
add_device_batched_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
}
}
#endif
#ifdef CK_ENABLE_FP8
if constexpr(is_same_v<ADataType, f8_t> && is_same_v<BDataType, f8_t> &&
is_same_v<CDataType, bhalf_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_comp_default_instances(
op_ptrs);
add_device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v1_default_instances(
op_ptrs);
add_device_batched_gemm_xdl_universal_f8_f8_bf16_mk_nk_mn_mem_v2_default_instances(
op_ptrs);
}
}
#endif
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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