mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-05-26 08:00:13 +00:00
introducing ck_tile! (#1216)
* enable gfx940 * switch between intrinsic mfma routines on mi100/200 and mi300 * fix mfma_int8 on MI300 * disable 2 int8 examples on MI300 * Update cmake-ck-dev.sh * restore gitignore file * modify Jenkinsfile to the internal repo * Bump rocm-docs-core from 0.24.0 to 0.29.0 in /docs/sphinx Bumps [rocm-docs-core](https://github.com/RadeonOpenCompute/rocm-docs-core) from 0.24.0 to 0.29.0. - [Release notes](https://github.com/RadeonOpenCompute/rocm-docs-core/releases) - [Changelog](https://github.com/RadeonOpenCompute/rocm-docs-core/blob/develop/CHANGELOG.md) - [Commits](https://github.com/RadeonOpenCompute/rocm-docs-core/compare/v0.24.0...v0.29.0) --- updated-dependencies: - dependency-name: rocm-docs-core dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> * initial enablement of gfx950 * fix clang format * disable examples 31 and 41 int8 on gfx950 * add code * fix build wip * fix xx * now can build * naming * minor fix * wip fix * fix macro for exp2; fix warpgemm a/b in transposedC * unify as tuple_array * Update the required Python version to 3.9 * Update executable name in test scripts * re-structure tuple/array to avoid spill * Merge function templates * Fix format * Add constraint to array<> ctor * Re-use function * Some minor changes * remove wrong code in store_raw() * fix compile issue in transpose * Rename enum Rename 'cood_transform_enum' to 'coord_transform_enum' * let more integral_constant->constant, and formating * make sure thread_buffer can be tuple/array * temp fix buffer_store spill * not using custom data type by default, now we can have ISA-level same code as opt_padding * fix compile error, fp8 not ready now * fix fp8 duplicated move/shift/and/or problem * Default use CK_TILE_FLOAT_TO_FP8_STOCHASTIC rounding mode * fix scratch in fp8 kernel * update some readme * fix merge from upstream * sync with upstream * sync upstream again * sync 22 * remove unused * fix clang-format * update README of ck_tile example * fix several issue * let python version to be 3.8 as minimal * remove ck_tile example from default cmake target like all/install/check * remove mistake * 1).support receipe in generate.py 2).use simplified mask type 3).change left/right to pass into karg * fix some bug in group-mode masking and codegen. update README * F8 quantization for FMHA forward (#1224) * Add SAccElementFunction, PComputeElementFunction, OAccElementFunction in pipeline * Add element function to fmha api * Adjust P elementwise function * Fix bug of elementwise op, our elementwise op is not inout * Add some elementwise op, prepare to quantization * Let generate.py can generate different elementwise function * To prevent compiler issue, remove the elementwise function we have not used. * Remove f8 pipeline, we should share the same pipeline even in f8 * Remove remove_cvref_t * Avoid warning * Fix wrong fp8 QK/KV block gemm setting * Check fp8 rounding error in check_err() * Set fp8 rounding error for check_err() * Use CK_TILE_FLOAT_TO_FP8_STANDARD as default fp8 rounding mode * 1. codgen the f8 api and kernel 2. f8 host code * prevent warning in filter mode * Remove not-in-use elementwise function kargs * Remove more not-in-use elementwise function kargs * Small refinements in C++ source files * Use conditional_t<> to simplify code * Support heterogeneous argument for binary function types * Re-use already-existing scales<> functor template * Fix wrong value produced by saturating * Generalize the composes<> template * Unify saturates<> implementation * Fix type errors in composes<> * Extend less_equal<> * Reuse the existing template less_equal<> in check_err() * Add equal<float> & equal<double> * Rename check_err() parameter * Rename check_err() parameter * Add FIXME comment for adding new macro in future * Remove unnecessary cast to void * Eliminate duplicated code * Avoid dividing api pool into more than 2 groups * Use more clear variable names * Use affirmative condition in if stmt * Remove blank lines * Donot perfect forwarding in composes<> * To fix compile error, revert generate.py back to4439cc107d* Fix bug of p element function * Add compute element op to host softmax * Remove element function in api interface * Extract user parameter * Rename pscale and oscale variable * rename f8 to fp8 * rename more f8 to fp8 * Add pipeline::operator() without element_functor * 1. Remove deprecated pipeline enum 2. Refine host code parameter * Use quantization range as input * 1. Rename max_dtype to dtype_max. 2. Rename scale to scale_s 3.Add init description * Refine description * prevent early return * unify _squant kernel name in cpp, update README * Adjust the default range. * Refine error message and bias range * Add fp8 benchmark and smoke test * fix fp8 swizzle_factor=4 case --------- Co-authored-by: Po Yen Chen <PoYen.Chen@amd.com> Co-authored-by: carlushuang <carlus.huang@amd.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: illsilin <Illia.Silin@amd.com> Co-authored-by: Illia Silin <98187287+illsilin@users.noreply.github.com> Co-authored-by: Jing Zhang <jizha@amd.com> Co-authored-by: zjing14 <zhangjing14@gmail.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Po-Yen, Chen <PoYen.Chen@amd.com> Co-authored-by: rocking <ChunYu.Lai@amd.com> [ROCm/composable_kernel commit:db376dd8a4]
This commit is contained in:
carlushuang
@@ -0,0 +1,64 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename AElementOp = ck_tile::identity,
|
||||
typename BElementOp = ck_tile::identity,
|
||||
typename BinaryElementOp = ck_tile::plus<AccDataType>>
|
||||
CK_TILE_HOST void reference_batched_elementwise(const HostTensor<ADataType>& a_b_m_n,
|
||||
const HostTensor<BDataType>& b_b_m_n,
|
||||
HostTensor<CDataType>& c_b_m_n,
|
||||
const AElementOp& a_element_op = {},
|
||||
const BElementOp& b_element_op = {},
|
||||
const BinaryElementOp& binary_element_op = {})
|
||||
{
|
||||
const ck_tile::index_t N = c_b_m_n.mDesc.get_lengths()[2];
|
||||
|
||||
const bool broadcast_a_dim_b = (a_b_m_n.get_lengths()[0] == 1);
|
||||
const bool broadcast_a_dim_m = (a_b_m_n.get_lengths()[1] == 1);
|
||||
const bool broadcast_a_dim_n = (a_b_m_n.get_lengths()[2] == 1);
|
||||
|
||||
const bool broadcast_b_dim_b = (b_b_m_n.get_lengths()[0] == 1);
|
||||
const bool broadcast_b_dim_m = (b_b_m_n.get_lengths()[1] == 1);
|
||||
const bool broadcast_b_dim_n = (b_b_m_n.get_lengths()[2] == 1);
|
||||
|
||||
auto f = [&](auto batch, auto m) {
|
||||
for(ck_tile::index_t n = 0; n < N; ++n)
|
||||
{
|
||||
AccDataType v_a{};
|
||||
{
|
||||
ck_tile::index_t i_b = (broadcast_a_dim_b ? 0 : batch);
|
||||
ck_tile::index_t i_m = (broadcast_a_dim_m ? 0 : m);
|
||||
ck_tile::index_t i_n = (broadcast_a_dim_n ? 0 : n);
|
||||
|
||||
v_a = ck_tile::type_convert<AccDataType>(a_element_op(a_b_m_n(i_b, i_m, i_n)));
|
||||
}
|
||||
|
||||
AccDataType v_b{};
|
||||
{
|
||||
ck_tile::index_t i_b = (broadcast_b_dim_b ? 0 : batch);
|
||||
ck_tile::index_t i_m = (broadcast_b_dim_m ? 0 : m);
|
||||
ck_tile::index_t i_n = (broadcast_b_dim_n ? 0 : n);
|
||||
|
||||
v_b = ck_tile::type_convert<AccDataType>(b_element_op(b_b_m_n(i_b, i_m, i_n)));
|
||||
}
|
||||
|
||||
c_b_m_n(batch, m, n) = ck_tile::type_convert<CDataType>(binary_element_op(v_a, v_b));
|
||||
}
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(
|
||||
std::thread::hardware_concurrency());
|
||||
}
|
||||
} // namespace ck_tile
|
||||
50
include/ck_tile/host/reference/reference_batched_gemm.hpp
Normal file
50
include/ck_tile/host/reference/reference_batched_gemm.hpp
Normal file
@@ -0,0 +1,50 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename AElementOp = ck_tile::identity,
|
||||
typename BElementOp = ck_tile::identity,
|
||||
typename ACCElementOp = ck_tile::identity>
|
||||
CK_TILE_HOST void reference_batched_gemm(const HostTensor<ADataType>& a_b_m_k,
|
||||
const HostTensor<BDataType>& b_b_n_k,
|
||||
HostTensor<CDataType>& c_b_m_n,
|
||||
const AElementOp& a_element_op = {},
|
||||
const BElementOp& b_element_op = {},
|
||||
const ACCElementOp& acc_element_op = {})
|
||||
{
|
||||
const int N = b_b_n_k.mDesc.get_lengths()[1];
|
||||
const int K = b_b_n_k.mDesc.get_lengths()[2];
|
||||
|
||||
auto f = [&](auto batch, auto m) {
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
AccDataType v_acc = 0;
|
||||
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
ADataType v_a = a_element_op(a_b_m_k(batch, m, k));
|
||||
BDataType v_b = b_element_op(b_b_n_k(batch, n, k));
|
||||
|
||||
v_acc += ck_tile::type_convert<AccDataType>(v_a) *
|
||||
ck_tile::type_convert<AccDataType>(v_b);
|
||||
}
|
||||
|
||||
c_b_m_n(batch, m, n) = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
|
||||
}
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f, c_b_m_n.mDesc.get_lengths()[0], c_b_m_n.mDesc.get_lengths()[1])(
|
||||
std::thread::hardware_concurrency());
|
||||
}
|
||||
} // namespace ck_tile
|
||||
32
include/ck_tile/host/reference/reference_batched_masking.hpp
Normal file
32
include/ck_tile/host/reference/reference_batched_masking.hpp
Normal file
@@ -0,0 +1,32 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename CDataType, typename MaskingType>
|
||||
CK_TILE_HOST void reference_batched_masking(HostTensor<CDataType>& c_b_m_n, const MaskingType& mask)
|
||||
{
|
||||
const int M = c_b_m_n.mDesc.get_lengths()[1];
|
||||
const int N = c_b_m_n.mDesc.get_lengths()[2];
|
||||
|
||||
auto f = [&](auto batch) {
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
for(int m = 0; m < M; ++m)
|
||||
{
|
||||
if(mask.IsOutOfBound(m, n))
|
||||
c_b_m_n(batch, m, n) = -ck_tile::numeric<CDataType>::infinity();
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f,
|
||||
c_b_m_n.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
|
||||
}
|
||||
} // namespace ck_tile
|
||||
71
include/ck_tile/host/reference/reference_batched_softmax.hpp
Normal file
71
include/ck_tile/host/reference/reference_batched_softmax.hpp
Normal file
@@ -0,0 +1,71 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename ADataType,
|
||||
typename CompDataType,
|
||||
typename BDataType,
|
||||
typename CompElementOp = ck_tile::identity>
|
||||
CK_TILE_HOST void reference_batched_softmax(
|
||||
const HostTensor<ADataType>& a_b_m_n,
|
||||
HostTensor<BDataType>& b_b_m_n,
|
||||
const CompElementOp& comp_element_op = {},
|
||||
std::optional<std::reference_wrapper<HostTensor<CompDataType>>> lse_b_m = std::nullopt)
|
||||
{
|
||||
const int N = a_b_m_n.mDesc.get_lengths()[2];
|
||||
|
||||
auto f = [&](auto batch, auto m) {
|
||||
CompDataType v_max = -ck_tile::numeric<CompDataType>::infinity();
|
||||
|
||||
// max
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
|
||||
|
||||
v_max = v_max < v_a ? v_a : v_max;
|
||||
}
|
||||
|
||||
CompDataType v_exp_sum = 0;
|
||||
// validate v_max if all the elements within a row are -INF
|
||||
if(std::isinf(v_max) && v_max < 0)
|
||||
{
|
||||
v_max = ck_tile::type_convert<CompDataType>(0.f);
|
||||
}
|
||||
|
||||
// sum
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
|
||||
|
||||
v_exp_sum += ck_tile::exp(v_a - v_max);
|
||||
}
|
||||
|
||||
// if sum is zero(masked), or nan/inf(other computation error), don't do divide
|
||||
CompDataType inv_sum = (v_exp_sum == 0.f ? 1.f : 1.f / v_exp_sum);
|
||||
|
||||
// elementwise
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const CompDataType v_a = ck_tile::type_convert<CompDataType>(a_b_m_n(batch, m, n));
|
||||
const CompDataType v_b = ck_tile::exp(v_a - v_max) * inv_sum;
|
||||
|
||||
b_b_m_n(batch, m, n) = ck_tile::type_convert<BDataType>(comp_element_op(v_b));
|
||||
}
|
||||
// lse
|
||||
if(lse_b_m)
|
||||
{
|
||||
lse_b_m->get()(batch, m) = v_max + ck_tile::log(v_exp_sum);
|
||||
}
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f, b_b_m_n.mDesc.get_lengths()[0], b_b_m_n.mDesc.get_lengths()[1])(
|
||||
std::thread::hardware_concurrency());
|
||||
}
|
||||
} // namespace ck_tile
|
||||
50
include/ck_tile/host/reference/reference_gemm.hpp
Normal file
50
include/ck_tile/host/reference/reference_gemm.hpp
Normal file
@@ -0,0 +1,50 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename ADataType,
|
||||
typename BDataType,
|
||||
typename AccDataType,
|
||||
typename CDataType,
|
||||
typename AElementOp = ck_tile::identity,
|
||||
typename BElementOp = ck_tile::identity,
|
||||
typename ACCElementOp = ck_tile::identity>
|
||||
CK_TILE_HOST void reference_gemm(const HostTensor<ADataType>& a_m_k,
|
||||
const HostTensor<BDataType>& b_n_k,
|
||||
HostTensor<CDataType>& c_m_n,
|
||||
const AElementOp& a_element_op = {},
|
||||
const BElementOp& b_element_op = {},
|
||||
const ACCElementOp& acc_element_op = {})
|
||||
{
|
||||
const int N = b_n_k.mDesc.get_lengths()[0];
|
||||
const int K = b_n_k.mDesc.get_lengths()[1];
|
||||
|
||||
auto f = [&](auto m) {
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
AccDataType v_acc = 0;
|
||||
|
||||
for(int k = 0; k < K; ++k)
|
||||
{
|
||||
ADataType v_a = a_element_op(a_m_k(m, k));
|
||||
BDataType v_b = b_element_op(b_n_k(n, k));
|
||||
|
||||
v_acc += ck_tile::type_convert<AccDataType>(v_a) *
|
||||
ck_tile::type_convert<AccDataType>(v_b);
|
||||
}
|
||||
|
||||
c_m_n(m, n) = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
|
||||
}
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f,
|
||||
c_m_n.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
|
||||
}
|
||||
} // namespace ck_tile
|
||||
61
include/ck_tile/host/reference/reference_im2col.hpp
Normal file
61
include/ck_tile/host/reference/reference_im2col.hpp
Normal file
@@ -0,0 +1,61 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename T>
|
||||
CK_TILE_HOST void reference_im2col(HostTensor<T>& in_mtx_host_ref,
|
||||
const HostTensor<T>& in_host,
|
||||
int /*N*/,
|
||||
int /*K*/,
|
||||
int C,
|
||||
int /*Y*/,
|
||||
int X,
|
||||
int Hi,
|
||||
int Wi,
|
||||
int Ho,
|
||||
int Wo,
|
||||
int ConvStrideH,
|
||||
int ConvStrideW,
|
||||
int ConvDilationH,
|
||||
int ConvDilationW,
|
||||
int InLeftPadH,
|
||||
int InLeftPadW,
|
||||
int /*InRightPadH*/,
|
||||
int /*InRightPadW*/)
|
||||
{
|
||||
int GemmM = in_mtx_host_ref.get_lengths()[0];
|
||||
int GemmK = in_mtx_host_ref.get_lengths()[1];
|
||||
|
||||
for(int gemm_m = 0; gemm_m < GemmM; ++gemm_m)
|
||||
{
|
||||
int mtmp = gemm_m;
|
||||
int n = mtmp / (Ho * Wo);
|
||||
mtmp -= n * Ho * Wo;
|
||||
int ho = mtmp / Wo;
|
||||
int wo = mtmp - ho * Wo;
|
||||
|
||||
for(int gemm_k = 0; gemm_k < GemmK; ++gemm_k)
|
||||
{
|
||||
int ktmp = gemm_k;
|
||||
int y = ktmp / (X * C);
|
||||
ktmp -= y * X * C;
|
||||
int x = ktmp / C;
|
||||
int c = ktmp - x * C;
|
||||
|
||||
int hi = y * ConvDilationH + ho * ConvStrideH - InLeftPadH;
|
||||
int wi = x * ConvDilationW + wo * ConvStrideW - InLeftPadW;
|
||||
|
||||
bool inbound = (hi >= 0 && hi < Hi && wi >= 0 && wi < Wi);
|
||||
|
||||
in_mtx_host_ref(gemm_m, gemm_k) = inbound ? in_host(n, hi, wi, c) : 0;
|
||||
}
|
||||
}
|
||||
}
|
||||
} // namespace ck_tile
|
||||
32
include/ck_tile/host/reference/reference_reduce.hpp
Normal file
32
include/ck_tile/host/reference/reference_reduce.hpp
Normal file
@@ -0,0 +1,32 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename ADataType, typename AccDataType, typename BDataType>
|
||||
CK_TILE_HOST void reference_reduce(const HostTensor<ADataType>& a_m_n, HostTensor<BDataType>& b_m)
|
||||
{
|
||||
auto f = [&](auto m) {
|
||||
const int N = a_m_n.mDesc.get_lengths()[1];
|
||||
|
||||
AccDataType v_acc = 0;
|
||||
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const ADataType v_a = a_m_n(m, n);
|
||||
|
||||
v_acc += v_a;
|
||||
}
|
||||
|
||||
b_m(m) = ck_tile::type_convert<BDataType>(v_acc);
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f, b_m.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
|
||||
}
|
||||
} // namespace ck_tile
|
||||
51
include/ck_tile/host/reference/reference_softmax.hpp
Normal file
51
include/ck_tile/host/reference/reference_softmax.hpp
Normal file
@@ -0,0 +1,51 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include "ck_tile/core.hpp"
|
||||
#include "ck_tile/host/host_tensor.hpp"
|
||||
#include <thread>
|
||||
|
||||
namespace ck_tile {
|
||||
|
||||
template <typename ADataType, typename AccDataType, typename BDataType>
|
||||
CK_TILE_HOST void reference_softmax(const HostTensor<ADataType>& a_m_n,
|
||||
HostTensor<BDataType>& b_m_n)
|
||||
{
|
||||
auto f = [&](auto m) {
|
||||
const int N = a_m_n.mDesc.get_lengths()[1];
|
||||
|
||||
AccDataType v_max = ck_tile::numeric<ADataType>::Lowest();
|
||||
|
||||
// max
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const ADataType v_a = a_m_n(m, n);
|
||||
|
||||
v_max = v_max < v_a ? v_a : v_max;
|
||||
}
|
||||
|
||||
AccDataType v_exp_sum = 0;
|
||||
|
||||
// sum
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const ADataType v_a = a_m_n(m, n);
|
||||
|
||||
v_exp_sum += ck_tile::exp(v_a - v_max);
|
||||
}
|
||||
|
||||
// elementwise
|
||||
for(int n = 0; n < N; ++n)
|
||||
{
|
||||
const ADataType v_a = a_m_n(m, n);
|
||||
|
||||
b_m_n(m, n) = ck_tile::exp(v_a - v_max) / v_exp_sum;
|
||||
}
|
||||
};
|
||||
|
||||
make_ParallelTensorFunctor(f,
|
||||
b_m_n.mDesc.get_lengths()[0])(std::thread::hardware_concurrency());
|
||||
}
|
||||
} // namespace ck_tile
|
||||
Reference in New Issue
Block a user