Skip to content

GitLab

Unverified Commit 29dcb956 authored by Illia Silin's avatar Illia Silin Committed by GitHub
Browse files

Merge pull request #33 from ROCm/lwpck-1292 

Merge from the public repo.
parents 29deceb6 cbcc844e
Hide whitespace changes
Inline Side-by-side
Showing with 1854 additions and 71 deletions
library/src/tensor_operation_instance/gpu/normalization_fwd/device_groupnorm_fwd_swish_f16_instance.cpp
View file @ 29dcb956
......@@ -11,7 +11,7 @@ namespace instance {
using Swish = ck::tensor_operation::element_wise::Swish;
void add_device_normalization_fwd_rank_5_3_swish_f16_instances(
std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, Swish, 5, 3>>>&
std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, Swish, 5, 3>>>&
instances)
{
add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm2d_fwd_f16_instance.cpp
View file @ 29dcb956
......@@ -11,7 +11,7 @@ namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_fwd_rank_2_1_f16_instances(
std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, Pass, 2, 1>>>&
std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, Pass, 2, 1>>>&
instances)
{
add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization_fwd/device_layernorm4d_fwd_f16_instance.cpp
View file @ 29dcb956
......@@ -11,7 +11,7 @@ namespace instance {
using Pass = ck::tensor_operation::element_wise::PassThrough;
void add_device_normalization_fwd_rank_4_3_f16_instances(
std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F32, Pass, 4, 3>>>&
std::vector<std::unique_ptr<DeviceNormalizationFwd<F16, F16, F16, F16, F16, Pass, 4, 3>>>&
instances)
{
add_device_operation_instances(instances,
......
library/src/tensor_operation_instance/gpu/normalization_fwd/normalization_fwd_instance_common.hpp
View file @ 29dcb956
......@@ -23,24 +23,24 @@ using device_normalization_f16_instances =
// clang-format off
std::tuple <
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
// clang-format on
>;
......@@ -49,31 +49,31 @@ using device_normalization_splitk_f16_instances =
// clang-format off
std::tuple <
// XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYSrcVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdScalarPerVector>
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 2, 1, 2, 1, 2, 1, 2, 2, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 4, 1, 4, 1, 4, 1, 4, 4, 1>, // irregular size
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 128, 1, 128, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 2, 16, 1, 8, 1, 8, 1, 8, 8, 2>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 256, 1, 256, 1, 32, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 512, 1, 512, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 8, 1, 8, 1, 8, 1, 8, 8, 1>,
DeviceNormalizationFwdSplitKImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 1024, 1, 1024, 1, 16, 1, 8, 1, 8, 1, 8, 8, 1>
// clang-format on
>;
template <typename OutElementwise, index_t Rank, index_t Reduce>
using device_normalization_f16_generic_instance = std::tuple<
// clang-format off
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F32, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
DeviceNormalizationFwdImpl<F16, F16, F16, F32, F16, F16, OutElementwise, Rank, Reduce, 64, 1, 64, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/permute_scale/CMakeLists.txt 0 → 100644
View file @ 29dcb956
add_instance_library(device_permute_scale_instance
device_permute_scale_instances.cpp)
library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.cpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using F16 = ck::half_t;
using F32 = float;
using Pass = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
// clang-format off
using device_permute_scale_f16_instances =
std::tuple <
DeviceElementwiseImpl<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4, 1, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4, 8, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4, 4, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4, 2, ck::Sequence<1>, ck::Sequence<1>>
>;
using device_permute_scale_f32_instances = std::tuple<
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4, 1, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4, 8, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4, 4, ck::Sequence<1>, ck::Sequence<1>>,
DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4, 2, ck::Sequence<1>, ck::Sequence<1>>
>;
// clang-format on
void add_device_permute_scale_f16_instances(
std::vector<std::unique_ptr<
DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4>>>& instances)
{
add_device_operation_instances(instances, device_permute_scale_f16_instances{});
}
void add_device_permute_scale_f32_instances(
std::vector<std::unique_ptr<
DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4>>>& instances)
{
add_device_operation_instances(instances, device_permute_scale_f32_instances{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp
View file @ 29dcb956
......@@ -22,13 +22,13 @@ using S = ck::Sequence<Is...>;
using NHWGC = ck::tensor_layout::convolution::NHWGC;
using GKYXC = ck::tensor_layout::convolution::GKYXC;
using NHWGK = ck::tensor_layout::convolution::NHWGK;
using GK = ck::tensor_layout::convolution::G_K;
using G_K = ck::tensor_layout::convolution::G_K;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Relu = ck::tensor_operation::element_wise::Relu;
using TanH = ck::tensor_operation::element_wise::TanH;
using GK_Tuple = ck::Tuple<GK>;
using GK_GK_Tuple = ck::Tuple<GK, GK>;
using GK_Tuple = ck::Tuple<G_K>;
using GK_GK_Tuple = ck::Tuple<G_K, G_K>;
using I32_Tuple = ck::Tuple<int32_t>;
using F32_Tuple = ck::Tuple<float>;
using I32_F32_Tuple = ck::Tuple<int32_t, float>;
......
library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
View file @ 29dcb956
set(DEVICE_SOFTMAX_INSTANCES)
list(APPEND DEVICE_SOFTMAX_INSTANCES
add_instance_library(device_softmax_instance
device_softmax_f16_f16_instance_rank3_reduce1.cpp
device_softmax_f16_f16_instance_rank3_reduce2.cpp
device_softmax_f16_f16_instance_rank3_reduce3.cpp
......@@ -14,4 +13,3 @@ list(APPEND DEVICE_SOFTMAX_INSTANCES
device_softmax_f32_f32_instance_rank4_reduce2.cpp
device_softmax_f32_f32_instance_rank4_reduce3.cpp
device_softmax_f32_f32_instance_rank4_reduce4.cpp)
add_instance_library(device_softmax_instance ${DEVICE_SOFTMAX_INSTANCES})
library/src/tensor_operation_instance/gpu/transpose/device_transpose_instances_3d.cpp
View file @ 29dcb956
......@@ -19,22 +19,14 @@ void add_device_transpose_f16_instances(
std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 5>>>&
instances)
{
#ifdef CK_ENABLE_FP16
add_device_operation_instances(instances, device_transpose_f16_instances{});
#else
ignore = instances;
#endif
}
void add_device_transpose_f32_instances(
std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 5>>>&
instances)
{
#ifdef CK_ENABLE_FP32
add_device_operation_instances(instances, device_transpose_f32_instances{});
#else
ignore = instances;
#endif
}
} // namespace instance
......
library/src/utility/CMakeLists.txt
View file @ 29dcb956
## utility
set(UTILITY_SOURCE
add_library(utility STATIC
device_memory.cpp
host_tensor.cpp
convolution_parameter.cpp
)
add_library(utility STATIC ${UTILITY_SOURCE})
add_library(composable_kernel::utility ALIAS utility)
set_target_properties(utility PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_compile_options(utility PRIVATE ${CMAKE_COMPILER_WARNINGS})
target_include_directories(utility PUBLIC
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>"
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/utility>"
)
if(WIN32)
target_compile_definitions(utility PUBLIC NOMINMAX)
endif()
rocm_install(
TARGETS utility
......
profiler/include/profiler/profile_gemm_add_impl.hpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = Add;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Add>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_gemm_add_relu_impl.hpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_relu_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddRelu = ck::tensor_operation::element_wise::AddRelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddRelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddRelu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_gemm_add_silu_impl.hpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/gemm_add_silu.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename ELayout>
bool profile_gemm_add_silu_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddRelu = ck::tensor_operation::element_wise::AddSilu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddRelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
ALayout,
BLayout,
ck::Tuple<D0Layout>,
ELayout,
ADataType,
BDataType,
ck::Tuple<D0DataType>,
EDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddSilu>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD0},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
std::string op_name = op_ptr->GetTypeString();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
}
}
else
{
std::cout << op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_gemm_impl.hpp
View file @ 29dcb956
......@@ -42,7 +42,9 @@ int profile_gemm_impl(int do_verification,
int K,
int StrideA,
int StrideB,
int StrideC)
int StrideC,
int n_warmup,
int n_iter)
{
bool pass = true;
......@@ -165,8 +167,8 @@ int profile_gemm_impl(int do_verification,
std::string op_name = op_ptr->GetTypeString();
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, 10, 50});
float avg_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
std::size_t flop = std::size_t(2) * M * N * K;
......@@ -296,7 +298,7 @@ int profile_gemm_impl(int do_verification,
}
}
return pass ? 0 : 1;
return pass;
}
} // namespace profiler
......
profiler/include/profiler/profile_gemm_splitk_impl.hpp
View file @ 29dcb956
......@@ -42,7 +42,9 @@ bool profile_gemm_splitk_impl(int do_verification,
int StrideA,
int StrideB,
int StrideC,
int KBatch)
int KBatch,
int n_warmup,
int n_iter)
{
bool pass = true;
......@@ -143,7 +145,7 @@ bool profile_gemm_splitk_impl(int do_verification,
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 32, 36, 40, 64, 96, 128};
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 19, 20, 32, 38};
if(KBatch > 0)
{
......@@ -177,7 +179,8 @@ bool profile_gemm_splitk_impl(int do_verification,
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_verification)
{
......@@ -200,8 +203,8 @@ bool profile_gemm_splitk_impl(int do_verification,
std::string op_name = op_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
float ave_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
std::size_t flop = std::size_t(2) * M * N * K;
......
profiler/include/profiler/profile_grouped_gemm_impl.hpp
View file @ 29dcb956
......@@ -42,7 +42,9 @@ bool profile_grouped_gemm_impl(int do_verification,
const std::vector<int>& StrideAs,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs,
int kbatch = 1)
int kbatch = 1,
int n_warmup = 1,
int n_iter = 10)
{
bool pass = true;
......@@ -261,7 +263,8 @@ bool profile_grouped_gemm_impl(int do_verification,
for(std::size_t i = 0; i < gemm_descs.size(); i++)
c_device_buf[i]->SetZero();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
if(do_verification)
{
......@@ -307,8 +310,8 @@ bool profile_grouped_gemm_impl(int do_verification,
pass = pass && instance_pass;
}
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
float ave_time = invoker_ptr->Run(
argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
if(time_kernel)
{
......
profiler/include/profiler/profile_groupnorm_bwd_data_impl.hpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/groupnorm_bwd_data.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp"
namespace ck {
namespace profiler {
template <typename DYDataType,
typename XDataType,
typename GammaDataType,
typename MeanInvStdDataType,
typename ComputeDataType,
typename DXDataType>
bool profile_groupnorm_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> length)
{
// we don't need DGamma and DBeta here, just for reference class
using DGammaDataType = DXDataType;
using DBetaDataType = DXDataType;
if(length.size() != 5)
return false;
index_t N = length[0];
index_t G = length[3];
index_t C = length[4];
std::vector<index_t> reduce_dim = {1, 2, 4};
std::vector<index_t> gammaLength = {G, C};
Tensor<DYDataType> dy(length);
Tensor<XDataType> x(length);
Tensor<GammaDataType> gamma({G, C});
Tensor<MeanInvStdDataType> mean({N, G});
Tensor<MeanInvStdDataType> inv_std({N, G});
Tensor<DXDataType> dx(length);
Tensor<DXDataType> host_dx(length);
Tensor<DGammaDataType> host_dgamma({G, C});
Tensor<DBetaDataType> host_dbeta({G, C});
std::vector<index_t> strideDy =
std::vector<ck::index_t>{dy.mDesc.GetStrides().begin(), dy.mDesc.GetStrides().end()};
std::vector<index_t> strideX = strideDy;
std::vector<index_t> strideDx = strideDy;
std::vector<index_t> strideGamma = {0, 0, 0, C, 1};
std::vector<index_t> strideMeanInvStd = {G, 0, 0, 1, 0};
switch(init_method)
{
case 0:
dy.GenerateTensorValue(GeneratorTensor_1<DYDataType>{});
x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
mean.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
inv_std.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
dx.GenerateTensorValue(GeneratorTensor_1<DXDataType>{});
break;
case 1:
dy.GenerateTensorValue(GeneratorTensor_2<DYDataType>{-5, 5});
x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
mean.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{-5, 5});
inv_std.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{-5, 5});
dx.GenerateTensorValue(GeneratorTensor_2<DXDataType>{-5, 5});
break;
default:
dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0, 1});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{-0.5, 0.5});
inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{-0.5, 0.5});
dx.GenerateTensorValue(GeneratorTensor_3<DXDataType>{-0.5, 0.5});
}
DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
DeviceMem dx_dev(sizeof(DXDataType) * dx.mDesc.GetElementSpaceSize());
dy_dev.ToDevice(dy.mData.data());
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
mean_dev.ToDevice(mean.mData.data());
inv_std_dev.ToDevice(inv_std.mData.data());
// add device normalization instances
using DeviceOp = ck::tensor_operation::device::DeviceNormalizationBwdData<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DXDataType,
5,
3>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
if(do_verification)
{
using ReferenceInstance =
ck::tensor_operation::host::ReferenceGroupnormBwd<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
DXDataType,
ComputeDataType>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, length);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
strideDy,
strideX,
strideGamma,
strideMeanInvStd,
strideMeanInvStd,
strideDx,
reduce_dim,
dy_dev.GetDeviceBuffer(),
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
mean_dev.GetDeviceBuffer(),
inv_std_dev.GetDeviceBuffer(),
dx_dev.GetDeviceBuffer());
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
++num_kernel;
}
else
{
if(time_kernel)
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
}
continue;
}
size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes = dy.mDesc.GetElementSize() * sizeof(DYDataType) +
x.mDesc.GetElementSize() * sizeof(XDataType) +
gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
mean.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
inv_std.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
dx.mDesc.GetElementSize() * sizeof(DXDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
dx_dev.FromDevice(dx.mData.data());
bool pass = ck::utils::check_err(
dx.mData, host_dx.mData, "Error: Incorrect results", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "dy : ", dy.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "host_dx : ", host_dx.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "dx : ", dx.mData, ",") << std::endl;
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
return false;
}
else
{
if(time_kernel)
std::cout << "pass" << std::endl;
}
}
}
if(time_kernel)
{
LogRange(std::cout << "length = ", length, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dim, ",") << std::endl;
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s,"
<< best_instance_name << std::endl;
}
if(num_kernel == 0)
{
std::cout << "Error: No kernel is applicable" << std::endl;
return false;
}
return true;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/groupnorm_bwd_gamma_beta.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp"
namespace ck {
namespace profiler {
template <typename DYDataType,
typename XDataType,
typename MeanInvStdDataType,
typename ComputeDataType,
typename DGammaDataType,
typename DBetaDataType>
bool profile_groupnorm_bwd_gamma_beta_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> length)
{
// we don't need GammaDataType and DXDataType here, just for reference class
using GammaDataType = DYDataType;
using DXDataType = DYDataType;
if(length.size() != 5)
return false;
index_t N = length[0];
index_t G = length[3];
index_t C = length[4];
std::vector<index_t> reduce_dim = {0, 1, 2};
std::vector<index_t> gamma_beta_length = {G, C};
Tensor<DYDataType> dy(length);
Tensor<XDataType> x(length);
Tensor<GammaDataType> gamma(gamma_beta_length); // dummy tensor, for reference
Tensor<MeanInvStdDataType> mean({N, G});
Tensor<MeanInvStdDataType> inv_std({N, G});
Tensor<DGammaDataType> dgamma(gamma_beta_length);
Tensor<DBetaDataType> dbeta(gamma_beta_length);
Tensor<DXDataType> host_dx(length); // dummy tensor, for reference
Tensor<DGammaDataType> host_dgamma(gamma_beta_length);
Tensor<DBetaDataType> host_dbeta(gamma_beta_length);
std::vector<index_t> strideDy =
std::vector<ck::index_t>{dy.mDesc.GetStrides().begin(), dy.mDesc.GetStrides().end()};
std::vector<index_t> strideX =
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()};
std::vector<index_t> strideDGamma{dgamma.mDesc.GetStrides().begin(),
dgamma.mDesc.GetStrides().end()};
std::vector<index_t> strideDBeta{dbeta.mDesc.GetStrides().begin(),
dbeta.mDesc.GetStrides().end()};
std::vector<index_t> strideMeanInvStd = {G, 0, 0, 1, 0};
switch(init_method)
{
case 0:
dy.GenerateTensorValue(GeneratorTensor_1<DYDataType>{});
x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
mean.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
inv_std.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
dgamma.GenerateTensorValue(GeneratorTensor_1<DGammaDataType>{});
dbeta.GenerateTensorValue(GeneratorTensor_1<DBetaDataType>{});
break;
case 1:
dy.GenerateTensorValue(GeneratorTensor_2<DYDataType>{-5, 5});
x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
mean.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{-5, 5});
inv_std.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{0, 5});
dgamma.GenerateTensorValue(GeneratorTensor_2<DGammaDataType>{-5, 5});
dbeta.GenerateTensorValue(GeneratorTensor_2<DBetaDataType>{-5, 5});
break;
default:
dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0, 1});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{-0.5, 0.5});
inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0, 0.5});
dgamma.GenerateTensorValue(GeneratorTensor_3<DGammaDataType>{-0.5, 0.5});
dbeta.GenerateTensorValue(GeneratorTensor_3<DBetaDataType>{-0.5, 0.5});
}
DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
dy_dev.ToDevice(dy.mData.data());
x_dev.ToDevice(x.mData.data());
mean_dev.ToDevice(mean.mData.data());
inv_std_dev.ToDevice(inv_std.mData.data());
// add device normalization instances
using DeviceOp =
ck::tensor_operation::device::DeviceNormalizationBwdGammaBeta<DYDataType,
XDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
5,
3>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
if(do_verification)
{
using ReferenceInstance =
ck::tensor_operation::host::ReferenceGroupnormBwd<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
DXDataType,
ComputeDataType>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, length);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
std::size_t num_bytes = dy.mDesc.GetElementSize() * sizeof(DYDataType) +
x.mDesc.GetElementSize() * sizeof(XDataType) +
mean.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
inv_std.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
dgamma.mDesc.GetElementSize() * sizeof(DGammaDataType) +
dbeta.mDesc.GetElementSize() * sizeof(DBetaDataType);
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
strideDy,
strideX,
strideMeanInvStd,
strideMeanInvStd,
gamma_beta_length,
strideDGamma,
strideDBeta,
reduce_dim,
dy_dev.GetDeviceBuffer(),
x_dev.GetDeviceBuffer(),
mean_dev.GetDeviceBuffer(),
inv_std_dev.GetDeviceBuffer(),
dgamma_dev.GetDeviceBuffer(),
dbeta_dev.GetDeviceBuffer());
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
++num_kernel;
}
else
{
if(time_kernel)
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
}
continue;
}
size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
dgamma_dev.FromDevice(dgamma.mData.data());
dbeta_dev.FromDevice(dbeta.mData.data());
bool pass =
ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "dy : ", dy.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "host_dgamma : ", host_dgamma.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "dgamma : ", dgamma.mData, ",") << std::endl;
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
return false;
}
else
{
if(time_kernel)
std::cout << "pass" << std::endl;
}
}
}
if(time_kernel)
{
LogRange(std::cout << "length = ", length, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dim, ",") << std::endl;
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s,"
<< best_instance_name << std::endl;
}
if(num_kernel == 0)
{
std::cout << "Error: No kernel is applicable" << std::endl;
return false;
}
return true;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_layernorm_bwd_data_impl.hpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/layernorm_bwd_data.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp"
namespace ck {
namespace profiler {
template <typename DYDataType,
typename XDataType,
typename GammaDataType,
typename MeanInvStdDataType,
typename ComputeDataType,
typename DXDataType,
index_t Rank>
bool profile_layernorm_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> length)
{
// we don't need DGamma and DBeta here, just for reference class
using DGammaDataType = DXDataType;
using DBetaDataType = DXDataType;
if(length.size() != Rank || Rank < 2)
return false;
// Assume normalize dimension except for batch (first) dimension
std::vector<index_t> reduce_length{length.begin() + 1, length.end()};
std::vector<index_t> reduce_dim;
for(int i = 1; i < Rank; ++i)
reduce_dim.push_back(i);
Tensor<DYDataType> dy(length);
Tensor<XDataType> x(length);
Tensor<GammaDataType> gamma(reduce_length);
Tensor<MeanInvStdDataType> mean({length[0]});
Tensor<MeanInvStdDataType> inv_std({length[0]});
Tensor<DXDataType> dx(length);
Tensor<DXDataType> host_dx(length);
Tensor<DGammaDataType> host_dgamma(reduce_length);
Tensor<DBetaDataType> host_dbeta(reduce_length);
std::vector<index_t> strideDy =
std::vector<ck::index_t>{dy.mDesc.GetStrides().begin(), dy.mDesc.GetStrides().end()};
std::vector<index_t> strideX = strideDy;
std::vector<index_t> strideDx = strideDy;
std::vector<index_t> strideGamma = strideDy;
strideGamma[0] = 0;
std::vector<index_t> strideMeanInvStd{Rank, 0};
strideMeanInvStd[0] = 1;
switch(init_method)
{
case 0:
dy.GenerateTensorValue(GeneratorTensor_1<DYDataType>{});
x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
gamma.GenerateTensorValue(GeneratorTensor_1<GammaDataType>{});
mean.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
inv_std.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
dx.GenerateTensorValue(GeneratorTensor_1<DXDataType>{});
break;
case 1:
dy.GenerateTensorValue(GeneratorTensor_2<DYDataType>{-5, 5});
x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
gamma.GenerateTensorValue(GeneratorTensor_2<GammaDataType>{-5, 5});
mean.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{-5, 5});
inv_std.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{-5, 5});
dx.GenerateTensorValue(GeneratorTensor_2<DXDataType>{-5, 5});
break;
default:
dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0, 1});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{-0.5, 0.5});
mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{-0.5, 0.5});
inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{-0.5, 0.5});
dx.GenerateTensorValue(GeneratorTensor_3<DXDataType>{-0.5, 0.5});
}
DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
DeviceMem dx_dev(sizeof(DXDataType) * dx.mDesc.GetElementSpaceSize());
dy_dev.ToDevice(dy.mData.data());
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
mean_dev.ToDevice(mean.mData.data());
inv_std_dev.ToDevice(inv_std.mData.data());
constexpr int NumReduceDim = Rank - 1;
// add device normalization instances
using DeviceOp = ck::tensor_operation::device::DeviceNormalizationBwdData<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DXDataType,
Rank,
NumReduceDim>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
if(do_verification)
{
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernormBwd<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
DXDataType,
ComputeDataType>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, length);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
strideDy,
strideX,
strideGamma,
strideMeanInvStd,
strideMeanInvStd,
strideDx,
reduce_dim,
dy_dev.GetDeviceBuffer(),
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
mean_dev.GetDeviceBuffer(),
inv_std_dev.GetDeviceBuffer(),
dx_dev.GetDeviceBuffer());
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
++num_kernel;
}
else
{
if(time_kernel)
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
}
continue;
}
size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes = dy.mDesc.GetElementSize() * sizeof(DYDataType) +
x.mDesc.GetElementSize() * sizeof(XDataType) +
gamma.mDesc.GetElementSize() * sizeof(GammaDataType) +
mean.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
inv_std.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
dx.mDesc.GetElementSize() * sizeof(DXDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
dx_dev.FromDevice(dx.mData.data());
bool pass = ck::utils::check_err(
dx.mData, host_dx.mData, "Error: Incorrect results", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "dy : ", dy.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "host_dx : ", host_dx.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "dx : ", dx.mData, ",") << std::endl;
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
return false;
}
else
{
if(time_kernel)
std::cout << "pass" << std::endl;
}
}
}
if(time_kernel)
{
LogRange(std::cout << "length = ", length, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dim, ",") << std::endl;
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s,"
<< best_instance_name << std::endl;
}
if(num_kernel == 0)
{
std::cout << "Error: No kernel is applicable" << std::endl;
return false;
}
return true;
}
} // namespace profiler
} // namespace ck
profiler/include/profiler/profile_layernorm_bwd_gamma_beta_impl.hpp 0 → 100644
View file @ 29dcb956
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include "ck/ck.hpp"
#include "ck/library/tensor_operation_instance/gpu/layernorm_bwd_gamma_beta.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp"
namespace ck {
namespace profiler {
template <typename DYDataType,
typename XDataType,
typename MeanInvStdDataType,
typename ComputeDataType,
typename DGammaDataType,
typename DBetaDataType,
index_t Rank>
bool profile_layernorm_bwd_gamma_beta_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> length)
{
// we don't need GammaDataType and DXDataType here, just for reference class
using GammaDataType = DYDataType;
using DXDataType = DYDataType;
if(length.size() != Rank || Rank < 2)
return false;
// Assume normalize dimension for first dimension
// Layernorm 2D, input = [M, K], reduce on M axis
// Layernorm 4D, input = [N, H, W, C], redice on N axis
constexpr int NumReduceDim = Rank - 1;
std::vector<index_t> reduce_dim = {0};
std::vector<index_t> invarient_length{length.begin() + 1, length.end()};
Tensor<DYDataType> dy(length);
Tensor<XDataType> x(length);
Tensor<GammaDataType> gamma(invarient_length); // dummy tensor, for reference
Tensor<MeanInvStdDataType> mean({length[0]});
Tensor<MeanInvStdDataType> inv_std({length[0]});
Tensor<DGammaDataType> dgamma(invarient_length);
Tensor<DBetaDataType> dbeta(invarient_length);
Tensor<DXDataType> host_dx(length); // dummy tensor, for reference
Tensor<DGammaDataType> host_dgamma(invarient_length);
Tensor<DBetaDataType> host_dbeta(invarient_length);
std::vector<index_t> strideDy =
std::vector<ck::index_t>{dy.mDesc.GetStrides().begin(), dy.mDesc.GetStrides().end()};
std::vector<index_t> strideX = strideDy;
std::vector<index_t> strideDGamma{dgamma.mDesc.GetStrides().begin(),
dgamma.mDesc.GetStrides().end()};
std::vector<index_t> strideDBeta{dbeta.mDesc.GetStrides().begin(),
dbeta.mDesc.GetStrides().end()};
std::vector<index_t> strideMeanInvStd{Rank, 0};
strideMeanInvStd[0] = 1;
switch(init_method)
{
case 0:
dy.GenerateTensorValue(GeneratorTensor_1<DYDataType>{});
x.GenerateTensorValue(GeneratorTensor_1<XDataType>{});
mean.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
inv_std.GenerateTensorValue(GeneratorTensor_1<MeanInvStdDataType>{});
dgamma.GenerateTensorValue(GeneratorTensor_1<DGammaDataType>{});
dbeta.GenerateTensorValue(GeneratorTensor_1<DBetaDataType>{});
break;
case 1:
dy.GenerateTensorValue(GeneratorTensor_2<DYDataType>{-5, 5});
x.GenerateTensorValue(GeneratorTensor_2<XDataType>{-5, 5});
mean.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{-5, 5});
inv_std.GenerateTensorValue(GeneratorTensor_2<MeanInvStdDataType>{0, 5});
dgamma.GenerateTensorValue(GeneratorTensor_2<DGammaDataType>{-5, 5});
dbeta.GenerateTensorValue(GeneratorTensor_2<DBetaDataType>{-5, 5});
break;
default:
dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0, 1});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0, 1});
mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{-0.5, 0.5});
inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0, 0.5});
dgamma.GenerateTensorValue(GeneratorTensor_3<DGammaDataType>{-0.5, 0.5});
dbeta.GenerateTensorValue(GeneratorTensor_3<DBetaDataType>{-0.5, 0.5});
}
DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
dy_dev.ToDevice(dy.mData.data());
x_dev.ToDevice(x.mData.data());
mean_dev.ToDevice(mean.mData.data());
inv_std_dev.ToDevice(inv_std.mData.data());
// add device normalization instances
using DeviceOp =
ck::tensor_operation::device::DeviceNormalizationBwdGammaBeta<DYDataType,
XDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
Rank,
NumReduceDim>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;
if(do_verification)
{
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernormBwd<DYDataType,
XDataType,
GammaDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
DXDataType,
ComputeDataType>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, length);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
}
std::size_t num_bytes = dy.mDesc.GetElementSize() * sizeof(DYDataType) +
x.mDesc.GetElementSize() * sizeof(XDataType) +
mean.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
inv_std.mDesc.GetElementSize() * sizeof(MeanInvStdDataType) +
dgamma.mDesc.GetElementSize() * sizeof(DGammaDataType) +
dbeta.mDesc.GetElementSize() * sizeof(DBetaDataType);
int num_kernel = 0;
for(auto& inst_ptr : instance_ptrs)
{
auto argument_ptr = inst_ptr->MakeArgumentPointer(length,
strideDy,
strideX,
strideMeanInvStd,
strideMeanInvStd,
invarient_length,
strideDGamma,
strideDBeta,
reduce_dim,
dy_dev.GetDeviceBuffer(),
x_dev.GetDeviceBuffer(),
mean_dev.GetDeviceBuffer(),
inv_std_dev.GetDeviceBuffer(),
dgamma_dev.GetDeviceBuffer(),
dbeta_dev.GetDeviceBuffer());
if(inst_ptr->IsSupportedArgument(argument_ptr.get()))
{
++num_kernel;
}
else
{
if(time_kernel)
{
std::cout << inst_ptr->GetTypeString() << " skipped due to unsupported argument: ";
LogRange(std::cout << "input lengths = ", length, ", ") << std::endl;
}
continue;
}
size_t workspace_sz = inst_ptr->GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
inst_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = inst_ptr->MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
float gb_per_sec = num_bytes / 1.E6 / avg_time;
if(time_kernel)
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
<< inst_ptr->GetTypeString() << std::endl;
if(avg_time < best_avg_time)
{
best_instance_name = inst_ptr->GetTypeString();
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
dgamma_dev.FromDevice(dgamma.mData.data());
dbeta_dev.FromDevice(dbeta.mData.data());
bool pass =
ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "dy : ", dy.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "host_dgamma : ", host_dgamma.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "dgamma : ", dgamma.mData, ",") << std::endl;
}
if(!pass)
{
std::cout << inst_ptr->GetTypeString() << " failed verification: ";
LogRange(std::cout << "lengths = [", length, ", ") << "]." << std::endl;
return false;
}
else
{
if(time_kernel)
std::cout << "pass" << std::endl;
}
}
}
if(time_kernel)
{
LogRange(std::cout << "length = ", length, ",") << ", ";
LogRange(std::cout << "reduce dims ", reduce_dim, ",") << std::endl;
std::cout << "best perf = " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s,"
<< best_instance_name << std::endl;
}
if(num_kernel == 0)
{
std::cout << "Error: No kernel is applicable" << std::endl;
return false;
}
return true;
}
} // namespace profiler
} // 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