Skip to content

GitLab

Commit 0b11569f authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into batched_gemm_c_permute

parents e8d3a0fb fa9a0a5c
Hide whitespace changes
Inline Side-by-side
Showing with 174 additions and 23 deletions
include/ck/tensor_operation/gpu/device/device_normalization.hpp 0 → 100644
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <vector>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
struct DeviceNormalization : public BaseOperator
{
// inLengths: input tensor extent(s) from high to low dimension
// inStrides: input tensor stride(s) from high to low dimension
// reduceDims: the dimension(s) the normalization operation is applied
// alpha: typeless pointer in host memory storing the alpha scaling value of type AccDataType
// beta: typeless pointer in host memory storing the beta scaling value of type AccDataType
// in_dev: typeless const pointer in device memory storing the input tensor
// out_dev: typeless pointer in device memory storing the output tensor
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
const std::vector<int> reduceDims,
const void* alpha,
const void* beta,
const void* in_dev,
void* out_dev) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
virtual index_t GetRank() const = 0;
virtual index_t GetNumReduceDim() const = 0;
};
using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
......
include/ck/tensor_operation/gpu/device/device_pool2d_fwd_nhwc_nhwc.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
......
include/ck/tensor_operation/gpu/device/device_reduce.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
......
include/ck/tensor_operation/gpu/device/device_reduce_common.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
......
include/ck/tensor_operation/gpu/device/device_reduce_multiblock.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
......
include/ck/tensor_operation/gpu/device/device_reduce_threadwise.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
......
include/ck/tensor_operation/gpu/device/device_softmax.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
......@@ -6,6 +9,7 @@
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_softmax.hpp"
......@@ -30,8 +34,15 @@ template <typename InDataType,
index_t InSrcVectorDim,
index_t InSrcVectorSize,
index_t OutDstVectorSize>
struct DeviceSoftmax : public BaseOperator
struct DeviceSoftmax : public DeviceNormalization
{
static constexpr index_t kRank = Rank;
static constexpr index_t kNumReduceDim = NumReduceDim;
virtual index_t GetRank() const override { return kRank; }
virtual index_t GetNumReduceDim() const override { return kNumReduceDim; }
using PassThrough = tensor_operation::element_wise::PassThrough;
// Used for freeloading of some handy functions from DeviceReduceMultiBlock
......@@ -58,18 +69,33 @@ struct DeviceSoftmax : public BaseOperator
using GridDesc_M_K = decltype(Reduction::MakeSrc2dDescriptor({1}, {1}, 1, 1));
using GridwiseReduce = GridwiseSoftmax_mk_to_mk<InDataType,
OutDataType,
AccDataType,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
InSrcVectorDim,
InSrcVectorSize,
OutDstVectorSize>;
using GridwiseSoftmaxGeneric = GridwiseSoftmax_mk_to_mk<InDataType,
OutDataType,
AccDataType,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
InSrcVectorDim,
InSrcVectorSize,
OutDstVectorSize,
false>;
using GridwiseSoftmaxSweepOnce = GridwiseSoftmax_mk_to_mk<InDataType,
OutDataType,
AccDataType,
GridDesc_M_K,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
InSrcVectorDim,
InSrcVectorSize,
OutDstVectorSize,
true>;
struct Argument : public Reduction::Argument
{
......@@ -118,8 +144,19 @@ struct DeviceSoftmax : public BaseOperator
const auto out_grid_desc_m_k = Reduction::MakeSrc2dDescriptor(
arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
const auto kernel_main =
kernel_softmax<GridwiseReduce, InDataType, OutDataType, AccDataType, GridDesc_M_K>;
bool sweep_once =
in_grid_desc_m_k.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
const auto kernel_main = sweep_once ? kernel_softmax<GridwiseSoftmaxSweepOnce,
InDataType,
OutDataType,
AccDataType,
GridDesc_M_K>
: kernel_softmax<GridwiseSoftmaxGeneric,
InDataType,
OutDataType,
AccDataType,
GridDesc_M_K>;
float avg_time = 0;
......@@ -164,24 +201,34 @@ struct DeviceSoftmax : public BaseOperator
return true;
};
// inLengths: input tensor extent(s) from high to low dimension
// inStrides: input tensor stride(s) from high to low dimension
// reduceDims: the dimension(s) the softmax normalization operate on
// alpha: typeless pointer in host memory storing the alpha scaling value as type AccDataType
// beta: typeless pointer in host memory storing the beta scaling value as type AccDataType
// in_dev: typeless const pointer in device memory storing the input tensor
// out_dev: typeless pointer in device memory storing the output tensor
std::unique_ptr<BaseArgument> MakeArgumentPointer(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
const std::vector<int> reduceDims,
AccDataType alpha,
AccDataType beta,
const void* alpha,
const void* beta,
const void* in_dev,
void* out_dev)
void* out_dev) override
{
return std::make_unique<Argument>(inLengths,
inStrides,
reduceDims,
alpha,
beta,
*static_cast<const AccDataType*>(alpha),
*static_cast<const AccDataType*>(beta),
static_cast<const InDataType*>(in_dev),
static_cast<OutDataType*>(out_dev));
};
std::unique_ptr<BaseInvoker> MakeInvokerPointer() { return std::make_unique<Invoker>(); };
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>();
};
std::string GetTypeString() const override
{
......
include/ck/tensor_operation/gpu/device/device_unary_elementwise.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
......
include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
......
include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/reduction_operator.hpp"
......
include/ck/tensor_operation/gpu/device/tensor_layout.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
......
include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
......@@ -8,8 +11,8 @@ namespace element_wise {
struct Add
{
template <typename T>
__host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
template <typename Y, typename X0, typename X1>
__host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
template <>
__host__ __device__ constexpr void
......@@ -25,6 +28,13 @@ struct Add
y = x0 + x1;
};
template <>
__host__ __device__ constexpr void
operator()<half_t>(half_t& y, const float& x0, const half_t& x1) const
{
y = type_convert<half_t>(x0) + x1;
};
// Question: should half_t be supported ?
template <>
__host__ __device__ constexpr void
......
include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
......
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/math.hpp"
......
include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/reduction_common.hpp"
......
include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
......
include/ck/tensor_operation/gpu/grid/gridwise_5ary_Elementwise_1d.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_description/cluster_descriptor.hpp"
......
include/ck/tensor_operation/gpu/grid/gridwise_binary_elementwise_1d.hpp
View file @ 0b11569f
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
......
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