Skip to content

GitLab

Commit defa2071 authored by Adam Osewski's avatar Adam Osewski
Browse files

Merge branch 'develop' into aosewski/ggemm_multi_d2

parents 28a68428 f2398f61
Hide whitespace changes
Inline Side-by-side
Showing with 1131 additions and 146 deletions
example/62_conv_fwd_activ/run_convnd_fwd_activ_example.inc 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
void print_helper_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
bool run_convnd_fwd_example(int argc, char* argv[])
{
print_helper_msg();
bool do_verification = true;
// Use floats for SoftRelu by default to avoid overflow after e^x.
int init_method =
std::is_same_v<OutElementOp, ck::tensor_operation::element_wise::SoftRelu> ? 2 : 1;
bool time_kernel = false;
// Following shapes are selected to avoid overflow. Expect inf in case of
// size increase for some elementwise ops.
ck::utils::conv::ConvParam conv_param{
3, 1, 16, 128, 8, {3, 3, 3}, {17, 17, 17}, {2, 2, 2}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}};
if(argc == 1)
{
// use default
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv);
}
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
const auto run = [&]() {
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(
conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(
conv_param);
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(
conv_param);
return run_grouped_conv_fwd<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
DeviceGroupedConvNDFwdActivInstance>(do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
};
if(conv_param.num_dim_spatial_ == 3)
{
return run();
}
return false;
}
example/63_layernorm4d_fwd/CMakeLists.txt 0 → 100644
View file @ defa2071
add_example_executable(example_layernorm4d_fwd_fp16 layernorm4d_fwd_fp16.cpp)
add_example_executable(example_layernorm4d_fwd_splitk_fp16 layernorm4d_fwd_splitk_fp16.cpp)
example/63_layernorm4d_fwd/common.hpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_fwd_impl.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_normalization_fwd_splitk_impl.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_common_util.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_layernorm.hpp"
example/63_layernorm4d_fwd/layernorm4d_fwd_fp16.cpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationFwdImpl<XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // XYVectorDim (0=M, 1=K)
8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_layernorm4d_fwd_example.inc"
int main() { return run_layernorm4d_fwd_example<DeviceInstance>(); }
example/63_layernorm4d_fwd/layernorm4d_fwd_splitk_fp16.cpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
using DeviceInstance = ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl<
XDataType,
GammaDataType,
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim,
256, // BlockSize
8, // ClusterM
32, // ClusterK
1, // SliceM
8, // SliceK
1, // XYVectorDim (0=M, 1=K)
8, // XScalarPerVector
1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_layernorm4d_fwd_example.inc"
int main() { return run_layernorm4d_fwd_example<DeviceInstance>(); }
example/63_layernorm4d_fwd/run_layernorm4d_fwd_example.inc 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
template <typename DeviceInstance>
int run_layernorm4d_fwd_example()
{
bool time_kernel = false;
ck::index_t N = 256;
ck::index_t H = 16;
ck::index_t W = 16;
ck::index_t C = 8;
Tensor<XDataType> x({N, H, W, C});
Tensor<GammaDataType> gamma({H, W, C});
Tensor<BetaDataType> beta({H, W, C});
Tensor<YDataType> y({N, H, W, C});
Tensor<SaveMeanInvStdDataType> save_mean({N});
Tensor<SaveMeanInvStdDataType> save_inv_std({N});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
beta.GenerateTensorValue(GeneratorTensor_3<BetaDataType>{0.0, 1.0});
DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
#ifdef SAVE_MEAN_INV_STD
DeviceMem save_mean_dev(sizeof(SaveMeanInvStdDataType) * save_mean.mDesc.GetElementSpaceSize());
DeviceMem save_inv_std_dev(sizeof(SaveMeanInvStdDataType) *
save_inv_std.mDesc.GetElementSpaceSize());
#endif
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
beta_dev.ToDevice(beta.mData.data());
auto device_instance = DeviceInstance{};
auto argument_ptr = device_instance.MakeArgumentPointer(
{N, H, W, C},
std::vector<ck::index_t>{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()},
{0, W * C, C, 1},
{0, W * C, C, 1},
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
{1, 2, 3},
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_dev.GetDeviceBuffer(),
save_inv_std_dev.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
PassThrough{});
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
{
std::cout << "The runtime parameters are not supported" << std::endl;
return 1;
};
size_t workspace_sz = device_instance.GetWorkSpaceSize(argument_ptr.get());
DeviceMem workspace_dev(workspace_sz);
device_instance.SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
auto invoker_ptr = device_instance.MakeInvokerPointer();
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
bool pass = true;
{
Tensor<YDataType> host_y({N, H, W, C});
Tensor<SaveMeanInvStdDataType> host_save_mean({N});
Tensor<SaveMeanInvStdDataType> host_save_inv_std({N});
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
ComputeDataType,
PassThrough,
Rank,
NumReduceDim>;
ReferenceInstance ref;
auto ref_argument = ref.MakeArgument(x,
gamma,
beta,
host_y,
host_save_mean,
host_save_inv_std,
PassThrough{},
{N, H, W, C},
{1, 2, 3},
1e-4);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data());
pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results (y)", 1e-3, 1e-3);
#ifdef SAVE_MEAN_INV_STD
save_mean_dev.FromDevice(save_mean.mData.data());
save_inv_std_dev.FromDevice(save_inv_std.mData.data());
pass &= ck::utils::check_err(
save_mean, host_save_mean, "Error: Incorrect results (mean)", 1e-3, 1e-3);
pass &= ck::utils::check_err(
save_inv_std, host_save_inv_std, "Error: Incorrect results (inv_std)", 1e-3, 1e-3);
#endif
}
return (pass ? 0 : 1);
}
example/CMakeLists.txt
View file @ defa2071
......@@ -11,31 +11,27 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
if(DEFINED DTYPES)
foreach(source IN LISTS FILE_NAME)
set(test 0)
foreach(type IN LISTS DTYPES)
if(type MATCHES "fp16")
set(type1 "_f16")
elseif(type MATCHES "fp32")
set(type1 "_f32")
elseif(type MATCHES "fp8")
set(type1 "_f8")
elseif(type MATCHES "bf16")
set(type1 "_b16")
elseif(type MATCHES "fp64")
set(type1 "_f64")
elseif(type MATCHES "int8")
set(type1 "_i8")
endif()
if("${source}" MATCHES "${type}" OR "${source}" MATCHES "${type1}")
#if filename matches any selected type, exit type loop and do no exclude the file from the list
set(test 0)
break()
elseif((source MATCHES "fp8" OR source MATCHES "fp32" OR source MATCHES "fp64" OR source MATCHES "bf16" OR source MATCHES "int8" OR source MATCHES "fp16" OR
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
NOT(source MATCHES type OR source MATCHES type1))
#if filename contains a type which doesn't match any selected type, mark it for removal
set(test 1)
endif()
endforeach()
if((source MATCHES "_fp16" OR source MATCHES "_f16") AND NOT "fp16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp32" OR source MATCHES "_f32") AND NOT "fp32" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp64" OR source MATCHES "_f64") AND NOT "fp64" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp8" OR source MATCHES "_f8") AND NOT "fp8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf8" OR source MATCHES "_bf8") AND NOT "bf8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf16" OR source MATCHES "_b16") AND NOT "bf16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_int8" OR source MATCHES "_i8") AND NOT "int8" IN_LIST DTYPES)
set(test 1)
endif()
if(test EQUAL 1)
message("removing example source file ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
......@@ -74,31 +70,27 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
if(DEFINED DTYPES)
foreach(source IN LISTS FILE_NAME)
set(test 0)
foreach(type IN LISTS DTYPES)
if(type MATCHES "fp16")
set(type1 "_f16")
elseif(type MATCHES "fp32")
set(type1 "_f32")
elseif(type MATCHES "fp8")
set(type1 "_f8")
elseif(type MATCHES "bf16")
set(type1 "_b16")
elseif(type MATCHES "fp64")
set(type1 "_f64")
elseif(type MATCHES "int8")
set(type1 "_i8")
endif()
if("${source}" MATCHES "${type}" OR "${source}" MATCHES "${type1}")
#if filename matches any selected type, exit type loop and do no exclude the file from the list
set(test 0)
break()
elseif((source MATCHES "fp8" OR source MATCHES "fp32" OR source MATCHES "fp64" OR source MATCHES "bf16" OR source MATCHES "int8" OR source MATCHES "fp16" OR
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
NOT(source MATCHES type OR source MATCHES type1))
#if filename contains a type which doesn't match any selected type, mark it for removal
set(test 1)
endif()
endforeach()
if((source MATCHES "_fp16" OR source MATCHES "_f16") AND NOT "fp16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp32" OR source MATCHES "_f32") AND NOT "fp32" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp64" OR source MATCHES "_f64") AND NOT "fp64" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_fp8" OR source MATCHES "_f8") AND NOT "fp8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf8" OR source MATCHES "_bf8") AND NOT "bf8" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_bf16" OR source MATCHES "_b16") AND NOT "bf16" IN_LIST DTYPES)
set(test 1)
endif()
if((source MATCHES "_int8" OR source MATCHES "_i8") AND NOT "int8" IN_LIST DTYPES)
set(test 1)
endif()
if(test EQUAL 1)
message("removing example ${source} ")
list(REMOVE_ITEM FILE_NAME "${source}")
......
include/ck/ck.hpp
View file @ defa2071
......@@ -66,6 +66,10 @@
#define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8
#elif defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__)
#define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8_GFX11
#endif
// MFMA instruction
......
include/ck/host_utility/hip_check_error.hpp
View file @ defa2071
......@@ -3,8 +3,10 @@
#pragma once
#include <sstream>
#include <hip/hip_runtime.h>
// To be removed, which really does not tell the location of failed HIP functional call
inline void hip_check_error(hipError_t x)
{
if(x != hipSuccess)
......@@ -15,3 +17,16 @@ inline void hip_check_error(hipError_t x)
throw std::runtime_error(ss.str());
}
}
#define HIP_CHECK_ERROR(retval_or_funcall) \
do \
{ \
hipError_t _tmpVal = retval_or_funcall; \
if(_tmpVal != hipSuccess) \
{ \
std::ostringstream ostr; \
ostr << "HIP Function Failed (" << __FILE__ << "," << __LINE__ << ") " \
<< hipGetErrorString(_tmpVal); \
throw std::runtime_error(ostr.str()); \
} \
} while(0)
include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
View file @ defa2071
......@@ -33,7 +33,8 @@ template <index_t NumDimM,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
typename CDEElementwiseOperation,
typename ComputeDataType = ADataType>
struct DeviceContractionMultipleD : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
......
include/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp
View file @ defa2071
......@@ -14,11 +14,12 @@ namespace device {
/**
* \brief Convolution Tensor Rearrange.
*
* This Device operator supports conversion image ([G, N, Di, Hi, Wi, C]) to
* the gemm problem([N * Do * Ho * Wo, Z * Y * X * C]) (Image to Column) and
* conversion gemm form to the image (Column to Image).
*
* Note that G must be equal to 1.
* This Device operator supports converting an image to
* the GEMM representation (Image to Column) and
* converting a GEMM form to the image (Column to Image).
* Supported layouts:
* [G, N, Di, Hi, Wi, C] <-> [G, N * Do * Ho * Wo, Z * Y * X * C]
* [N, Di, Hi, Wi, G, C] <-> [N * Do * Ho * Wo, G, Z * Y * X * C]
*
* \tparam NDimSpatial Number of spatial dimensions.
* \tparam ImageLayout Input Layout.
......@@ -39,13 +40,14 @@ struct DeviceConvTensorRearrange : public BaseOperator
*
* \param p_in A pointer to the device memory of the input image.
* \param p_out A pointer to the device memory of the output.
* \param G Convolution number of groups.
* \param N Convolution batch size.
* \param C Convolution number of channels.
* \param input_spatial_lengths Input spatial lengths.
* \param filter_spatial_lengths Filter spatial lengths.
* \param output_spatial_lengths Output spatial lengths.
* \param image_g_n_c_wis_strides Image strides in order [G, N, C, D, H, W].
* \param gemm_m_k_strides Gemm form strides.
* \param gemm_g_m_k_strides Gemm form strides.
* \param conv_filter_strides Convolution filter strides.
* \param conv_filter_dilations Convolution filter dilations.
* \param input_left_pads Convolution left pads.
......@@ -55,13 +57,14 @@ struct DeviceConvTensorRearrange : public BaseOperator
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in,
void* p_out,
const ck::index_t G,
const ck::index_t N,
const ck::index_t C,
const std::array<index_t, NDimSpatial>& input_spatial_lengths,
const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<index_t, NDimSpatial>& output_spatial_lengths,
const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_strides,
const std::array<index_t, 2>& gemm_m_k_strides,
const std::array<index_t, 3>& gemm_g_m_k_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
......
include/ck/tensor_operation/gpu/device/device_elementwise_scale.hpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <memory>
#include <array>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename InDataTypeTuple,
typename OutDataTypeTuple,
typename ElementwiseOperation,
typename UnaryOperation,
typename Scale,
index_t NumDim>
struct DeviceElementwise : public BaseOperator
{
static constexpr int NumInput = InDataTypeTuple::Size();
static constexpr int NumOutput = OutDataTypeTuple::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
const std::array<const void*, NumInput> in_dev_buffers,
const std::array<void*, NumOutput> out_dev_buffers,
ElementwiseOperation elementwise_op,
UnaryOperation unary_op,
Scale scale_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
}; // namespace device
template <typename InDataTypeTuple,
typename OutDataTypeTuple,
typename ElementwiseOperation,
typename UnaryOperation,
typename Scale,
index_t NumDim>
using DeviceElementwisePtr = std::unique_ptr<DeviceElementwise<InDataTypeTuple,
OutDataTypeTuple,
ElementwiseOperation,
UnaryOperation,
Scale,
NumDim>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <array>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/utility/is_detected.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename T>
using is_tuple = decltype(std::declval<T&>().IsTuple());
/**
* \brief Grouped Convolution Forward
*
* \details
* input : input image A[G, N, C, Hi, Wi], A1[G, N, C, Hi, Wi]...
* input : weight B[G, K, C, Y, X], B1[G, K, C, Y, X]...
* input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
* output : output image E[G, N, K, Ho, Wo]
*
* C = a_op(A, A1...) * b_op(B, B1...)
* E = cde_op(C, D0, D1, ...)
*
* \tparam NDimSpatial Number of spatial dimensions.
* \tparam ALayout Input layout (also for a1, a2...).
* \tparam BLayout Weight layout (also for b1, b2...).
* \tparam DsLayout Ds layouts.
* \tparam ELayout Output layout.
* \tparam ADataType Input data type. Pass tuple if there is multiple A.
* \tparam BDataType Weight data type. Pass tuple if there is multiple B.
* \tparam DsDataType D data types.
* \tparam EDataType Output data type.
* \tparam AElementwiseOperation A elementwise operation.
* \tparam BElementwiseOperation B elementwise operation.
* \tparam CDEElementwiseOperation CDE elementwise operation.
* \tparam ComputeType Compute data type (default: ADataType, first if tuple passed).
*/
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename ComputeType =
decltype(UnpackDataType<is_detected<is_tuple, ADataType>::value,
Number<0>,
ADataType>())> // ComputeType is InputType by default (first
// in tuple for MultiAB), unpack if tuple was
// passed
struct DeviceGroupedConvFwdMultipleABD : public BaseOperator
{
static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value;
static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value;
static constexpr index_t NumATensor = GetNumABTensors<isMultiA, ADataType>();
static constexpr index_t NumBTensor = GetNumABTensors<isMultiB, BDataType>();
static constexpr index_t NumDTensor = DsDataType::Size();
static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
// If DataType is tuple, user has to pass std::array with pointers.
using APointers =
std::conditional_t<isMultiA, std::array<const void*, NumATensor>&, const void*>;
using BPointers =
std::conditional_t<isMultiB, std::array<const void*, NumBTensor>&, const void*>;
/**
* \brief Make argument pointer for grouped conv fwd.
*
* \param p_a A pointer to the input (std::array<const void*, NumA> with
pointers for multiple A).
* \param p_b A pointer to the weight (std::array<const void*, NumA> with
pointers for multiple B).
* \param p_ds A pointers to the Ds.
* \param p_e A pointers to the output.
* \param a_g_n_c_wis_lengths Input lengths [G, N, C, Spatial...] (for 3d).
* \param a_g_n_c_wis_strides Input strides [G, N, C, Spatial...] (for 3d).
* \param b_g_k_c_xs_lengths Weight lengths [G, K, C, Spatial...] (for 3d).
* \param b_g_k_c_xs_strides Weight strides [G, K, C, Spatial...] (for 3d).
* \param ds_g_n_k_wos_lengths Ds lengths [G, N, K, Spatial...] (for 3d).
* \param ds_g_n_k_wos_strides Ds strides [G, N, K, Spatial...] (for 3d).
* \param e_g_n_k_wos_lengths Output lengths [G, N, K, Spatial...] (for 3d).
* \param e_g_n_k_wos_strides Output strides [G, N, K, Spatial...] (for 3d).
* \param conv_filter_strides Convolution filter strides.
* \param conv_filter_dilations Convolution filter dilations.
* \param input_left_pads Input left paddings.
* \param input_right_pads Input right paddings.
* \param a_element_op A elementwise operation object.
* \param b_element_op B elementwise operation object.
* \param cde_element_op CDE elementwise operation object.
* \return Pointer to the argument.
*/
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
APointers p_a,
BPointers p_b,
const std::array<const void*, NumDTensor>& p_ds,
void* p_e,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
View file @ defa2071
......@@ -3,21 +3,33 @@
#pragma once
#include <array>
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// Convolution Forward:
// input : input image A[G, N, C, Hi, Wi],
// input : weight B[G, K, C, Y, X],
// input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
// output : output image E[G, N, K, Ho, Wo]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
/**
* \brief Grouped Convolution Forward
*
* \note This structure is deprecated (left for backwards compatibility). Please use
* DeviceGroupedConvFwdMultipleABD.
*
* \tparam NDimSpatial Number of spatial dimensions.
* \tparam ALayout Input layout (also for a1, a2...).
* \tparam BLayout Weight layout (also for b1, b2...).
* \tparam DsLayout Ds layouts.
* \tparam ELayout Output layout.
* \tparam ADataType Input data type. Pass tuple if there is multiple A.
* \tparam BDataType Weight data type. Pass tuple if there is multiple B.
* \tparam DsDataType D data types.
* \tparam EDataType Output data type.
* \tparam AElementwiseOperation A elementwise operation.
* \tparam BElementwiseOperation B elementwise operation.
* \tparam CDEElementwiseOperation CDE elementwise operation.
* \tparam ComputeType Compute data type (default: ADataType, first if tuple passed).
*/
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
......@@ -30,36 +42,25 @@ template <index_t NDimSpatial,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
typename ComputeType = ADataType>
struct DeviceGroupedConvFwdMultipleD : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
static_assert(NumDTensor == DsLayout::Size(), "wrong! Inconsistent NumDTensor");
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
const void* p_a, // input image
const void* p_b, // weight
const std::array<const void*, NumDTensor>& p_ds,
void* p_e, // output image
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
typename ComputeType =
decltype(UnpackDataType<is_detected<is_tuple, ADataType>::value,
Number<0>,
ADataType>())> // ComputeType is InputType by default (first
// in tuple for MultiAB), unpack if tuple was
// passed
using DeviceGroupedConvFwdMultipleD = DeviceGroupedConvFwdMultipleABD<NDimSpatial,
ALayout,
BLayout,
DsLayout,
ELayout,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation,
ComputeType>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/device_normalization_bwd_gamma_beta.hpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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 {
template <typename DYDataType,
typename XDataType,
typename MeanInvStdDataType,
typename DGammaDataType,
typename DBetaDataType,
index_t Rank,
index_t NumReduceDim>
struct DeviceNormalizationBwdGammaBeta : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const std::vector<index_t> inLengths,
const std::vector<index_t> dyStrides,
const std::vector<index_t> xStrides,
const std::vector<index_t> meanStrides,
const std::vector<index_t> invStdStrides,
const std::vector<index_t> outLengths,
const std::vector<index_t> dgammaStrides,
const std::vector<index_t> dbetaStrides,
const std::vector<index_t> reduceDims,
const void* p_dy,
const void* p_x,
const void* p_mean,
const void* p_invStd,
void* p_dgamma,
void* p_dbeta) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename DYDataType,
typename XDataType,
typename MeanInvStdDataType,
typename DGammaDataType,
typename DBetaDataType,
index_t Rank,
index_t NumReduceDim>
using DeviceNormalizationBwdGammaBetaPtr =
std::unique_ptr<DeviceNormalizationBwdGammaBeta<DYDataType,
XDataType,
MeanInvStdDataType,
DGammaDataType,
DBetaDataType,
Rank,
NumReduceDim>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/device_normalization.hpp → include/ck/tensor_operation/gpu/device/device_normalization_fwd.hpp
View file @ defa2071
......@@ -19,7 +19,7 @@ template <typename XDataType,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim>
struct DeviceNormalization : public BaseOperator
struct DeviceNormalizationFwd : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const std::vector<index_t> lengths,
......@@ -50,14 +50,14 @@ template <typename XDataType,
typename YElementwiseOperation,
index_t Rank,
index_t NumReduceDim>
using DeviceNormalizationPtr = std::unique_ptr<DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
YElementwiseOperation,
Rank,
NumReduceDim>>;
using DeviceNormalizationFwdPtr = std::unique_ptr<DeviceNormalizationFwd<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
YElementwiseOperation,
Rank,
NumReduceDim>>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
View file @ defa2071
......@@ -17,15 +17,18 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
#include "ck/host_utility/io.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
// Image to column for input layout NDHWC:
// input : image converted to the gemm problem [N * Do * Ho * Wo, Z * Y * X * C]
// output : image [N, Di, Hi, Wi, C]
// Column to Image:
// input : gemm form [G, N * Do * Ho * Wo, Z * Y * X * C]
// output : input image [G, N, Di, Hi, Wi, C]
// input : gemm form [N * Do * Ho * Wo, G, Z * Y * X * C]
// output : input image [N, Di, Hi, Wi, G, C]
template <index_t NDimSpatial,
typename ImageLayout,
typename InputDataType,
......@@ -43,6 +46,14 @@ struct DeviceColumnToImageImpl
OutputDataType,
conv_tensor_rearrange_op::ColumnToImage>
{
static constexpr bool is_NSpatialGC =
std::is_same_v<ImageLayout, tensor_layout::convolution::NWGC> ||
std::is_same_v<ImageLayout, tensor_layout::convolution::NHWGC> ||
std::is_same_v<ImageLayout, tensor_layout::convolution::NDHWGC>;
static constexpr bool is_GNSpatialC =
std::is_same_v<ImageLayout, tensor_layout::convolution::GNWC> ||
std::is_same_v<ImageLayout, tensor_layout::convolution::GNHWC> ||
std::is_same_v<ImageLayout, tensor_layout::convolution::GNDHWC>;
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -90,7 +101,7 @@ struct DeviceColumnToImageImpl
const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<index_t, NDimSpatial>& output_spatial_lengths,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, 2>& gemm_m_k_strides,
const std::array<index_t, 3>& gemm_g_m_k_strides,
const std::array<index_t, NDimSpatial>& independent_filters,
const std::array<index_t, NDimSpatial>& effs)
{
......@@ -100,23 +111,23 @@ struct DeviceColumnToImageImpl
C * ck::accumulate_n<index_t>(
filter_spatial_lengths.begin(), NDimSpatial, 1, std::multiplies<>());
const index_t NStride = DoHoWo * gemm_m_k_strides[I0] * gemm_m_k_strides[I1];
const index_t NStride = DoHoWo * gemm_g_m_k_strides[I1] * gemm_g_m_k_strides[I2];
// Calculate the appropriate stride for each set of independent filters
// in each dimension
const index_t WStride =
math::integer_divide_ceil(effs[XIdx], conv_filter_strides[XIdx]) * gemm_m_k_strides[I0];
const index_t WStride = math::integer_divide_ceil(effs[XIdx], conv_filter_strides[XIdx]) *
gemm_g_m_k_strides[I1];
const index_t HStride = math::integer_divide_ceil(effs[YIdx], conv_filter_strides[YIdx]) *
output_spatial_lengths[XIdx] * gemm_m_k_strides[I0];
output_spatial_lengths[XIdx] * gemm_g_m_k_strides[I1];
const index_t DStride = math::integer_divide_ceil(effs[ZIdx], conv_filter_strides[ZIdx]) *
output_spatial_lengths[YIdx] * output_spatial_lengths[XIdx] *
gemm_m_k_strides[I0];
gemm_g_m_k_strides[I1];
// Create descriptor for independent filters in each dimension and
// then merge them into column form
if constexpr(NDimSpatial == 1)
{
const auto desc_gemm_form =
make_naive_tensor_descriptor(make_tuple(N, independent_filters[XIdx], CZYX),
make_tuple(NStride, WStride, gemm_m_k_strides[I1]));
make_tuple(NStride, WStride, gemm_g_m_k_strides[I2]));
const auto desc_gemm_form_merged_filters = transform_tensor_descriptor(
desc_gemm_form,
make_tuple(make_merge_transform(make_tuple(N, independent_filters[XIdx])),
......@@ -130,7 +141,7 @@ struct DeviceColumnToImageImpl
{
const auto desc_gemm_form = make_naive_tensor_descriptor(
make_tuple(N, independent_filters[YIdx], independent_filters[XIdx], CZYX),
make_tuple(NStride, HStride, WStride, gemm_m_k_strides[I1]));
make_tuple(NStride, HStride, WStride, gemm_g_m_k_strides[I2]));
const auto desc_gemm_form_merged_filters = transform_tensor_descriptor(
desc_gemm_form,
make_tuple(make_merge_transform(
......@@ -149,7 +160,7 @@ struct DeviceColumnToImageImpl
independent_filters[YIdx],
independent_filters[XIdx],
CZYX),
make_tuple(NStride, DStride, HStride, WStride, gemm_m_k_strides[I1]));
make_tuple(NStride, DStride, HStride, WStride, gemm_g_m_k_strides[I2]));
const auto desc_gemm_form_merged_filters = transform_tensor_descriptor(
desc_gemm_form,
make_tuple(make_merge_transform(make_tuple(N,
......@@ -262,24 +273,27 @@ struct DeviceColumnToImageImpl
ThreadClusterLengths,
ScalarPerVector,
InMemoryDataOperationEnum::Add,
Block2ETileMap>;
Block2ETileMap,
ComputePtrOffsetOfStridedBatch<>>;
struct Argument : public BaseArgument
{
Argument(const void* p_in, // input image
void* p_out, // output image
const ck::index_t G,
const ck::index_t N,
const ck::index_t C,
const std::array<index_t, NDimSpatial>& input_spatial_lengths,
const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<index_t, NDimSpatial>& output_spatial_lengths,
const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_strides,
const std::array<index_t, 2>& gemm_m_k_strides,
const std::array<index_t, 3>& gemm_g_m_k_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
const std::array<index_t, NDimSpatial>& input_right_pads)
: C_(C),
: G_(G),
C_(C),
X_(filter_spatial_lengths[NDimSpatial - I1]),
p_in_{static_cast<const InputDataType*>(p_in)},
p_out_{static_cast<OutputDataType*>(p_out)},
......@@ -289,6 +303,9 @@ struct DeviceColumnToImageImpl
input_left_pads_{input_left_pads},
input_right_pads_{input_right_pads}
{
compute_ptr_offset_of_batch_.BatchStrideA_ = gemm_g_m_k_strides[I0];
compute_ptr_offset_of_batch_.BatchStrideC_ = image_g_n_c_wis_strides[I0];
const index_t x_eff =
(filter_spatial_lengths[XIdx] - 1) * conv_filter_dilations[XIdx] + 1;
const index_t y_eff =
......@@ -354,7 +371,7 @@ struct DeviceColumnToImageImpl
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
gemm_m_k_strides,
gemm_g_m_k_strides,
independent_filters,
effs);
const auto out_grid_desc_m_k =
......@@ -387,10 +404,9 @@ struct DeviceColumnToImageImpl
// Memory offsets to next set of independent filters,
// move to independent filters in each dimension
const index_t in_offset =
x_idx * gemm_m_k_strides[0] +
y_idx * gemm_m_k_strides[0] * output_spatial_lengths[XIdx] +
z_idx * gemm_m_k_strides[0] * output_spatial_lengths[YIdx] *
output_spatial_lengths[XIdx];
(x_idx + y_idx * output_spatial_lengths[XIdx] +
z_idx * output_spatial_lengths[YIdx] * output_spatial_lengths[XIdx]) *
gemm_g_m_k_strides[I1];
// Move to independent filters in appropriate dimensions
const index_t out_offset =
x_offset_with_pad * image_g_n_c_wis_strides[spatial_offset + XIdx] +
......@@ -417,6 +433,7 @@ struct DeviceColumnToImageImpl
}
}
const ck::index_t G_;
const ck::index_t C_;
const ck::index_t X_;
......@@ -434,6 +451,8 @@ struct DeviceColumnToImageImpl
std::vector<const InputDataType*> p_in_container_;
std::vector<OutputDataType*> p_out_container_;
ComputePtrOffsetOfStridedBatch<> compute_ptr_offset_of_batch_;
};
struct Invoker : public BaseInvoker
......@@ -451,6 +470,7 @@ struct DeviceColumnToImageImpl
OutputGridDesc,
OutputDataType,
Block2ETileMap,
ComputePtrOffsetOfStridedBatch<>,
GridwiseTensorRearrangeKernel>;
// Execute each set of independent filters
......@@ -460,7 +480,7 @@ struct DeviceColumnToImageImpl
BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, InputGridDesc>(
arg.out_grid_desc_m_k_container_[i]);
const index_t grid_size =
block_2_tile_map.CalculateGridSize(arg.in_grid_desc_m_k_container_[i]);
block_2_tile_map.CalculateGridSize(arg.in_grid_desc_m_k_container_[i]) * arg.G_;
elapsed_time += launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
......@@ -470,7 +490,9 @@ struct DeviceColumnToImageImpl
arg.p_in_container_[i],
arg.out_grid_desc_m_k_container_[i],
arg.p_out_container_[i],
block_2_tile_map);
arg.G_,
block_2_tile_map,
arg.compute_ptr_offset_of_batch_);
}
return elapsed_time;
}
......@@ -485,8 +507,7 @@ struct DeviceColumnToImageImpl
bool IsSupportedArgument(const Argument& arg)
{
using namespace tensor_layout::convolution;
if constexpr(!(std::is_same_v<ImageLayout, GNWC> || std::is_same_v<ImageLayout, GNHWC> ||
std::is_same_v<ImageLayout, GNDHWC>))
if constexpr(!(is_NSpatialGC || is_GNSpatialC))
{
return false;
}
......@@ -534,13 +555,14 @@ struct DeviceColumnToImageImpl
static auto MakeArgument(const void* p_in, // input image
void* p_out, // output image
const ck::index_t G,
const ck::index_t N,
const ck::index_t C,
const std::array<index_t, NDimSpatial>& input_spatial_lengths,
const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<index_t, NDimSpatial>& output_spatial_lengths,
const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_strides,
const std::array<index_t, 2>& gemm_m_k_strides,
const std::array<index_t, 3>& gemm_g_m_k_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
......@@ -548,13 +570,14 @@ struct DeviceColumnToImageImpl
{
return Argument{static_cast<const InputDataType*>(p_in),
static_cast<OutputDataType*>(p_out),
G,
N,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
image_g_n_c_wis_strides,
gemm_m_k_strides,
gemm_g_m_k_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......@@ -566,13 +589,14 @@ struct DeviceColumnToImageImpl
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in, // input image
void* p_out, // output image
const ck::index_t G,
const ck::index_t N,
const ck::index_t C,
const std::array<index_t, NDimSpatial>& input_spatial_lengths,
const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
const std::array<index_t, NDimSpatial>& output_spatial_lengths,
const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_strides,
const std::array<index_t, 2>& gemm_m_k_strides,
const std::array<index_t, 3>& gemm_g_m_k_strides,
const std::array<index_t, NDimSpatial>& conv_filter_strides,
const std::array<index_t, NDimSpatial>& conv_filter_dilations,
const std::array<index_t, NDimSpatial>& input_left_pads,
......@@ -580,13 +604,14 @@ struct DeviceColumnToImageImpl
{
return std::make_unique<Argument>(static_cast<const InputDataType*>(p_in),
static_cast<OutputDataType*>(p_out),
G,
N,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
image_g_n_c_wis_strides,
gemm_m_k_strides,
gemm_g_m_k_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
View file @ defa2071
......@@ -385,9 +385,11 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
// desc for blockwise copy
using AsGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeAsGridDescriptor_AK0_M_AK1(AsGridDesc_M_K{}))>;
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAsGridDescriptor_AK0_M_AK1(
AsGridDesc_M_K{}))>;
using BsGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(GridwiseGemm::MakeBsGridDescriptor_BK0_N_BK1(BsGridDesc_N_K{}))>;
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBsGridDescriptor_BK0_N_BK1(
BsGridDesc_N_K{}))>;
using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
DsGridDesc_M_N{}))>;
......@@ -397,7 +399,7 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
// block-to-e-tile map
using Block2ETileMap =
remove_cvref_t<decltype(GridwiseGemm::MakeBlock2ETileMap(EGridDesc_M_N{}))>;
remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
// Argument
struct Argument : public BaseArgument
......@@ -429,7 +431,7 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
bs_grid_desc_bk0_n_bk1_{},
ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
e_grid_desc_mblock_mperblock_nblock_nperblock_{},
block_2_etile_map_{GridwiseGemm::MakeBlock2ETileMap(e_grid_desc_m_n_)},
block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
cde_element_op_{cde_element_op}
......@@ -481,10 +483,10 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
block_2_etile_map_))
{
as_grid_desc_ak0_m_ak1_ =
GridwiseGemm::MakeAsGridDescriptor_AK0_M_AK1(as_grid_desc_m_k_);
GridwiseGemm::MakeDefaultAsGridDescriptor_AK0_M_AK1(as_grid_desc_m_k_);
bs_grid_desc_bk0_n_bk1_ =
GridwiseGemm::MakeBsGridDescriptor_BK0_N_BK1(bs_grid_desc_n_k_);
GridwiseGemm::MakeDefaultBsGridDescriptor_BK0_N_BK1(bs_grid_desc_n_k_);
ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
View file @ defa2071
......@@ -145,7 +145,8 @@ template <index_t NumDimM,
index_t CShuffleNXdlPerWavePerShuffle,
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()>
typename ComputeDataType = ADataType,
LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceContractionMultipleD_Xdl_CShuffle
: public DeviceContractionMultipleD<NumDimM,
NumDimN,
......@@ -156,7 +157,8 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
EDataType,
AElementwiseOperation,
BElementwiseOperation,
CDEElementwiseOperation>
CDEElementwiseOperation,
ComputeDataType>
{
using DeviceOp = DeviceContractionMultipleD_Xdl_CShuffle;
......@@ -310,8 +312,6 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({{}}, {{}}))>;
using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N({}, {}));
using ComputeDataType = ADataType;
// GridwiseGemm
using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
ADataType, // TODO: distinguish A/B datatype
......@@ -595,7 +595,9 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
return false;
}
if(ck::get_device_name() != "gfx90a" && std::is_same<ADataType, double>::value)
if(ck::get_device_name() != "gfx90a" && ck::get_device_name() != "gfx940" &&
ck::get_device_name() != "gfx941" && ck::get_device_name() != "gfx942" &&
std::is_same<ADataType, double>::value)
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp 0 → 100644
View file @ defa2071
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/math.hpp"
#include "ck/utility/sequence.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_3d.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/host_utility/stream_utility.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename InDataTypeTuple,
typename OutDataTypeTuple,
typename ElementwiseOperation,
index_t NumDim_m, // choose how to set dims
index_t NumDim_n,
index_t NumDim_k,
index_t MPerThread,
index_t NPerThread,
index_t KPerThread,
typename InScalarPerVectorSeq,
typename OutScalarPerVectorSeq>
struct DeviceElementwise3dImpl : public DeviceElementwise<InDataTypeTuple,
OutDataTypeTuple,
ElementwiseOperation,
NumDim_m + NumDim_n + NumDim_k>
{
static constexpr index_t NumDim = NumDim_m + NumDim_n + NumDim_k;
static constexpr int NumInput = InDataTypeTuple::Size();
static constexpr int NumOutput = OutDataTypeTuple::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static_assert(NumInput == InScalarPerVectorSeq::Size() &&
NumOutput == OutScalarPerVectorSeq::Size(),
"Tuple size is inconsistent with the number of in/out!");
static auto GenerateInDataTypePointerTuple()
{
return generate_tuple(
[&](auto I) {
using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
return static_cast<const DataType*>(nullptr);
},
Number<NumInput>{});
}
static auto GenerateOutDataTypePointerTuple()
{
return generate_tuple(
[&](auto I) {
using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
return static_cast<DataType*>(nullptr);
},
Number<NumOutput>{});
}
using InDataTypePointerTuple = decltype(GenerateInDataTypePointerTuple());
using OutDataTypePointerTuple = decltype(GenerateOutDataTypePointerTuple());
template <typename Desc_MNK>
static auto PadDescriptor_MNK(Desc_MNK desc_mnk,
index_t gridSize,
index_t blockSize,
index_t num_threads_m,
index_t num_threads_n,
index_t num_threads_k)
{
std::ignore = blockSize;
std::ignore = gridSize;
const auto m = desc_mnk.GetLength(I0);
const auto n = desc_mnk.GetLength(I1);
const auto k = desc_mnk.GetLength(I2);
const index_t loop_step_m = num_threads_m * MPerThread;
const index_t loop_step_n = num_threads_n * NPerThread;
const index_t loop_step_k = num_threads_k * KPerThread;
const auto pad_m = math::integer_least_multiple(m, loop_step_m) - m;
const auto pad_n = math::integer_least_multiple(n, loop_step_n) - n;
const auto pad_k = math::integer_least_multiple(k, loop_step_k) - k;
const auto desc_mnk_pad =
transform_tensor_descriptor(desc_mnk,
make_tuple(make_right_pad_transform(m, pad_m),
make_right_pad_transform(n, pad_n),
make_right_pad_transform(k, pad_k)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return desc_mnk_pad;
}
static auto MakeDescriptor_MNK(const std::array<index_t, NumDim>& lengths,
const std::array<index_t, NumDim>& stride,
index_t gridSize,
index_t blockSize,
index_t num_threads_m,
index_t num_threads_n,
index_t num_threads_k)
{
auto tupleOfShape = generate_tuple([&](auto I) { return lengths[I]; }, Number<NumDim>{});
auto tupleOfStride = generate_tuple([&](auto I) { return stride[I]; }, Number<NumDim>{});
// nd desc - [s0, s1, s2, ...]
const auto desc = make_naive_tensor_descriptor(tupleOfShape, tupleOfStride);
constexpr auto mDimIds = typename arithmetic_sequence_gen<0, NumDim_m, 1>::type();
constexpr auto nDimIds =
typename arithmetic_sequence_gen<NumDim_m, NumDim_m + NumDim_n, 1>::type();
constexpr auto kDimIds =
typename arithmetic_sequence_gen<NumDim_m + NumDim_n, NumDim, 1>::type();
const auto mLengths = get_container_subset(tupleOfShape, mDimIds);
const auto nLengths = get_container_subset(tupleOfShape, nDimIds);
const auto kLengths = get_container_subset(tupleOfShape, kDimIds);
// merge nd to 3d desc - [s0 * s1 * ...]
if constexpr(NumDim > 3)
{
const auto desc_mnk = transform_tensor_descriptor(
desc,
make_tuple(make_merge_transform(mLengths),
make_merge_transform(nLengths),
make_merge_transform(kLengths)),
make_tuple(mDimIds, nDimIds, kDimIds),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
return PadDescriptor_MNK(
desc_mnk, gridSize, blockSize, num_threads_m, num_threads_n, num_threads_k);
}
else
return PadDescriptor_MNK(
desc, gridSize, blockSize, num_threads_m, num_threads_n, num_threads_k);
}
template <index_t TupleSize>
static auto GenerateInOutGrid3dDescTuple(Number<TupleSize>)
{
return generate_tuple(
[&](auto) {
if constexpr(NumDim > 3)
{
return MakeDescriptor_MNK({1, 1, 1}, {1, 1, 1}, 1, 1, 1, 1, 1);
}
else
{
return MakeDescriptor_MNK({1}, {1}, 1, 1, 1, 1, 1);
};
},
Number<TupleSize>{});
}
using OutGrid3dDescTuple = decltype(GenerateInOutGrid3dDescTuple(Number<NumOutput>{}));
using InGrid3dDescTuple = decltype(GenerateInOutGrid3dDescTuple(Number<NumInput>{}));
using GridwiseElementwise = GridwiseElementwise_3D<InGrid3dDescTuple,
OutGrid3dDescTuple,
InDataTypePointerTuple,
OutDataTypePointerTuple,
ElementwiseOperation,
MPerThread,
NPerThread,
KPerThread,
InScalarPerVectorSeq,
OutScalarPerVectorSeq>;
struct Argument : public BaseArgument
{
Argument(const std::array<index_t, NumDim> lengths,
const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
const std::array<const void*, NumInput> in_dev_buffers,
const std::array<void*, NumOutput> out_dev_buffers,
ElementwiseOperation elementwise_op)
: lengths_(lengths),
inStridesArray_(inStridesArray),
outStridesArray_(outStridesArray),
elementwise_op_(elementwise_op),
blockSize_(256)
{
static_assert(NumDim_m > 0, "");
static_assert(NumDim_n > 0, "");
static_assert(NumDim_k > 0, "");
in_dev_buffers_ = generate_tuple(
[&](auto I) {
using DataType = remove_cvref_t<decltype(InDataTypeTuple{}[I])>;
return static_cast<const DataType*>(in_dev_buffers[I.value]);
},
Number<NumInput>{});
out_dev_buffers_ = generate_tuple(
[&](auto I) {
using DataType = remove_cvref_t<decltype(OutDataTypeTuple{}[I])>;
return static_cast<DataType*>(out_dev_buffers[I.value]);
},
Number<NumOutput>{});
}
InDataTypePointerTuple in_dev_buffers_;
OutDataTypePointerTuple out_dev_buffers_;
std::array<index_t, NumDim> lengths_;
std::array<std::array<index_t, NumDim>, NumInput> inStridesArray_;
std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray_;
ElementwiseOperation elementwise_op_;
index_t blockSize_;
};
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
index_t gridSize = getAvailableComputeUnitCount(stream_config) * arg.blockSize_;
index_t num_threads_m = gridSize / (16 * 16);
index_t num_threads_n = 16;
index_t num_threads_k = 16;
auto in_grid_3d_desc_tuple = generate_tuple(
[&](auto I) {
return MakeDescriptor_MNK(arg.lengths_,
arg.inStridesArray_[I.value],
gridSize,
arg.blockSize_,
num_threads_m,
num_threads_n,
num_threads_k);
},
Number<NumInput>{});
auto out_grid_3d_desc_tuple = generate_tuple(
[&](auto I) {
return MakeDescriptor_MNK(arg.lengths_,
arg.outStridesArray_[I.value],
gridSize,
arg.blockSize_,
num_threads_m,
num_threads_n,
num_threads_k);
},
Number<NumOutput>{});
const auto kernel = kernel_elementwise_3d<GridwiseElementwise,
InGrid3dDescTuple,
OutGrid3dDescTuple,
InDataTypePointerTuple,
OutDataTypePointerTuple,
ElementwiseOperation>;
float elapsed_time = launch_and_time_kernel(stream_config,
kernel,
dim3(gridSize),
dim3(arg.blockSize_),
0,
in_grid_3d_desc_tuple,
out_grid_3d_desc_tuple,
arg.in_dev_buffers_,
arg.out_dev_buffers_,
arg.elementwise_op_,
num_threads_m,
num_threads_n,
num_threads_k);
return elapsed_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
if(pArg == nullptr)
return false;
if(pArg->lengths_.back() % MPerThread != 0)
return false;
auto IsScalarPerVectorValid = [&](const std::array<index_t, NumDim>& lengths,
const std::array<index_t, NumDim>& strides,
index_t scalarPerVector,
index_t vectorDim) {
if(strides[vectorDim] == 1 &&
(lengths[vectorDim] % scalarPerVector == 0 ||
lengths[vectorDim] % scalarPerVector == lengths[vectorDim]))
{
return true;
}
if(strides[vectorDim] >= scalarPerVector)
{
return true;
}
return false;
};
bool valid = true;
static_for<0, NumInput, 1>{}([&](auto I) {
valid = valid && IsScalarPerVectorValid(pArg->lengths_,
pArg->inStridesArray_[I.value],
InScalarPerVectorSeq::At(I),
NumDim_m - 1);
});
static_for<0, NumOutput, 1>{}([&](auto I) {
valid = valid && IsScalarPerVectorValid(pArg->lengths_,
pArg->outStridesArray_[I.value],
OutScalarPerVectorSeq::At(I),
NumDim - 1);
});
return valid;
}
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const std::array<index_t, NumDim> lengths,
const std::array<std::array<index_t, NumDim>, NumInput> inStridesArray,
const std::array<std::array<index_t, NumDim>, NumOutput> outStridesArray,
const std::array<const void*, NumInput> in_dev_buffers,
const std::array<void*, NumOutput> out_dev_buffers,
ElementwiseOperation elementwise_op) override
{
return std::make_unique<Argument>(lengths,
inStridesArray,
outStridesArray,
in_dev_buffers,
out_dev_buffers,
elementwise_op);
}
static auto MakeInvoker() { return Invoker{}; }
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>();
}
}; // namespace device
} // namespace device
} // namespace tensor_operation
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment