Skip to content

GitLab

"web/vscode:/vscode.git/clone" did not exist on "79ba0399d8d70bc655269fc3318455a70d14e180"
Unverified Commit bf98b476 authored by Bartłomiej Kocot's avatar Bartłomiej Kocot Committed by GitHub
Browse files

Add bilinear conv fwd and bwd data instances (#1164)

parent a78be3f6
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1093 additions and 71 deletions
example/62_convnd_activ/unary/convnd_fwd_xdl_clippedrelu_fp16.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::ClippedRelu;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_elu_fp16.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::Elu;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_leakyrelu_fp16.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::LeakyRelu;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_pow_fp16.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::Power;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_relu_fp16.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::Relu;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_sigmoid_fp16.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::Sigmoid;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_softrelu_fp16.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::SoftRelu;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
example/62_convnd_activ/unary/convnd_fwd_xdl_tanh_fp16.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include "convnd_fwd_activ_unary_common.hpp"
using OutElementOp = ck::tensor_operation::element_wise::TanH;
using DeviceGroupedConvNDActivInstance = DeviceGroupedConvNDFwdInstance<OutElementOp>;
#include "../run_convnd_activ_example.inc"
int main(int argc, char* argv[]) { return !run_convnd_example(argc, argv); }
include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
View file @ bf98b476
......@@ -165,7 +165,7 @@ struct Subtract
struct Bilinear
{
Bilinear(float alpha, float beta) : alpha_(alpha), beta_(beta){};
Bilinear(float alpha = 1.f, float beta = 1.f) : alpha_(alpha), beta_(beta){};
template <typename Y, typename X0, typename X1>
__host__ __device__ constexpr void operator()(Y&, const X0&, const X1&) const;
......@@ -184,6 +184,14 @@ struct Bilinear
y = alpha_ * x0 + beta_ * x1;
};
template <>
__host__ __device__ constexpr void
operator()<int8_t, int8_t, int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
{
y = type_convert<int8_t>(alpha_ * type_convert<float>(x0) +
beta_ * type_convert<float>(x1));
};
template <>
__host__ __device__ constexpr void
operator()<half_t, half_t, half_t>(half_t& y, const half_t& x0, const half_t& x1) const
......@@ -221,7 +229,8 @@ struct Bilinear
__host__ __device__ constexpr void operator()<std::int8_t, std::int32_t, std::int8_t>(
std::int8_t& y, const std::int32_t& x0, const std::int8_t& x1) const
{
y = type_convert<std::int8_t>(x0 + ck::type_convert<std::int32_t>(x1));
y = type_convert<int8_t>(alpha_ * type_convert<float>(x0) +
beta_ * type_convert<float>(x1));
};
float alpha_;
......
library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -25,25 +25,35 @@ template <ck::index_t NDimSpatial,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation,
ck::index_t NumAElementwiseTensor = 0,
ck::index_t NumBElementwiseTensor = 0,
ck::index_t NumDElementwiseTensor = 0,
typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
struct ReferenceConvBwdData : public device::BaseOperator
{
// Argument
struct Argument : public device::BaseArgument
{
Argument(Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight,
const Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
Argument(
Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight,
const Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors,
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors,
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors)
: input_{input},
weight_{weight},
output_{output},
elementwise_a_tensors_{elementwise_a_tensors},
elementwise_b_tensors_{elementwise_b_tensors},
elementwise_d_tensors_{elementwise_d_tensors},
conv_strides_{conv_filter_strides},
conv_dilations_{conv_filter_dilations},
in_left_pads_{input_left_pads},
......@@ -58,6 +68,10 @@ struct ReferenceConvBwdData : public device::BaseOperator
const Tensor<WeiDataType>& weight_;
const Tensor<OutDataType>& output_;
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors_;
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors_;
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors_;
std::vector<index_t> conv_strides_;
std::vector<index_t> conv_dilations_;
std::vector<index_t> in_left_pads_;
......@@ -106,26 +120,46 @@ struct ReferenceConvBwdData : public device::BaseOperator
{
for(std::size_t k = 0; k < K; ++k)
{
float v_out = 0;
float v_wei = 0;
arg.out_element_op_(
v_out, ck::type_convert<float>(arg.output_(g, n, k, wo)));
arg.wei_element_op_(
v_wei, ck::type_convert<float>(arg.weight_(g, k, c, x)));
v_acc += v_out * v_wei;
OutDataType v_out;
WeiDataType v_wei;
ExecuteElementwiseOp(arg.out_element_op_,
arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
v_out,
arg.output_(g, n, k, wo),
g,
n,
k,
wo);
ExecuteElementwiseOp(arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
arg.weight_(g, k, c, x),
g,
k,
c,
x);
v_acc += ck::type_convert<float>(v_out) *
ck::type_convert<float>(v_wei);
}
}
}
}
float v_in;
arg.in_element_op_(v_in, v_acc);
arg.input_(g, n, c, wi) = ck::type_convert<InDataType>(v_in);
InDataType v_acc_converted = ck::type_convert<InDataType>(v_acc);
InDataType& v_in = arg.input_(g, n, c, wi);
ExecuteElementwiseOp(arg.in_element_op_,
arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
v_in,
v_acc_converted,
g,
n,
c,
wi);
};
make_ParallelTensorFunctor(f_ncw,
......@@ -175,20 +209,34 @@ struct ReferenceConvBwdData : public device::BaseOperator
{
for(std::size_t k = 0; k < K; ++k)
{
float v_out = 0;
float v_wei = 0;
OutDataType v_out;
WeiDataType v_wei;
arg.out_element_op_(
ExecuteElementwiseOp(
arg.out_element_op_,
arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
v_out,
ck::type_convert<float>(
arg.output_(g, n, k, ho, wo)));
arg.wei_element_op_(
arg.output_(g, n, k, ho, wo),
g,
n,
k,
ho,
wo);
ExecuteElementwiseOp(
arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
ck::type_convert<float>(
arg.weight_(g, k, c, y, x)));
v_acc += v_out * v_wei;
arg.weight_(g, k, c, y, x),
g,
k,
c,
y,
x);
v_acc += ck::type_convert<float>(v_out) *
ck::type_convert<float>(v_wei);
}
}
}
......@@ -197,11 +245,18 @@ struct ReferenceConvBwdData : public device::BaseOperator
}
}
float v_in;
arg.in_element_op_(v_in, v_acc);
arg.input_(g, n, c, hi, wi) = ck::type_convert<InDataType>(v_in);
InDataType v_acc_converted = ck::type_convert<InDataType>(v_acc);
InDataType& v_in = arg.input_(g, n, c, hi, wi);
ExecuteElementwiseOp(arg.in_element_op_,
arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
v_in,
v_acc_converted,
g,
n,
c,
hi,
wi);
};
make_ParallelTensorFunctor(f_nchw,
......@@ -270,20 +325,37 @@ struct ReferenceConvBwdData : public device::BaseOperator
{
for(std::size_t k = 0; k < K; ++k)
{
float v_out = 0;
float v_wei = 0;
OutDataType v_out;
WeiDataType v_wei;
arg.out_element_op_(
ExecuteElementwiseOp(
arg.out_element_op_,
arg.elementwise_a_tensors_,
Number<NumAElementwiseTensor>{},
v_out,
ck::type_convert<float>(arg.output_(
g, n, k, do_, ho, wo)));
arg.wei_element_op_(
arg.output_(g, n, k, do_, ho, wo),
g,
n,
k,
do_,
ho,
wo);
ExecuteElementwiseOp(
arg.wei_element_op_,
arg.elementwise_b_tensors_,
Number<NumBElementwiseTensor>{},
v_wei,
ck::type_convert<float>(
arg.weight_(g, k, c, z, y, x)));
v_acc += v_out * v_wei;
arg.weight_(g, k, c, z, y, x),
g,
k,
c,
z,
y,
x);
v_acc +=
ck::type_convert<float>(v_out) *
ck::type_convert<float>(v_wei);
}
}
}
......@@ -295,11 +367,19 @@ struct ReferenceConvBwdData : public device::BaseOperator
}
}
float v_in;
arg.in_element_op_(v_in, v_acc);
arg.input_(g, n, c, di, hi, wi) = ck::type_convert<InDataType>(v_in);
InDataType v_acc_converted = ck::type_convert<InDataType>(v_acc);
InDataType& v_in = arg.input_(g, n, c, di, hi, wi);
ExecuteElementwiseOp(arg.in_element_op_,
arg.elementwise_d_tensors_,
Number<NumDElementwiseTensor>{},
v_in,
v_acc_converted,
g,
n,
c,
di,
hi,
wi);
};
make_ParallelTensorFunctor(f_ncdhw,
......@@ -325,6 +405,36 @@ struct ReferenceConvBwdData : public device::BaseOperator
}
};
template <typename... Args,
typename ElementwiseOp,
typename ElementwiseTensor,
typename NumTensor,
typename T>
static void ExecuteElementwiseOp(ElementwiseOp& elementwise_op,
ElementwiseTensor& elementwise_tensors,
NumTensor,
T& y,
const T& x,
Args... dims)
{
if constexpr(NumTensor::value == 0)
{
elementwise_op(y, x);
}
else if constexpr(NumTensor::value == 1)
{
elementwise_op(y, x, elementwise_tensors[0](dims...));
}
else if constexpr(NumTensor::value == 2)
{
elementwise_op(y, x, elementwise_tensors[0](dims...), elementwise_tensors[1](dims...));
}
else
{
throw std::runtime_error("ElementOp not supported in reference.");
}
}
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
......@@ -333,16 +443,20 @@ struct ReferenceConvBwdData : public device::BaseOperator
bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
static auto MakeArgument(Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight,
const Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
static auto MakeArgument(
Tensor<InDataType>& input,
const Tensor<WeiDataType>& weight,
const Tensor<OutDataType>& output,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
const std::array<Tensor<InDataType>, NumAElementwiseTensor>& elementwise_a_tensors = {},
const std::array<Tensor<WeiDataType>, NumBElementwiseTensor>& elementwise_b_tensors = {},
const std::array<Tensor<OutDataType>, NumDElementwiseTensor>& elementwise_d_tensors = {})
{
return Argument{input,
weight,
......@@ -353,7 +467,10 @@ struct ReferenceConvBwdData : public device::BaseOperator
input_right_pads,
in_element_op,
wei_element_op,
out_element_op};
out_element_op,
elementwise_a_tensors,
elementwise_b_tensors,
elementwise_d_tensors};
}
static auto MakeInvoker() { return Invoker{}; }
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data_bilinear.hpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
F32,
F32,
Tuple<F32>,
F32,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
template <ck::index_t NumDimSpatial,
typename OutLayout,
typename WeiLayout,
typename InLayout,
typename OutDataType,
typename WeiDataType,
typename InDataType,
typename ComputeTypeA,
typename ComputeTypeB>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<
NumDimSpatial,
OutLayout,
WeiLayout,
Tuple<InLayout>,
InLayout,
OutDataType,
WeiDataType,
Tuple<InDataType>,
InDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear,
ComputeTypeA,
ComputeTypeB>>
{
using DeviceOp =
DeviceGroupedConvBwdDataMultipleD<NumDimSpatial,
OutLayout,
WeiLayout,
Tuple<InLayout>,
InLayout,
OutDataType,
WeiDataType,
Tuple<InDataType>,
InDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear,
ComputeTypeA,
ComputeTypeB>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 3)
{
if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
is_same_v<OutLayout, NDHWGK>)
{
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16> && is_same_v<ComputeTypeA, F16> &&
is_same_v<ComputeTypeB, F16>)
{
add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_f16_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP32
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32> && is_same_v<ComputeTypeA, F32> &&
is_same_v<ComputeTypeB, F32>)
{
add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_f32_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16> && is_same_v<ComputeTypeA, BF16> &&
is_same_v<ComputeTypeB, BF16>)
{
add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
op_ptrs);
}
#endif
}
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bilinear.hpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <vector>
#include <memory>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
ck::Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<NDHWGK>,
NDHWGK,
F16,
F16,
ck::Tuple<F16>,
F16,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<NDHWGK>,
NDHWGK,
F32,
F32,
ck::Tuple<F32>,
F32,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
#ifdef CK_ENABLE_INT8
void add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<NDHWGK>,
NDHWGK,
int8_t,
int8_t,
ck::Tuple<int8_t>,
int8_t,
PassThrough,
PassThrough,
Bilinear>>>& instances);
#endif
template <ck::index_t NumDimSpatial,
typename InLayout,
typename WeiLayout,
typename DLayouts,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename DDataTypes,
typename OutDataType,
typename ComputeType>
struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<
NumDimSpatial,
InLayout,
WeiLayout,
DLayouts,
OutLayout,
InDataType,
WeiDataType,
DDataTypes,
OutDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear,
ComputeType>>
{
using DeviceOp =
DeviceGroupedConvFwdMultipleABD<NumDimSpatial,
InLayout,
WeiLayout,
DLayouts,
OutLayout,
InDataType,
WeiDataType,
DDataTypes,
OutDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::Bilinear,
ComputeType>;
static auto GetInstances()
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, NDHWGC> &&
is_same_v<WeiLayout, GKZYXC> && is_same_v<OutLayout, NDHWGK> &&
DLayouts::Size() == 1 && is_same_v<tuple_element_t<0, DLayouts>, NDHWGK>)
{
#ifdef CK_ENABLE_FP32
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t> && is_same_v<ComputeType, half_t>)
{
add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
is_same_v<WeiDataType, ck::bhalf_t> && is_same_v<OutDataType, ck::bhalf_t>)
{
add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instances(
op_ptrs);
}
#endif
}
return op_ptrs;
}
};
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/CMakeLists.txt 0 → 100644
View file @ bf98b476
set(GROUPED_CONV3D_BWD_DATA_BILINEAR
xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp)
add_instance_library(device_grouped_conv3d_bwd_data_bilinear_instance ${GROUPED_CONV3D_BWD_DATA_BILINEAR})
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
// g, k]
void add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
Bilinear>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_bilinear_bf16_instances<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(instances,
device_grouped_conv_bwd_data_xdl_bilinear_bf16_instances<
3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
// g, k]
void add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
F16,
F16,
Tuple<F16>,
F16,
PassThrough,
PassThrough,
Bilinear>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_bilinear_f16_instances<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_bilinear_f16_instances<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data_bilinear/xdl/device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_bilinear_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
// g, k]
void add_device_grouped_conv3d_bwd_data_xdl_bilinear_ndhwgk_gkzyxc_ndhwgc_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
F32,
F32,
Tuple<F32>,
F32,
PassThrough,
PassThrough,
Bilinear>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_bilinear_f32_instances<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_data_xdl_bilinear_f32_instances<3,
NDHWGK,
GKZYXC,
Tuple<NDHWGC>,
NDHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/CMakeLists.txt 0 → 100644
View file @ bf98b476
set(GROUPED_CONV3D_FWD_BILINEAR
xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp)
add_instance_library(device_grouped_conv3d_fwd_bilinear_instance ${GROUPED_CONV3D_FWD_BILINEAR})
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd_bilinear/xdl/device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp 0 → 100644
View file @ bf98b476
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_bilinear_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_fwd_xdl_bilinear_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
ck::Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
ck::Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
Bilinear>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_bilinear_bf16_instances<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_bilinear_bf16_instances<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
ConvFwd1x1P0>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_bilinear_bf16_instances<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
ConvFwd1x1S1P0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment