Skip to content

GitLab

"...resnet50_tensorflow.git" did not exist on "91fafd7bc77f2bb809f98cf325781db236e1150f"
Commit 87101e7d authored by Artur Wojcik's avatar Artur Wojcik
Browse files

Merge branch 'uif2-initial' into uif2-migraphx

parents d305c079 87e80971
Hide whitespace changes
Inline Side-by-side
Showing with 1255 additions and 4 deletions
client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 87101e7d
...@@ -13,3 +13,7 @@ target_compile_features(client_grouped_conv3d_bwd_weight_fp16 PRIVATE cxx_std_17 ...@@ -13,3 +13,7 @@ target_compile_features(client_grouped_conv3d_bwd_weight_fp16 PRIVATE cxx_std_17
add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp) add_executable(client_grouped_conv3d_bwd_weight_fp32 grouped_conv3d_bwd_weight_fp32.cpp)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations) target_link_libraries(client_grouped_conv3d_bwd_weight_fp32 PRIVATE composable_kernel::device_operations)
target_compile_features(client_grouped_conv3d_bwd_weight_fp32 PRIVATE cxx_std_17) target_compile_features(client_grouped_conv3d_bwd_weight_fp32 PRIVATE cxx_std_17)
add_executable(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp)
target_link_libraries(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE composable_kernel::device_operations)
target_compile_features(client_grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8 PRIVATE cxx_std_17)
client_example/11_grouped_conv_bwd_weight/common.hpp
View file @ 87101e7d
...@@ -85,7 +85,9 @@ template <ck::index_t NumDimSpatial, ...@@ -85,7 +85,9 @@ template <ck::index_t NumDimSpatial,
typename OutDataType, typename OutDataType,
typename InLayout, typename InLayout,
typename WeiLayout, typename WeiLayout,
typename OutLayout> typename OutLayout,
typename AComputeType = InDataType,
typename BComputeType = AComputeType>
bool run_grouped_conv_bwd_weight( bool run_grouped_conv_bwd_weight(
const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths, const std::array<ck::index_t, NumDimSpatial + 3>& input_lengths,
const std::array<ck::index_t, NumDimSpatial + 3>& input_strides, const std::array<ck::index_t, NumDimSpatial + 3>& input_strides,
...@@ -113,7 +115,9 @@ bool run_grouped_conv_bwd_weight( ...@@ -113,7 +115,9 @@ bool run_grouped_conv_bwd_weight(
OutDataType, OutDataType,
PassThrough, PassThrough,
PassThrough, PassThrough,
PassThrough>; PassThrough,
AComputeType,
BComputeType>;
// get device op instances // get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances(); DeviceOp>::GetInstances();
......
client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16_comp_bf8_fp8.cpp 0 → 100644
View file @ 87101e7d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using InLayout = ck::tensor_layout::convolution::NDHWGC;
using WeiLayout = ck::tensor_layout::convolution::GKZYXC;
using OutLayout = ck::tensor_layout::convolution::NDHWGK;
using AComputeType = ck::bf8_t;
using BComputeType = ck::f8_t;
static constexpr ck::index_t NumDimSpatial = 3;
static constexpr ck::index_t G = 8;
static constexpr ck::index_t N = 64;
static constexpr ck::index_t K = 128;
static constexpr ck::index_t C = 128;
static constexpr ck::index_t Z = 3;
static constexpr ck::index_t Y = 3;
static constexpr ck::index_t X = 3;
static constexpr ck::index_t Di = 28;
static constexpr ck::index_t Hi = 28;
static constexpr ck::index_t Wi = 3;
static constexpr ck::index_t Do = 28;
static constexpr ck::index_t Ho = 28;
static constexpr ck::index_t Wo = 3;
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_lengths{G, N, C, Di, Hi, Wi};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> filter_lengths{G, K, C, Z, Y, X};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_lengths{G, N, K, Do, Ho, Wo};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> input_strides{
N * Di * Hi * Wi * C, Di* Hi* Wi* C, 1, Hi* Wi* C, Wi* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> weights_strides{
K * Z * Y * X * C, Z* Y* X* C, 1, Y* X* C, X* C, C};
static constexpr std::array<ck::index_t, NumDimSpatial + 3> output_strides{
N * Do * Ho * Wo * K, Do* Ho* Wo* K, 1, Ho* Wo* K, Wo* K, K};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_strides{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> conv_filter_dilations{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1, 1};
static constexpr std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1, 1};
int main()
{
return run_grouped_conv_bwd_weight<NumDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InLayout,
WeiLayout,
OutLayout,
AComputeType,
BComputeType>(input_lengths,
input_strides,
filter_lengths,
weights_strides,
output_lengths,
output_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads)
? EXIT_SUCCESS
: EXIT_FAILURE;
}
example/44_elementwise_permute/CMakeLists.txt
View file @ 87101e7d
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp) add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp) add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
add_example_executable(example_elementwise_permute_4D_fp32_row elementwise_permute_4D_fp32_row.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_row elementwise_permute_4D_fp16_row.cpp)
add_example_executable(example_elementwise_permute_4D_fp32_col elementwise_permute_4D_fp32_col.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_col elementwise_permute_4D_fp16_col.cpp)
add_example_executable(example_elementwise_permute elementwise_permute.cpp) add_example_executable(example_elementwise_permute elementwise_permute.cpp)
add_example_executable(example_elementwise_permute_3d elementwise_permute_3d.cpp) add_example_executable(example_elementwise_permute_3d elementwise_permute_3d.cpp)
example/44_elementwise_permute/elementwise_permute_4D_fp16_col.cpp 0 → 100644
View file @ 87101e7d
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
8, // MPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
std::size_t N = A_nchw.mDesc.GetLengths()[0];
std::size_t C = A_nchw.mDesc.GetLengths()[1];
std::size_t H = A_nchw.mDesc.GetLengths()[2];
std::size_t W = A_nchw.mDesc.GetLengths()[3];
for(std::size_t w = 0; w < W; ++w)
for(std::size_t h = 0; h < H; ++h)
for(std::size_t c = 0; c < C; ++c)
for(std::size_t n = 0; n < N; ++n)
{
ADataType tmp_val;
// auto a_val = A_nchw(n, c, h, w);
auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
functor_b(tmp_val, a_val);
// functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {4, 2, 1, 8};
std::vector<std::size_t> nhwc = {4, 1, 8, 2};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 1.f;
auto i = 0;
for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
{
a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
(h * nchw[3]) + w] = i;
i++;
}
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {1,
static_cast<int>(nchw[0]),
static_cast<int>(nchw[0] * nchw[1]),
static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
std::array<ck::index_t, 4> b_strides = {1,
static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
static_cast<int>(nhwc[0]),
static_cast<int>(nhwc[0] * nhwc[1])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp16_row.cpp 0 → 100644
View file @ 87101e7d
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F16;
using BDataType = F16;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
8, // MPerThread
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
{
ADataType tmp_val;
auto a_val = A_nchw(n, c, h, w);
functor_b(tmp_val, a_val);
functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {16, 128, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 128};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 2.f;
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
static_cast<int>(nchw[2] * nchw[3]),
static_cast<int>(nchw[3]),
1};
std::array<ck::index_t, 4> b_strides = {static_cast<int>(nhwc[1] * nhwc[2] * nhwc[3]),
1,
static_cast<int>(nhwc[2] * nhwc[3]),
static_cast<int>(nhwc[3])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp32_col.cpp 0 → 100644
View file @ 87101e7d
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F32;
using BDataType = F32;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
1, // MPerThread
ck::Sequence<1>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
std::size_t N = A_nchw.mDesc.GetLengths()[0];
std::size_t C = A_nchw.mDesc.GetLengths()[1];
std::size_t H = A_nchw.mDesc.GetLengths()[2];
std::size_t W = A_nchw.mDesc.GetLengths()[3];
for(std::size_t w = 0; w < W; ++w)
for(std::size_t h = 0; h < H; ++h)
for(std::size_t c = 0; c < C; ++c)
for(std::size_t n = 0; n < N; ++n)
{
ADataType tmp_val;
auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
functor_b(tmp_val, a_val);
functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {5, 4, 2, 3};
std::vector<std::size_t> nhwc = {5, 2, 3, 4};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 1.f;
auto i = 0;
for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
{
a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
(h * nchw[3]) + w] = i;
i++;
}
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {1,
static_cast<int>(nchw[0]),
static_cast<int>(nchw[0] * nchw[1]),
static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
std::array<ck::index_t, 4> b_strides = {1,
static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
static_cast<int>(nhwc[0]),
static_cast<int>(nhwc[0] * nhwc[1])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
example/44_elementwise_permute/elementwise_permute_4D_fp32_row.cpp 0 → 100644
View file @ 87101e7d
#include <iostream>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
#include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
using F16 = ck::half_t;
using F32 = float;
using ADataType = F32;
using BDataType = F32;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
using Scale = ck::tensor_operation::element_wise::Scale;
using DeviceElementwisePermuteInstance =
ck::tensor_operation::device::DeviceElementwiseImpl<ck::Tuple<ADataType>, // InDataTypeTuple
ck::Tuple<BDataType>, // OutDataTypeTuple
PassThrough, // ElementwiseOp
UnaryOp, // UnaryOp
Scale, // Scalar
4, // NumDim
8, // MPerThread
ck::Sequence<8>, // InScalarPerVectorSeq
ck::Sequence<1>>; // OutScalarPerVectorSeq
template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
void host_elementwise4D(HostTensorB& B_nhwc,
const HostTensorA& A_nchw,
FunctorA functor_a,
FunctorB functor_b,
float scale)
{
for(std::size_t n = 0; n < A_nchw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_nchw.mDesc.GetLengths()[1]; ++c)
for(std::size_t h = 0; h < A_nchw.mDesc.GetLengths()[2]; ++h)
for(std::size_t w = 0; w < A_nchw.mDesc.GetLengths()[3]; ++w)
{
ADataType tmp_val;
auto a_val = A_nchw(n, c, h, w);
functor_b(tmp_val, a_val);
functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
}
}
int main()
{
bool do_verification = true;
bool time_kernel = true;
std::vector<std::size_t> nchw = {16, 128, 32, 64};
std::vector<std::size_t> nhwc = {16, 32, 64, 128};
Tensor<ADataType> a(nchw);
Tensor<BDataType> b(nhwc);
float scale = 2.f;
a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
std::array<ck::index_t, 4> ab_lengths;
std::array<ck::index_t, 4> a_strides = {static_cast<int>(nchw[1] * nchw[2] * nchw[3]),
static_cast<int>(nchw[2] * nchw[3]),
static_cast<int>(nchw[3]),
1};
std::array<ck::index_t, 4> b_strides = {static_cast<int>(nhwc[1] * nhwc[2] * nhwc[3]),
1,
static_cast<int>(nhwc[2] * nhwc[3]),
static_cast<int>(nhwc[3])};
ck::ranges::copy(nchw, ab_lengths.begin());
auto broadcastPermute = DeviceElementwisePermuteInstance{};
auto argument = broadcastPermute.MakeArgumentPointer(ab_lengths,
{a_strides},
{b_strides},
input,
output,
PassThrough{},
UnaryOp{},
Scale{scale});
if(!broadcastPermute.IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"The runtime parameters seems not supported by the device instance, exiting!");
};
std::cout << "A (nchw): " << a.mDesc << std::endl;
std::cout << "B (nhwc): " << b.mDesc << std::endl;
auto broadcastPermute_invoker_ptr = broadcastPermute.MakeInvokerPointer();
float ave_time =
broadcastPermute_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
Tensor<BDataType> host_b(nhwc);
host_elementwise4D(host_b, a, PassThrough{}, UnaryOp{}, scale);
pass &=
ck::utils::check_err(b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
}
return pass ? 0 : 1;
}
include/ck/tensor_operation/gpu/device/device_elementwise_scale.hpp 0 → 100644
View file @ 87101e7d
// 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/impl/device_elementwise_scale_impl.hpp 0 → 100644
View file @ 87101e7d
// 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_scale.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_elementwise_1d_scale.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,
typename UnaryOperation,
typename Scale,
index_t NumDim,
index_t MPerThread,
typename InScalarPerVectorSeq,
typename OutScalarPerVectorSeq>
struct DeviceElementwiseImpl : public DeviceElementwise<InDataTypeTuple,
OutDataTypeTuple,
ElementwiseOperation,
UnaryOperation,
Scale,
NumDim>
{
static constexpr int NumInput = InDataTypeTuple::Size();
static constexpr int NumOutput = OutDataTypeTuple::Size();
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_M>
static auto PadDescriptor_M_1d(Desc_M desc_m, index_t gridSize, index_t blockSize)
{
constexpr auto I0 = Number<0>{};
const auto m = desc_m.GetLength(I0);
const index_t loop_step = gridSize * blockSize * MPerThread;
const auto pad = math::integer_least_multiple(m, loop_step) - m;
const auto desc_m_pad =
transform_tensor_descriptor(desc_m,
make_tuple(make_right_pad_transform(m, pad)),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0>{}));
return desc_m_pad;
}
static auto MakeDescriptor_M(const std::array<index_t, NumDim>& lengths,
const std::array<index_t, NumDim>& stride,
index_t gridSize,
index_t blockSize)
{
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);
// merge nd to 1d desc - [s0 * s1 * ...]
if constexpr(NumDim > 1)
{
const auto desc_m = transform_tensor_descriptor(
desc,
make_tuple(make_merge_transform(tupleOfShape)),
make_tuple(generate_sequence_v2([&](auto I) { return I; }, Number<NumDim>{})),
make_tuple(Sequence<0>{}));
return PadDescriptor_M_1d(desc_m, gridSize, blockSize);
}
else
return PadDescriptor_M_1d(desc, gridSize, blockSize);
}
template <index_t TupleSize>
static auto GenerateInOutGrid1dDescTuple(Number<TupleSize>)
{
return generate_tuple(
[&](auto) {
if constexpr(NumDim > 1)
{
return MakeDescriptor_M({1, 1}, {1, 1}, 1, 1);
}
else
{
return MakeDescriptor_M({1}, {1}, 1, 1);
};
},
Number<TupleSize>{});
};
using InGrid1dDescTuple = decltype(GenerateInOutGrid1dDescTuple(Number<NumInput>{}));
using OutGrid1dDescTuple = decltype(GenerateInOutGrid1dDescTuple(Number<NumOutput>{}));
using GridwiseElementwise = GridwiseElementwise_1D<InGrid1dDescTuple,
OutGrid1dDescTuple,
InDataTypePointerTuple,
OutDataTypePointerTuple,
ElementwiseOperation,
UnaryOperation,
Scale,
MPerThread,
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,
UnaryOperation unary_op,
Scale scale_op)
: lengths_(lengths),
inStridesArray_(inStridesArray),
outStridesArray_(outStridesArray),
elementwise_op_(elementwise_op),
unary_op_(unary_op),
scale_op_(scale_op),
blockSize_(256)
{
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_;
UnaryOperation unary_op_;
Scale scale_op_;
index_t blockSize_;
};
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
index_t gridSize = getAvailableComputeUnitCount(stream_config);
auto in_grid_1d_desc_tuple = generate_tuple(
[&](auto I) {
return MakeDescriptor_M(
arg.lengths_, arg.inStridesArray_[I.value], gridSize, arg.blockSize_);
},
Number<NumInput>{});
auto out_grid_1d_desc_tuple = generate_tuple(
[&](auto I) {
return MakeDescriptor_M(
arg.lengths_, arg.outStridesArray_[I.value], gridSize, arg.blockSize_);
},
Number<NumOutput>{});
const auto kernel = kernel_elementwise_1d<GridwiseElementwise,
InGrid1dDescTuple,
OutGrid1dDescTuple,
InDataTypePointerTuple,
OutDataTypePointerTuple,
ElementwiseOperation,
UnaryOperation,
Scale>;
float elapsed_time = launch_and_time_kernel(stream_config,
kernel,
dim3(gridSize),
dim3(arg.blockSize_),
0,
in_grid_1d_desc_tuple,
out_grid_1d_desc_tuple,
arg.in_dev_buffers_,
arg.out_dev_buffers_,
arg.elementwise_op_,
arg.unary_op_,
arg.scale_op_);
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);
}
};
static bool IsSupportedArgument(const Argument& arg)
{
if(arg.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) {
if(strides.back() == 1 && lengths.back() % scalarPerVector == 0)
return true;
if(strides.back() != 1 && scalarPerVector == 1)
return true;
return false;
};
bool valid = true;
static_for<0, NumInput, 1>{}([&](auto I) {
if(!IsScalarPerVectorValid(
arg.lengths_, arg.inStridesArray_[I.value], InScalarPerVectorSeq::At(I)))
valid = false;
});
static_for<0, NumOutput, 1>{}([&](auto I) {
if(!IsScalarPerVectorValid(
arg.lengths_, arg.outStridesArray_[I.value], OutScalarPerVectorSeq::At(I)))
valid = false;
});
return valid;
};
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto
MakeArgument(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)
{
return Argument{lengths,
inStridesArray,
outStridesArray,
in_dev_buffers,
out_dev_buffers,
elementwise_op,
unary_op,
scale_op};
}
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) override
{
return std::make_unique<Argument>(lengths,
inStridesArray,
outStridesArray,
in_dev_buffers,
out_dev_buffers,
elementwise_op,
unary_op,
scale_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
include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
View file @ 87101e7d
...@@ -360,8 +360,8 @@ struct UnarySquare ...@@ -360,8 +360,8 @@ struct UnarySquare
template <typename T> template <typename T>
__host__ __device__ void operator()(T& y, const T& x) const __host__ __device__ void operator()(T& y, const T& x) const
{ {
static_assert(is_same_v<T, float> || is_same_v<T, double> || is_same_v<T, int32_t> || static_assert(is_same_v<T, float> || is_same_v<T, half_t> || is_same_v<T, double> ||
is_same_v<T, int8_t> is_same_v<T, int32_t> || is_same_v<T, int8_t>
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4 #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
|| is_same_v<T, int4_t> || is_same_v<T, int4_t>
#endif #endif
......
include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d_scale.hpp 0 → 100644
View file @ 87101e7d
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_description/cluster_descriptor.hpp"
#include "ck/utility/data_type.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
template <typename GridwiseElementwise1dFunctor,
typename InGrid1dDescTuple,
typename OutGrid1dDescTuple,
typename InDataTypePointerTuple,
typename OutDataTypePointerTuple,
typename ElementwiseOperation,
typename UnaryOperation,
typename Scale>
__global__ void kernel_elementwise_1d(const InGrid1dDescTuple in_grid_1d_desc_tuple,
const OutGrid1dDescTuple out_grid_1d_desc_tuple,
const InDataTypePointerTuple p_in_global_tuple,
const OutDataTypePointerTuple p_out_global_tuple,
const ElementwiseOperation elementwise_op,
const UnaryOperation unary_op,
const Scale scale_op)
{
GridwiseElementwise1dFunctor::Run(in_grid_1d_desc_tuple,
out_grid_1d_desc_tuple,
p_in_global_tuple,
p_out_global_tuple,
elementwise_op,
unary_op,
scale_op);
}
template <typename InGrid1dDescTuple,
typename OutGrid1dDescTuple,
typename InDataTypePointerTuple,
typename OutDataTypePointerTuple,
typename ElementwiseOperation,
typename UnaryOperation,
typename Scale,
index_t MPerThread,
typename InScalarPerVectorSeq,
typename OutScalarPerVectorSeq>
struct GridwiseElementwise_1D
{
static constexpr index_t NumInput = InDataTypePointerTuple::Size();
static constexpr index_t NumOutput = OutDataTypePointerTuple::Size();
static_assert(NumInput == InScalarPerVectorSeq::Size() &&
NumOutput == OutScalarPerVectorSeq::Size() &&
NumInput == InGrid1dDescTuple::Size() &&
NumOutput == OutGrid1dDescTuple::Size(),
"Tuple size is inconsistent with the number of in/out!");
static constexpr auto I0 = Number<0>{};
static constexpr auto thread_buffer_desc_m =
make_naive_tensor_descriptor_packed(make_tuple(Number<MPerThread>{}));
using PassThroughOp = tensor_operation::element_wise::PassThrough;
__device__ static void Run(const InGrid1dDescTuple in_grid_1d_desc_tuple,
const OutGrid1dDescTuple out_grid_1d_desc_tuple,
const InDataTypePointerTuple p_in_global_tuple,
const OutDataTypePointerTuple p_out_global_tuple,
const ElementwiseOperation elementwise_op,
const UnaryOperation unary_op,
const Scale scale_op)
{
const index_t thread_global_id = get_thread_global_1d_id();
auto in_thread_buf_tuple = generate_tuple(
[&](auto I) {
using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
using DataType = remove_cv_t<remove_pointer_t<DataTypePointer>>;
return StaticBuffer<AddressSpaceEnum::Vgpr, DataType, MPerThread, true>{};
},
Number<NumInput>{});
auto out_thread_buf_tuple = generate_tuple(
[&](auto I) {
using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
using DataType = remove_pointer_t<DataTypePointer>;
return StaticBuffer<AddressSpaceEnum::Vgpr, DataType, MPerThread, true>{};
},
Number<NumOutput>{});
auto in_global_buf_tuple = generate_tuple(
[&](auto I) {
static_assert(in_grid_1d_desc_tuple[I].GetNumOfDimension() == 1);
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_in_global_tuple[I], in_grid_1d_desc_tuple[I].GetElementSpaceSize());
},
Number<NumInput>{});
auto out_global_buf_tuple = generate_tuple(
[&](auto I) {
static_assert(out_grid_1d_desc_tuple[I].GetNumOfDimension() == 1);
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_out_global_tuple[I], out_grid_1d_desc_tuple[I].GetElementSpaceSize());
},
Number<NumOutput>{});
const auto thread_global_offset = make_multi_index(thread_global_id * MPerThread);
const index_t blockSize = get_block_size();
const index_t blockPerGrid = get_grid_size();
const auto M = in_grid_1d_desc_tuple[I0].GetLength(I0);
const index_t loop_step = blockPerGrid * blockSize * MPerThread;
const auto loop_step_index = make_multi_index(loop_step);
auto in_global_load_tuple = generate_tuple(
[&](auto I) {
using DataTypePointer = remove_cvref_t<decltype(InDataTypePointerTuple{}[I])>;
using DataType = remove_cv_t<remove_pointer_t<DataTypePointer>>;
return ThreadwiseTensorSliceTransfer_v2<DataType,
DataType,
decltype(in_grid_1d_desc_tuple[I]),
decltype(thread_buffer_desc_m),
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
InScalarPerVectorSeq::At(
I), // ScalarPerVector
1, // SrcScalarStrideInVector
false>{in_grid_1d_desc_tuple[I],
thread_global_offset};
},
Number<NumInput>{});
auto out_global_store_tuple = generate_tuple(
[&](auto I) {
using DataTypePointer = remove_cvref_t<decltype(OutDataTypePointerTuple{}[I])>;
using DataType = remove_pointer_t<DataTypePointer>;
return ThreadwiseTensorSliceTransfer_v1r3<DataType,
DataType,
decltype(thread_buffer_desc_m),
decltype(out_grid_1d_desc_tuple[I]),
PassThroughOp,
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
OutScalarPerVectorSeq::At(I),
InMemoryDataOperationEnum::Set,
1,
false>(
out_grid_1d_desc_tuple[I], thread_global_offset, PassThroughOp{});
},
Number<NumOutput>{});
index_t num_iter = M / (loop_step);
do
{
static_for<0, NumInput, 1>{}([&](auto I) {
in_global_load_tuple(I).Run(in_grid_1d_desc_tuple[I],
in_global_buf_tuple[I],
thread_buffer_desc_m,
make_tuple(I0),
in_thread_buf_tuple(I));
in_global_load_tuple(I).MoveSrcSliceWindow(in_grid_1d_desc_tuple[I],
loop_step_index);
});
static_for<0, MPerThread, 1>{}([&](auto iM) {
// get reference to in data
auto uop_data_refs = generate_tie(
// return type should be lvalue
[&](auto I) -> auto& { return in_thread_buf_tuple(I)(iM); },
Number<NumInput>{});
// get reference to dst data
auto out_data_refs = generate_tie(
// return type should be lvalue
[&](auto I) -> auto& { return out_thread_buf_tuple(I)(iM); },
Number<NumOutput>{});
unpack2(unary_op, uop_data_refs, uop_data_refs);
auto sop_in_data_refs = generate_tie(
// return type should be lvalue
[&](auto I) -> auto& { return in_thread_buf_tuple(I)(iM); },
Number<NumInput>{});
auto sop_out_data_refs = generate_tie(
// return type should be lvalue
[&](auto I) -> auto& { return in_thread_buf_tuple(I)(iM); },
Number<NumInput>{});
unpack2(scale_op, sop_out_data_refs, sop_in_data_refs);
const auto in_data_refs = generate_tie(
// return type should be lvalue
[&](auto I) -> const auto& { return in_thread_buf_tuple(I)(iM); },
Number<NumInput>{});
unpack2(elementwise_op, out_data_refs, in_data_refs);
});
static_for<0, NumOutput, 1>{}([&](auto I) {
out_global_store_tuple(I).Run(thread_buffer_desc_m,
make_tuple(I0),
out_thread_buf_tuple[I],
out_grid_1d_desc_tuple[I],
out_global_buf_tuple(I));
out_global_store_tuple(I).MoveDstSliceWindow(out_grid_1d_desc_tuple[I],
loop_step_index);
});
} while(--num_iter);
}
};
} // 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