Skip to content

GitLab

Commit 9dce6851 authored by Jing Zhang's avatar Jing Zhang
Browse files

merge develop

parents 3cc57101 5d37d7bf
Hide whitespace changes
Inline Side-by-side
Showing with 1343 additions and 330 deletions
example/08_conv3d_fwd/CMakeLists.txt 0 → 100644
View file @ 9dce6851
add_example_executable(example_conv3d_fwd_xdl conv3d_fwd_xdl.cpp)
example/09_convnd_fwd/CMakeLists.txt 0 → 100644
View file @ 9dce6851
add_example_executable(example_convnd_fwd_xdl convnd_fwd_xdl.cpp)
example/11_convnd_fwd_xdl/convnd_fwd_xdl.cpp → example/09_convnd_fwd/convnd_fwd_xdl.cpp
View file @ 9dce6851
......@@ -2,7 +2,6 @@
#include <iostream>
#include <numeric>
#include <type_traits>
#include "config.hpp"
#include "conv_utils.hpp"
#include "device.hpp"
......
example/10_conv2d_bwd_data/CMakeLists.txt 0 → 100644
View file @ 9dce6851
add_example_executable(example_conv2d_bwd_data_xdl conv2d_bwd_data_xdl.cpp)
example/7_conv2d_fwd_xdl_bias_relu_atomic_add/README.md → example/10_conv2d_bwd_data/README.md
View file @ 9dce6851
# Instructions for ```conv_xdl_bias_relu_add``` Example
# Instructions for ```conv2d_bwd_data_xdl``` Example
## Docker script
```bash
......@@ -13,7 +13,7 @@ rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```conv_xdl_bias_relu_add```
## Build ```conv2d_bwd_data_xdl```
```bash
mkdir build && cd build
```
......@@ -30,32 +30,50 @@ cmake \
```
```bash
make -j conv_xdl_bias_relu_add
make -j conv2d_bwd_data_xdl
```
## Run ```conv_xdl_bias_relu_add```
## Run ```conv2d_bwd_data_xdl```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
#arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
./example/conv_xdl_bias_relu_add 0 1 5
./bin/conv2d_bwd_data_xdl 0 1 5
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
Result
```
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
in_n_c_hi_wi: dim 4, lengths {128, 256, 71, 71}, strides {1290496, 1, 18176, 256}
wei_k_c_y_x: dim 4, lengths {256, 256, 3, 3}, strides {2304, 1, 768, 256}
out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
bias_k: dim 1, lengths {256}, strides {1}
resi_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
arg.a_grid_desc_k0_m_k1_{216, 165888, 8}
arg.b_grid_desc_k0_n_k1_{216, 256, 8}
arg.c_grid_desc_m_n_{ 165888, 256}
arg.c0_grid_desc_m_n_{ 165888, 256}
arg.c1_grid_desc_m_n_{ 165888, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
arg.a_grid_desc_k0_m_k1_container_{128, 175232, 8}
arg.b_grid_desc_k0_n_k1_container_{128, 256, 8}
arg.c_grid_desc_m_n_container_{ 175232, 256}
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 4, 2, 2, 4, 2 )
launch_and_time_kernel: grid_dim {2738, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.71779 ms, 85.4396 TFlops, 194.2 GB/s
Start running 1 times...
arg.a_grid_desc_k0_m_k1_container_{64, 175232, 8}
arg.b_grid_desc_k0_n_k1_container_{64, 256, 8}
arg.c_grid_desc_m_n_container_{ 175232, 256}
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 4, 2, 2, 4, 2 )
launch_and_time_kernel: grid_dim {2738, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
arg.a_grid_desc_k0_m_k1_container_{64, 175232, 8}
arg.b_grid_desc_k0_n_k1_container_{64, 256, 8}
arg.c_grid_desc_m_n_container_{ 175232, 256}
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 4, 2, 2, 4, 2 )
launch_and_time_kernel: grid_dim {2738, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
arg.a_grid_desc_k0_m_k1_container_{32, 175232, 8}
arg.b_grid_desc_k0_n_k1_container_{32, 256, 8}
arg.c_grid_desc_m_n_container_{ 175232, 256}
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( 2738, 4, 2, 2, 4, 2 )
launch_and_time_kernel: grid_dim {2738, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 1 times...
Perf: 2.45966 ms, 79.5597 TFlops, 169.325 GB/s
```
example/10_conv2d_bwd_data/conv2d_bwd_data_xdl.cpp 0 → 100644
View file @ 9dce6851
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "tensor_layout.hpp"
#include "element_wise_operation.hpp"
#include "device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp"
#include "reference_conv_bwd_data.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvBwdDefault =
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization_t::Default;
using DeviceConvBwdDataInstance = ck::tensor_operation::device::
DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
ConvBwdDefault, // ConvolutionBackwardDataSpecialization_t
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
8, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<2, 0, 1>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1>, // BBlockTransferSrcAccessOrder
1, // BBlockTransferSrcVectorDim
2, // BBlockTransferSrcScalarPerVector
8, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
7,
1>; // GemmCThreadTransferDstScalarPerVector
using ReferenceConvBwdInstance = ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 256;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 71;
ck::index_t Wi = 71;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 19)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
C = std::stoi(argv[6]);
Y = std::stoi(argv[7]);
X = std::stoi(argv[8]);
Hi = std::stoi(argv[9]);
Wi = std::stoi(argv[10]);
conv_stride_h = std::stoi(argv[11]);
conv_stride_w = std::stoi(argv[12]);
conv_dilation_h = std::stoi(argv[13]);
conv_dilation_w = std::stoi(argv[14]);
in_left_pad_h = std::stoi(argv[15]);
in_left_pad_w = std::stoi(argv[16]);
in_right_pad_h = std::stoi(argv[17]);
in_right_pad_w = std::stoi(argv[18]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
}
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
// tensor layout
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W) {
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
};
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X));
Tensor<InDataType> in_n_c_hi_wi_host_result(f_host_tensor_descriptor(N, C, Hi, Wi));
Tensor<InDataType> in_n_c_hi_wi_device_result(f_host_tensor_descriptor(N, C, Hi, Wi));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi_host_result.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) *
in_n_c_hi_wi_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
// do GEMM
auto conv = DeviceConvBwdDataInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
std::vector<ck::index_t>{{Hi, Wi}},
std::vector<ck::index_t>{{Y, X}},
std::vector<ck::index_t>{{Ho, Wo}},
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
if(!conv.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem");
}
float ave_time = invoker.Run(argument, nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
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;
if(do_verification)
{
auto ref_conv = ReferenceConvBwdInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
wei_k_c_y_x,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
check_error(in_n_c_hi_wi_host_result, in_n_c_hi_wi_device_result);
}
}
example/11_conv2d_bwd_wgt/CMakeLists.txt 0 → 100644
View file @ 9dce6851
add_example_executable(example_conv2d_bwd_wgt_xdl conv2d_bwd_wgt_xdl.cpp)
example/9_conv2d_fwd_xdl_int8/README.md → example/11_conv2d_bwd_wgt/README.md
View file @ 9dce6851
# Instructions for ```conv2d_fwd_xdl``` Example
# Instructions for ```conv2d_wrw_xdl``` Example
## Docker script
```bash
......@@ -13,7 +13,7 @@ rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```conv2d_fwd_xdl```
## Build ```conv2d_wrw_xdl```
```bash
mkdir build && cd build
```
......@@ -30,28 +30,29 @@ cmake \
```
```bash
make -j conv2d_fwd_xdl
make -j conv2d_wrw_xdl
```
## Run ```conv2d_fwd_xdl_int8```
## Run ```conv2d_wrw_xdl```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
#arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
./example/conv2d_fwd_xdl_int8 0 1 5
#arg4: is show log (0=no, 1=yes)
#arg5 to 19: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx, split-k
./example/conv2d_fwd_xdl 0 1 5 0 4
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
Result
```
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
arg.a_grid_desc_k0_m_k1_{216, 165888, 8}
arg.b_grid_desc_k0_n_k1_{216, 256, 8}
arg.c_grid_desc_m_n_{ 165888, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
in_n_c_hi_wi: dim 4, lengths {128, 1024, 14, 14}, strides {200704, 1, 14336, 1024}
wei_k_c_y_x: dim 4, lengths {256, 1024, 3, 3}, strides {9216, 1, 3072, 1024}
out_n_k_ho_wo: dim 4, lengths {128, 256, 6, 6}, strides {9216, 1, 1536, 256}
arg.a_grid_desc_kbatch_k0_m_k1_{4, 144, 256, 8}
arg.b_grid_desc_kbatch_k0_n_k1_{4, 144, 9216, 8}
arg.c_grid_desc_m_n_{ 256, 9216}
launch_and_time_kernel: grid_dim {576, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 1.43206 ms, 102.486 TFlops, 232.947 GB/s
Perf: 0.401084 ms, 54.2112 TFlops, 145.75 GB/s
```
example/11_conv2d_bwd_wgt/conv2d_bwd_wgt_xdl.cpp 0 → 100644
View file @ 9dce6851
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "tensor_layout.hpp"
#include "element_wise_operation.hpp"
#include "device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
#include "reference_conv_backward_weight.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InLayout = ck::tensor_layout::convolution::NHWC;
using WeiLayout = ck::tensor_layout::convolution::KYXC;
using OutLayout = ck::tensor_layout::convolution::NHWK;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
// clang-format off
using DeviceConvWrWInstance = ck::tensor_operation::device::
DeviceConv2dWrWXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
using ReferenceConvWrwInstance = ck::tensor_operation::host::
ReferenceConvWrw<InDataType, WeiDataType, OutDataType, InElementOp, WeiElementOp, OutElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
int do_log = 0;
int split_k = 4;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 1024;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 14;
ck::index_t Wi = 14;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 0;
ck::index_t in_left_pad_w = 0;
ck::index_t in_right_pad_h = 0;
ck::index_t in_right_pad_w = 0;
if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
}
else if(argc == 21)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
N = std::stoi(argv[6]);
K = std::stoi(argv[7]);
C = std::stoi(argv[8]);
Y = std::stoi(argv[9]);
X = std::stoi(argv[10]);
Hi = std::stoi(argv[11]);
Wi = std::stoi(argv[12]);
conv_stride_h = std::stoi(argv[13]);
conv_stride_w = std::stoi(argv[14]);
conv_dilation_h = std::stoi(argv[15]);
conv_dilation_w = std::stoi(argv[16]);
in_left_pad_h = std::stoi(argv[17]);
in_left_pad_w = std::stoi(argv[18]);
in_right_pad_h = std::stoi(argv[19]);
in_right_pad_w = std::stoi(argv[20]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4: is show log (0=no, 1=yes)\n");
printf("arg5: split-k \n");
printf("arg6 to 19: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
}
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
// tensor layout
auto f_host_tensor_descriptor = [](std::size_t N_,
std::size_t C_,
std::size_t H,
std::size_t W,
auto layout) {
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::KYXC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x_host_result(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<WeiDataType> wei_k_c_y_x_device_result(
f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
}
wei_k_c_y_x_device_result.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{0});
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) *
wei_k_c_y_x_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x_device_result.mData.data());
// do GEMM
auto conv = DeviceConvWrWInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
std::vector<ck::index_t>{{Hi, Wi}},
std::vector<ck::index_t>{{Y, X}},
std::vector<ck::index_t>{{Ho, Wo}},
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{},
split_k);
if(!conv.IsSupportedArgument(argument))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
}
float ave_time = invoker.Run(argument, nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
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;
if(do_verification)
{
auto ref_conv = ReferenceConvWrwInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x_host_result,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
if(do_log)
{
LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
}
check_error(wei_k_c_y_x_host_result, wei_k_c_y_x_device_result);
}
}
example/12_reduce/CMakeLists.txt 0 → 100644
View file @ 9dce6851
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
example/12_reduce/reduce_blockwise.cpp 0 → 100644
View file @ 9dce6851
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_reduce_blockwise.hpp"
#include "host_reduce_util.hpp"
#include "host_generic_reduction.hpp"
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
using namespace ck;
using namespace ck::tensor_operation::device;
using InDataType = half_float::half;
using OutDataType = half_float::half;
using AccDataType = float;
using kInDataType = ck::half_t;
using kOutDataType = ck::half_t;
using kAccDataType = float;
constexpr int Rank = 4;
using ReduceDims_ = ck::Sequence<0, 1, 2>;
constexpr ReduceTensorOp_t ReduceOpId = ReduceTensorOp_t::NORM2;
constexpr NanPropagation_t NanOpt = NanPropagation_t::PROPAGATE_NAN;
constexpr bool PropagateNan = (NanOpt == NanPropagation_t::NOT_PROPAGATE_NAN) ? false : true;
constexpr ReduceTensorIndices_t IndicesOpt = ReduceTensorIndices_t::NO_INDICES;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
using DeviceReduceInstance = DeviceReduceBlockWise<kInDataType,
kAccDataType,
kOutDataType,
Rank,
ReduceDims_,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
false,
256,
4,
64,
1,
1,
0,
1,
1>;
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
{"scales", required_argument, nullptr, 'S'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
class SimpleAppArgs
{
template <typename T>
static T getSingleValueFromString(const std::string& valueStr)
{
std::istringstream iss(valueStr);
T ret;
iss >> ret;
return (ret);
};
template <typename T>
static std::vector<T> getTypeValuesFromString(const char* cstr_values)
{
std::string valuesStr(cstr_values);
std::vector<T> values;
std::size_t pos = 0;
std::size_t new_pos;
new_pos = valuesStr.find(',', pos);
while(new_pos != std::string::npos)
{
const std::string sliceStr = valuesStr.substr(pos, new_pos - pos);
T val = getSingleValueFromString<T>(sliceStr);
values.push_back(val);
pos = new_pos + 1;
new_pos = valuesStr.find(',', pos);
};
std::string sliceStr = valuesStr.substr(pos);
T val = getSingleValueFromString<T>(sliceStr);
values.push_back(val);
return (values);
};
private:
int option_index = 0;
public:
std::vector<size_t> inLengths;
std::vector<float> scales;
bool do_verification = false;
int init_method = 1;
int nrepeat = 5;
public:
void show_usage(const char* cmd)
{
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
<< std::endl;
std::cout << "--scales or -S, comma separated two float values for alpha and beta"
<< std::endl;
std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
"comparing with the host-based reduction"
<< std::endl;
};
int processArgs(int argc, char* argv[])
{
unsigned int ch;
while(1)
{
ch = getopt_long(argc, argv, "D:S:v:l:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
{
case 'D':
if(!optarg)
throw std::runtime_error("Invalid option format!");
inLengths = getTypeValuesFromString<size_t>(optarg);
break;
case 'S':
if(!optarg)
throw std::runtime_error("Invalid option format!");
scales = getTypeValuesFromString<float>(optarg);
break;
case 'v':
if(!optarg)
throw std::runtime_error("Invalid option format!");
do_verification = static_cast<bool>(std::atoi(optarg));
break;
case '?':
if(std::string(long_options[option_index].name) == "help")
{
show_usage(argv[0]);
return (-1);
};
break;
default: show_usage(argv[0]); return (-1);
};
};
if(optind + 2 > argc)
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
init_method = std::atoi(argv[optind++]);
nrepeat = std::atoi(argv[optind]);
if(scales.empty())
{
scales.push_back(1.0f);
scales.push_back(0.0f);
};
return (0);
};
};
template <int Rank, typename ReduceDims>
static std::vector<int> get_reduce_dims()
{
std::vector<int> resDims;
static_for<0, ReduceDims::Size(), 1>{}([&](auto i) { resDims.push_back(ReduceDims::At(i)); });
return (resDims);
};
template <int Rank, typename ReduceDims>
static std::vector<int> get_invariant_dims()
{
std::vector<int> resDims;
unsigned int incFlag = 0;
static_for<0, ReduceDims::Size(), 1>{}(
[&](auto i) { incFlag = incFlag | (0x1 << ReduceDims::At(i)); });
for(int dim = 0; dim < Rank; dim++)
{
if(incFlag & (0x1 << dim))
continue;
resDims.push_back(dim);
};
return (resDims);
};
int main(int argc, char* argv[])
{
using namespace ck::host_reduce;
SimpleAppArgs args;
if(args.processArgs(argc, argv) < 0)
return (-1);
constexpr bool op_support_indices =
(ReduceOpId == ReduceTensorOp_t::MIN || ReduceOpId == ReduceTensorOp_t::MAX ||
ReduceOpId == ReduceTensorOp_t::AMAX);
constexpr bool NeedIndices =
(op_support_indices && (IndicesOpt != ReduceTensorIndices_t::NO_INDICES));
// if input is half type, no reason to use float for indiced reduction operation and must use
// float for non-indiced reduction operation for accuracy
constexpr bool invalid_reduce_1 =
std::is_same<InDataType, ck::half_t>::value &&
((!op_support_indices && !std::is_same<AccDataType, float>::value) ||
(op_support_indices && !std::is_same<AccDataType, ck::half_t>::value));
// if input is float type, no reason to use double for indiced reduction operation
constexpr bool invalid_reduce_2 =
std::is_same<InDataType, float>::value &&
(op_support_indices && !std::is_same<AccDataType, float>::value);
// indices option can only be used when it is really needed
constexpr bool invalid_reduce_3 =
(!op_support_indices && IndicesOpt != ReduceTensorIndices_t::NO_INDICES);
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3);
if constexpr(invalid_reduce)
std::cout << "Reduction setting is not supported, exiting!" << std::endl;
Tensor<InDataType> in(args.inLengths);
const std::vector<int> InvariantDims = get_invariant_dims<Rank, ReduceDims_>();
const std::vector<int> ReduceDims = get_reduce_dims<Rank, ReduceDims_>();
std::vector<size_t> outLengths;
if(InvariantDims.empty())
outLengths.push_back(1);
else
for(auto dim : InvariantDims)
outLengths.push_back(args.inLengths[dim]);
Tensor<OutDataType> out_ref(outLengths);
Tensor<OutDataType> out(outLengths);
Tensor<int> out_indices_ref(outLengths);
Tensor<int> out_indices(outLengths);
auto inStrides = in.mDesc.GetStrides();
auto outStrides = out.mDesc.GetStrides();
size_t invariant_total_length = out.mDesc.GetElementSize();
size_t reduce_total_length = in.mDesc.GetElementSize() / invariant_total_length;
float alpha = args.scales[0];
float beta = args.scales[1];
std::size_t num_thread = std::thread::hardware_concurrency();
if(args.do_verification)
{
switch(args.init_method)
{
case 0:
in.GenerateTensorValue(GeneratorTensor_1<InDataType>{}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_1<InDataType>{}, num_thread);
break;
case 1:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
break;
default:
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{1, 5}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_2<InDataType>{1, 5}, num_thread);
}
if(beta != 0.0f)
for(size_t i = 0; i < out_ref.mDesc.GetElementSpace(); i++)
out.mData[i] = out_ref.mData[i];
};
// these buffers are usually provided by the user application
DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
in_dev.ToDevice(in.mData.data());
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int) : 0;
DeviceMem out_indices_dev(indicesSizeInBytes);
if(args.do_verification)
{
ReductionHost<InDataType, AccDataType, OutDataType, ReduceOpId, PropagateNan, NeedIndices>
hostReduce(in.mDesc, out_ref.mDesc, InvariantDims, ReduceDims);
hostReduce.Run(
alpha, in.mData.data(), beta, out_ref.mData.data(), out_indices_ref.mData.data());
};
const auto i_inLengths = to_int_vector(args.inLengths);
const auto i_inStrides = to_int_vector(inStrides);
const auto i_outLengths = to_int_vector(outLengths);
const auto i_outStrides = to_int_vector(outStrides);
auto reduce = DeviceReduceInstance{};
auto wsSizeInBytes = reduce.GetWorkspaceSizeInBytes(i_inLengths);
DeviceMem ws_dev(wsSizeInBytes);
auto argument_ptr =
reduce.MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
InElementwiseOperation{static_cast<int>(reduce_total_length)},
AccElementwiseOperation{static_cast<int>(reduce_total_length)});
if(!reduce.IsSupportedArgument(argument_ptr.get()))
{
std::cout
<< "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
<< std::endl;
};
std::string reduce_name = reduce.GetTypeString();
auto invoker_ptr = reduce.MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), args.nrepeat);
std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InDataType) +
invariant_total_length * sizeof(OutDataType);
float gb_per_sec = num_bytes / 1.E6 / avg_time;
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
<< std::endl;
if(args.do_verification)
{
out_dev.FromDevice(out.mData.data());
check_error(out_ref, out);
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
check_indices(out_indices_ref, out_indices);
};
};
}
example/13_pool2d_fwd/CMakeLists.txt 0 → 100644
View file @ 9dce6851
add_example_executable(example_pool2d_fwd pool2d_fwd.cpp)
example/13_pool2d_fwd/pool2d_fwd.cpp 0 → 100644
View file @ 9dce6851
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_reduce_util.hpp"
#include "device_tensor.hpp"
#include "tensor_layout.hpp"
#include "reduction_operator.hpp"
#include "device_pool2d_fwd_nhwc_nhwc.hpp"
using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
using InLayout = ck::tensor_layout::convolution::NHWC;
using OutLayout = ck::tensor_layout::convolution::NHWC;
#if 1
static constexpr auto ReduceOpId = ck::ReduceTensorOp_t::MAX;
#else
static constexpr auto ReduceOpId = ck::ReduceTensorOp_t::AVG;
#endif
static constexpr bool NeedIndices = false;
static constexpr bool PropagateNan = false;
using DevicePoolFwdInstance =
ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
InDataType, // InDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
ReduceOpId,
NeedIndices,
64, // BlockSize
64, // ReduceMThreadClusterSize
1, // ReduceKThreadClusterSize
4, // ReduceMThreadSliceSize
1, // ReduceKThreadSliceSize
4>; // InSrcOutDstVectorSize
template <typename InDataType,
typename OutDataType,
typename AccDataType,
ck::ReduceTensorOp_t ReduceOpId,
bool PropagateNan,
bool NeedIndices>
static void pool_host_verify(const Tensor<InDataType>& in,
Tensor<OutDataType>& out,
Tensor<int>& out_indices,
const std::array<ck::index_t, 2>& window_spatial_lengths,
const std::array<ck::index_t, 2>& window_strides,
const std::array<ck::index_t, 2>& in_left_pads,
const std::array<ck::index_t, 2>& /*in_right_pads*/)
{
using namespace ck::host_reduce;
const int divider = window_spatial_lengths[0] * window_spatial_lengths[1];
const auto PreUnaryOp = PreUnaryOpFn<AccDataType, ReduceOpId>(divider);
const auto PosUnaryOp = PosUnaryOpFn<AccDataType, ReduceOpId>(divider);
if constexpr(!NeedIndices)
{
auto opReduce = ReduceOpFn<AccDataType, ReduceOpId>();
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
for(int y = 0; y < window_spatial_lengths[0]; ++y)
{
int hi = ho * window_strides[0] + y - in_left_pads[0];
for(int x = 0; x < window_spatial_lengths[1]; ++x)
{
int wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
PreUnaryOp(currVal);
binop_with_nan_check<AccDataType, PropagateNan>(opReduce, accuVal, currVal);
}
}
}
PosUnaryOp(accuVal);
out(n, c, ho, wo) = accuVal;
};
make_ParallelTensorFunctor(f_nchw,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
}
else
{
auto opReduce = ReduceOpFn2<AccDataType, ReduceOpId>();
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
int accuIndex = 0;
for(int y = 0; y < window_spatial_lengths[0]; ++y)
{
int hi = ho * window_strides[0] + y - in_left_pads[0];
for(int x = 0; x < window_spatial_lengths[1]; ++x)
{
int wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
int currIndex = y * window_spatial_lengths[1] + x;
PreUnaryOp(currVal);
binop_with_nan_check2<AccDataType, PropagateNan>(
opReduce, accuVal, currVal, accuIndex, currIndex);
}
}
}
PosUnaryOp(accuVal);
out(n, c, ho, wo) = accuVal;
out_indices(n, c, ho, wo) = accuIndex;
};
make_ParallelTensorFunctor(f_nchw,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
};
}
int main(int argc, char* argv[])
{
using namespace ck::host_reduce;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
// Pool shape
ck::index_t N = 128;
ck::index_t C = 192;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 71;
ck::index_t Wi = 71;
ck::index_t window_stride_h = 2;
ck::index_t window_stride_w = 2;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 16)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
N = std::stoi(argv[4]);
C = std::stoi(argv[5]);
Y = std::stoi(argv[6]);
X = std::stoi(argv[7]);
Hi = std::stoi(argv[8]);
Wi = std::stoi(argv[9]);
window_stride_h = std::stoi(argv[10]);
window_stride_w = std::stoi(argv[11]);
in_left_pad_h = std::stoi(argv[12]);
in_left_pad_w = std::stoi(argv[13]);
in_right_pad_h = std::stoi(argv[14]);
in_right_pad_w = std::stoi(argv[15]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
}
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - Y) / window_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - X) / window_stride_w + 1;
const std::array<ck::index_t, 2> window_spatial_lengths{{Y, X}};
const std::array<ck::index_t, 2> window_strides{{window_stride_h, window_stride_w}};
const std::array<ck::index_t, 2> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::array<ck::index_t, 2> input_right_pads{{in_right_pad_h, in_right_pad_w}};
// tensor layout
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWC>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<int> out_indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<int> out_indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
default: in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_c_ho_wo_device.mDesc.GetElementSpace());
DeviceMem out_indices_device_buf(sizeof(int) *
out_indices_n_c_ho_wo_device.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
auto pool = DevicePoolFwdInstance{};
auto invoker_ptr = pool.MakeInvokerPointer();
auto argument_ptr =
pool.MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<int*>(out_indices_device_buf.GetDeviceBuffer()),
N,
C,
std::array<ck::index_t, 2>{{Hi, Wi}},
std::array<ck::index_t, 2>{{Y, X}},
std::array<ck::index_t, 2>{{Ho, Wo}},
window_strides,
input_left_pads,
input_right_pads);
if(!pool.IsSupportedArgument(argument_ptr.get()))
{
throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
"not support this problem");
}
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
std::size_t num_btype =
sizeof(InDataType) * (N * C * Hi * Wi) + sizeof(OutDataType) * (N * C * Ho * Wo);
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;
if(do_verification)
{
pool_host_verify<InDataType,
OutDataType,
AccDataType,
ReduceOpId,
PropagateNan,
NeedIndices>(in_n_c_hi_wi,
out_n_c_ho_wo_host,
out_indices_n_c_ho_wo_host,
window_spatial_lengths,
window_strides,
input_left_pads,
input_right_pads);
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
check_error(out_n_c_ho_wo_host, out_n_c_ho_wo_device);
if constexpr(NeedIndices)
{
out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
// check_indices(out_indices_n_c_ho_wo_host, out_indices_n_c_ho_wo_device);
};
}
}
example/14_grouped_gemm/CMakeLists.txt 0 → 100644
View file @ 9dce6851
add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
example/7_conv2d_fwd_xdl_bias_relu_atomic_add/conv2d_fwd_xdl_bias_relu_atomic_add.cpp deleted 100644 → 0
View file @ 3cc57101
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "tensor_layout.hpp"
#include "device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation.hpp"
#include "convolution_utility.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InLayout = ck::tensor_layout::convolution::NHWC;
using WeiLayout = ck::tensor_layout::convolution::KYXC;
using OutLayout = ck::tensor_layout::convolution::NHWK;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddRelu;
static constexpr auto MemoryAtomicAdd = ck::InMemoryDataOperationEnum_t::AtomicAdd;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization_t::Default;
// clang-format off
using DeviceConvFwdInstance = ck::tensor_operation::device::
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
// clang-format off
// | InData| WeiData| OutData| AccData| In| Wei| Out| Out| ConvForward| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
// | | | | | Operation| Operation| Operation| DataOperation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
// | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
<InDataType, WeiDataType, OutDataType, AccDataType, InElementOp, WeiElementOp, OutElementOp, MemoryAtomicAdd, ConvFwdDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 8, 1, 1,32>, 2>;
// clang-format on
template <typename TIn,
typename TWei,
typename TOut,
typename InElementOp,
typename WeiElementOp,
typename OutElementOp>
void host_reference_calculation(const Tensor<TIn>& in_n_c_hi_wi,
const Tensor<TWei>& wei_k_c_y_x,
Tensor<TOut>& out_n_k_ho_wo,
const Tensor<TOut>& bias_k,
const std::vector<ck::index_t>& conv_strides,
const std::vector<ck::index_t>& conv_dilations,
const std::vector<ck::index_t>& in_left_pads,
const std::vector<ck::index_t>& /* in_right_pads */,
const InElementOp& in_element_op,
const WeiElementOp& wei_element_op,
const OutElementOp& out_element_op)
{
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
float v_acc = 0;
for(int c = 0; c < wei_k_c_y_x.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < wei_k_c_y_x.mDesc.GetLengths()[2]; ++y)
{
int hi = ho * conv_strides[0] + y * conv_dilations[0] - in_left_pads[0];
for(int x = 0; x < wei_k_c_y_x.mDesc.GetLengths()[3]; ++x)
{
int wi = wo * conv_strides[1] + x * conv_dilations[1] - in_left_pads[1];
if(hi >= 0 && hi < in_n_c_hi_wi.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in_n_c_hi_wi.mDesc.GetLengths()[3])
{
float v_in;
float v_wei;
in_element_op(v_in, static_cast<const float>(in_n_c_hi_wi(n, c, hi, wi)));
wei_element_op(v_wei, static_cast<const float>(wei_k_c_y_x(k, c, y, x)));
v_acc += v_in * v_wei;
}
}
}
}
float v_out;
out_element_op(v_out, v_acc, static_cast<float>(bias_k(k)));
out_n_k_ho_wo(n, k, ho, wo) += v_out;
};
make_ParallelTensorFunctor(f_nchw,
out_n_k_ho_wo.mDesc.GetLengths()[0],
out_n_k_ho_wo.mDesc.GetLengths()[1],
out_n_k_ho_wo.mDesc.GetLengths()[2],
out_n_k_ho_wo.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 192;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 71;
ck::index_t Wi = 71;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 1;
ck::index_t in_left_pad_w = 1;
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
}
else if(argc == 19)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
C = std::stoi(argv[6]);
Y = std::stoi(argv[7]);
X = std::stoi(argv[8]);
Hi = std::stoi(argv[9]);
Wi = std::stoi(argv[10]);
conv_stride_h = std::stoi(argv[11]);
conv_stride_w = std::stoi(argv[12]);
conv_dilation_h = std::stoi(argv[13]);
conv_dilation_w = std::stoi(argv[14]);
in_left_pad_h = std::stoi(argv[15]);
in_left_pad_w = std::stoi(argv[16]);
in_right_pad_h = std::stoi(argv[17]);
in_right_pad_w = std::stoi(argv[18]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
}
const std::vector<ck::index_t> conv_filter_strides{conv_stride_h, conv_stride_w};
const std::vector<ck::index_t> conv_filter_dilations{conv_dilation_h, conv_dilation_w};
const std::vector<ck::index_t> input_left_pads{in_left_pad_h, in_left_pad_w};
const std::vector<ck::index_t> input_right_pads{in_right_pad_h, in_right_pad_w};
const auto output_spatial_lengths =
ck::tensor_operation::ConvolutionUtility::ComputeOutputSpatialLengths({Hi, Wi},
{Y, X},
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads);
const ck::index_t Ho = output_spatial_lengths[0];
const ck::index_t Wo = output_spatial_lengths[1];
// tensor layout
auto f_host_tensor_descriptor = [](std::size_t N_,
std::size_t C_,
std::size_t H,
std::size_t W,
auto layout) {
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::KYXC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo_host_result(
f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
Tensor<OutDataType> out_n_k_ho_wo_device_result(
f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
// bias: assume contiguous 1d vector
Tensor<OutDataType> bias_k(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(K)})));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo_host_result.mDesc << std::endl;
std::cout << "bias_k: " << bias_k.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
out_n_k_ho_wo_host_result.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
bias_k.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
out_n_k_ho_wo_host_result.GenerateTensorValue(GeneratorTensor_3<OutDataType>{-0.5, 0.5});
bias_k.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
}
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) *
out_n_k_ho_wo_device_result.mDesc.GetElementSpace());
DeviceMem bias_device_buf(sizeof(OutDataType) * bias_k.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
out_device_buf.ToDevice(out_n_k_ho_wo_host_result.mData.data());
bias_device_buf.ToDevice(bias_k.mData.data());
auto conv = DeviceConvFwdInstance{};
auto invoker = conv.MakeInvoker();
auto argument =
conv.MakeArgument(static_cast<const InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
if(!conv.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device operator with the specified compilation parameters does "
"not support this problem");
}
float ave_time = invoker.Run(argument, nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo) + sizeof(OutDataType) * (K);
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;
if(do_verification)
{
host_reference_calculation(in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo_host_result,
bias_k,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
out_device_buf.FromDevice(out_n_k_ho_wo_device_result.mData.data());
check_error(out_n_k_ho_wo_host_result, out_n_k_ho_wo_device_result);
}
}
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