ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-06 05:28:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'origin/develop' into myamlak/cgemm

Browse Source
This commit is contained in:
myamlak
2022-05-17 10:23:36 +00:00
parent ffe12e2ea6 9f71ff48e2
commit e00a943e34
162 changed files with 3223 additions and 2327 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
example/01_gemm/gemm_xdl_bf16.cpp
View File

@@ -88,9 +88,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
@@ -105,13 +105,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -125,7 +125,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
@@ -198,7 +198,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
@@ -232,7 +232,7 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_f32_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;

16
example/01_gemm/gemm_xdl_fp16.cpp
View File

@@ -56,9 +56,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
@@ -73,13 +73,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -93,7 +93,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
@@ -171,7 +171,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
@@ -196,7 +196,7 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;

16
example/01_gemm/gemm_xdl_int8.cpp
View File

@@ -83,9 +83,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
@@ -100,13 +100,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -120,7 +120,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
@@ -194,7 +194,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
@@ -219,7 +219,7 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;

20
example/02_gemm_alpha_beta/gemm_xdl_alpha_beta.cpp
View File

@@ -86,9 +86,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBias2D<AD
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
@@ -106,13 +106,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
alpha = std::stof(argv[4]);
beta = std::stof(argv[5]);
@@ -121,7 +121,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -138,7 +138,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, alpha, beta\n");
exit(0);
}
@@ -216,7 +216,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
@@ -246,6 +246,8 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;
}

18
example/03_gemm_bias_relu/gemm_xdl_bias_relu.cpp
View File

@@ -83,9 +83,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBiasActiv
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
@@ -100,13 +100,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -120,7 +120,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
@@ -206,7 +206,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
@@ -232,6 +232,8 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;
}

18
example/04_gemm_bias_relu_add/gemm_xdl_bias_relu_add.cpp
View File

@@ -83,9 +83,9 @@ using ReferenceGemmInstance =
CElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
@@ -101,13 +101,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -122,7 +122,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, StrideC1\n");
exit(0);
}
@@ -218,7 +218,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * M +
@@ -250,6 +250,8 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;
}

2
example/06_conv2d_fwd_bias_relu/CMakeLists.txt
View File

@@ -1,2 +1,2 @@
add_example_executable(example_conv2d_fwd_xdl_bias_relu conv2d_fwd_xdl_bias_relu.cpp)
target_link_libraries(example_conv2d_fwd_xdl_bias_relu PRIVATE conv_fwd_util)
target_link_libraries(example_conv2d_fwd_xdl_bias_relu PRIVATE conv_util)

113
example/06_conv2d_fwd_bias_relu/conv2d_fwd_xdl_bias_relu.cpp
View File

@@ -7,7 +7,7 @@
#include "check_err.hpp"
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "device.hpp"
#include "device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp"
#include "device_tensor.hpp"
@@ -93,7 +93,7 @@ void PrintUseMsg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: run kernel # of times (>1)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "Following arguments:\n"
<< " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n"
@@ -120,40 +120,40 @@ ck::utils::conv::ConvParams ParseConvParams(int argc, char* argv[])
ck::utils::conv::ConvParams params;
int arg_idx = 4;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
@@ -165,9 +165,9 @@ int main(int argc, char* argv[])
{
using namespace ck::utils::conv;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
const int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
@@ -176,7 +176,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
if(argc >= 5)
@@ -184,21 +184,21 @@ int main(int argc, char* argv[])
params = ParseConvParams(argc, argv);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
@@ -211,7 +211,7 @@ int main(int argc, char* argv[])
get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
// bias: assume contiguous 1d vector
Tensor<OutDataType> bias(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K)})));
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K_)})));
std::cout << "input: " << input.mDesc << std::endl;
std::cout << "weights: " << weights.mDesc << std::endl;
@@ -248,16 +248,16 @@ int main(int argc, char* argv[])
static_cast<const WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
@@ -269,18 +269,18 @@ int main(int argc, char* argv[])
"not support this problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops(
params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype =
get_btype<InDataType, WeiDataType, OutDataType>(params.N,
params.C,
params.K,
params.input_spatial_lengths,
params.filter_spatial_lengths,
get_btype<InDataType, WeiDataType, OutDataType>(params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths) +
sizeof(OutDataType) * (params.K);
sizeof(OutDataType) * (params.K_);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
@@ -296,16 +296,17 @@ int main(int argc, char* argv[])
weights,
host_output,
bias,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data());
ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
return ck::utils::check_err(device_output.mData, host_output.mData) ? 0 : 1;
}
return 0;
}

5
example/07_conv2d_fwd_bias_relu_add/CMakeLists.txt
View File

@@ -1,2 +1,3 @@
add_example_executable(example_conv2d_fwd_xdl_bias_relu_add conv2d_fwd_xdl_bias_relu_add.cpp)
target_link_libraries(example_conv2d_fwd_xdl_bias_relu_add PRIVATE conv_fwd_util)
# FIXME: should fix validation failure
add_example_executable_no_testing(example_conv2d_fwd_xdl_bias_relu_add conv2d_fwd_xdl_bias_relu_add.cpp)
target_link_libraries(example_conv2d_fwd_xdl_bias_relu_add PRIVATE conv_util)

115
example/07_conv2d_fwd_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp
View File

@@ -7,7 +7,7 @@
#include "check_err.hpp"
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "device.hpp"
#include "device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp"
#include "device_tensor.hpp"
@@ -90,7 +90,7 @@ void PrintUseMsg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: run kernel # of times (>1)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "Following arguments:\n"
<< " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n"
@@ -117,40 +117,40 @@ ck::utils::conv::ConvParams ParseConvParams(int argc, char* argv[])
ck::utils::conv::ConvParams params;
int arg_idx = 4;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
@@ -162,9 +162,9 @@ int main(int argc, char* argv[])
{
using namespace ck::utils::conv;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
const int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
@@ -173,7 +173,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
if(argc >= 5)
@@ -181,21 +181,21 @@ int main(int argc, char* argv[])
params = ParseConvParams(argc, argv);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
@@ -209,7 +209,7 @@ int main(int argc, char* argv[])
// bias: assume contiguous 1d vector
Tensor<OutDataType> bias(
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K)})));
HostTensorDescriptor(std::vector<std::size_t>({static_cast<std::size_t>(params.K_)})));
// residual: assume same layout as output tensor
Tensor<OutDataType> residual(get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
@@ -259,16 +259,16 @@ int main(int argc, char* argv[])
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(bias_device_buf.GetDeviceBuffer()),
static_cast<const OutDataType*>(resi_device_buf.GetDeviceBuffer()),
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op);
@@ -280,20 +280,20 @@ int main(int argc, char* argv[])
"not support this problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops(
params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype =
get_btype<InDataType, WeiDataType, OutDataType>(params.N,
params.C,
params.K,
params.input_spatial_lengths,
params.filter_spatial_lengths,
get_btype<InDataType, WeiDataType, OutDataType>(params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths) +
sizeof(OutDataType) * (params.K) +
sizeof(OutDataType) * (params.K_) +
sizeof(OutDataType) *
(params.N * params.K * output_spatial_lengths[0] * output_spatial_lengths[1]);
(params.N_ * params.K_ * output_spatial_lengths[0] * output_spatial_lengths[1]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
@@ -310,17 +310,18 @@ int main(int argc, char* argv[])
host_output,
bias,
residual,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data());
ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
return ck::utils::check_err(device_output.mData, host_output.mData) ? 0 : 1;
}
return 0;
}

8
example/09_convnd_fwd/CMakeLists.txt
View File

@@ -1,6 +1,6 @@
add_example_executable(example_convnd_fwd_xdl convnd_fwd_xdl.cpp)
target_link_libraries(example_convnd_fwd_xdl PRIVATE conv_fwd_util)
add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp)
target_link_libraries(example_convnd_fwd_xdl_int8 PRIVATE conv_fwd_util)
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
target_link_libraries(example_convnd_fwd_xdl_fp16 PRIVATE conv_fwd_util)
target_link_libraries(example_convnd_fwd_xdl_fp32 PRIVATE conv_util)
target_link_libraries(example_convnd_fwd_xdl_int8 PRIVATE conv_util)
target_link_libraries(example_convnd_fwd_xdl_fp16 PRIVATE conv_util)

113
example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
View File

@@ -5,7 +5,7 @@
#include "check_err.hpp"
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "device.hpp"
#include "device_tensor.hpp"
#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
@@ -43,10 +43,10 @@ template <ck::index_t NumDimSpatial>
using DeviceConvNDFwdInstance = ck::tensor_operation::device::
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
// clang-format off
InDataType, //
InDataType, //
WeiDataType, //
OutDataType, //
AccDataType, //
AccDataType, //
InElementOp, // Input Elementwise Operation
WeiElementOp, // Weights Elementwise Operation
OutElementOp, // Output Elementwise Operation
@@ -110,7 +110,7 @@ void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: run kernel # of times (>1)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
@@ -137,40 +137,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, cha
ck::utils::conv::ConvParams params;
int arg_idx = 5;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
@@ -182,9 +182,9 @@ int main(int argc, char* argv[])
{
using namespace ck::utils::conv;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
@@ -193,7 +193,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
}
@@ -202,21 +202,21 @@ int main(int argc, char* argv[])
params = parse_conv_params(num_dim_spatial, argc, argv);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
@@ -256,16 +256,16 @@ int main(int argc, char* argv[])
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
@@ -277,16 +277,16 @@ int main(int argc, char* argv[])
"not support this Conv problem");
}
float ave_time = invoker->Run(argument.get(), nrepeat);
float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops(
params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype = get_btype<InDataType, WeiDataType, OutDataType>(
params.N,
params.C,
params.K,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -302,18 +302,18 @@ int main(int argc, char* argv[])
auto ref_argument = ref_conv.MakeArgument(input,
weights,
host_output,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data());
ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
return ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f) ? 0 : 1;
};
switch(num_dim_spatial)
@@ -338,4 +338,5 @@ int main(int argc, char* argv[])
}
}
}
return 0;
}

118
example/09_convnd_fwd/convnd_fwd_xdl.cpp → example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
View File

@@ -5,7 +5,7 @@
#include "check_err.hpp"
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "device.hpp"
#include "device_tensor.hpp"
#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
@@ -39,10 +39,10 @@ template <ck::index_t NumDimSpatial>
using DeviceConvNDFwdInstance = ck::tensor_operation::device::
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
// clang-format off
InDataType, //
InDataType, //
WeiDataType, //
OutDataType, //
AccDataType, //
AccDataType, //
InElementOp, // Input Elementwise Operation
WeiElementOp, // Weights Elementwise Operation
OutElementOp, // Output Elementwise Operation
@@ -107,7 +107,7 @@ void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: run kernel # of times (>1)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
@@ -134,40 +134,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, cha
ck::utils::conv::ConvParams params;
int arg_idx = 5;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
@@ -179,9 +179,9 @@ int main(int argc, char* argv[])
{
using namespace ck::utils::conv;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
@@ -190,7 +190,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
}
@@ -199,21 +199,21 @@ int main(int argc, char* argv[])
params = parse_conv_params(num_dim_spatial, argc, argv);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
@@ -255,16 +255,16 @@ int main(int argc, char* argv[])
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
@@ -276,16 +276,16 @@ int main(int argc, char* argv[])
"not support this Conv problem");
}
float ave_time = invoker->Run(argument.get(), nrepeat);
float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops(
params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype =
get_btype<InDataType, WeiDataType, OutDataType>(params.N,
params.C,
params.K,
params.input_spatial_lengths,
params.filter_spatial_lengths,
get_btype<InDataType, WeiDataType, OutDataType>(params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -301,18 +301,23 @@ int main(int argc, char* argv[])
auto ref_argument = ref_conv.MakeArgument(input,
weights,
host_output,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data());
ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
return ck::utils::check_err(device_output.mData,
host_output.mData,
"Error: incorrect results!",
1e-5f,
1e-4f)
? 0
: 1;
};
switch(num_dim_spatial)
@@ -337,4 +342,5 @@ int main(int argc, char* argv[])
}
}
}
return 0;
}

113
example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
View File

@@ -5,7 +5,7 @@
#include "check_err.hpp"
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "device.hpp"
#include "device_tensor.hpp"
#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
@@ -45,10 +45,10 @@ template <ck::index_t NumDimSpatial>
using DeviceConvNDFwdInstance = ck::tensor_operation::device::
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
// clang-format off
InDataType, //
InDataType, //
WeiDataType, //
OutDataType, //
AccDataType, //
AccDataType, //
InElementOp, // Input Elementwise Operation
WeiElementOp, // Weights Elementwise Operation
OutElementOp, // Output Elementwise Operation
@@ -112,7 +112,7 @@ void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: run kernel # of times (>1)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
@@ -139,40 +139,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, cha
ck::utils::conv::ConvParams params;
int arg_idx = 5;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
@@ -184,9 +184,9 @@ int main(int argc, char* argv[])
{
using namespace ck::utils::conv;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
@@ -195,7 +195,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
}
@@ -204,21 +204,21 @@ int main(int argc, char* argv[])
params = parse_conv_params(num_dim_spatial, argc, argv);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
@@ -258,16 +258,16 @@ int main(int argc, char* argv[])
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
@@ -279,16 +279,16 @@ int main(int argc, char* argv[])
"not support this Conv problem");
}
float ave_time = invoker->Run(argument.get(), nrepeat);
float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops(
params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype = get_btype<InDataType, WeiDataType, OutDataType>(
params.N,
params.C,
params.K,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -304,18 +304,18 @@ int main(int argc, char* argv[])
auto ref_argument = ref_conv.MakeArgument(input,
weights,
host_output,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data());
ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
return ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f) ? 0 : 1;
};
switch(num_dim_spatial)
@@ -340,4 +340,5 @@ int main(int argc, char* argv[])
}
}
}
return 0;
}

2
example/10_conv2d_bwd_data/CMakeLists.txt
View File

@@ -1,2 +1,2 @@
add_example_executable(example_conv2d_bwd_data_xdl conv2d_bwd_data_xdl.cpp)
target_link_libraries(example_conv2d_bwd_data_xdl PRIVATE conv_fwd_util)
target_link_libraries(example_conv2d_bwd_data_xdl PRIVATE conv_util)

20
example/10_conv2d_bwd_data/conv2d_bwd_data_xdl.cpp
View File

@@ -77,9 +77,9 @@ using ReferenceConvBwdInstance = ck::tensor_operation::host::ReferenceConvBwdDat
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// Conv shape
ck::index_t N = 128;
@@ -102,13 +102,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = 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]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
@@ -130,7 +130,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
@@ -214,7 +214,7 @@ int main(int argc, char* argv[])
"not support this Conv problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
@@ -249,6 +249,10 @@ int main(int argc, char* argv[])
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
ck::utils::check_err(in_n_c_hi_wi_device_result.mData, in_n_c_hi_wi_host_result.mData);
return ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData)
? 0
: 1;
}
return 0;
}

2
example/11_conv2d_bwd_weight/CMakeLists.txt
View File

@@ -1,2 +1,2 @@
add_example_executable(example_conv2d_bwd_weight_xdl conv2d_bwd_weight_xdl.cpp)
target_link_libraries(example_conv2d_bwd_weight_xdl PRIVATE conv_fwd_util)
target_link_libraries(example_conv2d_bwd_weight_xdl PRIVATE conv_util)

19
example/11_conv2d_bwd_weight/conv2d_bwd_weight_xdl.cpp
View File

@@ -82,9 +82,9 @@ using ReferenceConvBwdWeightInstance =
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int do_log = 0;
int split_k = 4;
@@ -109,7 +109,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
}
@@ -117,7 +117,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
@@ -141,7 +141,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\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, "
@@ -246,7 +246,7 @@ int main(int argc, char* argv[])
return 1;
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
@@ -291,6 +291,9 @@ int main(int argc, char* argv[])
LogRangeAsType<float>(std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
}
ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData);
return ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData)
? 0
: 1;
}
return 0;
}

2
example/12_reduce/CMakeLists.txt
View File

@@ -1 +1 @@
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp -D 16,64,32,960 -v 1 1 10)

20
example/12_reduce/reduce_blockwise.cpp
View File

@@ -116,10 +116,9 @@ class SimpleAppArgs
std::vector<size_t> inLengths;
std::vector<float> scales;
bool do_verification = false;
int init_method = 1;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
public:
void show_usage(const char* cmd)
@@ -135,7 +134,7 @@ class SimpleAppArgs
std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
"value, 3=decimal value)"
<< std::endl;
std::cout << "Arg2 -- number of repeats to run the kernel" << std::endl;
std::cout << "Arg2 -- time kernel (0=n0, 1=yes)" << std::endl;
};
int processArgs(int argc, char* argv[])
@@ -182,7 +181,7 @@ class SimpleAppArgs
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
init_method = std::atoi(argv[optind++]);
nrepeat = std::atoi(argv[optind]);
time_kernel = std::atoi(argv[optind]);
if(scales.empty())
{
@@ -352,7 +351,7 @@ int main(int argc, char* argv[])
auto invoker_ptr = reduce.MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), args.nrepeat);
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, args.time_kernel});
std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InDataType) +
invariant_total_length * sizeof(OutDataType);
@@ -362,16 +361,17 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
<< std::endl;
bool pass = true;
if(args.do_verification)
{
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
pass &= ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
pass &= ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
};
};
return pass ? 0 : 1;
}

22
example/13_pool2d_fwd/pool2d_fwd.cpp
View File

@@ -149,9 +149,9 @@ int main(int argc, char* argv[])
{
using namespace ck::host_reduce;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// Pool shape
ck::index_t N = 128;
@@ -171,13 +171,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = 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]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
C = std::stoi(argv[5]);
@@ -196,7 +196,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
@@ -271,7 +271,7 @@ int main(int argc, char* argv[])
"not support this problem");
}
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
@@ -285,6 +285,7 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
pool_host_verify<InDataType,
@@ -302,14 +303,15 @@ int main(int argc, char* argv[])
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
pass &= ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
if constexpr(NeedIndices)
{
out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
// ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
// out_indices_n_c_ho_wo_host.mData);;
pass &= ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
out_indices_n_c_ho_wo_host.mData);
};
}
return pass ? 0 : 1;
}

16
example/14_gemm_xdl_requant_relu_requant/gemm_xdl_requant_relu_requant_int8.cpp
View File

@@ -105,9 +105,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
@@ -125,13 +125,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -145,7 +145,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
@@ -219,7 +219,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
@@ -244,7 +244,7 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;

17
example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
View File

@@ -60,21 +60,21 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
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("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
@@ -202,7 +202,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -211,6 +211,7 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
@@ -227,9 +228,9 @@ int main(int argc, char* argv[])
c_element_op);
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
}
}
return 0;
return pass ? 0 : 1;
}

58
example/16_gemm_reduce/gemm_reduce_xdl_fp16.cpp
View File

@@ -4,6 +4,7 @@
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
@@ -58,9 +59,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 1;
bool do_verification = true;
int init_method = 1;
int nrepeat = 5;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
@@ -79,13 +80,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -99,7 +100,7 @@ int main(int argc, char* argv[])
{
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
@@ -192,30 +193,13 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
// warm up
invoker.Run(argument);
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker.Run(argument);
timer.End();
total_time += timer.GetElapsedTime();
}
float ave_time = total_time / nrepeat;
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
@@ -228,6 +212,7 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
@@ -264,10 +249,19 @@ int main(int argc, char* argv[])
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
}
check_error(c_m_n_host_result, c_m_n_device_result);
check_error(d0_m_host_result, d0_m_device_result);
check_error(d1_m_host_result, d1_m_device_result);
pass &= ck::utils::check_err(
c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c");
pass &= ck::utils::check_err(d0_m_device_result.mData,
d0_m_host_result.mData,
"Error: Incorrect results d0",
1e-3,
1e-3);
pass &= ck::utils::check_err(d1_m_device_result.mData,
d1_m_host_result.mData,
"Error: Incorrect results d1",
1e-3,
1e-3);
}
return 0;
return pass ? 0 : 1;
}

2
example/17_convnd_bwd_data_xdl/CMakeLists.txt
View File

@@ -1,2 +1,2 @@
add_example_executable(example_convnd_bwd_data_xdl convnd_bwd_data_xdl.cpp)
target_link_libraries(example_convnd_bwd_data_xdl PRIVATE conv_fwd_util)
target_link_libraries(example_convnd_bwd_data_xdl PRIVATE conv_util)

114
example/17_convnd_bwd_data_xdl/convnd_bwd_data_xdl.cpp
View File

@@ -6,7 +6,7 @@
#include <half.hpp>
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
@@ -87,7 +87,7 @@ void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
<< "arg3: run kernel # of times (>1)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
@@ -105,40 +105,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
ck::utils::conv::ConvParams params;
int arg_idx = 5;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
@@ -165,25 +165,25 @@ DeviceConvBwdDataBasePtr get_conv_instance(int num_dim_spatial)
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
params.C = 128;
params.C_ = 128;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc > 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
// check args number
int conv_args = 3 + num_dim_spatial * 6;
@@ -202,21 +202,21 @@ int main(int argc, char* argv[])
exit(1);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
@@ -263,16 +263,16 @@ int main(int argc, char* argv[])
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
@@ -284,16 +284,16 @@ int main(int argc, char* argv[])
"not support this Conv problem");
}
float ave_time = invoker->Run(argument.get(), nrepeat);
float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = ck::utils::conv::get_flops(
params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
params.N,
params.C,
params.K,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
@@ -310,10 +310,10 @@ int main(int argc, char* argv[])
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
wei_k_c_y_x,
out_n_k_ho_wo,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
@@ -322,7 +322,10 @@ int main(int argc, char* argv[])
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);
return ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData)
? 0
: 1;
};
switch(num_dim_spatial)
@@ -347,4 +350,5 @@ int main(int argc, char* argv[])
}
}
}
return 0;
}

73
example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
View File

@@ -4,6 +4,7 @@
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
@@ -57,18 +58,18 @@ using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 1;
bool do_verification = true;
int init_method = 1;
int nrepeat = 5;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t M = 2048;
ck::index_t N = 1920;
ck::index_t K = 2048;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
ck::index_t StrideA = 2048;
ck::index_t StrideB = 2048;
ck::index_t StrideC = 1920;
ck::index_t BatchCount = 4;
@@ -80,13 +81,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -96,13 +97,13 @@ int main(int argc, char* argv[])
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
BatchCount = std::stoi(argv[9]);
BatchCount = std::stoi(argv[10]);
}
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("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, BatchCount\n");
exit(0);
}
@@ -204,30 +205,13 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
// warm up
invoker.Run(argument);
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker.Run(argument);
timer.End();
total_time += timer.GetElapsedTime();
}
float ave_time = total_time / nrepeat;
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
@@ -241,6 +225,7 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< batched_gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
@@ -264,7 +249,7 @@ int main(int argc, char* argv[])
for(int n = 0; n < N; ++n)
{
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(m, n));
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
@@ -277,10 +262,18 @@ int main(int argc, char* argv[])
}
}
check_error(c_g_m_n_host_result, c_g_m_n_device_result);
check_error(d0_g_m_host_result, d0_g_m_device_result);
check_error(d1_g_m_host_result, d1_g_m_device_result);
pass &= ck::utils::check_err(c_g_m_n_host_result.mData, c_g_m_n_device_result.mData);
pass &= ck::utils::check_err(d0_g_m_device_result.mData,
d0_g_m_host_result.mData,
"Error: Incorrect results! D0",
1e-3,
1e-3);
pass &= ck::utils::check_err(d1_g_m_device_result.mData,
d1_g_m_host_result.mData,
"Error: Incorrect results! D1",
1e-3,
1e-3);
}
return 0;
return pass ? 0 : 1;
}

12
example/19_cgemm/cgemm_xdl_bf16.cpp
View File

@@ -88,9 +88,9 @@ using ReferenceCGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// CGEMM shape
ck::index_t M = 3840;
@@ -105,13 +105,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
@@ -223,7 +223,7 @@ int main(int argc, char* argv[])
"not support this CGEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(8) * M * N * K;
std::size_t num_btype = std::size_t(2) * sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +

13
example/CMakeLists.txt
View File

@@ -19,9 +19,18 @@ include_directories(BEFORE
add_custom_target(examples)
function(add_example_executable EXAMPLE_NAME)
function(add_example_executable EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_NAME} ${ARGN})
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
add_dependencies(examples ${EXAMPLE_NAME})
add_dependencies(check ${EXAMPLE_NAME})
endfunction(add_example_executable EXAMPLE_NAME)
function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
add_dependencies(examples ${EXAMPLE_NAME})
endfunction(add_example_executable EXAMPLE_NAME)