Skip to content

GitLab

Commit a3b4c5cb authored by wangshaojie6's avatar wangshaojie6
Browse files

merge develop branch and add gridwise pipeline v3

parents 48918ab9 1677cf70
Hide whitespace changes
Inline Side-by-side
Showing with 1565 additions and 420 deletions
example/07_conv2d_fwd_bias_relu_add/CMakeLists.txt
View file @ a3b4c5cb
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)
example/07_conv2d_fwd_bias_relu_add/conv2d_fwd_xdl_bias_relu_add.cpp
View file @ a3b4c5cb
......@@ -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));
......@@ -224,10 +224,10 @@ int main(int argc, char* argv[])
{
case 0: break;
case 1:
input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
residual.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-2, 2});
weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-2, 2});
bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
residual.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-2, 2});
break;
default:
input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
......@@ -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;
}
example/09_convnd_fwd/CMakeLists.txt
View file @ a3b4c5cb
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)
# FIXME: re-enable this exampe as test when SWDEV-335738 is fixed
add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp)
target_link_libraries(example_convnd_fwd_xdl_fp64 PRIVATE conv_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)
example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
View file @ a3b4c5cb
......@@ -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{});
......@@ -277,21 +277,21 @@ 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;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " << conv->GetTypeString()
<< std::endl;
if(do_verification)
......@@ -302,40 +302,38 @@ 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)
{
case 3: {
auto ref_conv = ReferenceConvNDFwdInstance<3>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
case 2: {
auto ref_conv = ReferenceConvNDFwdInstance<2>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
case 1: {
auto ref_conv = ReferenceConvNDFwdInstance<1>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
return 0;
}
example/09_convnd_fwd/convnd_fwd_xdl.cpp → example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
View file @ a3b4c5cb
......@@ -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;
......@@ -301,40 +301,43 @@ 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)
{
case 3: {
auto ref_conv = ReferenceConvNDFwdInstance<3>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
case 2: {
auto ref_conv = ReferenceConvNDFwdInstance<2>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
case 1: {
auto ref_conv = ReferenceConvNDFwdInstance<1>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
return 0;
}
example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp 0 → 100644
View file @ a3b4c5cb
#include <cstdlib>
#include <iostream>
#include <numeric>
#include <type_traits>
#include "check_err.hpp"
#include "config.hpp"
#include "conv_util.hpp"
#include "device.hpp"
#include "device_tensor.hpp"
#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
#include "element_wise_operation.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "reference_conv_fwd.hpp"
#include "tensor_layout.hpp"
namespace {
using InDataType = double;
using WeiDataType = double;
using OutDataType = double;
using AccDataType = double;
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 ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
using DeviceConvFwdBasePtr =
ck::tensor_operation::device::DeviceConvFwdPtr<InElementOp, WeiElementOp, OutElementOp>;
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, //
WeiDataType, //
OutDataType, //
AccDataType, //
InElementOp, // Input Elementwise Operation
WeiElementOp, // Weights Elementwise Operation
OutElementOp, // Output Elementwise Operation
ConvFwdDefault, // ConvForwardSpecialization
NumDimSpatial, // NumDimSpatial
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
2, // K1
16, // MPerXDL
16, // NPerXDL
4, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
2, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
2, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockTransferAddExtraN
7, // CThreadTransferSrcDstVectorDim
1>; // CThreadTransferDstScalarPerVector
// clang-format on
template <ck::index_t NumDimSpatial>
using ReferenceConvNDFwdInstance = ck::tensor_operation::host::ReferenceConvFwd<InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
NumDimSpatial>;
DeviceConvFwdBasePtr get_conv_instance(int num_dim_spatial)
{
switch(num_dim_spatial)
{
case 3: {
return std::make_unique<DeviceConvNDFwdInstance<3>>();
}
case 2: {
return std::make_unique<DeviceConvNDFwdInstance<2>>();
}
case 1: {
return std::make_unique<DeviceConvNDFwdInstance<1>>();
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
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"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
<< " <filter spatial dimensions>, (ie Y, X for 2D)\n"
<< " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
<< " <strides>, (ie Sy, Sx for 2D)\n"
<< " <dilations>, (ie Dy, Dx for 2D)\n"
<< " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
<< " <right padding>, (ie RightPy, RightPx for 2D)\n"
<< std::endl;
}
ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, char* argv[])
{
// (N, K, C) + num_dim_spatial * 6 (filter, input, strides, dilations, pad left, pad right)
int conv_args = 3 + num_dim_spatial * 6;
int cmdline_nargs = conv_args + 5;
if(cmdline_nargs != argc)
{
print_use_msg();
exit(0);
}
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.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.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.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_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.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_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++]);
}
return params;
}
} // anonymous namespace
int main(int argc, char* argv[])
{
using namespace ck::utils::conv;
bool do_verification = 0;
int init_method = 0;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
if(argc >= 5)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
}
if(argc >= 6)
{
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_)};
input_dims.insert(std::end(input_dims),
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_)};
filter_dims.insert(std::end(filter_dims),
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_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
Tensor<InDataType> input(get_input_host_tensor_descriptor(input_dims, num_dim_spatial));
Tensor<WeiDataType> weights(get_filters_host_tensor_descriptor(filter_dims, num_dim_spatial));
Tensor<OutDataType> host_output(
get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
Tensor<OutDataType> device_output(
get_output_host_tensor_descriptor(output_dims, num_dim_spatial));
std::cout << "input: " << input.mDesc << std::endl;
std::cout << "weights: " << weights.mDesc << std::endl;
std::cout << "output: " << host_output.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
weights.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
case 2:
input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
weights.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
break;
default:
input.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
weights.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem in_device_buf(sizeof(InDataType) * input.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * weights.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * device_output.mDesc.GetElementSpace());
in_device_buf.ToDevice(input.mData.data());
wei_device_buf.ToDevice(weights.mData.data());
// do GEMM
auto conv = get_conv_instance(num_dim_spatial);
auto invoker = conv->MakeInvokerPointer();
auto argument =
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_,
output_spatial_lengths,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
if(!conv->IsSupportedArgument(argument.get()))
{
throw std::runtime_error(
"wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem");
}
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);
std::size_t num_btype =
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;
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 verify_f = [&input, &weights, &host_output, &params, &out_device_buf, &device_output](
const auto& ref_conv) {
auto ref_invoker = ref_conv.MakeInvoker();
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_,
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);
};
switch(num_dim_spatial)
{
case 3: {
auto ref_conv = ReferenceConvNDFwdInstance<3>();
verify_f(ref_conv);
break;
}
case 2: {
auto ref_conv = ReferenceConvNDFwdInstance<2>();
verify_f(ref_conv);
break;
}
case 1: {
auto ref_conv = ReferenceConvNDFwdInstance<1>();
verify_f(ref_conv);
break;
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
}
example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
View file @ a3b4c5cb
......@@ -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;
......@@ -304,40 +304,38 @@ 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)
{
case 3: {
auto ref_conv = ReferenceConvNDFwdInstance<3>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
case 2: {
auto ref_conv = ReferenceConvNDFwdInstance<2>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
case 1: {
auto ref_conv = ReferenceConvNDFwdInstance<1>();
verify_f(ref_conv);
break;
return verify_f(ref_conv);
}
default: {
throw std::runtime_error("Unsupported number of spatial dimensions provided!");
}
}
}
return 0;
}
example/10_conv2d_bwd_data/CMakeLists.txt
View file @ a3b4c5cb
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)
example/10_conv2d_bwd_data/conv2d_bwd_data_xdl.cpp
View file @ a3b4c5cb
......@@ -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;
}
example/11_conv2d_bwd_weight/CMakeLists.txt
View file @ a3b4c5cb
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)
example/11_conv2d_bwd_weight/conv2d_bwd_weight_xdl.cpp
View file @ a3b4c5cb
......@@ -68,7 +68,7 @@ using DeviceConvBwdWeightInstance = ck::tensor_operation::device::
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
S<1, 32, 1, 8>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
......@@ -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;
}
example/12_reduce/CMakeLists.txt
View file @ a3b4c5cb
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
add_example_executable(example_reduce_blockwise_two_call reduce_blockwise_two_call.cpp)
example/12_reduce/README.md
View file @ a3b4c5cb
......@@ -5,23 +5,38 @@
# -D <xxx> : input 4-d tensor lengths
# -v <x> : verification (0=no, 1=yes)
#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg2: run kernel # of times (>1)
./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 10
#arg2: time kernel (0=no, 1=yes)
./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 1
```
Result
```
./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 1
launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 3 times...
Perf: 0.23536 ms, 267.32 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
error: 0
max_diff: 0, 529, 529
root@dc-smc-18:/data/composable_kernel/Build3# bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 10
launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
Warm up
Warm up 1 time
Start running 10 times...
Perf: 0.282592 ms, 222.641 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
```
# Instructions for ```example_reduce_blockwise_two_call```
## Run ```example_reduce_blockwise_two_call```
```bash
#arg1: verification (0=no, 1=yes(
#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg3: time kernel (0=no, 1=yes)
./bin/example_reduce_blockwise_two_call 1 2 1
Result
```
./bin/example_reduce_blockwise_two_call 1 2 1
launch_and_time_kernel: grid_dim {204800, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
launch_and_time_kernel: grid_dim {6400, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 0.23392 ms, 268.966 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
error: 0
max_diff: 0, 528, 528
Perf: 2.1791 ms, 771.42 GB/s, DeviceReduceBlockWise<256,M_C32_S1,K_C8_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1> => DeviceReduceBlockWise<256,M_C256_S1,K_C1_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1>
```
example/12_reduce/reduce_blockwise.cpp
View file @ a3b4c5cb
......@@ -12,8 +12,8 @@
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_reduce_blockwise.hpp"
#include "host_reduce_util.hpp"
#include "device_reduce_multiblock.hpp"
#include "host_common_util.hpp"
#include "host_reduction.hpp"
#include "reduction_enums.hpp"
......@@ -30,9 +30,8 @@ constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
constexpr NanPropagation NanOpt = NanPropagation::PROPAGATE_NAN;
constexpr bool PropagateNan = (NanOpt == NanPropagation::NOT_PROPAGATE_NAN) ? false : true;
constexpr ReduceTensorIndices IndicesOpt = ReduceTensorIndices::NO_INDICES;
constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using InElementwiseOperation =
......@@ -40,86 +39,44 @@ using InElementwiseOperation =
using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
using DeviceReduceInstance = DeviceReduceBlockWise<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
false,
256,
4,
64,
1,
1,
0,
1,
1>;
using DeviceReduceInstance = DeviceReduceMultiBlock<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
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;
std::vector<size_t> inLengths = {16, 64, 32, 960};
std::vector<float> scales = {1.0f, 0.0f};
bool do_verification = false;
int init_method = 1;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
public:
void show_usage(const char* cmd)
......@@ -127,24 +84,24 @@ class SimpleAppArgs
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;
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=no, 1=yes)" << std::endl;
};
int processArgs(int argc, char* argv[])
{
unsigned int ch;
using ck::host_common::getTypeValuesFromString;
int ch;
while(1)
{
ch = getopt_long(argc, argv, "D:S:v:l:", long_options, &option_index);
ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
......@@ -155,12 +112,6 @@ class SimpleAppArgs
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!");
......@@ -182,7 +133,7 @@ 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 = static_cast<bool>(std::atoi(argv[optind]));
if(scales.empty())
{
......@@ -196,23 +147,21 @@ class SimpleAppArgs
int main(int argc, char* argv[])
{
using namespace ck::host_reduce;
const std::vector<int> reduceDims{0, 1, 2};
const std::vector<int> invariantDims{3};
SimpleAppArgs args;
if(args.processArgs(argc, argv) < 0)
return (-1);
if(argc > 1)
{
if(args.processArgs(argc, argv) < 0)
return (-1);
};
constexpr bool op_support_indices =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool NeedIndices =
(op_support_indices && (IndicesOpt != ReduceTensorIndices::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 =
......@@ -226,8 +175,7 @@ int main(int argc, char* argv[])
(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::NO_INDICES);
constexpr bool invalid_reduce_3 = (!op_support_indices && OutputIndex);
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3);
......@@ -295,51 +243,54 @@ int main(int argc, char* argv[])
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
DeviceMem out_indices_dev(indicesSizeInBytes);
DeviceMem out_index_dev(indicesSizeInBytes);
if(args.do_verification)
{
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOpId,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
Rank,
NumReduceDim,
PropagateNan,
NeedIndices>
OutputIndex>
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);
std::vector<ck::index_t> i_inLengths;
std::vector<ck::index_t> i_inStrides;
std::vector<ck::index_t> i_outLengths;
std::vector<ck::index_t> i_outStrides;
i_inLengths.assign(args.inLengths.begin(), args.inLengths.end());
i_inStrides.assign(inStrides.begin(), inStrides.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
auto reduce = DeviceReduceInstance{};
auto wsSizeInBytes = reduce.GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
auto argument_ptr =
reduce.MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
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)});
auto argument_ptr = reduce.MakeArgumentPointer(
i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
out_index_dev.GetDeviceBuffer(),
InElementwiseOperation{static_cast<int32_t>(reduce_total_length)},
AccElementwiseOperation{static_cast<int32_t>(reduce_total_length)});
if(!reduce.IsSupportedArgument(argument_ptr.get()))
{
......@@ -352,7 +303,7 @@ 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 +313,19 @@ 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 = pass && ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
if(OutputIndex)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
out_index_dev.FromDevice(out_indices.mData.data());
pass = pass && ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
};
};
return (pass ? 0 : 1);
}
example/12_reduce/reduce_blockwise_two_call.cpp 0 → 100644
View file @ a3b4c5cb
#include <iostream>
#include <numeric>
#include <sstream>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "check_err.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_multiblock.hpp"
#include "host_common_util.hpp"
#include "host_reduction.hpp"
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
using namespace ck;
using namespace ck::tensor_operation::device;
using InOutDataType = ck::half_t;
using InOutDataType = ck::half_t;
using AccDataType = float;
constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false;
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 PassThroughOp = tensor_operation::element_wise::UnaryIdentic<AccDataType, AccDataType>;
using DeviceReduceInstance_1 = DeviceReduceMultiBlock<InOutDataType,
AccDataType,
InOutDataType,
5, // Rank
1, // NumReduceDim
ReduceOperation,
InElementwiseOperation,
PassThroughOp,
InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
256,
32,
8,
1,
1,
1, // vector dim
1,
1>;
using DeviceReduceInstance_2 = DeviceReduceMultiBlock<InOutDataType,
AccDataType,
InOutDataType,
4, // Rank
1, // NumReduceDim
ReduceOperation,
PassThroughOp,
AccElementwiseOperation,
InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
256,
128,
2,
1,
1,
1, // vector dim
1,
1>;
static bool do_verify;
static int init_method;
static float alpha;
static float beta;
static bool time_kernel;
int main(int argc, char* argv[])
{
// used by the device reduction
const std::vector<int> reduceDims_1 = {4};
const std::vector<int> invariantDims_1 = {0, 1, 2, 3};
const std::vector<int> reduceDims_2 = {3};
const std::vector<int> invariantDims_2 = {0, 1, 2};
// used by the host reduction
const std::vector<int> reduceDims = {3, 4};
const std::vector<int> invariantDims = {0, 1, 2};
const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
// input lengths of the second reduction, which is also the output lengths of the first
// reduction
const std::vector<size_t> inLengths_2 = {64, 320, 80, 4};
const std::vector<size_t> outLengths = {64, 320, 80};
if(argc == 1)
{
do_verify = true;
init_method = 2;
time_kernel = true;
}
else if(argc == 4)
{
do_verify = static_cast<bool>(argv[1]);
init_method = atoi(argv[2]);
time_kernel = static_cast<bool>(atoi(argv[3]));
}
else
{
std::ostringstream ostr;
ostr << "Wrong parameter! " << std::endl
<< "Usage: " << argv[0] << "[verify 0/1] init_method time_kernel" << std::endl;
throw std::runtime_error(ostr.str());
};
alpha = 1.0f;
beta = 0.0f;
Tensor<InOutDataType> in_1(inLengths_1);
Tensor<InOutDataType> out_ref(outLengths);
Tensor<InOutDataType> in_2(inLengths_2); // also the output tensor of the first reduction
Tensor<InOutDataType> out(outLengths);
auto inStrides_1 = in_1.mDesc.GetStrides();
auto inStrides_2 = in_2.mDesc.GetStrides();
auto outStrides = out.mDesc.GetStrides();
size_t invariant_total_length = out.mDesc.GetElementSize();
size_t reduce_total_length = in_1.mDesc.GetElementSize() / invariant_total_length;
std::size_t num_thread = 1;
if(do_verify)
{
switch(init_method)
{
case 0: break;
case 1:
in_1.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
break;
case 2:
in_1.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
break;
default:
in_1.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
num_thread);
}
if(beta != 0.0f)
for(size_t i = 0; i < out_ref.mDesc.GetElementSpace(); i++)
out.mData[i] = out_ref.mData[i];
};
DeviceMem in_1_dev(sizeof(InOutDataType) * in_1.mDesc.GetElementSpace());
DeviceMem in_2_dev(sizeof(InOutDataType) * in_2.mDesc.GetElementSpace());
DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpace());
in_1_dev.ToDevice(in_1.mData.data());
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
if(do_verify)
{
ReductionHost<InOutDataType,
AccDataType,
InOutDataType,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
5, // Rank
2, // NumReduceDim
PropagateNan,
OutputIndex>
hostReduce(in_1.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(alpha, in_1.mData.data(), beta, out_ref.mData.data(), nullptr);
};
std::vector<ck::index_t> i_inLengths_1;
std::vector<ck::index_t> i_inStrides_1;
std::vector<ck::index_t> i_inLengths_2;
std::vector<ck::index_t> i_inStrides_2;
std::vector<ck::index_t> i_outLengths;
std::vector<ck::index_t> i_outStrides;
i_inLengths_1.assign(inLengths_1.begin(), inLengths_1.end());
i_inStrides_1.assign(inStrides_1.begin(), inStrides_1.end());
i_inLengths_2.assign(inLengths_2.begin(), inLengths_2.end());
i_inStrides_2.assign(inStrides_2.begin(), inStrides_2.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
auto reduce_1 = DeviceReduceInstance_1{};
auto argument_ptr_1 = reduce_1.MakeArgumentPointer(
i_inLengths_1,
i_inStrides_1,
i_inLengths_2,
i_inStrides_2,
reduceDims_1,
1.0f,
0.0f,
in_1_dev.GetDeviceBuffer(),
nullptr,
in_2_dev.GetDeviceBuffer(),
nullptr,
InElementwiseOperation{static_cast<int32_t>(reduce_total_length)},
PassThroughOp{});
if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
{
std::cout
<< "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
<< std::endl;
};
auto invoker_ptr_1 = reduce_1.MakeInvokerPointer();
auto reduce_2 = DeviceReduceInstance_2{};
auto argument_ptr_2 = reduce_2.MakeArgumentPointer(
i_inLengths_2,
i_inStrides_2,
i_outLengths,
i_outStrides,
reduceDims_2,
alpha,
beta,
in_2_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
nullptr,
PassThroughOp{},
AccElementwiseOperation{static_cast<int32_t>(reduce_total_length)});
if(!reduce_2.IsSupportedArgument(argument_ptr_2.get()))
{
std::cout
<< "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
<< std::endl;
};
auto invoker_ptr_2 = reduce_2.MakeInvokerPointer();
float avg_time_1 = invoker_ptr_1->Run(argument_ptr_1.get(), StreamConfig{nullptr, time_kernel});
float avg_time_2 = invoker_ptr_2->Run(argument_ptr_2.get(), StreamConfig{nullptr, time_kernel});
std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
invariant_total_length * sizeof(InOutDataType);
float gb_per_sec = num_bytes / 1.E6 / (avg_time_1 + avg_time_2);
std::cout << "Perf: " << avg_time_1 + avg_time_2 << " ms, " << gb_per_sec << " GB/s, "
<< reduce_1.GetTypeString() << " => " << reduce_2.GetTypeString() << std::endl;
bool pass = true;
if(do_verify)
{
out_dev.FromDevice(out.mData.data());
pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
};
return (pass ? 0 : 1);
}
example/13_pool2d_fwd/CMakeLists.txt
View file @ a3b4c5cb
add_example_executable(example_pool2d_fwd pool2d_fwd.cpp)
add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)
example/13_pool2d_fwd/README.md
View file @ a3b4c5cb
# Instructions for ```example_pool2d_fwd``` Example
# Instructions for ```example_pool2d_fwd``` Examples
## Run ```example_pool2d_fwd```
## Run ```example_pool2d_fwd_fp16```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg3: run kernel # of times (>1)
#arg3: time kernel (0=no, 1=yes)
#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
./bin/example_pool2d_fwd 1 1 10
./bin/example_pool2d_fwd_fp16 1 1 1
```
Result
......@@ -14,9 +14,28 @@ Result
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1}
Warm up
Warm up 1 time
Start running 10 times...
Perf: 0.415453 ms, 1.37996 TFlops, 749.726 GB/s
error: 0
max_diff: 0, 1, 1
Perf: 0.397436 ms, 1.44252 TFlops, 783.713 GB/s
```
## Run ```example_pool2d_fwd_fp32```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg3: time kernel (0=no, 1=yes)
#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
./bin/example_pool2d_fwd_fp32 1 1 1
```
Result
```
./bin/example_pool2d_fwd_fp32 1 1 1
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 1.01823 ms, 0.563045 TFlops, 611.8 GB/s
```
example/13_pool2d_fwd/pool2d_fwd.cpp → example/13_pool2d_fwd/pool2d_fwd_common.hpp
View file @ a3b4c5cb
#pragma once
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include "check_err.hpp"
#include "config.hpp"
......@@ -10,89 +8,67 @@
#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"
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "reduction_functions_accumulate.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::MAX;
#else
static constexpr auto ReduceOpId = ck::ReduceTensorOp::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
#include "device_pool2d_fwd_nhwc_nhwc.hpp"
template <typename InDataType,
typename OutDataType,
typename AccDataType,
typename IndexDataType,
ck::ReduceTensorOp ReduceOpId,
bool PropagateNan,
bool NeedIndices>
bool OutputIndex>
static void pool_host_verify(const Tensor<InDataType>& in,
Tensor<OutDataType>& out,
Tensor<int>& out_indices,
Tensor<IndexDataType>& 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 int32_t divider = window_spatial_lengths[0] * window_spatial_lengths[1];
const int divider = window_spatial_lengths[0] * window_spatial_lengths[1];
using ReduceOperation = typename ck::reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using InElementwiseOperation = typename ck::
reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation = typename ck::
reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
const auto PreUnaryOp = PreUnaryOpFn<AccDataType, ReduceOpId>(divider);
const auto PosUnaryOp = PosUnaryOpFn<AccDataType, ReduceOpId>(divider);
const InElementwiseOperation in_elementwise_op(divider);
const AccElementwiseOperation acc_elementwise_op(divider);
if constexpr(!NeedIndices)
if constexpr(!OutputIndex)
{
auto opReduce = ReduceOpFn<AccDataType, ReduceOpId>();
using Accumulation =
ck::detail::AccumulateWithNanCheck<PropagateNan, ReduceOperation, AccDataType>;
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
auto accuVal = ReduceOperation::GetIdentityValue();
for(int y = 0; y < window_spatial_lengths[0]; ++y)
for(ck::index_t 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)
ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
for(ck::index_t 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])
ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]))
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
PreUnaryOp(currVal);
in_elementwise_op(currVal, currVal);
binop_with_nan_check<AccDataType, PropagateNan>(opReduce, accuVal, currVal);
Accumulation::Calculate(accuVal, currVal);
}
}
}
PosUnaryOp(accuVal);
acc_elementwise_op(accuVal, accuVal);
out(n, c, ho, wo) = accuVal;
};
......@@ -105,33 +81,34 @@ static void pool_host_verify(const Tensor<InDataType>& in,
}
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)
using Accumulation = ck::detail::AccumulateWithIndexAndNanCheck<PropagateNan,
ReduceOperation,
AccDataType,
IndexDataType>;
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOperation::GetIdentityValue();
IndexDataType accuIndex = 0;
for(ck::index_t 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)
ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
{
int wi = wo * window_strides[1] + x - in_left_pads[1];
ck::index_t 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;
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
IndexDataType currIndex = y * window_spatial_lengths[1] + x;
PreUnaryOp(currVal);
in_elementwise_op(currVal, currVal);
binop_with_nan_check2<AccDataType, PropagateNan>(
opReduce, accuVal, currVal, accuIndex, currIndex);
Accumulation::Calculate(accuVal, currVal, accuIndex, currIndex);
}
}
}
PosUnaryOp(accuVal);
acc_elementwise_op(accuVal, accuVal);
out(n, c, ho, wo) = accuVal;
out_indices(n, c, ho, wo) = accuIndex;
......@@ -145,62 +122,44 @@ static void pool_host_verify(const Tensor<InDataType>& in,
};
}
int main(int argc, char* argv[])
template <typename InDataType,
typename OutDataType,
typename AccDataType,
typename IndexDataType,
typename InLayout,
typename OutLayout,
ck::ReduceTensorOp ReduceOpId,
bool PropagateNan,
bool OutputIndex>
bool pool_test(bool do_verification,
int init_method,
bool time_kernel,
ck::index_t N,
ck::index_t C,
ck::index_t Y,
ck::index_t X,
ck::index_t Hi,
ck::index_t Wi,
ck::index_t window_stride_h,
ck::index_t window_stride_w,
ck::index_t in_left_pad_h,
ck::index_t in_left_pad_w,
ck::index_t in_right_pad_h,
ck::index_t in_right_pad_w)
{
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);
}
using DevicePoolFwdInstance =
ck::tensor_operation::device::DevicePool2dFwd_Input_N_Hi_Wi_C_Output_N_Ho_Wo_C<
InDataType, // InDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
ReduceOpId,
OutputIndex,
64, // BlockSize
64, // ReduceMThreadClusterSize
1, // ReduceKThreadClusterSize
4, // ReduceMThreadSliceSize
1, // ReduceKThreadSliceSize
4>; // InSrcOutDstVectorSize
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;
......@@ -228,9 +187,11 @@ int main(int argc, char* argv[])
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<IndexDataType> 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{}));
Tensor<IndexDataType> 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;
......@@ -245,25 +206,25 @@ int main(int argc, char* argv[])
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) *
DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
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);
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<IndexDataType*>(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()))
{
......@@ -271,7 +232,7 @@ 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;
......@@ -285,14 +246,17 @@ 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,
OutDataType,
AccDataType,
IndexDataType,
ReduceOpId,
PropagateNan,
NeedIndices>(in_n_c_hi_wi,
OutputIndex>(in_n_c_hi_wi,
out_n_c_ho_wo_host,
out_indices_n_c_ho_wo_host,
window_spatial_lengths,
......@@ -302,14 +266,16 @@ 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 = pass && ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
if constexpr(NeedIndices)
if constexpr(OutputIndex)
{
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 = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
out_indices_n_c_ho_wo_host.mData);
};
}
}
return (pass);
};
example/13_pool2d_fwd/pool2d_fwd_fp16.cpp 0 → 100644
View file @ a3b4c5cb
#include <iostream>
#include <cstdlib>
#include "config.hpp"
#include "tensor_layout.hpp"
#include "reduction_enums.hpp"
#include "pool2d_fwd_common.hpp"
using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
using IndexDataType = int32_t;
using InLayout = ck::tensor_layout::convolution::NHWC;
using OutLayout = ck::tensor_layout::convolution::NHWC;
#if 1
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
#else
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
#endif
static constexpr bool OutputIndex = false;
static constexpr bool PropagateNan = false;
int main(int argc, char* argv[])
{
bool do_verification;
int init_method;
bool time_kernel;
// 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 == 1)
{
do_verification = true;
init_method = 1;
time_kernel = true;
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = static_cast<bool>(std::stoi(argv[3]));
}
else if(argc == 16)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = static_cast<bool>(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: time kernel (0=no, 1=yes)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
}
bool pass = pool_test<InDataType,
OutDataType,
AccDataType,
IndexDataType,
InLayout,
OutLayout,
ReduceOpId,
PropagateNan,
OutputIndex>(do_verification,
init_method,
time_kernel,
N,
C,
Y,
X,
Hi,
Wi,
window_stride_h,
window_stride_w,
in_left_pad_h,
in_left_pad_w,
in_right_pad_h,
in_right_pad_w);
return (pass ? 0 : 1);
}
example/13_pool2d_fwd/pool2d_fwd_fp32.cpp 0 → 100644
View file @ a3b4c5cb
#include <iostream>
#include <cstdlib>
#include "config.hpp"
#include "tensor_layout.hpp"
#include "reduction_enums.hpp"
#include "pool2d_fwd_common.hpp"
using InDataType = float;
using OutDataType = float;
using AccDataType = float;
using IndexDataType = int32_t;
using InLayout = ck::tensor_layout::convolution::NHWC;
using OutLayout = ck::tensor_layout::convolution::NHWC;
#if 1
static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
#else
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
#endif
static constexpr bool OutputIndex = false;
static constexpr bool PropagateNan = false;
int main(int argc, char* argv[])
{
bool do_verification;
int init_method;
bool time_kernel;
// 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 == 1)
{
do_verification = true;
init_method = 1;
time_kernel = true;
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = static_cast<bool>(std::stoi(argv[3]));
}
else if(argc == 16)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = static_cast<bool>(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: time kernel (0=no, 1=yes)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
}
bool pass = pool_test<InDataType,
OutDataType,
AccDataType,
IndexDataType,
InLayout,
OutLayout,
ReduceOpId,
PropagateNan,
OutputIndex>(do_verification,
init_method,
time_kernel,
N,
C,
Y,
X,
Hi,
Wi,
window_stride_h,
window_stride_w,
in_left_pad_h,
in_left_pad_w,
in_right_pad_h,
in_right_pad_w);
return (pass ? 0 : 1);
}
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