ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-11 17:00:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

update build

Browse Source
This commit is contained in:
Chao Liu
2019-07-05 15:33:58 -05:00
parent c15ff3c825
commit 3276a5e9b9
9 changed files with 90 additions and 88 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
driver/include/conv_common.hpp
View File

@@ -1,14 +1,14 @@
#ifndef CK_CONV_COMMON_HPP
#define CK_CONV_COMMON_HPP
#ifndef CONV_COMMON_HPP
#define CONV_COMMON_HPP
#include "ConstantTensorDescriptor.hpp"
using namespace ck;
// this is ugly, only for 4d
template <class InDesc, class WeiDesc>
constexpr auto get_convolution_output_default_4d_tensor_descriptor(InDesc, WeiDesc)
{
using namespace ck;
constexpr auto in_desc = InDesc{};
constexpr auto wei_desc = WeiDesc{};
@@ -45,6 +45,8 @@ template <class InDesc,
constexpr auto get_convolution_with_padding_output_default_4d_tensor_descriptor(
InDesc, WeiDesc, ConvStrides, ConvDilations, LowerPads, UpperPads)
{
using namespace ck;
constexpr auto in_desc = InDesc{};
constexpr auto wei_desc = WeiDesc{};
@@ -84,6 +86,8 @@ constexpr auto get_convolution_with_padding_output_default_4d_tensor_descriptor(
template <class InDesc, class WeiDesc, class OutDesc>
constexpr std::size_t calculate_convolution_flops(InDesc, WeiDesc, OutDesc)
{
using namespace ck;
constexpr auto wei_desc = WeiDesc{};
constexpr auto out_desc = OutDesc{};
@@ -107,6 +111,8 @@ constexpr std::size_t calculate_convolution_flops(InDesc, WeiDesc, OutDesc)
template <class Float, class InDesc, class WeiDesc, class OutDesc>
constexpr std::size_t calculate_convolution_memory_size(Float, InDesc, WeiDesc, OutDesc)
{
using namespace ck;
constexpr auto wei_desc = WeiDesc{};
constexpr auto out_desc = OutDesc{};

6
driver/include/device.hpp
View File

@@ -1,11 +1,9 @@
#ifndef CK_DEVICE_HPP
#define CK_DEVICE_HPP
#ifndef DEVICE_HPP
#define DEVICE_HPP
#include <memory>
#include "config.hpp"
using namespace ck;
struct DeviceMem
{
DeviceMem() = delete;

6
driver/include/device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp
View File

@@ -6,8 +6,6 @@
#include "gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp"
#include "gridwise_convolution_implicit_gemm_v3_nchw_cyxk_nkhw_lds_double_buffer.hpp"
using namespace ck;
template <class T, class InDesc, class WeiDesc, class OutDesc>
void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
const Tensor<T>& in_nchw,
@@ -17,6 +15,8 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
Tensor<T>& out_nkhw,
index_t nrepeat)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
@@ -135,7 +135,6 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
WeiBlockCopyClusterLengths_C_K,
WeiBlockCopyDataPerAccess_K>{};
#if 1
float time = launch_kernel(run_gridwise_convolution_kernel<decltype(gridwise_conv), T>,
dim3(GridSize),
dim3(BlockSize),
@@ -149,7 +148,6 @@ void device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(InDesc,
(float)calculate_convolution_flops(InDesc{}, WeiDesc{}, OutDesc{}) /
(std::size_t(1000) * 1000 * 1000) / time);
usleep(std::min(time * 1000, float(10000)));
#endif
}
out_nkhw_device_buf.FromDevice(out_nkhw.mData.data());

4
driver/include/device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp
View File

@@ -6,8 +6,6 @@
#include "gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp"
#include "gridwise_convolution_implicit_gemm_v4_nchw_kcyx_nkhw_lds_double_buffer.hpp"
using namespace ck;
template <class T,
class InDesc,
class WeiDesc,
@@ -24,6 +22,8 @@ void device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(InDesc,
ConvDilations,
index_t nrepeat)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};

47
driver/include/tensor.hpp
View File

@@ -1,5 +1,5 @@
#ifndef CK_TENSOR_HPP
#define CK_TENSOR_HPP
#ifndef TENSOR_HPP
#define TENSOR_HPP
#include <thread>
#include <vector>
@@ -8,6 +8,7 @@
#include <utility>
#include <cassert>
#include <iostream>
#include "common_header.hpp"
template <class Range>
std::ostream& LogRange(std::ostream& os, Range&& range, std::string delim)
@@ -269,4 +270,46 @@ struct Tensor
std::vector<T> mData;
};
// this is ugly, only for 4d
template <class TConstTensorDesc>
void ostream_ConstantTensorDescriptor(TConstTensorDesc, std::ostream& os = std::cout)
{
using namespace ck;
static_assert(TConstTensorDesc::nDim == 4, "nDim is not 4");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto desc = TConstTensorDesc{};
os << "Lengths: {" << desc.GetLength(I0) << ", " << desc.GetLength(I1) << ", "
<< desc.GetLength(I2) << ", " << desc.GetLength(I3) << "}, "
<< "Strides: {" << desc.GetStride(I0) << ", " << desc.GetStride(I1) << ", "
<< desc.GetStride(I2) << ", " << desc.GetStride(I3) << "}" << std::endl;
}
// this is ugly, only for 4d
template <class TConstTensorDesc>
auto make_TensorDescriptor(TConstTensorDesc)
{
using namespace ck;
static_assert(TConstTensorDesc::nDim == 4, "nDim is not 4");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto desc = TConstTensorDesc{};
std::initializer_list<index_t> lengths = {
desc.GetLength(I0), desc.GetLength(I1), desc.GetLength(I2), desc.GetLength(I3)};
std::initializer_list<index_t> strides = {
desc.GetStride(I0), desc.GetStride(I1), desc.GetStride(I2), desc.GetStride(I3)};
return TensorDescriptor(lengths, strides);
}
#endif

75
driver/src/driver.cpp
View File

@@ -14,8 +14,6 @@
#include "device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw.hpp"
#include "device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw.hpp"
using namespace ck;
struct GeneratorTensor_1
{
template <class... Is>
@@ -65,44 +63,6 @@ struct GeneratorTensor_Checkboard
}
};
// this is ugly, only for 4d
template <class TConstTensorDesc>
void ostream_ConstantTensorDescriptor(TConstTensorDesc, std::ostream& os = std::cout)
{
static_assert(TConstTensorDesc::nDim == 4, "nDim is not 4");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto desc = TConstTensorDesc{};
os << "Lengths: {" << desc.GetLength(I0) << ", " << desc.GetLength(I1) << ", "
<< desc.GetLength(I2) << ", " << desc.GetLength(I3) << "}, "
<< "Strides: {" << desc.GetStride(I0) << ", " << desc.GetStride(I1) << ", "
<< desc.GetStride(I2) << ", " << desc.GetStride(I3) << "}" << std::endl;
}
// this is ugly, only for 4d
template <class TConstTensorDesc>
auto make_TensorDescriptor(TConstTensorDesc)
{
static_assert(TConstTensorDesc::nDim == 4, "nDim is not 4");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto desc = TConstTensorDesc{};
std::initializer_list<index_t> lengths = {
desc.GetLength(I0), desc.GetLength(I1), desc.GetLength(I2), desc.GetLength(I3)};
std::initializer_list<index_t> strides = {
desc.GetStride(I0), desc.GetStride(I1), desc.GetStride(I2), desc.GetStride(I3)};
return TensorDescriptor(lengths, strides);
}
template <class TIn,
class TWei,
class TOut,
@@ -416,6 +376,8 @@ void check_error(const Tensor<T>& ref, const Tensor<T>& result)
int main(int argc, char* argv[])
{
using namespace ck;
#if 0
constexpr index_t N = 8;
constexpr index_t C = 16;
@@ -611,7 +573,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 0
#elif 1
// 1x1 filter, 8x8 image
// cudnn@V100 77%, ck@V100 76%, ck@P100 79%, ck@VII 51%
constexpr index_t N = 128;
@@ -787,7 +749,7 @@ int main(int argc, char* argv[])
constexpr index_t HPad = 0;
constexpr index_t WPad = 0;
#elif 1
#elif 0
// 1x1 filter, 7x7 image
// cudnn@V100 49%, ck@V100 50%, ck@P100 61%, ck@VII 52%
constexpr index_t N = 128;
@@ -859,30 +821,31 @@ int main(int argc, char* argv[])
#endif
}
#if 1
#if 0
device_convolution_direct_v2_nchw_kcyx_nkhw
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v1_chwn_cyxk_khwn
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v1_nchw_cyxk_nkhw
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v2_chwn_cyxk_khwn
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 0
device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw
device_convolution_implicit_gemm_v3_nchw_cyxk_nkhw(
in_nchw_desc, in_nchw, wei_kcyx_desc, wei_kcyx, out_nkhw_desc, out_nkhw_device, nrepeat);
#elif 1
device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw
#endif
(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
device_convolution_implicit_gemm_v4_nchw_kcyx_nkhw(in_nchw_desc,
in_nchw,
wei_kcyx_desc,
wei_kcyx,
out_nkhw_desc,
out_nkhw_device,
ConvStrides{},
ConvDilations{},
nrepeat);
#elif 0
device_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded(in_nchw_desc,
in_nchw,