ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-12 09:16:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'direct_fp16'

Browse Source
This commit is contained in:
Chao Liu
2019-03-22 16:46:41 -05:00
parent 2c9b8c2432 18a81e356e
commit fdaaaa500c
24 changed files with 1953 additions and 834 deletions
Download Patch File Download Diff File Expand all files Collapse all files

82
driver/device_direct_convolution_2.hpp → driver/device_direct_convolution_2_nchw_kcyx_nkhw.hpp
View File

@@ -1,16 +1,16 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "gridwise_direct_convolution_2.hip.hpp"
#include "gridwise_direct_convolution_2_nchw_kcyx_nkhw.hip.hpp"
template <class T, class InDesc, class WeiDesc, class OutDesc>
void device_direct_convolution_2(InDesc,
const Tensor<T>& in,
WeiDesc,
const Tensor<T>& wei,
OutDesc,
Tensor<T>& out,
unsigned nrepeat)
void device_direct_convolution_2_nchw_kcyx_nkhw(InDesc,
const Tensor<T>& in,
WeiDesc,
const Tensor<T>& wei,
OutDesc,
Tensor<T>& out,
unsigned nrepeat)
{
std::size_t data_sz = sizeof(T);
DeviceMem in_device_buf(data_sz * in.mDesc.GetElementSpace());
@@ -46,6 +46,27 @@ void device_direct_convolution_2(InDesc,
constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2;
constexpr unsigned InBlockCopyDataPerRead = 2;
constexpr unsigned WeiBlockCopyDataPerRead = 4;
constexpr unsigned BlockSize = 128;
#elif 1
// 3x3, 34x34, 128 thread, fp16
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 4;
constexpr unsigned HoPerBlock = 2;
constexpr unsigned WoPerBlock = 32;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2;
constexpr unsigned InBlockCopyDataPerRead = 2;
constexpr unsigned WeiBlockCopyDataPerRead = 4;
constexpr unsigned BlockSize = 128;
#endif
@@ -57,27 +78,30 @@ void device_direct_convolution_2(InDesc,
for(unsigned i = 0; i < nrepeat; ++i)
{
float time = launch_kernel(gridwise_direct_convolution_2<T,
InDesc,
WeiDesc,
OutDesc,
NPerBlock,
KPerBlock,
CPerBlock,
HoPerBlock,
WoPerBlock,
NPerThread,
KPerThread,
CPerThread,
HoPerThread,
WoPerThread,
BlockSize,
GridSize>,
dim3(GridSize),
dim3(BlockSize),
static_cast<T*>(in_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_device_buf.GetDeviceBuffer()));
float time =
launch_kernel(gridwise_direct_convolution_2_nchw_kcyx_nkhw<T,
InDesc,
WeiDesc,
OutDesc,
NPerBlock,
KPerBlock,
CPerBlock,
HoPerBlock,
WoPerBlock,
NPerThread,
KPerThread,
CPerThread,
HoPerThread,
WoPerThread,
InBlockCopyDataPerRead,
WeiBlockCopyDataPerRead,
BlockSize,
GridSize>,
dim3(GridSize),
dim3(BlockSize),
static_cast<T*>(in_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms\n", time);
usleep(std::min(time * 1000, float(10000)));

211
driver/device_direct_convolution_2_vectorized_nchw_kcyx_nkhw.hpp Normal file
View File

@@ -0,0 +1,211 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw.hip.hpp"
template <class TInWei, class TOut, class InDesc, class WeiDesc, class OutDesc>
void device_direct_convolution_2_vectorized_nchw_kcyx_nkhw(InDesc,
const Tensor<TInWei>& in_nchw,
WeiDesc,
const Tensor<TInWei>& wei_kcyx,
OutDesc,
Tensor<TOut>& out_nkhw,
unsigned nrepeat)
{
// this suppose in / wei data type is int8x4
constexpr unsigned NVector = 4;
using accum_t = int32_t;
using vector_t = vector_type<TInWei, NVector>;
using vector_mem_t = typename vector_t::MemoryType;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto in_nchw_desc = InDesc{};
constexpr auto wei_kcyx_desc = WeiDesc{};
constexpr auto out_nkhw_desc = OutDesc{};
constexpr unsigned Hi = in_nchw_desc.GetLength(I2);
constexpr unsigned Wi = in_nchw_desc.GetLength(I3);
constexpr unsigned N = out_nkhw_desc.GetLength(I0);
constexpr unsigned Ho = out_nkhw_desc.GetLength(I2);
constexpr unsigned Wo = out_nkhw_desc.GetLength(I3);
constexpr unsigned K = wei_kcyx_desc.GetLength(I0);
constexpr unsigned C = wei_kcyx_desc.GetLength(I1);
constexpr unsigned Y = wei_kcyx_desc.GetLength(I2);
constexpr unsigned X = wei_kcyx_desc.GetLength(I3);
// vectorized input
auto in_nchw_vec_desc = make_ConstantTensorDescriptor(Sequence<N, C / NVector, Hi, Wi>{});
ostream_ConstantTensorDescriptor(in_nchw_vec_desc, std::cout << "in_nchw_vec_desc: ");
Tensor<vector_mem_t> in_nchw_vec(make_TensorDescriptor(in_nchw_vec_desc));
auto f_vectorized_nchw = [&](auto n, auto c, auto h, auto w) {
#if 0
in_nchw_vec(n, c, h, w) = in_nchw(n, c, h, w);
#elif 0
in_nchw_vec(n, c, h, w) =
vector_t::Pack(in_nchw(n, 2 * c, h, w), in_nchw(n, 2 * c + 1, h, w));
#elif 1
in_nchw_vec(n, c, h, w) = vector_t::Pack(in_nchw(n, 4 * c, h, w),
in_nchw(n, 4 * c + 1, h, w),
in_nchw(n, 4 * c + 2, h, w),
in_nchw(n, 4 * c + 3, h, w));
#endif
};
make_ParallelTensorFunctor(f_vectorized_nchw, N, C / NVector, Hi, Wi)(
std::thread::hardware_concurrency());
// vectorize weight
auto wei_kcyx_vec_desc = make_ConstantTensorDescriptor(Sequence<K, C / NVector, Y, X>{});
ostream_ConstantTensorDescriptor(wei_kcyx_vec_desc, std::cout << "wei_kcyx_vec_desc: ");
Tensor<vector_mem_t> wei_kcyx_vec(make_TensorDescriptor(wei_kcyx_vec_desc));
auto f_vectorized_kcyx = [&](auto k, auto c, auto y, auto x) {
#if 0
wei_kcyx_vec(k, c, y, x) = wei_kcyx(k, c, y, x);
#elif 0
wei_kcyx_vec(k, c, y, x) =
vector_t::Pack(wei_kcyx(k, 2 * c, y, x), wei_kcyx(k, 2 * c + 1, y, x));
#elif 1
wei_kcyx_vec(k, c, y, x) = vector_t::Pack(wei_kcyx(k, 4 * c, y, x),
wei_kcyx(k, 4 * c + 1, y, x),
wei_kcyx(k, 4 * c + 2, y, x),
wei_kcyx(k, 4 * c + 3, y, x));
#endif
};
make_ParallelTensorFunctor(f_vectorized_kcyx, K, C / NVector, Y, X)(
std::thread::hardware_concurrency());
//
DeviceMem in_nchw_vec_device_buf(sizeof(vector_mem_t) * in_nchw_vec.mDesc.GetElementSpace());
DeviceMem wei_kcyx_vec_device_buf(sizeof(vector_mem_t) * wei_kcyx_vec.mDesc.GetElementSpace());
DeviceMem out_nkhw_device_buf(sizeof(TOut) * out_nkhw.mDesc.GetElementSpace());
in_nchw_vec_device_buf.ToDevice(in_nchw_vec.mData.data());
wei_kcyx_vec_device_buf.ToDevice(wei_kcyx_vec.mData.data());
out_nkhw_device_buf.ToDevice(out_nkhw.mData.data());
#if 0
// 3x3, 34x34, 128 thread, fp32, vector = 1
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 4;
constexpr unsigned HoPerBlock = 2;
constexpr unsigned WoPerBlock = 32;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 2;
constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2;
constexpr unsigned InBlockCopyDataPerRead = 2;
constexpr unsigned WeiBlockCopyDataPerRead = 2;
constexpr unsigned BlockSize = 128;
#elif 0
// 3x3, 34x34, 128 thread, fp32, vector = 2
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 2;
constexpr unsigned HoPerBlock = 2;
constexpr unsigned WoPerBlock = 32;
constexpr unsigned NPerThread = 2;
constexpr unsigned KPerThread = 4;
constexpr unsigned CPerThread = 1;
constexpr unsigned HoPerThread = 2;
constexpr unsigned WoPerThread = 2;
constexpr unsigned InBlockCopyDataPerRead = 2;
constexpr unsigned WeiBlockCopyDataPerRead = 2;
constexpr unsigned BlockSize = 128;
#elif 0
// 3x3, 34x34, 128 thread, int8, vector = 4
constexpr unsigned NPerBlock = 2;
constexpr unsigned KPerBlock = 32;
constexpr unsigned CPerBlock = 8;
constexpr unsigned HoPerBlock = 4;
constexpr unsigned WoPerBlock = 32;
constexpr unsigned NPerThread = 1;
constexpr unsigned KPerThread = 8;
constexpr unsigned CPerThread = 2;
constexpr unsigned HoPerThread = 4;
constexpr unsigned WoPerThread = 2;
constexpr unsigned InBlockCopyDataPerRead = 2;
constexpr unsigned WeiBlockCopyDataPerRead = 2;
constexpr unsigned BlockSize = 128;
#elif 1
// 1x1, 32x32, 128 thread, int8, vector = 4
constexpr unsigned NPerBlock = 1;
constexpr unsigned KPerBlock = 64;
constexpr unsigned CPerBlock = 16;
constexpr unsigned HoPerBlock = 4;
constexpr unsigned WoPerBlock = 32;
constexpr unsigned NPerThread = 1;
constexpr unsigned KPerThread = 8;
constexpr unsigned CPerThread = 2;
constexpr unsigned HoPerThread = 4;
constexpr unsigned WoPerThread = 2;
constexpr unsigned InBlockCopyDataPerRead = 2;
constexpr unsigned WeiBlockCopyDataPerRead = 2;
constexpr unsigned BlockSize = 128;
#endif
constexpr unsigned GridSize =
(N / NPerBlock) * (K / KPerBlock) * (Ho / HoPerBlock) * (Wo / WoPerBlock);
printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);
for(unsigned i = 0; i < nrepeat; ++i)
{
float time = launch_kernel(
gridwise_direct_convolution_2_vectorized_nchw_kcyx_nkhw<TInWei,
TOut,
accum_t,
decltype(in_nchw_vec_desc),
decltype(wei_kcyx_vec_desc),
decltype(out_nkhw_desc),
NVector,
NPerBlock,
KPerBlock,
CPerBlock,
HoPerBlock,
WoPerBlock,
NPerThread,
KPerThread,
CPerThread,
HoPerThread,
WoPerThread,
InBlockCopyDataPerRead,
WeiBlockCopyDataPerRead,
BlockSize,
GridSize>,
dim3(GridSize),
dim3(BlockSize),
static_cast<TInWei*>(in_nchw_vec_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(wei_kcyx_vec_device_buf.GetDeviceBuffer()),
static_cast<TInWei*>(out_nkhw_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms\n", time);
usleep(std::min(time * 1000, float(10000)));
}
out_nkhw_device_buf.FromDevice(out_nkhw.mData.data());
}

37
driver/device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp
View File

@@ -74,7 +74,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
wei_cyxk_device_buf.ToDevice(wei_cyxk.mData.data());
out_khwn_device_buf.ToDevice(out_khwn.mData.data());
#if 1
#if 0
// for 3x3, 34x34
constexpr unsigned NPerBlock = 16;
constexpr unsigned KPerBlock = 64;
@@ -213,7 +213,7 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr unsigned WoPerThread = 1;
constexpr unsigned BlockSize = 128;
#elif 1
#elif 0
// for 1x1, 28x28
constexpr unsigned NPerBlock = 16;
constexpr unsigned KPerBlock = 128;
@@ -245,6 +245,39 @@ void device_implicit_gemm_convolution_1_chwn_cyxk_khwn(InDesc,
constexpr unsigned OutThreadCopyDataPerWrite = 2;
constexpr unsigned BlockSize = 128;
#elif 1
// for 1x1, 14x14
constexpr unsigned NPerBlock = 16;
constexpr unsigned KPerBlock = 128;
constexpr unsigned CPerBlock = 8;
constexpr unsigned HoPerBlock = 2;
constexpr unsigned WoPerBlock = 2;
constexpr unsigned NPerThread = 4;
constexpr unsigned KPerThread = 16;
constexpr unsigned CPerThread = 1;
constexpr unsigned HoPerThread = 1;
constexpr unsigned WoPerThread = 1;
constexpr unsigned InBlockCopy_ThreadPerDimC = 8;
constexpr unsigned InBlockCopy_ThreadPerDimH = 2;
constexpr unsigned InBlockCopy_ThreadPerDimW = 2;
constexpr unsigned InBlockCopy_ThreadPerDimN = 4;
constexpr unsigned InBlockCopyDataPerRead = 4;
constexpr unsigned WeiBlockCopyDataPerRead = 4;
constexpr unsigned GemmMPerThreadSubC = 4;
constexpr unsigned GemmNPerThreadSubC = 4;
constexpr unsigned GemmMLevel0Cluster = 4;
constexpr unsigned GemmNLevel0Cluster = 2;
constexpr unsigned GemmMLevel1Cluster = 2;
constexpr unsigned GemmNLevel1Cluster = 4;
constexpr unsigned GemmKPerThreadLoop = 1;
constexpr unsigned OutThreadCopyDataPerWrite = 2;
constexpr unsigned BlockSize = 128;
#endif

71
driver/device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp
View File

@@ -1,6 +1,7 @@
#pragma once
#include <unistd.h>
#include "device.hpp"
#include "gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn.hip.hpp"
#include "gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer.hip.hpp"
template <class T, class InDesc, class WeiDesc, class OutDesc>
@@ -209,39 +210,43 @@ void device_implicit_gemm_convolution_2_chwn_cyxk_khwn(InDesc,
for(unsigned i = 0; i < nrepeat; ++i)
{
float time =
launch_kernel(gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer<
GridSize,
BlockSize,
T,
decltype(in_chwn_desc),
decltype(wei_cyxk_desc),
decltype(out_khwn_desc),
BPerBlock,
KPerBlock,
CPerBlock,
BPerThread,
KPerThread,
GemmThreadPerColumnPerCluster,
GemmThreadPerRowPerCluster,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
InBlockCopyThreadPerDim0,
InBlockCopyThreadPerDim1,
WeiBlockCopyThreadPerDim0,
WeiBlockCopyThreadPerDim1,
InBlockCopyDataPerRead,
WeiBlockCopyDataPerRead>,
dim3(GridSize),
dim3(BlockSize),
static_cast<T*>(in_chwn_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_cyxk_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_khwn_device_buf.GetDeviceBuffer()));
float time = launch_kernel(
#if 0
gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn
#else
gridwise_implicit_gemm_convolution_2_chwn_cyxk_khwn_lds_double_buffer
#endif
<GridSize,
BlockSize,
T,
decltype(in_chwn_desc),
decltype(wei_cyxk_desc),
decltype(out_khwn_desc),
BPerBlock,
KPerBlock,
CPerBlock,
BPerThread,
KPerThread,
GemmThreadPerColumnPerCluster,
GemmThreadPerRowPerCluster,
GemmMPerThreadSubC,
GemmNPerThreadSubC,
GemmMLevel0Cluster,
GemmNLevel0Cluster,
GemmMLevel1Cluster,
GemmNLevel1Cluster,
GemmKPerThreadLoop,
InBlockCopyThreadPerDim0,
InBlockCopyThreadPerDim1,
WeiBlockCopyThreadPerDim0,
WeiBlockCopyThreadPerDim1,
InBlockCopyDataPerRead,
WeiBlockCopyDataPerRead>,
dim3(GridSize),
dim3(BlockSize),
static_cast<T*>(in_chwn_device_buf.GetDeviceBuffer()),
static_cast<T*>(wei_cyxk_device_buf.GetDeviceBuffer()),
static_cast<T*>(out_khwn_device_buf.GetDeviceBuffer()));
printf("Elapsed time : %f ms\n", time);
usleep(std::min(time * 1000, float(10000)));

227
driver/driver.hip.cpp
View File

@@ -7,10 +7,11 @@
#include "tensor.hpp"
#include "ConstantTensorDescriptor.hip.hpp"
#include "conv_common.hip.hpp"
#include "device_direct_convolution_1.hpp"
#include "device_direct_convolution_2.hpp"
//#include "device_direct_convolution_1.hpp"
#include "device_direct_convolution_2_nchw_kcyx_nkhw.hpp"
//#include "device_direct_convolution_2_vectorized_nchw_kcyx_nkhw.hpp"
#include "device_implicit_gemm_convolution_1_chwn_cyxk_khwn.hpp"
#include "device_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hpp"
//#include "device_implicit_gemm_convolution_1_chwn_cyxk_khwn_padded.hpp"
#include "device_implicit_gemm_convolution_2_chwn_cyxk_khwn.hpp"
struct GeneratorTensor_1
@@ -34,25 +35,6 @@ struct GeneratorTensor_2
}
};
struct GeneratorTensor_3
{
template <class... Is>
double operator()(Is... is)
{
#if 0
std::initializer_list<std::size_t> ls = {static_cast<std::size_t>(is)...};
return std::accumulate(ls.begin(), ls.end(), std::size_t(0));
#elif 1
assert(sizeof...(Is) > 0);
std::initializer_list<std::size_t> ids = {static_cast<std::size_t>(is)...};
std::vector<std::size_t> lens(sizeof...(Is), 100);
std::vector<std::size_t> strides(sizeof...(Is), 1);
std::partial_sum(lens.rbegin(), lens.rbegin() + (sizeof...(Is)-1), strides.rbegin() + 1);
return std::inner_product(ids.begin(), ids.end(), strides.begin(), std::size_t(0)) + 1;
#endif
}
};
struct GeneratorTensor_Checkboard
{
template <class... Ts>
@@ -106,9 +88,12 @@ auto make_TensorDescriptor(TConstTensorDesc)
return TensorDescriptor(lengths, strides);
}
template <class T, class LowerPads, class UpperPads>
void host_direct_convolution(
const Tensor<T>& in_nchw, const Tensor<T>& wei_kcyx, Tensor<T>& out, LowerPads, UpperPads)
template <class TIn, class TWei, class TOut, class LowerPads, class UpperPads>
void host_direct_convolution(const Tensor<TIn>& in_nchw,
const Tensor<TWei>& wei_kcyx,
Tensor<TOut>& out_nkhw,
LowerPads,
UpperPads)
{
unsigned h_pad_low = LowerPads{}.Get(Number<0>{});
unsigned w_pad_low = LowerPads{}.Get(Number<1>{});
@@ -129,26 +114,29 @@ void host_direct_convolution(
if(hi >= 0 && hi < in_nchw.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in_nchw.mDesc.GetLengths()[3])
{
v += in_nchw(n, c, hi, wi) * wei_kcyx(k, c, y, x);
v += double(in_nchw(n, c, hi, wi)) * double(wei_kcyx(k, c, y, x));
}
}
}
}
out(n, k, ho, wo) = v;
out_nkhw(n, k, ho, wo) = v;
};
auto f_par = make_ParallelTensorFunctor(f,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3]);
out_nkhw.mDesc.GetLengths()[0],
out_nkhw.mDesc.GetLengths()[1],
out_nkhw.mDesc.GetLengths()[2],
out_nkhw.mDesc.GetLengths()[3]);
f_par(std::thread::hardware_concurrency());
}
template <class T, class LowerPads, class UpperPads>
void host_winograd_3x3_convolution(
const Tensor<T>& in_nchw, const Tensor<T>& wei_kcyx, Tensor<T>& out, LowerPads, UpperPads)
template <class TIn, class TWei, class TOut, class LowerPads, class UpperPads>
void host_winograd_3x3_convolution(const Tensor<TIn>& in_nchw,
const Tensor<TWei>& wei_kcyx,
Tensor<TOut>& out_nkhw,
LowerPads,
UpperPads)
{
constexpr std::size_t HoPerTile = 2;
constexpr std::size_t WoPerTile = 2;
@@ -162,8 +150,8 @@ void host_winograd_3x3_convolution(
std::size_t Y = wei_kcyx.mDesc.GetLengths()[2];
std::size_t X = wei_kcyx.mDesc.GetLengths()[3];
std::size_t HO = out.mDesc.GetLengths()[2];
std::size_t WO = out.mDesc.GetLengths()[3];
std::size_t HO = out_nkhw.mDesc.GetLengths()[2];
std::size_t WO = out_nkhw.mDesc.GetLengths()[3];
unsigned h_pad_low = LowerPads{}.Get(Number<0>{});
unsigned w_pad_low = LowerPads{}.Get(Number<1>{});
@@ -177,11 +165,11 @@ void host_winograd_3x3_convolution(
std::size_t HTile = (HO + HoPerTile - 1) / HoPerTile;
std::size_t WTile = (WO + WoPerTile - 1) / WoPerTile;
Tensor<T> in_hold({N, C, HTile, WTile, HiPerTile, WiPerTile});
Tensor<T> in_transform({N, C, HTile, WTile, HiPerTile, WiPerTile});
Tensor<T> wei_transform({K, C, HiPerTile, WiPerTile});
Tensor<T> out_transform({N, K, HTile, WTile, HiPerTile, HiPerTile});
Tensor<T> out_hold({N, K, HTile, WTile, HoPerTile, WoPerTile});
Tensor<double> in_hold({N, C, HTile, WTile, HiPerTile, WiPerTile});
Tensor<double> in_transform({N, C, HTile, WTile, HiPerTile, WiPerTile});
Tensor<double> wei_transform({K, C, HiPerTile, WiPerTile});
Tensor<double> out_transform({N, K, HTile, WTile, HiPerTile, HiPerTile});
Tensor<double> out_hold({N, K, HTile, WTile, HoPerTile, WoPerTile});
auto f_in_hold = [&](auto n, auto c, auto htile, auto wtile) {
for(int j = 0; j < HiPerTile; ++j)
@@ -198,7 +186,7 @@ void host_winograd_3x3_convolution(
}
else
{
in_hold(n, c, htile, wtile, j, i) = T(0);
in_hold(n, c, htile, wtile, j, i) = TIn(0);
}
}
}
@@ -259,49 +247,61 @@ void host_winograd_3x3_convolution(
};
auto f_wei_transform = [&](auto k, auto c) {
wei_transform(k, c, 0, 0) = wei_kcyx(k, c, 0, 0);
wei_transform(k, c, 0, 1) =
0.5 * wei_kcyx(k, c, 0, 0) + 0.5 * wei_kcyx(k, c, 0, 1) + 0.5 * wei_kcyx(k, c, 0, 2);
wei_transform(k, c, 0, 2) =
0.5 * wei_kcyx(k, c, 0, 0) - 0.5 * wei_kcyx(k, c, 0, 1) + 0.5 * wei_kcyx(k, c, 0, 2);
wei_transform(k, c, 0, 3) = wei_kcyx(k, c, 0, 2);
wei_transform(k, c, 0, 0) = double(wei_kcyx(k, c, 0, 0));
wei_transform(k, c, 0, 1) = 0.5 * double(wei_kcyx(k, c, 0, 0)) +
0.5 * double(wei_kcyx(k, c, 0, 1)) +
0.5 * double(wei_kcyx(k, c, 0, 2));
wei_transform(k, c, 0, 2) = 0.5 * double(wei_kcyx(k, c, 0, 0)) -
0.5 * double(wei_kcyx(k, c, 0, 1)) +
0.5 * double(wei_kcyx(k, c, 0, 2));
wei_transform(k, c, 0, 3) = double(wei_kcyx(k, c, 0, 2));
wei_transform(k, c, 1, 0) =
0.5 * wei_kcyx(k, c, 0, 0) + 0.5 * wei_kcyx(k, c, 1, 0) + 0.5 * wei_kcyx(k, c, 2, 0);
wei_transform(k, c, 1, 1) = 0.25 * wei_kcyx(k, c, 0, 0) + 0.25 * wei_kcyx(k, c, 0, 1) +
0.25 * wei_kcyx(k, c, 0, 2) + 0.25 * wei_kcyx(k, c, 1, 0) +
0.25 * wei_kcyx(k, c, 1, 1) + 0.25 * wei_kcyx(k, c, 1, 2) +
0.25 * wei_kcyx(k, c, 2, 0) + 0.25 * wei_kcyx(k, c, 2, 1) +
0.25 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 1, 2) = 0.25 * wei_kcyx(k, c, 0, 0) - 0.25 * wei_kcyx(k, c, 0, 1) +
0.25 * wei_kcyx(k, c, 0, 2) + 0.25 * wei_kcyx(k, c, 1, 0) -
0.25 * wei_kcyx(k, c, 1, 1) + 0.25 * wei_kcyx(k, c, 1, 2) +
0.25 * wei_kcyx(k, c, 2, 0) - 0.25 * wei_kcyx(k, c, 2, 1) +
0.25 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 1, 3) =
0.5 * wei_kcyx(k, c, 0, 2) + 0.5 * wei_kcyx(k, c, 1, 2) + 0.5 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 1, 0) = 0.5 * double(wei_kcyx(k, c, 0, 0)) +
0.5 * double(wei_kcyx(k, c, 1, 0)) +
0.5 * double(wei_kcyx(k, c, 2, 0));
wei_transform(k, c, 1, 1) =
0.25 * double(wei_kcyx(k, c, 0, 0)) + 0.25 * double(wei_kcyx(k, c, 0, 1)) +
0.25 * double(wei_kcyx(k, c, 0, 2)) + 0.25 * double(wei_kcyx(k, c, 1, 0)) +
0.25 * double(wei_kcyx(k, c, 1, 1)) + 0.25 * double(wei_kcyx(k, c, 1, 2)) +
0.25 * double(wei_kcyx(k, c, 2, 0)) + 0.25 * double(wei_kcyx(k, c, 2, 1)) +
0.25 * double(wei_kcyx(k, c, 2, 2));
wei_transform(k, c, 1, 2) =
0.25 * double(wei_kcyx(k, c, 0, 0)) - 0.25 * double(wei_kcyx(k, c, 0, 1)) +
0.25 * double(wei_kcyx(k, c, 0, 2)) + 0.25 * double(wei_kcyx(k, c, 1, 0)) -
0.25 * double(wei_kcyx(k, c, 1, 1)) + 0.25 * double(wei_kcyx(k, c, 1, 2)) +
0.25 * double(wei_kcyx(k, c, 2, 0)) - 0.25 * double(wei_kcyx(k, c, 2, 1)) +
0.25 * double(wei_kcyx(k, c, 2, 2));
wei_transform(k, c, 1, 3) = 0.5 * double(wei_kcyx(k, c, 0, 2)) +
0.5 * double(wei_kcyx(k, c, 1, 2)) +
0.5 * double(wei_kcyx(k, c, 2, 2));
wei_transform(k, c, 2, 0) =
0.5 * wei_kcyx(k, c, 0, 0) - 0.5 * wei_kcyx(k, c, 1, 0) + 0.5 * wei_kcyx(k, c, 2, 0);
wei_transform(k, c, 2, 1) = 0.25 * wei_kcyx(k, c, 0, 0) + 0.25 * wei_kcyx(k, c, 0, 1) +
0.25 * wei_kcyx(k, c, 0, 2) - 0.25 * wei_kcyx(k, c, 1, 0) -
0.25 * wei_kcyx(k, c, 1, 1) - 0.25 * wei_kcyx(k, c, 1, 2) +
0.25 * wei_kcyx(k, c, 2, 0) + 0.25 * wei_kcyx(k, c, 2, 1) +
0.25 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 2, 2) = 0.25 * wei_kcyx(k, c, 0, 0) - 0.25 * wei_kcyx(k, c, 0, 1) +
0.25 * wei_kcyx(k, c, 0, 2) - 0.25 * wei_kcyx(k, c, 1, 0) +
0.25 * wei_kcyx(k, c, 1, 1) - 0.25 * wei_kcyx(k, c, 1, 2) +
0.25 * wei_kcyx(k, c, 2, 0) - 0.25 * wei_kcyx(k, c, 2, 1) +
0.25 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 2, 3) =
0.5 * wei_kcyx(k, c, 0, 2) - 0.5 * wei_kcyx(k, c, 1, 2) + 0.5 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 2, 0) = 0.5 * double(wei_kcyx(k, c, 0, 0)) -
0.5 * double(wei_kcyx(k, c, 1, 0)) +
0.5 * double(wei_kcyx(k, c, 2, 0));
wei_transform(k, c, 2, 1) =
0.25 * double(wei_kcyx(k, c, 0, 0)) + 0.25 * double(wei_kcyx(k, c, 0, 1)) +
0.25 * double(wei_kcyx(k, c, 0, 2)) - 0.25 * double(wei_kcyx(k, c, 1, 0)) -
0.25 * double(wei_kcyx(k, c, 1, 1)) - 0.25 * double(wei_kcyx(k, c, 1, 2)) +
0.25 * double(wei_kcyx(k, c, 2, 0)) + 0.25 * double(wei_kcyx(k, c, 2, 1)) +
0.25 * double(wei_kcyx(k, c, 2, 2));
wei_transform(k, c, 2, 2) =
0.25 * double(wei_kcyx(k, c, 0, 0)) - 0.25 * double(wei_kcyx(k, c, 0, 1)) +
0.25 * double(wei_kcyx(k, c, 0, 2)) - 0.25 * double(wei_kcyx(k, c, 1, 0)) +
0.25 * double(wei_kcyx(k, c, 1, 1)) - 0.25 * double(wei_kcyx(k, c, 1, 2)) +
0.25 * double(wei_kcyx(k, c, 2, 0)) - 0.25 * double(wei_kcyx(k, c, 2, 1)) +
0.25 * double(wei_kcyx(k, c, 2, 2));
wei_transform(k, c, 2, 3) = 0.5 * double(wei_kcyx(k, c, 0, 2)) -
0.5 * double(wei_kcyx(k, c, 1, 2)) +
0.5 * double(wei_kcyx(k, c, 2, 2));
wei_transform(k, c, 3, 0) = wei_kcyx(k, c, 2, 0);
wei_transform(k, c, 3, 1) =
0.5 * wei_kcyx(k, c, 2, 0) + 0.5 * wei_kcyx(k, c, 2, 1) + 0.5 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 3, 2) =
0.5 * wei_kcyx(k, c, 2, 0) - 0.5 * wei_kcyx(k, c, 2, 1) + 0.5 * wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 3, 3) = wei_kcyx(k, c, 2, 2);
wei_transform(k, c, 3, 0) = double(wei_kcyx(k, c, 2, 0));
wei_transform(k, c, 3, 1) = 0.5 * double(wei_kcyx(k, c, 2, 0)) +
0.5 * double(wei_kcyx(k, c, 2, 1)) +
0.5 * double(wei_kcyx(k, c, 2, 2));
wei_transform(k, c, 3, 2) = 0.5 * double(wei_kcyx(k, c, 2, 0)) -
0.5 * double(wei_kcyx(k, c, 2, 1)) +
0.5 * double(wei_kcyx(k, c, 2, 2));
wei_transform(k, c, 3, 3) = double(wei_kcyx(k, c, 2, 2));
};
auto f_out_transform = [&](auto n, auto k, auto htile, auto wtile) {
@@ -354,7 +354,7 @@ void host_winograd_3x3_convolution(
for(int i = 0; i < WoPerTile; ++i)
{
std::size_t wo = WoPerTile * wtile + i;
out(n, k, ho, wo) = out_hold(n, k, htile, wtile, j, i);
out_nkhw(n, k, ho, wo) = out_hold(n, k, htile, wtile, j, i);
}
}
};
@@ -372,20 +372,25 @@ void host_winograd_3x3_convolution(
template <class T>
void check_error(const Tensor<T>& ref, const Tensor<T>& result)
{
// printf("\n");
float error = 0;
float max_diff = -1;
float ref_value = 0, result_value = 0;
for(int i = 0; i < ref.mData.size(); ++i)
{
error += std::abs(ref.mData[i] - result.mData[i]);
float diff = std::abs(ref.mData[i] - result.mData[i]);
error += std::abs(double(ref.mData[i]) - double(result.mData[i]));
float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
if(max_diff < diff)
{
max_diff = diff;
ref_value = ref.mData[i];
result_value = result.mData[i];
}
// printf("{%f, %f}", double(ref.mData[i]), double(result.mData[i]));
}
// printf("\n");
std::cout << "error: " << error << std::endl;
std::cout << "max_diff: " << max_diff << ", " << ref_value << ", " << result_value << std::endl;
@@ -404,7 +409,7 @@ int main(int argc, char* argv[])
constexpr unsigned HPad = 0;
constexpr unsigned WPad = 0;
#elif 1
#elif 0
// 3x3, 34x34
constexpr unsigned N = 64;
constexpr unsigned C = 256;
@@ -563,6 +568,30 @@ int main(int argc, char* argv[])
constexpr unsigned HPad = 2;
constexpr unsigned WPad = 2;
#elif 0
// 1x1 filter, 32x32 image
constexpr unsigned N = 64;
constexpr unsigned C = 256;
constexpr unsigned HI = 32;
constexpr unsigned WI = 32;
constexpr unsigned K = 512;
constexpr unsigned Y = 1;
constexpr unsigned X = 1;
constexpr unsigned HPad = 0;
constexpr unsigned WPad = 0;
#elif 1
// 1x1 filter, 14x14 image
constexpr unsigned N = 128;
constexpr unsigned C = 2048;
constexpr unsigned HI = 14;
constexpr unsigned WI = 14;
constexpr unsigned K = 512;
constexpr unsigned Y = 1;
constexpr unsigned X = 1;
constexpr unsigned HPad = 0;
constexpr unsigned WPad = 0;
#endif
auto lower_pads = Sequence<HPad, WPad>{};
@@ -577,10 +606,12 @@ int main(int argc, char* argv[])
ostream_ConstantTensorDescriptor(wei_kcyx_desc, std::cout << "wei_kcyx_desc: ");
ostream_ConstantTensorDescriptor(out_nkhw_desc, std::cout << "out_nkhw_desc: ");
Tensor<float> in_nchw(make_TensorDescriptor(in_nchw_desc));
Tensor<float> wei_kcyx(make_TensorDescriptor(wei_kcyx_desc));
Tensor<float> out_nkhw_host(make_TensorDescriptor(out_nkhw_desc));
Tensor<float> out_nkhw_device(make_TensorDescriptor(out_nkhw_desc));
using in_data_t = float;
using out_data_t = float;
Tensor<in_data_t> in_nchw(make_TensorDescriptor(in_nchw_desc));
Tensor<in_data_t> wei_kcyx(make_TensorDescriptor(wei_kcyx_desc));
Tensor<out_data_t> out_nkhw_host(make_TensorDescriptor(out_nkhw_desc));
Tensor<out_data_t> out_nkhw_device(make_TensorDescriptor(out_nkhw_desc));
std::size_t num_thread = std::thread::hardware_concurrency();
@@ -601,9 +632,13 @@ int main(int argc, char* argv[])
#elif 1
in_nchw.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
#elif 1
in_nchw.GenerateTensorValue(GeneratorTensor_2{-2, 2}, num_thread);
wei_kcyx.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
#elif 0
in_nchw.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
auto gen_wei = [](auto... is) {
return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
};
wei_kcyx.GenerateTensorValue(gen_wei, num_thread);
#endif
}
@@ -611,7 +646,9 @@ int main(int argc, char* argv[])
#if 0
device_direct_convolution_1
#elif 0
device_direct_convolution_2
device_direct_convolution_2_nchw_kcyx_nkhw
#elif 0
device_direct_convolution_2_vectorized_nchw_kcyx_nkhw
#elif 1
device_implicit_gemm_convolution_1_chwn_cyxk_khwn
#elif 0
@@ -633,7 +670,6 @@ int main(int argc, char* argv[])
if(do_verification)
{
#if 1
if(Y == 3 && X == 3)
{
host_winograd_3x3_convolution(in_nchw, wei_kcyx, out_nkhw_host, lower_pads, upper_pads);
@@ -643,7 +679,6 @@ int main(int argc, char* argv[])
host_direct_convolution(in_nchw, wei_kcyx, out_nkhw_host, lower_pads, upper_pads);
}
check_error(out_nkhw_host, out_nkhw_device);
#endif
#if 0
LogRange(std::cout << "in_nchw : ", in_nchw.mData, ",") << std::endl;