Skip to content

GitLab

"git@developer.sourcefind.cn:orangecat/ollama.git" did not exist on "b8d5036e3328ebfd16dcc94c60beb46496ba8112"
Commit 16d78f5d authored by Jing Zhang's avatar Jing Zhang
Browse files

added host conv bwd and wrw

parent b62bf8c3
Hide whitespace changes
Inline Side-by-side
Showing with 203 additions and 46 deletions
host/host_tensor/include/conv_common.hpp
View file @ 16d78f5d
......@@ -12,6 +12,13 @@ enum ConvTensorLayout
NHWCc
};
enum ConvDirection
{
Forward,
BackwardData,
BackwardWeights
};
template <typename... InDesc,
typename... WeiDesc,
typename ConvStrides,
......
host/host_tensor/include/host_conv.hpp
View file @ 16d78f5d
......@@ -8,83 +8,233 @@ template <typename TIn,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void host_direct_convolution(const Tensor<TIn>& in,
const Tensor<TWei>& wei,
void host_direct_convolution(Tensor<TIn>& in,
Tensor<TWei>& wei,
Tensor<TOut>& out,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads&,
const ConvTensorLayout layout = ConvTensorLayout::NCHW)
const ConvTensorLayout layout = ConvTensorLayout::NCHW,
const ConvDirection dir = ConvDirection::Forward)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
double v = 0;
for(int c = 0; c < wei.mDesc.GetLengths()[1]; ++c)
{
for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
if(dir == ConvDirection::Forward)
{
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
double v = 0;
for(int c = 0; c < wei.mDesc.GetLengths()[1]; ++c)
{
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
{
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
{
v += static_cast<const double>(in(n, c, hi, wi)) *
static_cast<const double>(wei(k, c, y, x));
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
v += static_cast<const double>(in(n, c, hi, wi)) *
static_cast<const double>(wei(k, c, y, x));
}
}
}
}
}
out(n, k, ho, wo) = v;
};
out(n, k, ho, wo) = v;
};
auto f_nhwc = [&](auto n, auto ho, auto wo, auto k) {
double v = 0;
for(int c = 0; c < wei.mDesc.GetLengths()[3]; ++c)
{
for(int y = 0; y < wei.mDesc.GetLengths()[1]; ++y)
auto f_nhwc = [&](auto n, auto ho, auto wo, auto k) {
double v = 0;
for(int c = 0; c < wei.mDesc.GetLengths()[3]; ++c)
{
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[2]; ++x)
for(int y = 0; y < wei.mDesc.GetLengths()[1]; ++y)
{
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[1] && wi >= 0 &&
wi < in.mDesc.GetLengths()[2])
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[2]; ++x)
{
v += static_cast<const double>(in(n, hi, wi, c)) *
static_cast<const double>(wei(k, y, x, c));
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[1] && wi >= 0 &&
wi < in.mDesc.GetLengths()[2])
{
v += static_cast<const double>(in(n, hi, wi, c)) *
static_cast<const double>(wei(k, y, x, c));
}
}
}
}
}
out(n, ho, wo, k) = v;
};
out(n, ho, wo, k) = v;
};
if(layout == ConvTensorLayout::NCHW)
if(layout == ConvTensorLayout::NCHW)
{
make_ParallelTensorFunctor(f_nchw,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
else if(layout == ConvTensorLayout::NHWC)
{
make_ParallelTensorFunctor(f_nhwc,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
else
{
throw std::runtime_error("wrong! not supported layout");
}
}
else if(dir == ConvDirection::BackwardData)
{
make_ParallelTensorFunctor(f_nchw,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
auto f_nchw = [&](auto n, auto c, auto hi, auto wi) {
double v = 0;
for(int k = 0; k < wei.mDesc.GetLengths()[0]; ++k)
{
for(int y = 0; y < wei.mDesc.GetLengths()[2]; ++y)
{
int ho = (hi - y * conv_dilations[I0] + in_left_pads[I0]) / conv_strides[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[3]; ++x)
{
int wo =
(wi - x * conv_dilations[I1] + in_left_pads[I1]) / conv_strides[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
v += static_cast<const double>(out(n, k, ho, wo)) *
static_cast<const double>(wei(k, c, y, x));
}
}
}
}
in(n, c, hi, wi) = v;
};
auto f_nhwc = [&](auto n, auto hi, auto wi, auto c) {
double v = 0;
for(int k = 0; k < wei.mDesc.GetLengths()[0]; ++k)
{
for(int y = 0; y < wei.mDesc.GetLengths()[1]; ++y)
{
int ho = (hi - y * conv_dilations[I0] + in_left_pads[I0]) / conv_strides[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[2]; ++x)
{
int wo =
(wi - x * conv_dilations[I1] + in_left_pads[I1]) / conv_strides[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[1] && wi >= 0 &&
wi < in.mDesc.GetLengths()[2])
{
v += static_cast<const double>(out(n, ho, wo, k)) *
static_cast<const double>(wei(k, y, x, c));
}
}
}
}
in(n, hi, wi, c) = v;
};
if(layout == ConvTensorLayout::NCHW)
{
make_ParallelTensorFunctor(f_nchw,
in.mDesc.GetLengths()[0],
in.mDesc.GetLengths()[1],
in.mDesc.GetLengths()[2],
in.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
else if(layout == ConvTensorLayout::NHWC)
{
make_ParallelTensorFunctor(f_nhwc,
in.mDesc.GetLengths()[0],
in.mDesc.GetLengths()[1],
in.mDesc.GetLengths()[2],
in.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
else
{
throw std::runtime_error("wrong! not supported layout");
}
}
else if(layout == ConvTensorLayout::NHWC)
else if(dir == ConvDirection::BackwardWeights)
{
make_ParallelTensorFunctor(f_nhwc,
out.mDesc.GetLengths()[0],
out.mDesc.GetLengths()[1],
out.mDesc.GetLengths()[2],
out.mDesc.GetLengths()[3])(std::thread::hardware_concurrency());
auto f_kcyx = [&](auto k, auto c, auto y, auto x) {
double v = 0;
for(int n = 0; n < out.mDesc.GetLengths()[0]; ++n)
{
for(int ho = 0; ho < out.mDesc.GetLengths()[2]; ++ho)
{
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int wo = 0; wo < wei.mDesc.GetLengths()[3]; ++wo)
{
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
v += static_cast<const double>(in(n, c, hi, wi)) *
static_cast<const double>(out(n, k, ho, wo));
}
}
}
}
wei(k, c, y, x) = v;
};
auto f_kyxc = [&](auto k, auto y, auto x, auto c) {
double v = 0;
for(int n = 0; n < out.mDesc.GetLengths()[0]; ++n)
{
for(int ho = 0; ho < out.mDesc.GetLengths()[1]; ++ho)
{
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int wo = 0; wo < wei.mDesc.GetLengths()[2]; ++wo)
{
int wi = wo * conv_strides[I1] + x * conv_dilations[I1] - in_left_pads[I1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[1] && wi >= 0 &&
wi < in.mDesc.GetLengths()[2])
{
v += static_cast<const double>(in(n, hi, wi, c)) *
static_cast<const double>(out(n, ho, wo, k));
}
}
}
}
wei(k, y, x, c) = v;
};
if(layout == ConvTensorLayout::NCHW)
{
make_ParallelTensorFunctor(f_kcyx,
wei.mDesc.GetLengths()[0],
wei.mDesc.GetLengths()[1],
wei.mDesc.GetLengths()[2],
wei.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
else if(layout == ConvTensorLayout::NHWC)
{
make_ParallelTensorFunctor(f_kyxc,
wei.mDesc.GetLengths()[0],
wei.mDesc.GetLengths()[1],
wei.mDesc.GetLengths()[2],
wei.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
}
else
{
throw std::runtime_error("wrong! not supported layout");
}
}
else
{
throw std::runtime_error("wrong! not supported layout");
throw std::runtime_error("wrong! not supported direction");
}
}
......
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