Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "38086572c83a3882961ef5b14e66e7a01de54ddf"
Unverified Commit e823d518 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

ckProfiler and device-level XDL GEMM operator (#48) 

* add DeviceGemmXdl

* update script

* fix naming issue

* fix comment

* output HostTensorDescriptor

* rename

* padded GEMM for fwd v4r4r4 nhwc

* refactor

* refactor

* refactor

* adding ckProfiler

* adding ckProfiler

* refactor

* fix tuning parameter bug

* add more gemm instances

* add more fp16 GEMM instances

* fix profiler driver

* fix bug in tuning parameter

* add fp32 gemm instances

* small fix

* refactor

* rename

* refactor gemm profiler; adding DeviceConv and conv profiler

* refactor

* fix

* add conv profiler

* refactor

* adding more GEMM and Conv instance

* Create README.md

Add build instruction for ckProfiler

* Create README.md

Add Readme for gemm_xdl example

* Update README.md

Remove build instruction from top most folder

* Update README.md

* clean up
parent 6014185a
Expand all Show whitespace changes
Inline Side-by-side
Showing with 1167 additions and 672 deletions
host/driver_offline/src/conv_wrw_driver_offline.cpp
View file @ e823d518
......@@ -11,7 +11,6 @@
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "conv_common.hpp"
#include "host_conv_bwd_weight.hpp"
#include "device_tensor.hpp"
#include "device_convolution_backward_weight_implicit_gemm_v4r4r2_xdlops_nchw_kcyx_nkhw.hpp"
#include "device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_nhwc_kyxc_nhwk.hpp"
......@@ -19,6 +18,15 @@
#include "device_convolution_backward_weight_implicit_gemm_v4r4r4_xdlops_atomic_nhwc_kyxc_nhwk.hpp"
#include "device_convolution_backward_weight_implicit_gemm_v4r4r5_xdlops_atomic_nhwc_kyxc_nhwk.hpp"
enum ConvTensorLayout
{
NCHW,
NHWC,
CHWN,
NCHWc,
NHWCc
};
#define USE_DYNAMIC_MODE 1
#define USE_CONV_WRW_V4R4R2_XDL_NCHW 0
#define USE_CONV_WRW_V4R4R4_XDL_NHWC 0
......@@ -35,6 +43,92 @@ enum ConvBackwardWeightAlgo
V4R4R5XDLATOMICNHWC, // 4
};
template <typename TOut,
typename TIn,
typename TWei,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void host_convolution_backward_weight(const Tensor<TOut>& out,
const Tensor<TIn>& in,
Tensor<TWei>& wei,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads&,
const ConvTensorLayout layout = ConvTensorLayout::NCHW)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
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 < out.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 < out.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");
}
}
int main(int argc, char* argv[])
{
using namespace ck;
......@@ -414,7 +508,7 @@ int main(int argc, char* argv[])
if(do_verification)
{
host_direct_convolution_backward_weights(out,
host_convolution_backward_weight(out,
in,
wei_host,
make_tuple(conv_stride_h, conv_stride_w),
......
host/host_tensor/include/conv_common.hpp
View file @ e823d518
......@@ -3,15 +3,6 @@
#include "tensor_descriptor.hpp"
enum ConvTensorLayout
{
NCHW,
NHWC,
CHWN,
NCHWc,
NHWCc
};
template <typename... InDesc,
typename... WeiDesc,
typename ConvStrides,
......
host/host_tensor/include/host_conv.hpp
View file @ e823d518
#pragma once
#include "host_tensor.hpp"
#include "conv_common.hpp"
template <typename TIn,
typename TWei,
......@@ -8,19 +9,16 @@ template <typename TIn,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void host_direct_convolution(const Tensor<TIn>& in,
void host_conv_nchw_kcyx_nkhw(const Tensor<TIn>& in,
const 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 InRightPads&)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I0 = ck::Number<0>{};
constexpr auto I1 = ck::Number<1>{};
auto f_nchw = [&](auto n, auto k, auto ho, auto wo) {
double v = 0;
......@@ -44,281 +42,9 @@ void host_direct_convolution(const Tensor<TIn>& in,
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)
{
int hi = ho * conv_strides[I0] + y * conv_dilations[I0] - in_left_pads[I0];
for(int x = 0; x < wei.mDesc.GetLengths()[2]; ++x)
{
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;
};
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");
}
}
template <typename TIn, typename TWei, typename TOut, typename InLeftPads, typename InRightPads>
void host_winograd_3x3_convolution(const Tensor<TIn>& in_nchw,
const Tensor<TWei>& wei_kcyx,
Tensor<TOut>& out_nkhw,
InLeftPads,
InRightPads)
{
using namespace ck;
constexpr std::size_t HoPerTile = 2;
constexpr std::size_t WoPerTile = 2;
std::size_t N = in_nchw.mDesc.GetLengths()[0];
std::size_t C = in_nchw.mDesc.GetLengths()[1];
std::size_t K = wei_kcyx.mDesc.GetLengths()[0];
std::size_t Y = wei_kcyx.mDesc.GetLengths()[2];
std::size_t X = wei_kcyx.mDesc.GetLengths()[3];
std::size_t Ho = out_nkhw.mDesc.GetLengths()[2];
std::size_t Wo = out_nkhw.mDesc.GetLengths()[3];
index_t h_pad_low = InLeftPads{}.Get(Number<0>{});
index_t w_pad_low = InLeftPads{}.Get(Number<1>{});
std::size_t HiPerTile = HoPerTile + Y - 1;
std::size_t WiPerTile = WoPerTile + X - 1;
std::size_t HTile = (Ho + HoPerTile - 1) / HoPerTile;
std::size_t WTile = (Wo + WoPerTile - 1) / 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)
{
int hi = HoPerTile * htile + j - h_pad_low;
for(int i = 0; i < WiPerTile; ++i)
{
int wi = WoPerTile * wtile + i - w_pad_low;
if(hi >= 0 && hi < in_nchw.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in_nchw.mDesc.GetLengths()[3])
{
in_hold(n, c, htile, wtile, j, i) = in_nchw(n, c, hi, wi);
}
else
{
in_hold(n, c, htile, wtile, j, i) = TIn(0);
}
}
}
};
auto f_in_transform = [&](auto n, auto c, auto htile, auto wtile) {
in_transform(n, c, htile, wtile, 0, 0) =
in_hold(n, c, htile, wtile, 0, 0) - in_hold(n, c, htile, wtile, 0, 2) -
in_hold(n, c, htile, wtile, 2, 0) + in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 0, 1) =
in_hold(n, c, htile, wtile, 0, 1) + in_hold(n, c, htile, wtile, 0, 2) -
in_hold(n, c, htile, wtile, 2, 1) - in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 0, 2) =
-in_hold(n, c, htile, wtile, 0, 1) + in_hold(n, c, htile, wtile, 0, 2) +
in_hold(n, c, htile, wtile, 2, 1) - in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 0, 3) =
in_hold(n, c, htile, wtile, 0, 1) - in_hold(n, c, htile, wtile, 0, 3) -
in_hold(n, c, htile, wtile, 2, 1) + in_hold(n, c, htile, wtile, 2, 3);
in_transform(n, c, htile, wtile, 1, 0) =
in_hold(n, c, htile, wtile, 1, 0) - in_hold(n, c, htile, wtile, 1, 2) +
in_hold(n, c, htile, wtile, 2, 0) - in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 1, 1) =
in_hold(n, c, htile, wtile, 1, 1) + in_hold(n, c, htile, wtile, 1, 2) +
in_hold(n, c, htile, wtile, 2, 1) + in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 1, 2) =
-in_hold(n, c, htile, wtile, 1, 1) + in_hold(n, c, htile, wtile, 1, 2) -
in_hold(n, c, htile, wtile, 2, 1) + in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 1, 3) =
in_hold(n, c, htile, wtile, 1, 1) - in_hold(n, c, htile, wtile, 1, 3) +
in_hold(n, c, htile, wtile, 2, 1) - in_hold(n, c, htile, wtile, 2, 3);
in_transform(n, c, htile, wtile, 2, 0) =
-in_hold(n, c, htile, wtile, 1, 0) + in_hold(n, c, htile, wtile, 1, 2) +
in_hold(n, c, htile, wtile, 2, 0) - in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 2, 1) =
-in_hold(n, c, htile, wtile, 1, 1) - in_hold(n, c, htile, wtile, 1, 2) +
in_hold(n, c, htile, wtile, 2, 1) + in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 2, 2) =
in_hold(n, c, htile, wtile, 1, 1) - in_hold(n, c, htile, wtile, 1, 2) -
in_hold(n, c, htile, wtile, 2, 1) + in_hold(n, c, htile, wtile, 2, 2);
in_transform(n, c, htile, wtile, 2, 3) =
-in_hold(n, c, htile, wtile, 1, 1) + in_hold(n, c, htile, wtile, 1, 3) +
in_hold(n, c, htile, wtile, 2, 1) - in_hold(n, c, htile, wtile, 2, 3);
in_transform(n, c, htile, wtile, 3, 0) =
in_hold(n, c, htile, wtile, 1, 0) - in_hold(n, c, htile, wtile, 1, 2) -
in_hold(n, c, htile, wtile, 3, 0) + in_hold(n, c, htile, wtile, 3, 2);
in_transform(n, c, htile, wtile, 3, 1) =
in_hold(n, c, htile, wtile, 1, 1) + in_hold(n, c, htile, wtile, 1, 2) -
in_hold(n, c, htile, wtile, 3, 1) - in_hold(n, c, htile, wtile, 3, 2);
in_transform(n, c, htile, wtile, 3, 2) =
-in_hold(n, c, htile, wtile, 1, 1) + in_hold(n, c, htile, wtile, 1, 2) +
in_hold(n, c, htile, wtile, 3, 1) - in_hold(n, c, htile, wtile, 3, 2);
in_transform(n, c, htile, wtile, 3, 3) =
in_hold(n, c, htile, wtile, 1, 1) - in_hold(n, c, htile, wtile, 1, 3) -
in_hold(n, c, htile, wtile, 3, 1) + in_hold(n, c, htile, wtile, 3, 3);
};
auto f_wei_transform = [&](auto k, auto c) {
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 * 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 * 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) = 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) {
for(int j = 0; j < HiPerTile; ++j)
{
for(int i = 0; i < WiPerTile; ++i)
{
double v = 0;
for(int c = 0; c < C; ++c)
{
v += in_transform(n, c, htile, wtile, j, i) * wei_transform(k, c, j, i);
}
out_transform(n, k, htile, wtile, j, i) = v;
}
}
};
auto f_out_hold = [&](auto n, auto k, auto htile, auto wtile) {
out_hold(n, k, htile, wtile, 0, 0) =
out_transform(n, k, htile, wtile, 0, 0) + out_transform(n, k, htile, wtile, 0, 1) +
out_transform(n, k, htile, wtile, 0, 2) + out_transform(n, k, htile, wtile, 1, 0) +
out_transform(n, k, htile, wtile, 1, 1) + out_transform(n, k, htile, wtile, 1, 2) +
out_transform(n, k, htile, wtile, 2, 0) + out_transform(n, k, htile, wtile, 2, 1) +
out_transform(n, k, htile, wtile, 2, 2);
out_hold(n, k, htile, wtile, 0, 1) =
out_transform(n, k, htile, wtile, 0, 1) - out_transform(n, k, htile, wtile, 0, 2) -
out_transform(n, k, htile, wtile, 0, 3) + out_transform(n, k, htile, wtile, 1, 1) -
out_transform(n, k, htile, wtile, 1, 2) - out_transform(n, k, htile, wtile, 1, 3) +
out_transform(n, k, htile, wtile, 2, 1) - out_transform(n, k, htile, wtile, 2, 2) -
out_transform(n, k, htile, wtile, 2, 3);
out_hold(n, k, htile, wtile, 1, 0) =
out_transform(n, k, htile, wtile, 1, 0) + out_transform(n, k, htile, wtile, 1, 1) +
out_transform(n, k, htile, wtile, 1, 2) - out_transform(n, k, htile, wtile, 2, 0) -
out_transform(n, k, htile, wtile, 2, 1) - out_transform(n, k, htile, wtile, 2, 2) -
out_transform(n, k, htile, wtile, 3, 0) - out_transform(n, k, htile, wtile, 3, 1) -
out_transform(n, k, htile, wtile, 3, 2);
out_hold(n, k, htile, wtile, 1, 1) =
out_transform(n, k, htile, wtile, 1, 1) - out_transform(n, k, htile, wtile, 1, 2) -
out_transform(n, k, htile, wtile, 1, 3) - out_transform(n, k, htile, wtile, 2, 1) +
out_transform(n, k, htile, wtile, 2, 2) + out_transform(n, k, htile, wtile, 2, 3) -
out_transform(n, k, htile, wtile, 3, 1) + out_transform(n, k, htile, wtile, 3, 2) +
out_transform(n, k, htile, wtile, 3, 3);
};
auto f_out = [&](auto n, auto k, auto htile, auto wtile) {
for(int j = 0; j < HoPerTile; ++j)
{
std::size_t ho = HoPerTile * htile + j;
for(int i = 0; i < WoPerTile; ++i)
{
std::size_t wo = WoPerTile * wtile + i;
out_nkhw(n, k, ho, wo) = out_hold(n, k, htile, wtile, j, i);
}
}
};
std::size_t num_thread = std::thread::hardware_concurrency();
make_ParallelTensorFunctor(f_in_hold, N, C, HTile, WTile)(num_thread);
make_ParallelTensorFunctor(f_in_transform, N, C, HTile, WTile)(num_thread);
make_ParallelTensorFunctor(f_wei_transform, K, C)(num_thread);
make_ParallelTensorFunctor(f_out_transform, N, K, HTile, WTile)(num_thread);
make_ParallelTensorFunctor(f_out_hold, N, K, HTile, WTile)(num_thread);
make_ParallelTensorFunctor(f_out, N, K, HTile, WTile)(num_thread);
}
host/host_tensor/include/host_conv_bwd_data.hpp deleted 100644 → 0
View file @ 6014185a
#pragma once
#include "host_tensor.hpp"
template <typename TIn,
typename TWei,
typename TOut,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void host_direct_convolution_backward_data(Tensor<TIn>& in,
const Tensor<TWei>& wei,
const Tensor<TOut>& out,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads& /* in_right_pads */,
const ConvTensorLayout layout = ConvTensorLayout::NCHW)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
auto f_nchw = [&](auto n, auto c, auto hi, auto wi) {
std::size_t K = wei.mDesc.GetLengths()[I0];
std::size_t Y = wei.mDesc.GetLengths()[I2];
std::size_t X = wei.mDesc.GetLengths()[I3];
std::size_t Ho = out.mDesc.GetLengths()[I2];
std::size_t Wo = out.mDesc.GetLengths()[I3];
double v = 0;
for(int y = 0; y < Y; ++y)
{
int h_tmp = hi + in_left_pads[I0] - y * conv_dilations[I0];
if(h_tmp % conv_strides[I0] == 0)
{
int ho = h_tmp / conv_strides[I0];
if(ho >= 0 && ho < Ho)
{
for(int x = 0; x < X; ++x)
{
int w_tmp = wi + in_left_pads[I1] - x * conv_dilations[I1];
if(w_tmp % conv_strides[I1] == 0)
{
int wo = w_tmp / conv_strides[I1];
if(wo >= 0 && wo < Wo)
{
for(int k = 0; k < K; ++k)
{
v += out(n, k, ho, wo) * wei(k, c, y, x);
}
}
}
}
}
}
}
in(n, c, hi, wi) = v;
};
auto f_nhwc = [&](auto n, auto hi, auto wi, auto c) {
std::size_t K = wei.mDesc.GetLengths()[I0];
std::size_t Y = wei.mDesc.GetLengths()[I1];
std::size_t X = wei.mDesc.GetLengths()[I2];
std::size_t Ho = out.mDesc.GetLengths()[I1];
std::size_t Wo = out.mDesc.GetLengths()[I2];
double v = 0;
for(int y = 0; y < Y; ++y)
{
int h_tmp = hi + in_left_pads[I0] - y * conv_dilations[I0];
if(h_tmp % conv_strides[I0] == 0)
{
int ho = h_tmp / conv_strides[I0];
if(ho >= 0 && ho < Ho)
{
for(int x = 0; x < X; ++x)
{
int w_tmp = wi + in_left_pads[I1] - x * conv_dilations[I1];
if(w_tmp % conv_strides[I1] == 0)
{
int wo = w_tmp / conv_strides[I1];
if(wo >= 0 && wo < Wo)
{
for(int k = 0; k < K; ++k)
{
v += out(n, ho, wo, k) * 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");
}
}
host/host_tensor/include/host_conv_bwd_weight.hpp deleted 100644 → 0
View file @ 6014185a
#pragma once
#include "host_tensor.hpp"
template <typename TOut,
typename TIn,
typename TWei,
typename ConvStrides,
typename ConvDilations,
typename InLeftPads,
typename InRightPads>
void host_direct_convolution_backward_weights(
const Tensor<TOut>& out,
const Tensor<TIn>& in,
Tensor<TWei>& wei,
const ConvStrides& conv_strides,
const ConvDilations& conv_dilations,
const InLeftPads& in_left_pads,
const InRightPads&,
const ConvTensorLayout layout = ConvTensorLayout::NCHW)
{
using namespace ck;
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
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 < out.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 < out.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");
}
}
host/host_tensor/include/host_gemm.hpp
View file @ e823d518
......@@ -157,3 +157,26 @@ void host_gemm(const Tensor<AType>& a,
throw std::runtime_error("wrong! not supported layout");
}
}
template <typename AType, typename BType, typename CType>
void host_gemm_mk_kn_mn(const Tensor<AType>& a_m_k,
const Tensor<BType>& b_k_n,
Tensor<CType>& c_m_n)
{
auto f_mk_kn_mn = [&](auto m, auto n) {
const int K = a_m_k.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a_m_k(m, k)) * static_cast<const double>(b_k_n(k, n));
}
c_m_n(m, n) = v;
};
make_ParallelTensorFunctor(f_mk_kn_mn,
c_m_n.mDesc.GetLengths()[0],
c_m_n.mDesc.GetLengths()[1])(std::thread::hardware_concurrency());
}
host/host_tensor/include/host_tensor.hpp
View file @ e823d518
......@@ -120,6 +120,8 @@ struct HostTensorDescriptor
return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
}
friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc);
private:
std::vector<std::size_t> mLens;
std::vector<std::size_t> mStrides;
......@@ -224,7 +226,7 @@ struct Tensor
Tensor(const HostTensorDescriptor& desc) : mDesc(desc), mData(mDesc.GetElementSpace()) {}
template <typename G>
void GenerateTensorValue(G g, std::size_t num_thread = 1)
void GenerateTensorValue(G g, std::size_t num_thread = std::thread::hardware_concurrency())
{
switch(mDesc.GetNumOfDimension())
{
......
host/host_tensor/src/host_tensor.cpp
View file @ e823d518
......@@ -34,6 +34,21 @@ const std::vector<std::size_t>& HostTensorDescriptor::GetLengths() const { retur
const std::vector<std::size_t>& HostTensorDescriptor::GetStrides() const { return mStrides; }
std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc)
{
os << "dim " << desc.GetNumOfDimension() << ", ";
os << "lengths {";
LogRange(os, desc.GetLengths(), ", ");
os << "}, ";
os << "strides {";
LogRange(os, desc.GetStrides(), ", ");
os << "}";
return os;
}
void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream& os)
{
os << "dim " << desc.GetNumOfDimension() << ", ";
......
profiler/CMakeLists.txt 0 → 100644
View file @ e823d518
include_directories(BEFORE
include
${PROJECT_SOURCE_DIR}/host/host_tensor/include
${PROJECT_SOURCE_DIR}/device/include
${PROJECT_SOURCE_DIR}/device_operation/include
${PROJECT_SOURCE_DIR}/profiler/include
${PROJECT_SOURCE_DIR}/composable_kernel/include
${PROJECT_SOURCE_DIR}/composable_kernel/include/utility
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_description
${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_operation
${PROJECT_SOURCE_DIR}/composable_kernel/include/problem_transform
${PROJECT_SOURCE_DIR}/external/rocm/include
)
# device_gemm_instance
set(DEVICE_GEMM_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/device_gemm_xdl_instance_f32_f32_f32_mk_kn_mn.cpp;
${PROJECT_SOURCE_DIR}/device_operation/device_gemm_xdl_instance_f32_f32_f32_mk_nk_mn.cpp;
${PROJECT_SOURCE_DIR}/device_operation/device_gemm_xdl_instance_f32_f32_f32_km_kn_mn.cpp;
${PROJECT_SOURCE_DIR}/device_operation/device_gemm_xdl_instance_f32_f32_f32_km_nk_mn.cpp;
${PROJECT_SOURCE_DIR}/device_operation/device_gemm_xdl_instance_f16_f16_f16_mk_kn_mn.cpp;
${PROJECT_SOURCE_DIR}/device_operation/device_gemm_xdl_instance_f16_f16_f16_mk_nk_mn.cpp;
${PROJECT_SOURCE_DIR}/device_operation/device_gemm_xdl_instance_f16_f16_f16_km_kn_mn.cpp;
${PROJECT_SOURCE_DIR}/device_operation/device_gemm_xdl_instance_f16_f16_f16_km_nk_mn.cpp;
)
add_library(device_gemm_instance SHARED ${DEVICE_GEMM_INSTANCE_SOURCE})
target_include_directories(device_gemm_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_compile_features(device_gemm_instance PUBLIC)
set_target_properties(device_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
install(TARGETS device_gemm_instance LIBRARY DESTINATION lib)
# device_conv_instance
set(DEVICE_CONV_INSTANCE_SOURCE
${PROJECT_SOURCE_DIR}/device_operation/device_conv_xdl_instance_f32_f32_f32_nhwc_kyxc_nhwk.cpp;
${PROJECT_SOURCE_DIR}/device_operation/device_conv_xdl_instance_f16_f16_f16_nhwc_kyxc_nhwk.cpp;
)
add_library(device_conv_instance SHARED ${DEVICE_CONV_INSTANCE_SOURCE})
target_include_directories(device_conv_instance SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
target_compile_features(device_conv_instance PUBLIC)
set_target_properties(device_conv_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
install(TARGETS device_conv_instance LIBRARY DESTINATION lib)
# ck_profiler
set(PROFILER_SOURCE profiler.cpp gemm_profiler.cpp conv_profiler.cpp)
add_executable(ckProfiler ${PROFILER_SOURCE})
target_link_libraries(ckProfiler PRIVATE host_tensor)
target_link_libraries(ckProfiler PRIVATE device_gemm_instance device_conv_instance)
profiler/README.md 0 → 100644
View file @ e823d518
## Docker script
```bash
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```ckProfiler```
```bash
mkdir build && cd build
```
```bash
# Need to Specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
```bash
make -j ckProfiler
```
## Profile GEMM kernels
```bash
#arg1: tensor operation (gemm=GEMM)
#arg2: data type (0=fp32, 1=fp16)
#arg3: matrix layout (0=NN, 1=NT, 2=TN, 3=TT)
#arg4: verification (0=no, 1=yes)
#arg5: initialization (0=no init, 1=integer value, 2=decimal value)
#arg6: print matrix value (0=no, 1=yes)
#arg7: run kernel # of times (>1)
#arg8 to 13: M, N, K, StrideA, StrideB, StrideC
##################### op datatype layout verify init log repeat M___ N___ K___ StrideA StrideB StrideC
./profiler/ckProfiler gemm 1 1 1 1 0 5 3840 4096 4096 4096 4096 4096
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```bash
a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
....
Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
```
## Profile forward convolution kernels
```bash
#arg1: tensor operation (conv=Convolution)
#arg2: data type (0=fp32, 1=fp16)
#arg3: input tensor layout (0=NCHW, 1=NHWC)
#arg4: weight tensor layout (0=KCYX, 1=KYXC)
#arg5: output tensor layout (0=NKHW, 1=NHWK)
#arg6: verification (0=no, 1=yes)
#arg7: initialization (0=no init, 1=integer value, 2=decimal value)
#arg8: print matrix value (0=no, 1=yes)
#arg9: run kernel # of times (>1)
#arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx
##################### op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
./profiler/ckProfiler conv 1 1 1 1 1 1 0 5 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1
```
Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
```
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
....
Best Perf: 1.42509 ms, 102.988 TFlops, 234.086 GB/s
```
profiler/conv_profiler.cpp 0 → 100644
View file @ e823d518
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_conv.hpp"
enum ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
enum ConvInputLayout
{
NCHW, // 0
NHWC, // 1
};
enum ConvWeightLayout
{
KCYX, // 0
KYXC, // 1
};
enum ConvOutputLayout
{
NKHW, // 0
NHWK, // 1
};
int conv_profiler(int argc, char* argv[])
{
if(argc != 25)
{
printf("arg1: tensor operation (conv=Convolution)\n");
printf("arg2: data type (0=fp32, 1=fp16)\n");
printf("arg3: input tensor layout (0=NCHW, 1=NHWC)\n");
printf("arg4: weight tensor layout (0=KCYX, 1=KYXC)\n");
printf("arg5: output tensor layout (0=NKHW, 1=NHWK)\n");
printf("arg6: verification (0=no, 1=yes)\n");
printf("arg7: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg8: print matrix value (0=no, 1=yes)\n");
printf("arg9: run kernel # of times (>1)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
const int data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const int in_layout = static_cast<ConvInputLayout>(std::stoi(argv[3]));
const int wei_layout = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
const int out_layout = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const int nrepeat = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
const ck::index_t C = std::stoi(argv[12]);
const ck::index_t Y = std::stoi(argv[13]);
const ck::index_t X = std::stoi(argv[14]);
const ck::index_t Hi = std::stoi(argv[15]);
const ck::index_t Wi = std::stoi(argv[16]);
const ck::index_t conv_stride_h = std::stoi(argv[17]);
const ck::index_t conv_stride_w = std::stoi(argv[18]);
const ck::index_t conv_dilation_h = std::stoi(argv[19]);
const ck::index_t conv_dilation_w = std::stoi(argv[20]);
const ck::index_t in_left_pad_h = std::stoi(argv[21]);
const ck::index_t in_left_pad_w = std::stoi(argv[22]);
const ck::index_t in_right_pad_h = std::stoi(argv[23]);
const ck::index_t in_right_pad_w = std::stoi(argv[24]);
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
if(data_type == ConvDataType::F32_F32_F32 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv<2,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv<2,
ck::half_t,
ck::half_t,
ck::half_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
nrepeat,
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else
{
throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
}
return 1;
}
profiler/gemm_profiler.cpp 0 → 100644
View file @ e823d518
This diff is collapsed.
profiler/include/profile_conv.hpp 0 → 100644
View file @ e823d518
This diff is collapsed.
profiler/include/profile_gemm.hpp 0 → 100644
View file @ e823d518
This diff is collapsed.
profiler/profiler.cpp 0 → 100644
View file @ e823d518
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
int gemm_profiler(int, char*[]);
int conv_profiler(int, char*[]);
int main(int argc, char* argv[])
{
if(strcmp(argv[1], "gemm") == 0)
{
return gemm_profiler(argc, argv);
}
else if(strcmp(argv[1], "conv") == 0)
{
return conv_profiler(argc, argv);
}
else
{
printf("arg1: tensor operation (gemm=GEMM, conv=Convolution)\n");
return 0;
}
}
script/cmake-rocm.sh
View file @ e823d518
......@@ -8,11 +8,11 @@ MY_PROJECT_INSTALL=../install.dir
cmake \
-D CMAKE_INSTALL_PREFIX=${MY_PROJECT_INSTALL} \
-D HALF_INCLUDE_DIR="/root/workspace/external/half/include" \
-D BUILD_DEV=ON \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 -O3 --amdgpu-target=gfx908 -mllvm --amdgpu-spill-vgpr-to-agpr=0 -gline-tables-only -save-temps=$PWD" \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 -ftemplate-backtrace-limit=0 -mllvm --amdgpu-spill-vgpr-to-agpr=0 -gline-tables-only -save-temps=$PWD" \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_VERBOSE_MAKEFILE:BOOL=ON \
${MY_PROJECT_SOURCE}
script/conv_driver.sh 0 → 100755
View file @ e823d518
This diff is collapsed.
script/example_gemm_xdl.sh 0 → 100755
View file @ e823d518
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=1
make -j gemm_xdl
DRIVER="./example/gemm_xdl"
VERIFY=$1
INIT=$2
LOG=$3
REPEAT=$4
######### verify init log repeat M___ N___ K___ StrideA StrideB StrideC
#$DRIVER $VERIFY $INIT $LOG $REPEAT 960 1024 1024 1024 1024 1024
#$DRIVER $VERIFY $INIT $LOG $REPEAT 1024 1024 1024 1024 1024 1024
#$DRIVER $VERIFY $INIT $LOG $REPEAT 1920 2048 2048 2048 2048 2048
$DRIVER $VERIFY $INIT $LOG $REPEAT 3840 4096 4096 4096 4096 4096
#$DRIVER $VERIFY $INIT $LOG $REPEAT 7680 8192 8192 8192 8192 8192
script/gemm_driver.sh 0 → 100755
View file @ e823d518
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=0
make -j gemm_driver_offline
DRIVER="./host/driver_offline/gemm_driver_offline"
LAYOUT=$1
ALGO=$2
VERIFY=$3
INIT=$4
LOG=$5
REPEAT=$6
M01=$7
N01=$8
######### layout algo verify init log repeat M___ N___ K___ M01_ N01_
#$DRIVER $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 960 1024 1024 $M01 $N01
#$DRIVER $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 1024 1024 1024 $M01 $N01
#$DRIVER $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 1920 2048 2048 $M01 $N01
$DRIVER $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 3840 4096 4096 $M01 $N01
#$DRIVER $LAYOUT $ALGO $VERIFY $INIT $LOG $REPEAT 7680 8192 8192 $M01 $N01
script/run.sh deleted 100755 → 0
View file @ 6014185a
This diff is collapsed.
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