Skip to content

GitLab

"test/vscode:/vscode.git/clone" did not exist on "bac32ec1f777e4ed65a363e1c93ccd7f4beb990b"
Unverified Commit b491ebf3 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

FP16 data in-register transpose (#41) 

* start fixing 16bit data packing

* adding StaticTensor

* adding StaticTensor

* adding StaticTensor

* add missing constexpr

* adding static tensor

* adding static tensor

* adding transpose

* add inline asm for transpose 2x2 of half_t

* add general transpose_vectors(), but have unnecessary register initialization using v_mov

* fix unnecessary register initialization in transpose_vector by using more pass-by-reference

* add hardcoded logic for NHWC wrw

* improve asm for v_pack

* make ThreadwiseTensorSliceTransfer_v3r2 support any tensor

* tweak

* reorganize file
parent e823d518
Show whitespace changes
Inline Side-by-side
Showing with 311 additions and 183 deletions
device_operation/include/gemm_common.hpp deleted 100644 → 0
View file @ e823d518
#ifndef GEMM_COMMON_HPP
#define GEMM_COMMON_HPP
enum GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
#endif
host/driver_offline/include/device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk_1x1.hpp
View file @ b491ebf3
...@@ -104,7 +104,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk ...@@ -104,7 +104,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4; constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 4;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1; constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 1 #elif 0
// [M, N, K0, K1] = [256, 128, 4, 8], C = 128, for fp16 // [M, N, K0, K1] = [256, 128, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256; constexpr index_t BlockSize = 256;
...@@ -132,7 +132,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk ...@@ -132,7 +132,7 @@ void device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk
constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8; constexpr index_t GemmBBlockTransferDstScalarPerVector_GemmK1 = 8;
constexpr index_t GemmCThreadTransferDstScalarPerVector = 1; constexpr index_t GemmCThreadTransferDstScalarPerVector = 1;
#elif 0 #elif 1
// [M, N, K0, K1] = [128, 256, 4, 8], C = 128, for fp16 // [M, N, K0, K1] = [128, 256, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256; constexpr index_t BlockSize = 256;
......
host/driver_offline/src/gemm_driver_offline.cpp
View file @ b491ebf3
...@@ -10,7 +10,6 @@ ...@@ -10,7 +10,6 @@
#include "device.hpp" #include "device.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
#include "host_tensor_generator.hpp" #include "host_tensor_generator.hpp"
#include "gemm_common.hpp"
#include "host_gemm.hpp" #include "host_gemm.hpp"
#include "device_tensor.hpp" #include "device_tensor.hpp"
#include "device_gemm_xdlops_mk_kn_mn.hpp" #include "device_gemm_xdlops_mk_kn_mn.hpp"
...@@ -31,6 +30,18 @@ ...@@ -31,6 +30,18 @@
#define USE_GEMM_XDL_KM_KN_NM 0 #define USE_GEMM_XDL_KM_KN_NM 0
#define USE_GEMM_XDL_KM_NK_NM 0 #define USE_GEMM_XDL_KM_NK_NM 0
enum GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM // 7
};
enum GemmAlgo enum GemmAlgo
{ {
Xdl_MK_KN_MN, // 0 Xdl_MK_KN_MN, // 0
...@@ -43,6 +54,161 @@ enum GemmAlgo ...@@ -43,6 +54,161 @@ enum GemmAlgo
Xdl_KM_NK_NM, // 7 Xdl_KM_NK_NM, // 7
}; };
template <typename AType, typename BType, typename CType>
void host_gemm(const Tensor<AType>& a,
const Tensor<BType>& b,
Tensor<CType>& c,
const GemmMatrixLayout layout)
{
if(layout == GemmMatrixLayout::MK_KN_MN)
{
auto f_mk_kn_mn = [&](auto m, auto n) {
const int K = a.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(k, n));
}
c(m, n) = v;
};
make_ParallelTensorFunctor(f_mk_kn_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::MK_NK_MN)
{
auto f_mk_nk_mn = [&](auto m, auto n) {
const int K = a.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
}
c(m, n) = v;
};
make_ParallelTensorFunctor(f_mk_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::KM_KN_MN)
{
auto f_km_kn_mn = [&](auto m, auto n) {
const int K = a.mDesc.GetLengths()[0];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(k, n));
}
c(m, n) = v;
};
make_ParallelTensorFunctor(f_km_kn_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::KM_NK_MN)
{
auto f_km_nk_mn = [&](auto m, auto n) {
const int K = a.mDesc.GetLengths()[0];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(n, k));
}
c(m, n) = v;
};
make_ParallelTensorFunctor(f_km_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::MK_KN_NM)
{
auto f_mk_kn_nm = [&](auto n, auto m) {
const int K = a.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(k, n));
}
c(n, m) = v;
};
make_ParallelTensorFunctor(f_mk_kn_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::MK_NK_NM)
{
auto f_mk_nk_nm = [&](auto n, auto m) {
const int K = a.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
}
c(n, m) = v;
};
make_ParallelTensorFunctor(f_mk_nk_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::KM_KN_NM)
{
auto f_km_kn_nm = [&](auto n, auto m) {
const int K = a.mDesc.GetLengths()[0];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(k, n));
}
c(n, m) = v;
};
make_ParallelTensorFunctor(f_km_kn_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::KM_NK_NM)
{
auto f_km_nk_nm = [&](auto n, auto m) {
const int K = a.mDesc.GetLengths()[0];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(n, k));
}
c(n, m) = v;
};
make_ParallelTensorFunctor(f_km_nk_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else
{
throw std::runtime_error("wrong! not supported layout");
}
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
using namespace ck; using namespace ck;
......
host/host_tensor/include/host_gemm.hpp
View file @ b491ebf3
#pragma once #pragma once
#include "host_tensor.hpp" #include "host_tensor.hpp"
#include "gemm_common.hpp"
template <typename AType, typename BType, typename CType>
void host_gemm(const Tensor<AType>& a,
const Tensor<BType>& b,
Tensor<CType>& c,
const GemmMatrixLayout layout)
{
if(layout == GemmMatrixLayout::MK_KN_MN)
{
auto f_mk_kn_mn = [&](auto m, auto n) {
const int K = a.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(k, n));
}
c(m, n) = v;
};
make_ParallelTensorFunctor(f_mk_kn_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::MK_NK_MN)
{
auto f_mk_nk_mn = [&](auto m, auto n) {
const int K = a.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
}
c(m, n) = v;
};
make_ParallelTensorFunctor(f_mk_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::KM_KN_MN)
{
auto f_km_kn_mn = [&](auto m, auto n) {
const int K = a.mDesc.GetLengths()[0];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(k, n));
}
c(m, n) = v;
};
make_ParallelTensorFunctor(f_km_kn_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::KM_NK_MN)
{
auto f_km_nk_mn = [&](auto m, auto n) {
const int K = a.mDesc.GetLengths()[0];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(n, k));
}
c(m, n) = v;
};
make_ParallelTensorFunctor(f_km_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::MK_KN_NM)
{
auto f_mk_kn_nm = [&](auto n, auto m) {
const int K = a.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(k, n));
}
c(n, m) = v;
};
make_ParallelTensorFunctor(f_mk_kn_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::MK_NK_NM)
{
auto f_mk_nk_nm = [&](auto n, auto m) {
const int K = a.mDesc.GetLengths()[1];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
}
c(n, m) = v;
};
make_ParallelTensorFunctor(f_mk_nk_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::KM_KN_NM)
{
auto f_km_kn_nm = [&](auto n, auto m) {
const int K = a.mDesc.GetLengths()[0];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(k, n));
}
c(n, m) = v;
};
make_ParallelTensorFunctor(f_km_kn_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else if(layout == GemmMatrixLayout::KM_NK_NM)
{
auto f_km_nk_nm = [&](auto n, auto m) {
const int K = a.mDesc.GetLengths()[0];
double v = 0;
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(n, k));
}
c(n, m) = v;
};
make_ParallelTensorFunctor(f_km_nk_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
std::thread::hardware_concurrency());
}
else
{
throw std::runtime_error("wrong! not supported layout");
}
}
template <typename AType, typename BType, typename CType> template <typename AType, typename BType, typename CType>
void host_gemm_mk_kn_mn(const Tensor<AType>& a_m_k, void host_gemm_mk_kn_mn(const Tensor<AType>& a_m_k,
......
profiler/gemm_profiler.cpp
View file @ b491ebf3
...@@ -9,13 +9,30 @@ ...@@ -9,13 +9,30 @@
#include "device.hpp" #include "device.hpp"
#include "host_tensor.hpp" #include "host_tensor.hpp"
#include "host_tensor_generator.hpp" #include "host_tensor_generator.hpp"
#include "gemm_common.hpp"
#include "host_gemm.hpp" #include "host_gemm.hpp"
#include "device_tensor.hpp" #include "device_tensor.hpp"
#include "device_base.hpp" #include "device_base.hpp"
#include "device_gemm_xdl.hpp" #include "device_gemm_xdl.hpp"
#include "profile_gemm.hpp" #include "profile_gemm.hpp"
enum GemmMatrixLayout
{
MK_KN_MN, // 0
MK_NK_MN, // 1
KM_KN_MN, // 2
KM_NK_MN, // 3
MK_KN_NM, // 4
MK_NK_NM, // 5
KM_KN_NM, // 6
KM_NK_NM, // 7
};
enum GemmDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
};
int gemm_profiler(int argc, char* argv[]) int gemm_profiler(int argc, char* argv[])
{ {
if(argc != 14) if(argc != 14)
......
script/profile_conv.sh 0 → 100755
View file @ b491ebf3
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=0
make -j ckProfiler
DRIVER="./profiler/ckProfiler"
OP=$1
DATATYPE=$2
IN_LAYOUT=$3
WEI_LAYOUT=$4
OUT_LAYOUT=$5
VERIFY=$6
INIT=$7
LOG=$8
REPEAT=$9
# test
######## op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads Desired_grid_size__
$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 128 256 192 3 3 71 71 2 2 1 1 1 1 1 1 $DESIRED_GRID_SIZE
#N=${10}
# Resnet50
######## op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads Desired_grid_size__
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 1024 1 1 14 14 2 2 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 1024 1 1 14 14 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 58 58 2 2 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 30 30 2 2 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 256 1 1 56 56 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 256 1 1 56 56 2 2 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 16 16 2 2 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 512 1 1 28 28 2 2 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 512 1 1 28 28 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 1 1 56 56 1 1 1 1 0 0 0 0 $DESIRED_GRID_SIZE
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1 $DESIRED_GRID_SIZE
# SSD
######## op datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads Desired_grid_size__
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 64 3 7 7 300 300 2 2 1 1 3 3 3 3
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 64 64 3 3 75 75 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 64 64 3 3 75 75 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 64 64 3 3 75 75 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 64 64 3 3 75 75 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 64 64 3 3 75 75 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 64 64 3 3 75 75 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 64 1 1 75 75 2 2 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 64 3 3 75 75 2 2 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 128 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 128 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 128 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 128 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 128 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 128 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 128 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 128 1 1 38 38 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 128 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 256 1 1 38 38 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 512 256 3 3 38 38 2 2 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 512 1 1 19 19 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 512 256 3 3 19 19 2 2 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 512 1 1 10 10 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 128 3 3 10 10 2 2 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 256 1 1 5 5 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 128 3 3 5 5 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 128 256 1 1 3 3 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 256 128 3 3 3 3 1 1 1 1 0 0 0 0
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 340 256 3 3 38 38 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 510 512 3 3 19 19 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 510 512 3 3 10 10 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 510 256 3 3 5 5 1 1 1 1 1 1 1 1
#$DRIVER $OP $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT 120 340 256 3 3 3 3 1 1 1 1 1 1 1 1
script/profile_gemm.sh 0 → 100755
View file @ b491ebf3
#!/bin/bash
## GPU visibility
export HIP_VISIBLE_DEVICES=0
make -j ckProfiler
DRIVER="./profiler/ckProfiler"
OP=$1
DATATYPE=$2
LAYOUT=$3
VERIFY=$4
INIT=$5
LOG=$6
REPEAT=$7
######## op datatype layout verify init log repeat M___ N___ K___ StrideA StrideB StrideC
#$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $REPEAT 256 256 256 256 256 256
#$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $REPEAT 960 1024 1024 1024 1024 1024
#$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $REPEAT 1024 1024 1024 1024 1024 1024
#$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $REPEAT 1920 2048 2048 2048 2048 2048
$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $REPEAT 3840 4096 4096 4096 4096 4096
#$DRIVER $OP $DATATYPE $LAYOUT $VERIFY $INIT $LOG $REPEAT 7680 8192 8192 8192 8192 8192
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