Skip to content

GitLab

Commit 8c4e33f1 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into v5r1_add

parents 5aed38d4 3737bb03
Show whitespace changes
Inline Side-by-side
Showing with 556 additions and 135 deletions
composable_kernel/include/utility/tuple_helper.hpp
View file @ 8c4e33f1
......@@ -6,22 +6,6 @@
namespace ck {
template <typename... Ts>
struct is_known_at_compile_time<Tuple<Ts...>>
{
__host__ __device__ static constexpr bool IsKnownAtCompileTime()
{
return container_reduce(
Tuple<Ts...>{},
[](auto x, bool r) {
return is_known_at_compile_time<remove_cvref_t<decltype(x)>>::value & r;
},
true);
}
static constexpr bool value = IsKnownAtCompileTime();
};
template <typename F, index_t N>
__host__ __device__ constexpr auto generate_tuple(F&& f, Number<N>)
{
......@@ -29,6 +13,13 @@ __host__ __device__ constexpr auto generate_tuple(F&& f, Number<N>)
typename arithmetic_sequence_gen<0, N, 1>::type{});
}
template <typename F, index_t N>
__host__ __device__ constexpr auto generate_tie(F&& f, Number<N>)
{
return unpack([&f](auto&&... xs) { return tie(f(xs)...); },
typename arithmetic_sequence_gen<0, N, 1>::type{});
}
namespace detail {
template <typename F, typename X, index_t... Is>
......
composable_kernel/include/utility/type.hpp
View file @ 8c4e33f1
......@@ -31,21 +31,6 @@ using remove_cvref_t = remove_cv_t<std::remove_reference_t<T>>;
template <typename T>
inline constexpr bool is_pointer_v = std::is_pointer<T>::value;
template <typename T>
struct is_known_at_compile_time;
template <>
struct is_known_at_compile_time<index_t>
{
static constexpr bool value = false;
};
template <typename T, T X>
struct is_known_at_compile_time<integral_constant<T, X>>
{
static constexpr bool value = true;
};
template <typename Y, typename X, typename enable_if<sizeof(X) == sizeof(Y), bool>::type = false>
__host__ __device__ constexpr Y as_type(X x)
{
......
device_operation/include/device_gemm_xdl.hpp
View file @ 8c4e33f1
......@@ -3,7 +3,6 @@
#include <iostream>
#include "device.hpp"
#include "gemm_common.hpp"
#include "device_base.hpp"
#include "device_gemm.hpp"
#include "common_header.hpp"
......
device_operation/include/gemm_common.hpp deleted 100644 → 0
View file @ 5aed38d4
#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
external/rocm/include/bfloat16_dev.hpp
View file @ 8c4e33f1
......@@ -31,7 +31,7 @@ extern "C" {
#endif
#ifdef __HIP_PLATFORM_HCC__
#define EXECUTION_SPECIFIER __device__
#define EXECUTION_SPECIFIER __device__ __host__
#else
#define EXECUTION_SPECIFIER
#endif // MIOPEN_BACKEND_HIP
......
host/driver_offline/include/device_convolution_backward_data_implicit_gemm_v4r1r2_xdlops_nhwc_kyxc_nhwk_1x1.hpp
View file @ 8c4e33f1
......@@ -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 GemmCThreadTransferDstScalarPerVector = 1;
#elif 1
#elif 0
// [M, N, K0, K1] = [256, 128, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
......@@ -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 GemmCThreadTransferDstScalarPerVector = 1;
#elif 0
#elif 1
// [M, N, K0, K1] = [128, 256, 4, 8], C = 128, for fp16
constexpr index_t BlockSize = 256;
......
host/driver_offline/src/conv_bwd_driver_offline.cpp
View file @ 8c4e33f1
......@@ -325,30 +325,30 @@ int main(int argc, char* argv[])
// no initialization
break;
case 1:
out.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
out.GenerateTensorValue(GeneratorTensor_1<out_data_t>{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1<in_data_t>{}, num_thread);
break;
case 2:
out.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
out.GenerateTensorValue(GeneratorTensor_1<out_data_t>{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2<in_data_t>{-5, 5}, num_thread);
break;
case 3:
out.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
out.GenerateTensorValue(GeneratorTensor_2<out_data_t>{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1<in_data_t>{}, num_thread);
break;
case 4:
out.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
out.GenerateTensorValue(GeneratorTensor_2<out_data_t>{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2<in_data_t>{-5, 5}, num_thread);
break;
case 5:
out.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
out.GenerateTensorValue(GeneratorTensor_3<out_data_t>{0.0, 1.0}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_3<in_data_t>{-0.5, 0.5}, num_thread);
break;
default:
out.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
out.GenerateTensorValue(GeneratorTensor_2<out_data_t>{1, 5}, num_thread);
auto gen_wei = [](auto... is) {
return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
return GeneratorTensor_2<in_data_t>{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
};
wei.GenerateTensorValue(gen_wei, num_thread);
}
......
host/driver_offline/src/conv_fwd_driver_offline.cpp
View file @ 8c4e33f1
......@@ -76,6 +76,13 @@ void host_convolution_forward(const Tensor<TIn>& in,
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])
{
if constexpr(is_same<TIn, ushort>::value)
{
v += bfloat16_to_float(in(n, c, hi, wi)) *
bfloat16_to_float(wei(k, c, y, x));
}
else
{
v += static_cast<const double>(in(n, c, hi, wi)) *
static_cast<const double>(wei(k, c, y, x));
......@@ -83,7 +90,16 @@ void host_convolution_forward(const Tensor<TIn>& in,
}
}
}
}
if constexpr(is_same<TOut, ushort>::value)
{
out(n, k, ho, wo) = float_to_bfloat16(v);
}
else
{
out(n, k, ho, wo) = v;
}
};
auto f_nhwc = [&](auto n, auto ho, auto wo, auto k) {
......@@ -98,6 +114,13 @@ void host_convolution_forward(const Tensor<TIn>& in,
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])
{
if constexpr(is_same<TIn, ushort>::value)
{
v += bfloat16_to_float(in(n, hi, wi, c)) *
bfloat16_to_float(wei(k, y, x, c));
}
else
{
v += static_cast<const double>(in(n, hi, wi, c)) *
static_cast<const double>(wei(k, y, x, c));
......@@ -105,7 +128,15 @@ void host_convolution_forward(const Tensor<TIn>& in,
}
}
}
}
if constexpr(is_same<TOut, ushort>::value)
{
out(n, ho, wo, k) = float_to_bfloat16(v);
}
else
{
out(n, ho, wo, k) = v;
}
};
if(layout == ConvTensorLayout::NCHW)
......@@ -223,10 +254,14 @@ int main(int argc, char* argv[])
using in_data_t = float;
using acc_data_t = float;
using out_data_t = float;
#elif 1
#elif 0
using in_data_t = half_t;
using acc_data_t = float;
using out_data_t = half_t;
#elif 1
using in_data_t = ushort;
using acc_data_t = float;
using out_data_t = ushort;
#elif 1
using in_data_t = int8_t;
using acc_data_t = int32_t;
......@@ -292,30 +327,30 @@ int main(int argc, char* argv[])
// no initialization
break;
case 1:
in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
in.GenerateTensorValue(GeneratorTensor_1<in_data_t>{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1<in_data_t>{}, num_thread);
break;
case 2:
in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
in.GenerateTensorValue(GeneratorTensor_1<in_data_t>{}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2<in_data_t>{-5, 5}, num_thread);
break;
case 3:
in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
in.GenerateTensorValue(GeneratorTensor_2<in_data_t>{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_1<in_data_t>{}, num_thread);
break;
case 4:
in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
in.GenerateTensorValue(GeneratorTensor_2<in_data_t>{-5, 5}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_2<in_data_t>{-5, 5}, num_thread);
break;
case 5:
in.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
in.GenerateTensorValue(GeneratorTensor_3<in_data_t>{0.0, 1.0}, num_thread);
wei.GenerateTensorValue(GeneratorTensor_3<in_data_t>{-0.5, 0.5}, num_thread);
break;
default:
in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
in.GenerateTensorValue(GeneratorTensor_2<in_data_t>{1, 5}, num_thread);
auto gen_wei = [](auto... is) {
return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
return GeneratorTensor_2<in_data_t>{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
};
wei.GenerateTensorValue(gen_wei, num_thread);
}
......
host/driver_offline/src/conv_wrw_driver_offline.cpp
View file @ 8c4e33f1
......@@ -297,30 +297,30 @@ int main(int argc, char* argv[])
// no initialization
break;
case 1:
in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
out.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
in.GenerateTensorValue(GeneratorTensor_1<in_data_t>{}, num_thread);
out.GenerateTensorValue(GeneratorTensor_1<out_data_t>{}, num_thread);
break;
case 2:
in.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
out.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
in.GenerateTensorValue(GeneratorTensor_1<in_data_t>{}, num_thread);
out.GenerateTensorValue(GeneratorTensor_2<out_data_t>{-5, 5}, num_thread);
break;
case 3:
in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
out.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
in.GenerateTensorValue(GeneratorTensor_2<in_data_t>{-5, 5}, num_thread);
out.GenerateTensorValue(GeneratorTensor_1<out_data_t>{}, num_thread);
break;
case 4:
in.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
out.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
in.GenerateTensorValue(GeneratorTensor_2<in_data_t>{-5, 5}, num_thread);
out.GenerateTensorValue(GeneratorTensor_2<out_data_t>{-5, 5}, num_thread);
break;
case 5:
in.GenerateTensorValue(GeneratorTensor_3<float>{-0.1, 0.1}, num_thread);
out.GenerateTensorValue(GeneratorTensor_3<float>{-0.1, 0.1}, num_thread);
in.GenerateTensorValue(GeneratorTensor_3<in_data_t>{-0.1, 0.1}, num_thread);
out.GenerateTensorValue(GeneratorTensor_3<out_data_t>{-0.1, 0.1}, num_thread);
break;
default:
in.GenerateTensorValue(GeneratorTensor_2{1, 5}, num_thread);
in.GenerateTensorValue(GeneratorTensor_2<in_data_t>{1, 5}, num_thread);
auto gen_out = [](auto... is) {
return GeneratorTensor_2{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
return GeneratorTensor_2<out_data_t>{1, 5}(is...) * GeneratorTensor_Checkboard{}(is...);
};
out.GenerateTensorValue(gen_out, num_thread);
}
......
host/driver_offline/src/gemm_driver_offline.cpp
View file @ 8c4e33f1
......@@ -10,7 +10,6 @@
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "gemm_common.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdlops_mk_kn_mn.hpp"
......@@ -31,6 +30,18 @@
#define USE_GEMM_XDL_KM_KN_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
{
Xdl_MK_KN_MN, // 0
......@@ -43,6 +54,161 @@ enum GemmAlgo
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[])
{
using namespace ck;
......@@ -73,10 +239,14 @@ int main(int argc, char* argv[])
using ab_data_t = float;
using acc_data_t = float;
using c_data_t = float;
#elif 1
#elif 0
using ab_data_t = half_t;
using acc_data_t = float;
using c_data_t = half_t;
#elif 1
using ab_data_t = ushort;
using acc_data_t = float;
using c_data_t = ushort;
#elif 1
using ab_data_t = int8_t;
using acc_data_t = int32_t;
......@@ -155,24 +325,24 @@ int main(int argc, char* argv[])
// no initialization
break;
case 1:
a.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
b.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
a.GenerateTensorValue(GeneratorTensor_1<ab_data_t>{}, num_thread);
b.GenerateTensorValue(GeneratorTensor_1<ab_data_t>{}, num_thread);
break;
case 2:
a.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
b.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
a.GenerateTensorValue(GeneratorTensor_1<ab_data_t>{}, num_thread);
b.GenerateTensorValue(GeneratorTensor_2<ab_data_t>{-5, 5}, num_thread);
break;
case 3:
a.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
b.GenerateTensorValue(GeneratorTensor_1{}, num_thread);
a.GenerateTensorValue(GeneratorTensor_2<ab_data_t>{-5, 5}, num_thread);
b.GenerateTensorValue(GeneratorTensor_1<ab_data_t>{}, num_thread);
break;
case 4:
a.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
b.GenerateTensorValue(GeneratorTensor_2{-5, 5}, num_thread);
a.GenerateTensorValue(GeneratorTensor_2<ab_data_t>{-5, 5}, num_thread);
b.GenerateTensorValue(GeneratorTensor_2<ab_data_t>{-5, 5}, num_thread);
break;
default:
a.GenerateTensorValue(GeneratorTensor_3<float>{0.0, 1.0}, num_thread);
b.GenerateTensorValue(GeneratorTensor_3<float>{-0.5, 0.5}, num_thread);
a.GenerateTensorValue(GeneratorTensor_3<ab_data_t>{0.0, 1.0}, num_thread);
b.GenerateTensorValue(GeneratorTensor_3<ab_data_t>{-0.5, 0.5}, num_thread);
}
#if USE_GEMM_XDL_MK_KN_MN
......
host/host_tensor/include/host_gemm.hpp
View file @ 8c4e33f1
#pragma once
#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,
template <>
void host_gemm<ushort, ushort, ushort>(const Tensor<ushort>& a,
const Tensor<ushort>& b,
Tensor<ushort>& c,
const GemmMatrixLayout layout)
{
if(layout == GemmMatrixLayout::MK_KN_MN)
......@@ -17,10 +16,10 @@ void host_gemm(const Tensor<AType>& a,
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(k, n));
v += bfloat16_to_float(a(m, k)) * bfloat16_to_float(b(k, n));
}
c(m, n) = v;
c(m, n) = float_to_bfloat16(v);
};
make_ParallelTensorFunctor(f_mk_kn_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
......@@ -35,10 +34,10 @@ void host_gemm(const Tensor<AType>& a,
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
v += bfloat16_to_float(a(m, k)) * bfloat16_to_float(b(n, k));
}
c(m, n) = v;
c(m, n) = float_to_bfloat16(v);
};
make_ParallelTensorFunctor(f_mk_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
......@@ -53,10 +52,10 @@ void host_gemm(const Tensor<AType>& a,
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(k, n));
v += bfloat16_to_float(a(k, m)) * bfloat16_to_float(b(k, n));
}
c(m, n) = v;
c(m, n) = float_to_bfloat16(v);
};
make_ParallelTensorFunctor(f_km_kn_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
......@@ -71,10 +70,10 @@ void host_gemm(const Tensor<AType>& a,
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(n, k));
v += bfloat16_to_float(a(k, m)) * bfloat16_to_float(b(n, k));
}
c(m, n) = v;
c(m, n) = float_to_bfloat16(v);
};
make_ParallelTensorFunctor(f_km_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
......@@ -89,10 +88,10 @@ void host_gemm(const Tensor<AType>& a,
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(k, n));
v += bfloat16_to_float(a(m, k)) * bfloat16_to_float(b(k, n));
}
c(n, m) = v;
c(n, m) = float_to_bfloat16(v);
};
make_ParallelTensorFunctor(f_mk_kn_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
......@@ -107,10 +106,10 @@ void host_gemm(const Tensor<AType>& a,
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
v += bfloat16_to_float(a(m, k)) * bfloat16_to_float(b(n, k));
}
c(n, m) = v;
c(n, m) = float_to_bfloat16(v);
};
make_ParallelTensorFunctor(f_mk_nk_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
......@@ -125,10 +124,10 @@ void host_gemm(const Tensor<AType>& a,
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(k, n));
v += bfloat16_to_float(a(k, m)) * bfloat16_to_float(b(k, n));
}
c(n, m) = v;
c(n, m) = float_to_bfloat16(v);
};
make_ParallelTensorFunctor(f_km_kn_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
......@@ -143,10 +142,10 @@ void host_gemm(const Tensor<AType>& a,
for(int k = 0; k < K; ++k)
{
v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(n, k));
v += bfloat16_to_float(a(k, m)) * bfloat16_to_float(b(n, k));
}
c(n, m) = v;
c(n, m) = float_to_bfloat16(v);
};
make_ParallelTensorFunctor(f_km_nk_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
......
host/host_tensor/include/host_tensor.hpp
View file @ 8c4e33f1
......@@ -333,4 +333,41 @@ void check_error(const Tensor<T>& ref, const Tensor<T>& result)
std::cout << "max_diff: " << max_diff << ", " << ref_value << ", " << result_value << std::endl;
}
float bf16_to_f32(ushort src_val)
{
typedef union
{
ushort x, y;
float f32;
} bf16_f32_t;
bf16_f32_t v;
v.x = 0;
v.y = src_val;
return v.f32;
}
template <>
void check_error<ushort>(const Tensor<ushort>& ref, const Tensor<ushort>& result)
{
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(bf16_to_f32(ref.mData[i]) - bf16_to_f32(result.mData[i]));
float diff = std::abs(bf16_to_f32(ref.mData[i]) - bf16_to_f32(result.mData[i]));
if(max_diff < diff)
{
max_diff = diff;
ref_value = bf16_to_f32(ref.mData[i]);
result_value = bf16_to_f32(result.mData[i]);
}
}
std::cout << "error: " << error << std::endl;
std::cout << "max_diff: " << max_diff << ", ref: " << ref_value << ", res: " << result_value
<< std::endl;
}
#endif
host/host_tensor/include/host_tensor_generator.hpp
View file @ 8c4e33f1
......@@ -4,6 +4,7 @@
#include <cmath>
#include "config.hpp"
template <typename T>
struct GeneratorTensor_1
{
int value = 1;
......@@ -15,6 +16,30 @@ struct GeneratorTensor_1
}
};
template <>
struct GeneratorTensor_1<ushort>
{
float value = 1.0;
template <typename... Is>
ushort operator()(Is...)
{
return float_to_bfloat16(value);
}
};
template <>
struct GeneratorTensor_1<int8_t>
{
int8_t value = 1;
template <typename... Is>
int8_t operator()(Is...)
{
return value;
}
};
struct GeneratorTensor_0
{
int value = 0;
......@@ -26,6 +51,7 @@ struct GeneratorTensor_0
}
};
template <typename T>
struct GeneratorTensor_2
{
int min_value = 0;
......@@ -38,6 +64,33 @@ struct GeneratorTensor_2
}
};
template <>
struct GeneratorTensor_2<ushort>
{
int min_value = 0;
int max_value = 1;
template <typename... Is>
ushort operator()(Is...)
{
float tmp = (std::rand() % (max_value - min_value)) + min_value;
return float_to_bfloat16(tmp);
}
};
template <>
struct GeneratorTensor_2<int8_t>
{
int min_value = 0;
int max_value = 1;
template <typename... Is>
int8_t operator()(Is...)
{
return (std::rand() % (max_value - min_value)) + min_value;
}
};
template <typename T>
struct GeneratorTensor_3
{
......@@ -53,6 +106,39 @@ struct GeneratorTensor_3
}
};
template <>
struct GeneratorTensor_3<ushort>
{
float min_value = 0;
float max_value = 1;
template <typename... Is>
ushort operator()(Is...)
{
float tmp = float(std::rand()) / float(RAND_MAX);
float fp32_tmp = min_value + tmp * (max_value - min_value);
return float_to_bfloat16(fp32_tmp);
}
};
template <>
struct GeneratorTensor_3<int8_t>
{
float min_value = 0;
float max_value = 1;
template <typename... Is>
int8_t operator()(Is...)
{
int8_t min_tmp = static_cast<int8_t>(min_value);
int8_t max_tmp = static_cast<int8_t>(max_value);
return (std::rand() % (max_tmp - min_tmp)) + min_tmp;
}
};
struct GeneratorTensor_Checkboard
{
template <typename... Ts>
......
profiler/gemm_profiler.cpp
View file @ 8c4e33f1
......@@ -9,13 +9,30 @@
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "gemm_common.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_gemm_xdl.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[])
{
if(argc != 14)
......
script/profile_conv.sh 0 → 100755
View file @ 8c4e33f1
#!/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 @ 8c4e33f1
#!/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